Seasonal water deficits can constrain food provision in humid-subtropical agricultural regions, but their spatial patterns are difficult to diagnose when grain statistics are reported by administrative units and drought is represented by a single index. We developed a mass-constrained and interpretable framework to diagnose county-statistics-constrained relative food-provision shortfall (FPL) at 1-km resolution in the Heng–Shao drought corridor, China. County grain production from 2013 to 2024 was allocated using target-year cropland area and a target-year-excluded, cropland-conditioned Landsat Enhanced Vegetation Index (EVI) reference, while FPL was defined relative to a target-year-excluded county-level Theil–Sen statistical reference. Compound drought signals were represented by vapour-pressure-deficit anomaly (D_vpd), soil-moisture anomaly (D_soil), and climatic-water-deficit anomaly (D_cwd); their associations with FPL were evaluated using spatially blocked Categorical Boosting (CatBoost), cross-fitted SHapley Additive exPlanations (SHAP) and TreeSHAP. FPL was strongly right-skewed and spatially localised, with an uncapped mean of 0.036 and a skewness of 4.047. CatBoost achieved pooled out-of-fold coefficients of determination (R²) of 0.329–0.340 across 10–20 km spatial blocks, and adding the joint hydroclimatic predictor set comprising Dvpd, Dsoil and Dcwd increased R² by 0.166 beyond the background predictors. Cropland-landscape and hydroclimatic variables formed the two largest SHAP-attributed predictor groups, accounting for 34.67% and 33.72% of total attribution, respectively. The three hydroclimatic indicators also exhibited differentiated nonlinear responses and recurrent joint-attribution structures. These findings show that retaining complementary hydroclimatic dimensions alongside cropland-landscape context provides a spatially explicit diagnostic basis for characterising within-county relative food-provision shortfall.