Combining satellite and local rain gauge data to improve weather forecast for local smallholder farmers in the African Sahel.
A Computer Science master research project by Tom Siebring – supervision by Anna Bon (Vrije Universiteit Amsterdam), January 2024, in collaboration with UPEC, Université Paris-Est Créteil and André Baart, Babafla.
Rainfall information is extremely important for smallholder farmers in West-Africa, especially in the light of climate change and adaptation to new patterns of rain-fed irrigation. However, conventional, numerical weather predictions have shown to be inaccurate for large regions in Africa, due to scarcity of meteorological observatoriums such as weather stations, radars and radiosondes in low resource environments. Recently, Machine Learning models have been trained on satellite data, however, these models often fail to address the unique climatic characteristics of the West African Savanna. The contribution of this study is a Deep Learning model that uses a combination of satellite and local weather station data from the field. The deep learning model is trained on infrared satellite imagery and rain gauge measurements from 61 weather stations, with rainfall data over a period of 5 years from the Trans-African Hydro-Meteorological Observatory (TAHMO). Furthermore, we include in the study the effects of vapor and dust, and the performance of rainfall measurement and prediction. First, instead of relying on NWP, we use deep learning on satellite imagery. Second, thanks to TAHMO, hundreds of affordable weather stations have been installed in this region over the past ten years. This allows for training weather models on local measurements, instead of using measurement estimates from weather products such as the Integrated Multi-satellitE Retrievals for GPM (IMERG).