Modeling Actual Evapotranspiration Under Multicollinearity Using Remote Sensing Data with Machine Learning and Explainable AI

Accurate estimation of actual evapotranspiration (ETa) is essential for sustainable water resources management and agricultural planning. This study investigated the influence of reference evapotranspiration (ET₀) on ETa estimation using Random Forest (RF), Bagging, Least Squares Boosting (LSBoost), Generalized Additive Models (GAM), and Multiple Linear Regression (MLR). Global solar radiation (Rs), land surface temperature (LST), NDVI, and soil moisture (SM) were used as predictors under two scenarios: with and without ET₀. Variable importance and model interpretability were evaluated using Permutation Feature Importance (PFI) and SHAP analyses. Results showed that ET₀ was the dominant predictor when included in the models, exhibiting the highest importance and contribution values across all approaches. When ET₀ was excluded, the explan-atory structure shifted primarily toward Rs, followed by SM, LST, and NDVI. Although model accuracy generally decreased without ET₀, RF and Bagging maintained stable performance (R² ≈ 0.97), whereas LSBoost, GAM, and MLR exhibited more pronounced reductions in predictive accuracy. The analyses further revealed that SM consistently enhanced model performance by reducing prediction errors and improving robustness, particularly in the absence of ET₀. Overall, the findings demonstrate that both model selection and the physical representativeness of input variables are critical for reliable ETa estimation. Moreover, ET₀ drives an atmosphere-controlled evaporative demand regime, whereas Rs becomes the primary controlling factor when ET₀ is unavailable.