Spatial Heterogeneity of PM_2.5 Influencing Factors in China: A Study Integrating Geographically Weighted Regression and K-Nearest Neighbors

In meteorology and many other fields, spatial data heterogeneity will render stationary spatial models ineffective. To describe the spatial heterogeneity of China's air quality data in January 2024 precisely, this paper proposes a residual prediction method based on geographically weighted regression (GWR) and the k-nearest neighbors (KNN) algorithm-the GWR-KNN method. The method first applies GWR for preliminary fitting in an effort to model the non-stationary relationship between the spatial data and its relating independent variables. It subsequently adopts the KNN algorithm for local prediction of residuals from the GWR model with the aim of capturing residual local patterns without sacrificing computational efficiency and also enhancing model prediction accuracy. Simulation studies demonstrate that compared to the standard and stationary models, the improved method more effectively captures spatial heterogeneity in the data. Empirical analysis using locally regression-fitted coefficients reveals significant regional variation in the impacts of P M₁₀, S O₂, NO₂, CO₂, and O₃ concentrations on PM_2.5 concentrations in China in January 2024, showing pronounced spatial heterogeneity. In summary, the proposed wide-area regression-KNN model not only effectively explains the heterogeneous nature of spatial data like PM2.5 concentrations but also has greater predictive power than traditional static models in both simulation and empirical study. The approach provides region-specific policies for air pollution control in China.