FE-Kriging A Feature-Enhanced Spatiotemporal Model for PM2.5 Exposure Assessment with Environmental Equity Implications in Chengdu

Long-term exposure to spatially heterogeneous PM2.5 concentrations threatens urban public health and environmental equity. While spatial dynamic weighted kriging methods have been widely adopted for PM2.5 exposure assessment, their accuracy is limited in complex geographic settings because of the inadequate integration of multisource pollution characteristics, dynamic weighting mechanisms, and physical diffusion constraints. To address these gaps, this study proposes a feature-enhanced kriging (FE-kriging ) model, which synergizes three innovative modules: (1) random forest-based feature selection to identify key PM2.5drivers, (2) dynamic spatial attenuation weights (via KNN-constructed matrices) to capture spatiotemporal dependencies, and (3) Gaussian diffusion constraints to align predictions with industrial emission physics. When applied to Chengdu’s PM2.5 data, FE-Kriging achieved superior performance (MSE=0.1228, RMSE=0.3504, MAE=0.2036, R²=0.9935), outperforming traditional kriging by 12.3% in R². Leveraging high-precision exposure estimates, we further quantified demographic disparities via an age‒sex dual-weighting system. Spatial analyses revealed pollution hotspots (Getis-Ord Gi*), whereas bivariate Moran’s I decoupled socioeconomic-natural determinants of exposure heterogeneity. District-level GDP-exposure elasticity models highlight the marginal impacts of economic growth, and stratified regression reveals inequities across demographic groups. Finally, city/district Gini coefficients provide scalable metrics for environmental equity assessment. This study advances PM2.5exposure modelling through physics-informed machine learning and offers policymakers a unified framework to target intervention priorities—balancing pollution control with equitable health outcomes.