Long-term spatiotemporal evolution, relative hotspot persistence, and predictive attribution of PM2.5 across China.

China's PM2.5 pollution has declined markedly since 2013, but whether its pace, spatial hierarchy, and predictive structure changed concurrently remains unclear. We combined the 1-km ChinaHigh PM2.5 product for 2000-2024 with homogenized ground observations at 1269 fixed sites and analysed segmented trends, meteorological normalization evaluated at randomly held-out sites, fixed-population exposure, repeated spatial-block validation of global XGBoost and a block-centred geographically weighted variant (GeoXGBoost), and cross-fitted SHAP effects. The national mean fell from 48.20 μgm-3 in 2013 to 24.83 in 2024, while the Theil-Sen decline slowed from -3.489 μgm-3yr-1 during 2014-2018 to -1.030 during 2019-2024. Meteorological normalization preserved this ordering. Although 75.2-88.2% of early hotspot area remained in the corresponding late-period upper tail, fixed-2020-population exposure decreased by 19.57 μgm-3 between 2000-2004 and 2020-2024. Across 40 held-out folds from eight spatial partitions, mean R2 was 0.208 for XGBoost and 0.175 for GeoXGBoost; the partition-level paired difference was 0.033 (bootstrap 95% confidence interval, 0.015-0.048). Forward-year R2 ranged from 0.794 to 0.842 with the year index and from -0.228 to 0.275 without it. Low spatial scores may reflect clustered monitoring, environmental differences among held-out regions, and missing spatially resolved processes. Spatial out-of-fold SHAP rankings nevertheless replicated for ground observations (Spearman ρ = 0.983), and the strongest interactions linked temperature with humidity and latitude. These findings separate persistent spatial ranking from declining absolute exposure and identify geographic, temporal, and response-level stress tests as safeguards for generalizing predictive PM2.5 attribution.