Continuity of multi-sensor AVHRR time series for vegetation monitoring. Case of application to the AMMA project. Ivan VALLEJO VALL January 14, 2005

Ivan Vallejo Vall From July to December 2004

Welcoming firms CESBIO CNES MEDIAS-France 18, avenue Edouard Belin 31401 Toulouse cedex 4 Internship tutor at CESBIO Mr Eric MOUGIN

University T´el´ecom Paris (ENST) 46, rue Barrault 75013 Paris Internship tutor at T´el´ecom Paris Mr Michel ROUX

Acknowledgements This paper as well as the whole internship would have not been possible without the computer skills and patience of Bastien Miras. To him I owe the solving of most of the computer bugs and the optimistic mood in which I always faced unexpected problems. I would like also to thank Patrice Bicheron, for taking the time to help me understand all the algorithms and also to proofread this report several times and give me his valuable advice. Olivier Hagolle has been very important for the correct development of the internship, as well as very calm and clear in the moments were things seemed not to work. If the study advanced at a fast pace and in the good direction, it is greatly due to him. Of course I have to mention all members of the team POSTEL at MEDIASFrance. All of them have been very helpful and attentive. The same can be said of all other members of MEDIAS-France, with whom I have spent most of the internship’s time and to whom I am grateful for the good working atmosphere. Finally, I would like to acknowledge to all the persons responsible of plotting and guiding this internship their outstanding job. The internship’s goal was clear from the beginning and the understanding between CNES, CESBIO and MEDIASFrance sharp and fast. They all made me feel that the work I was doing was meaningful to them and their firms.

Mystical dance, which yonder starry sphere Of planets and of fixed in all her wheels Resembles nearest, mazes intricate, Eccentric, intervolved, yet regular Then most, when most irregular they seem, And in their motions harmony divine John Milton, Paradise Lost

Contents 1 Introduction 1.1 Enterprises . . . . . . . 1.1.1 CESBIO . . . . . 1.1.2 CNES . . . . . . 1.1.3 MEDIAS-France 1.2 Internship’s outline . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

2 Internship 2.1 Description of the chain . . . . . . . . . . 2.1.1 Overview . . . . . . . . . . . . . . 2.1.2 Level 1 and 2 . . . . . . . . . . . . 2.1.3 Level 3a . . . . . . . . . . . . . . . 2.1.4 Level 3b . . . . . . . . . . . . . . . 2.2 Accomplished tasks . . . . . . . . . . . . . 2.2.1 Intersatellite calibration . . . . . . 2.2.2 Validation of navigation’s accuracy 2.2.3 Production of output data sets . . 2.3 Evaluation of the results . . . . . . . . . .

. . . . .

. . . . . . . . . .

. . . . .

. . . . . . . . . .

. . . . .

. . . . . . . . . .

. . . . .

. . . . . . . . . .

. . . . .

. . . . . . . . . .

. . . . .

. . . . . . . . . .

. . . . .

. . . . . . . . . .

. . . . .

. . . . . . . . . .

. . . . .

. . . . . . . . . .

. . . . .

. . . . . . . . . .

. . . . .

. . . . . . . . . .

. . . . .

. . . . . . . . . .

. . . . .

6 . 8 . 8 . 11 . 16 . 18

. . . . . . . . . .

19 19 19 21 25 26 27 27 32 41 41

. . . . . . . . . .

3 Conclusion 47 3.1 Perspectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47 3.2 Personal balance . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 A Symbols and acronyms

50

B Sensors’ Description 51 B.1 AVHRR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51 B.2 VEGETATION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53 B.3 POLDER . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53 C Attitudes C.1 Simulator of attitudes

56 . . . . . . . . . . . . . . . . . . . . . . . . . 57

Abstract The advanced very high resolution radiometer (AVHRR) time series are among the most commonly used wide field-of-view products. Their availability and their global coverage has motivated many scientists to investigate interannual variability and trends in land surface conditions. The reliability of the results remains contested because of the degradation of AVHRR shortwave channels and the variation of illumination conditions. This paper evaluates the performances of a new set of processing algorithms for AVHRR products based on surface reflectances. The Sahel is chosen as a zone of study during the period of November 2002 - February 2003. The criterion used to evaluate the quality of the corrections are founded on the analysis of simultaneous imagery from the AVHRR16 and AVHRR17 sensors, onboard NOAA-L and NOAA-M satellites. This study shows that the performances of current AVHRR products can be improved using the algorithms detailed in this paper. The normalized difference vegetation index (NDVI) of composited images is corrected so that artifact trends are assessed to be inferior to that of the real ones by a factor of at least 2. AVHRR surface reflectances are corrected to reach a level of quality similar to that of VEGETATION ones. Finally, the representability of the case of study is discussed and a procedure to continue on the reprocessing of the AVHRR archive is suggested.

Chapter 1 Introduction The advanced very high resolution radiometer (AVHRR) on board NOAA polar orbiting satellites (POES) has been furnishing global observations since 1981. The AVHRR was originally designed as a weather imaging sensor. Nevertheless, the spectral response of channel 1 and 2 made the sensor also useful for quantitative studies of Earth’s environment (see Figure 1 and Appendix B.1 for details on AVHRR technical characteristics). The long term availability of AVHRR data sets has proved to be a unique advantage. It has rendered the use of AVHRR time series prevalent in the scientific domain of interannual variability and trends in land surface conditions [1]. Many papers have been published on the subject, such as [2] and [3] on vegetation activity and photosynthetic trends or [4] on climate variations. However, the use of AVHRR time series as a basis to scientific research encounters serious challenges, specially when dealing with long-term trends. The overwhelming amount of data and its reliability remain the most important ones to face. Most studies use the normalized difference vegetation index (NDVI) as an indicator of the photosynthetic activity. Therefore, the greenness evolution of a certain zone is inferred form the reflectances measured by channel 1 and 2. Accuracy is consequently dependent on the calibration of these two channels. Figure 1: Normalized response of AVHRR The age of the POES system of satelchannels 1 and 2. Channel 1 covers the red lites makes it difficult to intercalibrate part of the spectrum and channel 2 the near the succeeding satellites, which have difinfrared. Thus they serve well as and NDVI ferent characteristics : radiometric reindicator. sponse, viewing configuration, etc. Moreover, the aging causes satellites prior to

7 NOAA-151 to drift and thus the equator-crossing times vary with time (see Figure 2(a)). For instance, that of the afternoon satellites shifts to a later hour. Therefore observations are made progressively later until a new satellite takes the place of the old one, producing an abrupt drop back to an earlier hour. These alterations of solar illumination conditions have an important impact on the measured radiance. Other common error sources are instrumental noise and navigation errors. Several NDVI data sets are available, being the PAL [5] and the GIMMS [6] ones two of the more commonly used. Notwithstanding all the efforts undertaken to correct the fluctuations, artifact trends remain with a magnitude comparable to that of the real ones in a 20 years perspective. Figure 2(b) shows the remaining trends in a desert area that should have a plain NDVI evolution. As pointed in [1], post-launch calibration is mandatory. Furthermore, available data sets may display misleading trends. In order to profit from the great potential of the AVHRR temporal series, a new approach is necessary to improve data reliability. In the framework of the european project Cyclopes, MEDIAS-France and CNES have developed a new image processing chain, with a new set of algorithms at its core. Those algorithms were conceived for the sensor VEGETATION, but can equally be applied to AVHRR. Figure 2: (a) Equator crossing time for afterBefore taking in charge the repro- noon passes. (b) NDVI in the Arabian desert cessing of the whole AVHRR archive, for four different data sets. Data are smoothed preliminary tests have to be made to and offset for clarity. determine the quality of the new processing method. This internship is intended to evaluate the quality of the images produced and to improve and develop the processing chain, so that it reaches the standards necessary to the use of these products in long-term scientific research. The CESBIO provides the scientific savoir-faire necessary to understand and evaluate the results.

1

For NOAA KLM series (15, 16 and 17) the phenomenon of orbital drifts was already known before the satellites were launched, so they were modified in order to rectify the orbits when necessary.

1.1. ENTERPRISES

1.1

8

Enterprises

Three firms converge on the internship: CESBIO, CNES and MEDIAS-France. They have complementary expertises which make it possible to reach a goal of such complexity. From the algorithms developed at CNES2 through the computer science skills of MEDIAS-France, to the scientific knowledge of CESBIO, all parts of the process are reviewed during the internship. A good coordination between the different actors as well as the understanding of their singular interests is crucial to the success of the internship.

1.1.1

CESBIO

Center for the Study of the Biosphere from Space (CESBIO) is a mixed research unit made up of four different research institutions: UPS, CNRS, CNES and IRD. It has the objective to develop the knowledge on continental biosphere dynamics and functioning at various temporal and spatial scales. Some of his main missions are: • To do research in the domains of observation and modelling of the continental surfaces. • To participate in the specification of space missions and processing of remotely sensed data. • To provide the interface between physical and biological sciences. • To be in direct contact with socio-economic community. It organizes itself in work groups and projects. There are three work groups divided by their subject of study: • Remote sensing data and extraction of structural and functional information of the surface. • Modeling of ecosystem functioning. • Modeling of the biospheric component of the hydrological cycle. Projects are more representative than work groups. They are divided by their zone of study: • South-West. • South-Med. • Global.

2

Concretely, the algorithms were developed by the division Physique de la Mesure Optique.

1.1. ENTERPRISES

9

The internship is ascribed to the global project.

Global project The objective of the global project is to understand and model vegetation cycles of continental surfaces and their evolution in order to predict global changes. That means to find answer several key questions: • Is it possible to quantify ’Global change’ ? • How do we distinguish between interannual fluctuation and sustainable evolutions? • What is the impact on earth’s resources? • How can be distinguished climatic and anthropic impacts? The project endeavours to answer those questions with the following methods: • Satellite data. • Models. • Ground measurements and external data to calibrate and validate the developed models and interpret the results. The final goal is to understand and control the water, energy and carbon fluxes of the sea/surface interface at a global scale. The effort is concentrated on three regions: • The peri-arctic zone, which is very sensitive to climatic changes and is under relatively low anthropogenic influences. Biomass and snow cover are functional parameters studied within the SIBERIA II project. • The circum-saharienne zone, which is an exemple of a region under the strong influence of climate and anthropogenic effects, because of its position between a desert ecosystem to the north and a humid savannah to the south. The Sahelian zone is the object of particular attention. These studies are conducted at the WestAfrican scale with large view optical sensors (AVHRR, VEGETATION) . Also microwave sensors (ERS) have been used to extract surface characteristics since 1991. Moreover, CESBIO is participating in the European project AMMA, which aims to study the influence of the continental surface on the establishment of the monsoon regime in West-Africa. It studies also the feed-backs of the interannual variability of the monsoon with the vegetation dynamic functioning. • The European zone, as an example of a region where the anthropogenic effects are more important than the natural climate variability. This project is related to studies at regional scale such as those of the South-West project. The internship lies within the European project AMMA.

1.1. ENTERPRISES

10

AMMA The goals relating to the ”Continental Surfaces” of the AMMA project are to describe the evolution of terrestrial surfaces during the 50 last years while distinguishing between anthropic and climatic forces and to study the possibility of an impact of this evolution on monsoon. The studies in this domain should advance our comprehension of the role that continental surfaces play in the installation, evolution and interannual variability of West African monsoon. They should also make it possible to quantify the impact on the surface properties of climatic variations on various time scales which characterize monsoon. It is thus necessary: • To study in detail the evolution of the biomass, of the vegetation cover (types of vegetation, density), the moisture and the surface temperature on local and global scales, by using different available archives, in particular satellite (optics and microwave). • To understand how the surface parameters act on the dynamics of the cycle of continental water. To tackle this strong multidisciplinary problem the task was divided accordingly to the space scale characteristics of the various interactions which come into play: • On small space scales (lower than 10 km) the goal is to understand the processes which intervene in the evolution of the heterogeneity of the surface generated by hydrology or the vegetation. These scales are also those of the genesis of the convection and the operation of the convective systems. The studies on a small scale will be validated on basins of the area. • On medium space scales, the goal is to understand the seasonal memory of continental surface. It is on these scales that the interaction of the North-South gradient of the vegetation, the progression of the convective clusters and that of the rainy season on West Africa will be analyzed. This aspect of the project will cover studies which will be made on the large water basins of the area by thus offering also another place of collaboration between the communities. It is particularly the case of the basins of Niamey and Ou located in the window CATCH. The first basin is located in the sahelian zone covering 12000 km2, the second is in a Sudanese environment with a surface of 14000 km2. The internship finds its place in the medium scale data analysis of the Sahel.

1.1. ENTERPRISES

1.1.2

11

CNES

The French space agency, CNES, was created in 1961 by the French government. The CNES is a public establishment with an industrial and commercial profile, in charge of the development of French space activities. The CNES’s mission is to make proposals to the government concerning the orientations of space policy in France. It then carries out the chosen programs in collaboration with its partners in the industry, research, and defence sectors. The budget managed by CNES is of 1876 million euros (adding up the state subsidy of 1412 million euros and CNES’s own resources 464 million euros). The main objective of CNES is to develop the space utilities in order to satisfy the public, military and scientific community requirements and to support the emergence and spreading of new applications, creative poles and employment in the the space domain. Another important goal of CNES is to give life to the spatial science politics. For example, in fields related to Earth observation and oceanography it has obtained successful results allowing a better knowledge of our planet.

Roles The CNES conducts French space policy in two complementary directions: • By participating in the programs of the European Space Agency (ESA) in which it plays a major role. • By carrying out a dynamic national program, to guarantee strong industrial competitiveness worldwide. The CNES has 2,500 employees spread over its different sites and generates industrial activities representing more than 10,000 jobs in France.

Activities French space policy is implemented in areas of strategic, economic and scientific importance: • Access to space, where CNES has geared its strategy toward assuring a full spectrum of launchers for Europe, with Ariane, Soyuz and Vega operated from the Guiana Space Center. • Space applications for the consumer market, through development of space technologies for digital television, navigation, broadband Internet, positioning, telemedicine and education (e.g. AGORA and Galileo). • Sustainable development, encompassing environmental applications, natural resource management, climatology, and natural and man-made hazards (e.g. SPOT, Jason, Envisat, GMES, Demeter, Argos, METOP, IASI and Pleiades). • Science and technology research, in conjunction with European and international research organizations, to support exploration of the Universe, efforts to discover the origins of life, fundamental physics and microgravity experiments. E.g., Rosetta, COROT, Odin, Cassini-Huygens, Mars-Express, Microscope and Pharao.

1.1. ENTERPRISES

12

• Defence activities, which drive programmatic choices designed to meet military requirements (Helios and Essaim) and define how dual-use systems such as Pleiades are structured.

Subsidiaries and holdings CNES is a shareholder of eight limited liability companies (ARIANESPACE, CLS, INTESPACE, NOVESPACE, SCOT, SPOTIMAGE, CERFACS and SIMKO), a partner in a Private Limited Liability Company (DERSI), a member of four Economic Interest Groups (GDTA, MEDES, PROSPACE, SATEL CONSEIL), five Public Interest Groups and a non-trading company. The CNES is one of the main spatial centers in Europe and the most important contributor to the European Spatial Agency (ESA). In figure 1.1.1 it is shown all the participations of CNES in the space sector.

Programs The CNES is responsible for applying the French space policy and making proposals within ESA. The CNES is therefore involved in several fields of activity. As a guarantee of Europe’s independence in space, launching activities are given top priority. Simultaneously, CNES fosters the development and use of new applications and plays an essential role in space sciences, such as Earth observation, exploration of the universe or microgravity research. The most important activities of CNES are: • To facilitate the access to space with the Ariane program and the development of the Guyana launch center. • To promote operational and commercial applications of spatial techniques, like Earth observation (Spot, ERS, Meteosat, IASI,...) and telecommunications (T´el´ecom2, Stentor, Galileo, ...) • Scientific programs in partnership with research organisms in European and international co-operation (Topex-Pos´eidon, Jason, Envisat, Polder 2, Cluster, Odin, Corot, Cassini-Huygens, Rosetta, Mars-Express). • Research activities and training experiences for the International Spatial Station.

CNES’s centers To accomplish its assigned missions, CNES calls on the specialist expertise of its 2,500 employees in four centers. The Paris Headquarters: Head Office, in Paris, is the central hub responsible for organizing and running the agency. It shapes and promotes CNES policy in conjugation with the Ministries of Research and Defence. It also works with the technical centers to define CNES’s strategy and relations with outside partners. Nowadays, about 250 persons are working in the Paris establishment of CNES.

1.1. ENTERPRISES

Figure 1.1.1: Subsidiaries and holdings of CNES.

13

1.1. ENTERPRISES

14

The Launch Division (DLA): Located in Evry, the Evry Space Center leads the Ariane launcher developments on behalf of the European Space Agency. It also offers ongoing support to Arianespace, which is in charge of industrial production, marketing and launches. It is also paving the way for future developments through innovative launcher concepts and advanced propulsion systems. The Toulouse Space Center (CST): The Toulouse Space Center is a technical and operational center unique in size and scope. It works closely with industry to develop complete space systems from conception through to commissioning. The CST conducts all satellite positioning and orbit control operations for which CNES is responsible. It is also responsible for balloon activities. The Guiana Space Center (CSG): The Guiana Space Center (French Guiana) is Europe’s spacecraft hub in Kourou, dedicated to the Ariane program and soon to the Soyuz and Vega launchers. It coordinates all the launch support facilities, including radar tracking of satellites, reception and processing of data from the launch vehicle. The CSG is responsible for all aspects of range safety during satellite launches.

The Toulouse spatial center The spatial center of Toulouse was created in 1968. Nowadays it is a very important technical and operational institution for its activities in space. The CNES in Toulouse develops all the spatial systems and also takes part in the realization of scientific or instrumental projects and pilots the application of research programs such as Spot, H´elios, T´el´ecom, Agros or Cospas-Sarsat. This center leads the setting and deployment of satellites in orbit. Thus, the spatial center of Toulouse develops an important research and technological program. Today, about 2500 persons are working in this center. The CST (Center Spatial of Toulouse) mission is defined by the following points: 1. Control of the basic techniques necessary to the development of space systems allowing the French space research to continue its action in a national, European and international context. 2. Control and exploitation of operation systems for the account of national and institutional customers. In this environment CNES studies, develops, operates and qualifies all the spatial systems. However the future of spatial systems requires also the development of instruments, spatial engines and the ground sector. This program of research and development is divided in two parts: • The R &D themes: radiocommunications, observation sciences, orbital infrastructures and spatial transport. • The generic R &D: system design and development, technique, technologies and electronic components.

1.1. ENTERPRISES

15

Service Physique de la Mesure Optique (SI/MO) Its domains of expertise are the physical principles of measurements and the processing techniques of optical and infrared instruments: multi and hyperspectral imaging, spectro-imaging, spectroscopes and probing devices or alerting instruments. The service exercises its responsibilities in the accounting for the scientific needs, in the scientific quality expertise, in the physics of measurements and in the calibration and intercalibration of instrumental chains. It assures the access to the calculating tools needed for those functions, such as the atmospheric models. The service is also in charge of the development of processing algorithms and the activity system/quality of the associated product for wide field imaging devices. This second activity deploys itself in tasks such as: technical analysis of mission and translation of system’s specifications, simulations, studies of feasibility, system modeling, establishment and control of balance sheets, system checking over during their operative life span, technological experiments related to image quality, etc. In concrete terms, the service deals with the radiometric aspects of measuring chains. It stands responsible for the absolute calibration and the intercalibration of the instruments. From the internship’s point of view, this department of CNES has developed the algorithms that sustain the chain. Their main interest is to prove the utility of these algorithms and also to support the establishment of the thematic pole POSTEL.

Figure 1.1.2: The main entrance of the CST

1.1. ENTERPRISES

1.1.3

16

MEDIAS-France

In France, seven institutions (CNES, CNRS, IRD, METEO-FRANCE, Toulouse University, CLS and SPOT-IMAGE) are supporting a public, non-profit corporation, headquartered in Toulouse, called MEDIAS-France, with a view to co-ordinate French contributions within the network. Medias-France works in close collaboration with national laboratories and many other partners in France and abroad. MEDIAS-France supports research projects on a regional and a global environment, from a sustainable development point of view, particularly within the Mediterranean Basin (MEDI) and Subtropical Africa (AS), in cooperation with Figure 1.1.3: Partners of MEDIASnational (Eclipse, GICC, IFB, ORE, RTE, Zones- France Ateliers,etc.) and international (IGBP, WCRP, IHDP, Diversitas, FP6,etc.) programs.

Structure This support shows itself in services such as: constitution and management of databases, supply of biogeophysic products produced from space observation data (Postel project), formation and information (summer schools, workshops, seminars, Medias newsletter, etc.) or coordination and consultancy (constitution and management of Websites, administrative support, management of programs, etc.).

Domains of expertise In its area of geographical interest and with research themes relating to global change, MEDIAS-France develops its activity in two technical axes: Databases: data management and coordination of scientific data networks Observation of land surfaces: satellite data processing and production of final biophysical indicators. E.g., leaf area index, fraction of vegetation, albedo, downward radiation flux, surface moisture and temperature, etc. The internship is circumscribed to the second domain, in the framework of the POSTEL project.

POSTEL POSTEL is a national project concerning the development of a pole of thematic competences, in the area of continental surfaces. It is intended to lead to the future installation of a European Center of Services of Biogeophysical Parameters (CSP). The national organizations CNES, METEO-France, IRD, CNRS and INRA are presently considering their support to POSTEL. A multi-organism convention is under preparation. POSTEL has the purpose to be a national ”module” in an European

1.1. ENTERPRISES

17

network. It falls under the European program GMES through the FP6 GMES Integrated Project called geoland. The activities entering within the framework of POSTEL are: • To set up the Center of Services of Biogeophysic parameters (CSP). • To build its interfaces upstream with Centers of Expertise and downstream with users. • To take part in European and national precursory projects (geoland, CYCLOPES, POLDER, AMMA among others), viewed as milestones in the construction of a future CSP. • To provide products to the users, via these precursory projects. Although projects share many resources and often overlap each other, the internship was conceived to take place in the environment of the AMMA project, in which takes also part CESBIO.

1.2. INTERNSHIP’S OUTLINE

1.2

18

Internship’s outline

CESBIO is interested in the study of vegetation trends in the Sahel zone, in a long-term (decades) perspective. The only source that can provide such an amount of continual data is the AVHRR sensor. Thus, CESBIO is willing to test Cyclopes processing chain to evaluate the possibility of reprocessing and so improving the AVHRR temporal series. MEDIAS-France and CNES are engaged in the development and improvement of Cyclopes. Cyclopes takes as input AVHRR and VEGETATION data files. These products are classiFigure 1.2.1: Zone of study over the Sahel refied in three levels according to the gion. definition given by the Comittee on Earth Observation Satellites (CEOS) [7]. Level 1 products provide users with geolocated and calibrated top of atmosphere (TOA) reflectances acquired during a time span not exceeding one orbit. Level 1 processing does not make any assumption on the physical nature of the observed target: it only corrects for sensor artifacts. Level 2 converts TOA reflectances into land surface reflectances. Finally, level 3 products are the chain’s final products, i.e, land surface reflectances corrected and composed over a certain time period. These products are useful for applications that do not require daily observations: they provide global or regional maps that minimize data volume and cloud cover. For the case of study, Cyclopes was to be tested for the processessing of AVHRR level 1b data, i.e., GAC 1b products for AVHRR (see [8] and [9] for details about data formats), into level 3 AVHRR composite images. A period of 4 months, from November 2002 to February 2003, was designed to test the performance of the processing method. At that period, several AVHRR sensors were working at the same time, obtaining images of the same spot from different satellites and therefore at different hours. That fact makes it possible to easily compare the robustness of Cyclopes, because composite products of different satellites should render the same results. That is so due to the fact level 3 reflectances are normalized to a common hour and viewing geometry, thanks to the application of a Bidirectional Compositing (BDC) algorithm. Moreover, that allows also to determine the degree of intersatellite correction that can be achieved. That is, the correction of the abrupt discontinuities that can be seen when a change of satellite occurs (see Figure 2). AVHRR 15, 16 and 17 data were collected from the NOAA site in order to realize the study. The geographical zone taken into account was that of Figure 1.2.1. The internship was divided in three stages: a first one concerning documentation and bibliography, undergone at CESBIO, a second one centered on development and a third one on synthesis and analysis, the two last of them effectuated at MEDIAS-France. Some specific processing programs needed to be launched at CNES.

Chapter 2 Internship 2.1 2.1.1

Description of the chain Overview

Cyclopes was first developed to bring images acquired by the AVHRR and VEGETATION instruments to the same reference. That is, to correct using the same algorithms both sets of images and normalize them to a single reference. Appendixes B.1 and B.2 detail the characteristics of both sensors. The output products of the processing chain are FAPAR, NDVI, LAI and biome images, which all are biophysical indicators and so final products for a researcher. In the case of study of the internship, the chain is used to obtain the NDVI images of different AVHRR sensors working at the same time. I.e., they are normalized to a single solar time. It is important to remark that Cyclopes produces other biophysical products apart from the NDVI, so the evaluation of the reflectances themselves would be also of interest. The chain was conceived taking into account the overwhelming amount of data to be processed. An effort was made not only to develop the algorithms but also to engineer the software. The result is a complex arborescence of binary files written in C, C++, KSH and PYTHON scripts, to be executed in a Unix/Linux environment. The process is divided in three independent and automatized parts: Level 1 & 2: processing of AVHRR GAC files into projected TOA images (level 1) and conversion of those TOA reflectances into surface reflectances (level 2). Level 3a: fit of a Bidirectional Reflectance Distribution Function (BRDF), which turns the surface reflectances into normalized reflectances (level 3a). Level 3b: computation of biophysical indexes. Level 3b provides no feature apart from the plain calculus of the indicators. The other two levels are the core of the new processing technique. Level 3a enforces the BRDF model while level 1 & 2 apply all the corrections, i.e., data projection, atmospheric rectifications, calibration, cloud screening, etc. Summarily, the three most sensitive or innovative parts of the process are:

2.1. DESCRIPTION OF THE CHAIN

20

• The intersatellite calibration (level 1). It takes as reference data from the POLDER sensor, on board the ADEOS satellite. • The cloud screening enhanced algorithm for VEGETATION (level 2). A more standard one is applied for AVHRR data. • The application of a BRDF model to correct the solar illumination conditions (level 3). A quick overview of the whole process will be given. Afterwards, during the presentation of the internship’s progress those parts that were more problematic will be discussed in detail. For more precisions about the chain functioning see [10] and [11].

Figure 2.1.1: Processing stages of Cyclopes

2.1. DESCRIPTION OF THE CHAIN

2.1.2

21

Level 1 and 2

Stage 0 The input products for Cyclopes are level 1b GAC data files, as can be fetched from NOAA’s web site [12]. It must be taken into account that although they are free of charges there’s a downloading delay which brings the real transfer rate down to 1 day of recorded GAC data per hour. When dealing with long time periods or with several satellites at the same time, the downloading process must be scheduled in advance. It must be also planned how to stock the data: a compressed file containing all orbits for a given day, with a global coverage, has a size of about 850 Mbytes, to which it has to be added the disk space needed to stock the resulting images of each level of Cyclopes.

Stage 1 It consists of several steps: a first one of data formatting, a second one of calibration and a last step of data projection (i.e., navigation). Concerning the task of calibrating the digital numbers (that is, direct read-outs from the sensor), different processes are implemented for the thermal and the red/infrared channels. Channel 3, 4 and 5 (thermal channels) contain the Earth scene count. A series of transformations must be realized to obtain a temperature out of the radiance. Firstly, a linear model is used to compute the linear radiance: NLIN = Slope C10 + Intercept

(2.1)

C10 corresponds to the 10 bit digital number. The slope and intercept values are read directly from GAC files and may vary during the orbit. Secondly, the Earth scene radiance is calculated taking into account the non-linear effects observed in channel 4 and 5 (2.2 and 2.3 are not necessary for channel 3): 2 NCOR = a0 + a1 NLIN + a2 NLIN

(2.2)

NB = NLIN + NCOR

(2.3)

At last, to convert the Earth scene radiance into the equivalent blackbody temperature the following two steps are necessary: TB∗ =

c2 νC ln[1 + (

3 c1 νC NB )]

T = A0 + B 0 TB∗

(2.4)

(2.5)

Accordingly, a set of six parameters (a0 , a1 , a2 , νC , A0 , B 0 ) is needed for each of the three channels. They are provided by NOAA itself and can be found at [9], along with all the constants

2.1. DESCRIPTION OF THE CHAIN

22

For the red and infrared channels a simpler calibration process is enforced because of the absence of onboard calibration devices:  low low Xi < Xithreshold  Gi Xi + Of fi Li = i = 1, 2 (2.6)  high Gi Xi + Of fihigh Xi ≥ Xithreshold where Xi is the digital number. A salient feature of the AVHRR/3 is the use of dual gain detection circuitry in the visible and near-infrared channels to enhance radiometric resolution at the lower end of the dynamic range of the albedo. It is the same technique as 2.1 but taking into consideration the split-gain feature. The slope, intercept and cross-over points of each channel (5 parameters) are read directly in the data stream. The calibration of AVHRR red and infrared channels is insufficient. It is so because AVHRR was conceived only for meteorological applications, which do not require the accuracy that vegetation monitoring does for surface reflectances. Thus, a second phase of calibration has been added in order to improve the results. It consists in a cross calibration of the reflectances provided by the sensors. As explained in [13], several desert sites of 100x100 km2 are chosen as targets. Observations of those sites by each sensor are calibrated taking as reference data from POLDER. Thanks to the large spectral and directional capability of POLDER (see Appendix B.3 for more information), a quasi complete coverage of the hemisphere for solar zenith angles between 10◦ and 50◦ is available. If we formulate the relation between the digital number and the radiance as follows: DNV = α Radiance = AK (t) (ρK cos θS ) (2.7) where θS is the solar zenith angle, the calibration consists in determining the good AK (t) for each t. If we write the final coefficient as that necessary to correct the pre-launch calibration: AK,deserts (t) ∆AK (t) = (2.8) AK,ref (t) we can state the deviation from the reference calibration as a function of the digital number, the reflectance measured by POLDER and θS : t = t1   ∆AK =  AK,ref ∗ ρP OLDER ∗ cos θS  DN

(2.9)

Equation 2.9 is computed for all data available. A statistical expo linear model is fitted to characterize the variations with time of the instrument:  if d < d0  α eδ (d−d0 ) + β AK,deserts /AK,ref = (2.10)  α δ (d − d0 ) + β + α if d ≥ d0 where d is the current day and d0 is that taken as reference. Figure 2.1.2 is an example of the curves obtained. After the calibration process, it is necessary to project the data into a cartographic grid. That is, to transform the instrument’s geometry into that of the observed scene. This is undertaken by the program Navigate. It was developed by Daniel Baldwin

2.1. DESCRIPTION OF THE CHAIN

23

Figure 2.1.2: Light blue circles show the calibration coefficients ∆AK (t) averaged over one month for: (a) AVHRR14 channel 1, (b) AVHRR14 channel 2. The red dashed lines correspond to the sensor official calibration, the blue dashed lines to a linear fit and the green ones to an expo linear fit of the enhanced calibration coefficients.

(Colorado Center for Astrodynamics Research) and later slightly modified by CNES. It carries out a plate-carr´ee projection of the orbits with a linear interpolation, at a 4 km resolution. Moreover, it takes into account the timedrift of the on board satellite clock. As shown in Figure 2.1.3, there’s a difference between the GMT time and the internal clock, which is watched by the Satellite Operations Control Center and kept between ± 0.5 seconds. An erroneous estimation of a scanline’s time of acquisition causes a bad location of that line, thus badly assigning the digital numbers. Navigate corrects the drift if it is given as an input parameter.

Stage 2 Data must be corrected in order to remove as much as possible the effects of atmospheric conditions of acquisition on measured reflectances. The set of auxiliary data used in Cyclopes consists in: • Ozone data files containing information about its concentration. One file per day is available. • Meteorological data files where the sea-level pressure and water vapour information is found. There’s one file for each 6 hours.

2.1. DESCRIPTION OF THE CHAIN

24

Figure 2.1.3: Typical timedrift of NOAA onboard clock.

• The digital land model and the terrestrial cover file. All auxiliary data are resampled to match the resolution of the reflectances. A bilinear interpolator (linear in longitude and in latitude) is used, except for meteorological files that are interpolated with a trilinear one (to profit from data prior and posterior to the time of acquisition). Sea level pressure files are corrected to obtain the ground level pressure of those points above the sea. The digital land model and the terrestrial cover file are processed in order to build the land/sea mask. That is, a map where sea pixels are displayed in opposition to land ones.

Stage 3 A whole stage of the chain is devoted to the construction of the cloud’s mask. In order to detect the cloudy pixels, the thermal channels are exploited. The CLAVR [14] algorithm is employed. In short, it imposes a set of thresholds to reflectances and to temperatures. Those enforced on reflectances screen unnatural high values due to cloud reflected acquisitions. Those imposed on temperatures try to ascertain cloudy pixels by their low temperatures. It is worth remarking that the cloud detection stage is one of the most sensitive parts of the chain and that it has a great effect on its final outcome. This issue is discussed in detail in Section 2.3.4.

Stage 4/5/6 Water vapour, pressure and ozone files are taken as input for the atmospheric corrections phase. So is the cloud’s mask, to prevent the cloudy pixels from being processed, and the angle files (i.e., view zenith angle, view azimuth angle, solar zenith angle and solar azimuth angle). The angle’s information is read directly from the GAC files. The goal of this stage is to convert the TOA reflectances into surface reflectances. The inversion is accomplished using the SMAC algorithm (see [15] for more details on

2.1. DESCRIPTION OF THE CHAIN

25

SMAC). Apart from the before mentioned inputs, a set of SMAC coefficients is needed. They concern parameters that intervene in the calculus of the atmospheric functions. They are provided by CNES. The last two phases take care of the resampling and tiling of the images. It is thus possible to change the resolution to coarser ones (i.e., greater than 4km) and to divided the image into the several zones of study. See Appendix B.1 for details about AVHRR GAC data resolution.

2.1.3

Level 3a

In the AVHRR time series one terrestrial target is viewed once or twice a day. The viewing conditions of a target change during short time periods (e.g, 10 days) and yet it may be considered that the target surface remains invariant. To integrate multiple observations into a single robust one, surface reflectances must be corrected using a bidirectional model. The objective is to produce a series of images in which all of them suffer the same effects due to the geometry of observation. Without this rectification, the trends observed in the surface reflectances of AVHRR temporal series may be partially induced by different acquisition conditions. The Figure 2.1.4: Diagram of solar illuminamagnitude of these effects can lead to large tion angles. errors, in particular when observing the phonological evolution of vegetation on a regional scale [16] [17]. The BRDF model applied in Cyclopes is that discussed in [18]. It consists in three kernels fitted with three subsequent parameters: ρ(θS , θV , φ) = k0 + k1 f1 (θS , θV , φ) + k2 f2 (θS , θV , φ)

(2.11)

where: f1 (θS , θV , φ) =

i 1 h 1 (π − φ) cos φ + sin φ tan θS tan θV − tan θS + 2π π  p + tan θV + tan θ2 S + tan θ2 V − 2 tan θV tan θS cos φ

f2 (θS , θV , φ) =

h π i 1 4 1 ( − ξ) cos ξ + sin ξ − 3π cos θS + cos θV 2 3

(2.12)

(2.13)

and: cos ξ = cos θS cos θV + sin θS sin θV cos φ

(2.14)

The number of parameters is chosen as low as possible in order to reduce data needed to fit the model. A larger number of parameters would be problematic when dealing

2.1. DESCRIPTION OF THE CHAIN

26

Figure 2.1.5: Diagrams of functions f1 and f2 for three solar angles θS as a function of θV in the principal plane (φ = 0◦ = 180◦ ). Positive (negative) θV correspond to forward (backward) scattering.

with cloudy days and in general with AVHRR data. On the other hand, a model such as [19] that only needs two parameters does not account for the relative azimuth between solar and viewing angles dependence. The first parameter, k0 , represents the surface reflectance when both sun and sensor are at the nadir. The k1 parameter determines the magnitude of geometrical and shadowing effects. The f1 component is modelled by vertical opaque protrusions placed on a flat horizontal plane and reflecting according Lambert’s law. They represent mainly irregularities of bare soil surfaces but may also account for structured features of low transmittance canopies. As seen in Figure 2.1.5, below θS = 51◦ the function has a local maximum in the backscattering direction. It is strongly dependent on the azimuth φ. The k2 parameter establishes the magnitude of volume scattering effects. The f1 function is modelled as a collection of randomly located facets absorbing and scattering radiation. They represent leaves of canopies characterized by a nonnegligeable transmittance, but can also model dust and fine structures. It is shown in Figure 2.1.5 that it has a minimum in the principal plane in the forward scattering side and that it increases with θV when it is sufficiently large. Unlike f1 , it has little dependence on the azimuth φ. Level 3a realizes the synthesis of the images obtained in level 2. The bidirectional compositing program normalizes the reflectances to nadir viewing geometry and to a certain solar time, producing one synthesis for each time period taken as a basis to the composition. This compositing period can be adjusted. In order to correctly make the regression at least four observations are necessary. For more information about level 3a see [11]

2.1.4

Level 3b

The last level of the chain makes the conversion from level 3a products to biophysical products: FAPAR, NDVI, LAI, biomes, etc. The output formats are those detailed in

2.2. ACCOMPLISHED TASKS

27

[10] and [20].

2.2

Accomplished tasks

In order to obtain the set of final images necessary to evaluate the quality ofCyclopes’ corrections, several tasks had to be previously carried out. First, computer programs had to be updated to take into account all satellites involved in the case of study. It was also mandatory to become acquainted with the structure and functioning of the whole set of programming files, so as to be able to modify them later on. That entailed also the knowledge of all the algorithms and processing techniques applied inCyclopes. A second part was devoted to intersatellite calibration. That is, to assure the coherence of measured radiances belonging to different satellites. A third one was dedicated to navigation, i.e, to ascertain the good location of all pixels. A good navigation assures that images agree with each other, so that each pixel corresponds to the same spot of the terrestrial globe. Finally, the process of production itself had to be managed. All results from the previous parts had to be introduced in Cyclopes and the processing launched. As the first task accomplished, it concerned mainly programming. Cyclopes was developed to process VEGETATION 1 & 2 and AVHRR 14 and 16 products. In order to perform the study the chain needed to be updated for AVHRR 15 and AVHRR 17. The changes that were carried out involved the recuperation of time drift files for AVHRR 15 and 17 as well as files containing their thermal calibration coefficients. Furthermore, the code was locally changed to solve a few problematic issues that had arisen. For instance, channel 3 was removed from the code beyond level 2, so as to avoid the difficulty of determining if it was operating as a thermal or a as a reflectance channel. See Appendix B.1 for more details on channel 3a/3b characteristics. It was also decided that night acquisitions could be neglected because they were of no use. It was at this stage that while simulating the output of the chain with no intersatellite calibration (it had not been yet calculated for any of the sensors of study) it was noticed that AVHRR15 passing hours were too early in the morning or to late in the evening. Figure 2.2.1 shows the solar zenith angles of AVHRR15 compared to those of AVHRR17. AVHRR16 solar zenith angles are similar to those of AVHRR17 (see Appendix B.1 for details on satellites’ passing hours). It was therefore decided to continue the study with AVHRR16 and 17.

2.2.1

Intersatellite calibration

The intercalibration algorithm discussed in Section 2.1.2 is central to Cyclopes’ utility. Data from different satellites must be brought to a common reference in order to supply a continuous set of acquisitions. It is for that purpose that the enhanced algorithm of intercalibration developed by the CNES has to be tested and the performance ascertained.

2.2. ACCOMPLISHED TASKS

28

Figure 2.2.1: Solar zenith angles of AVHRR15 (thick line) and AVHRR17 (fine line). Computed from the equator crossing time directly retrieved from the orbit’s information. The red dashed line shows the limit of daytime.

Overview The process of intercalibration is accomplished in four stages: Extraction: Data are extracted directly from GAC 1b AVHRR files, searching in each orbit the pixels that belong to one of the sites and storing them in a different file. For each site, 15x15 pixels are extracted taking as the center the middle of the site. All desert sites are larger than 100x100 km2 while the resolution of AVHRR GAC products is of 4km at nadir. It is thus assured that the approximately 60x60 km2 have an homogeneous behavior. Data processing: The extracted files are filtered to remove cloudy observations and the mean of all pixels belonging to the same orbit and the same site is calculated. Reformatting: Data are reformatted in order to comply with the standards of Sade database. Sade database: Data are inserted in the Sade database, developed by CNES and only available there. Sade enforces the expolinear fit and produces as output the coefficients of the model. In order to do so, all output files from the reformatting stage must be transferred by means of ftp to a UNIX machine in CNES.

Implementation A first process of data conversion from the GAC format to binary files is carried out for latitudes, longitudes, acquisition angles and the channels themselves. Then latitude and

2.2. ACCOMPLISHED TASKS

29

longitude files are explored for a set of points that contain one of the sites in their interior. GAC files contain only one longitude and latitude point per each 8, beginning at number 5. The program reads by pairs of two longitude and latitude data thus reconstructing at each iteration a polygon enclosing a certain portion of the viewed scene. The algorithm used to determine whether a polygon contains the center of a desert site or not is the following: we test each site by considering a horizontal ray emanating from it to the right. If the number of times this ray intersects the line segments making up the polygon is even then the point is outside the polygon. Whereas if the number of intersections is odd then the point lies inside the polygon (see Figure 2.2.2 for some examples). While testing the algorithm for several orbits, it was Figure 2.2.2: Examples perceived that it did not take into consideration the time of determining whether a change at 180◦ of longitude. Accordingly, the orbits that point lies within a polypassed through that zone suffered from a change of sign in gon. Source [21]. longitude (two successive points could be at 180◦ and -180◦ ) that was misinterpreted and caused false detections. That was the case of AVHRR16, because it had been downloaded with a global coverage at [12]. On the contrary, AVHRR17 was downloaded specifically for the purpose of the internship, and therefore only the concerned zone of study was recovered. In order to avoid further problems, a new constraint was added to remove those scan lines that passed through the time change line. Note that each orbit covers a daytime and nighttime observation of opposite zones of the globe. Therefore it may be that orbits passing at night by the time change zone contain information about the sites. Another important feature of the extraction procedure is that of screening cloudy pixels. As explained in 2.1.2, thermal channels are calibrated following NOAA’s official calibration. Once temperatures are obtained, a superior threshold is imposed on channel 1 reflectance and an inferior one on the temperature calculated from channels 4 and 5. The set of programs available to enforce the whole process were those used for AVHRR14. They consisted in binary C files, scripts KSH, two MATLAB files and some necessary data files (e.g., table with the latitude and longitude of the sites or meteorological files). The scripts had to be executed manually to enchain the different parts of the process. More information about the original programs can be found at [22]. The adaptation of available program files to those needed to process AVHRR16 and 17 required many changes because data set structures differ between NOAA KLM (i.e., 15, 16 and 17) and the previous ones. That is, byte ordering of header and data records may change substantially (see [9], Section 8.3.1.1 for more details). At the same time, an effort was made to compact as much as possible all the proceedings so that the process was as easy to launch as possible for future users. An internal documentation paper was written for MEDIAS-France. Nevertheless, due to the fact that MEDIAS-France does not have a MATLAB license (IDL is used instead) many file transfers between CNES and MEDIAS-France remained necessary. It is important to bear in mind that LINUX systems (those at MEDIAS-France) and UNIX ones (those at CNES) have different byte orderings: little endian for LINUX and big endian for UNIX.

2.2. ACCOMPLISHED TASKS

30

Figure 2.2.3: Near infrared and red (visible) channel profiles for a site located in the desert of Mali. The left plot shows a well calibrated set and the right plot a badly calibrated one.

Results’ analysis Many iterations were necessary to reach a reliable calibration. Indeed, misunderstandings about data record coding, about the expolinear function or the fact of not having implemented a cloud screening method caused the images to be wrongly calibrated. In order to test the performance of each set of calibration coefficients, desert sites were extracted from level 3a images (i.e., syntheses of reflectances after all corrections applied) and the profiles of channel 1 (red) and channel 2 (near infrared) plotted. It was done in this manner because deserts have the property of being very stable in time and in space, in addition of having feeble directional effects. Thus, in desert sites it is mainly the calibration that determines the measured value. Figure 2.2.3 shows the results for well and wrong calibrated data sets. Figures 2.2.4 and 2.2.5 show the final results for the intercalibration of AVHRR16 and AVHRR17. The exponential term of the model was deemed to be non reliable because of the too short period taken into account. In other terms, we cannot predict the time evolution in a long time scale by only processing four months. For example, AVHRR16 was launched four years ago and is still in service. To obtain a better fit of the variation of its calibration along time, a suitable procedure would be to take a time span of three years and extract data of one month out of each consecutive three. Thus more spaced acquisitions would be available. Finally, the coefficients used were those of the constant term. That is, α + β in Equation 2.10, which in the case of study are:

AVHRR16 AVHRR17

Channel 1 1.0174 1.08

Channel 2 0.9475 0.918

2.2. ACCOMPLISHED TASKS

31

Figure 2.2.4: The green lines display the expolinear fit for AVHRR16 red and infrared channels. Light blue circles show the averaged coefficient for each 10 days.

Figure 2.2.5: The Blue dashed lines display the expolinear fit for AVHRR17 red and infrared channels. Light blue circles show the singular values for each day and site extracted.

2.2. ACCOMPLISHED TASKS

2.2.2

32

Validation of navigation’s accuracy

Overview AVHRR GAC 1b files provide a set of digital numbers that are associated to a certain time of acquisition. Each time determines a definite position in the satellite track and also a concrete viewing geometry. Therefore data may be projected into a physical location. As said in Section 2.1.2, Navigate undertakes that task in the Cyclopes processing chain. That is to say that it navigates the data. To have all pixels of an image at 4km of resolution shifted one column to the right may not be very important when searching for zonal trends. On the contrary, it is crucial when comparing products acquired by two different satellites. Pixel mismatching alters quality assessment results, because pixels are compared to others that do not correspond to the same spot. It may be worse if the two satellites have discrepancies in the sense of the shifting. It is therefore necessary to evaluate the quality of the navigation procedure before continuing with the processing. The maximum shift acceptable at a given resolution is of ± 0.5 pixels, to thus assure a correct data location.

Implementation CNES has developed a set of computing tools called Medicis that calculates the pixel shifting between images. For each two matching images, Medicis takes one as reference and begins by dividing it in small vignettes. Figure 2.2.6 shows a sketch of the proceedings. The spacing and the number of vignettes is given as an input parameter. For each vignette in the image of reference a zone of research is defined in the second image and a measure of resemblance is systematically calculated for all points belonging to that zone. The objective is to determine the placement of the equivalent vignette in the secondary image. Two resemblance criteria are given as a choice by Medicis: one based on linear correlations and a second one based on histogram estimates (algorithms derived from the f-Divergence family). The former is the one chosen in the case of our study. It basically consist in the search of the correlation peak , prior application of two thresholds to remove those correlation values that are too low to be deemed valid results. The first threshold is applied to the single correlation coefficient of the two central pixels, and the second one to the correlation obtained for the whole vignette. By this procedure the corresponding vignette in the secondary image is determined. Two shift values are assigned per vignette: one concerning the vertical shift (lines) and a second one the horizontal shift (rows). Medicis computes the mean shift value for lines and rows of all vignettes of the image. For more details about Medicis’ programming characteristics see [23] for through technical information. To estimate the pixel shift in the case of our study, AVHRR16 and 17 reflectances were extracted after the first step of stage 5 of Cyclopes. That is, navigated images of orbits that have been corrected (atmospheric corrections, cloud screening) and tiled over the zone of study. The resolution is the original one for GAC data: 4km. A level 3 VEGETATION synthesis of that period was taken as image of reference, degraded to 4km of resolution in order to match that of AVHRR images.

2.2. ACCOMPLISHED TASKS

33

Figure 2.2.6: Procedure of comparison between two images used to estimate the pixel shift.

The VEGETATION sensor onboard SPOT 4 and 5 is considered to be much more reliable for data location than AVHRR. That is so because VEGETATION products have an original resolution of 1km instead of 4km as AVHRR product. Moreover, unlike AVHRR, VEGETATION was conceived as a vegetation monitoring device. AVHRR on the contrary was designed for meteorological applications. See Appendix B.2 for more information about VEGETATION characteristics. The comparison is enforced between VEGETATION’s B2 channel and AVHRR’s channel 1. Usually, infrared images are chosen for that type of comparisons because they are more sensitive to landscape singularities and thus easier to find the correspondences between them. Nevertheless, the visible channels were preferred in the case of study because they have a closer spectral response for AVHRR and VEGETATION.

Results’ Analysis Figures 2.2.7 and 2.2.8 show the first results obtained. Shifts are bounded as follows:

[−1.6, 1.2] rows in AVHRR16 [0.75, 1.75] lines in AVHRR16 [0.75, 2.75] rows in AVHRR17 [−0.75, −1.75] lines in AVHRR17

It is obvious that shift differences between the two satellites are too important to continue the processing at the same resolution. Indeed, both satellites have different behaviors, so the direct comparison between the two of them would produce misleading results. Moreover, important variations in magnitude and sense of the shift occur for

2.2. ACCOMPLISHED TASKS

34

Figure 2.2.7: Mean shift values for columns and lines of AVHRR16 data. For each image it is calculated one mean shift value for lines and one for rows. Only days with more than 20 valid vignettes are taken into account. When more than one image is available per day, only the values of that which had more valid vignettes are considered.

Figure 2.2.8: Same as Figure 2.2.7 but for AVHRR17 data. The upper line corresponds to AVHRR17 shifts in rows and the lower one to its shift values in lines.

2.2. ACCOMPLISHED TASKS

35

Figure 2.2.9: Mean shift values for AVHRR16 data as in Figure 2.2.7. Instead of being plotted against the day of acquisition data is plotted against the orbit’s beginning hour.

the same satellite and between short periods of time. That would cause Cyclopes to mix in the compositing phase data belonging to different pixels. Change to a coarser resolution would mean to degrade images to 16km. That is, to divide by four the magnitude of the observed shifts. 8Km resolution wouldn’t be enough to assure shift values of about ± 0.5 pixels. The abrupt periodicity seen in Figures 2.2.7 and 2.2.8 is in fact the mark of a relation between the passing hour of the satellite and the measured shift. It was so inferred from the set of continuous maximum peaks. They correspond to opposite geometries of acquisition, i.e., when the satellite sees the same spot from one side or the other. Figure 2.2.11 shows an example of two opposite satellite swaths. Therefore, same data can be displayed also as in Figures 2.2.9 and 2.2.10. Indeed, they prove that there is a strong relation between satellite passing hour and shift.

Figure 2.2.10: Same as Figure 2.2.9 but for AVHRR17 data.

2.2. ACCOMPLISHED TASKS

36

Figure 2.2.11: Two orbits that see the zone of study from different sides. Pixels in the west/center of the zone are seen from the two viewing geometries.

At the same time, different hours of observation mean changes in view zenith angles (VZA). This approach suggests that there may be a relation between VZAs and measured shift. To prove it, shift values must be computed over specific pixels instead of over the whole image. Thus, a value of VZA can be associated to each pixel, which would be impossible to do for a whole image because VZAs vary for each pixel. In order to plot the shift/VZA graph, a set of landmarks was chosen so as to be representative of the whole zone of study and yet be all placed in salient geographical spots. Figures 2.2.12 and 2.2.13 show the results. Unlike the graphs shown before they do not account for a mean value per day but for a set of individual values for the whole period. Therefore variability is much higher. However, it can be stated that except for AVHRR16 row’s shift all other graphs display clear biases.

Figure 2.2.12: Shift values plotted against VZAs for AVHRR16 data. Negative angles correspond to the same VZA as positive ones but with a difference of about 180◦ in azimuth.

2.2. ACCOMPLISHED TASKS

37

Figure 2.2.13: Same as Figure 2.2.12 but for AVHRR17 data.

Figure 2.2.14 gives more information about the placement of landmarks for AVHRR16 row’s shift. It is thus shown that only landmarks belonging to the west side of the zone of study have been seen by the satellite from the two different sides. That is, from view azimuth angles differing 180◦ (signaled in the graphs with a change of sign). Therefore it may be considered that fit to be the most reliable one along with that of Figure 2.2.12 for shifts in AVHRR16 rows.

Figure 2.2.14: Same as Figure 2.2.12 but displaying in different colors shift measures in the west, center or east of the zone of study.

Biases and in general trends relating VZA s and shifts can usually be explained by wrong assumptions of satellite pointing. For instance, we may suppose in the navigation program that the sensor points straight down and then sweeps to the sides, while it

2.2. ACCOMPLISHED TASKS

38

may be slightly tilted to the front. That would cause each scanline to be seen before it actually was computed to be and thus it would be wrongly assigned. Satellite attitudes are used to measure pointing errors following the three axes. Appendix C reviews the concepts of pitch, roll and yaw. Moreover, it shows that when dealing with small angles pitch accounts for constant shifts in lines, roll for constant shifts in rows and yaw for linear trends in line shifts. From that point of view, Figures 2.2.12 and 2.2.13 can be corrected by applying a certain pitch and roll angle so as to remove the biases. As for the linear trend in AVHRR16’s rows, no combination of pitch/roll/yaw angles can explain it and at the same time be consistent with the shift bias observed for AVHRR16’s lines. Concerning these data location errors, it was found a memorandum in NOAA’s website on the subject of along track errors for NOAA-16. A pitch type attitude correction of -0.44◦ was suggested. A max scan angle of ± 55.25◦ instead of ± 55.37◦ was also counseled. The max scan angle is the maximum viewing angle allowed from the satellite. It is closely related to the view zenith angle. On the other hand, a simulator of attitudes was developed in IDL language. The program reproduces a satellite placed above the the equator, at the same altitude as NOAA16 and 17. The satellite points straight down and then sweeps to the left and to the right n spots, which limits the scan angle to ± (∆ θscan · Nspots ). By defining the number of steps Nspots and the angle ∆ θscan of each step, the max scan angle is set. The nomenclature mimics that used in Cyclopes. In Appendix C.1 there’s more information concerning the exact formulae as well as the conventions followed. Tests were carried out first to AVHRR16 data because there was NOAA’s memorandum about it. The testing method was the same as explained in 2.2.2 and results were checked directly in the graphs displaying the mean values of images along time. Those obtained only with the corrections suggested by NOAA show that they are not adapted to our zone and period of study (see Figure 2.2.15). However, the max scan angle limitation was kept as it distinctly improved the results.

Figure 2.2.15: Different set of corrections applied to the graphs of mean shift values along time for AVHRR16.

2.2. ACCOMPLISHED TASKS

39

Figure 2.2.16: Same as Figure 2.2.15 but for AVHRR17.

Next, it was determined with the simulator of attitudes the pitch and roll values that should correct Figures 2.2.12 and 2.2.13. The fact of not having a complete set of observations (one for each VZA) in those graphs forced the exact values to be adjusted by hand. Figures 2.2.15 and 2.2.16 show the whole set of corrections and Figures 2.2.17 and 2.2.18 the final results given by the simulator of attitudes. Remark that the simulator results are given in function of the scan angle instead of the view zenith angle. The best correction is that of AVHRR17 shift in rows. Indeed, it is the one which has a better match between Figures 2.2.12 and 2.2.18. Notwithstanding, Figures 2.2.19 and 2.2.20 show that all shifts are corrected to 3/4 of pixel. The final pitch and roll values in radians are:

AVHRR16 AVHRR17

roll -0.001264503 -0.003338278

pitch 0.005420579 0.006409866

Figure 2.2.17: Simulation with IDL of the theoretical corrections for AVHRR16 rendered by the final values of attitudes chosen.

2.2. ACCOMPLISHED TASKS

Figure 2.2.18: Same as Figure 2.2.17 but for AVHRR17.

Figure 2.2.19: Final shift results at 4km of resolution for AVHRR16.

Figure 2.2.20: Final shift results at 4km of resolution for AVHRR17.

40

2.3. EVALUATION OF THE RESULTS

41

In conclusion, results allow us to change to a resolution of 8km and be sure that navigation is correct enough to continue with the processing. CESBIO agrees in degrading the images to 8km, as it was initially 8km the planned resolution.

2.2.3

Production of output data sets

All previous tasks were necessary in order to assure the correct functioning of the processing chain and the validity of the results. The coefficients of intersatellite calibration were introduced in Cyclopes. To take also into account the correction of attitudes, Navigate has the possibility to specify in a file the pitch, roll and yaw values to apply. These values are treated as constants and thus enforced in all orbit files, which means that all images belonging to the same satellite are corrected with the same attitudes. The images used to study the quality of the final products were level 3a synthesis, i.e., surface reflectances normalized to a common viewing geometry. The hour of normalisation was chosen to be 10.30h. As shown in Appendix B.1, AVHRR17 crosses the equator at about 10:00, so it sees the zone of study at about the hour of normalization. On the other hand, AVHRR16 sees it later, at about 14:30. It was preferred to normalize to the passing hour of one of the two, thus only extrapolating with the BRDF the reflectances of the other One synthesis is produced each 10 days, each of them being a composition of reflectances measured 15 days before and after. That amounts to a total of 12 syntheses for the whole period of November 2002- February 2003. Four sets of images were produced: two for AVHRR16 at 8km and 16km of resolution, and two more for AVHRR17.

2.3

Evaluation of the results

The final objective of the internship is to determine whether the new algorithms implemented in Cyclopes are capable of bringing data from AVHRR16 and AVHRR17 to a same reference. Theoretically, two sets of images acquired by AVHRR16 and AVHRR17 in the same day should produce the same results but for the errors they introduce, because they are seeing the same scene. Two syntheses covering the same period should therefore render equal images if the corrections were perfect. Differences between syntheses are in consequence due to artifacts not rectified by the processing method. It is thus possible to ascertain the quality of the corrections.

Quality Criteria In [11] the concept of Normalized Reflectance Difference (NRD) is developed: N RD = 2

ρ16 − ρ17 ρ16 + ρ17

(2.15)

The difference between corresponding reflectances of AVHRR16 and 17 channels is normalized by the mean value of them. Equation 2.15 can be also written switching ρ16 and ρ17 : it is not the sign that matters but the magnitude and the distribution of the anomalies.

2.3. EVALUATION OF THE RESULTS

42

The NRD is calculated in a pixel per pixel basis, i.e., two synthesis of the same day are processed to obtain the NRD for each pixel. The calculus is repeated for all dates of synthesis and for channel 1 and 2. The mean value of each date is computed in order to obtain the profile of channel 1 and channel 2 evolution with time. This is a good estimate of the degree of intersatellite calibration, since mean values over images should reveal biases between satellites. The objective is to correct the abrupt drops seen in Figure 2 (b). Assuming that errors affecting AVHRR16 and AVHRR17 data sets are statistically independent [11], the standard deviation of noise can be stated as: σN RD σN OISE = √ 2

(2.16)

Again, σN OISE can be computed for each date and for channels 1 and 2. The standard deviation provides an estimate of the magnitude of artifact trends in syntheses, thus fixing the threshold of reliable detectable trends. Mean NRD and σN OISE are therefore the two criteria chosen to evaluate the quality of channel 1 and 2 reflectances. However, CESBIO is mainly interested in the performance of Cyclopes for NDVI images. NRD may display unrealistic high values when applied to the NDVI of desert zones. That is so because NRD normalizes by the mean value, which is almost 0 in deserts for the NDVI. The preponderance of deserts in the zone of study makes it necessary to define another set of estimators for NDVI images. Instead of normalizing the difference it is computed directly: dif fN DV I = N DV I16 − N DV I17 and: σnoiseN DV I =

σdif fN DV I √ 2

(2.17)

(2.18)

Results and discussion For all graphs obtained, it was remarked that syntheses at 8km and 16km of resolution rendered nearly the same results (see Figure 2.3.2). The trade-off between resolution and quality does not justify therefore the choice of the coarser resolution. Thus, all figures thereafter present only the results for a resolution of 8km. Reflectances are examined first because it is from them that the NDVI and other biophysical products are obtained. In order to determine whether the results are good enough or need to be improved they are compared to those obtained in [11] for VEGETATION products. It is taken into account the fact that VEGETATION is a much more precise sensor when used for vegetation monitoring, and thus it is expected to obtain worse results for AVHRR images. Furthermore, results are also checked so that computed NDVIs reach the level required by posterior long-term trend analysis. Indeed, a variation of about 0.05 NDVI units for a period of 20 years would be typically expected, which forces noise levels to be inferior. For interannual trends the same magnitude of NDVI variation is awaited. Figure 2.3.1 shows the profiles of two desert sites and two other ones corresponding to zones covered with vegetation. They ascertain the good precision of intersatellite calibration.

2.3. EVALUATION OF THE RESULTS

43

Figure 2.3.1: Evolution of channel 1 (visible) and channel 2 (infrared) for AVHRR16 and 17 in two desert sites.

Next, Figure 2.3.2 displays the results for mean values of NRD over the zone of study. In [11] it is taken as a criterion for VEGETATION images to keep differences below 5%, which in our case is achieved except for one synthesis of channel 1.

Figure 2.3.2: Profile of mean values of NRD for channel 1 (left) and channel 2 (right). Light blue displays results at 8km resolution and dark blue at 16km.

As for values of σN OISE , Figure 2.3.3 presents the profiles for both channels. Channel 1 displays percentages that are clearly too high while those of channel 2 are in the limit of acceptance. Furthermore, Figure 2.3.8 shows that they do not reach the level of correction needed for the NDVIs, as the noise level is equal to that of expected trends. The explanation of the bad results obtained for channel 1 can be found directly in the syntheses. Figure 2.3.4 shows two syntheses of the zone for AVHRR16 and AVHRR17. Although they correspond to the same date, AVHRR17’s image displays clouds not screened in the strip of land by the coast. On the contrary, AVHRR16’s image has no trace of clouds. When the NRD is computed, both images will display very high differences in that zone. Moreover, σN OISE will be more sensitive to those

2.3. EVALUATION OF THE RESULTS

44

Figure 2.3.3: Profile of the standard deviation of NRD for channel 1 (above) and channel 2 (below).

discrepancies, since the standard deviation has a quadratic development that magnifies the weight of anomalies. Same problems are found in channel 2 but with a lower degree of importance. The issue of clouds in the southern coast of Sahel had already been arisen on previous studies of the same zone made by CESBIO. It was thus decided from the beginning of the internship that it was better to obtain reliable images by imposing stricter cloud screening parameters than to produced cloud distorted ones. Indeed, cloudy acquisitions produce unrealistic high reflectances that are then processed together with normal ones in the compositing stage of Cyclopes, thus altering the fit of the BRDF. Even if it is

Figure 2.3.4: Channel 1 syntheses for AVHRR16 (left) and AVHRR17 (center) corresponding to the 5th of January 2003. Bright pixels display high values of reflectances, due to low levels of vegetation (deserts) or clouds (as for instance the southern strip of land). At the right the cloud mask developed a posteriori for that date.

2.3. EVALUATION OF THE RESULTS

45

Figure 2.3.5: Profile of the standard deviation of NRD for channel 1 (right) and channel 2 (left) enhanced with the application of a cloud mask a posteriori.

Figure 2.3.6: Profile of mean values of NRD for channel 1 (left) and channel 2 (right) enhanced with the application of a cloud mask a posteriori.

at the cost of lower percentages of valid pixels, it is preferable not to take into account cloudy pixels. To ascertain the real statistics that could be achieved with a better cloud screening procedure, a cloud mask was built directly from reflectances obtained in the syntheses. A maximum threshold is imposed on differences between channel 1 and 2, so that pixels that differed more than 200 units (reflectances are coded from 0-2000) are deemed cloudy. Afterwards the mask is dilated. An example of a mask can be seen in Figure 2.3.4. Figure 2.3.6 shows the results obtained with the masking. The improvement is clear yet there is also a pixel loss of about 10%. These results are in concordance with those obtained in [11] for VEGETATION. Furthermore, replotting of the NRD’s profiles with the application of the cloud mask (Figure 2.3.5) improves the ones obtained in Figure 2.3.2. It is thus possible to assure that differences are kept below 5%, which means that it fullfils the same criterion imposed on VEGETATION products in [11].

2.3. EVALUATION OF THE RESULTS

46

Concerning the results for the absolute NDVI difference, Figure 2.3.7 shows their profiles without masking as well as with masking. The mean value of the absolute NDVI for the whole period is worsened by the application of the mask. In fact, variability is reduced yet the mean value is higher in magnitude since syntheses without cloud masking oscillate around 0. Nevertheless, results are better than those that can be seen in Figure 2 (b) for the GIMMS and PAL datasets. They display drops between succeeding satellites that may go as far as 0.02 and 0.03 NDVI units, even though they are restricted to a desert zone in the Arabian desert. In our case of study, differences are always below 0.02 NDVI units and have a mean value of 0.01.

Figure 2.3.7: Profile of differences of NDVIs without masking (left) and with the application of a mask (right) .

As for the standard deviations of the absolute NDVI, Figures 2.3.8 prove the necessity of applying the cloud mask. With the masking, result’s mean value is around 0.02 NDVI units. It can be thus assured that a trend of 0.05 NDVI units cannot be caused by artifact trends.

Figure 2.3.8: Profile of the standard deviation of differences of NDVIs without masking (left) and with the application of a mask (right) .

Chapter 3 Conclusion 3.1

Perspectives

The objective of the internship was to estimate the level of correction that Cyclopes could achieve. For that purpose a case of study was proposed from which some conclusions can be deduced. However, the task of determining the degree of representability of this case in the scope of the whole AVHRR archive is much more complex. Concerning the case of study, it has been proved that the performance of Cyclopes for AVHRR data may match that achieved for VEGETATION data. That is to say, AVHRR data can be corrected so as to approximate the level of reliability (at its own resolution) of a sensor originally designed for vegetation monitoring. Bearing in mind the extension of the AVHRR archive, it is a promising advancement and a good indicator of the potential of Cyclopes. It can be also ascertained that for the given period of time and the satellites chosen, biases between data acquired by different satellites can be reduced. Also, the level of artifact trends can be assured to be below that of expected trends. Nevertheless, it is important to remark that final results are greatly dependent on the precision of cloud screening. A method of cloud masking a posteriori has been used to reach the presented results because the one implemented by Cyclopes has proved to be insufficient for the zone of study. It is therefore necessary to revise the actual method and enhance it for future studies. If a better cloud screening can be achieved, results may easily be better since the method a posteriori is only partial and does not prevent many cloudy pixels to be used in the fit of the BRDF. It should also be considered when dealing with longer time periods a more automatic procedure for the determination of attitudes. Moreover, it may be necessary to apply attitudes not constant through time when processing satellites with several years of life span. This problem should be tackled before undertaking a comprehensive processing of the AVHRR archive. Concerning the perspectives of Cyclopes, it is yet necessary to study the representability of the case chosen, i.e., it must be ascertained if it can be generalized spatially and temporally. To extend the study to a bigger time period (e.g. 20 years) would mean to process data from more than two AVHRR sensors. As the resulting reflectances are intercalibrated, the performances should be equal for other satellites in the same conditions, i.e., that have passing hours that differ a maximum of 4h and the same sensor with the same

3.2. PERSONAL BALANCE

48

calibration. All AVHRR have similar characteristics (see Figure 1 for different spectral responses of channel 1 and 2), apart from singular events that may have caused errors for a given sensor during a short time period. However, as it is shown in Appendix B.1 , passing hours vary. Only those corresponding to daytime must be taken into account. In summary, NOAA-17 is the one that passes earlier (10:00) and NOAA-7 the one that does later (15:00), 30 minutes after NOAA-16. Therefore the case of study is quite representative of the worse differences of passing hours of the whole archive. On the other hand, satellites prior to NOAA-15 shift their equator crossing time (see Figure 2 (a)). This phenomenon was not present in our case of study. Theoretically, the intersatellite calibration variation along time should correct the effects of the drift, yet the extent of the real achieved correction remains untested. Considering the application of Cyclopes to other geographical zones, the Sahel is not a specially favorable area. As seen in Figure 2.3.4, the southern strip of coast is often cloudy, more than what is normal in other latitudes. If zones of study of the same size are processed separately, atmospheric corrections should at least have the same performance (except zones with ice and other particular phenomena). Moreover, results should be enhanced if a better cloud screening is achieved. In conclusion, it is not yet possible to give a definite response to whether Cyclopes will improve the quality of available AVHRR data sets. However, the results exposed in this paper prove that for the case of study it does so which is a prior step to prove it for other cases. Moreover, the accomplished tasks were necessary not only to reach the results but also to establish a precedent of the necessary steps to undertake for each new satellite to process. Now, all elements necessary to take the decision of whether going further or not are provided: the expected results, the time necessary to integrate each new satellite and those parts of the processing that should be enhanced in order to improve the results.

3.2

Personal balance

The internship was well planned in advance so the objectives and the procedure were clear from the beginning. Furthermore, I always felt the internship was of strategic importance for MEDIAS-France, the firm were 5 out of 6 months of the internship took place. Accordingly, its development was closely followed and the results taken into account and used in discussions where the future continuation of that line of development was decided on the affirmative. The organisation between the three involved firms was fluid and prompt. Except a first phase of documentation at CESBIO, the rest of the internship was developed at MEDIAS-France. Some processes needed to be launched at CNES. Meetings between the three parts took place regularly, mainly to inform and discuss with Laurent Kergoat and Eric Mougin of CESBIO the advancement of the internship. They provided the advice and the point of view of users of biophysical products, i.e., of scientific researchers that use images of the the Sahel to monitor vegetation changes. At MEDIAS-France I worked in close contact with the group POSTEL, specially with Bastien Miras and Patrice Bicheron. Olivier Hagolle from CNES was the person who guided the internship, since his department had developed all the algorithms of Cyclopes and he had already realized a similar study for VEGETATION instead of

3.2. PERSONAL BALANCE

49

AVHRR’s products. We were in continuous contact through e-mail and telephone, apart from some days I went to CNES or he came to MEDIAS-France. In brief, the internship led me to work with very different business cultures. Mostly, all the computing skills of MEDIAS-France and the scientific savoir-faire of CESBIO were new to me and so were their approaches of the subject. The formation and procedure of Olivier Hagolle from CNES was more in concordance with the one I had received in T´el´ecom Paris. I learned a lot about Linux/Unix environments because MEDIAS-France is engaged in the use of open source software. Furthermore, it was often necessary to go into the code in order to modify the programs, that were of a complexity proportional to the huge amount of data to be processed. That pushed me to make an effort on computing science, since all programs that I had used before didn’t have to account for the engineering of such a big ensemble of routines. Thus I got acquainted with KSH and PYTHON scripts, IDL, envi as well as further my knowledge of C and C++. On the other hand, the questions which CESBIO wanted to be answered at the end of the stage were asked from the point of view of a scientific researcher. That forced me to pass from a level of detail lower than that which I was used to (computer science engineering) to a higher and more specific one. Thus in the phase of documentation I read scientific papers on the subject of vegetation monitoring, global temperature change and so on. I reviewed them during the writing of the paper in order to reflect their questions and try to answer them as much as the results allowed me to do. It is therefore possible that this paper gives a wrong impression of how I spent my time. Indeed, during the internship most time was devoted to computer issues and discussions about the algorithms and principles that were behind them. However, the writing of the paper has been also an important part of the apprenticeship. It has forced me to go a little above the typical technical report and try to approach the issues that could be of interest for scientific researchers and research project managers. To close, I will add that this internship has given me the confidence to believe that I’m prepared to work on any subject that arises my curiosity and interest.

Appendix A Symbols and acronyms AMMA AVHRR BRDF CLAVR CYCLOPES FAPAR GAC GIMMS LAC LAI NDVI NOAA NRD PAL POES POLDER POSTEL SMAC SZA TOA VEGETATION VZA

African Monsoon Multidisciplinary Analyses Avanced Very High Resolution Radiometer Bi-directional Reflectance Distribution Function CLouds for AVHRR Carbon cYcle and Change in Land Observational Products from an Ensemble of Satellites Fraction of Obsorved Photosynthetically Active Radiation Global Area Coverage Global Inventory Monitoring and Modeling Systems Local Area Coverage Leaf Area Index Normalized Difference Vegetation Index. ρ raRed −ρRed N DV I = ρInf Inf raRed +ρRed National Ocean and Atmosphere Administration Normalized Reflectance Difference Pathfinder AVHRR Land Polar Orbiting satellitES POLarisation and Directionality of the Earth Reflectance Pˆole d’ Observation des Surfaces Terrestres aux Echelles Larges Simplified Method for Atmospheric Correction Solar Zenith Angle Top Of Atmosphere The medium resolution sensor onboard SPOT4 and SPOT5 satellites View Zenith Angle

Appendix B Sensors’ Description B.1

AVHRR

The Advanced Very High Resolution Radiometer (AVHRR) sensor is an optical instrument. It provides multi-spectral imaging by sensing reflected sunlight and thermal emissions. The AVHRR sensor is nominally a five channel scanning sensor. Band 1 is the visible band, Band 2 is the near infrared band, and Bands 3, 4 and 5 are the thermal bands. Table B.1 indicates the differences of wavelength on each band for each satellite.

Acronym 1

NOAA Satellite 7, 9, 11, 14 (µm) 0.58-0.68

NOAA Satellite 15, 16, 17 (µm) 0.58-0.68

2

0.725-1.10

0.725-1.10

3A 3B

3.55-3.93

1.58-1.64 3.55-3.93

4

10.30-11.30

10.30-11.30

5

11.50-12.50

11.50-12.50

Potential Applications Daytime cloud surface mapping, vegetation Daytime cloud surface mapping, vegetation Vegetation Sea/land night surface temperature, nighttime cloud mapping Sea/land day and night surface temperature, nighttime cloud mapping Sea/land day and night surface temperature, nighttime cloud mapping

Table B.1: Wavelengths and applications of AVHRR bands.

B.1. AVHRR

52

Several sensors have been launched for 20 years.Table B.2 gives the temporal coverage since TIROS-N.

Satellite number N 6 7 8 9 10 11 12 14 15 16 17

Launch date 0/13/1978 06/27/1979 06/23/1981 03/28/1983 12/12/1984 09/17/1986 09/24/1988 05/13/1991 12/30/1994 05/13/1998 09/21/2000 06/24/2002

Ascending node 15:00 19:30 14:30 19:30 14:20 19:30 13:40 19:30 13:40 19:30 14:00 22:00

Descending node 03:00 07:30 02:30 07:30 02:20 07:30 01:40 07:30 01:40 07:30 02:00 10:00

Service dates 10/19/1978-01/30/1980 06/27/1979-11/16/1986 08/24/1981-06/07/1986 05/03/1983-10/31/1985 02/25/1985-05/11/1994 11/17/1986-Present 11/08/1988-09/13/1994 05/14/1991-12/15/1994 12/30/1994-Present 05/13/1998-Present 09/21/2000-Present 06/24/2002-Present

Table B.2: Service dates and equator crossing times of AVHRR series [24].

The sensor has a small field of view, scanning across the earth by the continuous 360 degree rotation of a flat scanning mirror. All the spectral channels are registered so that they all measure energy from the same spot on the earth at the same time. AVHRR data are broadcasted continually as well as tape-recorded onboard the spacecraft for readout at a NOAA receiving centre. The LAC and GAC forms of transmission and area coverage are explained below (HRPT and APT are not described): LAC (Local Area Coverage) LAC is nominally 1 km resolution AVHRR imagery recorded with the on board tape recorder for subsequent transmission during the overpass of a station controlled by NOAA. Owing to the large number of data bits, only about 11+ minutes of LAC can be accommodated on a single recorder. LAC imagery can only be obtained from NOAA/NESDIS and only in their formats. GAC (Global Area Coverage) GAC data is lower resolution (4 km) AVHRR imagery. It is derived on board the NOAA satellite by sub-sampling and averaging the nominal 1 km resolution AVHRR imagery. It provides daily global coverage which is recorded on a satellite tape recorder and then transmitted to a ground station controlled by NOAA. 115 minutes of this lower resolution imagery can be stored on a recorder, enough to cover an entire orbit of data acquisition. GAC imagery can only be obtained from NOAA/NESDIS and only in their formats.

B.2. VEGETATION

53

The ground resolution is approximately 1.1 km at the satellite nadir from the nominal orbit altitude of about 850 km. The width of off-nadir pixels increases from 1.1 km to about 5 km at the most extreme viewing angle at the edge of the 3000 km imaging swath. The orientation of the scan lines is perpendicular to the satellite orbit track and the speed of rotation of the scan mirror is selected so that adjacent scan lines are contiguous at the sub-satellite (nadir) position. The satellite speed and scan mirror rotation rate result in an along track pixel height of about 1.1 km. The LAC spatial resolution is loosely said to be 1 km resolution. For GAC data, out of every 5 normal across track LAC pixels, the 5 bands of the first four are individually averaged and all the bands of the fifth pixel are ignored. In the along track direction, only every third normal line of LAC/HRPT pixels is considered; the two intervening lines are ignored. The GAC spatial resolution is loosely said to 4 km resolution. The AVHRR data used in the CYCLOPES project are in GAC data.

B.2

VEGETATION

Since April 1998, the VEGETATION sensors have been operational on board the SPOT 4 and 5 earth observation satellite system. They provide a global observation of the world on a daily basis. The instrumental concept relies on a linear array of 1728 CCD detectors with a large field of view optic (101◦ ) and four optical spectral bands described in table B.3.

Acronym B0 B2 B3 SWIR

Center (nm) 450 645 835 1665

Width (nm) 40 70 110 170

Potential Applications Vegetation, ocean color Vegetation Vegetation Vegetation

Table B.3: Wavelengths and applications of VEGETATION bands.

The spatial resolution is 1.15km at nadir and presents minimum variations for offnadir observations. The 2200 km swath width implies a maximum off nadir observation angle of about 50.5◦ . About 90% of the equatorial areas are imaged each day, the remaining 10% being imaged the next day. For latitudes higher than 35% (North and South), all regions are acquired at least once a day. The multi-temporal registration is about 300 meters.

B.3

POLDER

The POLDER instrument is a radiometer designed to measure the directionality and polarization of the sunlight scattered by the ground plus atmosphere system. The instrument is made of bi-dimensional CCD matrix, a rotating wheel that carries filters and polarizers, and a wide field of view lens (114◦ ). The field of view seen by the CCD matrix is ±43◦ along track and ±5◦ across track. The view zenith angles seen at surface

B.3. POLDER

54

level are larger due to Earth curvature, ±50◦ along track and ±61◦ across track (±70◦ in the matrix diagonal). The pixel size on the ground is about 6 km at nadir for the ADEOS altitude of 800 km. It degrades slightly with view angles by 21% for a view angle of 60◦ . The rotating wheel carries polarized filters. The most important ones are listed in table B.4.

Acronym 443P

Center (nm) 444.5

Width (nm) 20

443NP 490NP 565NP 670P

444.9 492.2 564.5 670.2

20 20 20 20

763NP 765NP

763.3 763.1

10 40

865P

907.7

20

910NP

860.8

40

Potential Applications Vegetation, aerosols, earth radiation budget Ocean colour Ocean colour Ocean colour, vegetation Vegetation, aerosol, earth radiation budget Cloud top pressure Vegetation, aerosols, cloud top pressure Vegetation, aerosols, earth radiation budget Water vapour amount

Table B.4: Wavelengths and applications of POLDER bands.

Images of the same band are acquired every 19.6s, which permits a large overlap between successive images. During the satellite overpass, a surface target is viewed up to 14 times, with a different viewing angle at each time. The directional configuration changes each day due to orbital shift between successive days. Therefore after a few days, assuming favourable atmospheric conditions, the slices of measurements provide a sampling of the BRDF in the limits of the instrument field of view. Data from different view angles are geometrically registered to within an average accuracy of 0.1 pixels (one single orbit) and 0.2 pixels (different orbits). The multi-temporal registration is 0.4 pixels for multi-temporal registration and 1 pixel for absolute registration.

B.3. POLDER

55

Figure B.3.1: Geometrical sampling of the BRDF available over a site in the desert of Libya for each sensor: top POLDER, center AVHRR and bottom VEGETATION. Angles in the polar coordinates correspond to φrelative , lengths to VZA (marked by different concentric circumferences) and different symbols indicate different SZA classes.

Appendix C Attitudes Satellite attitudes correspond to changes in the pointing direction of the sensor. Mathematically, they are characterized by rotations around the axes of coordinates. They are usually treated as the combination of three rotations, each one corresponding to one of the axes: pitch, roll and yaw. Figure 1 shows the sense of each rotation as well as the effect it has on the acquired images.

Figure 1: Effects of pitch, roll and yaw on the pointing direction and on the viewed scene.

If the pointing direction of the sensor is badly determined, measured radiances are assigned to erroneous locations, thus producing a general shift of pixels in the reconstructed scene. The observed shift varies according to the type of attitude. It is therefore possible to retrace the combination of attitudes that produce a certain shift in a set of images, if the shifts are well determined. In this manner images can be corrected so as to remove the effects of attitudes. Pitch causes the satellite to point forward or backward, thus altering the vertical placement of pixels. An important shift in lines is observed as well a feeble horizontal one. The shift in rows is caused by the roundness of the earth and is negligible except for very important values of pitch, which seldom occur in satellites. It is important to

C.1. SIMULATOR OF ATTITUDES

57

remark that the effects of pitch do vary depending on the scan angle, as they are more important in side viewed spots than in those that are closer to the nadir. Roll has a similar effect than pitch. Instead of altering the vertical placement of pixels, it concerns only the horizontal one. In sum, roll causes a shift of rows in the observed scene which is more important for high scan angles than for low ones. Yaw causes both rows and columns to shift. For small angles the effect it has on the vertical placement of pixels predominates, thus producing a shift in lines. Line shifts caused by yaw are softer than those caused by pitch. Those in rows, although also negligible for small angles, are much more important than those induced by pitch.

C.1

Simulator of attitudes

The program, written in IDL language, simulates a satellite placed on the equator to which different attitudes are applied separately. Figure C.1.1 shows a picture of the system as well as the conventions followed for the axes and the angles.

Figure C.1.1: Sketch of the simulated scene and the conventions of angles and axes.

First, a complete swath of the satellite is simulated without the application of any attitude. In order to do so, each pointing direction is represented by a straight line going from the satellite to the earth. The direction of each line is determined by the θscan . In fact, a series of discrete scan angles is calculated from two parameters: Nspots and ∆ θscan . The latter sets the discrete angular step of each new viewing line and the former the number of steps of a complete sweep to one side. Once the set of θscan is determined, each line of the 3-D model of figure C.1.1 is characterized by the following equation: (x, y, z) = (0, 0, h + R) + k (− sin θscan sin φ, sin θscan cos φ, − cos θscan )

(C.1)

where h is the altitude of the satellite, R the radius of the earth and k the unknown variable to be solved. The equation of the earth is x2 +y 2 +z 2 = R2 , and so the system of equations is completed and can be solved. Each solution corresponds to the intersection

C.1. SIMULATOR OF ATTITUDES

58

of a viewing line and the earth, i.e., the coordinates of the pixel seen for a certain θscan . With this method the coordinates of all pixels of the viewed scene are computed. Secondly, a rotation of the axes is applied to the set of viewing vectors. The following matrices of rotation are used:   1 0 0 Mx =  0 cos γroll sin γroll  (C.2) 0 − sin γroll cos γroll   cos γpitch 0 sin γpitch  0 1 0 My =  (C.3) − sin γpitch 0 cos γpitch   cos γyaw sin γyaw 0 Mz =  − sin γyaw cos γyaw 0  (C.4) 0 0 1 The new vectors correspond to the viewing lines of the satellite under the influence of a certain pitch, roll or yaw angle. Interceptions with the earth are recalculated and so the new coordinates of the pixels are obtained. The difference of these values and those obtained without any attitudes correspond to the shifts values. That is, the difference according to the X axis is the line shift and that according to the Y axis the row shift.

Bibliography [1] Garik G. Gutman. On the use of long-term global data of land reflectances and vegetation indices derived from the advanced very high resolution radiometer. Journal of geophysical research, 104(D6):6241–6255, March 1999. [2] Daniel A. Slayback, Jorge E. Pinzon, Sietse O. Los, and Compton J.Tucker. Northern hemisphere photosynthetic trends 1982-99. Global Change Biology, pages 1–15, 2003. [3] Nikolay V. Shabanov, Liming Zhou, Yuri Knyazikhin, Ranga B. Myneni, and Compton J. Tucker. Analysis of interannual changes in northern vegetation activity observed in avhrr data from 1981 to 1994. IEE Transactions on geoscience and remote sensing, 40(1), January 2002. [4] L. Zhou, R. K. Kaufmann, Y. Tian, R. B. Myneni, and C. J. Tucker. Analysis of interannual changes in northern vegetation activity observed in avhrr data from 1981 to 1994. Journal of geophysical research, 108(D1), 2003. [5] M. E. James and S. N. V. Kalluri. The pathfinder avhrr land data set: An improved coarse-resolution data set for terrestrial monitoring. Int. J. Remote Sensing, 15:3347–3364, 1994. [6] C. J. Tucker. History of the use of avhrr data for land applications. Ed. Dodrecht, 1996. In hte book Advances in the use of of NOAA AVHRR Data for Land Application, by G. D’Souza. [7] Kingwell et al. Technical report of the ceos. Technical report, CEOS, 1996. [8] NOAA Polar Orbiter Data User’s guide. URL: http://www2.ncdc.noaa.gov/docs/podug. [9] NOAA KLM User’s guide. URL: http://www2.ncdc.noaa.gov/docs/klm/. [10] A. de Pereyra. Manuel d’utilisation: maquette d’une chaˆıne de production de LAI et FPAR ` a partir de produits AVHRR GAC 1b et VGT-P. [11] O. Hagolle, A. Lobo, P. Maisongrande, F. Cabot, B. Duchemin, and A. de Pereyra. Quality assessment and improvement of temporally composited products of remotely sensed imagery by combination of vegetation 1 & 2 images. Remote Sensing of Environment, 94(2):172–186, 2004.

BIBLIOGRAPHY

60

[12] NOAA Comprehensive Large Array-data Stewardship System. URL: http://www.saa.noaa.gov/nsaa/products/welcome. [13] O. Hagolle and F. Cabot. Cyclopes level 1 processing: algorithm theoretical basis document. [14] Stowe, L.L.and E.P. McClain, R. Carey, P. Pellegrino, G.G. Gutman, P. Davis, and C. Long. Global distribution of cloud cover derived from noaa/avhrr operational satellite data. Advances in space research, 3:51–54, 1991. [15] H. Rahman and G. Dedieu. smac: A simplified method for atmospheric correction of satellite measurements in the solar spectrum. Int. J. Remote sensing, 15(1):123– 143, 1994. [16] G. Gutman. The derivation of vegetation indices from avhrr data. Int. J. Remote Sensing, 8:1235–1243, 1987. [17] Jean louis Roujean, Marc Leroy, Pierre-Yves Deschamps, and A. Podaire. Evidence of surface reflectance bidirectional effects from a noaa/avhrr multitemporal data set. Int. J. Remote Sensing, 13:685–698, 1992. [18] Jean louis Roujean, Marc Leroy, and Pierre-Yves Deschamps. A bidirectional reflectance model of the earth’s surface for the correction of remote sensing data. Journal of geophysical research, 97(D18):20455–20468, December 1992. [19] M. Minnaert. The reciprocity principle in lunar photometry. Astrophysical Journal, 93:403–410, 1941. [20] C. Tison. Mise en coh´erence d’images AVHRR et VEGETATION. Synth`ese d’un produit biophysique de v´eg´etation pour les deux capteurs. Technical report, Internship report CNES, February 2002. [21] Paul Burke. Determining if a point lies on the interior of a polygon, 1987. From the website of the author, who is a teacher at The Centre for Astrophysics and Supercomputing, Swinburne, Australia. URL:http://astronomy.swin.edu.au/ pbourke/geometry/insidepoly/. [22] Fran¸cois Cabot. Extraction de donn´ees pour les sites d´esertiques ou comment nourrir SADE. Technical report, CNES, 2004. [23] Anne Chanie and Thomas Feuvrier. Manuel utilisateur MEDICIS. Technical report, CNES, 2004. [24] US Geological Survey’s guide of AVHRR. URL:http://edc.usgs.gov/guides/avhrr.html.