Environmental Sensing Analytical Techniques for Earth Observation

Également Samuel, Coralie, Jennifer Leslie, Julie Mohammad, Alain Sylvie, Héloïse… Stéphane, La De, Alimentaire, De L'ouest
2012 unpublished
he African population is now estimated at 831 million and is projected to reach 3.8 billion in 2100. In addition to those human demography changes, deep-rooted environmental changes, including climate change will probably negatively affect agricultural production in a region dominated by rainfed farming systems. Consequently, agricultural production monitoring has to be strengthened in order to reduce the population' vulnerability to food insecurity and to allow the implementation of effective
more » ... daptation and risk mitigation measures. Remotely sensed observations give synoptic, timely and objective information on the state of Earth surfaces and thus constitute a reliable source of information for agricultural monitoring at a regional scale and its use in the framework of the food security monitoring systems (EWS) might be reinforced. This thesis investigates different methodological approaches based on moderate resolution remote sensing products from MODIS for the monitoring and characterization of agricultural production in West Africa. In particular three aspects are addressed: cropland, yield monitoring and biomass production trends. First, we assess the quality and reliability of the MODIS Land Cover (LC) product for locating and estimating crop areas at a regional scale. Using agricultural statistics, we show that the MODIS LC product allows a good estimation of crop acreage and dynamics at regional and national scales. Then, concerning the cropland spatial distribution, our findings highlight a strong relationship between the MODIS LC product user accuracy and the fragmentation of agricultural landscapes (R²=0.62). Based on these results, we produced a regional map of the MODIS LCP accuracy for cropland classes at regional scale. In addition, we used the Pareto Boundary method to isolate the part of incompressible errors (12% for omission errors and 20% for commission errors) due to the low resolution of the data and the high heterogeneity of African agricultural landscapes, from the part that could be directly linked to the performance of the adopted classification algorithm. The second part of the thesis is devoted to yield monitoring in West Africa focusing on a South-West Niger site. In a move towards spatialized yield estimation, two approaches based on remote sensing and the SARRA-H crop model were tested and compared: (i) an empirical statistical model derived from MODIS NDVI to estimate aboveground biomass and the CWSI to estimate the Harvest Index, and (ii) the SARRA-H crop model using satellite rainfall estimates products as input data. When compared to official agricultural statistics, the first model allows a good year-to-year yield variability estimation (r=0.82) owing to an implicit integration of yield limiting and reducing factors. The second method gives more of an indication about exploitable yield. In the final part, the analysis of MODIS NDVI time series allowed a better characterization of biomass production dynamics at regional and local scales. At the scale of the Sahelian region, we found that greening trends, meaning an increase in biomass production, are induced mainly by climatic factors, while the browning trends seems to be linked either to a combination of climate and human impacts or to human activities only. At local scale (South West Niger), we further analyzed the main drivers of biomass production changes by relating NDVI trends to a set of potential drivers using the RandomForest algorithm. The results revealed that biomass production dynamics are determined both by annual rainfall as well as soil type and land accessibility constraints. The methodological developments made and results obtained in this thesis open new perspectives for monitoring agriculture at regional scale and thus might contribute to strengthening the EWS effectiveness in both their monitoring and warning functions. Moreover, the upcoming availability of Sentinel-2 data with higher spatio-temporal resolutions should contribute to significantly strengthen our approach. ABSTRACT 15