Daily evapotranspiration estimation in Germany through multi-source data fusion and hybrid machine learning models

Accurate estimation of daily evapotranspiration (ET) is vital for sustainable water resource management and irrigation practices. However, achieving high accuracy remains challenging due to the complex physical nature of ET, its sensitivity to environmental variability, and uncertainties in both models and input data. This study evaluates five models for daily ET estimation by integrating remote sensing (RS) and meteorological data across three land-use types in Germany. The tested models include the Artificial Neural Network (ANN), Adaptive Neuro-Fuzzy Inference System (ANFIS), TabTransformer, Random Forest (RF), and the physically based OPTRAM-ET model. Sentinel-2 and Landsat-8 data, together with meteorological observations and ERA-5 reanalysis data, were used as inputs, and ET estimates were validated against observations from Integrated Carbon Observation System (ICOS) stations. The results showed that AI models generally outperformed both RF and OPTRAM-ET. Among all models, the ANN achieved the best overall performance. Using only satellite data, the ANN achieved an RMSE of 0.98, while combining satellite and meteorological data significantly improved performance, resulting in an RMSE of 0.60, a correlation coefficient of 0.91, and a coefficient of determination (R2) of 0.83. Model performance varied across land-use types, with the highest accuracy observed in grassland and the lowest in cropland. SHAP analysis revealed that ET was mainly influenced by radiative energy, vegetation condition, and photosynthetically active radiation (PAR), with the Enhanced Vegetation Index (EVI) consistently identified as the most important predictor. Validation against the Standardized Precipitation-Evapotranspiration Index (SPEI) showed that ANN-estimated ET explained up to 40% of the variance in the drought index, compared to 16% for MODIS ET data. These findings highlight the potential of ANN-based ET estimation for improving drought monitoring and supporting more effective water and irrigation management.