Spatiotemporal continuous PM2.5 concentrations inversion based on multisource data and hybrid model

Abstract

Air pollution affects public health, destroys ecological environment and even aggravates climate change, and PM2.5 is an important factor causing pollution. Therefore, accurate and continuous spatiotemporal PM2.5 concentration inversion is of great significance for air pollution control. Nevertheless, current high temporal resolution inversions are predominantly conducted using Top-of-atmosphere reflectance, nighttime concentration inversion cannot be realized in the visible band. Moreover, existing hybrid models often rely on linear integrated, increasing the risk of overfitting. Therefore, we propose an improved stacking model that contains XGBoost and CatBoost model, stacked by elastic net regression. By fusing features from multi-source data and accounting for spatiotemporal heterogeneity, the model enables the simultaneous inversion of daytime and nighttime PM2.5 concentrations. Compared to using Top-of-atmosphere reflectance alone, the daytime inversion results demonstrate a 1.08% increase in R 2 and a reduction of 1.34µg/m 3 in RMSE. Meanwhile, the nighttime inversion results show a 2.04% improvement in R 2 and reduce the MAE to 9.54µg/m 3 when spatiotemporal features are used. Applying the model to the Beijing-Tianjin-Hebei region enables spatiotemporal continuous inversion of PM2.5 concentrations in the region.