Enhancing Evapotranspiration Estimates by Integrating Sentinel-3 LST and Sentinel-2 Data Through Machine Learning-Based Downscaling

Evapotranspiration (ET) plays a significant role in monitoring and forecasting drought. High spatiotemporal resolution ET data enables precise decision-making in agricultural management and drought assessment. The synergy of daily Sentinel-3 Land Surface Temperature (LST) and Sentinel-2 multispectral images offers opportunities for more accurate ET estimation. In this study, we estimate and assess ET of Busia County, Kenya using Earth observation data, machine learning, and Two-Source Energy Balance (TSEB) model. A random forest model is employed with an input of Sentinel-2 spectral indices to downscale Sentinel-3 1-km LST images to 20-m. An optimized random forest model produced 20-m LST with accuracies 0.88, 0.82, and 0.55 for R
²
, RMSE, and MAE. The finer resolution of LST and ERA5 meteorological data serve as inputs for TSEB, estimating daily actual ET at 20-m. The final ET product (20m, Sentinel 3 timesteps) is compared with WaPOR and Gleam products to evaluate the derived ET products.