Integrating Landsat and Physics-Informed Neural Networks for reference evapotranspiration estimation

Evapotranspiration (ET) is a key component of the hydrological cycle with important implications for water management and climate modeling. Despite the robustness of physically based evapotranspiration models, their application at high spatial resolution remains limited by point-scale forcing and the coarse representation of key meteorological drivers, particularly near-surface wind speed. In this study, we present a fully satellite-based framework to estimate reference evapotranspiration (ETo) at 30 m resolution by integrating Landsat observations with a physics-informed neural network (PINN). Near-surface wind speed at 2 m is first estimated to the Landsat scale using a Random Forest model that leverages static terrain and land-cover information together with dynamically retrieved land surface temperature, net radiation, and vapor-pressure deficit. These high-resolution meteorological fields are then used to drive a PINN constrained by the FAO-56 Penman–Monteith and Priestley–Taylor formulations, which are embedded as complementary physical losses to ensure consistency with both aerodynamic and radiative controls on ETo. The approach is evaluated across eight eddy covariance flux-tower sites spanning cropland, grassland, and forest ecosystems in Asia and Europe. Results demonstrate strong agreement with tower-based Penman–Monteith ETo (R = 0.80–0.97; RMSE = 0.52–1.43 mm/d), with the highest accuracy observed over homogeneous croplands and larger, yet systematic, deviations during short-duration high-flux periods in heterogeneous and structurally complex canopies. Spatial comparison with ERA5-Land ETo further highlights the added value of high-resolution, satellite-driven estimates in capturing sub-grid variability. These results indicate that physics-informed learning provides a robust and scalable pathway for canopy-scale ETo mapping in heterogeneous landscapes.