Abstract
Accurately quantifying water and carbon fluxes is essential for assessing the hydroclimatic vulnerability of semi-arid ecosystems. However, conventional data-driven models often lack transparency in capturing complex non-linear flux responses to environmental constraints. In this study, we employed five machine learning algorithms (XGBoost, CatBoost, Random Forest, LightGBM, and NGBoost) coupled with SHapley Additive exPlanations (SHAP) to simulate daily gross primary production (GPP), evapotranspiration (ET), net ecosystem exchange (NEE), and ecosystem respiration (Reco) at a specific semi-arid site by integrating satellite-derived solar-induced chlorophyll fluorescence (SIF) and leaf area index (LAI) with local meteorological and soil variables. Integrating multi-source remote sensing data substantially improved model performance in capturing the spatiotemporal variability of GPP and ET (R² = 0.72–0.80), reducing centered root mean square difference (CRMSD) for GPP simulations by approximately 20%. Among the evaluated algorithms, CatBoost achieved the highest accuracy for GPP and Reco, while XGBoost performed optimally for ET and NEE. However, overall predictive accuracies for Reco and NEE remained lower than those for GPP and ET. Rather than inferring unobserved ecophysiological mechanisms, SHAP analysis extracted quantitative benchmarks, revealing a strong predictive complementarity: SIF emerged as the foremost functional predictive driver for GPP beyond a zero-crossing threshold of 0.24 W m−2 μm−1 sr−1, whereas LAI governed ET as a fundamental structural constraint, shifting to a promotional role at a strict threshold of 1.32 m2 m−2. Furthermore, sample-level clustering analysis objectively demonstrated that extreme high-flux events are non-linearly co-determined by specific quantitative combinations of SIF and LAI. Ultimately, this study demonstrates the potential of interpretable machine learning in disentangling the non-linear roles of functional and structural vegetation indices. The derived quantitative benchmarks provide a rigorous, data-driven framework for evaluating site-specific ecosystem responses to hydroclimatic stressors.