Estimating South Asian PM 2.5 (2017–2023) via Transfer Learning of a Model Trained on Chinese MERSI/FY-3 Satellite Data

Against the backdrop of accelerating global industrialization and urbanization, fine particulate matter (PM_2.5) remains a critical determinant in air quality management. This study introduces an innovative ML-CN+ML-SA hybrid transfer learning model to estimate spatiotemporal distributions of ground-level PM_2.5 concentrations across South Asia from 2017 to 2023. In this article, we integrate MERSI/FY-3 apparent reflectance (TOA) data, ground observation data et al, and use the XGBoost algorithm to build transfer learning frameworks. Compared with the standalone ML-CN, the ML-CN+ML-SA model significantly improving correlation with ground measurements from 0.55 to 0.8. Subsequent analysis reveals significant seasonal modulation of PM_2.5 concentrations by monsoon dynamics, exhibiting marked reductions during summer monsoon periods contrasted with pronounced wintertime accumulation. The research results of this paper will help South Asian countries and other areas with lack of ground station data to effectively estimate ground PM_2.5 and provide research direction, and also help the relevant governments in South Asian countries to provide reference for air pollution control policies.