A satellite-driven hybrid machine learning framework for daily pixel-scale actual evapotranspiration estimation

Actual evapotranspiration (AET) is a key component of land-atmosphere mass and energy exchange and plays a critical role in water balance and surface energy modelling. Accurate estimation of AET is often limited by sparse or point-scale measurements. This study developed and evaluated three satellite-driven hybrid machine learning models, namely VMD-M5 Tree, VMD-XGB, and LLSVM-HHO, for daily pixel-scale AET estimation across 11 subbasins in northeastern Iran. Meteorological parameters and MODIS-derived satellite imagery were employed as predictors in three scenarios under two modelling structures. In the first structure, the RMSE values for the three models, VMD-M5, VMD-XGB, and LSSVM-HHO, were obtained in the ranges of 0.213–0.972, 0.25–0.908, and 0.2–0.734 mm.day−1, respectively. In the second structure, where spatial transferability was evaluated by training the models on data from nine subbasins and testing them on data from the other two subbasins, the RMSE and NSE values for the VMD-M5, VMD-XGB, and LSSVM-HHO models were obtained in the ranges of (0.33–1.76, 0.59–0.96), (0.33–1.67, 0.17–0.97), and (0.32–1.65, 0.23–0.97), respectively. These results highlight the potential of the proposed computational approach for high-resolution environmental modelling, while further evaluation in watersheds with different climatic and hydrological conditions is required to fully assess its generalizability.HighlightsA satellite-driven hybrid computational framework for pixel-scale daily AET estimationAssessing VMD-M5, VMD-XGB & LSSVM-HHO for capturing multi-scale dynamicsSatellite-derived MODIS data used to represent spatially distributed surface energy statesDemonstrated spatial transferability across independent subbasinsBetter performance in higher AET intervals across all three hybrid ML models