Estimating actual evapotranspiration from widely available meteorological data with a hybrid CNN–LSTM

Actual evapotranspiration (ETa) is critical for closing terrestrial water and energy budgets but remains poorly constrained where direct flux measurements are sparse. Many applications substitute reference evapotranspiration (ETo) for ETa, conflating atmospheric demand with realized flux. Here we develop a hybrid CNN–LSTM framework that estimates daily ETa directly from widely available station-type variables (radiation, temperature, vapor pressure deficit, wind, precipitation, pressure, soil moisture) plus static soil texture; ETo is used strictly as an auxiliary predictor rather than a surrogate target. Using 167 FLUXNET sites spanning diverse biomes, the model explains 92% of daily variance in withheld tower data (validation R2=0.92; RMSE=0.38 mm d−1), with comparable performance on an independent test set. Ablation indicates solar radiation, pressure, wind, and vapor pressure deficit are most informative; a coarse vegetation class adds noise, and ETo delivers only modest marginal benefit once constituent drivers are included. We deploy the model with ERA5 forcings to produce global ETa fields for 2012 and 2024. Spatial patterns align closely with ERA5, and daily performance is highest across snow-influenced continental climates, while arid and high-latitude regions show larger discrepancies, reflecting known uncertainties in both training data and reanalysis benchmarks. By leveraging standard meteorological observations, the approach enables temporally continuous ETa estimates without reliance on ETo surrogacy or dense satellite coverage, providing a practical benchmark for hydrologic modelling, drought monitoring, irrigation planning, and land-surface model evaluation.