Explainable machine learning for joint estimation of evapotranspiration and gross primary productivity in a winter wheat cropland in Northwest China

Under climate change, accurate estimation of actual crop evapotranspiration (ETc act) and gross primary productivity (GPP) is essential for understanding ecosystem water–carbon coupling and improving agricultural water management. Traditional models mainly rely on meteorological and structural variables but often fail to represent physiological processes such as photosynthesis and stomatal regulation. In this study, four consecutive years of joint observations were conducted in a winter wheat field in Yangling, Northwest China, using eddy covariance, tower-based solar-induced chlorophyll fluorescence (SIF), and meteorological sensors. SIF was integrated into a machine learning framework to estimate water–carbon fluxes, and the Shapley additive explanations (SHAP) method was used to quantify factor contributions across phenological stages. The inclusion of SIF improved model accuracy. For ETc act, R2 increased by 0.02 and RMSE decreased by 0.03 mm day−1. For GPP, R2 increased by 0.06 and RMSE decreased by 1.01 g C m−2 day−1. Among all models, the long short-term memory model performed best, achieving R2 values of 0.88 and 0.91 and RMSE values of 0.85 mm day−1 and 0.81 g C m−2 day−1 for ETc act and GPP, respectively. SHAP analysis indicated that net radiation was the dominant driver of ETc act, while the contribution of SIF increased with crop development. For GPP, SIF was the most important predictor in almost all stages. Taken together, integrating SIF with explainable machine learning enhanced both estimation accuracy and understanding of coupled water–carbon processes, providing support for precision irrigation and sustainable crop management.