Revealing the Nonlinear Relationships Between Gray-Green Spaces and PM2.5 Pollution Through Explainable Machine Learning: A Case Study of Nanjing

Urban PM2.5 pollution is influenced not only by emissions and meteorological conditions but also by gray-green spatial patterns. However, the joint and nonlinear relationships of built-up and green-space configurations with PM2.5 concentrations remain insufficiently understood. Taking Nanjing as a case study, we constructed eight composite indicators spanning coverage intensity, fragmentation and shape complexity, aggregation and connectivity, and morphological structure, and combined Random Forest, SHAP, and partial dependence analysis with nested spatial block cross-validation to examine summer and winter PM2.5 concentrations in 2024. The results showed that: (1) mean concentrations were 20.83μg/m3 in summer and 47.21μg/m3 in winter, with higher values in central areas. (2) GCI and GMSF ranked first and second in normalized SHAP relative importance in both seasons; BCI ranked third in summer and BACI in winter. Predicted concentrations decreased as GCI increased from approximately −1.0 to 0.5 standardized units before leveling off, while GMSF declined across the intermediate range and rebounded slightly at higher values. (3) SHAP values were predominantly negative when GCI and GMSF exceeded the sample-centered reference. These findings indicate that PM2.5 responses are associated with joint gray-green configurations and provide an interpretable basis for comparing spatial configurations in Nanjing.