Efficient Hyperparameter Optimization for Reference Evapotranspiration Estimation with Limited Parameters: A Comparison of Optuna and Grid Search in the Doukkala Region, Morocco

Abstract

Accurate estimation of reference evapotranspiration (ETo) is essential for irrigation scheduling and water resource management, particularly in semi-arid regions where meteorological data are often limited. In this study, we evaluate the performance of machine learning models—Support Vector Regression (SVR), Random Forest (RF), Extreme Gradient Boosting (XGB), and Artificial Neural Networks (ANN)—for daily ETo estimation in the Doukkala region of Morocco using only limited parameters (Julian day, maximum and minimum temperature, and relative humidity). Two hyperparameter optimization strategies, Grid Search and Optuna, were compared to assess their effectiveness and efficiency. Results show that all models achieved high predictive accuracy, with R² values ranging from 0.90 to 0.916 and RMSE between 0.53 and 0.56 mm/day. Optuna consistently matched or slightly outperformed Grid Search across all models while requiring fewer evaluations. For example, RF and SVR achieved R² of 0.9160 and 0.9118 respectively with Optuna, compared to 0.9156 and 0.9104 with Grid Search. Similarly, XGB improved from 0.9040 to 0.9137 with Optuna, while ANN performance remained stable around R² ≈ 0.913. These findings highlight the effectiveness of Optuna as a more efficient and flexible alternative to Grid Search for hyperparameter tuning in ETo modeling.