Abstract Considering the spatial non-stationary contributions of environment variables to PM2.5 variations, the geographically weighted regression (GWR) modeling method has been using to estimate PM2.5 concentrations widely. However, most of the GWR models in reported studies so far were established based on the screened predictors through pretreatment correlation analysis, and this process might cause the omissions of factors really driving PM2.5 variations. This study therefore developed a best subsets regression (BSR) enhanced principal component analysis-GWR (PCA-GWR) modeling approach to estimate PM2.5 concentration by fully considering all the potential variables' contributions simultaneously. The performance comparison experiment between PCA-GWR and regular GWR was conducted in the Beijing-Tianjin-Hebei (BTH) region over a one-year-period. Results indicated that the PCA-GWR modeling outperforms the regular GWR modeling with obvious higher model fitting- and cross-validation based adjusted R2 and lower RMSE. Meanwhile, the distribution map of PM2.5 concentration from PCA-GWR modeling also clearly depicts more spatial variation details in contrast to the one from regular GWR modeling. It can be concluded that the BSR enhanced PCA-GWR modeling could be a reliable way for effective air pollution concentration estimation in the coming future by involving all the potential predictor variables’ contributions to PM2.5 variations.

中文翻译：

---

一种改进的地理加权回归模型，用于大面积 PM 2.5 浓度估计

摘要 考虑到环境变量对PM2.5变化的空间非平稳贡献，地理加权回归（GWR）建模方法已被广泛用于估计PM2.5浓度。然而，迄今为止报道的研究中的大部分 GWR 模型都是基于通过预处理相关分析筛选出的预测因子建立的，这一过程可能会导致遗漏真正驱动 PM2.5 变化的因素。因此，本研究开发了一种最佳子集回归 (BSR) 增强主成分分析-GWR (PCA-GWR) 建模方法，通过同时充分考虑所有潜在变量的贡献来估计 PM2.5 浓度。PCA-GWR 与常规 GWR 的性能对比实验在京津冀 (BTH) 地区进行了一年的时间。结果表明，PCA-GWR 建模优于常规 GWR 建模，具有明显更高的基于模型拟合和交叉验证的调整 R2 和更低的 RMSE。同时，与常规 GWR 建模相比，PCA-GWR 建模的 PM2.5 浓度分布图也清楚地描绘了更多的空间变化细节。可以得出结论，BSR 增强的 PCA-GWR 建模可能是未来有效空气污染浓度估计的可靠方法，因为它涉及所有潜在的预测变量对 PM2.5 变化的贡献。结果表明，PCA-GWR 建模优于常规 GWR 建模，具有明显更高的基于模型拟合和交叉验证的调整 R2 和更低的 RMSE。同时，与常规 GWR 建模相比，PCA-GWR 建模的 PM2.5 浓度分布图也清楚地描绘了更多的空间变化细节。可以得出结论，BSR 增强的 PCA-GWR 建模可能是未来有效空气污染浓度估算的可靠方法，因为它涉及所有潜在预测变量对 PM2.5 变化的贡献。结果表明，PCA-GWR 建模优于常规 GWR 建模，具有明显更高的基于模型拟合和交叉验证的调整 R2 和更低的 RMSE。同时，与常规 GWR 建模相比，PCA-GWR 建模的 PM2.5 浓度分布图也清楚地描绘了更多的空间变化细节。可以得出结论，BSR 增强的 PCA-GWR 建模可能是未来有效空气污染浓度估计的可靠方法，因为它涉及所有潜在的预测变量对 PM2.5 变化的贡献。