A Paradigm Shift to Automated Machine Learning for Local and External Reference Evapotranspiration Estimation with Uncertainty Implication

Accurate estimation of reference evapotranspiration (ET0) can be decisive in agricultural, hydrological and meteorological applications. Although different machine learning (ML)-based models have been successfully applied for ET0 estimation under a wide spectrum of climatic conditions, most of these models present the crucial shortcoming of being site-specific. Hence, a thorough hyperparameter tuning would be necessary before translating such models to another domain with different data distributions. The hyperparameter tuning is a complex procedure that mainly depends on the operator’s experience. Automated ML might be a suitable approach to adapt the models’ architectures. The present study evaluated the performance of different automated ML algorithms, namely, neural architecture search (NAS), Optuna, enhanced grey wolf (EGWO), and quantum whale optimization (QWOA) algorithms coupled with random forest, neural networks, and light gradient boosting models for estimating daily ET0 at three different climatic regions (Cairo, Singapore, and London). For local validation, the NN-NAS model provided the most accurate results in Cairo (R2 = 0.969, RMSE = 0.432 mm/day) and Singapore (R2 = 0.657, RMSE = 0.596 mm/day), while NN-Optuna provided the highest performance accuracy in London (R2 = 0.941, RMSE = 0.370 mm/day). Hybrid AutoML models improved R2 by 5–15% and reduced RMSE by 10–20% compared to standalone models. In external validation, NN-NAS and NN-Optuna presented superior generalizability, with R2 values up to 0.899 and 0.680 in London and Cairo, respectively. Nonetheless, the performance of the hybrid models depended on the climatic conditions of the studied sites, where NN-NAS was the best model for the arid site, while NN-Optuna provided the highest accuracy in the temperate climate. Further, the analysis of variance confirmed significant differences among the performance accuracies of the developed model. The Shapley additive explanations (SHAP) analysis was performed to identify the variables’ effect on ET0 estimation, which suggested that solar radiation showed the highest impact in all three studied climatic contexts, although the degree of importance was climatic dependent. Finally, an external modeling scenario was conducted using exogenous data for estimating ET0 at the target sites, which confirmed the models’ ability.