Accurate quantification of crop evapotranspiration (ETc) is crucial for effective agricultural water management and climate-adaptive production. Despite advancements in estimation methods—from simplified models to data-driven technologies—achieving high-precision estimations remains a significant challenge. This study systematically evaluated eight ETc estimation methods, including the FAO dual crop coefficient method, the AquaCrop model, three machine learning models, and three coupled models, to assess the differences in prediction accuracy and robustness across various modeling approaches. Additionally, a multi-source coupled modeling framework integrating residual learning and physical constraints was proposed to address the limitations of physical models, which suffer from structural bias, and the high data dependency and training complexity of data-driven models. The results showed that the dual crop coefficient method performed less accurately than the mechanistically interpretable AquaCrop model (R² = 0.901), primarily due to its simplified representation of the crop-soil system dynamics. While the pure data-driven CNN-LSTM model approximated the AquaCrop model’s performance when sufficient data was available (R² = 0.893), its generalization ability deteriorated significantly with limited data (R² dropped to 0.640), highlighting its dependence on large datasets. In contrast, the coupled models, which incorporated physical priors and residual learning, leveraged physical constraints to reduce the mapping space required for deep learning fitting. This approach reduced reliance on large training datasets and decreased training cycles. By combining the structural knowledge of crop models with the nonlinear capabilities of machine learning at both the feature and output levels, the accuracy and robustness of the models were significantly improved. Notably, the connected embedded coupling model (CECM) achieved the best performance (R² = 0.924). This study demonstrates that synergistic modeling of physical mechanisms and data-driven approaches provides an ETc estimation pathway that balances interpretability with high predictive accuracy, offering valuable support for precision irrigation and agricultural water resource management.
扫码关注我们
求助内容:
应助结果提醒方式:
