Ensemble machine learning-based extrapolation of Penman-Monteith-Leuning evapotranspiration data

The Penman-Monteith-Leuning version 2 (PML-V2) evapotranspiration (ET) model, an advanced iteration of the classic Penman-Monteith (PM) model, is available globally via Google Earth Engine with a spatio-temporal resolution of 500 m and 8 days. PML-V2 improves canopy conductance estimation and incorporates carbon dioxide effects on transpiration via gross primary production. However, it faces limitations, particularly in calibration and the lack of pre-2000 data. This study applies several machine learning (ML) models—including a backpropagation neural network (BPNN), an adaptive neuro-fuzzy inference system (ANFIS), support vector regression (SVR), and long short-term memory (LSTM)—to simulate PML-V2 ET in the Ahar Chay basin, Northwestern Iran. The Seto mixed forest site in Japan, characterized by a contrasting ecosystem, served as a cross-validation site to further validate the methodology. Sensitivity analysis was performed to optimize the input variables and reduce uncertainty. Among the models, LSTM demonstrated superior performance, while an ensemble of shallow ML models increased prediction accuracy by up to 24 %. The optimal model was applied to extrapolate PML-V2 ET data for the period from 1983 to 2000. In the Ahar Chay basin, actual ET (AET) was estimated using the water balance equation, as direct observations were unavailable, and was evaluated via dynamic time warping from 2002 to 2016. Notably, neural ensemble ET and PML-V2 improved ET estimates by 55 % and 41 %, respectively, over the PM model, particularly during the growing season (April–September). At the Seto site, the methodology yielded a 39 % improvement over the PM model based on observed AET data. These findings have significant implications for ecohydrology, offering improved ET estimates for future projections and historical periods prior to 2000.