Interpretable machine learning for spatially differentiated nonlinear relationships between PM2.5 and landscape patterns in an urban-rural integrated zone in China.

Understanding the spatial heterogeneity of fine particulate matter (PM2.5) is critical for differentiated air-quality management in rapidly urbanizing regions. This study investigates spatially differentiated nonlinear associations between PM2.5 and landscape patterns in an urban-rural integrated zone of Xuchang, China. Digital elevation models, normalized difference vegetation indices, building footprints, and land-use/land-cover data are integrated with morphological spatial pattern analysis and FRAGSTATS metrics to characterize grey-green spaces (GGS). A Random Forest model with SHAP interpretation is used to quantify the relative importance, direction, and nonlinear effects of these critical factors, leading to the identification of three spatial patterns along urban-rural and topographic gradients. In high-elevation mountainous areas, elevation shows the strongest association with PM2.5, consistent with enhanced atmospheric dispersion and lower anthropogenic emissions. In low-elevation and high-density urban cores, mean building height exhibits an apparent inflection range of 35-40 m, above which its positive association with PM2.5 is strengthened. In peri-urban and county-town transition zones, GGS configuration is predominant: green-space porosity ranges of 0.18-0.22 in plains and 0.12-0.15 in hilly areas, together with greater blue-space connectivity, are associated with lower PM2.5 concentrations. These findings provide spatially explicit diagnostic references for differentiated air-quality governance, highlighting terrain and ecological continuity in mountainous areas, building layout and ventilation in dense urban cores, and green-space porosity and landscape connectivity in transition zones. The identified thresholds should be validated using independent datasets before being considered for planning standards.