Comparative Analysis of Crop Coefficient Approaches and Machine Learning Models for Predicting Water Requirements in Three Major Crops in Coastal Saline-Alkali Land

The accuracy of the crop coefficient approaches recommended by the FAO-56 guidelines for evapotranspiration (ET) in saline environments is limited due to complex soil–water–crop interactions, highlighting the need for advanced methods to improve ET estimation for water management in saline-alkali lands. To improve ET estimation for wheat, maize, and soybean in the Yellow River Delta, China, three machine-learning algorithms—gradient-boosting decision tree (GBDT), random forest (RF), and extreme gradient-boosting regression (XGBoost)—were applied alongside single- and dual-crop coefficient approaches (Kc-ETo). The results showed that increasing the input variables did not necessarily improve the ML model performance. The ML models outperformed Kc-ETo approaches, particularly for summer crops (maize and soybean), with the mean absolute error reduced by 26.4% to 80.9%, R2 increased by 5.6% to 11.2%, and root mean square error (RMSE) decreased by 22.4% to 98.1%. RF and XGBoost were more accurate than GBDT, with R2 increasing by 3.2% to 5.4% and RMSE decreasing by 22% to 57%. Scenario simulations showed increased ET with intensified emission scenarios for RF and GBDT, similar to Kc-ETo approaches. However, XGBoost simulated a significantly lower ET in high-emission scenarios, indicating potential unreliability for scenario predictions beyond the training dataset, especially in a saline-alkali environment with an increasingly complex background.