Machine Learning-Based Prediction of Guayule (Parthenium argentatum) Evapotranspiration Under Flood Irrigation and Arid Conditions

Accurate prediction of daily actual crop evapotranspiration (ETa) is crucial for effective irrigation scheduling and sustainable water resources management. Recently, machine learning (ML) has expanded to predict nonlinear environmental systems like ETa. Here, 40 ML-based ETa models were developed using ten input scenarios of daily standard agro-meteorological variables to predict guayule daily ETa at Maricopa and Eloy field sites in Pinal County, Arizona. Four ML models, Random Forest (RF), Light Gradient Boosting Machine (LGBM), Categorical Boosting (CatBoost), and Artificial Neural Network (ANN), were trained and tested against soil water balance (SWB)-derived ETa. Predictions were assessed using the coefficient of determination (R2), mean bias error (MBE), and normalized root-mean-squared error (RMSEn). Feature importance scores (FISs) for each input variable in predicting ETa were also determined. Mean temperature (Tmean) and reference evapotranspiration (ETo) were the most significant parameters affecting guayule daily ETa at both sites, with Tmean remaining dominant when ETo was excluded. During testing, LGBM overestimated guayule ETa by 0.473–0.704 mm·day−1 with R2 of 0.906–0.960 and RMSEn of 14.504–21.863%, making it the most suitable choice to predict guayule daily ETa at both sites. Findings suggest ML-based ETa models are promising for precision irrigation scheduling in water-scarce arid regions.