An optimized multi-source feature extraction model with black-winged kite algorithm for hourly seamless PM2.5 estimation

Fine particulate matter (PM_2.5) estimation is becoming of significant importance for public health protection and air quality management. However, it is difficult to extract effective data features and easily lead to data missing with single-source dataset. And the multi-source data PM_2.5 estimation method will also increase data dimensionality and nonlinear data coupling. In this paper, a novel hourly seamless PM_2.5 estimation method is proposed with leveraging multi-source data integration through attentive interpretable tabular learning network (TabNet)-based feature extraction and the categorical boosting (CatBoost) prediction model. Firstly, a multi-source dataset is constructed for the satellite and reanalysis data fusion to solve the missing spatio-temporal data problem and to realize the hourly seamless estimation. Secondly, the TabNet structural parameters are optimized by the black-winged kite algorithm (BKA), effectively extracting key feature information from the multi-source data. And then the hourly seamless PM_2.5 concentration is estimated by CatBoost with Bayesian optimization (BO), which improves the generalization performance of the whole model. Finally, the multi-source dataset of Beijing-Tianjin-Hebei (BTH) region is used for the experimental analysis and validation. The results demonstrate that BKA has superiority over other algorithms in optimizing TabNet feature extraction. The proposed method exhibits excellent performance in hourly seamless PM_2.5 estimation, with the R² of the day and night models reaching 0.91 and 0.92 and the RMSE as low as 12.22 μg/m³ and 11.70 μg/m³, which are better than other regression models. In addition, the hourly PM_2.5 estimated are generally consistent with observations and the applicability across various spatial locations is validated.