Deep Learning and Earth Observation for the Study of West African Rainfall: Observing rainfall processes through the lens of AI

West African food and economic safety are heavily reliant on agriculture, most of which is rainfed. Changing rainfall patterns induced by global warming jeopardize yields by unpredictable water availability. At the same time, a rapidly growing population leads to rising demands for food production. Accurate rainfall information is essential for farmers to adjust their crop management practices and avoid yield, thereby improving the overall resilience of the region. However, this information is largely lacking due to a sparse rain gauge distribution, limited resources and data transmission challenges. Added to this, existing satellite rainfall products show a particularly poor correlation with ground observations in West Africa.
This dissertation aims at improving rainfall information for farmers in the Sudanian Savanna bioclimatic region of West Africa. The Sudanian Savanna stretches across Africa as a broad belt, covering roughly from southern Mali in the north to northern Ghana in the south. Its West African area expands from Senegal in the west to Chad in the easte. Improving rainfall information will support climate resilience and food and economic safety in the region. This research leverages on the unique potential of Earth Observation satellites to provide rainfall information everywhere, due to their global coverage and ability to track atmospheric processes. To tackle the general poor performance of existing rainfall information products, this work proposes an alternative avenue. Particularly, it investigates the potential of Deep Learning (DL) methods to extract relationships between meteorological variables and raw satellite data that might be overlooked by traditional satellite rainfall retrieval methods....