Hyperparameter optimization of machine learning models for predicting actual evapotranspiration

Direct measurement of actual evapotranspiration (AET) using eddy covariance and lysimeters is challenging, particularly in large areas, due to high cost, technical complexity, and the need for specialized instrumentation. Consequently, AET data is limited, prompting the use of meteorological and soil features for prediction. This study develops and evaluates machine learning models for AET prediction based on two input combinations. The first group, selected through Pearson correlation, tolerance, and VIF scores to address multicollinearity, includes net CO₂, sensible heat flux, air temperature, relative humidity, and wind speed. The second group, chosen for practical applicability and more accessible, consists of soil surface temperature, air temperature, relative humidity, and wind speed.

Two predictive approaches are proposed: (i) deep learning models (LSTM, GRU, CNN) and (ii) classical machine learning models (SVR, RF). Hyperparameters were optimized using Bayesian optimization and compared with grid search. Bayesian optimization demonstrated higher performance and reduced computation time. Model performance was evaluated using statistical indicators (RMSE, MSE, MAE, R²). Deep learning methods outperformed classical methods, with LSTM achieving the best results (Bayesian optimization: RMSE=0.0230, MSE=0.0005, MAE=0.0139, R²=0.8861).

Performance decreased with fewer predictors. LSTM maintained superiority, achieving R²=0.8861 with five predictors and R²=0.8467 with four. LSTM also slightly outperformed SVR (R² = 0.8456) with fewer predictors. Overall, deep learning methods, especially with Bayesian optimization, have been shown to be more effective than classical machine learning methods for AET prediction. This findings encourage future research using varied input combinations and advanced modeling approaches for AET accurate prediction.