Spatiotemporal characteristics and driving factors of coal-burning hotspots in Shandong Province revealed by multi-source data and machine learning.

Coal combustion emissions significantly contribute to air pollution in China, especially in the residential sector, where they are widely dispersed and poorly tracked, making accurate identification difficult for traditional methods. This study proposes a framework combining the Deep Self-Attention Fusion Model (DSAF) with the Isolation Forest algorithm (IForest) to identify coal-burning hotspots using SO2 as a tracer. The framework was applied to analyze the spatiotemporal distribution of daily SO2 concentrations at 1 km × 1 km resolution in Shandong Province from 2022 to 2024 and to identify coal-burning hotspots. The DSAF model integrates satellite remote sensing, ground monitoring, and meteorological data for high-precision SO2 concentrations estimation with complete spatiotemporal coverage. The IForest algorithm combined with multidimensional spatial features identifies areas of anomalously high SO2 concentrations and generates a provincial-scale hotspots distribution map. Spatial patterns were characterized by integrating industrial facility locations and urban-rural data. Results indicate that the annual average SO2 concentration decreased from 10.15 μg/m3 in 2022 to 8.73 μg/m3 in 2024 (14.00% reduction), showing clear seasonal variations with higher concentrations in winter and spring. Non-heating season hotspots are primarily located in chemical, steel, and heavy industrial clusters, highly consistent with industrial facility distributions. During the heating season, hotspots coverage significantly expands, with notable increases in rural areas, indicating additive effects of civil bulk coal use on top of industrial emissions. The methods developed in this study provide technical support for refined coal emission management and coordinated control of secondary pollutants such as PM2.5.