Downscaling MODIS evapotranspiration into finer resolution using machine learning approach on a small scale, Ribb watershed, Ethiopia

Abstract

By monitoring evapotranspiration (ET), the exchange of water and energy between the soil, plants, and the atmosphere can be controlled. Routine estimations of ET on a daily, monthly, and seasonal basis can give relevant information on small-scale agricultural practices, such as the Ribb watershed in Ethiopia. However, MODIS sensors have recently given high temporal resolution ET products across large areas, but their low spatial resolution limits its application on a local scale. The primary goal of the study was to downscale the MODIS ET (1 km) product to a finer spatial resolution at the watershed level. The model’s 12 predictor variables (NDVI, EVI, LAI, FVC, SAVI, NDMI, NDWI, Albedo, emissivity, LST, and DEM: slope and elevation) were produced using the random forest (RF) algorithm using Sentinel-2 (S-2) 20 m and Landsat-8 (L-8) 30 m. The RF regression model was used to assess the relationship between predicted variables and downscaled MODIS ET. The FAO-PM ET model, developed from meteorological stations, was validated by \(\varvec{R}^{\varvec{2}}\) and RMSE for three seasons (rainy, post-rainy, and dry) in 2022. The results were in good agreement with MODIS ET, with an RMSE of 0.22 for S-2 and 0.28 for L-8. In the FAO-PM ET model, the downscaled result showed greater spatial details and better agreement with gage station readings (\(\varvec{R}^{\varvec{2}} \varvec{\approx } \varvec{0.88}\) and \(\varvec{0.82}\)). Thus, considering the effectiveness and simplicity of machine learning techniques, our study demonstrated the potential for ET downscaling. Furthermore, the study suggests integrating spatiotemporal time series data to reach higher resolution.