Spatially remotely sensed evapotranspiration estimates in Sahel region using an ensemble contextual model with automated heterogeneity assessment

Water scarcity and the inter-annual variability of water resources in semi-arid areas are limiting factors for agricultural production. The characterization of plant water use, together with water stress, can help us to monitor the impact of drought on agrosystems and ecosystems, especially in the Sahel region. Indeed, this region is identified as a ”hot spot” for climate change. In-situ measurements often are insufficient for accounting for spatial variability at large scales ($>100$ km) due to the scarcity of gauge networks. To tackle this issue, remotely sensed evaporation is often used. In this study, estimates using thermal infrared and visible data from MODIS/TERRA and AQUA are used. Spatially distributed estimates of the daily actual evapotranspiration (ETd) are simulated using the EVASPA S-SEBI Sahel (E3S) ensemble contextual method over a mesoscale area (145x145 km) in central Senegal. E3S uses a set of different methods in order to identify the dry and wet edges of the surface temperature/albedo scatterplot and therefore estimate the evaporative fraction (EF). However, contextual approaches assume the simultaneous presence of sufficient fully wet and fully dry pixels within the same satellite image. This assumption of heterogeneity does not always hold, especially in the Sahel, which is characterized by the alternation of dry and wet seasons due to the monsoon-influenced climate. To tackle this issue, E3S uses different sets of methods depending on the season, based on local knowledge. The present study thus aims at generalizing the approach by proposing a new version of E3S called ”E3S-V2”. This latter allows an automatic detection of different heterogeneity conditions. Therefore, a sensitivity analysis examining the effect of using different EF estimation methods over different spatial coverages was performed. It made it possible to identify relevant normalized indicators to determine the heterogeneity level, as well as to discriminate among the most adapted EF determination methods for each situation. From this analysis, an automated procedure of method selection according to the heterogeneity conditions is proposed. A local-scale evaluation was performed using eddy-covariance measurements in the Senegal Groundnut Basin. A spatialized evaluation was also performed using GLEAM and ERA5-Land, which are proven reference ETd products over the area. ”E3S-V2” simulations yield comparable performances with in-situ and reference products in our study area.