High-Resolution Estimation of Particulate Matter Concentrations based on Sentinel-2 Satellite Imagery Using Machine Learning and Deep Learning Approaches

Air pollution has intensified globally with the acceleration of industrialization and urbanization, posing significant threats to human health and ecosystem stability. In particular, particulate matter (PM) is a major hazardous component that has been strongly linked to the development and worsening of respiratory and cardiovascular diseases. PM concentrations exhibit substantial spatial variability even over short distances due to differences in emission sources, meteorological conditions, and land-use characteristics. Therefore, high-resolution spatial monitoring is essential for accurate exposure assessment, particularly in densely populated and environmentally vulnerable areas. However, existing ground-based monitoring networks are spatially unevenly distributed, thereby constraining their capacity to accurately represent fine-scale spatial variability and localized high-concentration events in urban environments. To address these limitations, this study aims to estimate PM concentrations at a spatial resolution of 20 m using high-resolution satellite imagery, thereby complementing ground-based observations and enabling detailed characterization of local PM variability and hotspots. Sentinel-2 satellite data were integrated with multiple ancillary datasets, including meteorological variables, a digital elevation model, and land cover information, to estimate PM concentrations. Multiple machine learning and deep learning algorithms were implemented and systematically compared, and XGBoost was identified as the optimal model. Model performance was evaluated using multiple statistical metrics. The results demonstrated high predictive performance for both PM10 and PM2.5 concentrations. For PM10, the model achieved a mean absolute error (MAE) of 6.684㎍/㎥, a root mean square error (RMSE) of 11.132㎍/㎥, and a determination of coefficient (R²) of 0.887. Similarly, PM2.5 estimation yielded an MAE of 4.094㎍/㎥, an RMSE of 6.648㎍/㎥, and an R² of 0.841. These findings confirm the feasibility and effectiveness of generating high-resolution PM concentration maps using Sentinel-2 satellite data. This study provides a robust framework for detailed assessment of urban-scale PM spatial distributions and offers valuable baseline data for population exposure assessment and the development of targeted air quality management policies.