TORUS 2 -- toward an open resource using services cloud computing for environmental data 9781119720539, 1119720532, 9781119720553, 1119720559

Citation preview

TORUS 2 – Toward an Open Resource Using Services

TORUS 2 – Toward an Open Resource Using Services Cloud Computing for Environmental Data

Edited by

Dominique Laffly

First published 2020 in Great Britain and the United States by ISTE Ltd and John Wiley & Sons, Inc.

Apart from any fair dealing for the purposes of research or private study, or criticism or review, as permitted under the Copyright, Designs and Patents Act 1988, this publication may only be reproduced, stored or transmitted, in any form or by any means, with the prior permission in writing of the publishers, or in the case of reprographic reproduction in accordance with the terms and licenses issued by the CLA. Enquiries concerning reproduction outside these terms should be sent to the publishers at the undermentioned address: ISTE Ltd 27-37 St George’s Road London SW19 4EU UK

John Wiley & Sons, Inc. 111 River Street Hoboken, NJ 07030 USA

www.iste.co.uk

www.wiley.com

© ISTE Ltd 2020 The rights of Dominique Laffly to be identified as the author of this work have been asserted by him in accordance with the Copyright, Designs and Patents Act 1988. Library of Congress Control Number: 2019956836 British Library Cataloguing-in-Publication Data A CIP record for this book is available from the British Library ISBN 978-1-78630-600-5

Contents

Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

xi

Part 1. Earth Science Remote Sensing . . . . . . . . . . . . . . . . . . . . . . .

xvii

Introduction to Part 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Dominique LAFFLY

xix

Chapter 1. A Brief History of Remote Sensing. . . . . . . . . . . . . . . . . . . Dominique LAFFLY

1

1.1. History . . . . . . . . . . . . . . . . . . . . . . 1.2. Fields of application . . . . . . . . . . . . . . . 1.3. Orbits, launchers and platforms . . . . . . . . 1.4. The acquired data are digital images . . . . . . 1.5. So what is remote sensing? Some definitions . 1.6. Notes . . . . . . . . . . . . . . . . . . . . . . . 1.7. References . . . . . . . . . . . . . . . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

1 8 10 12 14 19 21

Chapter 2. Physics of RS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Luca TOMASSETTI

23

2.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 2.2. Remote sensing . . . . . . . . . . . . . . . . . . . . . . 2.3. Fundamental properties of electromagnetic waves . . . 2.3.1. Wave equation and solution . . . . . . . . . . . . . 2.3.2. Quantum properties of electromagnetic radiation . 2.3.3. Polarization, coherence, group and phase velocity, the Doppler effect . . . . . . . . . . . . . . . . . . . . . .

. . . . . . .

. . . . .

. . . . . . .

. . . . .

. . . . . . .

. . . . .

. . . . . . .

. . . . .

. . . . . . .

. . . . .

. . . . . . .

. . . . .

. . . . . . .

. . . . .

. . . . . . .

. . . . .

. . . . . . .

. . . . .

. . . . . . .

. . . . .

. . . . . . .

. . . . .

. . . . . . .

. . . . .

. . . . .

23 23 29 29 30

. . . . . . . . . . . . .

31

vi

TORUS 2 – Toward an Open Resource Using Services

2.4. Radiation quantities . . . . . . . . . . . . . . . . . . . 2.4.1. Spectral quantities . . . . . . . . . . . . . . . . . 2.4.2. Luminous quantities . . . . . . . . . . . . . . . . 2.5. Generation of electromagnetic waves . . . . . . . . . 2.6. Detection of electromagnetic waves . . . . . . . . . . 2.7. Interaction of electromagnetic waves with matter . . 2.7.1. Overview . . . . . . . . . . . . . . . . . . . . . . 2.7.2. Interaction mechanisms . . . . . . . . . . . . . . 2.8. Solid surfaces sensing in the visible and near infrared 2.8.1. Wave-surface interaction mechanisms . . . . . . 2.9. Radiometric and geometric resolutions . . . . . . . . 2.10. References . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . .

31 33 34 34 37 38 38 39 41 43 45 46

Chapter 3. Image Quality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Dominique LAFFLY

47

3.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2. Image quality – geometry . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.1. Whiskbroom concept . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.2. Pushbroom concept . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.3. Full frame concept . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.4. Optical geometric distortions . . . . . . . . . . . . . . . . . . . . . 3.2.5. Relief distortions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.6. Inverse location model . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.7. Direct location model . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.8. Root Mean Square (RMS) validation . . . . . . . . . . . . . . . . . 3.2.9. Resampling methods . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.10. Image geometric quality to assume geographical space continuity 3.3. Image quality – radiometry . . . . . . . . . . . . . . . . . . . . . . . . . 3.3.1. Radiometric model of the instrument . . . . . . . . . . . . . . . . . 3.3.2. Radiometric equalization and calibration . . . . . . . . . . . . . . . 3.3.3. Radiometric signal noise reduction (SNR) . . . . . . . . . . . . . 3.3.4. Radiometric physical value . . . . . . . . . . . . . . . . . . . . . . 3.3.5. Image quality – resolution . . . . . . . . . . . . . . . . . . . . . . . 3.4. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.5. Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.6. References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . .

47 54 57 60 62 64 66 67 69 72 73 75 76 78 79 81 82 84 91 91 91

Chapter 4. Remote Sensing Products . . . . . . . . . . . . . . . . . . . . . . . . Van Ha PHAM, Viet Hung LUU, Anh PHAN, Dominique LAFFLY, Quang Hung BUI and Thi Nhat Thanh NGUYEN

95

4.1. Atmospheric observation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1.1. Introduction to common atmospheric gases and particles . . . . . . . . . .

95 95

Contents

4.1.2. Introduction to meteorological parameters 4.1.3. Atmospheric observation from satellite . . 4.2. Land observation. . . . . . . . . . . . . . . . . 4.2.1. Introduction . . . . . . . . . . . . . . . . . 4.2.2. Land cover/land use classification system 4.2.3. Legend . . . . . . . . . . . . . . . . . . . . 4.2.4. Data . . . . . . . . . . . . . . . . . . . . . 4.2.5. Methodology . . . . . . . . . . . . . . . . 4.2.6. Global land cover datasets . . . . . . . . . 4.3. Conclusion . . . . . . . . . . . . . . . . . . . . 4.4. References . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . .

103 107 128 128 129 134 134 137 154 158 158

Chapter 5. Image Processing in Spark . . . . . . . . . . . . . . . . . . . . . . . . Yannick LE NIR, Florent DEVIN, Thomas BALDAQUIN, Pierre MESLER LAZENNEC, Ji Young JUNG, Se-Eun KIM, Hyeyoung KWOON, Lennart NILSEN, Yoo Kyung LEE and Dominique LAFFLY

163

5.1. Introduction . . . . . . . . 5.2. Prediction map generation 5.2.1. Spark . . . . . . . . . . 5.2.2. Implementation . . . . 5.2.3. Naive method . . . . . 5.2.4. Advanced method . . . 5.3. Conclusion . . . . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . . . . . .

. . . . . . .

. . . . . . . . . . .

. . . . . . .

. . . . . . . . . . .

. . . . . . .

. . . . . . . . . . .

. . . . . . . . . . . . .

. . . . . . .

. . . . . . . . . . .

. . . . . . . . . . . . .

. . . . . . .

. . . . . . . . . . .

. . . . . . . . . . . . .

. . . . . . .

. . . . . . . . . . .

. . . . . . . . . . . . .

. . . . . . .

. . . . . . . . . . .

. . . . . . . . . . . . .

. . . . . . .

. . . . . . . . . . .

. . . . . . . . . . . . .

. . . . . . .

. . . . . . . . . . .

. . . . . . . . . . . . .

. . . . . . .

. . . . . . . . . . .

. . . . . . . . . . . . .

. . . . . . .

. . . . . . . . . . .

. . . . . . . . . . . . .

. . . . . . .

. . . . . . . . . . .

173

. . . . . . . . . . . . .

. . . . . . .

. . . . . . . . . . .

Chapter 6. Satellite Image Processing using Spark on the HUPI Platform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Vincent MORENO and Minh Tu NGUYEN . . . . . . . . . . . . .

. . . . . . .

. . . . . . . . . . .

163 164 164 165 167 168 171

. . . . . . . . . . . . .

. . . . . . .

. . . . . . . . . . .

. . . . . . .

6.1. Introduction . . . . . . . . . . . . . . . . . . . . . . 6.2. Presentation of GeoTrellis . . . . . . . . . . . . . . 6.3. Using GeoTrellis in Hupi-Notebook . . . . . . . . . 6.3.1. Some core concepts of GeoTrellis . . . . . . . . 6.3.2. Computation of NDVI . . . . . . . . . . . . . . 6.3.3. Compare two NDVI . . . . . . . . . . . . . . . 6.3.4. Descriptive statistics of NDVI per Tile . . . . . 6.3.5. K-means . . . . . . . . . . . . . . . . . . . . . . 6.4. Workflows in HDFS: automatize image processing 6.4.1. Create a jar . . . . . . . . . . . . . . . . . . . . 6.4.2. Monitor the Spark jobs . . . . . . . . . . . . . . 6.4.3. Tune performance of the Spark job . . . . . . . 6.4.4. Create a workflow in Hupi-Studio . . . . . . . .

. . . . . . .

. . . . . . . . . . .

vii

. . . . . . . . . . . . .

. . . . . . .

. . . . . . . . . . . . .

. . . . . . . . . . . . .

173 174 174 177 177 178 178 179 181 181 182 183 184

viii

TORUS 2 – Toward an Open Resource Using Services

6.5. Visualizations in Hupi-Front . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.6. Cloud service . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.7. Development . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

186 188 189

Chapter 7. Remote Sensing Case Studies . . . . . . . . . . . . . . . . . . . . . Van Ha PHAM, Thi Nhat Thanh NGUYEN and Dominique LAFFLY

191

7.1. Satellite AOD validation using R. . . . . . . . . 7.1.1. Introduction . . . . . . . . . . . . . . . . . . 7.1.2. Datasets . . . . . . . . . . . . . . . . . . . . 7.1.3. Validation methodology . . . . . . . . . . . 7.1.4. Experiments and results . . . . . . . . . . . 7.1.5. Conclusion . . . . . . . . . . . . . . . . . . . 7.2. Georeferencing satellite images . . . . . . . . . 7.2.1. Introduction . . . . . . . . . . . . . . . . . . 7.2.2. Georeferencing methods . . . . . . . . . . . 7.2.3. Datasets and methodology . . . . . . . . . . 7.2.4. Results and discussion . . . . . . . . . . . . 7.3. Conclusion . . . . . . . . . . . . . . . . . . . . . 7.4. Appendix: R source code of validation process . 7.5. References . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . . . . . . .

191 191 192 195 198 204 204 204 205 207 210 216 217 222

Conclusion to Part 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Dominique LAFFLY

225

Part 2. GIS Application and Geospatial Data Infrastructure . . . . . . . . . .

227

Chapter 8. Overview of GIS Application . . . . . . . . . . . . . . . . . . . . . . . Quang Huy MAN

229

8.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 8.2. Enterprise GIS for environmental management . . . . . . 8.3. GIS and decision-making in planning and management . 8.3.1. Data quality and control . . . . . . . . . . . . . . . . 8.3.2. Decision support systems (DSS) . . . . . . . . . . . 8.3.3. Integrating GIS with the DSS . . . . . . . . . . . . . 8.4. GIS for water-quality management. . . . . . . . . . . . . 8.5. GIS for land-use planning . . . . . . . . . . . . . . . . . . 8.6. Application of the technology in LUP and management . 8.6.1. Computers and software programs applied to LUP and management . . . . . . . . . . . . . . . . . . . 8.6.2. Application of GIS analysis and MCE in land-use planning and management . . . . . . . . . . . . . . . . . . . 8.7. References . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . .

. . . . . . . . .

229 230 232 233 233 234 235 236 240

. . . . . . . . . . . .

241

. . . . . . . . . . . . . . . . . . . . . . . .

242 243

Contents

Chapter 9. Spatial Data Infrastructure . . . . . . . . . . . . . . . . . . . . . . . . Quang Hung BUI, Quang Thang LUU, Duc Van HA, Tuan Dung PHAM, Sanya PRASEUTH and Dominique LAFFLY 9.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.2. Spatial data infrastructure . . . . . . . . . . . . . . . . . . . . . . 9.3. Components of spatial data infrastructure . . . . . . . . . . . . . 9.4. Open standards for spatial data infrastructure . . . . . . . . . . . 9.4.1. Open geospatial consortium (OGC) . . . . . . . . . . . . . 9.4.2. OGC’s open standards . . . . . . . . . . . . . . . . . . . . . 9.4.3. Usage of OGC’s open standards in SDI. . . . . . . . . . . . 9.5. Server architecture models for the National Spatial Data Infrastructure and Geospatial One-Stop (GOS) portal . . . . . . . . . 9.5.1. GOS portal architecture . . . . . . . . . . . . . . . . . . . . 9.5.2. Standards for GOS portal architecture . . . . . . . . . . . . 9.5.3. Taxonomy of geospatial server architecture . . . . . . . . . 9.5.4. Three reference architectures for server architecture model. 9.6. References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

ix

247

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

247 247 249 251 251 252 255

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

256 256 257 257 258 260

List of Authors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

263

Index. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

265

Summaries of other volumes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

267

Preface Why TORUS? Toward an Open Resource Using Services, or How to Bring Environmental Science Closer to Cloud Computing

Geography, Ecology, Urbanism, Geology and Climatology – in short, all environmental disciplines are inspired by the great paradigms of Science: they were first descriptive before evolving toward systemic and complexity. The methods followed the same evolution, from the inductive of the initial observations one approached the deductive of models of prediction based on learning. For example, the Bayesian is the preferred approach in this book (see Volume 1, Chapter 5), but random trees, neural networks, classifications and data reductions could all be developed. In the end, all the methods of artificial intelligence (IA) are ubiquitous today in the era of Big Data. We are not unaware, however, that, forged in Dartmouth in 1956 by John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon, the term artificial intelligence is, after a long period of neglect at the heart of the future issues of the exploitation of massive data (just like the functional and logical languages that accompanied the theory: LISP, 1958, PROLOG, 1977 and SCALA, today – see Chapter 8). All the environmental disciplines are confronted with this reality of massive data, with the rule of the 3+2Vs: Volume, Speed (from the French translation, “Vitesse”), Variety, Veracity, Value. Every five days – or even less – and only for the optical remote sensing data of the Sentinel 2a and 2b satellites, do we have a complete coverage of the Earth at a spatial resolution of 10 m for a dozen wavelengths. How do we integrate all this, how do we rethink the environmental disciplines where we must now consider at the pixel scale (10 m) an overall analysis of 510 million km2 or more than 5 billion pixels of which there are 1.53 billion for land only? And more important in fact, how do we validate automatic processes and accuracy of results?

xii

TORUS 2 – Toward an Open Resource Using Services

Dartmouth Summer Research Project on Artificial Intelligence, 1956

Figure P.1. At the beginnig of AI, Dartmouth Summer Research Project, 1956. Source: http://www.oezratty.net/wordpress/2017/semantique-intelligence-artificielle/

Including social network data, Internet of Things (IoT) and archive data, for many topics such as Smart Cities, it is not surprising that environmental disciplines are interested in cloud computing. Before understanding the technique (why this shape, why a cloud?), it would seem that to represent a node of connection of a network, we have, as of the last 50 years, drawn a potatoid freehand, which, drawn took the form of a cloud. Figure P.2 gives a perfect illustration on the left, while on the right we see that the cloud is now the norm (screenshot offered by a search engine in relation to the keywords: Internet and network). What is cloud computing? Let us remember that, even before the term was dedicated to it, cloud computing was based on networks (see Chapter 4), the Internet and this is: “since the 50s when users accessed, from their terminals, applications running on central systems” (Wikipedia). The cloud, as we understand it today, has evolved considerably since the 2000s; it consists of the mutualization of remote computing resources to store data and use services dynamically – to understand software – dedicated via browser interfaces.

Preface

xiii

Figure P.2. From freehand potatoid to the cloud icon. The first figure is a schematic illustration of a distributed SFPS switch. For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

This answers the needs of the environmental sciences overwhelmed by the massive data flows: everything is stored in the cloud, everything is processed in the cloud, even the results expected by the end-users recover them according to their needs. It is no wonder that, one after the other, Google and NASA offered in December 2016 – mid-term of TORUS! – cloud-based solutions for the management and processing of satellite data: Google Earth Engine and NASA Earth Exchange. But how do you do it? Why is it preferable – or not – for HPC (High Performance Computing) and GRIDS? How do we evaluate “Cloud & High Scalability Computing” versus “Grid & High-Performance Computing”? What are the costs? How do you transfer the applications commonly used by environmental science to the cloud? What is the added value for environmental sciences? In short, how does it work? All these questions and more are at the heart of the TORUS program developed to learn from each other, understand each other and communicate with a common language mastered: geoscience, computer science and information science; and the geosciences between them; computer science and information sciences. TORUS is not a research program. It is an action that aims to bring together too (often) remote scientific communities, in order to bridge the gap that now separates contemporary

xiv

TORUS 2 – Toward an Open Resource Using Services

computing from environmental disciplines for the most part. One evolving at speeds that cannot be followed by others, one that is greedy for data that others provide, one that can offer technical solutions to scientific questioning that is being developed by others and so on. TORUS is also the result of multiple scientific collaborations initiated in 2008–2010: between the geographer and the computer scientist, between France and Vietnam with an increasing diversity of specialties involved (e.g. remote sensing and image processing, mathematics and statistics, optimization and modeling, erosion and geochemistry, temporal dynamics and social surveys) all within various scientific and university structures (universities, engineering schools, research institutes – IRD, SFRI and IAE Vietnam, central administrations: the Midi-Pyrénées region and Son La district, France–Vietnam partnership) and between research and higher education through national and international PhDs. Naturally, I would like to say, the Erasmus+ capacity building program of the European Union appeared to be a solution adapted to our project: “The objectives of the Capacity Building projects are: to support the modernization, accessibility and internationalization of higher education in partner countries; improve the quality, relevance and governance of higher education in partner countries; strengthen the capacity of higher education institutions in partner countries and in the EU, in terms of international cooperation and the process of permanent modernization in particular; and to help them open up to society at large and to the world of work in order to reinforce the interdisciplinary and transdisciplinary nature of higher education, to improve the employability of university graduates, to give the European higher education more visibility and attractiveness in the world, foster the reciprocal development of human resources, promote a better understanding between the peoples and cultures of the EU and partner countries.”1 In 2015, TORUS – funded to the tune of 1 million euros for three years – was part of the projects selected in a pool of more than 575 applications and only 120 retentions. The partnership brings together (Figure P.3) the University of Toulouse 2 Jean Jaurès (coordinator – FR), the International School of Information Processing Sciences (EISTI – FR), the University of Ferrara in Italy, the Vrije University of Brussels, the National University from Vietnam to Hanoi, Nong Lam University in Ho Chi Minh City and two Thai institutions: Pathumthani’s Asian Institute of Technology (AIT) and Walaikak University in Nakhon Si Thammarat. 1 http://www.agence-erasmus.fr/page/developpement-des-capacites.

Preface

xv

Figure P.3. The heart of TORUS, partnership between Asia and Europe. For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

With an equal share between Europe and Asia, 30 researchers, teachersresearchers and engineers are involved in learning from each other during these three years, which will be punctuated by eight workshops between France, Vietnam, Italy, Thailand and Belgium. Finally, after the installation of the two servers in Asia (Asian Institute of Technology – Thailand; and Vietnam National University Hanoi – Vietnam), more than 400 cores will fight in unison with TORUS to bring cloud computing closer to environmental sciences. More than 400 computer hearts beat in unison for TORUS, as well as those of Nathalie, Astrid, Eleonora, Ann, Imeshi, Thanh, Sukhuma, Janitra, Kim, Daniel, Yannick, Florent, Peio, Alex, Lucca, Stefano, Hichem, Hung(s), Thuy, Huy, Le Quoc, Kim Loi, Agustian, Hong, Sothea, Tongchai, Stephane, Simone, Marco, Mario, Trinh, Thiet, Massimiliano, Nikolaos, Minh Tu, Vincent and Dominique. To all of you, a big thank you.

xvi

TORUS 2 – Toward an Open Resource Using Services

Structure of the book This book is divided into three volumes. Volume 1 raises the problem of voluminous data in geosciences before presenting the main methods of analysis and computer solutions mobilized to meet them. Volume 2 presents remote sensing, geographic information systems (GIS) and spatial data infrastructures (SDI) that are central to all disciplines that deal with geographic space. Volume 3 is a collection of thematic application cases representative of the specificities of the teams involved in TORUS and which motivated their needs in terms of cloud computing. Dominique LAFFLY January 2020

Part 1

Earth Science Remote Sensing

TORUS 2 – Toward an Open Resource Using Services: Cloud Computing for Environmental Data, First Edition. Edited by Dominique Laffly. © ISTE Ltd 2020. Published by ISTE Ltd and John Wiley & Sons, Inc.

Introduction to Part 1 From Above We Can See Earth Better

In 2013, Hansen et al. proposed a global analysis – to understand the Earth as a whole – of forest evolution from remote sensing data: “Quantification of global forest change has been lacking despite the recognized importance of forest ecosystem services. In this study, Earth observation satellite data were used to map global forest loss (2.3 million square kilometers) and gain (0.8 million square kilometers) from 2000 to 2012 at a spatial resolution of 30 meters [...] These results depict a globally consistent and locally relevant record of forest change”. Three centuries earlier, in 1794 at the Battle of Fleurus, the Montgolfier brothers allowed the military to see at a distance and in single look the battlefields thus opening the way to the observation of the Earth’s surface from the sky but “The history Remote sensing [remote contactless] begins in 1858 when Gaspard Felix Tournachon said Nadar (1820-1910) takes the first aerial photograph from an aerostat above the Kremlin Bicetre district in Paris”1: From aerostation with a photographer equipped with a tin-plate apparatus to acquire an oblique view of a small portion of space, remote sensing now consists of constellations of orbital platforms, equipped with various sensors with very high spatial and thematic precision that continuously acquire digital images covering the entire planet. Remote sensing, previously exceptional, is now becoming commonplace. “Remote sensing is the set of techniques used to determine the properties of objects, natural or artificial, from the radiation they emit or reflect” and the fundamental principle: the study of Matter-Radiation interactions. Remote viewing in a single glance and photography are the two pillars of remote sensing that are quickly captured by the military who fix cameras on balloons, kites, pigeons, planes and so on to cover more and more surface area, and be able to review the footage in time for strategic, operational and intelligence purposes (the u-2 incident, 1960 and the Cuban Missile Crisis in 1962). While the aerial photographic missions from the 1930s are a Chapter written by Dominique LAFFLY. 1 (Wikipedia https://en.wikipedia.org/wiki/Nadar).

xx

TORUS 2 – Toward an Open Resource Using Services

classic technique of Earth observation – in 1959 for countries that could afford it, the cameras were placed in orbit to see even further, cover larger surfaces and renew images more frequently (Keyhole-1). In 1960 the first satellite for civil applications (meteorology and weather forecasting) was placed in orbit by NASA – TIROS (Television InfraRed Observation Satellite) acquires images every 10 seconds with these two cameras scanning. In 1972, NASA again ushered in a new era of Earth observation, the first LANDSAT Earth observation satellite equipped with a point-topoint optico-electronic acquisition sensor (MSS – multi spectral scanner) was placed in polar orbit. In 1986 CNES launched SPOT1 in the same orbit with the first pick-up sensor “pushbroom” on board. Today, hundreds of sensors embedded on a dozen platforms provide in a continuous flow – or almost – images of the Earth at spatial resolutions, ranging from a few centimeters to a few kilometers, geographic coverage from punctual to global and in a variety of measurements that can be used to analyze the atmosphere, as well as land and hydrosphere – the three main applications of remote sensing. For a decade now a new era marks remote sensing by three major facts: imaging very very (ultra?) High Spatial Resolution is a part of remote sensing with UAV imagery; in situ observation – in situ sensing – tends to become widespread in many environmental domains; A-train (NASA) and Copernicus-Sentinel (ESA) offer free remote sensing imagery to the community of users, some of which (Sentinel 2 and Landsat OLI) cover the entire Earth in a matter of days with spatial resolutions of the order of 10 m to 30 m for the finest thematic diversity. What do we do with all these images that theoretically allow us, on a pixel scale, to generalize a thematic analysis to the whole of the Earth (see Ansen et al., Ibid.)? How do we integrate data from different sensors? How is the high frequency of time acquisition an added value? What computer and methodological tools are able to process such amounts of data? What are the needs of the end users? In addition, remote sensing is at the heart of the Big Data problem and their systematic exploitation (Artificial Intelligence and Cloud Computing). Remote sensing is also at the heart of the TORUS program as – all things considered – it is essential for NASA, ESA and Google who, in quick succession in 2014 and 2016, proposed cloud computing solutions (NASA Earth eXchange – NEX, Sentinel Cloud Services for Copernicus Users and Google Earth Engine) to deeply exploit this environmental information. Other platforms today offer autonomous cloud solutions such as EOS Platform that provides a complete cloud solution for interacting with geospatial data, creates an ecosystem allowing users to collaborate and share information with each other and changes the rules for data distributing and reselling. A last major fact in the field of remote sensing is that digital imaging, on the one hand, and the representation of the Earth, on the other hand, are today part of the everyday life of most citizens of the Earth. Royer et al. (2007) noted “at least two major factors have gradually changed our perception of things: the public availability of digital images with digital cameras (photoscopes) and the accessibility of satellite images via the Internet or in the media. In fact, not only has the digital image become as commonplace in our daily lives as telecommunications,

Introduction to Part 1

xxi

but it can also be easily transported over the Internet or through telephony. Satellite or aerial imagery is now part of our daily lives, as can be seen with every disaster of natural or human origin, followed almost directly, wherever in the world the tragedy takes place: tsunami in Asia, flooding in Louisiana, volcanic eruption in Reunion, terrorist attack in New York or drought in the Sahel. The tools also followed, like Google Earth (in 2006 alone the software was downloaded 100 million times)2, which literally made everyone discover the fabulous potential of satellite imagery and which also makes it very easy to integrate the various measured parameters by making high-performance visualization tools available.” Remote sensing is everywhere and is for everybody. However remote sensing is poorly mastered. Stork et al. explain “there exists a huge knowledge gap between the application and the understanding of satellite images. Remote sensing only plays a tangential role in schools, regardless of the political claims to strengthen the support for teaching on the subject” because “a lot of the computer software explicitly developed for school lessons has not yet been implemented due to its complexity. Thereby, the subject is either not at all integrated into the curriculum or does not pass the step of an interpretation of analogue images. In fact, the subject of remote sensing requires a consolidation of physics and mathematics as well as competences in the fields of media and methods, apart from the mere visual interpretation of satellite images”3. In the same idea, Difter et al. explains, “at most universities remote sensing is associated with Geography departments. Remote Sensing has a growing relevance in the modern information society. It represents a key technology as part of the aerospace industry and bears increasing economic relevance [...] Furthermore, remote sensing exceedingly influences everyday life, ranging from weather forecasts to reports on climate change or natural disasters. As an example, 80% of German students use the services of Google Earth [...] But studies have shown that only a fraction of them know more about the data they are working with”. The NASA EOSDIS annual customer satisfaction survey illustrates this situation perfectly. In the 2018 survey, which student, which user can really answer this question? “Which of the following web services are you/would you be interested in using? – OGC (e.g. WMS, WCS, WFS), – OPeNDAP (e.g. THREDDS, Hyrax), – REST based web calls, – SOAP based web calls, – Remote Procedure Call (RPC), – Programming Language Library, – Other (please specify).”

2 (Wikipedia https://en.wikipedia.org/wiki/Google_Earth). 3 (Wikipedia https://en.wikipedia.org/wiki/Remote_sensing).

xxii

TORUS 2 – Toward an Open Resource Using Services

Let us not forget, in light of contemporary remote sensing, that at the beginning, the question was rather: know what to do with these images? This question may come as a surprise because spatial images are now ubiquitous – we are talking here about images of satellites in polar orbit, that is to forget that remote sensing, in its early days in the mid-1980s, was of little interest: the images were rare, very expensive and delivered on magnetic tapes do you. They required the use of large computer systems with a very limited software offer – so new major costs. The analog restitutions were very limited and only a few “laser printers” in the world allowed it: we obtained a “positive” that could then be shot like a silver print or video, projected on a wall or screen. Alternatively, more costs the images were printed in small sections of 132 pixels wide on linear B&W printers or a pixel was represented by a character (Figure I.1). Under these conditions, only a few national and international institutions used them for scientific purposes, as well as the big oil companies.

Figure I.1. Silver print from a positive (https://esdac.jrc.ec.europa.eu/resourcetype/national-soil-maps-eudasm) and extracted from a B&W print with a linear printer from a 1973 MSS image. For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

However, since the origins of remote sensing, the main agencies have associated the production of images with the development of software and libraries of functions to facilitate the integration of data in communities of potential users. NASA supported the launch of Landsat satellites for remotely sensed data dissemination missions by partnering with major computer manufacturers to develop equipment (see COMTAL monitor by the pass) and software, while financing upgrading companies. CNES did the same as ESA, and today all those agencies, that no longer need to promote satellite imagery, offer free feature libraries to make the most of the images: – NASA Technology Transfer Program – remote sensing toolkit4: a range of software and libraries in the domain of environmental science (Earth, air, space, 4 https://software.nasa.gov/remotesensing/, see full catalog here: https://software.nasa.gov/ NASA_Software_Catalog_2017-18.pdf

Introduction to Part 1

xxiii

exoplanet, terrestrial environments, planetary atmospheric modeling, radiation shielding) and data and image processing (algorithms, data analysis, data processing); – ESA Earth Online – software tools5: more than 40 tools to use ESA data in the domains of agriculture, atmosphere, solid earth, water, land, oceans and coasts, snow and ice and natural disasters; – Theia Softwares: the Theia land data services center6 is a French national inter-agency organization designed to foster the use of images issued from the space observation of land surfaces. Theia is offering scientific communities and public policy actors a broad range of images at different scales, methods and services: OTB (Orfeo toolbox – CNES), Monteverdi (CNES), SNAP (Sentinel Application Platform – ESA), NEST (Next ESA SAT Toolbox – ESA), Guidos ToolBox (EU), PolSARpro (ESA), ASF MapReady (ESA), Micmac (IGN) and OLIVE platform (ESA). – Many other open geospatial tools are also available (with some for the cloud). OSGEO libraries7: GDAL (Geospatial Data Abstraction Library), GeoTrellis (Scala), GeoTools (Java) and so on. Geographical information systems such as QGIS, GRASS and SAGA8 that all now offer cloud platforms. Remote sensing is everywhere and is for everybody but remote sensing always needs “a consolidation of physics and mathematics as well as competences in the fields of media and methods” (Ibid.). In this part of the book we will recall the great principles inherent to remote sensing to ensure a common theoretical and methodological knowledge to researchers and teacher-researchers of the different disciplines involved in the TORUS program: – a bit of history, economic aspects and some definitions to never forgot that remote sensing programs are very expensive even when data are free; – launchers and orbits to never forgot how hard it is to leave the Earth’s gravity and to control the orbit; – measurement of electromagnetic radiation and digital images to never forget the fundamental rules of physics and the native digital format of the data; – image quality to never forget how remote sensing is an expression of great technical prowess and engineering; – the major fields of application to never forget why we use remote sensing, societal issues for the development of the planet (or it should be). See “100 Earth Shattering Remote Sensing Applications & Uses” published by GISGeography in 2018 and find yours: 5 https://earth.esa.int/web/guest/software-tools 6 http://www.theia-land.fr/en 7 https://www.osgeo.org/choose-a-project/development/libraries/ 8 https://qgis.org, https://grass.osgeo.org, http://www.saga-gis.org/en/index.html and see GISGeography for more details: https://gisgeography.com/free-gis-software/

xxiv

1 2 3

4

5 6

7 8 9 10 11 12 13 14

TORUS 2 – Toward an Open Resource Using Services

Determining soil moisture content using active and passive sensors from space, Mapping with laser precision using Light Detection and Ranging technology, Catching tax-evaders redhanded by locating new construction and building alterations, Spinning the globe with mapping services like Google Earth, Bing Maps and OpenStreetMaps, Predicting retail earnings and market share by counting cars in parking lot, Snapping aerial photos for military surveillance using messenger pigeons in World War II, Charging higher insurance premiums in flood-prone areas using radar, Doing the detective work for fraudulent crop insurance claims, Searching for aircrafts and saving lives after fatal crashes, Detecting oil spills for marine life and environmental preservation, Counting polar bears to ensure sustainable population levels, Uncovering habitat suitability and fragmentation for panda bears in protected areas, Identifying forest stands and tallying their area to estimate forest supplies, Navigating ships safely with the most optimal route,

15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31

Measuring wind speed and direction for wind farms, weather forecasting and surfers, Spying on enemies with reconnaissance satellites, Delineating and assessing the health of riparian zones to conserve lakes and rivers, Estimating surface elevation with the Shuttle Radar Topography Mission, Extracting mineral deposits with hyperspectral remote sensing, Watching algae grow as an indicator of environmental health, Forecasting weather to warn about natural disasters, Detecting land cover/use types for decision making, Monitoring the environment with the ESA’s Copernicus Program, Mapping soil types for agriculture planning, Preventing the spread of forest disease types, Fighting wildfires by planning firefighter dispatch, Monitoring air quality in the lower atmosphere, Assessing terrain stability using interferometry in the oil and gas sector, Unearthing ancient archaeological sites like the Mayans and ancient Egypt, Pinpointing your position on Earth with Global Positioning Satellites, Optimizing solar panel energy output with global horizontal irradiance,

Introduction to Part 1

32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51

Finding the driving factors that contribute to poverty, Observing the flow of ocean currents and circulation, Studying glacier melts and effects on sea levels, Providing a base map for visual reference and assisting orient the map reader, Snorkeling in an oasis of marine vegetation with the coastal channel, Tracking hazards for better response and recovery, Keeping tabs on the shift from rural to urban growth, Quantifying crop conditions with Normalized Difference Vegetation Index (NDVI), Preventing the degradation and loss of wetland ecosystems, Tracking sediment transport into rivers and lakes, Saving money and time on the farm with precision farming, Reversing illegal rainforest cutting in Brazil, Putting illegal boat dumping under the microscope, Inventorying and assessing rural road conditions with UAVs, Driving with no hands (autonomous vehicles), Measuring gravity with the GRACE satellites, Deriving elevation and contours using photogrammetry, Watching the aurora borealis from another angle, Comparing the past and present with human impact change, Planning an optimal telecom network capacity,

52 53 54 55 56 57

58 59 60 61 62 63 64 65 66 67 68 69 70

xxv

Tracking displaced refugees to help deliver aid and services, Covering the most ground in search of road cracks, Getting a top-down view when purchasing real estate, Keeping a watchful eye to prevent future atrocities from happening, Designing a lift irrigation system to supply water in India, Measuring the volume difference at a uranium enrichment site using 3D mapping, Helping provide clean drinking water with base maps, Monitoring active volcanoes using thermal remote sensing, Inventorying potential landslides with interferometry, Catching fish and improving long-term fisheries sustainability, Tracking the great distances of migratory birds and inspecting their prevalence, Preventing the spread of diseases in epidemiology, Recording video footage from satellites, Quantifying the damage after an earthquake, Looking at the Earth as an art masterpiece, Recognizing buildings easily with the bird’s-eye oblique view, Mapping the mysteries of our ocean floors, Understanding the human rights situation in North Korea, Comparing climatic factors from past to present,

xxvi

71 72 73 74 75 76 77 78 79 80 81 82 83 84

TORUS 2 – Toward an Open Resource Using Services

Monitoring the global sex trade situation in remote areas, Assessing fuel economy of vehicle emissions, Providing early warning signs for famine over large scales, Mapping regional economic activity at night, Studying geology of the Earth’s surface, Assessing the environmental change and promoting biodiversity in parks, Measuring albedo for Earth’s radiation budget, Locating groundwater activity for wells, Observing population growth in urban areas using land use change, Keeping a watchful eye on biodiversity, Keeping an inventory on cemeteries using UAVs, Predicting the occurrence of dinosaur tracks for paleontologists, Delineating watersheds using DEMs for hydrologists, Using habitat suitability models to predict the abundance of mosquitoes,

85

Using a least-cost analysis and vegetation to understand wildebeest migration, 86 Assisting cities manage assets and ensuring safety standards, 87 Calculating the depth of snowpack, 88 Planning spine-jarring black diamond ski runs with aspect data, 89 Improving efficiency and safety of air traffic control, 90 Spotting undeclared nuclear power plants automatically, 91 Narrowing down a search for a missing body, 92 Monitoring oil reserves by looking at floating oil roof tanks, 93 Finding ghost cities on the map, 94 Spotting swimming pools for late-night dives, 95 Reducing traffic jams using change detection, 96 Measuring the size of protests for journalists, 97 Measuring the rise of sea levels, 98 Creating an automated road network instantly, 99 Picking up on signals from submarines in shallow water, 100 Exploring, protecting and navigating in the arctic.

Introduction to Part 1

xxvii

This list illustrates how spatial imagery is now a source of information for all environmental sciences, from urban planning to health, through biology and meteorology. End users are just as numerous: teachers, farmers, planners and urbanists, doctors, tour operators, social scientists, economic and exact scientists, artists, territorial managers, computer scientists, journalists, and datajournalists and so on. Spatial and spectral resolutions, geographical coverage and acquisition repetition allow us to distinguish the fields of application from local to global and from static to pseudo-dynamic. In this diversity there are generally three main axes: the cryosphere and the hydrosphere; the land surfaces; the atmosphere. Teams associated with the TORUS program provide expertise in many of these areas, which will be discussed in detail in this book: the use of remote sensing to model the transport of pollutants and air pollution (Asian Institute of Technology and Vietnam National University, Hanoi); spatial imagery in erosion modeling, land cover mapping, monitoring the temporal evolution of landscape changes (Vrjie University of Brussels, Nong Lam Ho Chi Minh City University, University of Toulouse 2); methods of learning and artificial intelligence for the mapping of soil organic carbon from space imaging (University of Toulouse 2 and International School of Information Processing); webmapping applications and remote sensing (Vietnam National University, Hanoi, International School of Information Processing Sciences, Vrjie University of Brussels). Remote sensing is everywhere and is for everybody, and also because images are so beautiful, yes so beautiful (Figure I.2): Earth as Art as introduced by Lawrence Fried from NASA: “In 1960, the United States put its first Earth-observing environmental satellite into orbit around the planet. Over the decades, these satellites have provided invaluable information, and the vantage point of space has provided new perspectives on Earth. This book celebrates Earth’s aesthetic beauty in the patterns, shapes, colors, and textures of the land, oceans, ice, and atmosphere. Earth-observing environmental satellites can measure outside the visible range of light, so these images show more than what is visible to the naked eye. The beauty of Earth is clear, and the artistry ranges from the surreal to the sublime. Truly, by escaping Earth’s gravity we discovered its attraction. Earth as art – enjoy the gallery.”

xxviii

TORUS 2 – Toward an Open Resource Using Services

Figure I.2. Earth as Art from NASA9 (modified from: see footnote #12. For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

9 Modified from: https://www.nasa.gov/pdf/703154main_earth_art-ebook.pdf.

1 A Brief History of Remote Sensing

1.1. History It is with the invention of photography – with Joseph Nicéphore Niepce being one of the main actors (1765–1833) – and the aerial transport of balloons that bore the first images of the Earth seen from above, in the sense that we habitually hear. It is to Nadar that we owe this feat: “Before the winter of 1858, Nadar chartered a balloon and rented the services of a member of the Godard family, the famous dynasty of aeronauts. The laboratory is installed in the nacelle, the fixed camera and lens oriented towards the ground. Nadar rises, triggers his ‘horizontal guillotine’ acting as a shutter, immediately develops the image, but the plate comes out of the baths completely veiled. No visible image. Nadar tries several times, but to no avail. After ‘a last failure, [he] gives the order to release everything’ and offers himself a free balloon ascent to the valley of Bièvres, Petit-Bicêtre. The weather is calm and Nadar decides not to empty the ball, to try the next day a new experience. The appendage that releases the hydrogen is closed and the aerostat securely moored to an apple tree. In the morning, the photographer goes back in his basket, but the balloon refuses to leave the ground. The cold of the night has condensed the hydrogen molecules which have lost their ascensional force. Nadar rids the basket of all superfluous, embarks equipped with a plate that he has sensitized and this time rises 80 meters from the ground. It triggers, gives the order to go down the balloon and develops in the neighboring inn. This time, the hydrogen sulphide contained in the closed envelope of the balloon could not blacken the plate of the Chapter written by Dominique LAFFLY.

TORUS 2 – Toward an Open Resource Using Services: Cloud Computing for Environmental Data, First Edition. Edited by Dominique Laffly. © ISTE Ltd 2020. Published by ISTE Ltd and John Wiley & Sons, Inc.

2

TORUS 2 – Toward an Open Resource Using Services

photographer. Although faintly, an image appears and shows a farm, an inn and the gendarmerie […] We can distinguish perfectly the roof tiles and on the road a tapestry whose carter stopped short in front of the balloon. From this photographic feat, we only know the written description of Nadar’s hand in Les Mémoires du Géant and that are repeated the various texts recalling the essays of 1858. Nadar does not inform us about the angle of view whose perpendicularity is one of the ‘essential conditions that can [enable] him to fulfill the purpose he has proposed’, that is ‘to use the photograph for the lifting of plans’ for civil and military use. The initial project of making aerial photography a tool for mapping remains unresolved. One of the parameters forgotten by Nadar’s patent is the set of aerostatic constraints.”1 The first available archives go back to 1868, and the following figures present the photographs seen from the sky which will become generalized little by little2.

Figure 1.1. Aerial view of the district of Etoile, drawn from negative glass in wet collodion 24 × 30 cm, Nadar July 16, 1868 1 http://images-aero.blogspot.com/2009/11/1-nadar-1858.html. 2 See Études photographiques: https://journals.openedition.org/etudesphotographiques/479.

A Brief History of Remote Sensing

3

Figure 1.2. “La Bourse”, Paris, and its surroundings, A. Schelcher and A. Omer-Decugis, 1909

In 1907, Julius Neubronner, a German apothecary, and inventor, had the idea of using pigeons equipped with cameras to take aerial shots. His initial idea was to follow the routes taken by the pigeons, the snapshots were sold as a postcard until the army recovers its invention during World War I. Many nations used this process for espionage before it was abandoned and replaced by aviation.

4

TORUS 2 – Toward an Open Resource Using Services

Figure 1.3. Pigeon photographer and snapshot examples

In the early 20th Century, the plane gradually took the place of the balloons. The Wright brothers made their first flight in 1909 and World War I revealed the military and strategic importance of aviation, including the contribution of aerial photographs.

A Brief History of Remote Sensing

5

Figure 1.4. Device developed for aerial photography with an operator attached to the front of the aircraft near the engine (World War I)

Figure 1.5. Glaciers Lovén, Spisterg (80 °N, 12° E) Nadir image (1948) Oblique image (1936). Sources: NorskPolarInstitutt (Norway)

In parallel with military missions (the U-2 incident, 1960 and the Cuban Missile Crisis in 1962), civil applications were emerging, whose main purpose was cartography and map updating. The original oblique shots are quickly made perpendicular to the ground to reduce – but not eliminate – the many distortions caused by various factors detailed later.

6

TORUS 2 – Toward an Open Resource Using Services

Balloons, pigeons, kites, airplanes...observation platforms gain altitude. The U-2 plane was flying at 21,000 meters in 1950 and foreshadowed the era of satellites, the first of which would be launched by the Americans in 1959: Keyhole. This reconnaissance satellite has a camera embedded, and the pictures are stored in a capsule, regularly ejected from the satellite and recovered in flight by a military aircraft.

Figure 1.6. First image of the pentagon acquired by the KH1 satellite (Keyhole 1) in 1959 (Wikipedia https://fr.wikipedia.org/wiki/Corona_(satellite))

In 1960, the first meteorological satellite was put into orbit, again by the Americans: TIROS (Television InfraRed Observation Satellite). Although it lasted only 78 operational days, TIROS-1 demonstrated its usefulness (22,592 photos acquired and used for forecasting) and TIROS-2 was launched the same year, which is still in orbit although it is out of order. “The TIROS-1 was launched by NASA from Cape Canaveral in an elliptical orbit (696 km × 756 km) with a period of 98.3 minutes, tilted 48.4 degrees from the plane of the equator. It had a cylindrical shape with 18 sides on which 9,200 photovoltaic cells were mounted. It measured 1.07 meters by 0.56 meters, including objectives. It weighed 128.4 kilograms, including the storage battery and solid fuel for a thruster stabilizing rotation. TIROS-1 was equipped with two slow-scan television cameras taking pictures of the Earth under the satellite – up to a photo every ten seconds. The first was equipped with a wide-angle

A Brief History of Remote Sensing

7

lens with a field of view of 1,207 km on each side of the point under the satellite, and the second, a zoom with a viewing angle of 129 km. The imaging was preprogrammed and the photos were stored on two magnetic tape units, one for each camera, for later broadcast when the satellite was away from a receiving antenna. Each tape was 122 meters long, enough to record 32 photos. On the other hand, the images were sent live when passing over a receiving station and the ground control could then command to take pictures every 10 or 30 seconds. Two terrestrial stations were receiving TIROS-1 data. [...] Data transmitted from the satellite was picked up by either station and recorded on 35mm film for later reproduction. Meteorological technicians analyzed the cloud cover from these photos and produced maps by hand for faxing to the National Weather Service’s main weather station near Washington, DC. It was not until 1962, with the TIROS-4 and 5 that the photos were sent directly to the main center and to some offices throughout the United States”3.

Figure 1.7. First TIROS image (1960) transmitted by radio waves to a terrestrial reception station (Source: Wikipedia, see https://fr.wikipedia.org/wiki/TIROS-1)

3 The former USSR launched Sputnik 2 in 1957 carrying Laika the dog – one of the first animals in space – and in 1961, sent the first man, Yuri Gagarin. The USA’s Apollo program ran from 1961 until 1972 and ended the race to the Moon – Armstrong, 1969. The USA subsequently imposed their supremacy in the field of space and opened the era of scientific missions and the observation of the Earth: exploration of the solar system (Pioneer, 1973, Voyager 1 and 2, 1976–1977); microgravity experimentation (Skylab, Salyut, MIR, ISS, etc.); and space astronomy (Hubble Space Telescope, 1990).

8

TORUS 2 – Toward an Open Resource Using Services

The era of space conquest was launched. First, with prestige and international supremacy as the logic. In 1972, the first LANDSAT-1 Earth Observation Satellite equipped with an optical point-to-point acquisition (MSS) sensor was placed in the polar orbit. It inaugurated a new era of observation and mapping of landscapes and planning. In 1986, SPOT-1 introduced the first line-by-line (pushbrom) acquisition sensor and then appeared in the 1990s as the first 2D matrix detectors – which, remain marginal however.

Figure 1.8. Landsat MSS First Image – July 1972 – Salt Lake City (USA). Detail in vignette, pixel of 56 * 79 brought back to 50 * 50 m (NASA) and a few centimeters spatial resolution of actual imagery! For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

1.2. Fields of application The fields of application of remote sensing from above are varied and respond to various concerns, fundamental laws of physics and the origin of the universe, through telecommunications, GPS navigation, exploration of the solar system and the Earth observation that interests us here. For this, several hundred satellites and sensors are now operational.

A Brief History of Remote Sensing

9

Figure 1.9. 150 satellites, 300 instruments4. For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

The study and observation of the Earth can be separated into two main groups: – the “objects” of interest themselves: - the Earth’s surface in the geological sense; - landscapes, land use, development, farming systems, urban planning etc; - the hydrosphere (oceans, continental waters, etc.); - atmosphere; – their interfaces and cycles: - water cycle; - carbon cycle; - the dynamics of the landscapes. Images and other data provided by satellites do not constitute the science of Earth observation, they are information that complements other sources of data in the field, in the oceans, in the atmosphere and so on. Information complements the

4 From Satellite Imagery – From Acquisition Principles to Processing of Optical Images for Observing the Earth, C. Valorge (CNES), (2012).

10

TORUS 2 – Toward an Open Resource Using Services

models developed by scientists. The main interest of remote sensing is to offer a global and homogeneous vision. However, the constraints are important: the instruments are distant from the objects observed and the measured information is all the more disturbed by the crossing of the atmosphere; except for the geostationary orbits, the repetitiveness of the acquisitions is relatively limited, the oblique sights nevertheless allow us to reduce the steps of acquisitions to a few days; space systems are expensive and require several years of development; very high spatial resolution data are (very) expensive when distributed through commercial channels. In any case, satellite data is now ubiquitous in the Earth sciences and development, they are free access today for a great diversity of sensors and spatial resolution (Terra/Aqua, A-Train constellation, COPERNICUS European program, etc.). 1.3. Orbits, launchers and platforms One of the originalities of remote sensing space is that the satellites or platforms are placed in orbit around the Earth. These orbits answer to complex laws and are generally grouped into three types based on target application, and whether all or part of the terrestrial sphere is to be observed: – the geostationary orbit (GEO, GEostationary Orbit) is an orbit located at an altitude of 35,786 km above the equator of the Earth, in the equatorial plane and of zero orbital eccentricity. This is a special case of geosynchronous orbit. Most of the meteorological and telecommunications satellites are placed in such orbits; – the average orbit usually peaks at an altitude of 20,000 km with a 12-hour period (MEO, Medium Earth Orbit). The orbit located outside the Earth’s atmosphere is very stable. The signals sent by the satellite can be received over a large part of the Earth’s surface. This is the altitude chosen for navigation satellites such as GNSS (Global Navigation Satellite System); – Low Earth Orbit (LEO) is an area of Earth’s orbit up to 2,000 kilometers in altitude. This orbit is very important for space exploration, since it is here that there are many remote sensing and telecommunications satellites, as well as space stations. Among the low orbits, particular attention is given to the so-called sun-synchronous one, where the majority of the Earth observation satellites are located, such as Landsat, SPOT, IRS, PLEIADES, and SENTINEL. For an artificial satellite, the sun-synchronous orbit designates a geocentric orbit (around the Earth)

A Brief History of Remote Sensing

11

whose altitude and inclination is chosen so that the angle between the orbit plane and the direction of the sun remains approximately constant. A satellite placed in such an orbit passes over a given point on the Earth’s surface at the same local solar time. This orbit – called the polar orbit – is used by all satellites that make photographic observations in visible light, because the solar illumination of the observed location will be very variable from one image to another. A polar orbit that passes near the pole is low (between 600 and 1000 km) with a short periodicity (described every 96–110 minutes). The satellite cuts at the plane of the Earth’s equator about 12 times a day and flies over the scene 1500 hrs, local time around. To be placed in orbit, the platforms must escape the terrestrial attraction aboard “rockets” or launchers. These vary according to the chosen orbit and the mass of the satellite. Launchers are the result of significant financial and research investments that only major nations can support. Among them the US first, then Europe, the former USSR, China, Japan and India. Recently, private companies have come to compete with the launcher market, for example SpaceX. Among the main launchers, we currently distinguish: – Falcon 9 (SpaceX) – 22.8 t in low orbit or 8.3 t in transfer orbit; – Ariane 5 (Arianespace) – 10 t maximum capacity; – Vega (Arianespace) – 1.5 t in low orbit; – Soyuz (Arianespace) – 4 t up to 1400 km maximum; – Delta (Boeing) – up to 20 t in low orbit for the largest launcher in the series, 13 t in geostationary orbit; – Titan and Atlas (Lockheed Martin and ILS) – 20 t in low orbit for the largest launcher in the series, 8.7 t in geostationary orbit; – Proton (ILS) – 21 t in low orbit for the largest launcher in the series, 5–7 in geostationary orbit; – CGWIC (China) – 5.2 t maximum capacity. Launchers do not place the satellites directly in their final orbit but in a transfer orbit, especially for those destined for a geostationary orbit. It will then be the platforms themselves which will ensure their routing in their final orbit, which they need in this capacity of propulsion and guidance. This equipment can considerably increase the load and reduce the load for the sensors. One of the most delicate phases of a satellite’s life is to make it operational: deploying solar panels; pointing of acquisitions; starting all the equipment and so on. All these operations are done in

12

TORUS 2 – Toward an Open Resource Using Services

direct connection with ground receiving stations which are an integral part of the system. This is called a space segment and a ground control segment. There are sensors that can adapt to a dedicated platform (e.g. SPOT) sensors that are part of the platform (PLEIADES, for example). In all cases the systems must fulfill different functions and meet several constraints: – payload of propulsion energy; – recharging and storage of solar batteries; – satellite attitude control and update of the orbit (gyroscope and/or star tracker); – telecommunications (data transfer, control by the ground segment); – acquisition of data; – interface with the launcher. 1.4. The acquired data are digital images If at the origin of remote sensing the images were argentic photographs, very quickly they were replaced by electronic systems, to become exclusively digital today. Let us remember here that a digital image associates with each elementary point – contraction pixel of the terms picture element – a numerical account translated into gray level. It is in fact a matrix of numerical values associating with each pixel a quantity called radiometry. In general terms, a digital image is therefore a two-dimensional function sampled regularly and giving, at each sampling point, a value called radiometry proportional to the luminance emitted by the landscape: Image (i,j) = radiometry (x + i*Δx; y + j*Δy) with: Δx, Δy the respective sampling steps in line and column. There are generally three main areas for image quality: – all that is concerned with the precise positioning of the pixels: absolute location of the point on the Earth and relative location of the pixels between them. These aspects are grouped under the term of geometry of the images; – all that concerns the physical interpretation of gray levels is called radiometry of images; – all that concerns the capacity of the system to perceive the small details of the landscapes, to make the images clear, represents the study of the resolution of the images.

A Brief History of Remote Sensing

13

Figure 1.10. Variation of the spatial resolution according to the swath. The diagonal 5 represents the growth of the file size, from a few megabytes to several gigabytes

We will see below that many processes – concerning all or part of these three aspects related to image quality – are generally implemented by users. A primordial aspect of remote sensing escapes these; however, it is the despatization of images (not to mention the aspects related to the controls of the mission). Despatialization processes consist of correcting the defects of the images related to the acquisition of this one. They are based on a precise knowledge of the conditions of acquisitions that affect the orbit, the electronics, the excesses of the electronics and so on. The goal is to provide a data directly usable by the users. The images offer a compromise between spatial resolution, ground footprint, depth of coding, power transmission/reception torque, computation time and so on, where many technical, physical and financial parameters interfere which limit considerably, even forbid, certain future developments such as the race to centimeter resolution. The most accurate images were changed from a resolution of 56 * 79 m with Landsat MSS in 1972 to pixels of about 50 cm with the satellites GEOEYE, ORBVIEW, QUICKBIRD, WORLDVIEW or PLEAIDES, with a geographical coverage of 188 km from side about 10 km. At the same time, the mass of the satellites was multiplied by 25 and the size of the receiving stations divided by 50.

5 see: tandfonline.com/doi/abs/10.1080/01431168708948666.

14

TORUS 2 – Toward an Open Resource Using Services

The size of the files has increased from a few tens of megabytes to several gigabytes, parallel to the rise of microprocessors. The processing is increasingly greedy in pixel calculations to the point that, even with farms of several hundred microprocessors, despatization processing times, for example, were estimated at more than 45 min per image for PLEIADES (12,000 operations/pixel against 37 for SPOT1) at the beginning of the eponymous program in 2011. 1.5. So what is remote sensing? Some definitions Encyclopædia Britannica – Earth Exploration “This comprises measurements of electromagnetic radiation from the ground, usually of reflected energy in various spectral ranges measured from aircraft or satellites. Remote sensing encompasses aerial photography and other kinds of measurements that are generally displayed in the form of photograph-like images. Its applications involve a broad range of studies, including cartographic, botanical, geological, and military investigations. Remote-sensing techniques involve using combinations of images. Images from different flight paths can be combined to allow an interpreter to perceive features in three dimensions, while those in different spectral bands may identify specific types of rock, soil, vegetation, and other entities, where species have distinctive reflectance values in different spectral regions (i.e. tone signatures). Images taken at intervals make it possible to observe changes that occur over time, such as the seasonal growth of a crop or changes wrought by a storm or flood. Those taken at different times of the day or at different sun angles may reveal quite distinct features; for example, seafloor features in relatively shallow water in a calm sea can be mapped when the Sun is high. Radar radiation penetrates clouds and thus permits mapping from above them. Sidelooking airborne radar (SLAR) is sensitive to changes in land slope and surface roughness. By registering images from adjacent flight paths, synthetic stereo pairs may give ground elevations. Thermal infrared energy is detected by an opticalmechanical scanner. The detector is cooled by a liquid-nitrogen (or liquid-helium) jacket that encloses it, making the instrument sensitive at long wavelengths and isolating it from heat radiation from the immediate surroundings. A rotating mirror directs radiation coming from various directions onto the sensor. An image can be created by displaying the output in a form synchronized with the direction of the beam (as with a cathode-ray tube). Infrared radiation permits mapping surface temperatures to a precision of less than a degree and thus shows the effects of phenomena that produce temperature variations, such as groundwater movement. Landsat images are among the most commonly used. They are produced with data obtained from a multispectral scanner carried aboard certain U.S. Landsat satellites orbiting the Earth at an altitude of about 900 kilometres. Images covering an area of 185 kilometres square are available for every segment of the Earth’s surface.

A Brief History of Remote Sensing

15

Scanner measurements are made in four spectral bands: green and red in the visible portion of the spectrum, and two infrared bands. The data are usually displayed by arbitrarily assigning different colours to the bands and then superimposing these to make ‘false-colour’ images.” Encyclopædia Universalis (volume 22, pp. 205–211). “‘Remote sensing’ means a set of techniques implemented from airplanes, balloons, satellites, space shuttles, orbital stations and intended to study either the surface of the Earth (or other planets), the atmosphere, using the properties of electromagnetic waves emitted, reflected or diffracted by the different bodies observed. It makes it possible to inventory the terrestrial resources, to improve the meteorological forecasts and, more generally, to contribute to the geography. The origin of remote sensing merges with that of aerial photography, but the range of waves interesting for remote sensing is much wider than the visible range and goes from the ultraviolet (o, 3 μm) to the centimeter waves of the radar. On the one hand, we distinguish between passive techniques, where we simply record the natural energy emitted or reflected by the bodies, and, on the other hand, the active techniques, or we ‘illuminate’ the bodies to study, before recording the energy they send back to the detector. The atmosphere does not transmit electromagnetic radiation in a uniform manner; many are absorbed, especially by water vapor and carbon dioxide. Transmission ‘windows’ exist, on the one hand, in the band of the visible spectrum (0.4 to 0.7 μm), on the other hand, in the infrared (from 0.7 to 14 μm); farther on, from the millimeter waves, the atmosphere becomes almost transparent. These are the three areas particularly used by remote sensing.” Centre National des Études Spatiales (CNES, A. Alouges, “Remote sensing – definition” in acts of the summer school “Remote sensing of terrestrial resources”, Tarbes, 21/08-20/09 1973, pp. 5–7). “Remote sensing means any process of acquiring information about an object without the sensor or measuring device being in contact with that object [...]. However, we must be careful not to confuse the methodology of remote sensing with that, well known, of aerial photography. Indeed, in the case of photointerpretation of aerial photographs, the interpreter is, in a way, placed before the fait accompli. It has a shot that has been generally recorded in optimal technical conditions (season, purity of the atmosphere, quality of shooting) to obtain an excellent result, in the photographic sense. This information may be very clear; they

16

TORUS 2 – Toward an Open Resource Using Services

may also be masked more or less by useless elements; finally, it is possible that they are not there at all, and that their search is therefore in vain. Conversely, the goal of remote sensing is to promote tools and data that simplify, while making the work of users more efficient. From then on, it becomes clear that this supposes a thorough knowledge of the objects of interest, their properties vis-à-vis the physical phenomena that will serve as a support for the transfer of information to the sensor, and finally the technical possibilities of the instruments capable of collecting this information and translating it into assimilable terms by the interpreter [...]. In these terms, remote sensing appears to be an extremely general technique, in fact its historical development and its main applications have made it besides other sciences, other techniques, which, although proceeding by “teledetection” are not part of the subject that will be treated in this course (just think of astronomy, the sounding of the seabed by sonar, gravimetry techniques, aeromagnetism etc.). The physical limits of remote sensing are essentially related to the choice made among the various phenomena likely to transfer information from the object to the sensor: – electromagnetic radiation; – fields of electric, magnetic or gravitational forces; – acoustic or mechanical vibrations and so on. In practice, we will consider only the electromagnetic radiation, the other phenomena not being compatible with the distance and the vacuum which separates the object to be measured and the sensor. Of all the possible applications of remote sensing, only those related to the study of the Earth’s natural resources and the impacts of human societies will be considered here. With these restrictions, the remote sensing technique thus includes the study of the terrestrial surface with regard to electromagnetic radiations whose wavelengths can go from the ultraviolet to the microwaves, the study of the technical means allowing these radiations to be received and the study of data analysis methods received in order to extract the elaborate information sought”.

A Brief History of Remote Sensing

17

Canadian Center of Remote Sensing – Government of Canada “Remote sensing is the science (and to some extent, art) of acquiring information about the Earth’s surface without actually being in contact with it. This is done by sensing and recording reflected or emitted energy and processing, analyzing, and applying that information. In much of remote sensing, the process involves an interaction between incident radiation and the targets of interest. This is exemplified by the use of imaging systems where the following seven elements are involved. Note, however that remote sensing also involves the sensing of emitted energy and the use of non-imaging sensors.

1) energy source or illumination (A) – the first requirement for remote sensing is to have an energy source which illuminates or provides electromagnetic energy to the target of interest; 2) radiation and the atmosphere (B) – as the energy travels from its source to the target, it will come in contact with and interact with the atmosphere it passes through. This interaction may take place a second time as the energy travels from the target to the sensor; 3) interaction with the target (C) – once the energy makes its way to the target through the atmosphere, it interacts with the target depending on the properties of both the target and the radiation;

18

TORUS 2 – Toward an Open Resource Using Services

4) recording of energy by the sensor (D) – after the energy has been scattered by, or emitted from the target, we require a sensor (remote – not in contact with the target) to collect and record the electromagnetic radiation; 5) transmission, reception and processing (E) – the energy recorded by the sensor has to be transmitted, often in electronic form, to a receiving and processing station where the data are processed into an image (hardcopy and/or digital); 6) interpretation and analysis (F) – the processed image is interpreted, visually and/or digitally or electronically, to extract information about the target which was illuminated; 7) application (G) – the final element of the remote sensing process is achieved when we apply the information we have been able to extract from the imagery about the target in order to better understand it, reveal some new information or assist in solving a particular problem.” A Geographer (M. Robin, 1995, “Remote Sensing”, Nathan, pp. 5–8) “The purpose of remote sensing is to provide landscape information in the form of image data, using electromagnetic radiation as the vehicle for this information. The type and quality of image data obtained on the landscape is a function of both the type of sensor used and the complexity of the landscape [...] This aspect needs the attention of the geographer who wants relevant information on the landscape. Everett and Simonett in 1976 (Colwell, 1983) have attempted to formalize this aspect by an expression: S = f (I, E) where

S: information that the geographer manipulates, I: information obtained by the sensor, E: information from the landscape.

More specifically, the information I obtained by the sensor is a function of the spectral resolution (Rs), the radiometric resolution (Rr), the spatial resolution (Sr), the temporal resolution (Rt) and the number of channels used (C): I = f (Rs, Rr, Sr, Rt, C) Information E emanating from the landscape is a function of the number of classes of landscape or land use (Cp), the size and arrangement of these classes

A Brief History of Remote Sensing

19

(T, notion of texture and structure), kinetics of these classes (V, temporal variability), atmospheric constraints that can alter the information received by the sensor (A) and other aspects related to the originality of one landscape with respect to another (O), is: E = f (Cp, T, V, A, O) All branches of geography are likely to be enriched to one degree or another by the use of remote sensing. Some themes even require, to be well apprehended, the use of these techniques either because the scale required does not have to be necessarily fine (space oceanography, meteorology, inventory with medium or small scale, cartography of landscape devoid of any other source information etc.), or that the landscape lends itself well to a satellite approach: loose structure, well discriminable by imagery (large plot, forestry etc.). On the other hand, certain fields of application are still beyond the reach of remote sensing techniques in the current state of affairs (analysis scale unsuitable for certain urban themes, for example) or still belong to the field of research.” CNES|ONERA|IGN (2012) “Remote sensing is the set of techniques used to determine remotely the properties of objects, natural or artificial, from the radiation they emit or reflect.” 1.6. Notes NOTE 1.– “An Unmanned Aerial Vehicle (UAV), commonly known as a drone, is an aircraft without a human pilot aboard. UAVs are a component of an unmanned aircraft system (UAS); which include a UAV, a ground-based controller, and a system of communications between the two. The flight of UAVs may operate with various degrees of autonomy: either under remote control by a human operator or autonomously by on-board computers” (Wikipedia https://en.wikipedia.org/wiki/Unmanned_aerial_vehicle).

NOTE 2.– The Afternoon Constellation – A-Train – from NASA (https://atrain.nasa.gov). NOTE 3.– Copernicus is the new name for the Global Monitoring for Environment and Security program, previously known as GMES https://www.esa.int/Our_Activities/ Observing_the_Earth/Copernicus/Overview3). Available at: https://eos.com/platform/

20

TORUS 2 – Toward an Open Resource Using Services

NOTE 4.– “Mainframe computers dominated in the 1960s and early 1970s. Basically these machines had 32-bit or higher word capacity that allowed them to do large numerical calculations. They were very fast and by the 1970s they had network capability that permitted terminals to link into them. In comparison the 32-bit microcomputer of today was in its infancy and had only an 8-bit word size that allowed it to handle only numbers between 0 and 255, not a machine on which to calculate large numbers. The primary output device on the mainframe computer was the high-speed line printer, designed mainly for printing numerical information. It was definitely not built for graphic output”, (https://www.oneonta.edu/faculty/baumanpr/geosat2/RS%20History%20II/RS-HistoryPart-2.html).

NOTE 5.– The images of meteorological satellites were rapidly integrated by the major national and international meteorological agencies as early as the 1970s; agencies that had facilities for data processing.

NOTE 6.– The story of GoogleEarth is interesting as it perfectly illustrates the close links between the electronic component manufacturers, the game developers in search of 3D graphics power, the media capture and the movie industry, to become one of Google’s everyday life tools. “The core technology behind Google Earth was originally developed at Intrinsic Graphics in the late 1990s. At the time, the company was developing 3D gaming software libraries. As a demo of their 3D software, they created a spinning globe that could be zoomed into, similar to the Powers of Ten film. The demo was popular, but the board of Intrinsic wanted to remain focused on gaming, so in 1999, they created Keyhole, Inc., headed by John Hanke. Keyhole then developed a way to stream large databases of mapping data over the internet to client software, a key part of the technology, and acquired patchworks of mapping data from governments and other sources. The product, called “Keyhole EarthViewer”, was sold on CDs for use in fields such as real estate, urban planning, defense, and intelligence; users paid a yearly fee for the service. Despite making a number of capital deals with Nvidia and Sony, the small company was struggling to make payroll, and employees were leaving. Fortunes for the company changed in early 2003 when CNN received a discount for the software in exchange for placing the Keyhole logo on-air whenever the map was used. Keyhole did not expect it would amount to more than brief 5 or 10 second pre-recorded animation clips, but it was used extensively by Miles O’Brien live during the 2003 invasion of Iraq, allowing CNN and millions of viewers to follow the progress of the war in a way that had never been seen before. Public interest in the software exploded and Keyhole servers were not able to keep up with demand. Keyhole was soon contacted by the Central Intelligence Agency’s venture capital firm, In-Q-Tel, and the National Geospatial-Intelligence Agency, for use with defense mapping databases, which gave Keyhole a much-needed cash infusion. Intrinsic Graphics was sold in 2003 to Vicarious Visions after its gaming libraries did not sell well, and its core group of engineers and management transitioned to Keyhole with Hanke remaining at the head. At the time, Google was finding that over 25% of its searches were of a geospatial character, including searches for maps and directions. In October 2004, Google acquired Keyhole as part of a strategy to better serve its users” (Wikipedia https://en.wikipedia.org/wiki/Google_Earth).

A Brief History of Remote Sensing

21

1.7. References Alain R., François R. and Paul G. (2007). Trente-cinq ans d’observations spatiales de la Terre : de la photographie à l’électromagnétrie, Revue Télédétection, 7, (1–4), 65–88. CNES (2012). Satellite Imagery: From Acquisition Principles to Processing of Optical Images for Observing the Earth. Cépaduès Éditions, France. Ditter R., Haspel M., Jahn M., Kollar I., Siegmund A., Viehrig K., Volz D. and Siegmund A. (2012). Geospatial technologies in school – theoretical concept and practical implementation in K-12 schools. In International Journal of Data Mining, Modelling and Management (IJDMMM): FutureGIS: Riding the Wave of a Growing Geospatial Technology Literate Society, Vol. X. Hansen M.C., Potapov P.V., Moore R., Hancher M., Turubanova S.A., Tyukavina A., Thau D., Stehman S.V., Goetz S.J., Loveland T.R., Kommareddy A., Egorov A., Chini L., Justice C.O. and Townshend J.R.G. (2013). High-Resolution Global Maps of 21st-Century Forest Cover Change. Science 15 November, 342, (6160), 850–853. DOI: 10.1126/science. 1244693. Available at: https:// earthenginepartners.appspot.com/science2013-global-forest Stork E.J., Sakamoto S.O., and Cowan R.M. (1999). The integration of science explorations through the use of earth images in middle school curriculum. Proc. I IEEE Transactions on Geoscience and R emote Sensing 3 7 , 1801–1817.

2 Physics of RS

2.1. Introduction Although there are several books and manuals devoted to Remote Sensing (see for example, [JEN 07, LIL 94, SAB 78]) and to the physics involved in this discipline, (see for example, [ELA 06, COA 14, MAN 16]), here it is worth giving a brief introduction and summary of these topics, in order to better understand the research activities carried out in the TORUS project. This chapter is inspired by the work of Professor Paul R. Baumann [BAU 10] and follows the work of Elachi and van Zyl [ELA 06] in presenting the topics. 2.2. Remote sensing Remote sensing is the art and science of recording, measuring and analyzing information about a phenomenon from a distance. In order to study large areas of the Earth’s surface, scientists use devices known as remote sensors. These sensors are mounted on platforms such as helicopters, planes and satellites, that make it possible for the sensors to observe the Earth from above. Two types of sensors exist, namely passive and active. A passive sensor system needs an external energy source (Figure 2.1). In most cases, this source is the sun. These sensors generally detect reflected and emitted energy wavelengths from a phenomenon. An active sensor system provides its own energy source. As an example, a radar (radio detection and ranging) sensor sends out electromagnetic waves, in the radio or microwave domain, and records the reflection waves coming back from the surface. Passive systems are much more common than active systems.

Chapter written by Luca T OMASSETTI.

TORUS 2 – Toward an Open Resource Using Services: Cloud Computing for Environmental Data, First Edition. Edited by Dominique Laffly. © ISTE Ltd 2020. Published by ISTE Ltd and John Wiley & Sons, Inc.

24

TORUS 2 – Toward an Open Resource Using Services

Figure 2.1. Remote sensing with passive sensors

Electromagnetic energy is the means by which information is transmitted from an object to a sensor. Information could be encoded in the frequency content, intensity or polarization of the electromagnetic wave. The information is propagated by electromagnetic radiation at the velocity of light from the source, directly through free space or indirectly by reflection, scattering and re-radiation to the sensor. The interaction of electromagnetic waves with natural surfaces and atmospheres is strongly dependent on the frequency of the waves. Waves in different spectral bands tend to excite different interaction mechanisms such as electronic, molecular or conductive mechanisms. The electromagnetic spectrum is divided into a number of spectral regions. Most sensors record information about the Earth’s surface by measuring the transmission of energy from the surface, in different portions of the electromagnetic (EM) spectrum (Figure 2.2). The radio band covers the region of wavelengths longer than 10 cm (frequency less than 3 GHz). This region is used by active radio sensors such as imaging radars, altimeters and sounders, and, to a lesser extent, passive radiometers. The microwave band covers the neighboring region, down to a wavelength of 1 mm (300 GHz frequency). In this region, most of the interactions are governed by molecular rotation, particularly at the shorter wavelengths. This region is mostly used by microwave radiometers/spectrometers and radar systems. The infrared band covers the spectral region from 1 mm to 0.7 μm. This region is subdivided into subregions called submillimeter, far infrared, thermal infrared and near infrared. In this region, molecular rotation and vibration play important roles. Imagers, spectrometers, radiometers, polarimeters and lasers are used in this region

Physics of RS

25

for remote sensing. The same is true in the neighboring region, the visible region (0.7 – 0.4 μm) where electronic energy levels start to play a key role.

Figure 2.2. Electromagnetic spectrum. For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

In the next region, the ultraviolet (0.4 μm to 300 Å), electronic energy levels play the main role in wave-matter interaction. Ultraviolet sensors have been mainly used to study planetary atmospheres or to study surfaces with no atmosphere because of the opacity of gases at these short wavelengths. X-rays (300 Å to 0.3 Å) and gamma rays (shorter than 0.3 Å) have been used to an even lesser extent because of atmospheric opacity. Their use has been limited to low-flying aircraft platforms or to the study of planetary surfaces with no atmosphere (e.g. the Moon). As the Earth’s surface varies in nature, the transmitted energy also varies. This variation in energy allows images of the surface to be created. Human eyes see this variation in energy in the visible portion of the EM spectrum. Sensors detect variations in energy in both the visible and non-visible areas of the spectrum. Energy waves in certain sections of the EM spectrum pass easily through the atmosphere, while other types do not. The ability of the atmosphere to allow energy to pass through it is referred to as its transmissivity, and varies with the wavelength/type of radiation. The gases that comprise our atmosphere absorb energy in certain wavelengths while allowing energy with differing wavelengths to pass through. The areas of the EM spectrum that are absorbed by atmospheric gases such as water vapor, carbon dioxide and ozone are known as absorption bands (Figure 2.3).

26

TORUS 2 – Toward an Open Resource Using Services

In Figure 2.3 (Bottom), absorption bands (shown in brown) are represented by a low transmission value that is associated with a specific range of wavelengths. Trying to obtain remotely sensed imagery in the absorption bands is nearly impossible; thus, sensors are generally designed not to record information in these portions of the spectrum.

Figure 2.3. Top: types of electromagnetic waves that pass through the Earth’s atmosphere. Also known as atmospheric electromagnetic transmittance or opacity. Note that the atmosphere is “transparent” to (lets through) visible light, but “opaque” to (absorbs and stops) infrared radiation. Bottom: Transmission versus wavelength; atmospheric windows (in green) and absorption bands (in brown). For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

In contrast to the absorption bands, there are areas of the EM spectrum (shown in green in Figure 2.3 (Bottom) and described in Table 2.1) where the atmosphere is transparent (little or no absorption of energy) to specific wavelengths.

Physics of RS

27

These wavelength bands are known as atmospheric “windows” since they allow the energy to easily pass through the atmosphere to Earth’s surface. It is in these windows that sensors are used to gather information about Earth phenomena. Region name Gamma ray

Wavelength < 0.03 nm

X-ray

0.03 – 0.4 nm

Ultraviolet

0.03 – 0.4 μm

Photographic ultraviolet 0.3 – 0.4 μm 0.4 – 0.7 μm

Visible

Near- and mid-infrared 0.7 – 3.0 μm Thermal infrared

< 3 – 100 μm

Micowave or radar

0.1 – 100 cm

Radio

> 100 cm

Comments Entirely absorbed by the Earth’s atmosphere and not available for remote sensing. Entirely absorbed by the Earth’s atmosphere and not available for remote sensing. Wavelengths from 0.03 to 0.3 micrometers absorbed by ozone in the Earth’s atmosphere. Available for remote sensing the Earth, and can be imaged with cameras and sensors. Available for remote sensing the Earth, and can be imaged with cameras and sensors. Available for remote sensing the Earth, and can be imaged with cameras and sensors. Available for remote sensing the Earth. This wavelength cannot be captured by film cameras. Sensors are used to image this wavelength band. Longer wavelengths of this band can pass through clouds, fog and rain. Images using this band can be made with sensors that actively emit microwaves. Not normally used for remote sensing the Earth.

Table 2.1. Major regions of the electromagnetic spectrum and their availability for remote sensing

Most remote sensing instruments on aircraft or space-based platforms operate in one or more of these windows by making their measurements with detectors tuned to specific frequencies (wavelengths), that pass through the atmosphere. When a remote sensing instrument has a line-of-sight with an object that is reflecting sunlight or emitting heat, the instrument collects and records the radiant energy. While most remote sensing systems are designed to collect reflected energy, some sensors, especially those on meteorological satellites, directly measure absorption phenomena, such as those associated with carbon dioxide (CO2 ) and other gases. The atmosphere is nearly opaque to EM energy in part of the mid-IR and all of the far-IR regions. In the microwave region, by contrast, most of this radiation moves through unimpeded, so radar waves reach the surface (although weather radars are able to detect clouds and precipitation because they are tuned to observe backscattered radiation from liquid and ice particles). Traditional aerial photographs were black and white pictures based on camera and film technology. Such photographs related to one region of the EM spectrum. Satellite

28

TORUS 2 – Toward an Open Resource Using Services

images are generally captured using sensors and digital technology. A sensor often records simultaneously in several different regions of the spectrum, creating multiimages taken at the same time. The regions of the spectrum scanned are called bands. A band is identified in nanometers (nm). For example, an image or band scanned in the visible light region of the spectrum would be identified as 400 – 700 nm. With multi-bands, various color composite images can be created. Furthermore, bands of various widths on the spectral scale can be recorded. A single band image shows features in various gray tones; however, if several images are combined, they can form a color composite. Figure 2.4(B) is a true color composite image of Charleston, South Carolina. A true color composite is based on using the red, green and blue portions of the visible region of the EM spectrum. This type of composite relates to what the human eye would see whether a person was on a satellite or aircraft looking down at the Earth. Other band combinations form false color composites. Figures 2.4(A), (C) and (D) represent two different false color composites. A false color composite generally enhances certain features on an image, features that might not be as apparent on a true color composite. Using different color composites is one way that a remote sensing specialist detects features on the Earth.

Figure 2.4. Color composites. For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

Physics of RS

29

2.3. Fundamental properties of electromagnetic waves The behavior of electromagnetic waves in free space is governed by Maxwell’s equations, which below for reference: ∂B , ∂t ∂D ∇×H= − + J, ∂t ∇×E= −

[2.1] [2.2]

B = μ0 μr H,

[2.3]

D = 0 r E,

[2.4]

∇ · E = 0,

[2.5]

∇ · B = 0,

[2.6]

where E is the eletric vector, D is the displacement vector, H is the magnetic vector and B is the induction vector; μ0 , 0 are the permeability and permittivity of vacuum, and μr , r are the relative permeability and permittivity. Maxwell’s concept of electromagnetic waves is that a smooth wave motion exists in the magnetic and electric force fields. In any region in which there is a temporal change in the electric field, a magnetic field appears automatically, in that same region, as a conjugal partner and vice versa. This is expressed by the above coupled equations. 2.3.1. Wave equation and solution In homogeneous, isotropic and nonmagnetic media, Maxwell’s equations can be combined to derive the wave equation, in the case of a sinusoidal field: ∇2 E +

ω2 E=0 c2r

[2.7]

1 c =√ μ 0  0 μr  r μr  r

[2.8]

where cr = √

Usually, μr = 1 and r varies from 1 to 80 and is a function of the frequency. The solution for the differential equation [2.7] is given by: E = Aei(kr−ωt+φ)

[2.9]

where A is the wave amplitude, ω is the angular frequency, φ is the phase, and k √ is the wave vector in the propagation medium (k = 2π r /λ, λ = 2πc/ω is the

30

TORUS 2 – Toward an Open Resource Using Services

wavelength, c is the speed of light in vacuum). The wave frequency ν is defined as ν = ω/2π. Remote sensing instruments exploit different aspects of the solution to the wave equation, in order to learn more about the properties of the medium from which the radiation is being sensed. For example, the interaction of electromagnetic waves with natural surfaces and atmospheres is strongly dependent on the frequency of the waves. This will manifest itself in changes in the amplitude (the magnitude of A in equation [2.9]) of the received wave, as the frequency of the observation is changed. This type of information is recorded by multispectral instruments such as the LandSat Thematic Mapper and the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER). In other cases, we can infer information about the electrical properties and geometry of the surface, by observing the polarization (the vector components of A in equation [2.9]) of the received waves. This type of information is recorded by polarimeters and polarimetric radars. Doppler lidars and radars, conversely, measure the change in frequency between the transmitted and received waves, in order to infer the velocity with which an object or medium is moving. This information is contained in the angular frequency ω of the wave shown in equation [2.9]. The quantity kr − ωt + φ in equation [2.9] is known as the phase of the wave. This phase changes by 2π every time the wave moves through a distance equal to the wavelength λ. Measuring the phase of a wave therefore provides an extremely accurate way to measure the distance that the wave actually traveled. Interferometers exploit this property of the wave to accurately measure differences in the path length between a source and two collectors, allowing us to significantly increase the resolution with which the position of the source can be established. 2.3.2. Quantum properties of electromagnetic radiation At very short wavelengths, Maxwell’s formulation of electromagnetic radiation fails to account for certain significant phenomena that occur when the wave interacts with matter. In this case, a quantum description is more appropriate. The electromagnetic energy can be presented in a quantized form, as bursts of radiation with a quantized radiant energy Q, which is proportional to the frequency ν: Q = hν

[2.10]

where h = 6.626 × 10−34 J s, is the Planck’s constant. The radiant energy carried by the wave is not delivered to a receiver as if it is spread evenly over the wave, as Maxwell had visualized, but is delivered on a probabilistic basis. The probability that a wave train will make full delivery of its radiant energy at some place along the wave is proportional to the flux density of the wave at that place. If a very large number of wave trains are coexistent, then the overall average effect follows Maxwell’s equations.

Physics of RS

31

2.3.3. Polarization, coherence, group and phase velocity, the Doppler effect An electromagnetic wave consists of a coupled electric and magnetic force field. In free space, these two fields are at right angles to each other and transverse to the direction of propagation. The direction and magnitude of only one of the fields (usually the electric field) is sufficient to completely specify the direction and magnitude of the other field, using Maxwell’s equations. The polarization of the electromagnetic wave is contained in the elements of the vector amplitude A of the electric field in equation [2.9]. For a transverse electromagnetic wave, this vector is orthogonal to the direction in which the wave is propagating, and therefore, we can describe completely the amplitude of the electric field by writing A as a two-dimensional complex vector: ˆ + av ei δv v A = ah ei δh h ˆ

[2.11]

ˆ and v where h ˆ are the horizontal and vertical orthogonal basis vectors, respectively. The polarization states of the incident and reradiated waves play an important role in remote sensing. They provide an additional information source (in addition to the intensity and frequency) that may be used to study the properties of the radiating or scattering object. For example, at an incidence angle of 37◦ from vertical, an optical wave, polarized perpendicular to the plane of incidence, will reflect about 7.8% of its energy from a smooth water surface, whereas an optical wave polarized in the plane of incidence, will not reflect any energy from the same surface. All the energy will penetrate into the water; this is the Brewster effect. The coherence, group and phase velocity and the Doppler effect, although relevant in remote sensing, are beyond the scope of this book. A detailed discussion can be found in several Physics and Remote Sensing textbooks, for example [ELA 06, COA 14, MAN 16]. 2.4. Radiation quantities A number of quantities are commonly used to characterize electromagnetic radiation and its interaction with matter. These are briefly described below and summarized in Table 2.2. – Radiant energy. The energy carried by an electromagnetic wave. It is a measure of the capacity of the wave to do work by moving an object by force, heating it or changing its state. The amount of energy per unit volume is called radiant energy density.

32

TORUS 2 – Toward an Open Resource Using Services

– Radiant flux. The time rate at which radiant energy passes a certain location. It is closely related to the wave power, which refers to the time rate of doing work. The term flux is also used to describe the time rate of flow of quantized energy elements, such as photons. Then, the term photon flux is used. – Radiant flux density. It corresponds to the radiant flux intercepted by a unit area of a plane surface. The density for flux incident upon a surface is called irradiance. The density for flux leaving a surface is called exitance or emittance. – Solid angle. The solid angle Ω subtended by area A on a spherical surface is equal to the area A, divided by the square of the radius of the sphere (see Figure 2.5). – Radiant intensity. The radiant intensity of a point source in a given direction is the radiant flux per unit solid angle leaving the source in that direction. – Radiance. The radiant flux per unit solid angle leaving an extended source in a given direction per unit projected area in that direction (see Figure 2.5). If the radiance does not change as a function of the direction of emission, the source is called Lambertian. A piece of white matte paper, illuminated by diffuse skylight, is a good example of a Lambertian source. – Hemispherical reflectance. The ratio of the reflected exitance (or emittance) from a plane of material, to the irradiance on that plane. – Hemispherical transmittance. The ratio of the transmitted exitance, leaving the opposite side of the plane, to the irradiance. – Hemispherical absorptance. The flux density that is absorbed over the irradiance. The sum of the reflectance, transmittance and absorptance is equal to one. Quantity

Symbol

Radiant energy

Q

Radiant energy density

W

Radiant flux

Φ E (irradiance)

Radiant flux density

M (emittance)

Equation

Units J

W =

dQ dV

J/m3

Φ=

dQ dt

W

E, M =

dΦ dA

W/m2

Radiant intensity

I

I=

dΦ dΩ

W/steradian

Radiance

L

L=

dI dA cos θ

W/steradian m2

Hemispherical reflectance

ρ

ρ=

Mr E

Hemispherical absorptance

α

α=

Ma E

Hemispherical transmittance

τ

τ =

Mt E

Table 2.2. Radiation quantities

Physics of RS

33

Figure 2.5. Solid angle definition and concept of radiance (from [ELA 06])

2.4.1. Spectral quantities Any wave can be considered as being composed of a number of sinusoidal component waves or spectral components, each carrying a part of the radiant flux of the total wave form. The spectral band, over which these different components extend, is called the spectral width or bandwidth of the wave. The way in which the radiation quantities are distributed among the components of different wavelengths or frequencies is called the spectral distribution. All radiance quantities have equivalent spectral quantities that correspond to the density as a function of the wavelength or frequency. For instance, the spectral radiant flux Φ(λ) is the flux in a narrow spectral width around λ divided by the spectral width: Φ(λ) =

Flux in all waves in the band λ − Δλ to λ + Δλ 2Δλ

To obtain the total flux from a wave form covering the spectral band from λ1 to λ2 , the spectral radiant flux must be integrated over that band: ˆ Φ(λ1 to λ2 ) =

λ2

λ1

Φ(λ)dλ

[2.12]

34

TORUS 2 – Toward an Open Resource Using Services

2.4.2. Luminous quantities Luminous quantities are related to the ability of the human eye to perceive radiative quantities. The relative effectiveness of the eye in converting radiant flux of different wavelengths to visual response is called the spectral luminous efficiency V (λ). This is a dimensionless quantity that has a maximum of unity at about 0.55 μm and covers the spectral region from 0.4 to 0.7 μm. V (λ) is used as a weighting function in relating radiant quantities to luminous quantities. For instance, luminous flux Φ is related to the radiant spectral flux Φe (λ) by: ˆ Φv = 680

0

∞

Φe (λ)V (λ)dλ

[2.13]

where the factor 680 is to convert from radiant flux units (watts) to luminous flux units (lumens). Luminous quantities are also used in relation to sensors, other than the human eye. These quantities are usually referenced to a standard source, with a specific blackbody temperature. For instance, standard tungsten lamps operating at temperatures between 3200 K and 2850 K are used to test photoemissive tubes 2.5. Generation of electromagnetic waves Electromagnetic radiation is generated by transformation of energy from other forms such as kinetic, chemical, thermal, electrical, magnetic or nuclear. A variety of transformation mechanisms lead to electromagnetic waves over different regions of the electromagnetic spectrum. In general, the more organized (as opposed to random) the transformation mechanism is, the more coherent (or narrower in spectral bandwidth) the generated radiation is. Radio frequency waves are usually generated by periodic currents of electric charges in wires, electron beams or antenna surfaces. If two short, straight metallic wire segments are connected to the terminals of an alternating current generator, electric charges are moved back and forth between them. This leads to the generation of a variable electric and magnetic field near the wires and, to the radiation of an electromagnetic wave at the frequency of the alternating current. This simple radiator is called a dipole antenna. At microwave wavelengths, electromagnetic waves are generated using electron tubes that use the motion of high-speed electrons in specially designed structures, to generate a variable electric/magnetic field, which is then guided by waveguides to a radiating structure. At these wavelengths, electromagnetic energy can also be generated by molecular excitation, as is the case in masers. Molecules have different

Physics of RS

35

levels of rotational energy. If a molecule is excited, by some means, from one level to a higher one, it could drop back to the lower level by radiating the excess energy as an electromagnetic wave. Higher-frequency waves in the infrared and the visible spectra are generated by molecular excitation (vibrational or orbital), followed by decay. The emitted frequency is exactly related to the energy difference between the two energy levels of the molecules. The excitation of the molecules can be achieved by a variety of mechanisms such as electric discharges, chemical reactions or photonic illumination. Molecules in the gaseous state tend to have well-defined, narrow emission lines. In the solid phase, the close packing of atoms or molecules distorts their electron orbits, leading to a large number of different characteristic frequencies. In the case of liquids, the situation is compounded by the random motion of the molecules relative to each other. Lasers use the excitation of molecules and atoms and the selective decay between energy levels to generate narrow-bandwidth electromagnetic radiation over a wide range of the electromagnetic spectrum, ranging from UV to the high submillimeter. Heat energy is the kinetic energy of random motion of the particles of matter. The random motion results in excitation (electronic, vibrational or rotational) due to collisions, followed by random emission of electromagnetic waves during decay. Due to its random nature, this type of energy transformation leads to emission over a wide spectral band. If an ideal source (called a blackbody) transforms heat energy into radiant energy with the maximum rate permitted by thermodynamic laws, then the spectral emittance is given by Planck’s formula as: S(λ) =

2πhc2 1 λ5 ech/λkT − 1

[2.14]

where h is Planck’s constant, k is the Boltzmann constant, c is the speed of light, λ is the wavelength and T is the absolute temperature in degrees Kelvin. Figure 2.6 shows the spectral emittance of a number of blackbodies, with temperatures ranging from 2000 K (temperature of the sun’s surface) to 300 K (temperature of the Earth’s surface). The spectral emittance is maximum at the wavelength given by: λm =

a T

[2.15]

where a = 2898 μm K. The total energy emitted over the whole spectrum is given by the Stefan-Boltzmann law: S = σT 4

[2.16]

36

TORUS 2 – Toward an Open Resource Using Services

where σ = 5.669 × 10−8 Wm−2 K−4 . Thermal emission is usually unpolarized and extends through the total spectrum, particularly at the low-frequency end. Natural bodies are also characterized by their spectral emissivity (λ), which expresses the ability to emit radiation due to thermal energy conversion, relative to a blackbody with the same temperature. The properties of this emission mechanism will not be discussed in this chapter but can be found in [ELA 06], for example.

Figure 2.6. Spectral radiant emittance of a blackbody at various temperatures (from [ELA 06])

Physics of RS

37

Going to even higher energies, waves in the gamma ray regions are mainly generated in the natural environment by radioactive decay of uranium (U), thorium (Th) and potassium 40 (40 K). The radioisotopes found in nature, 238 U and 232 Th, are long-lived alpha emitters and parents of individual radioactive decay chains. Potassium is found in almost all surfaces of the Earth, and its isotope 40 K, which makes up 0.12% of natural potassium, has a half-life of 1.3 billion years. 2.6. Detection of electromagnetic waves The radiation emitted, reflected or scattered from a body, generates a radiant flux density in the surrounding space that contains information about the body’s properties. To measure the properties of this radiation, a collector is used, followed by a detector. The collector is a collecting aperture that intercepts part of the radiated field. In the microwave region, an antenna is used to intercept some of the electromagnetic energy. Examples of antennas include dipoles, an array of dipoles or dishes. In the case of dipoles, the surrounding field generates a current in the dipole, with an intensity proportional to the field intensity and a frequency equal to the field frequency. In the case of a dish, the energy collected is usually focused onto a limited area where the detector (or waveguide connected to the detector) is located. In the IR, visible and UV regions, the collector is usually a lens or a reflecting surface that focuses the intercepted energy onto the detector. Detection then occurs by transforming the electromagnetic energy into another form of energy such as heat, electric current or state change. Depending on the type of sensor, different properties of the field are measured. In the case of synthetic-aperture imaging radars, the amplitude, polarization, frequency and phase of the fields are measured at successive locations along the flight line. In the case of optical spectrometers, the energy of the field at a specific location is measured as a function of wavelength. In the case of radiometers, the main parameter of interest is the total radiant energy flux. In the case of polarimeters, the energy flux at different polarizations of the wave vector is measured. In the case of x-ray and gamma ray detection, the detector itself is usually the collecting aperture. As the particles interact with the detector material, ionization occurs, leading to light emission or charge release. Detection of the emitted light or generated current gives a measurement of the incident energy flux.

38

TORUS 2 – Toward an Open Resource Using Services

2.7. Interaction of electromagnetic waves with matter 2.7.1. Overview The interaction of electromagnetic waves with matter (e.g. molecular and atomic structures) calls into play a variety of mechanisms that are mainly dependent on the frequency of the wave (i.e. its photon energy) and the energy level structure of the matter. As the wave interacts with a certain material - be it gas, liquid or solid - the electrons, molecules and/or nuclei are put into motion (rotation, vibration or displacement), which leads to an exchange of energy between the wave and the material. This section gives just a simplified overview of the interaction mechanisms between waves and matter. Detailed discussions are out of the scope of this book. Atomic and molecular systems exist in certain stationary states with well-defined energy levels. In the case of isolated atoms, the energy levels are related to the orbits and spins of the electrons. These are called the electronic levels. In the case of molecules, there are additional rotational and vibrational energy levels that correspond to the dynamics of the constituent atoms relative to each other. Rotational excitations occur in gases where molecules are free to rotate. The exact distribution of the energy levels depends on the exact atomic and molecular structure of the material. In the case of solids, the crystalline structure also affects the energy level distribution. In the case of thermal equilibrium, the density of population Ni at a certain level i is proportional to (Boltzmann’s law): Ni ∼ e−Ei /kT

[2.17]

where Ei is the level energy, k is the Boltzmann constant and T is the absolute temperature. At absolute zero, all the atoms will be in the ground state. Thermal equilibrium requires that a level with higher energy be less populated than a level of lower energy. Let us assume that a wave of frequency ν is propagating in a material in which two of the energy levels i and j are such that: hν = Ej − Ei

[2.18]

This wave would then excite some of the population of level i to level j. In this process, the wave loses some of its energy and transfers it to the medium. The wave energy is absorbed. The rate pij of such an event happening, also called transition probability, is equal to: pij = Bij ν

[2.19]

Physics of RS

39

where n u is the wave energy density per unit frequency and Bij is a constant determined by the atomic (or molecular) system. Once excited to a higher level by absorption, the atoms may return to the original lower level directly, by spontaneous or stimulated emission, and in the process, they emit a wave at frequency ν, or they could cascade down to intermediate levels, and in the process, emit waves at frequencies lower than ν. Spontaneous emission could occur any time an atom is in an excited state, independent of the presence of an external field. The rate of downward transition from level j to level i is given by: pji = Aji

[2.20]

where Aji is the characteristic of the pair of energy levels involved. Stimulated emission corresponds to downward transition, which occurs as a result of the presence of an external field with the appropriate frequency. In this case, the emitted wave is in phase with the external field and will add energy to it coherently. This results in an amplification of the external field and energy transfer from the medium to the external field. The rate of downward transition is given by: pji = Bji ν

[2.21]

The relationships between Aji , Bji and Bij are known as the Einstein relations; in case of non-degenerate states: Bji = Bij Aji =

8πhν 3 n3 Bji c3

[2.22] [2.23]

where n is the index of refraction of the medium. 2.7.2. Interaction mechanisms Starting from the highest spectral region used in remote sensing, gamma- and x-ray interactions with matter call into play atomic and electronic forces, such as the photoelectric effect (absorption of photon with ejection of electron), the Compton effect (absorption of photon with ejection of electron and radiation of lower-energy photon) and the pair production effect (absorption of photon and generation of an electron-positron pair). The photon energy in this spectral region is larger than 40 eV. This spectral region is mainly used to sense the presence of radioactive materials. In the ultraviolet region (photon energy between 3 eV and 40 eV), the interactions call into play electronic excitation and transfer mechanisms, with their associated

40

TORUS 2 – Toward an Open Resource Using Services

spectral bands. This spectral region is mostly used for remote sensing of the composition of the upper layers of the Earth and planetary atmospheres. In the visible and near infrared (energy between 0.2 eV and 3 eV), vibrational and electronic energy transitions play the key role. In the case of gases, these interactions usually occur at well-defined spectral lines, which are broadened due to the gas pressure and temperature. In the case of solids, the closeness of the atoms in the crystalline structure leads to a wide variety of energy transfer phenomena with broad interaction bands. These include molecular vibration, ionic vibration, crystal field effects, charge transfer and electronic conduction, which are widely used in rock and vegetation remote sensing. In the mid-infrared region (8 – 14 μm), the Si–O fundamental stretching vibration provides diagnostics of the major types of silicates. This spectral region also corresponds to vibrational excitation in atmospheric gaseous constituents. In the thermal infrared, the emissions from the Earth’s and other planets’ surfaces and atmospheres are strongly dependent on the local temperature, and the resulting radiation is governed by Planck’s law. This spectral region provides information about the temperature and heat constant of the object under observation. In addition, a number of vibrational bands provide diagnostic information about the emitting object’s constituents. In the submillimeter region, a large number of rotational bands provide information about the atmospheric constituents. These bands occur all across this spectral region, making most planetary atmospheres completely opaque for surface observation. For some gases, such as water vapor and oxygen, the rotational band extends into the upper regions of the microwave spectrum. The interaction mechanisms in the lower-frequency end of the spectrum (ν < 20 GHz, λ > 1.5 cm) do not correspond to energy bands of specific constituents. Rather, they are collective interactions that result from electronic conduction and non-resonant magnetic and electric multipolar effects. As a wave interacts with a simple molecule, the resulting displacement of the electrons leads to the formation of an oscillating dipole that generates an electromagnetic field. This will result in a composite field, moving at a speed lower than the speed of light in vacuum. The effect of the medium is described by the index of refraction or the dielectric constant. In general, depending on the structure and composition of the medium, the dielectric constant could be anisotropic or could have a loss term that is a result of wave energy transformation into heat energy. In the case of an interface between two media, the wave is reflected or scattered, depending on the geometric shape of the interface. The physical properties of the interface and the dielectric properties of the two media are usually the major factors

Physics of RS

41

affecting the interaction of waves and matter in the microwave and radio frequency part of the spectrum. Thus, remote sensing in this region of the spectrum will mainly provide information about the physical and electrical properties of the object instead of its chemical properties, which are the major factors in the visible/infrared region, or its thermal properties, which are the major factors in the thermal infrared and upper microwave regions (see Table 2.3). Spectral region Main Interaction(s)

Remote sensing application(s)

Gamma rays, x-rays

Atomic processes

Mapping of radioactive materials

Ultraviolet

Electronic processes

Presence of H and He in atmospheres

Visible and near infrared

Electronic and vibration molecular Surface chemical composition, processes vegetation cover, biological properties

Mid-infrared

Vibrational, vibrational-rotational Surface and atmospheric chemical molecular processes composition

Thermal infrared

Thermal emission, vibrational and Surface and atmospheric temperature, rotational processes surface and atmospheric constituents

Microwave

Rotational processes, thermal emission, scattering, conduction

Radio frequency

Scattering, conduction, ionospheric Surface physical properties, sounding effect

Surface physical properties, atmospheric precipitation

Table 2.3. Wave-matter interaction mechanisms across the electromagnetic spectrum

In summary, a remote sensing system can be visualized (Figure 2.1) as a source of electromagnetic waves (e.g. the sun, a radio source) that illuminate the object being studied. An incident wave interacts with the object and the scattered wave is modulated by a number of interaction processes that contain the “fingerprint” of the object. In some cases, the object itself is the source and the radiated wave contains information about its properties. A part of the scattered or radiated wave is then collected by a collector, focused on a detector and its properties measured. An inverse process is then used to infer the properties of the object from the measured properties of the received wave. 2.8. Solid surfaces sensing in the visible and near infrared The visible and near-infrared regions of the electromagnetic spectrum have been the most commonly used in remote sensing of planetary surfaces. This is partially because this is the spectral region of maximum illumination by the sun and most widely available detectors (electro-optical and photographic). The sensor detects the

42

TORUS 2 – Toward an Open Resource Using Services

electromagnetic waves reflected by the surface and measures their intensity in different parts of the spectrum. By comparing the radiometric and spectral characteristics of the reflected waves, to the characteristics of the incident waves, the surface reflectivity is derived. This in turn is analyzed to determine the chemical and physical properties of the surface. The chemical composition and crystalline structure of the surface material has an effect on the reflectivity, because of the molecular and electronic processes that govern the interaction of waves with matter. The physical properties of the surface, such as roughness and slope, also affect the reflectivity, mainly due to geometric factors related to the source-surface-sensor relative angular configuration. Thus, information about the surface properties is acquired by measuring the modulation that the surface imprints on the reflected wave, by the process of wave-matter interactions, which will be discussed briefly in this section.

Figure 2.7. Sun illumination spectral irradiance; black-body emission at 6000 K, solar irradiation curve outside atmosphere and solar irradiation curve at sea level (showing absorption bands of chemicals)

By far the most commonly used source of illumination in the visible and near-infrared, is the sun. In the most simple terms, the sun emits approximately as a hot blackbody at 6000 K temperature. The solar illumination spectral irradiance at the Earth’s distance is shown in Figure 2.7. The total irradiance is measured at approximately 1370 W/m2 above the Earth’s atmosphere. This irradiance decreases as the square of the distance from the sun because of the spherical geometry. As the solar waves propagate through a planet’s atmosphere, they interact with the atmospheric constituents, leading to absorption in specific spectral regions,

Physics of RS

43

depending on the chemical composition of these constituents. Figure 2.7 shows the sun illumination spectral irradiance at the Earth’s surface. Strong absorption bands exist in the near infrared, particularly about 1.9, 1.4, 1.12, 0.95 and 0.76 μm. This is mainly due to the presence of water vapor (H2 O), carbon dioxide (CO2 ), and, to a lesser extent, oxygen (O2 ). In addition, scattering and absorption lead to a continuum of attenuation across the spectrum. Another important factor in visible and near-infrared remote sensing is the relative configuration of the sun-surface-sensor. Due to the tilt of the Earth’s rotation axis, relative to the plane of the ecliptic, the sun’s location in the sky varies as a function of the seasons and the latitude of the illuminated area. In addition, the ellipticity of the Earth’s orbit has to be taken into account. 2.8.1. Wave-surface interaction mechanisms When an electromagnetic wave interacts with a solid material, there are a number of mechanisms that affect the properties of the resulting wave. Some of these mechanisms operate over a narrow band of the spectral region, whereas others are wide-band and thus affect the entire spectrum from 0.3 to 2.5 μm. The narrow-band interactions are usually associated with resonant molecular and electronic processes. These mechanisms are strongly affected by the crystalline structure, leading to splitting, displacement and broadening of the spectral lines, into spectral bands. The wide-band mechanisms are usually associated with non-resonant electronic processes that affect the material index of refraction (i.e. velocity of light in the material). When an electromagnetic wave is incident on an interface between two materials (in the case of remote sensing, one of the materials is usually the atmosphere), some of the energy is reflected in the specular direction, some of it is scattered in all directions of the incident medium and some of it is transmitted through the interface (Figure 2.8). The transmitted energy is usually absorbed in the bulk material and either reradiated by electronic or thermal processes, or dissipated as heat. When the interface is very smooth, relative to the incident wavelength λ (i.e. λ >> interface roughness), the reflected energy is in the specular direction and the reflectivity is given by Snell’s law. The reflection coefficient is a function of the complex index of refraction n and the incidence angle θi . The expression of the reflection coefficient is given by: |Rh |2 =

sin2 (θi − θt ) , sin2 (θi + θt )

|Rv |2 =

tan2 (θi − θt ) tan2 (θi + θt )

[2.24]

44

TORUS 2 – Toward an Open Resource Using Services

for horizontally and vertically polarized incidence waves respectively. θt is the transmission angle given by Snell’s law: sin θi = n sin θt

[2.25]

Figure 2.8. Wave interaction with an interface (from [ELA 06])

In the case of normal incidence, the reflection coefficient becomes:     Nr + iNi − 1 n−1 = , Rh = Rv = R = n+1 Nr + iNi + 1 |R|2 =

(Nr − 1)2 + Ni2 (N r + 1)2 + Ni2

[2.26] [2.27]

where Nr and Ni are the real and imaginary parts of n respectively. In the real-world, most natural surfaces are rough, relative to the wavelength, and usually consist of particulates. Thus, scattering plays a major role and the particulates’ size distribution has a significant impact on the spectral signature. An adequate description of the scattering mechanism requires a rigorous solution of Maxwell’s equations, including multiple scattering. This is usually very complex, and a number of simplified techniques and empirical relationships are used. For instance, if the particles are small, relative to the wavelength, then Rayleigh’s law is applicable, and it adequately describes the scattering mechanism. In this case, the scattering cross section is a fourth power of the ratio a/λ (i.e. scattering is (a/λ)4 ) of the particles’ size, over the wavelength. The Rayleigh scattering explains the blue color of the sky; molecules and very fine dust scatter blue light about four times more

Physics of RS

45

than red light, making the sky appear blue. The direct sunlight, on the contrary, is depleted of the blue and therefore appears reddish. In the case of a particulate surface, the incident wave is multiply scattered, and some of the energy, that is reflected toward the incident medium, penetrates some of the particles. Thus, if the material has an absorption band, the reflected energy is depleted of energy in that band. Usually, as the particles get larger, the absorption features become more pronounced, even though the total reflected energy is decreased. This is commonly observed in measured spectra, and is the usual case in remote sensing in the visible and infrared regions. In the general case of natural surfaces, the reflectivity is modeled by empirical expressions. One such expression is the Minnaert law, which gives: B cos θs = B0 (cos θi cos θs )κ

[2.28]

where B is the apparent surface brightness, B0 is the brightness of an ideal reflector at the observation wavelength, κ is a parameter that describes darkening at zero phase angle, and θi and θs are the incidence and scattering angles, respectively. In the case of a Lambertian surface, κ = 1, thus equation [2.28] becomes B = B0 cos θi . 2.9. Radiometric and geometric resolutions Digital remote sensing deals with two types of resolution: radiometric (spectral) and geometric (spatial). Radiometric resolution is the number of levels that a sensor can record spectral information. Such information generally ranges from (0, 28 − 1) to (0, 216 − 1). These numbers are integer values (whole numbers). A single byte can hold one distinct integer value, ranging from (0,255). This value represents the degree of reflective or emitted energy recorded by a sensor, for a particular ground spot on the Earth’s surface. Although geographers take into account radiometric resolution when selecting imagery to study an environmental issue, they relate more to geometric resolution due to its spatial nature. Geometric resolution refers to the smallest amount of area on the Earth’s surface for which a sensor can record radiometric (spectral) information. Generally, this resolution is expressed in terms of a pixel (picture element). The pixel size of the Enhanced Thematic Mapper sensor on Landsat 7 is 30 m, which relates to an area of 30 m × 30 m on the Earth’s surface. In comparison, a sensor entitled AVHRR has a pixel of 1.1 km2 , while the panchromatic sensor on the QuickBird satellite possesses a 61 cm2 pixel size. Geometric resolutions vary greatly and are defined loosely as being low, moderate and high. The parameters associated with these designations change as finer

46

TORUS 2 – Toward an Open Resource Using Services

resolutions in imagery become available. Imagery identified as being of moderate resolution at one point in time might now be low resolution. An image with a one kilometer pixel size is viewed as being a low resolution image. Such an image provides a synoptic coverage of the Earth’s surface. Remote sensing imagery has many applications in cartography, land use and cover, agriculture, soils mapping, forestry, city planning, grassland management, archeology, military observations, meteorology, and geomorphology, among others. In order to use such imagery, we must have considerable knowledge about the Earth’s surface and a strong background in remote sensing data acquisition and analysis techniques. Since the Earth’s surface consists of a mosaic of environmental conditions, geographers, through their unique training in both the sciences and social sciences, are well qualified to undertake various remote sensing applications. 2.10. References [BAU 10] BAUMANN P.R., “Introduction to Remote Sensing”, 2010. [COA 14] C OAKLEY J R . J.A., YANG P., Atmospheric Radiation, John Wiley and Sons, Inc, New York, 2014. [ELA 06] E LACHI C., VAN Z YL J.J., Introduction to the Physics and Techniques of Remote Sensing, John Wiley and Sons, Inc, New York, 2006. [JEN 07] J ENSON J.R., Remote Sensing of the Environment: An Earth Resource Perspective, Prentice Hall, Pearson, London, 2007. [LIL 94] L ILLESAND T.M., K IEFER R.W., Remote Sensing and Image Interpretation, John Wiley and Sons, Inc, New York, 1994. [MAN 16] M ANOLAKIS D.G., L OCKWOOD R.B., C OOLEY T.W., Hyperspectral Imaging Remote Sensing, Cambridge University Press, Cambridge, 2016. [SAB 78] S ABINS F.F., Remote Sensing: Company US, 1978.

Principles and Interpretation, Freeman and

3 Image Quality

3.1. Introduction Imagine yourself on a frail boat in the open sea swept by a violent tropical storm and you draw – impassive, insensitive to the elements – straight lines so fine and regular that they could be printed with high definition… Not easy is it? One would marvel at such a feat. So, I invite you to be just as impressed when you watch a satellite image with very high spatial resolution – tens of centimeters – acquired in polar orbit or another at hectometric pixels but acquired from a geostationary orbit. This should not be as precise, as clean and as geometrically correct since the initial conditions of these images are comparable or even worse than those of the storm that was wobbling you at the beginning of this text. Furthermore, in space orbit, the physical conditions are extreme. There is no atmosphere to protect and regulate thermic balance (Figure 3.1), hence there is a strong deterioration of electronic equipment.

Figure 3.1. Thermic balance on the ground, in low orbit (200 km) and geostationary orbit (36,000 km) for black, white and yellow bodies. Modified from Laffly, 2017. For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip Chapter written by Dominique LAFFLY.

TORUS 2 – Toward an Open Resource Using Services: Cloud Computing for Environmental Data, First Edition. Edited by Dominique Laffly. © ISTE Ltd 2020. Published by ISTE Ltd and John Wiley & Sons, Inc.

48

TORUS 2 – Toward an Open Resource Using Services

“To experiment long-term effects of space exposure on materials, components and systems, NASA propose the Long Duration Exposure Facility program, or the LDEF. It was a school bus-sized cylindrical facility placed in low Earth orbit by Space Shuttle Challenger in April 1984. The original plan called for the LDEF to be retrieved in March 1985, but after a series of delays it was eventually returned to Earth by Columbia in January 1990. It successfully carried science and technology experiments for about 5.7 years, that have revealed a broad and detailed collection of space environmental data that has benefited NASA spacecraft designers to this day” (modified from Wikipedia 2020a). Figure 3.2 shows the same component before (left) and after (right) 69 months in space environment and illustrates why space needs specific electronic vacuum (cleanliness, outgassing, only radiative transfer, etc.), no in-orbit servicing/repairing, external envelope eroded by debris, atomic oxygen, specific thermal control and design. Also, in such conditions, we must remember that LANDSAT 5 was operational over 28 years…

Figure 3.2. LDEF prelaunch and postflight. For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

Remote sensing is above all an expression of great technical prowess and engineering. To see in detail from a distance, a telescope is necessary and the laws of the optics are formal: to decrease the resolution on the ground considerably increases the length of the focal. Hence, the choice to include mirrors in the Cassegrain telescope model which reduce the size and hold the instrument on a platform. Then, we have to leave the Earth to be placed in orbit, which is not simple and not only for reasons of celestial mechanics: there are few launchers available, the costs are prohibitive, the windows of fire are reduced and the waiting list is loaded where civilian Earth observation

Image Quality

49

missions are not given priority. Once in orbit – polar or geostationary – it is necessary to control in real time the platform on this orbit while integrating the variations of laces, rolls and pitches. Then, the sensor acquires each image in a few seconds because of the scanning acquisition technology, a few seconds during which the Earth continues to rotate and the platform wobbles in its orbit as well: at the equator, in 9 seconds – the time of acquisition of a SPOT image – the latitudinal distance traveled by the Earth is 4.16 km and the longitudinal distance of the satellite of 67.24 km; attitude differences deter the aim (imagine yourself looking at a target with big binoculars on your frail boat); the surface of the Earth is uneven while the image is acquired on a plane from where there are geometric deformations, to which are added those related to the angles of acquisition; optics and other prisms interfere with the signal already disturbed by the atmosphere; the quantization of the continuous signal in discrete values over a more or less extended interval deteriorates the contrasts (8 bits per pixel images from 1980 to 2000 to data at 12 bits per pixel today, i.e. 256 levels against 4,096); the sensors must be recalibrated daily to compensate for their drifting… In spite of all this, the images are geometrically precise and thematically contrasted, or – in addition to being so beautiful – the result of a multidisciplinary technological and scientific feat. We will discuss, in a synthetic way, the main aspects that allow a final rendering of quality. Because of the raw data acquired on the orbital to the user imagery (Figure 3.3) the way is long and difficult. In 1986, NASA introduced data level to identify image quality: – Level 0: reconstructed, unprocessed instrument and payload data at full resolution, with any and all communication artifacts (e.g. synchronization frames, communication headers, duplicate data) removed. – Level 1a: reconstructed, unprocessed instrument data at full resolution, timereferenced, and annotated with ancillary information, including radiometric and geometric calibration coefficients and georeferencing parameters (e.g. platform ephemeris) computed and appended but not applied to the level 0 data (or if applied, in a manner that level 0 is fully recoverable from level 1a data). – Level 1b: level 1a data that have been processed to sensor units (e.g. radar backscatter cross section, brightness temperature, etc.). Not all instruments have level 1b data, and level 0 data is not recoverable from level 1b data. – Level 2: derived geophysical variables (e.g. ocean wave height, soil moisture, ice concentration) at the same resolution and location as level 1 source data.

50

TORUS 2 – Toward an Open Resource Using Services

– Level 3: variables mapped on uniform space-time grid scales, usually with some completeness and consistency (e.g. missing points interpolated, complete regions mosaicked together from multiple orbits, etc.). – Level 4: model output or results from analyses of lower level data (i.e. variables that were not measured by the instruments but instead are derived from these measurements) (modified from Wikipedia 2020b).

Figure 3.3. From level 0 to level 1 (modified from Valorge 2012). For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

From the space to the end user, the way is controlled by the space and ground segments of the data acquisition system. Figure 3.4 illustrates these segments for Landsat 7; it is approximately similar for many systems such as SPOT, SENTINEL, Pleiades (see Note 1 in section 3.5), etc. The space segment – totally restricted to mission management office – controls: – the orbital geometry: orbit, satellite attitude and viewing directions; – the orbital sensor calibration: detector response, noise, amplification, quantization and modulation transfer modulation (MTF); – the orbital storage and transmission: compression, formatting and transmission. The ground segment control – also restricted to mission management office: – image preprocessing: decompression, unformatting; – image geometry: correction of imaging distortion, relief effect and over laying on a map; – image radiometry: equalization, deconvolution and denoising; – image storage and distribution: datacenter and network access.

Image Quality

51

Figure 3.4. Space and ground segments – control mission (modified from https://Landsat.usgs.gov/Landsat-7-data-users-handbook)

With Landsat, from 1972 to 2000s, the end user obtains an image at Level 1, with which they can apply geometric correction, georeferencing – and rarely orthorectification – data and calculate physical value and other indices. Today, all high spatial resolution space systems delivery directly images at Level 1c and then 2, 3 or 4 in GeoTIFF, JPEG2000 or HDF file format1. Images are orthorectified and transformed on top of atmosphere (TOA) reflectance – sometimes bottom of atmosphere (BOA) – with automatic processes including in the control mission ground segment. ESA Sentinel 2 collection is delivered automatically at Level 1c without a ground control point (Figure 3.5): “The goal of the geometric correction is to perform the temporal and spectral registration of all the images taken over by any target. To achieve this objective, the physical geometric model, which associates a viewing direction to any pixel, has to be refined. This physical geometric model combines position, attitude and date information, transformation matrices between different reference frames – satellite, instrument, focal planes and detectors – and for each pixel 1 70s, 80s and 90s file format wars are over today.

52

TORUS 2 – Toward an Open Resource Using Services

of each elementary detector of each band a viewing direction is defined. An automatic correlation processing between a reference band of the image to be refined and a reference image provides ground control points, allowing the online calibration of the viewing model and correcting for variations in attitude or position or thermoelastic deformation. This refined geometrical model is then applied to all the bands and is used to project them onto a cartographic reference frame. The reference image will be part of a worldwide georeferenced database of Sentinel-2 mono-spectral images that will be gradually built up from cloud free scenes. The georeferencing of all the images is performed through a global space-triangulation process using tie points between the different images and GCPs. Because of the parallax between odd and even detector modules, and between bands, the registration is sensitive to the Digital Elevation Model (DEM) used for the processing. A Shuttle Radar Topography Mission […] class DEM is necessary to reach the required performance. It is planned to use the SRTM DEM, complemented at higher and lower latitudes with DEMs such as the Canadian National DEM, the Greenland Geoscience Laser Altimeter System (GLAS) DEM […], the National Elevation Dataset (NED) […] and corrected in specific areas where the SRTM presents artefacts (e.g. over the Himalayas).” (ESA 2012). (Source: https://esamultimedia.esa.int/multimedia/publications/SP-1322_2/).

Geometric image quality requirements A priori accuracy of image location: 2 km max (3σ) Accuracy of image location: 20 m (3σ) Accuracy of image location: 12.5 m (3σ) Multitemporal registration: 3 m (2σ) for 10 m bands 6 m (2σ) for 20 m bands 18 m (2σ) for 60 m bands Multispectral registration for any pair of spectral bands: 3 m (3σ) for 10 m bands 6 m (3σ) for 20 m bands 18 m (3σ) for 60 m bands

Ground processing hypothesis No processing After image processing without GCPs After image processing with GCPs

After image processing with GCPs

After image processing with GCPs

Figure 3.5. ESA Sentinel 2 ground segment (see http://esamultimedia.esa.int/ multimedia/publications/SP-1322_2l). For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

Image Quality

53

“From April 2005 until December 2006 onboard Landsat 7 gyros performed extremely well”, and then NASA used imagery to define a world mosaic orthorectified reference system (Global Land Survey – GLS2000). This database and a collection of validated GCP are systematically used to automatically orthorectify all actual Landsat imagery by autocorrelation identification of GCP (Figure 3.6, in green and blue) and onboard an orbital model. And finally, “in 2016, the USGS started reorganizing the Landsat archive into a formal tiered data Collection structure. This data Collection structure ensures that Landsat Level-1 products provide a consistent archive of known data quality to support time series analyses and data ‘stacking’, while controlling continuous improvement of the archive and access to all data as they are acquired.” NASA specified that the “Global Land Survey (GLS) data sets were created [in collaboration with] the U.S. Geological Survey (USGS), and were designed to allow scientists and data users to have access to a consistent, terrain corrected, coordinated collection of data. Five epochs have been created, each using the primary Landsat sensor in use at the time: – GLS1975: uses Multispectral Scanner (MSS) data; – GLS1990: uses Landsat 4-5 Thematic Mapper (TM) data; – GLS2000: uses Landsat 7 Enhanced Thematic Mapper Plus (ETM+) data; – GLS2005: uses a combination of Landsat TM and ETM+ data; – GLS2010: uses a combination of Landsat TM and ETM+ data. The Landsat data incorporated into each GLS data set meets quality and cloud cover standards, and is processed to the following parameters: – Data Resolution (Reflective bands): 30 meters (Landsat MSS: 60 meters); – Data Format: GeoTIFF (Level-1 product), Full-resolution.jpg (Landsat Quick Look images); – Resampling: Cubic Convolution; – Projection: Universal Transverse Mercator (UTM); – Datum: WGS84. The Global Land Survey data sets are scene-based, and can be searched and downloaded at no charge from EarthExplorer or GloVis.”

54

TORUS 2 – Toward an Open Resource Using Services

Figure 3.6. NASA GLS products (modified from Rengarajan et al. 2015 and https:// Landsat.usgs.gov/global-Land-surveys-gls). For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

As we saw, the main remote sensing data agency deliver today high-quality level imagery with an integration of automatic processes in the production chain, and soon close to real time. It really is a new era of remote sensing where cloud computing plays a major role. Image quality control distinguished three main domains which we will detail here: – geometry: location and shape of each pixel on the ground; – radiometry: physical interpretation of measurements; – resolution: spatial frequencies, image’s representativeness with regard to the landscapes. 3.2. Image quality – geometry The geometry of a raw image acquired in orbit does not accurately represent the geometry of the Earth’s surface observed. Distortions are numerous. The causes are varied. They are related to the orbit, the sensor and the terrestrial relief…

Image Quality

55

In addition, the use of images is linked to cartographic representations of the Earth’s surface, i.e. a plane projection of an irregular ellipsoidal surface, which is not without geometric distortions (Figure 3.7).

Figure 3.7. Reference document, a map is the planimetric projection of the Earth’s surface. For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

To overcome these distortions, the main idea is to match each pixel of an image with the portion of the geographic space that it represents from an inverse localization model that directly gives latitudes and longitudes. The image acquisition is an intersection of the Earth’s surface with the viewing detector direction using a direct location model defined by: – T, pixel time-tagging; – S, the position of the satellite in a terrestrial reference (orbit in lat, long, z); – satellite attitude at T (pitch, roll, yaw); – viewing direction of detector at T (angle); – elevation of target pixel, relief of the observed Earth portion. T is given by synchronization with GPS atomic clock, S is given by Doris receiver (with Pleiades “Autonomous orbit determination is performed by a Doris receiver, which achieves an accuracy of about 1 m – on all three axes”). Satellite attitude is controlled by gyros and star tracker (Figure 3.8) with heavy satellites maneuvering capabilities, for example, the Attitude Determination and Control Subsystem (ACDS) of Spot 6 consists of: “the enhanced 3-axis stabilization attitude

56

TORUS 2 – Toward an Open Resource Using Services

control system, based on a set of 4 RW (Reaction Wheels) for fine-pointing with 3 MTQ (Magnetic Torquers) for off-loading. In case of very high agility requirements, the reaction wheels are replaced by Astrium patented CMGs (Control Moment Gyros) (see Note 2 in section 3.5). Attitude and orbit measurement is performed with a GPS and a Star Tracker (STR) for nominal operation, providing a pointing accuracy of up to 500 µrad (3σ) and a pointing knowledge of up to 30 µrad (3σ), depending mainly on STR accommodation and alignment accuracy. While standard precise attitude control is performed without the support of a gyro, an optional inertial measurement unit can be added for attitude control improvement. On-board orbit determination accuracies of < 10 m (1σ) are achieved, if the standard 1-frequency GPS design is applied. The 2-frequency GPS option provides on-board orbit determination accuracies of < 3 m (1σ). The SPOT-6 spacecraft is the first EO mission to utilize Sodern’s new generation HYDRA star tracker for guidance and navigation. With HYDRA, Sodern is offering a multi-head design which separates the optical head from the centralized processing unit. This eases the thermal control of the optical heads and allows continuous attitude measurement, whether the satellite is facing Earth, the Sun or other bright objects. Safe mode attitude sensing is based on a Magnetometer (MAG)/Sun Sensor (BASS) system or a Magnetometer/ Coarse Earth Sun Sensor (CESS) system. This provides two optional Safe mode attitude control principles: 1) Sun oriented, magnetic field spin controlled, or 2) Earth oriented.”

Figure 3.8. Determination of attitude and orbit (Fiber Optic Gyros), Control Moment Gyros (CMG) and Hydra Star Trackers – source: Airbus defense and Space, Astrium & ESA. For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

Image Quality

57

Today, the controls of the orbit and the gaps of attitude allow the position of the satellite to be known – and updated with high frequency – to a few meters for a ground error of the order of 4 m (Pleiade, SPOT 6 and 7) at 10–15 m (SPOT5, Landsat 7 and 8), directly obtained by the inverse location model. It was approximately 240 m for SPOT 4 in 1998. Once again, we must repeat: remote sensing is above all an expression of great technical prowess and engineering. The geometric quality of an image also depends on the acquisition technique of the sensor; there are three (see Figure 3.9): – whiskbroom concept: mechanical scanning perpendicular to the satellite’s directional motion (one or more detectors per spectral band); – pushbroom concept: along-track scanning (an array of detectors covering the lateral field of view); – full frame concept: 2D image (1 pixel is 1/n detectors).

Figure 3.9. Whiskbroom – pushbroom – full frame. For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

3.2.1. Whiskbroom concept Historically, the first MSS sensor used this technique through the rotation of a mirror. Today, Landsat 7 satellites still use this technique. With the MSS six detectors per band (24 in total) and 24 per band with TM (114 in total) scan few lines along columns of the image using orbital motion along the track.

58

TORUS 2 – Toward an Open Resource Using Services

Figure 3.10. Detector projection at the ETM+ focal plane

ETM+ Landsat 7 is more precise: “the relative position of all detectors from both focal planes with respect to their actual ground projection geometry […] The evennumbered detectors are arranged in a row normal to the scan direction while the odd-numbered detectors are arranged in a parallel row, off exactly one Instantaneous Field of View (IFOV) in the along-scan direction. This arrangement provides for a contiguous bank of 32, 16, and 8 detectors for Band 8, Bands 1–5 and 7, and Band 6 respectively. The detector arrays are swept left to right (forward) and right to left (reverse) by the scan mirror. With each sweep or scan, an additional 480 meters (32, 16, and 8 data lines at a time) of along-track image data is added to the acquired subinterval.” (Figure 3.11). The mechanisms of the mirror are very complex (specifically for motion compensation); hence, geometric disturbances and a limited integration time (and then trade-off radiometry/spatial resolution): “An MSS scene had an Instantaneous Field Of View (IFOV) of 68 meters in the cross-track direction by 83 meters in the along-track direction (223.0 by 272.3 feet respectively). To understand this concept, we must consider a ground scene composed of a single 83 by 83 meter area. The scan monitor sensor ensures that the cross-track optical scan is 185 km at nominal altitude regardless of mirror scan nonlinearity or other perturbations of mirror velocity. Cross-track image velocity was nominally 6.82 meters per microsecond. After 9.958 microseconds, the 83 by 83 meter image has moved 67.9 meters. The sample taken at this instant represented 15 meters of previous information and 68 meters of new information. Therefore, the effective IFOV of the MSS detector in the cross-track direction was considered to be 68 meters, which corresponds to a nominal picture element (pixel) ground area of 68 by 83 meters at the satellite nadir point. Using the

Image Quality

59

effective IFOV in area calculation eliminates the overlap in area between adjacent pixels.” (NASA). Figure 3.11 presents the MSS scanning complex system and geometry of the raw image (left) and after scan line compensation (right).

Figure 3.11. Whiskbroom complex geometry of acquisition

Landsat 7 ETM+ sensor Scan Line Corrector (SLC) failed on May 31, 2003. “The Scan Line Corrector (SLC) compensates for the forward motion. Subsequent efforts to recover the SLC were not successful, and the failure appears to be permanent. Without an operating SLC, the Enhanced Thematic Mapper Plus (ETM+) line of sight now traces a zig-zag pattern along the satellite ground track [Figure 3.12]. As a result, the imaged area is duplicated, with a width that increases toward the scene edge.”

Figure 3.12. SLC-Off (left)/SLC-On (right) Landsat 7 ETM+. For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

60

TORUS 2 – Toward an Open Resource Using Services

3.2.2. Pushbroom concept In 1986, CNES launched SPOT 1 with the first pick-up sensor “pushbroom” on board; today, most sensors use this technique as it offers main advantages: it is easy to use and flexible, has a high geometric quality according to guidance control, and a wide swath and long strips are accessible. Lines are simultaneously acquired by aligned detectors and columns are acquired during the satellite’s forward motion and according to guidance control. The main disadvantages are the complexity of the focal plane and problems with radiometric equalization. Operational Land Imager (OLI) Landsat 8 sensor is also a pushbroom one. “The OLI sensor, which has a five-year design life […] represents a significant technological advancement over Landsat 7’s Enhanced Thematic Mapper Plus (ETM+) sensor. Instruments on earlier Landsat satellites employed oscillating mirrors to sweep the detectors’ field of view across the swath width (‘whiskbroom’), but OLI instead uses long linear detector arrays with thousands of detectors per spectral band. Detectors aligned across the instrument focal planes collect imagery in a ‘pushbroom’ manner resulting in a more sensitive instrument with fewer moving parts. OLI has a four-mirror telescope and data generated by OLI are quantized to 12 bits, compared to the 8-bit data produced by the TM & ETM+ sensor.” The detectors are divided into 14 identical Sensor Chip Assemblies (SCAs) arranged in an alternating pattern along the centerline of the focal plane (Figure 3.13). “Each SCA consists of rows of detectors, a read-out integrated circuit (ROIC), and a nine-band filter assembly. Data are acquired from 6916 across-track detectors for each spectral band (494 detectors per SCA) with the exception of the 15 m panchromatic band that contains 13,832 detectors […]. Even and odd numbered detector columns are staggered and aligned with the satellite’s flight track. Evennumbered SCAs are the same as odd-numbered SCAs, only the order of the detector arrays is reversed top to bottom. The detectors on the odd and even SCAs are oriented such that they look slightly off nadir in the forward and aft viewing directions. This arrangement allows for a contiguous swath of imagery as the pushbroom sensor flies over the Earth, with no moving parts. There is one redundant detector per pixel in each VNIR band, and two redundant detectors per pixel in each SWIR band. The spectral response from each unique detector corresponds to an individual column of pixels within the Level-0 product […]. Silicon PIN (SiPIN) detectors collect the data for the visible and near-infrared spectral bands (bands 1 to 4 and 8) while Mercury–Cadmium–Telluride (MgCdTe) detectors are used for the shortwave infrared bands (bands 6, 7, and 9). There is an additional ‘blind’ band that is shielded from incoming light and used to track small electronic drifts. […] That is a total of 70,672 operating detectors that must be characterized and calibrated during nominal operations.” (ibid.)

Image Quality

61

Figure 3.13. OLI focal plane and odd/even SCA band arrangement (ibid.)

High spatial resolution pushbroom sensors (SPOT 6 and 7, Pleiades) use Time Delay Integration (TDI) because, when in orbit with very high spatial resolution, linear arrays are unable to collect very much light and are impossible to slow down (increase exposure time), and then the solution consists of successively observing the same ground pixel by the n lines of the sensor and transferring the charge collected at the same cell (Figure 3.14). “Five sensors of each kind are used simultaneously to obtain a wide field of view (about 20 km). Image reconstruction at the level of 4 inter-array areas (IAA) is one of the tasks of the ground segment.” (Source: https://www.int-arch-photogramm-remote-sens-spatial-inf-sci.net/XXXIXB1/537/2012/isprsarchives-XXXIX-B1-537-2012.pdf).

62

TORUS 2 – Toward an Open Resource Using Services

Figure 3.14. Principles of Pleiade TDI acquisition (modified from the above source). For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

3.2.3. Full frame concept In an ideal solution, 1 pixel is 1/n detector in a matrix of detectors and all pixels are acquired simultaneously. There is also high geometric quality and a high number of images is accessible. However, this technique is not yet used in remote sensing because of “the technological complexity of CCD [see Note 3 in section 3.5] matrices (number and size of detectors, data dumping), the orbital motion of the

Image Quality

63

satellite during imaging: motion compensation needed in high resolution and the radiometric calibration in the field.” (Valorge 2012) Image geometry is also influenced by the telescope and its optic system. Optics laws are formal (Figure 3.15): increase field of observation (FOV), reduce the focal length and then degrade ground resolution; increase ground resolution, enlarge focal length and then reduce FOV. The relationship between the diameter of the optic and size of the focal length imposes the use of Korsh or Ritchey-Chrétien telescopes (see Note 4 in section 3.5) with a set of mirrors that make it possible to condense the distance of the focal length: Pleiade with a 65 cm optical diameter needs a 12.9 m focal length reduced to 1.3 m with a Korsh telescope… that can not put on an orbital platform!

Figure 3.15. The laws of optics are formal. For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

Telescope’s prisms and lenses – even with the highest quality – disturb the path of light and also induce geometric distortions.

64

TORUS 2 – Toward an Open Resource Using Services

3.2.4. Optical geometric distortions The very high optical quality of remote sensing sensor limits aberration but not completely we mainly distinguish (Figure 3.16): – “Spherical aberration is an optical effect observed in an optical device (lens, mirror, etc.) that occurs due to the increased refraction of light rays when they strike a lens or a reflection of light rays when they strike a mirror near its edge, in comparison with those that strike nearer the center. It signifies a deviation of the device from the norm, i.e. it results in an imperfection of the produced image. – Comatic aberration, in an optical system, refers to aberration inherent to certain optical designs or due to imperfection in the lens or other components that result in off-axis point sources such as stars appearing distorted, appearing to have a tail (coma) like a comet. Specifically, coma is defined as a variation in magnification over the entrance pupil. In refractive or diffractive optical systems, especially those imaging a wide spectral range, coma can be a function of wavelength, in which case it is a form of chromatic aberration. – Astigmatism is one where rays that propagate in two perpendicular planes have different foci. If an optical system with astigmatism is used to form the image of a cross, the vertical and horizontal lines will be in sharp focus at two different distances. It is very limited with 2 and 3 mirrors telescope. This aberration can be corrected by a lens (see Hubble). – Chromatic distortion or spherochromatism is an effect resulting from dispersion in which there is a failure of a lens to focus all colors to the same convergence point. It occurs because lenses have different refractive indices for different wavelengths of light. The refractive index of transparent materials decreases with increasing wavelength in degrees unique to each. Chromatic aberration manifests itself as ‘fringes’ of color along boundaries that separate dark and bright parts of the image, because each color in the optical spectrum cannot be focused at a single common point. Since the focal length f of a lens is dependent on the refractive index n, different wavelengths of light will be focused on different positions.” (modified from Wikipedia 2020c). Wikipedia https://en.wikipedia.org/ wiki/Chromatic_aberration. The last but not least, some sensors can be inclined (off-pointing) in order to reduce the acquisition time which induces other geometric distortions further accentuated by the relief of the observed terrestrial surface. The ratio between the size of a pixel nadir acquired and the same pixel with different off-pointing degrees is given in Figure 3.17: with 32° off-pointing (maximum for SPOT), the size ratio is 1.46 (the 6,000 SPOT P pixels/column does not cover 60 km of flat area but 88 km).

Image Quality

Comatic aberration

Spherical aberration

Astigmatism aberration

Figure 3.16. Optical aberrations

Figure 3.17. Ratio/nadir pixel size according to the detector off-pointing (modified from https://www.nrcan.gc.ca/node/9407). For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

65

66

TORUS 2 – Toward an Open Resource Using Services

3.2.5. Relief distortions It’s easy to understand that, observed from above, the Earth seems to be flat even though we know it’s not. Relief impacts image geometry (Figure 3.18). Geometric distortions, called parallax errors MM’, depend on the acquisition angle α and elevation h regarding the reference plan: MM’= d.tan(α). To compensate for distortions, a digitized elevation model (DEM) is used in an orthorectification procedure applied to each pixel (DEM must approximately have the same spatial resolution than the image; see Note 5 in section 3.5).

Figure 3.18. Principles of orthorectification to eliminate relief geometric distortions

To resume geometric distortions caused by the type of sensors, the duration of acquisition, the velocity of the platform, the accuracy of the clock, frequencies of the clock, scan mirror control, geocentric point on the orbit (controlled by spatial geodesy and ground station), satellite orientation: pitch, roll, yaw (controlled by Star Tracker and Gyros), additional off-pointing using mirror, slippage of the line due to the Earth’s rotation, step deviation with off-pointing due to the Earth’s spherical shape, ground elevation, optical aberrations, etc., the square raw matrix acquired is not at all a square on the ground (Figure 3.19). All these distortions are to be corrected to finally obtain a true map from the raw image using two solutions: the inverse location model and the direct location model.

Image Quality

67

Figure 3.19. The true shape on the ground of a square matrix acquired from above (modified from SwissTopo)

3.2.6. Inverse location model The Inverse Location Model (ILM or Orbital Model, Physic Model) consists of determining for each pixel in the sampling grid, the equivalent position on the Earth in geographical coordinates X and Y (or latitude, longitude) using all parameters given by the system of acquisition (platform and sensor) and the knowledge of Earth’s shape, velocity and relief. Because these parameters are imperfect, the model is refined using ground control points and homologous points to improve the location through the re-estimation of the physical imaging parameters. Complex mathematical model and quiet – nothing to do for the user. The imaging parameters P10, …, Pn0 used to calculate the position and viewing directions are sampled, and the model is calculating iteratively until convergence between the intersection of the detector’s viewing direction at the moment of acquisition with the Earth (ellipsoid or DEM). This calculation is applied on a regular grid of the image (i.e. not for every pixel – Figure 3.20), and we finally obtain a direct location model, which can be written as follows [3.1]:

X = FX (l, p, h, P10 ,...,Pn0 )  Y = FY (l, p, h, P10 ,...,Pn0 ) where

X, Y: Lat, Long coordinates; Fx, Fy: functions using P10, …, Pn0 to determine X and Y.

[3.1]

68

TORUS 2 – Toward an Open Resource Using Services

The orbital pushbroom model is a very complex calculation process. Main agencies used them and some commercial applications support them for a few sensors but the use of these models is limited. The rational polynomial camera/coefficient (RPC) model is a generic sensor model that can be used with a wide range of sensors as an alternate sensor model (additionally, it is widely supported in commercial software packages). However, the use of the RPC model requires values for a set of coefficients used in the model. These coefficients are generated using the physical camera model. Thus, to use the RPC model, either: – the physical camera model (or knowledge of the interior configuration of the sensor) must be known; or – the coefficients used in the model must be provided by the data producer; – precision is limited by the accuracy of the imaging parameters: SPOT1-4: 350 m, SPOT5 HRG: 30 m, SPOT5 HRS: 15 m, Pleiades: 5 m, Landsat 8 level 1T: 12 m. MODIS Row & Line – pixels coordinates 700 600 500 400 300 200 100 0 0

100

200

300

400

MODIS equivalent Latitude & Longitude using Orbital Model 25 20 15 10 5 0 95

100

105

110

115

120

125

Figure 3.20. MODIS GCP from the sensor model (inverse location model). For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

Image Quality

69

3.2.7. Direct location model The Direct Location Model (DLM) consists of selecting a set of GCP to define a mathematical model between X and Y geographical coordinates and Line and Row pixel coordinates in the sampling grid (at the difference with the refined inverse location model that here only pixel and geographical coordinates are considered). These GCP matrices can be calculated from the inverse location model and be a set of points included in the image metadata (approximately 133 points for MODIS Terra/Aqua, 3,000 for VIIRS Suomi NPP). They are used to find the best adjustment method using a statistical modelization analysis, generally: – polynomial regression order 3 if small swath, validation with RMS (Root Mean Square) between given Latitude/Longitude and estimated ones (2): (2) LATest = a0 + a1 X + a2 Y + a3 X2 + a4 XY + a5 Y2 + a6 X3 + a7 X2Y + a8XY2 + a9 Y3 LONGest = b0 + b1 X + b2 Y + b3 X2 + b4 XY + b5 Y2 + b6 X3 + b7 X2Y + b8XY2 + b9 Y3 where: LATest ,LONGest: Latitude and Longitude are estimated, X, Y: line and row in the raw matrix, a0, …, an and b0, … bn: coefficients of the polynom; – thin plat spline (TPS), more efficient (Figure 3.21) with large swath but no RMS control: “the TPS arises from consideration of the integral of the square of the second derivative – this forms its smoothness measure. In the case where x is two dimensional, for interpolation, the TPS fits a mapping function f(x) between corresponding point-sets {Xi} and {Yi} that minimizes the following energy function.” (Wikipedia 2020d):

[3.2]

70

TORUS 2 – Toward an Open Resource Using Services

Figure 3.21. TPS more efficient solution than polynomial ones. For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

With ancient images or maps, archives from aerial photography or remote sensing and in situ data – in all case, without the inverse model location or sensor model – geometric correction and georeferencing may use different mathematical models: – polynomial and TPS, as we saw; – triangle-based rectification, also called rubber sheeting. “Once the triangle mesh has been generated and the spatial order of the control points is available, the geometric rectification can be done on a triangle-by-triangle basis. This triangle-

Image Quality

71

based method is appealing because it breaks the entire region into smaller subsets. If the geometric problem of the entire region is very complicated, the geometry of each subset can be much simpler and modeled through simple transformation. For each triangle, the polynomials can be used as the general transformation form between source and destination systems.” (ERDAS 1999): Delaunay triangles are defined from the GCP network; for each triangle an order 1 polynom is defined to estimate the Lat and Long from pixel coordinates (along arcs between triangles, the two polynoms give same coordinates). Figure 3.22 illustrates this method applied to in-situ sensing images we have to mosaic to survey the East Lovén glacier (Spitsbergen, Svalbard).

Figure 3.22. Signal processing steps requiring the preliminary measurement of the position of ground control points over the glacier surface. A significant issue is the interpretation of the pictures in order to locate these features and identify which flag is located at known positions. For a color version of this figure, see www.iste.co. uk/laffly/torus2.zip

72

TORUS 2 – Toward an Open Resource Using Services

3.2.8. Root Mean Square (RMS) validation

Polynomial geometric corrections are generally validated by RMS, calculated between the known coordinates on the ground and the estimated ones by the model [3.3]: =

(

−

) +(

−

)

[3.3]

where: n: GCP n; lat’ & lon’: estimated latitude and longitude; lat & lon: given latitude and longitude. The polynomial transformation is a very general method: it can be applied to any type of image and allows for successive corrections. It is thus a method that is a priori flexible and powerful. Its disadvantage is the ability to adjust just about everything, so also big mistakes, especially bad control points (Figure 3.23).

Figure 3.23. Do not use RMS blindly! For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

Image Quality

73

3.2.9. Resampling methods Geometric correction consists of reducing distortion relative to the acquisition system but georeferencing – generally associated in the process – consists of transforming the original sensor regular grid in a new one with distortions relative to the end coordinate system chosen: geographical (Latitude, Longitude and ellipsoidal height), projected (X, Y, Z). It’s easy to understand that a single 45° rotation defines a new grid in which it’s necessary to calculate a new value for the corresponding initial pixel – and it can be with a change of pixel size (Figure 3.24).

Figure 3.24. Main idea of resampling, initial grid (red) after a 45° rotation (black). For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

The main resampling methods are: – Nearest neighbor (NN): this is simply the value of the pixel of the raw image closest to the pixel to be resampled that is affected. This method is fast and very inexpensive in calculation… but provides images of poor quality: staircase effects, oversampling insignificant (the size of the pixels can be reduced only by a redundancy of information). However, it is the only method suitable for qualitative data or when the original physical value must be retained as it is subsequently integrated into a rigorous physical function.

74

TORUS 2 – Toward an Open Resource Using Services

Figure 3.25. Cubic convolution (ERDAS 1999). For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

– Cubic interpolation (CC): “Since a cubic, rather than a linear, function is used to weight the 16 input pixels [Figure 3.26], the pixels farther from (xr,yr) have exponentially less weight than those closer to (xr,yr). Several versions of the cubic convolution equation are used in the field. Different equations have different effects upon the output data file values. Some convolutions may have more of the effect of a low-frequency filter (like bilinear interpolation), serving to average and smooth the values. Others may tend to sharpen the image, like a high-frequency filter [3.4]” (ERDAS 1999): =

( − 1, +

+ ( + 1, +

− 2) ∗ ( ( − 1, +

− 2) + 1) + ( , +

∗ ( ( , + − 2) + 1) − 2) ∗ ( ( + 1, + − 2) + 1)

where: i: int (Xr); J: int (Yr); d(i,j): the distance between a pixel with coordinates (i,j) and (Xr,Yr); V(i,j): the data file value of a pixel (i,j); Vr: the output data file value.

− 2) [3.4]

Image Quality

75

A: -0.5 (a constant which differs in other applications of cubic convolution) f(x) = the following function [3.5]:

[3.5]

Figure 3.26. Grid (red) and image deformation after resampling (left NN, right CC). For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

3.2.10. Image geometric quality to assume geographical space continuity Ideally, geometrically corrected and georeferenced images are all in the same representation and can be mosaicked at will – acquired on the same date but separately or by different systems or at different dates – without revealing spatial discontinuities. Ideally, from the raw data digital counts or their apparent luminance equivalence, it is essential to operate a radiometric balance between the images to ensure uniformity and continuity of hue between the assembled images. Figure 3.27 (top) gives the example of an assembly of Pleiades scenes – both acquired by nadir – which also optimizes the mosaicing line to make it almost invisible, at the bottom of the figure, on the other hand, it gives an example, on the contrary, of the distortions encountered that are either large-scale or in urban areas where all the geometric distortions cannot be corrected in the absence of a 3D model of buildings.

76

TORUS 2 – Toward an Open Resource Using Services

Figure 3.27. Limits in image geometric quality

3.3. Image quality – radiometry

Basically, remote sensing measures the energy transported by photon energy associated with a photon emitted at a wavelength . “An Earth observation instrument measures a luminance according to the optical properties of the analyzed landscape, which are then modified by the presence of the atmosphere […]. Given the different atmospheric phenomena, the signal measured by an instrument at the ( )] can be decomposed as the sum [Figure 3.28] of top of the atmosphere [ several radiative components whose relative importance depends on the spectral domain considered [3.6]: ( )=

( )+

( )+

( )+

( )

[3.6]

( ): luminance resulting from the diffuse and transmitted solar where: radiation and from the Earth–Atmosphere coupling incident on the target and reflected in the direction of the sensor and the luminance reflected by the environment and then diffused in the direction of the sensor.

Image Quality

77

( ): luminance emitted by the target, of emissivity e, at the temperature T and transmitted in the direction of the sensor. ( ): luminance resulting from the scattering of solar radiation by particles suspended in the atmosphere in the direction of the sensor. ( ): luminance resulting from the radiative emission of the atmosphere in the direction of the sensor.”

Figure 3.28. The signal is the sum of several radiative components (modified from ibid.). For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

The terrestrial surfaces reflect the incident energy spectrally, selectively, and predominantly in the [0.4 μm; 2.5 μm]. Water surfaces (in all its forms), artificial surfaces, vegetation and bare soil (including rocky outcrops) are mainly distinguished. “The spectral signature of a material is composed of absorption bands characteristic of the presence of ions or molecules. Indeed, these absorb more or less strongly the incident radiation when it is in agreement with the energy of an electronic, vibrational, or rotational transition of the molecule.” (ibid.). On the ground to characterize a surface, we distinguish: the reflectance reflects the capacity of a body to reflect the incident energy; emissivity allows us to characterize the potential of a body to reflect light and the transmittance of the one to transmit this light. Radiative values are: – radiant flux (W); – flux density: irradiance, exitance (W⋅m-2); – intensity (W⋅sr-1); – radiance (W⋅m-2⋅sr-1). As we saw in Chapter 12, the apparent radiance (luminance) captured by the sensor ToA does not at all correspond to the ground reflectance. Landscape objects

78

TORUS 2 – Toward an Open Resource Using Services

interfere with the electromagnetic signal already modified during these atmospheric paths (Figure 3.29). It is also necessary to take into account artifacts related to acquisition technology because sensors capture reflected solar energy, convert these data to apparent radiance and then rescale this data into an 8-bit or 16-bit digital number (DN). We distinguish three main points: – radiometric model of the instrument (optic and electronic); – radiometric processing: equalization and calibration; – radiometric quality: signal noise reduction.

Figure 3.29. The light long path to the sensor. For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

3.3.1. Radiometric model of the instrument The photon flux passes through optical lenses and filters and is reflected by mirrors before reaching the sensor which generates wavelength separation. Optics collects the incident radiation, selects the spectral bands, transmits and focuses in the detection plane. We saw an upper optical aberration introduced by optics. The detection – the conversion of the number of photons in electrical values – is the

Image Quality

79

domain of the quantum detection, generally electronic CCD (Coupled Charged Device) that “count” the number of photons arriving on the detector of surface S during time Ti. Detectors depend of the system used (ibid.): – “array of 12,000 of 6.5 µm silicon detectors for SPOT5 (visible); – arrays of 300 of 13 µm Gallium Arsenide detectors for SPOT4/5 (MIR), connected end-to-end to form a line of 3000 detectors; – arrays of 300 of 30 µm HgCdTe detectors (thermal IR); – TDI of 32 lines and 4,000 columns of 12 µm for IKONOS (visible); – POLDER 242 x 548 matrix of 27 16 µm detectors (visible); – Matrix of 4,000 x 4,000 of 9 µm detectors for aerial imaging (visible).” The signal was amplified before digitalization (amplification step) and, lastly, the physical quantity measured for each pixel follows a linear radiometric model where coefficients only depend on the instrument and the detector, not on the landscape. These coefficients depend on the sensor type: – whiskbroom: 2 unique coefficients to be determined; – pushbroom or TDI: 2 coefficients per detector (per image column); – matrix: 2 coefficients for each pixel in the image. Consequently, there are differences between this model relative to the nonlinearity of the detectors, the radiometric noise and the time variation in these coefficients. However, the image must be as faithful as possible to the landscape and consequently needs correcting for instrumental imperfections. It’s the equalization step: “we want the function X = f(L) to be the same for the entire image. Physical interpretation of the grey levels measured by each pixel (knowledge of radiances, therefore the laws concerning f), it’s the calibration step.” (Valorge 2012). 3.3.2. Radiometric equalization and calibration

With a pushbroom sensor as SPOT 5 or similar, the radiometric model is (CNES 2012 – see Note 6 in section 3.5): = .

where

.

.

,

. +

: numerical count of the detector n of the register b with gain m; A: absolute calibration coefficient;

[3.7]

80

TORUS 2 – Toward an Open Resource Using Services

: global amplification gain; : inter-register amplification gain (mean of ,

= 1);

: inter-detector gain (mean on the register = 1);

L: equivalent apparent radiance observed;

: darkness coefficient of the detector n of the register b with gain m. and for each detector n of each register b, the coefficients , For each should be determined, and then equalization consists of calculating [3.8]: =

.

,

∝ .

.

,

,

[3.8]

“With a pushbroom sensor one or more detector arrays scan the landscape vertically below the satellite and consequently each image column is acquired by the same detector. Equalization has to integrate some correction of differences between inter-detector responses and the consideration of differential effects of reading registers. In all cases, the artefacts appear in columns.” (Valorge 2012). Calibration precision is estimated in absolute, in inter-band, in multi-temporal, and will determine the precision of the downstream applications. Each detector’s response is modeled linearly and is therefore characterized by two values (offset + slope). A minimum of two observations are needed, ones where the radiances entering the instrument are uniform in the field, enough to calculate these two values: – On the ground: use of integrating spheres and reference ground-radiometers (primary measurement standards) which inter-calibrate the sensors prior to launch (but variation in calibration between ground and flight – optical transmission, spectral response of the filters, etc. – due to the air-vacuum transfer, UVs, thermal conditions, etc.). – In orbit: use of onboard measurement standards (lamps, black bodies, solar sensors, etc.); observing landscapes whose radiance is theoretically known: glitter, Rayleigh, clouds, Moon, etc.; inter-calibration on uniform landscape (snowy sites in Greenland or Antarctica which are very stable); observation of landscapes characterized simultaneously on the ground (reflectance, atmospheric state, etc. – the Crau alluvial plain, White Sands Desert, etc.); comparison of all these methods (current precision, a few %). Figure 3.30 gives an example of calibration inter-array (left) and inter-bands (right) on a SPOT image (CNES 2012).

Image Quality

81

Figure 3.30. Calibration inter-array and inter-bands on a SPOT image. For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

3.3.3. Radiometric signal noise reduction (SNR) Radiometric noise is a standard deviation of the data in the image of a uniform landscape. As we saw, image quality is strongly linked with the system’s ability to distinguish targets with similar radiance and noise structures heavily dependent on the imaging of different modes – array, mainly column and inter-column noise (Figure 3.31). Sources of radiometric noise are (Valorge 2012): – photonic noise: an incoherent light is observed, the process of arrival of photons follows the Poisson distribution, and then the number of photons effectively received has a mean of N and a standard deviation Root(N); – electronic noise from the reading and amplification system; – quantization noise; – equalization noise (structured); – compression noise (structured).

82

TORUS 2 – Toward an Open Resource Using Services

Quantization and compression are adapted to the maximum rate imposed between onboard and ground. Also adapted to instrument performance: – if the analog noise is significant with regard to the quantization noise, bits are used to encode only noise, which is pointless; – if there is significant quantization noise, the instrument is underused. We seek similar levels of analog and quantization noise.

Figure 3.31. Signal noise reduction. For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

3.3.4. Radiometric physical value

The benefit of calibration and SNR also consists of deducing physical values from apparent luminance calculated from raw digital numbers. Figure 3.32 illustrates the possible workflow regarding using data from DN to Top of Atmosphere reflectance (ToA), simulated ground reflectance (Bottom of Atmosphere – BoA). ToA to BoA reflectance conversion is possible using a radiative transfer model of the atmosphere at the moment of the image acquisition (6S model or LibRadtran library – Sentinel 2 level 2, for example). The atmosphere parameters may be derived from image bands (cirrus, water vapor NIR, SWIR).

Image Quality

83

Figure 3.32. From DN to simulated ground reflectance. For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

“There are two formulas that can be used to convert DNs to radiance; the method you use depends on the scene calibration data available in the header file(s). One method uses the Gain and Bias (or Offset) values from the header file. The longer method uses the LMin and LMax spectral radiance scaling factors. The formula to convert DN to radiance using gain and bias values is: Lλ = gain* DN + bias

where: Lλ is the cell value as radiance; DN is the cell value digital number; gain is the gain value for a specific band; bias is the bias value for a specific band. The formula used in Lmin, Lmax process is as follows: Lλ = ((LMAXλ – LMINλ)/(QCALMAX–QCALMIN))*(QCAL–QCALMIN)+LMINλ

84

TORUS 2 – Toward an Open Resource Using Services

where: Lλ is the cell value as radiance; QCAL = digital number; LMINλ= spectral radiance scales to QCALMIN; LMAXλ = spectral radiance scales to QCALMAX; QCALMIN = the minimum quantized calibrated pixel value (typically = 1), QCALMAX = the maximum quantized calibrated pixel value (typically = 255 with 8-bit, 65535 with 16-bit). The last conversion consists of calculating the top-of-atmosphere (ToA) reflectance: pλ = π * Lλ * d 2 / ESUN λ * cos θ s

where: ρλ = unitless planetary reflectance; Lλ= spectral radiance (from the earlier step); d = Earth–Sun distance in astronomical units; ESUNλ = mean solar exoatmospheric irradiances; θs = solar zenith angle.” (NASA). In theory, BoA reflectance offers the possibility of directly using spectral ground signature to analyze landscape and to compare data from different systems and/or different dates. In practice, except for experimental studies, this is not possible because the vertical atmosphere composition description is not enough precisely. 3.3.5. Image quality – resolution

Image resolution is the capability of an imaging system to reconstitute the information contained in an observed scene, sharpness, for example. Valorge (2012) gives some common definitions: – “resolving power: minimum distance for distinguishing between 2 neighboring objects (lines, points…); – maximum perceptible number of pairs of lines by unit of length of a periodic pattern (maximum perceptible spatial frequency); – size of elementary detector projected onto the ground and/or sampling step: Instantaneous Field Of View (IFOV, angular size of the elementary detector).”

Image Quality

85

For the end user, image resolution is the number of bands and the pixel size. From Landsat MSS to Landsat OLI or from SPOT 1 to SPOT 7 in a few decades, the spatial resolution decreased from 56*79 m to 15*15 m (P), 30*30 m for Landsat and from 10*10 m (P), 20*20 m (XS) to 1.5*1.5 m (P), 6*6 m (XS) and spectral resolution to a few visible and near infrared bands to increase bands in the infrared domains. Since 2015, Sentinel 2 satellites offer 13 bands with 3 different spatial resolutions (10 m, 20 m and 60 m) for free. However, image resolution cannot be reduced to the spatial and spectral resolutions; it’s more complex. For example, data are quantized on 12 bits today against 8 bits before, pushbroom high resolution is TDI (Time Delay Integration), and the inverse location model and orbit are well controlled with gyros and star tracker… To understand the image resolution concept, it’s necessary to understand that the sensor has to distinguish a maximum of contrasts with high noise control with the best sampling rate. Each of the elements making up the instrument chain is considered to be a linear system, spatially invariant. Mathematically, the effect of each contributor i can then be modeled as the convolution of the entry e with a function hi called the point spread function (PSF). The function hi is by definition the image of a point acquired by contributor i, and the overall PSF is the convolution of the PSF of the various contributors (all elements of the image of the landscape). In physics, one pixel, one element of the image convoluted by a PSF, is a pulse, a point of light that enters in the instrument (Figure 3.33): a gate function or Dirac (the mathematical model of a pulse at x0 and y0 is a Dirac distribution d(x-x0,y-y0)). “The sampled input signal consists of a series of pulses to which correspond a succession of individual responses (output from CCD sensor). A pulse is an image whose value is zero, except at a point (Po). The pulse undergoes a degradation of its representation due to the distribution of the information, which is also distributed over a segment (evaluated by the Transfer Function Linear), a surface (2D). This spreading of the information is determined by the Point Spread Function (PSF).” (Wikipedia).

Figure 3.33. From the points of light to the comb of Dirac. For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

86

TORUS 2 – Toward an Open Resource Using Services

CNES (2012), explains “In the image, physically the response to a point of light is an image spot h(x,y) and mathematically: I(x,y) = h(x,y)* d(x-x0,y-y0). The generalization is that any landscape can be considered as the sum of Diracs: L ( x; y ) =

L

x0 , y0

x0 , y0

.δ ( x − x0 ; y − y0 )

As the acquisition system is linear and spatially invariant, the image of this landscape is the sum of the responses of each Dirac”: I ( x; y ) =

L

x0 , y0

x0 , y0 I ( x;

y) =

L

x0 , y0 h ( x,

y )δ ( x − x0 ; y − y0 ) = h( x, y ) L ( x, y )

x0 , y0

The instrument that receives energy true the optic system transforms the PSF (Point Spread Function) in a frequency domain function called MTF (Modulation Transfer Function) that is a Fourier transform of the PSF (Figure 3.34).

Figure 3.34. Spatio-temporal and spectral domains: the image is a convolution of the landscape by the MTF of the sensor (that is a Fourier transform of the PSF). For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

In the spectral domain, “the Fourier transform provides two separate domains, equivalent to each other, that contain all the information, and can give rise to all the reproduction or processing operations. The domain of Fourier lends itself best to

Image Quality

87

mathematical treatments, while the spatio-temporal domain is more in the spirit. The characterization of signals in the frequency domain provides information on bandwidth and spectral components, and facilitates the study of linear systems. A satellite image is considered as any periodic function of frequency which is constituted by the sum of a series (Fourier series) of terms of period and amplitude of its own. The first term is constant (fundamental frequency), the following are sinusoidal functions whose (harmonic) frequencies are integral multiples of the fundamental frequency (the model is easier to interpret in terms of spatial frequencies): – overall MTF: simple product of the elementary MTFs; – image spectrum: product of the MTF and of the spectrum of the landscape” (CNES 2012). To understand the image resolution concept, one is obliged to always jungle between both domains: spatio-temporal and spectral. Image can also be considered as a target sine wave with period p and amplitude A varying according to x values of A/2 with frequencies (fx=-1/p; fy=0) and (fx=1/p; fy=0). That corresponding to a low pass filter (Figure 3.35) and the control of MTF will directly impact the image quality (Figure 3.35): – high MTF: good determination of contrasts, sharpness; – low MTF: poor determination of contrasts, blurring.

Figure 3.35. Sampling frequencies and MTF control

88

TORUS 2 – Toward an Open Resource Using Services

The choice of sampling frequency fe is determined (Figure 3.36): – if fe is large: the original continuous signal can be reconstituted; – if fe is small: the original continuous signal is not reconstituted; – fe must be compared to the frequencies present in the unsampled signal.

Figure 3.36. The choice of a sampling frequency. For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

If the Shannon–Nyquist (Figure 3.37) condition is respected: – all the information is contained in the periodic pattern [-fe /2, fe /2]; – this information represents the spectrum of the continuous signal only if the terms of the sum do not overlap: the spectrum of the continuous signal = 0 outside [-fe /2, fe /2];

Image Quality

89

– f the spectrum [-fe /2, fe /2] sampling must respect fe ≥ 2fmax.

Figure 3.37. The Shannon–Nyquist sampling condition (Valorge 2012). For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

If the Shannon–Nyquist condition is not respected (Figure 3.38): – signals whose frequency fe/2 is not transmitted by the sampling: loss of information; – worse still, signals whose frequency ≥ fe/2 is interpreted as low frequency signals (i.e. as having a frequency lower than < fe/2): aliasing: false information.

Figure 3.38. Example of two-dimensional aliasing, 2 images of the same landscape, obtained with identical instrument MTFs, the first having double the sampling rate of the second (modified from Valorge 2012)

90

TORUS 2 – Toward an Open Resource Using Services

FTM in-flight measurement methods and focus defect may be operated from targets on the ground (Figure 3.39) dispersed by space agencies and also relative MTF measurement (with a higher resolution image).

Figure 3.39. Target on the ground can be used to calibrate MTF

In addition to image resolution control, FTM knowledge allows for high quality Panchromatic/Multispectral fusions for systems such as Spot, Pleiades, Ikonos, Quickbird, Landsat OLI or Worldwiew which have a highly resolved mode for a wide spectral band (P) and a less solved mode for several bands (XS); the acquisitions are simultaneous. The idea is to have in fine an image with a high spatial resolution (that of the P) with the radiometric diversity of the multispectral one. The simplest fusion techniques consist of an RGB-HSI-RGB (Red Green Blue and Hue Saturation Intensity) transformation. The first transformation consists of calculating the HSI components from an RGB color space defined by the multispectral bands. Then, the panchromatic band is substituted for the I component before a second transformation to find an RGB space. The bands must be perfectly georeferenced and resampled at the spatial resolution of the panchromatic band. Another solution consists of simulating the original landscape signal by a modelization of the inverse of the known MTF of the system of acquisition: deconvolution. This deconvolution is legitimate: no aliasing, the high frequencies amplified are “clean”. However, radiometric noise is amplified at the same time as useful high frequencies. The restoration pansharpened (Figure 3.40) method uses these characteristics of deconvolution as the sum of deconvolution and denoising (difference between original P and deconvoluted P).

Image Quality

91

Figure 3.40. Restoration fusion SPOT image (from Valorge 2012). For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

3.4. Conclusion This introduction to the main principles governing the quality of images reminds us of the technical and engineering prowess of the field of spatial imagery. One can once again only marvel at the high quality of remote sensing images and be convinced that before anything else, it is a discipline that requires a deep investment to even conceive the entirety of concepts and techniques mobilized. Now fed with this knowledge, it is time to approach the application, what can be done with all these images?

3.5. Notes NOTE 1.– The focal length of the Pléiade sensor is 12.9 m; it is 57.6 m for Hubble. For Pléiade, the telescope is Cassegrain type, it is a Korsh model with three mirrors that can compress the focal length to 1.3 m and the diameter to 65 cm to offer an instrument that at the size adapted to a satellite. NOTE 2.– “Innovative guidance techniques are used to circumvent the usual drawbacks of CMGs: instead of following a predefined attitude profile that leads the cluster to encounter singular points, the cluster is reoriented to take into account the satellite trajectory and the history of the cluster, in order to optimize the overall

92

TORUS 2 – Toward an Open Resource Using Services

SYSTEM. It has been shown that this reorientation strategy always makes it possible to avoid singularities, while ensuring a correct convergence of the guiding algorithm. This new approach makes it possible to use the entire torque capacity of the actuator cluster, which is approximately 3.2 times the capacity of a basic CMG (15 Nms) in roll and pitch. In addition, this open loop guidance can be performed autonomously from the analysis of the programming message. The management of the CMG cluster in the flight software is thus greatly simplified. Propulsion is integrated in a specific module and is only used in the orbit control maneuvering phases. The modes of acquisition and survival are based on the use of an attitude control by magnetic actuators according to a law in B point.”

NOTE 3.– Sensors are Charged Coupled Devices (CCD) (Wikipedia 2020e), “a photosensitive electronic component used to convert electromagnetic radiation (UV, visible or IR) into an analog electrical signal. This signal is then amplified, then digitized by an analog-digital converter and finally processed to obtain a digital image. The sensor is therefore the basic component of cameras and digital cameras, the equivalent of film in silver photography.” NOTE 4.– (Wikipedia 2020f) “A Ritchey–Chrétien telescope (RCT or simply RC) is a specialized variant of the Cassegrain telescope that has a hyperbolic primary mirror and a hyperbolic secondary mirror designed to eliminate off-axis optical errors (coma). The RCT has a wider field of view free of optical errors compared to a more traditional reflecting telescope configuration. Since the mid-20th Century, a majority of large professional research telescopes have been Ritchey–Chrétien configurations (Hubble Space telescope, Landsat ETM+, Landsat OLI). A threemirror astigmat is a telescope (Korsh, Schmidt-Cassegrain) built with three curved mirrors, enabling it to minimize all three main optical aberrations – spherical aberration, coma, and astigmatism. This is primarily used to enable wide fields of view, much larger than those possible with telescopes with just one or two curved surfaces. The Korsh telescope has the advantage of limiting the amount of parasitic light and of forming a flat image on the detector (Spot, Pleiade).” NOTE 5.– Before 1996, global DEM (Digitized Elevation Model) didn’t exist for civil application. If you needed one, you had to realize it yourself or pay for it (GPS on the ground, orthophotogrammetry). Completed in 1996, GTOPO30 is a global digital elevation model (DEM) with a horizontal grid spacing of 30 arc seconds (approximately 1 kilometer). The DEM was derived from several raster and vector sources of topographic information but the spatial resolution is not adapted for satellite imagery as Landsat or Spot. In 2009 with SRTM (Shuttle Radar Topographic Mission), NASA realized a global 90 m – and some part at 30 m – DEM available for free. GDEM (Global DEM) is a 30 m DEM from ASTER imagery today (version 2, 2014). It is more accurate than SRTM, and is also available for free. On the other hand, there are many national 1 m spatial resolutions

Image Quality

93

(which aren’t free) which are available from LIDAR mission. Since 2014, acquisitions from radar satellites TerraSAR-X and TanDEM-X have become available in the form of a uniform global coverage with a resolution of 12 m. Using High Spatial Resolution satellite imagery or aerial photos (from an archive or from UAV), it’s today relatively “easy” to create DEM using STEREOPIPE free software (NASA 2020b, and NASA Ames Intelligent Robotics Group 2019). NOTE 6.– CNES’ book, Satellite Imagery (2012), “was written for students and engineers wishing to understand the basic principles behind the acquisition of optical imagery for Earth observation and the ways in which the quality of the images can be optimized. Intended both for designers and downstream users, the book begins with a detailed explanation of the physical PRINCIPLES involved when a satellite acquires an optical image and then goes on to discuss image processing and its limits as well as the ultimate performance obtained. It also covers in depth the problems to be solved when designing and dimensioning observation systems so that the reader can become familiar with the various processes implemented for acquiring an optical image. The book describes a very wide range of subjects from fundamental physics (radiation, electronics, optics) to applied mathematics (frequency analysis), geometry and technological issues. It draws on work done over many years by engineers from CNES (the French Space Agency), the IGN (the French National Geographic Institute) and ONERA (the French Aerospace Laboratory) in the field of satellite optical imagery.” 3.6. References Atkinson (1985). Preliminary results of the effects of resampling on Thematic Mapper Imagery. ACSM-ASPRS Fall Convention Technical Papers, Indianapolis, USA. CNES (2012). Satellite Imagery: From Acquisition Principles to Processing of Optical Images for Observing the Earth. Editions Cépaduès, France. CNES (2015). Le bus [Online]. Available at: https://pleiades.cnes.fr/fr/PLEIADES/Fr/lien2_ sat.htm. eoPortal (2020). SPOT-6 and 7 [Online]. Available at: https://directory.eoportal.org/web/ eoportal/satellite-missions/content/-/article/spot-6-7. ERDAS (1999). ERDAS Field Guide, 5th edition. ERDAS Inc., Atlanta, USA. Available at: http://web.pdx.edu/~emch/ip1/FieldGuide.pdf. ISPRS (2019). The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences [Online]. Available at: https://www.isprs.org/publications/ archives.aspx. NASA (1986). Earth Observing System: Data and Information System. NASA Technical Memorandum 87777, 2a. Available at: http://hdl.handle.net/2060/19860021622.

94

TORUS 2 – Toward an Open Resource Using Services

NASA (2019). Landsat 7 Science Data Users Handbook [Online]. Available at: https:// landsat.gsfc.nasa.gov/wp-content/uploads/2016/08/Landsat7_Handbook.pdf. NASA (2020a). Search Results: All Fields similar to ‘Ldef’ [Online]. Available at: http:// nasaimages.lunaimaging.com/luna/servlet/view/search?search=SUBMIT&q=ldef&Quick SearchA=QuickSearchA. NASA (2020b). Neo-Geography Toolkit [Online]. Available at: https://ti.arc.nasa.gov/tech/ asr/groups/intelligent-robotics/ngt/. NASA Ames Intelligent Robotics Group (2019). The Ames Stereo Pipeline: NASA’s Open Source Automated Stereogrammetry Software [Online]. Available at: https://byss.arc. nasa.gov/stereopipeline/daily_build/asp_book.pdf. OSE (2012). OSE 2012 : Observation spatiale de l’environnement [Online]. Available at: https://oseocefr.wordpress.com. Rengarajan R., Sampath A., Storey J., Choate M. (2015). Validation of geometry of accuracy of global land survey (GLS) 2000 Data. Photogrammetric Engineering and Remote Sensing, 81(2), 131–141. TORUS (2015). TORUS [Online]. Available at: http://www.cloud-torus.com. USGS (2019). Landsat 8 (L8): Data Users Handbook, version 5.0 [Online]. Available at: https://prd-wret.s3-us-west-2.amazonaws.com/assets/palladium/production/atoms/files/ LSDS-1574_L8_Data_Users_Handbook-v5.0.pdf. USGS (2020a). Remote Sensing Technologies: Test Sites Catalog [Online]. Available at: https://calval.cr.usgs.gov/rst-resources/sites_catalog/. USGS (2020b). Landsat Level-1 Data Products [Online]. Available at: https://landsat.usgs. gov/landsat-level-1-standard-data-products. Wikipedia (2020a). Long Duration Exposure Facility [Online]. Available: https://en. wikipedia.org/wiki/Long_Duration_Exposure_Facility. Wikipedia (2020b). Remote sensing [Online]. Available at: https://en.wikipedia.org/wiki/ Remote_sensing. Wikipedia (2020c). Chromatic aberration [Online]. Available at: https://en.wikipedia.org/ wiki/Chromatic_aberration. Wikipedia (2020d). This plate spline [Online]. Available at: https://en.wikipedia.org/wiki/ Thin_plate_spline. Wikipedia (2020e). Charged-coupled device [Online]. Available at: https://en.wikipedia.org/ wiki/Charge-coupled_device. Wikipedia (2020f). Ritchey–Chrétien telescope [Online]. Available at: https://en.wikipedia. org/wiki/Ritchey%E2%80%93Chr%C3%A9tien_telescope.

4 Remote Sensing Products

4.1. Atmospheric observation 4.1.1. Introduction to common atmospheric gases and particles The atmosphere contains different gases, particles and aerosols, and provides various functions, especially the ability to sustain life. Although air is still primarily composed of oxygen and nitrogen, the concentrations of some pollutants and greenhouse gases have increased, which can cause global warming (Enviropedia 2019). The common atmospheric pollutants that are hazardous to human health include sulfur dioxide (SO2), carbon monoxide (CO), nitrogen dioxide (NO2), ozone (O3) and particulate matter (PM). The greenhouse gases that effect the atmosphere include carbon dioxide (CO2), methane (CH4), nitrous oxide (N2O) and other gases. Air pollution caused about 4.2 million premature deaths in 2016, according to the World Health Organization. Atmospheric pollution is linked to higher rates of cancer, heart disease, stroke and respiratory diseases. In the United States, more than 40% of the population are at risk of disease and premature death due to air pollution, according to the American Lung Association (NUNEZ 2019). Atmospheric pollution can also cause short-term problems such as sneezing and coughing, eye irritation, headaches and dizziness. Particulate matter smaller than 10 µm (classified as PM10 and, smaller, PM2.5) poses a higher health risk because it can enter the bloodstream. Air pollutants also contribute to climate change. Extreme weather and other impacts related to increased greenhouse gases can have a negative impact on human health (NUNEZ 2019).

Chapter written by Van Ha PHAM, Viet Hung LUU, Anh PHAN, Dominique LAFFLY, Quang Hung BUI and Thi Nhat Thanh NGUYEN.

TORUS 2 – Toward an Open Resource Using Services: Cloud Computing for Environmental Data, First Edition. Edited by Dominique Laffly. © ISTE Ltd 2020. Published by ISTE Ltd and John Wiley & Sons, Inc.

96

TORUS 2 – Toward an Open Resource Using Services

The origins and effects of atmospheric gases and particles on the environment and human health are summarized in Table 4.1. Pollutant

Environmental risks

Human health risks

– Automobiles – Fires – Industrial

– Contributes to smog formation

– Heart disease – Vision problems – Reduced physical and mental capabilities

Nitrogen oxides (NO and NO2)

– Automobiles – Electricity – Industrial

– Damages foliage – Contributes to smog formation

– Inflammation and irritation of breathing passages

Sulfur dioxide (SO2)

– Electricity – Fossil fuel combustion – Industrial – Automobiles

– Major cause of haze and acid rain formation – Reacts to form particulate matter

– Breathing difficulties – (particularly for people with asthma and heart disease)

Ozone (O3)

– Industrial – Automobiles – Gasoline vapors – Chemical solvents – Electrical

– Interferes with the respiration ability of plants – Increased susceptibility to other environmental stressors

– Reduced lung function – Irritation and inflammation of breathing passages

Particulate matter (PM)

– Fires – Smokestacks – Construction sites – Unpaved roads – Power plants – Automobiles

– Formation of haze – Acid rain

– Irritation of breathing passages – Aggravation of asthma – Irregular heartbeat

– Fossil fuel combustion

– Atmosphere warming – Increase in pollution and related diseases

– Slow human cognition

– Atmosphere warming

– Suffocation – Loss of consciousness – Rapid breathing/ increased breathing rate

Carbon monoxide (CO)

Carbon dioxide (CO2) Methane (CH4)

Common sources

– Livestock – Natural gas and petroleum – Landfills

Table 4.1. Common sources and effects of atmospheric gases and particles on the environment and human health (Edited from U.S. EPA)

Remote Sensing Products

97

4.1.1.1. Carbon monoxide (CO) CO is a toxic gas which has both direct and indirect impacts on the environment and human health (Luo et al. 2015). CO is formed by burning fossil fuels such as coal, oil and some other organic substances. Every year, around 600 million tons of CO are produced globally. CO is used as a tracer for air pollution with a strong relationship to climate. At very high levels, CO can cause dizziness, unconsciousness and death. When CO levels rise outdoors, this may be of particular interest to people with certain types of heart disease. They are especially susceptible to the effects of CO during exercise and short-term exposure to high CO can lead to reduced oxygen to the heart with angina (EPA 2019a). CO can be used for anthropogenic activity observation, emissions evaluation, and determining the transport and influence of these emissions downwind of sources. CO is also a major precursor for tropospheric O3 on a global scale. Figure 4.1 shows the time series of the global mean mole fraction of the CO retrievals. The black line represents satellite observation of MOPITTv6 XCO, while the blue and red lines represent the prior and posterior simulations (calculated from model simulations with the MOPITT prior profiles and averaging kernels) respectively. Compared to the MOPITT XCO, the prior XCO simulation is on average 15 and 17%. The global mean posterior XCO fits the observation irrespective of the OH field used (Yin et al. 2015).

Figure 4.1. Time series of the global monthly mean mole fraction in the CO column. For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

98

TORUS 2 – Toward an Open Resource Using Services

4.1.1.2. Nitrogen dioxide (NOx) Nitrogen oxide (NOx) is a term for nitric oxide (NO) and nitrogen dioxide (NO2). NOx is derived from burning fuel and from exhaust gas, from vehicles and power plants. The solubility of this waste, along with the increase in transport, means it has increased environmental pollution in cities. High concentrations of NOx can irritate and inflame your airway mucosa, causing asthma attacks and symptoms such as coughing and shortness of breath. Children and the elderly are also more affected and more likely to have respiratory infections, or react to allergens. NOx reacts with other chemicals to form ozone and particles physically, both of which are also harmful by inhalation due to the respiratory system. NOx reacts with other chemicals to form acid rain, damaging ecosystems such as lakes and forests. NOx in the atmosphere contributes to polluting nutrients in coastal waters (EPA 2019b). Figure 4.2 shows the global trend of NOx emission from 1960 to 2014. For comparison, the trends of NOx emission reported by PKU, EDGARV4.3, POET, RETRO and ACCMIP data sources are shown. The total NOx emissions reached a peak value recently in 2013 (Huang et al. 2017).

Figure 4.2. Global trending of NOx emission from 1906 to 2014 (source: Huang et al. 2017). For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

4.1.1.3. Sulfur dioxide (SO2) Sulfur dioxide (SO2) is a colorless gas with a strong, stifling odor. SO2 is an air pollutant with low concentration in the atmosphere, concentrated mainly in the troposphere. It is a component of the atmosphere that is associated with adverse respiratory effects, and can also form secondary sulfate particles through atmospheric reactions. Acid deposition is another detrimental effect of SO2. It is derived from volcanic eruptions, coal, oil, gas, plant biomass, sulfur ore, etc. SO2 is very toxic to humans and organisms, causing lung diseases. SO2, encountering oxygen and water in the air, forms acid, which, when concentrated in rain water,

Remote Sensing Products

99

causes acid rain. SO2 can harm the human respiratory system and cause breathing difficulties, especially for children, the elderly and people with asthma. High levels of SO2 in the air generally also lead to the formation of other sulfur oxides and react with other atmospheric compounds to form dust particles (PM) and cause additional health problems (EPA 2019c). Figure 4.3 shows the SO2 pollution trend at the global level from 1850 to 2010. Total annual SO2 emissions worldwide, by region in the world, lasting from 1850–2010, are calculated. A rapid increase in sulfur pollution is visible in Europe, followed by North America, in the mid-19th Century. The increase in SO2 emissions in the rest of the world has been delayed until the 20th Century. The emissions from Asia and Africa are relatively small compared to Europe and North America. Today, the annual trend of SO2 emissions in Europe and America continues to decline, while emissions in Asia and Africa are increasing (Ritchie and Roser 2017).

Figure 4.3. Global trend of SO2 pollution from 1850 to 2010 (source: Ritchie and Roser 2017). For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

4.1.1.4. Ozone (O3) Ozone (O3) is an oxygen gas whose molecules are made up of three oxygen atoms linked together. Ozone occurs naturally in the upper atmosphere, the stratosphere, forming a protective layer shielding us from the harmful ultraviolet rays of the sun. This beneficial ozone was destroyed partially by artificial chemicals, sometimes called the “hole in the ozone layer”. Ozone can form through chemical reactions between local air pollutants such as nitrous oxide (NOx), volatile organic compounds (VOC) and sunlight. The ozone layer usually absorbs 97–99%

100

TORUS 2 – Toward an Open Resource Using Services

of incoming UV-B radiation. Therefore, higher ozone concentrations in the stratosphere are very important to ensure that life (including humans) on the Earth’s surface is not exposed to harmful UV-B radiation. Ground ozone is an air pollutant because it is a major component of smog and it affects humans and the environment. These air pollutants are emitted from motor vehicle emissions, industrial processes, electricity and water and chemical solvents. Ozone is especially harmful for young people, the elderly and people with latent respiratory problems (EPA 2019d). The annual stratospheric ozone concentration in the Southern Hemisphere from 1979 to 2017 is shown in Figure 4.4. Ozone concentrations are collected through satellite as part of NASA’s Ozone Watch meteorological program. We find that from 1979 to the early 1990s the stratospheric ozone concentration in the Southern Hemisphere decreased to 100 DU. Since the 1990s, concentrations have continued below 100 DU. However, in the last few years since 2010, ozone levels have begun to recover slowly. In 2018, NASA’s Aura Program announced the first results showing clear early signs of ozone recovery (Ritchie and Roser 2018).

Figure 4.4. Averaged concentration of stratospheric ozone in the Southern Hemisphere from 1979 to 2017 (source: Ritchie and Roser 2018). For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

4.1.1.5. Particulate matter (PM) Particulate matter (PM) is a mixture of solids and liquids, including carbon, sulfate, nitrate, mineral dust and water suspended in the air. These particles are made up of hundreds of different chemicals and are a variety of sizes and shapes. The most harmful are small particles called PM10 and PM2.5. PM10 refers to particles with a diameter of less than 10 microns, while PM2.5 refers to particles with a diameter less than 2.5 microns, known as fine particles. The smallest fine particles, with a diameter less than 0.1 microns, are called ultrafine particles. Artificial PM mainly

Remote Sensing Products

101

comes from industrial processes, construction, emissions from diesel and gasoline engines, friction from brakes and tires, and dust from road surfaces. Diesel engines tend to produce more PM than equivalent gasoline engines. Natural particle resources include volcanoes, seawater, pollen and soil. PM is also formed in the atmosphere when gases such as nitrogen dioxide and sulfur dioxide are changed in the air by chemical reactions. Particulate matter can cause serious health problems. Some particles may go deep into the lungs or may even enter the bloodstream. Among these, particles less than 2.5 micrometers in diameter have the highest risk to health. The fine particles are also major causes of reduced visibility (haze) (EPA 2019e).

Figure 4.5. Distributions of PM2.5 concentrations globally and at the country level for the 10 most populous countries (source: EPA 2019e). For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

The distribution of PM2.5 concentrations globally and at the country level for the 10 most populous countries is examined in Figure 4.5. The plots include PM2.5 trends of all other countries along with all of the country-level PM2.5 for 1990, 1995, 2000, 2005, 2010 and 2013. Large proportional increases in mean PM2.5 concentrations were apparent in India, China, Brazil, Bangladesh and Pakistan, with decreases observed in the United States, Indonesia, Russia, Japan and Nigeria (Brauer et al. 2016). 4.1.1.6. Carbon dioxide (CO2) Carbon dioxide (CO2) is the most important artificial greenhouse gas which is a component of the atmosphere. It is present in the Earth’s atmosphere at low concentrations and acts as a greenhouse gas. CO2 and other greenhouse gases

102

TORUS 2 – Toward an Open Resource Using Services

contribute to global climate warming. CO2 with a concentration of 0.03% in the atmosphere is the raw material for photosynthesis, to produce primary biological productivity in green plants. Normally, the amount of CO2 produced naturally is equal to the amount of CO2 used for photosynthesis. CO2 is obtained from a variety of sources, including volcanic gases, combustion products of organic compounds and respiratory activity of aerobic organisms. Two types of human activities, burning fossil fuels and deforestation, have made this process unbalanced, adversely affecting the global climate (EPA 2019f). Figure 4.6 shows the average monthly carbon dioxide average on the global surface of the sea. Carbon dioxide data and other greenhouse gases were measured by the NOAA/Earth Research Laboratory’s Global Monitoring Laboratory for decades at a network of fully distributed air sampling locations (NOAA 2019). The red lines show the average monthly values while the black lines represent the same after adjusting for the average seasonal cycle. The average concentration in December 2018 was 409.36 ppb while the average concentration in December 2017 was 406.53 ppb. Today’s carbon dioxide level is higher than at any time for at least 800,000 years (Lindsey 2018). According to the State of the Climate in 2017 report, between 2016 and 2017, global annual mean carbon dioxide increased 2.2 ± 0.1 ppm, which was less than the increase between 2015 and 2016 (3.0 ppm per year). Carbon dioxide levels in the atmosphere are expected to increase to nearly a record level in 2019, according to the Met Office (Carrington 2019).

Figure 4.6. Global monthly mean CO2 concentration from 2014 to 2019 (source: NOAA)

Remote Sensing Products

103

4.1.1.7. Methane (CH4) Methane (CH4) is the second most important artificial greenhouse gas after CO2. It can slow down climate change for a short time. It provides a lever for slowing short-term climate change. Major anthropogenic sources include natural gas extraction and use, coal mining, landfills, livestock, rice cultivation, and biomass burning. Wetlands are the largest natural source. In addition, CH4 can be removed from the atmosphere by natural processes in the soil and chemical reactions in the atmosphere. CH4 promotes steam oxidation in the stratosphere. The increase in water vapor causes a much stronger greenhouse effect than the direct effect of CH4. Globally, over 60 percent of total CH4 emissions come from human activities (EPA 2019f). Figure 4.7 shows the global average monthly methane abundance determined from sea surface locations. The red lines show the global average monthly values that focus on the middle of each month. The black line shows the long-term trend in which the seasonal average period has been removed. The average concentration in December 2018 was 1,867.2 ppb, while the average concentration in December 2017 was 1,858.8 ppb.

Figure 4.7. Global monthly mean concentration of CH4 from 2015 to 2019 (source: NOAA)

4.1.2. Introduction to meteorological parameters Meteorology, climatology, atmospheric physics and atmospheric chemistry are subdisciplines of the atmospheric sciences. Meteorological phenomena are described and quantified by the variables of Earth’s atmosphere: temperature, air pressure, water

104

TORUS 2 – Toward an Open Resource Using Services

vapor, mass flow (source: https://en.wikipedia.org/wiki/Meteorology). Earth observation satellites provide important data about the Earth and its environment. Measurements of atmosphere include a number of parameters covering temperature, humidity, component gas concentration, clouds and radiation. In this section, we will focus on introducing some specific meteorological parameters including temperature, pressure, humidity, winds and precipitation. Other atmospheric parameters such as clouds and radiation will not be explored in this chapter. 4.1.2.1. Atmospheric temperature Atmospheric temperature profile data is a core requirement for weather forecast, along with humidity and precipitation. The data is used to track changes in global temperature, to determine the relationship between atmospheric parameters and climate behavior and to validate the atmospheric model worldwide.

Figure 4.8. Global atmosphere temperature trends between 1979 and 2005. For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

Remote Sensing Products

105

Figure 4.8 shows the satellite temperature trends from January 1979 to December 2005. The top image shows temperatures in the middle troposphere while the lower image shows temperatures in the lower stratosphere. The results indicated that the atmosphere near the Earth’s surface warmed during the period. The stratosphere image is dominated by blues and greens, indicating cooling (NASA 2007). 4.1.2.2. Atmospheric winds Atmospheric wind is of utmost importance for weather forecasts and for research on global climate change. Wind speed and direction are fundamental elements of the climate system that affects many other variables. Substantial information can be derived by these methods but quality control is difficult and vertical resolution is poor. Planned instruments for geostationary satellites promise improved information, but the limited vertical resolution and the problems of accurate height assignment of winds will remain areas to be improved. Figure 4.9 shows the global mean month wind speed for January and July (1984 – 2014). A clear seasonal cycle with high wind speeds being seen at high latitudes in the respective winters are observed. The summer–winter cycle in the Northern Hemisphere is much stronger than the Southern Hemisphere. Although the maximum values in both hemispheres are similar in the corresponding winters, the Southern Hemisphere has shown relatively high wind speeds during summer. (Young and Donelan 2018).

Figure 4.9. Monthly mean wind speed for January and July from radiometer (1984– 2014). For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

106

TORUS 2 – Toward an Open Resource Using Services

4.1.2.3. Atmospheric humidity Atmospheric humidity (water vapor) is a core requirement for global and regional weather forecast (NWP) models. Polar satellites provide information on the troposphere moisture with global coverage with acceptable accuracy, but with poor vertical resolution. Time series of global mean monthly precipitable water (PW) anomalies were derived from the GPS, radiosonde, and MWR data (Figure 4.10). It is evident that global mean PW increases generally with time over both land and oceans during recent decades. The trends are statistically significant at the 5% level for radiosonde and MWR but at the 10% level for GPS (Wang et al. 2016).

Figure 4.10. Time series of global mean monthly water vapor over (a) land (1995–2011) and (b) oceans (1988–2011) (source: Wang et al. 2016). For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

4.1.2.4. Atmospheric precipitation Water forms one of the most important constituents of the Earth’s atmosphere. The global water cycle is at the heart of the Earth’s climate system, and better predictions of its behavior are needed for monitoring climate variability and change, weather forecasting and sustainable development of the world’s water resources.

Remote Sensing Products

107

Information on liquid water and precipitation rate is used for initializing NWP models. Figure 4.11 shows the annual mean precipitation for the satellite (1979–2014). The maximum rainfall along and off the east coasts of the continents are noted extending northeastward into the mid-latitudes of the Southern Hemisphere. A weaker but continuous precipitation maximum is observed circling the hemisphere. There are significant variations even on the seasonal scale. The areas of subtropical minima and mid-latitude maxima also evidence seasonal variation in their magnitudes and positions (Adler et al. 2017).

Figure 4.11. Global mean precipitation (mm/day) during 1979–2014. For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

4.1.3. Atmospheric observation from satellite Currently, with the development of the satellite system, atmospheric products can be measured from several satellite instruments. A list of common atmospheric satellites is presented in Table 4.2.

NASA/JAXA

NOAA

NASA/JAXA/INPE

NASA

ESA

TRMM

NOAA15/16/17/18

TERRA

AQUA

ENVISAT

2002–2012

2002

2000

1998–2018

1997–2015

1996–2006

1995–2011

ESA

NASA

ERS-2

Operational period

Sponsoring agency

TOMS-EP

Satellite platform

48 km

AMSU-A

1.7 km 260 m

MERIS

1 × 1 km

GOMOS

AATSR X

X

X

X

48 km

AMSU-A

SCIAMACHY 32 × 215 km

X

X

0.25–1 km

MODIS 13 × 13 km

0.25–1 km

MODIS

AIRS

0.275–1.1 km

MISR

22 × 22 km

MOPITT

X

X X

X

X

20.3 km

X

X

HIRS/3

X

X

Meteorological parameters

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

O3 PM (AOD) CH4 CO2 Temperature Humidity Wind Precipitation

Gases and particles NO2 SO2 CO

ATOVS

16 km

1.1 km

AVHRR/3

AMSU-B

4.3 km

39 × 39 km

40 × 320 km

Spatial resolution

PR

TOMS

GOME

Instrument

108 TORUS 2 – Toward an Open Resource Using Services

Sponsoring agency

EUMETSAT/ESA

NASA

ESA/EUMETSAT

NASA/CNES

JAXA/MOE/NIES

NASA/NOAA

NASA

ESA/COM/NSO

Satellite platform

Meteosat8/9/10/11

AURA

METOPA/B/C

CALIPSO

GOSAT1/2

Suomi NPP/JPSS

OCO-2

Sentinel-5P

2017

2014

2011

2009

14 km

5.5 × 5.5 km X

X

X

X

X

X

X

X

X

Meteorological parameters

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

X

O3 PM (AOD) CH4 CO2 Temperature Humidity Wind Precipitation

Gases and particles NO2 SO2 CO

Table 4.2. Summary of the major satellite instruments

TROPOMI

2.25 km

CrIS OCO-2

1.1o km

0.75 × 0.75 km

VIIRS ATMS

0.5–1.5 km

TANSO-CAI

25 km

40 × 80 km

10.5 km

IASI 5 × 5 km

GOME-2

2006

2006

5.3 × 8.5 km

CALIOP

TES

2004

13 × 24 km

1–3 km

Spatial resolution

TANSO-FTS

OMI

2004

2006

SEVIRI

Instrument

2002

Operational period

Remote Sensing Products 109

110

TORUS 2 – Toward an Open Resource Using Services

4.1.3.1. ERS-2 ERS-2 is ESA’s second EO mission (after ERS-1), and was launched in 1995. It included two new instruments GOME and ATSR-2. ATSR-2 included 3 visible spectrum bands specialized for chlorophyll and vegetation. ERS-2 has a repeat cycle of 35 days. ERS-2 was finally depleted of all fuel on 5 September 2011. GOME is a cross-track scanning optical double spectrometer, a nadir viewing instrument. The double spectrometer operates in the spectral range of 240 – 790 nm with a good spectral resolution of 0.2 – 0.4 nm. The field of view may be varied in size from 40 km × 40 km to 320 km × 40 km. GOME is used to observe total amounts and profiles of ozone on a daily basis. It also measures H2O and other gases including NO2, OClO, BrO etc. GOME can be used to observe the distribution of atmospheric aerosols. The GOME/ERS-2 trace gas column densities and cloud properties (Level 2 product) are retrieved from GOME (ir)radiance and PMD data (Level 1 product). The table below includes the trace gas and cloud products and the corresponding wavelength regions. The GOME Level 2 data product comprises the product header, total column densities of ozone and nitrogen dioxide (DLR 2012).

Figure 4.12. GOME Level 2 data products

GOME/ERS-2 has three forward-scan pixels with a nominal resolution of 40 km × 320 km, and one back-scan pixel with a nominal resolution of 40 km × 960 km. GOME/ERS-2 total column products are generated at DLR’s D-PAF on behalf of ESA. These data could be ordered from the ESA ESRIN-EO Help Desk ([email protected]). Some GOME/ERS-2 data are shown in Figure 4.13.

Remote Sensing Products

111

Figure 4.13. O3 and NO2 data from GOME/ERS-2. For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

4.1.3.2. TOMS-EP The Total Ozone Mapping Spectrometer, onboard an Earth Probe Satellite (TOMS/EP) was launched in July 1996 to produce mapping of the global distribution of atmospheric ozone. It measures ozone indirectly by mapping ultraviolet light emitted by the Sun to that scattered from the Earth’s atmosphere back to the satellite. TOMS also measures sulfur dioxide released in volcanic eruptions. TOMS/EP resolution was 47 km at nadir with a swath width of 2750 km. Retrieval of total ozone is based on a comparison between the measured normalized radiances and radiances derived by radiative transfer calculations for different ozone amounts and the conditions of the measurement. Table 4.3 shows the list of atmospheric products from TOMS-EP. Figure 4.14 shows the total ozone zonal mean from TOMS-EP in 1998. These data are provided at GES DISC system (source: https://disc.gsfc.nasa.gov/). Products

TOMSEPL3

TOMSEPL3mtoz

TOMSEPL3dtoz TOMSEPL3ztoz

Description TOMS Earth-Probe Total Ozone (O3) Aerosol Index UV-Reflectivity UV-B Erythemal Irradiances Daily L3 Global 1 deg × 1.25 deg V008 TOMS Earth Probe Total Column Ozone Monthly L3 Global 1 deg × 1.25 deg Lat/Lon Grid V008 TOMS Earth Probe Total Column Ozone Daily L3 Global 1 deg × 1.25 deg Lat/Lon Grid V008 OMS EP Total Column Ozone Daily and Monthly Zonal Means V008

Temporal resolution

Spatial resolution

Start date– end date

Daily

1 × 1.25 deg

1996–2005

Monthly

1 × 1.25 deg

1996–2005

Monthly

1 × 1.25 deg

1996–2005

Monthly

5 deg

1996–2005

Table 4.3. TOMS-EP atmospheric product

112

TORUS 2 – Toward an Open Resource Using Services

Figure 4.14. TOMS ozone zonal mean for 1998. For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

4.1.3.3. TRMM The Tropical Rainfall Measuring Mission (TRMM) was developed by NASA and the Japan Aerospace Exploration Agency (JAXA) to observe rainfall for weather and climate research. The TRMM satellite launched in late November 1997 and stopped collecting data on April 15, 2015. The TRMM satellite produced over 17 years of valuable scientific data. TRMM carried 5 instruments: a 3-sensor rainfall suite (PR, TMI, VIRS) and 2 related instruments (LIS and CERES). The VIRS (of NOAA AVHRR heritage) provides high resolution observations on cloud coverage, cloud type, and cloud top temperatures while the TMI (of DMSP

Remote Sensing Products

113

SSM/I heritage) provides information on the integrated column precipitation content, cloud liquid water, cloud ice, rain intensity, and rainfall types. The PR is an electronically scanning radar operating at 13.8 GHz that measures the 3D rainfall distribution and defines the layer depth of the precipitation. CERES will measure the energy at the top of the atmosphere, as well as estimate energy levels within the atmosphere and at the Earth’s surface while the Lightning Imaging Sensor (IS) detects and locates lightning over the tropical region of the globe. TRMM delivered a unique 17-year dataset of global tropical rainfall and lightning. The data also supported operational applications such as flood and drought monitoring and weather forecasting. Table 4.4 shows the list of atmospheric product from TRMM. All data are provided at GES DISC system. Products

Instrument TRMM PR, TRMM TMI

TRMM_2B31.7

TRMM TMI TRMM_2A12.7

TRMM PR TRMM_2A21.7

TRMM PR TRMM_2A23.7

TRMM PR TRMM_2A25.7

TRMM PR, TRMM TMI GPM_2BCMB_TRMM.06

Temporal resolution

Spatial resolution

Start date– end date

TRMM Combined Precipitation Radar and Microwave Imager Rainfall Profile L2 1.5 hours V7

90 min

5 × 5 km

1997–2015

TRMM Microwave Imager Hydrometeor Profile L2 1.5 hours V7

90 min

5.1 × 5.1 km

1997–2015

TRMM Precipitation Radar Surface Cross-Section L2 1.5 hours V7

90 min

5 × 5 km

1997–2015

TRMM Precipitation Radar Rain Characteristics L2 1.5 hours V7

90 min

4 × 4 km

1997–2015

TRMM Precipitation Radar Rainfall Rate and Profile L2 1.5 hours V7

90 min

4 × 4 km

1997–2015

GPM PR and TMI on TRMM Combined Precipitation L2B 1.5 hours 5 km V06

90 min

5 × 5 km

1997–2015

Description

114

TORUS 2 – Toward an Open Resource Using Services

TRMM PR

GPM_2APR.06

TRMM TMI GPM_2AGPROFTRMMTMI_ CLIM.05

GPM PR on TRMM Reflectivity, Precipitation Characteristics and Rate, at Surface and Profile L2 1.5 hours V06 GPM TMI on TRMM (GPROF) Climate-based Radiometer Precipitation Profiling L2A 1.5 hours 13 km V05

90 min

5 × 5 km

1997–2015

90 min

13 × 5 km

1997–2015

Table 4.4. List of atmospheric products from TRMM

4.1.3.4. NOAA-15/16/17/18 NOAA-15 is one of the NASA-provided TIROS series of weather forecasting satellite run by NOAA. It was launched on May 13 1998, and is currently operational in a sun-synchronous orbit 807 km above the Earth, orbiting every 101 minutes. It carries 5 instruments including the AMSU-A and AMSU-B instruments, the AVHRR and High Resolution Infrared Radiation Sounder (HIRS/3) instruments and Space Environment Monitor (SEM/2). NOAA-15 replaced the decommissioned NOAA-12 in an afternoon equatorcrossing orbit. It provided support to environmental monitoring including Earth radiation, atmospheric ozone, aerosol distribution, sea surface temperature, vertical temperature and water profiles in the troposphere and stratosphere. NOAA-16 was launched on September 21 2000, in a sun-synchronous orbit 849 km above the Earth. NOAA-16 has the same suite of instruments as carried by NOAA-15 plus an SBUV/2 instrument as well. NOAA-17 was launched on June 24 2002 at 824 km above the Earth, orbiting every 101 minutes. It carries the AMSU, AVHRR and High Resolution Infrared Radiation Sounder (HRIS) instruments. The satellite was retired in 2013. NOAA-18 was launched on May 20 2005 at an altitude of 854 km, with an orbital period of 102 minutes. It is the first NOAA POES satellite to use MHS in place of AMSU-B. NOAA-19 was launched on February 6, 2009. NOAA-19 carries a suite of instruments that provide data for weather and climate predictions. It provides global observation of clouds and surface features and vertical profiles of atmospheric temperature and humidity as well as data on ozone distribution in the upper part of

Remote Sensing Products

115

the atmosphere, and near-Earth space environments – information important for the marine, aviation, power generation and agriculture, among others communities. Table 4.5 show the list of atmospheric product from NOAA satellite. All data are provided at GES DISC system. Products

Instrument

Description

Temporal resolution

Spatial resolution

Start date–end date

1 month

5° zonal

1970–2013

3 hours

0.25° × 0.25°

1998–2010

MSO3L3zm5.1

NOAA-16 SBUV/2, NOAA17 SBUV/2, NOAA-18 SBUV/2, NOAA19 SBUV/2

Multi-Satellite Merged Ozone (O3) Profile and Total Column 1 Month Zonal Mean L3 Global 5.0 degree Latitude Zones V1

TMI/TRMM precipitation and uncertainty (TMPA) L3 3 hour 0.25 degree × 0.25 degree V001

WC_MULTISEN _PREC_025.001

NOAA-15 AMSU-B, NOAA-16 AMSU-B, NOAA-17 AMSU-B, NOAA-18 MHS, NOAA-19 MHS

SBUV2N16L3zm.1

NOAA-16 SBUV/2 SBUV2/NOAA-16 Ozone (O3) Profile and Total Column Ozone 1 Month Zonal Mean L3 Global 5.0 degree Latitude Zones V1

1 month

5° zonal

2000–2013

SBUV2N17L3zm.1

NOAA-17 SBUV/2 SBUV2/NOAA-17 Ozone (O3) Profile and Total Column Ozone 1 Month Zonal Mean L3 Global 5.0 degree Latitude Zones V1

1 month

5° zonal

2002–2013

SBUV2N19L3zm.1

NOAA-19 SBUV/2 SBUV2/NOAA-19 Ozone (O3) Profile and Total Column Ozone 1 Month Zonal Mean L3 Global 5.0 degree Latitude Zones V1

1 month

5° zonal

2009–2013

SBUV2N16L2.1

NOAA-16 SBUV/2 SBUV2/NOAA-16 Ozone (O3) Nadir Profile and Total Column 1 Day L2 V1

1 hour

11.3° × 11.3°

2000–2013

SBUV2N16O3.008

NOAA-16 SBUV/2 SBUV2/NOAA-16 Level 2 Daily Ozone Profile and Total Column from CD-ROM V008

1 day

180 km × 180 km

2000–2003

SBUV2N17L2.1

NOAA-17 SBUV/2 SBUV2/NOAA-17 Ozone (O3) Nadir Profile and Total Column 1 Day L2 V1

1 hour

11.3° × 11.3°

2002–2013

SBUV2N19L2.1

NOAA-19 SBUV/2 SBUV2/NOAA-19 Ozone (O3) Nadir Profile and Total Column 1 Day L2 V1

1 hour

11.3° × 11.3°

2009–2013

Table 4.5. List of atmospheric products from NOAA satellite

4.1.3.5. TERRA TERRA (EOS AM-1) is a Sun-synchronous orbit passing over the equator in the morning. The satellite was launched on December 18 1999, and began collecting

116

TORUS 2 – Toward an Open Resource Using Services

data on February 24 2000. It was placed into a near-polar, sun-synchronous orbit at an altitude of 705 km (438 mi), with a 10:30am descending node. TERRA carries five sensors including ASTER (Advanced Spaceborne Thermal Emission and Reflection Radiometer), CERES (Clouds and the Earth’s Radiant Energy System), MISR (Multi-angle Imaging SpectroRadiometer), MODIS (Moderate-resolution Imaging Spectroradiometer) and MOPITT (Measurements of Pollution in the Troposphere). ASTER observes clouds, ice, water and the land surface at high-resolution using 3 different sensors: Shortwave Infrared (SWIR); Thermal Infrared (TIR); and Visible and Near Infrared (VNIR). They cover 14 multi-spectral bands from visible to the thermal infrared. The MISR instrument consists of nine separate digital cameras that collect data in four different spectral bands. One camera points toward the nadir, while the others provide forward and backward view angles at 26.1°, 45.6°, 60.0°, and 70.5°. Each region of the Earth’s surface is successively imaged by all nine cameras in each of four wavelengths (blue, green, red, and near-infrared). MISR employs a unique multiangle design that allows it to observe the atmosphere through different effective path lengths, leading to a very different aerosol retrieval algorithm (Martonchik et al. 1998). MISR provides a rich aerosol dataset to study long-term spatial and temporal trends of particle mass, composition, and other information. MISR-retrieved AOD is reported in its Level 2 aerosol data product (MIL2ASAE) on a 17.6 km × 17.6 km grid. The MISR aerosol algorithm retrieves AOD and aerosol type by analyzing MISR TOA radiances from 16 × 16 pixel patches of 1.1 km resolution (Kahn et al. 2001). MOPITT is a nadir sounding instrument that measures upwelling infrared radiation at 4.7 μm and 2.2–2.4 μm. It uses correlation spectroscopy to calculate total column observations and profiles of carbon monoxide in the lower atmosphere. Although observations of methane were also planned, to date no data have been released. The MOPITT is a nadir-viewing gas correlation radiometer operating in the 4.7 mm band of carbon monoxide. The MOPITT pixel is 22 × 22 km2 at nadir with a 29 pixel wide swath. MOPITT make measurements of tropospheric carbon monoxide (CO) on the global scale. MOPITT has been operational since March 2000. MOPITT Version 8 Level 2 and Level 3 products are now available for the entire MOPITT mission. Level 2 data include CO mixing ratio profiles and CO total column with 22 km horizontal resolution, 4 km vertical resolution and 10% precision. Level 3 products are gridded global CO distributions at 1 deg × 1 deg, daily and monthly averages.

Remote Sensing Products

117

MOPITT CO retrieval products have been refined continuously since the launch of TERRA. The MOPITT CO retrieval algorithm is based on the maximum likelihood method (Wang et al. 1999). The MODIS instrument has 36 channels with varying spatial resolution of 250, 500 and 1000 m, depending on the channel. The channels span the spectral range from 410 to 14,200 nm, and bandwidth varies from channel to channel. Two independent retrievals are conducted at 470 and 660 nm, and subsequently interpolated to 550 nm. The surface reflectances for the channels at 470 and 660 nm are estimated from measurements at 2.1 mm using empirical relationships. Table 4.6 presents a list of atmospheric products from the TERRA satellite. Products

Description

MODIS/Terra Aerosol 5-Min L2 Swath 10 km MODIS/Terra Aerosol 5-Min L2 MOD04_3K Swath 3 km MODIS/Terra Temperature and MOD07_L2 Water Vapor Profiles 5-Min L2 Swath 5 km MODIS/Terra Total Precipitable MOD05_L2 Water Vapor 5-Min L2 Swath 1 km and 5 km MOPITT Derived CO (Near and MOP02J.008 Thermal Infrared Radiances) V008 MOPITT Derived CO (Near MOP02N.008 Infrared Radiances) V008 MOPITT Derived CO (Thermal MOP02T.008 Infrared Radiances) V008 MISR Level 2 Version 3 Aerosol MIL2ASAE Data MOD04_L2

Temporal resolution

Spatial resolution

Start date– end date

Daily

10 km

2000

Daily

3 km

2000

Daily

5 km

2000

Daily

1–5 km

2000

Daily

22 km

2000

Daily

22 km

2000

Daily

22 km

2000

Daily

17.6 km

2000

Table 4.6. List of atmospheric products from Terra satellite

4.1.3.6. AQUA AQUA (EOS PM-1) is the second major component of the Earth Observing System (EOS) preceded by TERRA (launched 1999) and followed by AURA (launched 2004). AQUA carries six instruments including AMSR-E, MODIS, AMSU-A, AIRS, HSB and CERES. AIRS is a cross-track scanning grating spectrometer that covers the 3.7–16 mm spectral range with 2378 channels and a 13.5 km nadir field of view. CO retrievals are conducted at 4.7 mm with a spatial resolution of 45 km at nadir, yielding a profile with

118

TORUS 2 – Toward an Open Resource Using Services

0.5–1.5 DFS. AIRS provides mid-tropospheric carbon dioxide over ocean and land data. AIRS retrievals are based on cloud-cleared thermal infrared radiance and associated Level 2 geophysical profiles of temperature, water vapor and ozone. The Standard Retrieval Product (AIRS2STC) consists of retrieved estimates of CO2, plus estimates of the errors associated with the retrieval. The horizontal resolution is about 110 km (1 × 1 degree). An AIRS granule has been set as 6 minutes of data, 15 footprints cross track by 22 lines along track. Version 5 Level 3 carbon dioxide data have 2° latitude × 2.5° longitude grid boxes. AIRS3C2D is daily gridded data and AIRX3C2M is a monthly product at 2.5 × 2 deg grid cell size. The AIRS broad swath makes it able to map the global distribution of carbon dioxide every day. Detail about AIRS CO2 are showed in (Chahine et al. 2008, Olsen et al. 2008). The AIRS Version 5 (V5) tropospheric CO2 product is derived by the technique of vanishing partial derivatives (VPD) (Chahine et al. 2005). The VPD CO2 is obtained by an iterative process that minimizes the RMS residuals between the Level 2 cloudcleared radiances and forward-computed radiances from the retrieved Level 2 atmospheric state for selected channels. The process begins with the AIRS V5retrieved Level 2 atmospheric state and CO2 climatology and then separately perturbs the T(p), H2O(p), O3 (p) and CO2. The auxiliary T(p), H2O(p) and O3 (p) channels are used to accelerate the iteration process, i.e. reducing the number of iterations required. The solution is obtained when the residuals are individually minimized. Table 4.7 presents a list of atmospheric products from the AQUA satellite. Products

Description

Temporal resolution

Spatial resolution

Start date– end date 2002

MYD04_L2

MODIS/Aqua Aerosol 5-Min L2 Swath 10 km

Daily

10 km

MYD04_3K

MODIS/Aqua Aerosol 5-Min L2 Swath 3 km

Daily

3 km

MYD07_L2

MODIS/Aqua Temperature and Water Vapor Profiles 5-Min L2 Swath 5 km

Daily

5km

MYD05_L2

MODIS/Aqua Total Precipitable Water Vapor 5-Min L2 Swath 1 km and 5 km

Daily

1–5 km

AIRS2SUP.006

AIRS/Aqua L2 Support Retrieval (AIRS-only) V006

Daily

50 km

2002

AIRX2SUP.006

AIRS/Aqua L2 Support Retrieval (AIRS+AMSU) V006

Daily

50 km

2002–2016

AIRS2SPC.005

AIRS/Aqua L2 CO2 support retrieval (AIRS-only) V005

Daily

90 km

2010–2017

2002 2002

2002

Table 4.7. List of atmospheric products from AQUA satellite

Remote Sensing Products

119

4.1.3.7. ENVISAT ENVISAT, was launched in 2002, provided ice and ocean observations and added important new capabilities for understanding and monitoring our environment. ENVISAT was the largest and most complex satellite ever built in Europe. Its package of ten instruments made major contributions to the global study and monitoring of the Earth and its environment, including global warming, climate change, ozone depletion, and ocean and ice monitoring. More information on the ENVISAT mission can be found at http://envisat.esa.int. (Laur and Liebig, 2014). ENVISAT carried the SCanning Imaging Absorption spectrometer for Atmospheric CHartogaphY (SCIAMACHY) (Bovensmann et al. 1999), observed trace gases in the troposphere: NO2, SO2, CO, CH4, NMVOC, PM (AOD), CO2. SCIAMACHY measures backscattered solar radiation upwelling from the atmosphere, alternately in nadir and limb viewing geometry. Eight channels, comprised of grating optics and a linear diode array detector, measure the spectrum over 214–1750 nm at a resolution of 0.2–1.4 nm, and two spectral bands around 2.0 and 2.3 mm, having a spectral resolution of 0.2 nm. The typical spatial resolution of SCIAMACHY is 30 – 60 km2 (Martin 2008). SCIAMACHY is the first instrument on the ENVISAT satellite which showed the necessary observation of CO2 in the near-infrared (NIR) spectral regions (Buchwitz et al. 2007). The XCO2 has been retrieved using WFMDOAS (Schneising et al. 2008) and BESD retrieval algorithm (Reuter et al. 2016). The horizontal resolution (size of a single ground pixel) is typically 30 km along track (approximately north-south) and 60 km across track (approximately east-west). Full coverage in nadir is achieved at the equator in six days. The monthly XCO2 product has the spatial resolution of 0.5 × 0.5 degrees. All products are provided between 2002 and 2012. The product detail and quality is presented in many previous studies (Jones et al. 1912, Noël et al. 1999, Buchwitz et al. 2004, Richter et al. 2007, Hewitt et al. 2007, Buchwitz et al. 2007, Barkly et al. 2006). GOMOS is a medium resolution spectrometer covering the wavelength range from 250 nm to 950 nm. GOMOS performs multi-spectral observation of a selected set of star occultations of the Earth’s limb caused as a result of satellite motion. In addition to monitoring ozone, GOMOS monitors other atmospheric trace gases (H2O, NO2, NO3, and OClO), temperature and aerosols.

120

TORUS 2 – Toward an Open Resource Using Services

MERIS is a medium-spectral resolution, imaging spectrometer which scans the Earth’s surface by the so-called “push-broom” method. Linear CCD arrays provide spatial sampling in the across-track direction, while the satellite’s motion provides scanning in the along-track direction. The instrument’s 68.5° field of view around nadir covers a swath width of 1150 km. This wide field of view is shared between five identical optical modules arranged in a fan shape configuration. Table 4.8 presents a list of atmospheric products from the ENVISAT satellite. Products

Description

Temporal resolution

ENVISAT.GOM.NL GOMOS Level 2 - Atmospheric _2P constituents profiles

Daily

GOMOS Level 2 - Atmospheric NetCDF.GOMOS_U constituents profiles - Gridded FP_Gridded User Friendly Product

Daily

SCI_OL__2P

SCIAMACHY Total column densities and stratospheric profiles

Spatial resolution

Daily

Start date– end date 2002–2012

2002–2012

Table 4.8. List of atmospheric products from ENVISAT satellite

4.1.3.8. AURA AURA (EOS CH-1) is the third major component of the Earth Observing System (EOS) following TERRA (launched 1999) and AQUA (launched 2002). AURA follows on from the Upper Atmosphere Research Satellite (UARS). AURA is a multi-national NASA scientific research satellite which carries four instruments for studies of atmospheric chemistry: HIRDLS, MLS, OMI, TES. OMI uses two-dimensional CCD detectors to measure the solar radiation backscattered over 270–500 nm with a spectral resolution of 0.5 nm. The OMI spatial resolution is 13 × 24 km2 at nadir. Retrieval algorithms of tropospheric trace gases and their expected uncertainty for OMI are similar to those for GOME-1 and SCIAMACHY. For retrieving aerosols through OMI, two types of aerosol-retrieval algorithms are in use: the near-UV aerosol algorithm (OMAERUV) and the multiwavelength algorithm (OMAERO), with a pixel resolution of 13 × 24 km at nadir. The OMAERUV uses two UV wavelengths (354 and 388 nm) for retrieving aerosol extinction and absorption optical depth (AAOD), whereas the OMAERO is a multiwavelength algorithm having 19 channels (330–500 nm).

Remote Sensing Products

121

TES is a Fourier transform infrared emission spectrometer with high spectral resolution (0.1 cm-1) and coverage over a wide spectral range (650–3,050 cm-1). The TES nadir footprint is 5 × 8 km2 and has 71 observations per orbit (spaced approximately 175 km apart). Tropospheric O3 and CO are retrieved with an optimal estimation method. TES was designed to measure the global, vertical distribution of tropospheric ozone and ozone precursors such as carbon monoxide (Beer et al. 2001, Beer 2006). The footprint of each nadir observation is 5 km × 8 km, averaged over detectors. TES is on the EOS-AURA platform (source: http://aura.gsfc.nasa.gov/) in a near-polar, Sun-synchronous, 705 km altitude orbit. The ascending node equator crossings are near 1:45 pm local solar time. The TES CO2 data are provided in the period of 2005–2012. The spatial resolution of TES CO2 is 0.5 × 5 km at nadir (Herman et al. 2013). TES CO2 estimates are derived from thermal IR radiances measured aboard NASA’s AURA satellite. ENANOR is the algorithm for a single retrieval step for a subset of retrieved species using spectral micro-windows of L1b data. The development of TES CO2 data relied on a combination of guidance from validation with in situ CO2 data and predictive calculations of error and information. CO2 is estimated by iteratively minimizing a cost function using the Levenberg– Marquardt nonlinear least squares (NLLS) algorithm (Machida et al. 2010). Table 4.9 presents a list of atmospheric products from the AURA satellite. Products

Description

Temporal resolution

Spatial resolution

Start date– end date

ML2DGG.004

MLS/Aura Level 2 Diagnostics, Geophysical Parameter Grid V004

Daily

165 × 3 km

2004

OMAEROZ.003

OMI/Aura Aerosol product Multiwavelength Algorithm Zoomed 1Orbit L2 Swath 13 × 12 km V003

32 day

13 × 12 km

2004

OMIAuraAER.1

OMI/Aura Near UV Aerosol Index, Optical Depth and Single Scattering Albedo 1-Orbit L2 13 × 24 km

Daily

13 × 24 km

2004

OMO3PR.003

OMI/Aura Ozone (O3) Profile 1Orbit L2 Swath 13 × 48 km V003

Daily

13 × 48 km

2004

OMAERUV.003

OMI/Aura Near UV Aerosol Optical Depth and Single Scattering Albedo 1-orbit L2 Swath 13 × 24 km V003

Daily

13 × 48 km

2004

OMDOAO3.003

OMI/Aura Ozone (O3) DOAS Total Column 1-Orbit L2 Swath 13 × 24 km V003

Daily

13 × 24 km

2004

OMNO2.003

OMI/Aura Nitrogen Dioxide (NO2) Total and Tropospheric Column 1orbit L2 Swath 13 × 24 km V003

Daily

13 × 24 km

2004

122

TORUS 2 – Toward an Open Resource Using Services

OMSO2.003

OMI/Aura Sulphur Dioxide (SO2) Total Column 1-orbit L2 Swath 13 × 24 km V003

Daily

13 × 24 km

2004

ML2RHI.003

MLS/Aura Level 2 Relative Humidity With Respect To Ice V003

Daily

165 × 3 km

2004–2015

Daily

165 × 3 km

2004

MLS/Aura Near-Real-Time L2 ML2SO2_NRT.004 Sulfur Dioxide (SO2) Mixing Ratio V004 ML2SO2.003

MLS/Aura Level 2 Sulfur Dioxide (SO2) Mixing Ratio V003

Daily

165 × 3 km

2004–2015

ML2CO.004

MLS/Aura Level 2 Carbon Monoxide (CO) Mixing Ratio V004

Daily

165 × 3 km

2004

ML2O3.004

MLS/Aura Level 2 Ozone (O3) Mixing Ratio V004

Daily

165 × 3 km

2004

Table 4.9. List of atmospheric products from AURA satellite

4.1.3.9. METOP-A/B/C MetOp, launched in 2006 in partnership between ESA and EUMETSAT, is Europe’s first polar-orbiting satellite dedicated to operational meteorology. MetOp is a series of three satellites to be launched sequentially over 14 years, forming the space segment of EUMETSAT’s Polar System (EPS). MetOp carries a set of “heritage” instruments provided by the United States and new instruments to measure temperature humidity, wind speed and direction and atmospheric ozone profiles. The IASI instrument is designed to measure the infrared spectrum emitted by the Earth in the thermal infrared. The IASI instrument field of view is sampled by a matrix of 2 × 2 circular pixels of 12 km diameter each. Measurements are taken every 50 km at nadir with broad horizontal coverage. The system generates Level 1C Thinned (L1CT) radiance and Level 2 profile products. Currently, the IASI Level 2 products from the MetOp-2 satellite include CO2 on a global scale. The spatial resolution of IASI CO2 is 12 × 12 km. IASI CO2 products are retrieved using an inversion algorithm based on artificial neural networks (ANNs). The ANNs were trained with simulated radiances using the RTTOV model and the atmospheric composition profiles from the MOZART model (García et al. 2016). CO observations are also made by IASI on the MetOp-A satellite (George et al. 2009). These measurements are generally most sensitive to CO in the middle troposphere, with limited vertical profile information, available only under favorable

Remote Sensing Products

123

conditions (e.g. Deeter et al. 2004). Using solar reflection at 2.3 µm allows for retrieval of the total CO column with measurement sensitivity to all altitudes, including the lowermost troposphere. 4.1.3.10. CALIPSO The Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observation (CALIPSO) satellite combines an active lidar instrument with passive infrared and visible imagers to provide global clouds and aerosols. CALIPSO was launched on April 28, 2006 with the cloud profiling radar system on the CloudSat satellite. It retrieves the vertical distribution of aerosols and clouds on a global scale. The CALIPSO aerosol-retrieval process is derived on the basis of cluster analysis of a multiyear (1993–2002) AERONET dataset. Six AERONET aerosol clusters were considered based on observed physical and optical properties (Omar et al. 2005). Table 4.10 presents a list of atmospheric products from the CALIPSO satellite. Description

Temporal resolution

Spatial resolution

Start date–end date

CAL_LID_L2_05kmALayStandard-V4-20

CALIPSO Lidar Level 2.5 km Aerosol Layer Data

Daily

5 km

2006

CAL_LID_L2_05kmAProStandard-V4-20

CALIPSO Lidar Level 2 Aerosol Profile Data

Daily

5 km

Products

2006

Table 4.10. List of atmospheric products from CALIPSO satellite

4.1.3.11. GOSAT-1/2 In January 23 2009, Japan Aerospace Exploration Agency (JAXA), and collaborating institutions, launched Greenhouse gases Observing SATellite (GOSAT) to measure CO2 and CH4 concentrations (Kuze et al. 2009). GOSAT (Greenhouse gases Observing SATellite) was developed by the Japan Aerospace Exploration Agency (JAXA) and managed by the National Institute for Environmental Studies (NIES). It was launched on January 23 2009, for and data collection started in February 2009. The concentrations of CO2 and CH4 at the observation points are estimated using a global gas transportation model. The FTS SWIR data products provide the total amount of CO2 and CH4 columns obtained from the spectra mapped in the 1 to 3 band of FTS. The FTS TIR Level 2 data products are the vertical concentrations of CO2 and CH4 derived from the spectra mapped in the FTS range 4. The FTS SWIR Level 3 product is created by interpolating, extrapolating and smoothing the mean values of the FTS SWIR column at Level 2 for CO2 and CH4 monthly.

124

TORUS 2 – Toward an Open Resource Using Services

(Saitoh et al. 2009) provided an algorithm for retrieving CO2 concentrations from the TIR band of TANSO-FTS. GSAT SWIR CO2 retrieval algorithm was provided by (Yoshida et al. 2017). This algorithm uses FTS SWIR Level 1B data as input, then performs pre-processing, filtering, retrieval processing, and screening to calculate column amounts of carbon dioxide (CO2). Table 4.11 lists all types of the GOSAT data products provided for general users. The Level 1 data (FTS Level 1B, CAI Level 1B, and CAI Level 1B+ data) contain spectra and radiances acquired by the satellite. The higher level data products (FTS Level 2, CAI Level 2, FTS Level 3, CAI Level 3, Level 4A, and Level 4B data products) store atmospheric concentrations of CO2 and CH4. Data users can search and order these data products by accessing GDAS. Processing Level

Sensor/Band FTS

Product Name

Description

Unit

FTS L1B data

Radiance spectral data obtained by performing Fourier transform on interferogram data

FTS scene

CAI L1B data

Radiance data (band-to-band and geometric corrections applied / data mapping not performed)

L1B CAI L1B+

CAI

FTSSWIR L2 FTSTIR CAI

FTSSWIR

L3

L4B

Radiance data (band-to-band and geometric corrections applied / data mapping performed)

L2 CO2 column amount (SWIR)

CO2 column abundance data retrieved from SWIR radiance spectral data

L2 CH4 column amount (SWIR)

CH4 column abundance data retrieved from SWIR radiance spectral data

L2 H20 column amount (SWIR)*1

H2O column abundance data retrieved from SWIR radiance spectral data

L2 CO2profile (TIR)*2

CO2vertical profile data retrieved from TIR radiance spectral data

L2 CH4 profile (TIR)*2

CH4 vertical profile data retrieved from TIR radiance spectral data

L2 cloud flag

Cloud coverage data

L3 global CO2 distribution (SWIR)

CO2column-averaged mixing ratio data projected on a global map

L3 global CH4distribution (SWIR) L3 global radiance distribution

CAI

L4A

CAI L1B+ data

–

HDF5 CAI frame

1 – multiple scans

CAI frame

HDF5

Global (monthly CH4 column-averaged mixing ratio data projected average) on a global map Global radiance distribution data (3 days’ worth, including data for cloudy segments)

L3 global reflectance Clear-sky radiance data (composed only of cleardistribution sky segments from a month worth of data)

Global

L3 NDVI

Vegetation index global distribution data (cloudy Lat. 30º x Lon. 60º segments excluded)

L4A global CO2 flux

CO2flux per each of 64 global regions (monthly average)

64 regions across the globe, 1º mesh (annual)

CH4 flux per each of 43 global regions (monthly L4A global CH4 flux average)

43 regions across the globe, 1º mesh (annual)

–

Format

L4B global CO2distribution

Three-dimensional global distribution of CO2 concentration

L4B global CH4 distribution

Three-dimensional global distribution of CH4 concentration

Global 2.5º mesh (monthly)

Table 4.11. GOSAT data products. For a color version of this table, see www.iste.co.uk/laffly/torus2.zip

Text/ Net CDF

Net CDF

Remote Sensing Products

125

4.1.3.12. Suomi NPP/JPSS The Suomi National Polar-orbiting Partnership or Suomi NPP was launched in 2011 and continues to operate. The satellite orbits the Earth about 14 times each day. Its five imaging systems include: ATMS, CrIS, OMPS, VIIRS and CERES. The Visible Infrared Imaging Radiometer Suite (VIIRS) is a remote-sensing cross-track scanning radiometer as a successor to AVHRR and MODIS. VIIRS aerosol retrievals are made over M-bands (0.412–12.016 μm) while AOD is specifically retrieved at 550 nm. It has the capability of retrieving the Angstrom exponent and aerosol type over land and can distinguish aerosol fine and coarse mode fractions over ocean (Jackson et al. 2013). Table 4.12 shows the list of atmospheric product from NPP satellite. Products AERDB_L2_VIIRS _SNPP

Description

VIIRS/SNPP Deep Blue Aerosol L2 6-Min Swath 6 km GPM ATMS on SUOMI-NPP GPM_2AGPROFNP (GPROF) Radiometer Precipitation PATMS Profiling L2 1.5 hours 16 km V05 GPM ATMS on SUOMI-NPP GPM_2AGPROFNP (GPROF) Radiometer Precipitation PATMS_CLIM Profiling L2 1.5 hours 16 km V05 OMPS_NPP_NMSO OMPS/NPP PCA SO2 Total Column 2_PCA_L2.1 1-Orbit L2 swath 50 × 50 km V1 OMPS_NPP_NPBU OMPS-NPP L2 NP Ozone (O3) VO3_L2.2 Vertical Profile swath orbital V2 OMPS-NPP L2 NM Sulfur Dioxide OMPS_NPP_NMSO (SO2) Total and Tropospheric 2_L2.2 Column swath orbital V2 OMPS-NPP L2 NM Nitrogen Dioxide OMPS_NPP_NMN (NO2) Total and Tropospheric O2_L2.2 Column swath orbital V2 OMPS_NPP_NMT OMPS-NPP L2 NM Ozone (O3) Total O3_L2.2 Column swath orbital V2 OMPS_NPP_LP_L2 OMPS-NPP L2 LP Ozone (O3) _O3_DAILY.2 Vertical Profile swath daily 3slit V2.5 OMPS_NPP_LP_L2 MPS-NPP L2 LP Aerosol Extinction _AER675_DAILY.1.5 Vertical Profile swath daily 3slit V1.5

Temporal resolution

Spatial resolution

Start date– end date

Daily

6 km

2012

Daily

16 km

2014

Daily

16 km

2011

Daily

50 km

2012

Daily

50 km

2012

Daily

50 km

2012

Daily

50 km

2012

Daily

50 km

2012

Daily

125 × 2 km

2012

Daily

125 × 2 km

2012

Table 4.12. List of atmospheric products from NPP satellite

4.1.3.13. OCO-2 The Orbiting Carbon Observatory (OCO-2) is NASA’s first satellite to study atmospheric carbon dioxide from space. OCO-2 was successfully launched on

126

TORUS 2 – Toward an Open Resource Using Services

2 July 2014, and has gathered more than 2 years of observations (Eldering et al. 2017). OCO-2 records how much of the sunlight reflected off the Earth is absorbed by CO2 molecules in an air column instead of directly measuring concentrations of carbon dioxide in the atmosphere. OCO-2 makes measurements in three different spectral bands over four to eight different footprints of approximately 1.29 km × 2.25 km. The OCO-2 retrievals for XCO2 are created using the full physics algorithm that has been described previously (O’Dell et al. 2012). The retrieval algorithm is based on an optimal estimation scheme and an efficient radiative transfer technique that accounts for multiple scattering and polarization effects. Table 4.13 presents a list of atmospheric products from the OCO-2 satellite. Products

Description

OCO-2 Level 2 bias-corrected solarinduced fluorescence and other OCO2_L2_Lite_SIF.8r select fields from the IMAP-DOAS algorithm aggregated as daily files, Retrospective Processing V8r OCO-2 Level 2 geolocated XCO2 retrieval results and algorithm OCO2_L2_Diagnostic.8r diagnostic information, Retrospective Processing V8r OCO-2 Level 2 bias-corrected XCO2 and other select fields from OCO2_L2_Lite_FP.8r the full-physics retrieval aggregated as daily files, Retrospective Processing V8r OCO-2 Level 2 meteorological parameters interpolated from global OCO2_L2_Met.8 assimilation model for each sounding V8 OCO-2 Level 2 meteorological parameters interpolated from global OCO2_L2_Met.8r assimilation model for each sounding, Retrospective Processing V8r OCO-2 Level 2 geolocated XCO2 OCO2_L2_Standard.8 retrievals results, physical model V8 OCO-2 Level 2 geolocated XCO2 retrievals results, physical model, OCO2_L2_Standard.8r Retrospective Processing V8r OCO-2 Level 2 spatially ordered geolocated retrievals screened using OCO2_L2_IMAPDOAS.8 the IMAP-DOAS Preprocessor (IDP) V8

Temporal resolution

Spatial resolution

Start date–end date

16 days

2.25 × 1.29 km

2014

16 days

2.25 × 1.29 km

2014

16 days

2.25 × 1.29 km

2014

16 days

2.25 × 1.29 km

2014

16 days

2.25 × 1.29 km

2014

16 days

2.25 × 1.29 km

2014

16 days

2.25 × 1.29 km

2014

16 days

2.25 × 1.29 km

2014

Table 4.13. List of atmospheric products from OCO-2 satellite

Remote Sensing Products

127

4.1.3.14. Sentinel-5P Sentinel-5 Precursor (Sentinel-5P) was successfully launched on 13 October 2017 by ESA as part of the Copernicus Programme to close the gap in continuity of observations between ENVISAT and Sentinel-5. The satellite operates in an 824 km with a Local Time of Ascending node of 13:30 hours. It carries the TROPOspheric Monitoring Instrument (TROPOMI) instrument which is a spectrometer sensing ultraviolet (UV), visible (VIS), near infrared (NIR) and short-wavelength infrared (SWIR) to monitor ozone, methane, formaldehyde, aerosol, carbon monoxide, NO2 and SO2 in the atmosphere. It extends the capabilities of the OMI from the AURA satellite and the SCIAMACHY instrument from ENVISAT. TROPOMI will be taking measurements every second covering an area of approximately 2,600 km wide and 7 km long in a resolution of 7 × 7 km. Light will be separated into different wavelengths using grating spectrometers and then measured with four different detectors for respective spectral bands. The UV spectrometer has a spectral range of 270–320 nm, the visible light spectrometer has a range of 310–500 nm, the NIR spectrometer has a range of 675–775 nm and the SWIR spectrometer has a range of 2,305–2,385 nm. Table 4.14 presents a list of atmospheric products from the Sentinel-5P satellite. Products

Description

Temporal resolution

Spatial resolution

Start date–end date

S5P_L2__NO2___.1

Sentinel-5P TROPOMI Tropospheric NO2 1-Orbit L2 7 km × 3.5 km

Daily

7 × 3.5 km

2018

S5P_L2__CO____.1

Sentinel-5P TROPOMI Carbon Monoxide CO Column 1-Orbit L2 7 km × 7 km

Daily

7 × 7 km

2018

S5P_L2__O3_TOT.1

Sentinel-5P TROPOMI Total Ozone Column 1-Orbit L2 7 km × 3.5 km

Daily

7 × 3.5 km

2018

S5P_L2__SO2___.1

Sentinel-5P TROPOMI Sulphur Dioxide SO2 1-Orbit L2 7 km × 3.5 km

Daily

7 × 3.5 km

2018

S5P_L2__O3_TCL.1

Sentinel-5P TROPOMI Tropospheric Ozone Column

Daily

0.5° × 1.0°

2018

Table 4.14. List of atmospheric products from Sentinel-5P satellite

128

TORUS 2 – Toward an Open Resource Using Services

4.2. Land observation 4.2.1. Introduction Land observation is the monitoring of planet Earth’s surface. It involves gathering information and assessing the status of, and changes in, the physical and socioeconomic characteristics of land. The two terms “land cover” and “land use” describe two different but complementary aspects of land observation and are often used interchangeably and jointly. While land use documents how people utilize the land and socioeconomic activity, “land cover is the observed (bio)physical cover on the earth’s surface” (Gregorio and Jansen 2000). There may be multiple and alternate land uses for any point or place, but there can only be one land cover specification. Popular land cover types include grass, asphalt, trees, bare ground and water. Urban and agricultural land uses are two of the most commonly known land use classes. Land cover/land use can be determined by analyzing satellite and aerial imagery. To date, many attempts have been made to monitor land cover and land use at a large scale. Figure 4.15 shows an example of a CORINE land cover map over Europe in 1990 under the Copernicus Programme.

Figure 4.15. Land cover map of Europe in 1990 (source: European Environment Agency)

Assessment and monitoring of land cover and land use change (LCLUC) are essential requirements for the sustainable management of natural resources. It provides a set of indicators to best understand the current landscape and the change

Remote Sensing Products

129

over time, which help the land owner, the government planner and others to evaluate past management decisions. They also help in gaining insight into the possible effects of their current decisions before they are implemented. Figure 4.16 shows possible societal benefits of land cover map.

Figure 4.16. GEO societal benefits and land cover/ vegetation observations (source: FAO)

4.2.2. Land cover/land use classification system Classification is an abstract representation of the situation in the field using welldefined diagnostic criteria (e.g. the classifiers) (Gregorio and Jansen 2000). A classification system is the systematic framework which involves the definition of classes, their hierarchical relationship and the criteria used to distinguish them. One of the major issues on land cover/land use research is that different works define similar categories in different ways. Example 1: Areas without trees may be classified as forest cover “if the intention is to re-plant” (the United Kingdom and Ireland), while areas with many trees may not be labeled as forest “if the trees are not growing fast enough” (Norway and Finland). Example 2: Figure 4.17 shows an example of the classification of vegetation using the diagnostic criteria of “height” and “cover” which lead to a different perspective of the same feature in comparison to the use of “leaf phenology” and “leaf type”.

130

TORUS 2 – Toward an Open Resource Using Services

Figure 4.17. Example of description of a land cover using a different underlying principle (source: FAO)

A consistent land cover/land use classification system is in need, and should be: – scale independent: the classes at all levels of the system should be applicable at any scale or level of detail; – source independent: it is independent of the means used to collect information, whether satellite imagery, aerial photography, field survey or some combination of them is used. 4.2.2.1. Land cover classification In 2004, the Food and Agriculture Organization (FAO), UNEP and the Government of Italy developed the Land Cover Classification System (LCCS) for land cover mapping. The LCCS was developed to respond to the need for the consistent and reliable assessment of land cover resources through the use of standards; definitions; classifiers; methods; approaches; semantic interoperability and preparation of interoperable, scalable and interchangeable land cover products at various levels (Gregorio and Jansen 2000). The LCCS has been designed with two main phases. In the initial dichotomous phase, eight major land cover types are defined: – cultivated and managed terrestrial areas; – natural and semi-natural terrestrial vegetation; – cultivated aquatic or regularly flooded areas;

Remote Sensing Products

131

– natural and semi-natural aquatic or regularly flooded vegetation; – artificial surfaces and associated areas; – bare areas; – artificial waterbodies, snow and ice; – natural waterbodies, snow and ice. Table 4.15 presents the hierarchical construction of FAO LCCS in the dichotomous phase. Level 1

Level 2

Level 3 Managed terrestrial area

Terrestrial Primarily vegetated

Natural and semi-natural terrestrial vegetation Cultivated aquatic areas

Aquatic or regularly flooded

Terrestrial Primarily non-vegetated Aquatic or regularly flooded

Natural and semi-natural aquatic vegetation Artificial surfaces Bare areas Artificial water bodies, snow and ice Natural water bodies, snow and ice

Table 4.15. FAO LCCS in the dichotomous phase

The dichotomous phase is followed by a subsequent so-called modularhierarchical phase, in which land cover classes are created by the combination of sets of pre-defined classifiers. These classifiers are tailored to each of the eight major land cover types. The tailoring of classifiers in the second phase allows the use of most appropriate classifiers to define land cover classes derived from the major land cover types and at the same time reduces the likelihood of impractical combinations of classifiers.

132

TORUS 2 – Toward an Open Resource Using Services

In global scale, statistics of 14 land cover classes defined by FAO can be seen in Table 4.16. FAO code

Type

1992

2001

2015

Share

[6970]

Artificial surfaces (including urban and associated areas)

26.04

34.33

55.40

0.37%

[6971]

Herbaceous crops

1,716.22

1,749.58

1,712.15

11.50%

[6972]

Woody crops

162.86

181.32

199.90

1.34%

[6973]

Multiple or layered crops

[6974]

Tree-covered areas

4,434.92

4,393.70

4,335.00

29.11%

[6975]

Mangroves

18.06

18.39

18.74

0.13%

[6976]

Shrub-covered areas

1,685.00

1,669.65

1,627.34

10.93%

[6977]

Shrubs and/or herbaceous vegetation, aquatic or regularly flooded

202.61

194.77

185.39

1.24%

[6978]

Sparsely natural vegetated areas 891.78

878.69

868.07

5.83%

[6979]

Terrestrial barren land

2,001.25

2,000.87

1,884.00

12.65%

[6980]

Permanent snow and glaciers

78.59

84.32

84.29

0.57%

[6981]

Inland water bodies

432.60

435.00

444.57

2.98%

[6982]

Coastal water bodies and intertidal areas

[6983]

Grassland

1,793.65

1,806.50

1,801.14

12.09%

14,893.91

100%

Total land mass

Table 4.16. World land cover statistic by FAO

4.2.2.2. Land use classification Currently, there is no classification system specific for land use only. Land use classification schemes typically address both land use and land cover. The most popular one has been developed by the United States Geological Survey (USGS). The land use categories are arranged in a nested hierarchy of four levels. Level I represents the most general classification including broad land use categories (Table 4.17), and is commonly used for regional and other large-scale applications. Within each class of one level, a number of more detailed classes can be defined and

Remote Sensing Products

133

mapped. The restriction is that the classes within each level must be mutually exclusive and exhaustive. In total, there are four levels of classification. Table 4.17 shows an example of a two-level land use/land cover classification system defined by USGS. Level I

Level II 11. Residential 12. Commercial and services

Urban or built-up

13. Industrial 14. Transportation, communication and utilities 15. Industrial and commercial complexes 16. Mixed urban and built-up land 21. Cropland and pasture

Agriculture

22. Orchards, groves, vineyards, nurseries and ornamental horticultural areas 23. Confined feeding operations 31. Herbaceous rangeland

Rangeland

32. Shrub and brush rangeland 33. Mixed rangeland 41. Deciduous forest land

Forest land

42. Evergreen forest land 43. Mixed forest land 51. Streams and canals

Water

52. Lakes 53. Reservoirs 54. Bays and estuaries

Wetlands

61. Forested wetlands 62. Non-forested wetlands 71. Dry salt flats 72. Beaches

Barren

73. Sandy areas other than beaches 74. Bare exposed rock 75. Strip mines, quarries and gravel pits Table 4.17. Example of a two-level land use/land cover classification system (source: JR et al. 1976)

134

TORUS 2 – Toward an Open Resource Using Services

4.2.3. Legend A legend is the application of a classification in a specific area using a defined mapping scale and specific dataset (Figure 4.18). Therefore, a legend may contain only a proportion, or sub-set, of all possible classes of the classification. Thus, a legend is: – scale and cartographic representation-dependent: with an occurrence of mixed mapping units if the elements composing this unit are too small to be delineated independently; – data and mapping methodology-dependent.

Figure 4.18. Legend as an application of a classification in a particular area (source: FAO)

4.2.4. Data 4.2.4.1. Satellites and sensors Satellite images are images of Earth or other planets collected by observation satellites. The satellites are often operated by governmental agencies or businesses around the world. There are currently many Earth observation satellites, and they have common characteristics including: spatial resolution, spectral resolution,

Remote Sensing Products

135

radiometric resolution, and temporal resolution. A detailed description of each resolution is given below: – Spatial resolution: refers to the instantaneous field of view (IFOV) which is the area on the ground viewed by the satellite’s sensor. For example, the Landsat 8 satellite has a 30 m spatial resolution, which means that a Landsat 8’s pixel covers an area on the Earth’s surface of 30 m × 30 m. – Spectral resolution: spectral resolution describes the ability of the sensor to receive the Sun’s light. If conventional cameras on the phone can only obtain wavelengths in the visible range, including red, green and blue lights, many satellite sensors are able to sense many other wavelengths such as near-infrared, short-wave infrared and so on. For example, the TIRS sensor mounted on a Landsat 8 satellite can receive wavelengths ranging from 10.6 to 12.51 µm. – Radiometric resolution: the radiometric resolution of a sensor describes the ability to distinguish very small differences in light energy. A better radiometric resolution can detect small differences in reflection or energy output. – Temporal resolution: temporal resolution of a satellite is the time interval between two successive observations over the same area on the Earth’s surface. For example, the temporal resolution of a Landsat 8 satellite is 16 days. There are currently many Earth observation satellites having different spatial resolutions, temporal resolutions, radiometric resolutions and spectral resolutions. Table 4.18 compares some satellite images widely used for land observation. Spectral resolution (excluding panchromatic)

Radiometric resolution

Temporal resolution

250– 1,000 m

36 bands

12 bits

Daily

Optical

10 m

4 bands (green, red, near-IR, SWIR)

8 bits

2–3 days, depending on latitude

Landsat 8

Optical

30 m

10 bands (Coastal -> TIRS2)

12 bits

16 days

Sentinel 2A

Optical

10–20 m

12 bands (Coastal -> SWIR)

12 bits

10 days

Satellite image

Type

1

MODIS

Optical

2

SPOT 5

3

4

Typical spatial resolution

Table 4.18. Some featured satellite images

136

TORUS 2 – Toward an Open Resource Using Services

4.2.4.2. Field survey Land cover data can be obtained by field-surveying, which gathers data through direct observations made by surveyors in the field. During a field survey, surveyors visit the points and observe the land cover, land use, and environmental parameters they find on the ground. The surveyor then documents the land cover and land use according to a harmonized classification system. The strict requirement is that each of the surveyors applies the same methods when visiting the assigned geographical point. In addition, we can use a range of methods to collect information about environmental quality in a qualitative way. 4.2.4.3. Questionnaires A questionnaire is a research instrument consisting of a series of questions (or other types of prompts) for the purpose of gathering information from respondents. A questionnaire consists of a number of questions that the respondent has to answer in a set format. A distinction is made between open-ended and closed-ended questions. An open-ended question asks the respondent to formulate his or her own answer, whereas a closed-ended question has the respondent pick an answer from a given number of options. The response options for a closed-ended question should be exhaustive and mutually exclusive. For a land cover/land use field survey, the surveyors fill in a questionnaire with a series of land cover and land use parameters. An example of a questionnaire can be seen in Figure 4.19.

Figure 4.19. Land survey questionnaire

Remote Sensing Products

137

4.2.4.4. Annotated photograph The surveyor takes a series of photographs at each point, recording all of the different land cover and linear elements. They could photograph whole street scenes or focus on small details such as a pile of rubbish. An annotated photograph can be used to make a judgment about the quality of the environment, representing ground evidence for the calibration of satellite images and a register of points for specific surveys.

Figure 4.20. Photograph showing a field survey at the Obama geothermal field using a GPS antenna and Scintrex CG-3M gravimeter (source: Saibi, Nishijima, and Ehara)

4.2.5. Methodology 4.2.5.1. Satellite image mosaicing Image mosaicing refers to the process of aligning multiple images into larger compositions, which represent portions of a scene. This is very useful if our area of interest sits between two images along a satellite path or multiple data tiles. As shown in Figure 4.21, mosaicing involves various steps of image processing: – registration: establishing the geometric correspondences between a pair of images. This is done by estimating the geometric transformations which align the images to a reference image;

138

TORUS 2 – Toward an Open Resource Using Services

– reprojection: aligning images based on their geometric transformation into a common coordinate system; – stitching: painting images on a larger canvas by merging pixel values of the overlapping portions and retaining pixels where no overlap occurs; – blending: minimizing the discontinuities in the global appearance of the mosaic.

Figure 4.21. Image mosaicing (source: Ghosh and Kaabouch 2016)

Remote Sensing Products

139

4.2.5.2. Satellite image compositing Generally, compositing refers to the process of combining spatially overlapping images into a single image, based on an aggregation function. In cloud-prone areas, deriving high-resolution LCLUC maps from optical satellite imagery is challenging because of infrequent satellite revisits and lack of cloud-free data (Man et al. 2018). Thus, removing cloud effects from satellite images is a critical task when using the data to monitor LCLUC. The compositing techniques can produce cloud-free images from individual partly cloudy input scenes by selecting the cloud-free pixel for the same geographic location from images taken over a limited period of time (see Figure 4.22).

Figure 4.22. (a) Original surface reflectance images, (b) composite images (source: Man et al. 2018). For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

Cloud-free pixel selection is typically performed by selecting the best available observation for a given pixel (BAP) within a specified compositing period (e.g. one year). The process can be based on simple rules such as maximum value compositing (MVC). One popular criterion is using the maximum normalized difference vegetation index (NDVI). The algorithm is based on the theory that, for a given pixel over land, a higher NDVI usually indicates a lower cloud fraction. On a per-pixel basis, the pixel with the maximum NDVI in the compositing period is chosen for compositing. Another strategy is to use cloud masks. Cloud masks are first generated by cloud detection schemes, and then clear pixels are picked out according to the cloud masks. Usually, satellite imagery is provided with a cloud mask as an auxiliary product. Figure 4.22 shows an example of a MODIS cloud mask data.

140

TORUS 2 – Toward an Open Resource Using Services

Figure 4.23. (Left) Terra MODIS Band 31 data, (right) MODIS Cloud Mask product

4.2.5.3. Satellite image classification 4.2.5.3.1 Supervised classification 4.2.5.3.1.1 Sampling methods Most land cover mapping methods require the collection of ground reference data at the time when the remotely sensed data are acquired. Due to the high cost of collecting reference data, a crucial task is to obtain a subset with a sufficient number of training samples and high quality using the sampling method. Sampling is the process of selecting a representative group of data from the population under study. We can think of all pixels in the satellite image as a population. Each pixel belongs to different groups of specific land cover/land use type. It is impractical to select every pixel for training and/or assessing the

Remote Sensing Products

141

classification model as the number of pixels is very large. Therefore, we select samples or sub-groups of the population that is likely to be representative of the target population we are interested in. Random sampling and stratified sampling are two of the most popular yet simple sampling methods. 4.2.5.3.1.2. Random sampling Every sample in the entire dataset has an equal chance of being selected. The sample size in this sampling method should ideally be large enough so that it satisfies the law of the large number to guarantee the stable results. Example: we have a dataset with 450,000 samples of three land cover types. The number of samples for each type is as follows: Land cover

Number of samples

Paddy rice

100,000

Forest

200,000

Water

150,000

We want to sample a subset of 100,000 samples from the dataset. Then, we sample the dataset three times using the random sampling method. The number of samples for each land cover is as follows: Land cover

Random sampling #1

Random sampling #2

Random sampling #3

Paddy rice

15,000

50,000

16,000

Forest

26,000

24,000

56,000

Water

59,000

26,000

28,000

4.2.5.3.1.3. Stratified sampling Different groups of data (e.g. land cover types) that make up the entire dataset are identified. Then, proportions needed for the sample to be representative are selected. Example: we sample the dataset in the above example using stratified sampling.

142

TORUS 2 – Toward an Open Resource Using Services

Land cover

Stratified sampling

Paddy rice Forest Water

100,000 × 100,000 = 22,222 100,000 + 200,000 + 150,000 200,000 × 100,000 = 44,444 100,000 + 200,000 + 150,000 150,000 × 100,000 = 33,333 100,000 + 200,000 + 150,000

Stratified sampling can be done using different criteria (e.g. land cover type or geographic area). The advantage is that the sample should be highly representative of the target dataset, and therefore we can generalize from the results obtained. 4.2.5.3.1.4. Feature extraction 4.2.5.3.1.4.1. Spectral indices A wide range of spectral characteristics of land cover/land use type can be inferred through various spectral indices. Spectral indices enhance the spectral information and increase the separability of the classes of interest by combining the value of multiple bands into one. All of these factors result in an increase in the quality of the land use land cover (LULC) mapping produced (Ustuner et al. 2014). The main spectral bands that are used to calculate spectral indices are: – visible: - blue: 450–495 nm, - green: 495–570 nm, - red: 620–750 nm; – infrared: - near infrared (NIR): 750–900 nm, - short-wave infrared (SWIR): 900–3,000 nm; - thermal infrared (TIR): 3,000–14,000 nm Note that the wavelength domains vary from satellite to satellite as they incorporate different sensor types. Seven types of spectral indices along with its application are presented in Table 4.20

Remote Sensing Products

Indices Normalized difference vegetation index (NDVI)

Enhanced vegetation index (EVI)

Normalized difference water index (NDWI)

Equation =

=

2.5 × ( + 1×

1= 2=

− +

) − − 2×

− +

− +

+

143

Characteristics NDVI values range between 0 and 1 (due to the normalization procedure). Very low values of NDVI (< 0.1) correspond to barren areas of rock, sand or snow. Free-standing water tends to be in the very low positive to negative values. Soil tends to generate rather small NDVI values (0.1– 0.2). Sparse vegetation such as shrubs and grasslands may result in moderate NDVI values (0.2– 0.5). High NDVI values correspond to dense vegetation. The enhanced vegetation index (EVI) is the most common alternative vegetation index which reduces the influence of atmospheric conditions and corrects canopy background signals. NDWI1 is used to monitor changes in the water content of leaves. NDWI2 is used to monitor changes related to water content in water bodies.

144

TORUS 2 – Toward an Open Resource Using Services

Modified soil2 adjusted 2× = vegetation index (MSAVI2)

Soil-adjusted vegetation index (SAVI)

Soil-adjusted vegetation index optimized for agricultural monitoring (OSAVI)

+ 1) − 8 × ( 2

+ 1 − (2 ×

−

=

+

=

+

× (1 + )

− +

+ 0.16

−

MSAVI2 is mainly ) used in the analysis of plant growth, desertification research, grassland yield estimation, LAI assessment, analysis of soil organic matter, drought monitoring and the analysis of soil erosion (Xue et al. 2017). SAVI is a modification of the NDVI with a correction factor for soil brightness. The value of L is adjusted based on the amount of vegetation. L = 0.5 is the default value and works well in most situations. With L = 0, NDVI = SAVI. OSAVI is the soiladjusted vegetation index optimized for agricultural monitoring. OSAVI is more sensitive to vegetation and shows differences in vegetation better than SAVI.

Table 4.19. Spectral indices for land cover classification

4.2.5.3.1.4.2. Spatial-texture features Together with spectral indices, various spatial features can be used to extract information about the relationship of target pixel and its neighbors. Among those, standard deviation and the gray-level co-occurrence matrix (GLCM) are the most popular ones.

Remote Sensing Products

145

Standard deviation is a statistic that measures the dispersion of a dataset relative to its mean and is calculated as the square root of the variance. It is calculated as the square root of variance by determining the variation between each data point relative to the mean. If the data points are further from the mean, there is a higher deviation within the dataset; thus, the more spread out the data, the higher the standard deviation: where points.

∑

=

(

̅)

[4.1]

point, ̅ is the mean value and

is the value of

is the number of data

GLCM is a statistical method of examining texture that considers the spatial relationship of pixels is the gray-level co-occurrence matrix. The GLCM functions characterize the texture of an image by calculating how often pairs of pixels with specific values and in a specified spatial relationship occur in an image, creating a GLCM and then extracting statistical measures from this matrix. Figure 4.24 is an example of constructing a GLCM matrix. Considering the raster image , (left), we count the number of times that ( , ) = and ( ± 1, ± 1) = as , in the GLCM matrix (right).

Figure 4.24. GLCM matrix construction

We can then use several characteristics of

as features:

– maximum probability entry; – element difference moment of order : ∑ ∑ ( − ) – contrast: ∑ ∑ ( − ) – entropy: − ∑ ∑ – energy: ∑ ∑

,

;

log

;

;

– homogeneity: ∑ ∑

|

|

.

,

;

146

TORUS 2 – Toward an Open Resource Using Services

4.2.5.3.1.5. Classifier Classifiers are responses to assign each observation data a label based on the extracted features. The maximum likelihood classifier (MLC), the decision tree (DT) and the artificial neural network are the three most popular supervised classifiers used by the remote sensing community for land cover/land use mapping. 4.2.5.3.1.5.1. Maximum likelihood classifier The maximum likelihood classifier (MLC) is one of the most popular methods of classification in remote sensing, in which a pixel with the maximum likelihood is is defined as the posterior classified into the corresponding class. The likelihood probability of a pixel belonging to class : = ( | )=

( )× ( | )

[4.2]

( )

where ( ) is the prior probability of class , ( | ) is the conditional probability to observe X from class k, or the probability density function, and ( ) is the probability that X is observed: ( )=∑

Pixel

( )× ( | )

[4.3]

( ) is often treated as a normalization constant to ensure ∑ is classified as class by the rule:

( | ) sums to 1.

∈ if ( | ) > ( | ) for all ≠

[4.4]

MLC often assumes that the distribution of the data obeys a multivariate Gaussian distribution. We can then define the log likelihood (or discriminant function): ( ) = ln ( | ) = − ( −

)

( −

) − ln 2 − ln

[4.5]

Each pixel is then assigned to class by the rule: ∈ if

( )>

( ) for all ≠

[4.6]

Remote Sensing Products

147

4.2.5.3.1.5.2. Decision tree A decision tree (DT) has been used widely for land cover/land use classification by the remote sensing community. The tree is composed of a root node (formed from all data), a set of internal nodes (splits) and a set of terminal nodes (leaves). A dataset is classified by sequentially subdividing it into subsets according to the decision defined in each internal node. The decision function in each internal node is normally a simple rule (e.g. larger than A, less than B). The label is assigned to each observation according to the leaf node. Figure 4.25 shows a toy example of land cover mapping using a decision tree on spectral indices.

Figure 4.25. Toy example of a decision tree classifier for land cover mapping.

Constructing a decision tree is greedy. Figure 4.26 shows a pseudo-code for constructing the decision tree using the ID3 algorithm. Split (node, {observations} ): 1. ← the best attribute for splitting {observations}. 2. For each value of A, split {observations} to new child node. 3. For each subset in child node, if subset is pure: Stop else: Split(child node, {subset})

Figure 4.26. Pseudo-code of ID3 algorithm for constructing decision tree

148

TORUS 2 – Toward an Open Resource Using Services

The key component for constructing a decision tree is to find the best attribute at each step. The ID3 algorithm defines the best attribute using entropy (or information gain). Entropy is a fundamental theorem which is commonly used in information theory to measure important information relative to its size. Let be our set of observations containing samples of label; the entropy of relative to this classification is: ( )=∑

−

log

[4.7]

where is the number of samples for each label. We derive the original entropy of the population to measure the information gain of each attribute. For training set and its attribute , the formula of information gain is: ( , )=

( )−∑∈

|∆ | ( ) |∆ |

( )

[4.8]

We can see that the weather conditions may affect the decision to play sport. Given the column “Play Tennis” as our target class label which defines the decision to play tennis or not, and the other column shows weather conditions: outlook, temperature, humidity and wind. We want to know what the best day to play tennis will be.

Out of 14 samples, there are five “No” and nine “Yes” answers. The probability = 5/14 = 0.357, and the probability of “Yes” is = 9/14 = of “No” is 0.643. The entropy of the dataset is: ( ) = −0.357 × log 0.357 − 0.643 × log 0.643 = 0.940

Remote Sensing Products

149

For simplicity, we will give the name of each attribute in dataset: – O is the outlook attribute; – T is the temperature attribute; – H is the humidity attribute; – W is the wind attribute. At first iteration, we select the best attribute which has maximum information gain as the root of our decision tree: = −( ) log (

−

) = −( ) log

(

) = −( ) log

( , )

=

0.940 −

( )− × 0.971 +

log −

−

= 0.971

log

=0

log

= 0.971

+

(

×0+

)+

(

) =

× 0.971 = 0.246

Applying the similar calculations, we have information gain value for other attributes as follows: = 0.048

–

( ,

–

( , )

= 0.029

–

( , )

= 0.151

)

As seen, the outlook factor on decision produces the highest information gain. Thus, the outlook feature will be used as a root node. The decision tree technique has several characteristics which are advantages for remote sensing classification problems, including: – flexibility: the tree is able to handle both numerical and categorical data; – intuitive simplicity: the tree is simple to understand and visualize; – computational efficiency: the cost of using the tree is logarithmic in the number of data points used to train the tree;

150

TORUS 2 – Toward an Open Resource Using Services

– non-parametric: the tree requires no pre-defined parameter; – interpretable: the explanation for the classification of an observation is easily explained by the Boolean logic. 4.2.5.3.1.5.3. Artificial neural network Artificial neural networks (ANNs) are a group of statistical learning models that are inspired by biological neural networks in the human brain (Foody and Mathur 2004). The most widely used model is multilayer perceptron (MLP), a feed-forward neural network, due to how easy it is to understand and interpret. The backpropagation learning algorithm, introduced by Rumelhart et al., is the most popular algorithm used to train an MLP (Rumelhart, Hinton and Williams 1986). Figure 4.27 presents a three-layer perceptron with three inputs, two outputs and one hidden layer assembled by five neurons:

Figure 4.27. An example of MLP

Each neuron has several input links. The inputs are output values from neurons in the previous layer. At a particular neuron, the inputs are summed up with certain weights, plus a bias term. The sum is then transformed using an activation function , which may be different for different neurons. In other words, given the inputs x of the layer, the output y of the layer n + 1 are computed as: =∑

,

×

+

,

= ( )

[4.9]

Remote Sensing Products

151

There are different activation functions. Some standard functions are: Identify function: ( )=

[4.10]

Rectified linear unit (ReLU) function: 0, ( )= ,

{ val features = value.features.toArray val data = tifList.map(tif => { val (col, row) = tif.rasterExtent.mapToGrid(value.features(0), → value.features(1)) → tif.tile.get(col, row).toDouble }).toArray LabeledPoint(value.label, Vectors.dense(features ++ data)) }) }

Code 5.5. Create LabeledPoint collection

5.2.4. Advanced method The second processing method we have implemented aims to eliminate the preprocessing applied to TIF files which is expensive and not very optimized. For this reason, we have used the RasterFrames library developed by LocationTech, the same working group as GeoTrellis. 5.2.4.1. RasterFrames RasterFrames is a tool that unites all the machine learning tools of Spark MLlib using DataFrames with cartographic algebra and GeoTrellis tile operations. This tool responds exactly to the needs set out in the problem. On the other hand, its development is recent and remains subject to possible bugs and lack of documentation. RasterFrames is based on a type for all these operations called RasterFrame. This object is just a DataFrame with some invariants to store the different tiles and metadata. More precisely, a RasterFrame is a Tagged type of a DataFrame. In addition, RasterFrames provides various visualization tools to export results, as well as evaluators and processors to facilitate the implementation of Pipelines MLlib.

Image Processing in Spark

169

5.2.4.2. Implementation We have therefore implemented a solution based on RasterFrames that allows us to optimize pre-treatment. This method uses the same mechanisms as the previous methods by exploiting all the capabilities of the Spark cluster as well as GeoTrellis. We then apply the same processing pipeline as before on the DataFrame that we obtain by using RasterFrames. In order to create this DataFrame to drive our data, we perform three steps. The first one is to load the GeoTIF files to create a list of RasterFrame files, each containing a column identified by an ID, contained in bandColNames. To gather these different rasters into a single RasterFrame, the spatial tool Join is used, which allows us to merge the different bands according to a certain method (inner, outer, etc.) and to obtain the RasterFrame joinedRF, in code 5.6. def readTiff(name: String): SinglebandGeoTiff = SinglebandGeoTiff(s"./Path/$name") → val tileSize = 10 // Tile size for Geotrellis val bandCol = Array(("AGE_UTM1.tif","age"), ("wvnii_2013.tif","wvnii_2013"), ("irox_2013.tif","irox_2013"),...))/ val bandColNames = bandCol.unzip._2 val bandNumbers = bandColNames.indices val joinedRF = bandNumbers. map { b => (b, readTiff(bandCol(b)._1)) }. map { case (b, t) => t.projectedRaster.toRF(tileSize, tileSize, bandCol(b)._2) }. → reduce(_ spatialJoin _)

Code 5.6. Merge GeoTIF files

Then the same operation is performed on the GeoTIF containing the labels, specifying that it is the target column and contains NoData, in code 5.7. Finally, we can create the final DataFrame containing all the data identified by columns, in code 5.8.

170

TORUS 2 – Toward an Open Resource Using Services

val targetCol = "target" val target = readTiff("stock_nodata.tiff"). mapTile(_.convert(DoubleConstantNoDataCellType)). projectedRaster. toRF(tileSize, tileSize, targetCol)

Code 5.7. Load GeoTIF Label

val raster = joinedRF.spatialJoin(target)

Code 5.8. Create final DataFrame

Once the DataFrame is created, the processing pipeline must be set up to operate the data. For this purpose, it is necessary to define a pipeline with three transformers in order to make data compatible with the models, as previously attempted with the LabeledPoints. The first transformer allows us to retrieve each pixel for each tile, in code 5.9 import astraea.spark.rasterframes.ml.TileExploder val exploder = new TileExploder()

Code 5.9. Define a pipeline

A filter is then applied to remove missing data from all columns of our data set, in code 5.10. Once the data have been isolated and filtered, they must be transformed into a readable form by an estimator, i.e. a label column and a features column, in code 5.11. Finally, we can define a regression model, here a linear regression model, and assemble the pipeline and then train the model, in code 5.12. Once this model has been created, it is possible to test it, to make predictions on new TIFs or to save it. It should be noted that the process carried out in the pipeline to pre-process the data is closer to what we were trying to do with our naive approach. However, as we have seen, performance is much better due to an optimized implementation to work with a Spark and GeoTrellis cluster.

Image Processing in Spark

171

import astraea.spark.rasterframes.ml.NoDataFilter val noDataFilter = new NoDataFilter().setInputCols(bandColNames :+ targetCol) →

Code 5.10. Filtering data

import org.apache.spark.ml.feature.VectorAssembler val assembler = new VectorAssembler(). setInputCols(bandColNames). setOutputCol("features")

Code 5.11. Transformation before estimation

val regression = new LinearRegression(). setLabelCol(targetCol). setFeaturesCol(assembler.getOutputCol) val pipeline = new Pipeline(). setStages(Array(exploder, noDataFilter, assembler, regression)) → val model = pipeline.fit(raster)

Code 5.12. Add regression to the pipeline

5.3. Conclusion The purpose of this chapter was to explore the possibilities of removing the lock that consisted of the processing of very large images in order to potentially be able to apply machine learning algorithms. The main idea was to do this work on a Spark cluster to benefit from the power of parallelization that it allows us to consider. The way in which data are stored (DataFrame) has greatly facilitated the application of prediction on the data. The initial hypothesis has been verified using the recently developed RasterFrame solution.

6 Satellite Image Processing using Spark on the HUPI Platform

6.1. Introduction Raster processing has always been a very slow task; it is complicated, timeconsuming work. In fact, to be able to use these data, we need many intermediate steps in between. It is particularly problematic when working with Big Data. The question we wish to address in this chapter is as follows: “Is it possible to perform satellite image processing using Spark on the HUPI platform?” To do this, we will cover six key areas: – an introduction to GeoTrellis; – using GeoTrellis in Hupi-Notebook; – workflows in HDFS; – visualizations in Hupi-Front; – creating a cloud service; – development.

Chapter written by Vincent MORENO and Minh Tu NGUYEN.

TORUS 2 – Toward an Open Resource Using Services: Cloud Computing for Environmental Data, First Edition. Edited by Dominique Laffly. © ISTE Ltd 2020. Published by ISTE Ltd and John Wiley & Sons, Inc.

174

TORUS 2 – Toward an Open Resource Using Services

6.2. An introduction to GeoTrellis GeoTrellis is a Scala library and framework, using Spark to work with raster data. With GeoTrellis, we can read, write and operate raster data easily; we can also transform vector to raster or raster to vector, without any intermediate work. Furthermore, by using GeoTrellis, we can render raster data to formats such as PNG or JPG. We can also save and read data from HDFS as GeoTiff or PNG. More importantly, GeoTrellis can process at web-speed while working with raster data; therefore, it can be an important tool to help us work with Big Data in satellite image processing. More information about GeoTrellis can be found at the following links: – https://geotrellis.io/ – https://docs.geotrellis.io/en/latest/

6.3. Using GeoTrellis in Hupi-Notebook In this section, we will demonstrate how to use GeoTrellis in Hupi-Notebook. All notebooks mentioned in this chapter can be found in Hupi’s Github: https://github.com/hupi-analytics/formation/tree/master/Training-Torus/Notebooks To access the Hupi-Notebooks, the following direct link can be used: https://factory02-01-notebook.thai.cloud-torus.com/tree/THAILAND%20 workshop

Within Hupi-Notebook, there are five workshops: – some core concepts of GeoTrellis; – computation of NDVI; – compare two NDVI images;

Satellite Image Processing using Spark on the HUPI Platform

175

– descriptive statistics of NDVI per Tile; – K-means. To obtain input data, we can download GeoTiff images from USGS. Here, we need to choose the location and duration:

Then, choose type of Landsat and click “Results”.

176

TORUS 2 – Toward an Open Resource Using Services

Click the download icon to download and choose “GeoTiff images” as the format. We can also use images from Sentinel, however we need to convert JP2 format to GeoTiff in order to use GeoTrellis. The following steps show how to use images from Sentinel in Hupi-Notebook: – download data from Copernicus (JP2 images); – download SNAP (http://step.esa.int/main/download/); – in SNAP, click “File” > “Open product” > [search_jp2_file_in_repo_ Granule_then_L1C] > “Open”;

Satellite Image Processing using Spark on the HUPI Platform

177

– [select_image_in_product_explorer] > Raster > Data Conversion > Convert DataType > Save As GeoTiff; – save the GeoTiff in Hupi-Studio HDFS and read in Hupi-Notebook. 6.3.1. Some core concepts of GeoTrellis Notebook 1 (see link in section 6.3) Documentation: https://geotrellis.readthedocs.io/en/latest/guide/core-concepts.html Some principal concepts are: – Tile represents grid of “Cells”; – extent is like “Bounding Box”; – CRS is coordinate reference system; – GeoTiff is a combination of Time and ProjectedExtent (CRS + Extent); – in GeoTrellis, Raster means Tile + Extent. 6.3.2. Computation of NDVI Notebook 2 (see link in section 6.3) NDVI formula:

Here, inputs are two bands (red band and NIR band) from LC08_L1TP_125052_20171231_20180103_01_T1 (image taken from Landsat 8 on December 31, 2017). Outputs expected are: PNG and GeoTiff of NDVI.

178

TORUS 2 – Toward an Open Resource Using Services

Results of notebook in PNG:

6.3.3. Compare two NDVI Notebook 3 (see link in section 6.3) Here, inputs are GeoTiffs of NDVI of LC08_L1TP_125052_20171231_ 20180103_01_T1 and LT05_L1GS_125052_20070915_20161112_01_T2 We would like to compare NDVI in the same place between two years. Below is the PNG of this comparison. The reason for the non-superposition is because of the quality of the image coming from the satellites (another problem).

6.3.4. Descriptive statistics of NDVI per Tile Notebook 4 (see link in section 6.3) Here, we compute some descriptive statistics (min, max, mean, standard deviation) of NDVI for each Tile.

Satellite Image Processing using Spark on the HUPI Platform

179

Inputs are GeoTiff of NDVI of LC08_L1TP_125052_20171231_20180103_ 01_T1

Outputs expected are dataframe of all descriptive values of NDVI by Tile (this dataframe will be saved in MongoDB) and a PNG of meanNDVIByTile (see photo above). 6.3.5. K-means We try to create a K-means model by using raster data. To do this, we created four notebooks: – create MultiBandTile; – create a model with k clusters; – choose optimal k with the Elbow method; – visualize final K-means model. 6.3.5.1. Create MultiBandTile Notebook 5 (see link in section 6.3) Here, inputs are all SingleBand GeoTiff from HDFS of LC08_L1TP_ 125052_20171231_20180103_01_T1 (except the eighth band) The output expected is one MultiBand GeoTiff in HDFS.

180

TORUS 2 – Toward an Open Resource Using Services

6.3.5.2. Create a model with k clusters Notebook 6 (see link in section 6.3) Inputs is MultiBand GeoTiff from HDFS of LC08_L1TP_125052_20171231_ 20180103_01_T1 (except the eighth band). The outputs expected are K-means and PMML models (to save in HDFS). 6.3.5.2.1. Choose optimal k with the Elbow method Notebook 7 (see link in section 6.3) By using the Elbow method, we can determine optimal k for the K-means model. In our case, with k = 3–20, we have:

Figure 6.1. Graphic of Within Set Sum of Squared Error of k = 3 to 20

Here, we can see that k is optimal at six or nine clusters. If we extend to 30 clusters, we have:

K=9

Figure 6.2. Graphic of Within Set Sum of Squared Error of k = 3 to 30

Satellite Image Processing using Spark on the HUPI Platform

181

Here, we can see more clearly that optimal k is nine clusters. 6.3.5.2.2. Visualize final K-means model Notebook 8 (see link in section 6.3)

Figure 6.3. PNG of K-means model with nine clusters

6.4. Workflows in HDFS: automatize image processing Here are some steps to create a Workflow in Hupi-Studio. 6.4.1. Create a jar In Windows, we will need an editor text, sbt, Java, FileZilla, Putty. In general, we will have: Editor Text  Windows PowerShell  run”sbt assembly”  jar is created in /target/scala-2.11 (If your scala version is 2.11) To test the jar: FileZilla  Putty  run spark2-submit [spark_options] [path_jar] [arguments of jar]

182

TORUS 2 – Toward an Open Resource Using Services

Example: spark2-submit --master yarn --deploy-mode client --driver-memory 12g --numexecutors 10 --executor-cores 3 --executor-memory 12g --conf spark.kryoserializer. buffer.max=256m --conf spark.dynamicAllocation.enabled=false /home/minhtu.nguyen/kmeans-assembly-1.0.jar “LC08_L1TP_125052_20171231_20180103_ 01_T1” 100 “3,4”1 6.4.2. Monitor the Spark jobs We can monitor Spark jobs by using Hupi-Studio or Yarn – Yarn: https://factory02-yarn.thai.cloud-torus.com/cluster/apps

– Hupi-studio: http://factory02-studio.thai.cloud-torus.com/oozie/list_oozie_workflows/

In fact, when we run: spark2-submit --master yarn --deploy-mode client --driver-memory 12g --numexecutors 5 --executor-cores 3 --executor-memory 12g --conf spark.kryoserializer. buffer.max=256m --conf spark.dynamicAllocation.enabled=false /home/minhtu.nguyen/kmeans-assembly-1.0.jar “LC08_L1TP_125052_20171231_20180103_ 01_T1” 100 “3,4” 1 To learn more about Spark options see: https://spark.apache.org/docs/latest/running-onyarn.html.

Satellite Image Processing using Spark on the HUPI Platform

183

6.4.3. Tune performance of the Spark job Reference link: http://blog.cloudera.com/blog/2015/03/how-to-tune-your-apache-spark-jobspart-2/ Some basic tips: – do not forget to put options: --conf spark.dynamicAllocation.enabled=false; – to run in Yarn, always put --master yarn; – for the deploy mode, when we test in our local → --deploy-mode client. For the Workflows, to be able to find the jar in HDFS --> --deploy-mode cluster; – for the number of executors, first try with --num-executors 1 or 2 maximum and --executor-cores 1 to 3 cores, for example; – for executor-memory, try with some even numbers (like 4 or 6g) here, it depends on the jobs, but we can modify later if we need more resources; – if it throws an error, for example: “org.apache.spark.SparkException: Kryo serialization failed: Buffer overflow. Available: 0, required: 1 Serialization trace: values (org.apache.spark.sql.catalyst.expressions.GenericRow) otherElements (org.apache.spark.util.collection.CompactBuffer). To avoid this, increase spark.kryoserializer.buffer.max value”. → search for the options, read about it and increase its max value using: --conf spark.kryoserializer.buffer.max=256m;

184

TORUS 2 – Toward an Open Resource Using Services

– if it works successfully, but very slowly → observe in Yarn, see if there are GC Time (Garbage Collection) for executors, if yes → first consider decreasing the number of cores (if > 1) and try again, if it is not better → increase number of executors. Be moderate with your choice of options to minimize waste of resources. 6.4.4. Create a workflow in Hupi-Studio To create a workflow, we need to go to Hupi-Studio: – by URL : https://factory02-studio.thai.cloud-torus.com/ or; – by choosing Studio in the HUPI platform.

In Hupi-Studio, we can click “Workflows” > “Editors” > “Workflows”.

Then choose “Create”

Satellite Image Processing using Spark on the HUPI Platform

185

After clicking “Create”, we can see more options of Workflows. Here, to create a Spark job, we need to choose “ssh” job and “email” job which can notify us if the job failed.

After selecting the ssh action: – user and host → factory02@hupi-factory-02-01-01-01; – ssh command → spark2-submit; – arguments: one option/argument per line (we can keep the command when we test in local, but we need to change): – deploy-mode cluster; – path of jars (save jar to HDFS and put the good path in the workflow). For example in local: spark2-submit --master yarn --deploy-mode client --driver-memory 12g --numexecutors 10 --executor-cores 3 --executor-memory 12g --conf spark.kryoserializer. buffer.max=256m --conf spark.dynamicAllocation.enabled=false /home/minhtu.nguyen/kmeans-assembly-1.0.jar “LC08_L1TP_125052_20171231_20180103_ 01_T1” 100 “3,4” In workflow: spark2-submit --master yarn --deploy-mode cluster --driver-memory 12g --numexecutors 10 --executor-cores 3 --executor-memory 12g --conf spark.kryoserializer. buffer.max=256m --conf spark.dynamicAllocation.enabled=false hdfs://hupi-factory02-01-01-01/user/factory02/thailand_workshop/jars/kmeans-assembly-1.0.jar “LC08_ L1TP_125052_20171231_20180103_01_T1” 100 “3,4” After putting all arguments in ssh action, we can save the workflow!

186

TORUS 2 – Toward an Open Resource Using Services

6.5. Visualizations in Hupi-Front Reference link http://factory01.thai.cloud-torus.com/themes/thailand_workshop/ To go to Hupi-Link, we can click on Interactif or Link in the HUPI platform.

The image below shows how we can create a new endpoint, see, modify an existing endpoint or import a new endpoint.

Some examples of widgets: – in columns;

Satellite Image Processing using Spark on the HUPI Platform

187

– in area.

To create an endpoint prediction, we can go to “Predict”, instead of “Endpoint”.

Example of an endpoint prediction:

188

TORUS 2 – Toward an Open Resource Using Services

6.6. Cloud service By using the HUPI platform, we can create a cloud service – an API that computes descriptive statistics of NDVI per Tile by entering landsatName. We can test this service with Postman (to download Postman, see https://www.getpostman.com/).

In Hupi-Front, in Accounts, API-Token can be found in Link ID & Tokens:

Result in Postman when we call API “descriptiveStats_NDVI_by_Tile” and put “LC08-L1TP_125052_20171231_20180103_01_T1” as the: filter

Satellite Image Processing using Spark on the HUPI Platform

189

The full cloud service process in the Hupi platform

6.7. Development To create a complete cloud service, we need to create a module to retrieve USGS data. To be able to do that, the following must be considered:

190

TORUS 2 – Toward an Open Resource Using Services

– your organization’s primary purpose with regard to needing USGS data; how will you use the data you receive?; – EarthExplorer login username (new users can register at https://earthexplorer. usgs.gov/register/); – dataset(s) required: provide the EarthExplorer dataset name(s) – these can be found on EarthExplorer (Landsat on-demand datasets are not eligible for this access); – data product(s) required; – approximate number of scenes needed; – scripting capabilities: verify that PHP, PERL or another scripting language is known to access the data; – detailed justification for needing this access; – complete name, address and contact information for the entity requesting access and so on. In conclusion, with GeoTrellis and Spark, we are able to operate with raster data and create a complete cloud service in the HUPI platform.

7 Remote Sensing Case Studies

7.1. Satellite AOD validation using R 7.1.1. Introduction Aerosols are the solid and liquid particles suspended in the atmosphere (Bo 2016, Hongqing et al. 2014, Taylor et al. 2012). These particles can be either natural (desert and soil dust, sea salt particles, volcanic emissions, wildfire smoke, biogenic emissions) or anthropogenic (industrial emissions, biomass burning, traffic). Aerosols have been linked to increases in morbidity and mortality rates and to degradations of environmental quality in terms of acid rain and the reduction of visibility. The way to estimate the presence of aerosols in the atmosphere is the aerosol optical thickness (AOT) or aerosol optical depth (AOD). It is a unitless measure, and is defined as the integral of the atmospheric extinction coefficient from the surface to the top of atmosphere (TOA). Aerosol observations can be divided into two broad categories: in situ and remote sensing. In situ measurement systems collect aerosol particles on filters and analyze the filters in a laboratory (Lenoble et al. 2013). There, a chemical analysis can be performed, mass concentration determined and properties such as the refractive index of the bulk material or absorption properties can be directly measured. Aerosols can also be directly observed in situ by instruments that measure particles’ physical and optical properties. Many of these instruments make use of the particles’ scattering and absorption properties by shining a laser on the particles and then measuring the attenuation of the laser light or the amount or direction of the light scattered (Lenoble et al. 2013).

Chapter written by Van Ha PHAM, Thi Nhat Thanh NGUYEN and Dominique LAFFLY.

TORUS 2 – Toward an Open Resource Using Services: Cloud Computing for Environmental Data, First Edition. Edited by Dominique Laffly. © ISTE Ltd 2020. Published by ISTE Ltd and John Wiley & Sons, Inc.

192

TORUS 2 – Toward an Open Resource Using Services

Although very useful, in situ measurements can provide only limited results, both in time and in space. There is a strong need for aerosol observations on a global basis through the entire column of the atmosphere, which is best obtained by remote sensing. Satellites are increasingly used to obtain information on aerosol properties such as the aerosol optical depth, or the columnar concentration of particles and their sizes, taking advantage of the technical and scientific developments over the last decades. Initially, retrievals were obtained only for measurements over water. Several retrievals of aerosol optical depth from space-borne measurements were performed using observations from the multi-spectral scanner (MSS) onboard the Earth Resources Technology Satellite (ERTS-1) (Griggs 1975), the Advanced Very High Resolution Radiometer (AVHRR) (Stowe et al. 2002) onboard TIROS-N, the Stratospheric Aerosol Measurement (SAM) instrument onboard Nimbus-7, Ozone Monitoring Instrument (OMI) onboard TOMS and AVHRR on NOAA (National Oceanic and Atmospheric Administration) (Geogdzhayev et al. 2002, Hauser et al. 2005). Over land, the retrieval of aerosol properties is more complicated due to the relatively strong contribution of the land surface reflectance to the radiation measured at the top of the atmosphere. One of the first reliable retrievals of aerosol optical depth over land was made using the dual view of the Along-Track Scanning Radiometer (ATSR-2) (Veefkind et al. 1998), followed by retrievals using POLDER (Deuzé et al. 2001), MODIS (MODerate Resolution Imaging Spectroradiometer) (Kaufman et al. 1997) and MISR (Multiangle Imaging SpectroRadiometer) (Martonchik et al. 1998). SeaWiFS and MERIS have also been used for this purpose (von Hoyningen-Huene et al. 2003). In this chapter, we present a methodology to validate satellite-based AOD to ground-based AOD using the R programming language. The experiment is applied on MODIS and AERONET AOD over Vietnam. The following sections present a brief introduction of datasets. The validation methodology will be introduced in section 7.1.3. All validation steps on the R programming language will be shown in section 7.1.4. Finally, a conclusion and future work will be given. 7.1.2. Datasets In this chapter, we validate aerosol optical depth products from MODIS Terra/ Aqua instruments, the MODIS (Moderate Resolution Imaging Spectroradiometer) onboard the TERRA and AQUA satellites. This sensor produces satellite imagery products for land, atmospheric and oceanographic studies including aerosols. The TERRA satellite carrying MODIS moves from north to south and flies over the equator at 10:30 am. The AQUA satellite operates from the south to the north and crosses the equator approximately at 1:30 pm. The orbital time is about 99 minutes. The repeat cycle is 16 days with a swath width of 2,330 km (Qu et al. 2007).

Remote Sensing Case Studies

193

The MODIS sensor is designed with 36 bands ranging from 0.4 to 14.5 μm. There are two bands at 250 m resolution, five bands at 500 m and 29 other bands. The bands used for aerosol studies are 250 m with spectral bands of 620–670 nm and 841–876 nm, and 500 m with spectral bands of 459–479 nm, 545–565 nm, 1,230–1,250 nm, 1,628–1,652 nm and 2,105–2,155 nm (Qu et al. 2007). MODIS satellite data are classified into four levels: Level 0, Level 1B, Level 2 and Level 3. Level 0 data are raw data; Level 1B data are the data that have been processed by sensor units; Level 2 data are geophysical variables at the same location and resolution as Level 1; and Level 3 data include average daily data, average weekly and average monthly data. The aerosol product is estimated using two different algorithms: Deep Blue and Dark Target. The main data of this product are the aerosol optical thickness at 550 nm. In addition, the product also provides more information on scattering, quality information, Angstrom parameters and so on. MOD04_L2 and MYD04_L2 products provide airborne aerosol data with a resolution of 10 km × 10 km; meanwhile, MOD04_3K and MYD04_3K products (which appear in collections 6 and 6.1) provide aerosol data at 3 km resolution. Each image file is formatted in the Hierarchical Data Format (HDF) format, in which the major aerosol data are the combined ocean and land data at 550 nm “Depth Land And Ocean”, with the best quality on land and all quality over the ocean. AERONET AOD data are used to validate satellite AOD. AERONET (AErosol RObotic NETwork) is a network of ground observations established by NASA and PHOTONS (NASA and PHOTONS n.d.). Figure 7.1 shows the network of monitoring sites in the world.

Figure 7.1. Distribution of AERONET monitoring sites in the world. For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

194

TORUS 2 – Toward an Open Resource Using Services

Direct measurements are made in eight bands of 340, 380, 440, 500, 670, 870, 940 and 1020 nm. Optical depth is calculated from the attenuation of the direct radiation beam at each wavelength, based on the Beer–Bouguer law. Rayleigh scattering, ozone absorption and air pollutants are estimated and eliminated to extract aerosol optical thickness. Time frequency of measurements is about 15 minutes. Data from the AERONET monitoring stations are divided into three levels: Level 1.0, Level 1.5 and Level 2.0. Level 2.0 data include quality assurance. The products are divided into two categories: aerosol optical depth (v2, v3) and aerosol inversion. Level 1.0 data are provided in real time. After about 12 months, Level 2.0 data will be available on the AERONET website. AERONET offers a wide range of products based on different algorithms. There are three types of data for each product: time observation data, average daily data and monthly average data. Depending on the time period and station location of collected data, the product may have different levels of data. All data fields at Level 2.0 are presented in Table 7.1. No

Data field

Description

1

Date (dd-mm-yy)

Date of measurement

2

Time (hh:mm:ss)

Time of measurement

3

Julian_Day

Day in Julian format

4

AOD at 1640, 1020, 870, 675, 667, 555, 551, 532, 531, 500, 490, 443, 440, 412, 380, 340 nm

AOD at specific wavelength

5

Water (cm)

Water vapor

6

AOD Inversion 1640, 1020, 870, 675, 667, 555, 551, 532, 531, 500, 490, 443, 440, 412, 380, 340 nm

AOD inversion at specific wavelength

7

%WaterError

Water error ratio

8

Angstrom at 380–500, 440–675, 500–870, 340–440, 440–675 nm

Angstrom exponent

9

Last_Processing_Date (dd/mm/yyyy)

Processing date

10 Solar_Zenith_Angle

Solar zenith angle

Table 7.1. AERONET data fields at Level 2.0

Remote Sensing Case Studies

195

7.1.3. Validation methodology The AOD validation process consists of four steps: First, the aerosol products of the satellite (MODIS AOD) and the ground truth (AERONET AOD) are downloaded from the data sources. Second, satellite images are pre-processed and extracted into AOD data at 550 nm. This includes the extracting process (extract dataset) and reprojection process (GeoTransform and Resampling). AERONET data are extracted and interpolated to the AOD value at 550 nm. Third, MODIS and AERONET data are integrated by time and space. Finally, the integrated data are used to calculate statistics indices such as correlation coefficient (R), root-meansquare error (RMSE) or relative error (RE) and plot some on charts for quality evaluation. The validation methodology is described in Figure 7.2. AOD (MODIS AOD)

Ground truth AOD (AERONETAOD)

Data collecting

Data pre-processing: - Extract dataset - GeoTransform and Resampling - Normalization

- Statistic indices: R2,RMSE, RE - Draw plots

Figure 7.2. Satellite AOD validation process

196

TORUS 2 – Toward an Open Resource Using Services

7.1.3.1. Data pre-processing As mentioned in section 7.1.2, for satellite data, all MODIS AOD images are stored in the Hierarchical Data Format (HDF). Each satellite AOD image contains a lot of information: different datasets (data), information about the image size, projection and coordinates (metadata). The pre-processing process includes the following steps: First, the dataset that needs to be extracted is Optical_Depth_Land_And_Ocean. The extracted data are a two-dimensional matrix with each element in the matrix containing the data of a pixel. Other interested information includes ground control points (GCPs), offset values and scale factors. The AOD data matrix extracted is an integer, not the real AOD value. To ensure accurate values of AOD values, it must be adjusted based on the scale_factor and offset parameters. The actual AOD value at a pixel is calculated as follows: aot_value = scale_factor * stored_integer – offset

[7.1]

where aot_value is the actual AOD value and stored_integer is the AOD value stored in the matrix. Second, the reprojection process includes the georeferencing and resampling step. Georeferencing is the process of converting data from image coordinate to map coordinate, while resampling is the process of resizing and changing the resolution of the image (changing the number of pixels). In this chapter, the gdalwarp utility in the GDAL library was used for this purpose. To determine the coordinates of each cell on the grid, the Thin Plate Spline transform and Nearest Neighbor resampling methods are used. For ground truth data, since AERONET AODs have been measured at 500 nm, we interpolated AERONET AOD at this wavelength to AOD at 550 nm using the Angstrom exponent in the range of 440–675 nm (Mielonen et al. 2011) by using the following equation:

τ 0.5 μm

τ 0.55μm = e

−α 0.44 μm − 0.67 μm ln

0.5 0.55

[7.2]

where τ 0.55μm and τ 0.5μm are AERONET AOD at 550 and 500 nm. α 0.44μm −0.67μm is the Angstrom exponent in the range of 440–675 nm.

Remote Sensing Case Studies

197

7.1.3.2. Data integration Since data from satellites and data from ground observation stations have different spatial and temporal resolutions, there are constraints to integrate regarding time and space (Ichoku et al. 2002). For data integration, MODIS AOD data processing should be performed by averaging the AOD values surrounding the station location. For AERONET AOD, data are averaged in the series of AERONET data values that have time in T minutes from the MODIS data recorded time. After that, data are integrated at T minutes and in radius R to derive the corresponding AOD value of the observation station and the satellite AOD value. We proposed the threshold R as 25 km for MODIS AOD respectively and T as 30 minutes (i.e. duration of 60 minutes coinciding satellite overpasses) for AERONET AOD. 7.1.3.3. Data evaluation Validation of satellite aerosol products has been carried out by comparison of MODIS AOD to AERONET AOD in terms of accuracy. MODIS AOD and AERONET AOD are compared using the following criteria: arithmetic mean (x), standard deviation (σ), coefficient of determination (R2), root-mean-square error (RMSE) and relative error (RE) defined as follows:

x =

σ =

1 n

n

x

[7.3]

t

t =1

1 n

n

 (x

− x)2

t

[7.4]

t =1

2

 n    ( xt − x )( yt − y )   R 2 = n t =1 n  ( xt − x ) 2  ( yt − y ) 2 t =1

t =1

n

RMSE =

[7.5]

 (x

t

− yt ) 2

t =1

n

[7.6]

198

TORUS 2 – Toward an Open Resource Using Services

RE =

| xt − yt | .100% xt

[7.7]

where x and y are the arithmetic means of ground AOD and satellite AOD respectively; x and y are ground and satellite AOD at time t respectively and n is the number of samples. 7.1.4. Experiments and results 7.1.4.1. Experimental environment In this study, AOD data validation was performed using the R language and the GDAL library. The R language is used to call the GDAL library and perform satellite data pre-processing, converting them from HDF to GeoTiff format. Later, integration and validation processes were also performed on the R language under the support of various libraries. Experimental environments, libraries and the language used are listed in Table 7.2. No

Category

Content

Note

1

Operating system

Windows/Ubuntu/CentOS Any version

2

Language

R

Programing language

3

Library

GDAL

Raster pre-processing library

4

raster

R packages for manipulate raster data type (GeoTiff format)

5

stringr

R packages for manipulating string data type

6

gdalUtils

R packages for calling GDAL library

7

hydroGOF

R packages for calculating root-meansquare error

8

sqldf

R packages for querying R data frame object like SQL statement

Table 7.2. Experimental environments and libraries

Remote Sensing Case Studies

199

7.1.4.2. Experimental process 7.1.4.2.1. Data collecting AOD satellite data can be downloaded from the Internet1, such as from NASA, for example, with the selection of a spatial area, satellite type, type of image and date to search for images. In Figure 7.3, the search parameters include space (mapped area), time (time period for image search) and type of image (in NASAprovided image file types). Select product

Select spatial range

Select temporal range

List of results

Figure 7.3. Collect satellite AOD data. For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

Ground truth AOD data (AERONET) can be downloaded from the AERONET website2 with the selection of a station, data type and level of data. In Figure 7.4, the search parameters include geographic region, country/state, AERONET site, start time and end time (time period for image search) and type of image (in NASAprovided image file types).

1 https://ladsweb.modaps.eosdis.nasa.gov/search/ 2 https://aeronet.gsfc.nasa.gov/cgi-bin/webtool_opera_v2_new

200

TORUS 2 – Toward an Open Resource Using Services

Select site

Select product and data level

Select temporal range

List of results

Figure 7.4. Collect AERONET AOD data. For a color version of this figure, see www.iste.co.uk/laffly/torus2.zip

7.1.4.2.2. Data pre-processing AOD satellite data have absolutely no information about the coordinate reference system (CRS). In addition, this HDF image has a lot of information about other MODIS satellite bands. Therefore, we used the GDAL library with the R scripting language, combined with the GDAL Linux shell, to convert HDF to GeoTIFF; the steps are: Read the list of sub-datasets from the HDF: using the gdalinfo command with the input parameter as the HDF file, we will have a list of sub-datasets displayed and retrieve the AOD dataset we want to use. For example: gdalinfo MOD04_L2.A2010214.0300.051.2011120192340.hdf. Therefore, we can get the dataset name, read the AOD values into a pixel matrix array and recalculate the AOD value according to the formula (scale and offset depending on HDF type, where scale = 0.001 and offset = 0 − obtained from the gdalinfo command): new pixel value = current pixel value * scale + offset

Remote Sensing Case Studies

201

Set up the CRS: use the gdalwarp command to return to the reference frame from the latitude and coordinates of the world WGS84. gdalwarp -t_srs ‘+ proj = longlat + datum = WGS84’ -ot Float32 -tps “mecrator.tif” “output.tif” With this reference system, the GeoTiff image has angles with latitude and longitude coordinates, and AOD values are converted from integers to floats. The satellite data pre-processing code is described in the Appendix (section 7.4.). 7.1.4.2.3. Data integration After AOD data have been obtained from satellites and AERONET AOD, the data integration step is performed. The built-in data are output to a CSV file or saved to a database. To perform data integration, the satellite AOD data are taken from the radius R around the station; in this experiment, we chose the value R = 25 km. The averaging around the station is done via the extractRaster function of the R raster library. You can use coordinates (points) of AERONET station to extract satellite AOD value by the following equation: extract(x, y, buffer=NULL, fun=NULL, na.rm=TRUE) where x is the Raster object which is read from the satellite image in GeoTiff format, y is a numeric vector representing the coordinates of the station location, buffer is the radius of a buffer around each point from which to extract cell values and fun is the operator that calculates the average value of the pixels The AERONET AOD data from the ground truth station are taken according to the time before and after T minutes compared to the time the satellite crosses. In this chapter, we choose T = 30 minutes. This process can be performed in R by combining two functions: subset and mean, of which the subset is used to filter the data for a certain period of time, while the mean function is used to calculate the mean on filtered data: selected_data = subset(aeronet_data,aeronet_data$aqstime >= start_time & aeronet_data$aqstime = start_time & strptime(aeronet_data$aqstime,format=“%Y-%m-%d %H:%M:%S”) = start_time & aeronet_data$aqstime