LEVERAGING SATELLITE DATA FOR WATER USAGE INSIGHTS: A COMPARISON OF SEN-ET, EEFLUX, AND THE PRE-TRAINED U-NET MODEL SAR2ET USING UAV-DERIVED EVAPOTRANSPIRATION AS GROUND TRUTH

In addition to its central role in hydrology and climate studies, evapotranspiration (ET) is the most crucial parameter in understanding agriculture’s water use. Due to its critical role in managing water resources, this study will evaluate the three satellite-based ET estimation products, having Unmanned Aerial Vehicle (UAV)-derived ET as ground truth. The ground truth data is acquired from a height of 200 m over irrigated cotton fields with thermal orthomosaics based on Zenmuse H20T thermal camera integrated into Matrice 300 RTK UAV. We explore and compare three satellite-based ET data over these agricultural fields to uncover their usage potential in practice at time t. The first data is Landsat-based Analysis Ready Data (ARD) called the Earth Engine Evapotranspiration Flux (EEFlux), which provides 30 m resolution global ET data but with limited time resolution, 16 days. The second dataset is Sentinel-1-based ET data, generated using a pre-trained U-Net structure with the SAR2ET model, providing regular updates based on Sentinel-1 passes. The last one is based on the Sen-ET using Two-Source Energy Balance (TSEB) model and proposed for obtaining high (∼ 20 m) resolution ET data from Sentinel-2 and Sentinel-3 as inputs. A general underestimation and reduced variability —a narrower range of values— are observed for all satellite-based ET products compared to the RGB+thermal UAV-based ET values, likely due to the coarse spatial resolution that the spatial variability of ET cannot be detected. Compared to the Landsat-based ARD ET estimations, the Sentinel-2/-3-based ET estimations displayed higher agreement with UAV acquisitions. This could be related to calibrating the model with more site-specific conditions at acquisition time, leading to capturing the RGB+thermal UAV-based ET values. Given that SAR2ET is trained by EEFlux data, SAR2ET model-based ET estimations are, as expected, closer to the one from EEFlux.