Improving evapotranspiration estimation by integrating process-based biophysical variables into a deep learning approach

103 FLUXNET2015 flux towers distributed across diverse climatic and ecological regions. Accurate estimation of evapotranspiration (ET) is critical for understanding regional ecohydrological processes. Physically based models such as the Penman-Monteith-Leuning (PML) model are robust but often constrained by fixed parameterization schemes, while data-driven approaches such as Long Short-Term Memory (LSTM) networks can capture nonlinearities but depend heavily on training data. To address these limitations, this study developed a hybrid model (PML-LSTM) by integrating biophysical variables from PML simulation into LSTM network. Model performance was systematically evaluated against standalone PML and LSTM across three modelling levels: local (site), type (vegetation type), and group (forest and non-forest). The PML-LSTM model achieved superior performance, with median NSE values of 0.851, 0.913, and 0.933 during validation, surpassing both the PML model (0.843, 0.788, 0.766) and the LSTM model (0.818, 0.879, 0.873). Integrating biophysical information in the PML-LSTM model improved ET estimation accuracy and model generalization, leading to more robust spatiotemporal performance under leave-one-out cross-validation and extreme weather extrapolation experiments. Distinct model behaviors emerged under varying sample conditions: the PML model exhibited greater robustness under data-scarce conditions at the local level, while the LSTM and PML-LSTM models benefited from larger training datasets. This study highlights the potential of combining process-based and data-driven approaches to improve ET estimation and provides insights for regional ecohydrological modelling. • Biophysical variables simulated from PML were integrated into LSTM to improve ET estimation. • LSTM-based models could capture ET dynamics at multiple modelling levels. • The Hybrid model achieved superior ET estimation accuracy and spatiotemporal generalization.