LSTM and Temporal Fusion Transformers for daily evapotranspiration estimation using FLUXNET2015 and Google Earth Engine

This is the authors' abstract. We don't add key points for this paper.

Accurate estimation of actual evapotranspiration (ETₐ) is essential for hydrological modeling and drought assessment. Using 270,489 daily observations from 115 FLUXNET2015 sites (2000–2014), this study benchmarks deep-learning and machine-learning model for daily ETₐ prediction. Predictors from Google Earth Engine (GEE) include radiation, temperature, humidity, precipitation, soil moisture, NDVI, and seasonal indicators. Data-driven architectures (ANN, XGBoost, ANN-ET, LSTM, TFT) are evaluated using a strict leave-one-site-out protocol to ensure full spatial independence. Temporal models markedly outperform static approaches: LSTM and TFT reach mean R2 ≈ 0.82–0.83 and RMSE ≈ 0.50–0.52 mm day⁻1, versus R2 ≈ 0.30–0.38 for ANN, ANN-ET, and XGBoost. Calibrated physical ET products (MOD16, GLDAS, ERA5-Land) are compared globally after linear adjustment and show lower skill. Monte-Carlo perturbations yield narrow 95% prediction intervals and minimal RMSE drift for temporal models. Climate- and biome-stratified analyses show consistently stronger performance of temporal architectures, with lowest skill across all methods in tropical monsoon regions.