Evaluation of Artificial Intelligence Models for Reference Evapotranspiration Estimation in Precision Agriculture under Semi-Arid Climate Conditions

Accurate estimation of reference evapotranspiration (ET₀) is essential for irrigation scheduling in high-efficiency agricultural systems. This study compares the performance of two modeling approaches—specifically, the XGBoost algorithm and a long short-term memory (LSTM) neural network—for estimating daily ET₀ in a semi-arid region in northwestern Mexico, using meteorological data derived from the NASA POWER system between January 2020 and July 2024. Both models were trained on a homogeneous dataset of 1,827 days and evaluated using standard metrics, including MAE, RMSE, R², NSE, MAPE, and Willmott’s d coefficient, with separate analyses for dry and wet seasons. The XGBoost model achieved a MAE of 0.145 mm day⁻¹ and an R² of 0.983, outperforming the LSTM model across all metrics, which exhibited higher absolute error and greater seasonal variability. SHAP analysis identified shortwave solar radiation, vapor pressure deficit, and wind speed as the most influential variables, aligning with agrometeorological theory. Additionally, XGBoost demonstrated greater robustness to extreme events and lower systematic bias. These results position XGBoost as an accurate, interpretable, and computationally efficient tool for integration into automated decision-making systems and agricultural IoT platforms. Nonetheless, geographic and methodological limitations are acknowledged, suggesting the need for multi-regional validations and exploration of hybrid architectures in future research.