The contribution of VEGETATION /SPOT4 products to remote sensing applications for food security, early warning and environmental monitoring in the IGAD sub region G. Pierre, C. Crépeau, P. Bicheron, T. Bennouna and Valérie Caldairou SCOT, 8-10 rue Hermès, Parc Technologique du Canal, 31526 Ramonville Cedex, France Introduction The IGAD sub-region including Ethiopia, Sudan, Eritrea, Somalia, Kenya, Uganda, and Djibouti is highly dependant on rain-fed agricultural activities which accounts for approximately 60% of its GNP. Consequently, droughts or desertification phenomena have a strong impact on population and economies. In this framework, the overall objective of this project is to improve the use and dissemination of remote sensing data and other ancillary information for early warning systems in food security and environmental monitoring. The IGAD-project started by an assessment mission in the IGAD countries from January to March 1999, and several facts were observed. There is a little use of remotely sensed data in the various institutions in the sub-region. Moreover, most of the centers in the sub-region are poorly equipped and thus face great difficulties in their operations. In this regard, it is logical to answer the following questions: - What are the needs in terms of data, and what information are needed? - How these data will be provided to the interested institutions? On a technical point of view, one key of the success of this project relies on fully operational communication infrastructures to provide real-time data (information) to the end-. This infrastructure is essential in many aspects but principally it will guarantee availability of information to all states member in case of failure of their national sources. The advanced very high resolution radiometer (AVHRR) has been, until recently, the only satellite sensor used for remote sensing applications on early warning and environmental monitoring at a regional scale. Due to its improved quality concerning radiometric, geometric, and atmospheric processing, VEGETATION sensor brings a more accurate information than AVHRR, but does not have the AVHRR temporal depth, necessary for early warning and environmental studies. Consequently, the inter-comparison between AVHRR and VEGETATION data is important and represents the first step to constitute a virtual VEGETATION archive. In this paper, we first present the IGAD project architecture. In a second stage, we develop the interest of the VEGETATION data in the framework of this project and the manner to harmonize the use of these data. We conclude the study while providing results concerning the inter-comparison of 10-day NDVI derived from VEGETATION and AVHRR sensors. IGAD-Project architecture The success of the project relies on the capacity to provide real time data to the end-. The IGAD-project s approach will set up two simultaneous strategies to ensure an operational use of the information. In both of them, Meteorological Services (NMS) will play a central role in acquisition, processing and dissemination of the data (figure 1). - Network of data dissemination: the first strategy will put into service two sub-regional hubs (Kenya and Ethiopia) for acquisition and processing of satellite data. Sub-regional and national products will be available on a sub-regional file, and accessible by telecommunications links to all the NMS in the IGAD sub-region.
- Direct capabilities of satellite data: this second strategy will complement the reliability of the general system in equipping countries with national acquisition facilities (Sudan, Uganda, Kenya) where they are not existing, and in the other one where an upgrade of the installations is deemed necessary. This global architecture can be efficient only if the functioning of the key institutions involved in the process of producing the various data necessary to build an efficient information system is reinforced. Three different entities may be distinguished: NMS, Early Warning Units (EWU), and Environmental Units (EMU). The project will enhance the capacity to process and improve the quality of products directly derived from satellite, and will make them immediately available for the national institutions. NMS will act as provider for other national agencies interested in the data. The EWU and EMU will benefit from the project by having a direct connection to a national or regional that will disseminate the low-resolution products. The success of this project depends therefore on several factors: - the capacity to quickly disseminate the information, - the possibility to have an important temporal data archive: unlike VEGETATION sensor, AVHRR has this archive, the inter-comparison study presented here is then very important. - the obligation to generate products with the same methods in order to permit their comparison between countries in the whole IGAD sub-region. AVHRR-VEGETATION VEGETATION inter-comparison Inter-comparison study area and data processing One year of 10-day VEGETATION NDVI over the sub-region IGAD is not sufficient yet to generate products useful directly in environmental studies such as desertification, variation of biomass production estimates, or early warning which need data over large periods. In these cases, the constitution of a virtual VEGETATION archive is necessary. We will focus more precisely on the comparison of 10-day NDVI synthesis derived from the both instruments. The study area (figure 2) is located within the IGAD sub-region over an area of 1500 x 1500 km² (19 N, 25 E ; 5 N, 39 N) with a broad vegetation gradient from north to south. The landcover is characterised by different vegetation classes (10 classes following the USGS landcover classification (figure 3) realised with LANDSAT data). The classes are : Barren and sparsely vegetated, Shrubland, Grassland, Cropland and Grassland, Dryland and Cropland, Evergreen Broadleaf Forest, Cropland and Woodland, Urban, Water. The temporal NDVI series are acquired at a 10-day frequency following the MVC technique over 19 decades between may and october 1998 with the full spatial resolution. Such products named S10 are operationally delivered for VEGETATION by CTIV but not for AVHRR. Therefore, the AVHRR data have been processed through SPACE II/OSS software (Millot, 1995; Sahara and Sahel Observatory, 1997) which includes atmospheric corrections using the SMAC code (Rahman and Dedieu, 1994), cloud detection, and geometric corrections. The NDVI series are then established following the MVC technique. In order to enhance the geometric corrections in a more accurate way, a VEGETATION NDVI image is used as reference. Finally two comparable datasets with almost the similar processings are at our disposal. Inter-comparison Methodology For the both datasets, the whole studied area has been spatially sampled at a step of 1. It leads to a database of 175 NDVI temporal signatures. Our approach is composed of two stages. The first one relies on a coarse filtering of AVHRR-NDVI data: all the NDVI values greater than 0 and lower than 0.3 (for dryland) and 0.7 (for all other classes, except forest) have been kept. Then,
the AVHRR and VEGETATION NDVI temporal signatures have been analysed for some landcover types, and the linear regression between AVHRR and VEGETATION has been assessed. The second stage consists of a filtering improvement. Two kinds of outlines were considered: linear (width of 2 standard deviation around the initial regression line), and hyperbolic (eccentricity e and top distance a, symmetry axis about the first bisecting line). For each kind of outlines, two constraints are taken into account. The first one is rejecting a value if it is outside the template, and the second one is rejecting the entire temporal signature if more than 25 % of the signature is outside the template. For these 4 cases, the regression coefficients between AVHRR and VEGETATION are assessed. Inter-comparison results The left handside of figure 4 presents the first linear filtering which rejects points if not into the outline. This outline is defined around the initial regression line at more or less a distance equal to one standard deviation. The right handside of figure 4 presents the results for a hyperbolic outline centered at the mean of the both median values of VEGETATION and AVHRR NDVI distributions. The root mean square error (rmse) is better than before with a mean value over 7 classes equal to 7% including Cropland and Woodland class (not displayed here). The mean percentage of points present in the outline is of 59%. The means of all the director coefficients except "Barren or sparsely vegetated" class equal 0.92. The discrepancy between AVHRR and VEGETATION is thus much smaller than before. The figure 5 presents the same filters with an additional assumption of temporal constraint. In the both cases of linear or hyperbolic outlines, the improvement is confirmed with a decrease of rmse at 6% and a reasonable percentage of present points. We observe also an increase of director coefficients. All the results are summarised in table 1. Filters Temporal Constraint rmse (%) % points present in the template a b Coarse 0.13 0.58 0.20 Linear Hyperbolic Yes 0.056 57 0.83 0.07 No 0.13 75 0.76 0.18 Yes 0.066 62.5 0.94 0.03 No 0.06 59 0.92 0.05 Table 1: The correlation parameters obtained for each filtering method Example of elaborated products Due to the lack of ground data network within this sub-region, the recourse to satellite imagery is necessary for the vegetation monitoring. In this case, the primary product is the NDVI. We propose here a better interpretation of this original product with: - improvements of the mapping techniques, - choice of a standard and unique NDVI look up table (FAO and FEWS table), - addition of subsidiary data, - use of same projection for all products, The proposed products are given following the 3 maps : - Chronological analysis using successive 10-day NDVI synthesis, - differential analysis (NDVI comparison between two successive 10-day synthesis, two 10- day NDVI synthesis of two different years) - Subsidiary data for the NDVI interpretation : mean of annual rainfall, landcover types, population density, altitude.
The figure 6a presents the successive NDVI maps for decades of March and April. This representation of NDVI maps with the superimposed cartographic information and other subsidiary data allows to the end a quick chronological analysis of their interest zones. It also allows to visually identify areas where the evolution is abnormal, and to have a first interpretation of this anomaly. The figure 6b shows a comparative analysis of NDVI between two successive decades and between the studied decade and the same decade of the previous year. NDVI are always superimposed to subsidiary data. This representation allows to identify areas where vegetation activity increases, areas without any change, and areas where vegetation activity decreases. In this case the ancillary data permit to to give better interpretation of changes. Conclusions Since the beginning of the project, several tasks have been accomplished mainly for the assessments of national/regional needs as well as technical options and implementation of the project. From a technical point of view, the representatives of the IGAD states member have recently agreed the implementation modalities. From an applicative point of view, the AVHRR-VEGETATION intercomparison represents a critical step in the perspective of a virtual built up VEGETATION archive, necessary for the integration of VEGETATION data in early warning bulletins and numerous environmental studies. The obtained regression between AVHRR and VEGETATION shows a similar behaviour for the vegetation classes, but different for Barren or sparsely vegetated class. These differences between vegetated and non-vegetated surfaces may be partly due to the various widths of spectral bands between the both sensors AVHRR and VEGETATION, and further investigations are needed in that direction. The different filterings improve the results in terms of a rmse decrease. In particular, the temporal constraint leads to better regressions with an increase of the linearity between VEGETATION and AVHRR. A number of support projects identified for the environmental component of IGAD-project will benefit from the built up of virtual VEGETATION archive. References Rahman H, and G. Dedieu, SMAC: a simplified method for the atmospheric correction of satellite measurements in the solar spectrum, Int. J. of Rem. Sen., vol. 15, n 1, 123-143, 1994. Millot M., NOAA-AVHRR preprocessing, Workshop for central and eastern Europe on agrometeorological models. Theory and applications in the MARS project. Nov. 1994, Ispra Italy. Publication EUR 16008 EN of the office for official publications of the EU, Luxembourg, 173-179, 1995 Space II OSS, Système de traitement des images NOAA-AVHRR. Manuel Utilisateur, AGG- 96/001-SUM-001, version 1, Observatoire du Sahel au Sahara, 1997.
SUDAN NMS Upgraded ERITREA ISP ETHIOPIA DJIBOUTI NMS NEW NMSA Regional Upgraded ISP ISP Internet Internet UGANDA GTS- TCP/IP KENYA SOMALIA SPOT4/ VEGETATION ftp CTIV NEW NMS GTS- TCP/IP Regional KMD NEW Internet backbone International Fig. 1 Functioning architecture between the entities of the IGAD member states Figure 2 Studied area Figure 3 Landcover map
Figure 4 Filtering without any temporal constraint
Figure 6 a-b Chronological and differential 10-day VEGETATION NDVI series