Spatially remotely sensed evapotranspiration estimates in Sahel regions using an ensemble contextual model : structural uncertainty estimation and reduction

The Sahel region, identified as a "hot spot" for climate change, is characterized by a water scarcity and an inter-annual variability of water resources. Indeed, ongoing climate changes intensify the evaporative demand which could lead to more frequent period of droughts. Therefore, an important issue in these countries is to provide accurate estimation of evapotranspiration (ET) in a spatially distributed manner. The growing number of spatial ET products, including simple empirical equations (e.g., Penman-Monteith), land surface models (LSM), energy balance models, interpolated in-situ measurements, neural network approaches, or data fusion, form an interesting alternative in these areas scarcely gauged. However, until recently, there is no product combining simultaneously good spatiotemporal resolution (i.e., <1km, <daily) and good performances. Remote Sensing (RS) data in the thermal infrared domain, used in energy balance models, is particularly useful because it allows for spatial ET estimates at various space-time resolutions. A well-adapted method for the Sahelian context was proposed based on an ensemble contextual energy balance model combining thermal and visible satellite information (EVASPA S-SEBI Sahel method; E3S, Allies et al, 2020, 2022). This contextual method is based on the thermal contrast (hot/dry and cold/wet pixels) observed in a given thermal image to provide an ensemble of instantaneous estimation of evapotranspiration conditions. The applicability and accuracy of this approach suppose: (1) The presence of sufficient heterogeneity between dry and wet pixels within the same image and (2) the correct identification of the driest and wettest pixels, also known as dry and wet boundaries. These two hypotheses are rarely checked before computation within contextual models, leading to high uncertainties in ET estimation. Therefore, the aim of this study is firstly to allow for a systematic detection of the heterogeneity conditions and a dynamic selection of adapted methods for the determination of wet and dry boundaries by using only the image information without prior knowledge of local conditions. Secondly, our aim is also to assess the added value of using a thermal information from high spatial resolution (Landsat or Ecostress data) compared to medium resolution (Modis data) on the image heterogeneity and consequently on ET estimation. The proposed method shows higher performance in comparison with reference ET products in our study area in central Senegal, with a lower RMSE value (around 0.5 mm.day^-1) compared to eddy-covariance measurements. Moreover, it reduces significantly structural uncertainties by around 0.6 mm.day^-1 in dry season and around 0.4 mm.day^-1 in wet season. Thermal information from higher resolution data are expected to further improve ET simulation due to a higher perceived heterogeneity in satellite images. It could lead to more accurate estimates of surface water deficit in semi-arid areas. The use of high-resolution data also makes this study a good demonstrator for the upcoming thermal earth observation missions like TRISHNA (CNES/ISRO), which this work is part of, LSTM (ESA) and SBG (NASA).