Estimation de l'évapotranspiration en utilisant les méthodes d’intelligence artificielle cas des bassins versant algériens

Abstract

Accurate estimation of reference evapotranspiration (ET0) is critical for water resource management, agricultural planning, and climate studies. This study evaluates the performance of multiple machine learning and regression models in estimating ET0 across three watersheds with distinct climate types: Oued Rhumel (sub-humid), Soummam (sub-humid), and Cheliff (semi-arid). Using combinations of meteorological variables (e.g., Tmax, RHmean, Vitesse du vent, Tmin) and four modeling approaches (Multiple Regression, Bagged Trees, SVM, Neural Networks), we assess model accuracy through RMSE, MAE, MRE, and R² metrics. Results reveal that Neural Networks consistently outperform other models, achieving near-perfect R² values (1.000) with optimal variable combinations. While both sub-humid watersheds show high model performance, the semi-arid Cheliff basin exhibits greater sensitivity to variable selection, with simpler models (e.g., egression) showing higher errors under limited inputs. Climate type significantly influences model robustness, with sub-humid conditions allowing flexibility in variable choice, whereas semi-arid systems demand precise variable-model pairing. These findings underscore the need for climate-specific ET0 estimation strategies to enhance hydrological and agricultural applications in contrasting environments