Spatial Patterns and Nonlinear Associations Between Agricultural Comprehensive Risk and Grain Production Capacity in China: Evidence from Explainable Machine Learning

Climate change, more frequent extreme events, and tightening resource and environmental constraints are increasing uncertainty in grain production, making agricultural risk assessment essential for food security. Yet high realized production capacity does not necessarily imply low risk, and output-based assessments may obscure climatic, ecological, and socioeconomic vulnerabilities. Using a panel of 4140 city-year observations for 180 Chinese prefecture-level cities (2000–2022), we construct an Agricultural Comprehensive Risk Index (ACRI) spanning climate–meteorological exposure, ecological–environmental conditions, and socioeconomic vulnerability, and examine its joint spatial patterns and evolution with grain production capacity. Explainable machine learning (XGBoost + SHAP) identifies key predictors, nonlinear effects, interactions, and regional heterogeneity. We find pronounced spatial mismatches and asynchronous dynamics between risk and capacity. High-capacity, medium-to-high-risk areas concentrate in northeast, north, and parts of central/eastern China, while parts of southwest and western China show low capacity with medium-to-high risk. Ecological–environmental variables contribute 77.6% of global SHAP importance, with strongly nonlinear associations and non-additive interactions involving thermal, vegetation, and productivity factors. Over the sample period, predictions grow more dependent on thermal, ecological, and selected drought- and flood-related features, with clear differences between northern and southern China. These findings show that realized grain production capacity cannot substitute for agricultural risk assessment. Food-security governance should therefore evaluate capacity and risk jointly and target region-specific constraints.