An interpretable stacking ensemble model for monitoring wheat stripe rust severity by integrating solar-induced chlorophyll fluorescence and hyperspectral indices

Wheat stripe rust represents a major threat to global wheat production and food security, rendering early diagnosis and accurate severity assessment essential for containing its spread and minimizing yield losses. This study developed an interpretable ensemble learning framework for monitoring wheat stripe rust severity, based on a stacking regression approach integrated with SHapley Additive exPlanations (SHAP) for model interpretation. The stacking model incorporated Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM) as base learners and linear regression as the meta-learner to establish relationships between disease severity and ground-based canopy-scale predictors. These predictors included solar-induced chlorophyll fluorescence (SIF), hyperspectral vegetation indices (VIs), and canopy physiological-biochemical traits derived from wheat canopies under stripe rust stress. Across five-fold cross-validation, the stacking ensemble consistently outperformed all individual base learners, improving the coefficient of determination (R 2 ) by 2.81% and reducing the root mean square error (RMSE) by 7.46% relative to the best single model. Using SIF alone as input yielded an R 2 of 0.511, whereas integrating SIF, VIs, and canopy physiological parameters increased R 2 to 0.878, corresponding to an absolute gain (ΔR 2 ) of 0.367 and a relative increase of 71.8%, demonstrating the added value of integrating multi-source spectral and physiological information. Global SHAP analysis indicated that chlorophyll content (Cab) and a SIF yield component (SIFy A ) were the most influential predictors throughout disease progression, whereas the near-infrared reflectance of vegetation (NIRv) and other structural indices gained importance during mid-to-late stages of canopy degradation. Stage-specific SHAP interaction analysis further revealed distinct nonlinear interaction patterns among SIF, structural metrics, and physiological parameters under mild, moderate, and severe disease conditions, indicating a stage-dependent, coordinated response of wheat photosynthetic physiology to stripe rust stress. Overall, the proposed interpretable ensemble framework not only enables high-accuracy, remote sensing-based monitoring of stripe rust severity, but also quantitatively elucidates the synergistic mechanisms by which multi-source spectral and physiological information encodes disease progression.