Integration of Satellite and Ground-Based Data for High-Resolution PM2.5 Assessment and Hotspot Identification Using Machine Learning

High-resolution air quality monitoring is critical for effective pollution management in urban as well as rural areas. India’s sparse and expensive air monitoring infrastructure fails to capture localized PM2.5 dynamics for targeted actions, while low-cost alternatives have data quality concerns. Satellite observations have data gaps due to cloud cover, unfavorable meteorological conditions, and retrieval failures. To address these monitoring gaps, this study integrates data from reference-grade monitors, low-cost sensors (LCS), and satellite aerosol products, with meteorological and land-use features, employing machine learning (ML) to predict daily PM2.5 concentrations at high spatial resolution (1 km × 1 km) in Delhi. We employed a multistage ML approach involving LCS calibration, AOD gap-filling, and PM2.5 estimation incorporating meteorological parameters such as relative humidity (RH), temperature (T), planetary boundary layer height (PBLH), and land-use features such as normalized difference vegetation index (NDVI), road density (RD), and population density (PD). Random Forest evolved as the best-performing model for LCS calibration with R2 of 0.85–0.94 for most of the sites (85%) and a lower RMSE (13.45–54.58 μg/m3). The AOD filling approach enhanced the data points by 15.32% using the RF model with R2 of 0.97. In PM2.5 estimation, Extreme Gradient Boosting (XGB) outperformed RF with R2 and RMSE of 0.77 and 38.37 μg/m3, respectively, considering all the parameters. Further, high-resolution (1 km × 1 km) PM2.5 hotspots were identified at monthly, seasonal, and annual time scales for Delhi, providing hyperlocal insights to support targeted policy interventions. The proposed framework provides high-resolution air quality information that can be adapted in other Indian cities with similar monitoring challenges.