Satellite-driven prediction of fine particulate matter (PM 2.5 ) concentrations: machine learning and explainable artificial intelligence

Abstract Accurate estimation of fine particulate matter (PM 2.5 ) is essential for air quality management, particularly in regions with sparse ground-monitoring networks. This study provides a systematic review of literature published between 2015 and 2025, synthesizing advancements in integrating satellite-derived data with machine learning (ML) and deep learning (DL) for PM 2.5 concentration prediction. The review evaluates different satellite products, including aerosol optical depth (AOD) from low Earth orbit and geostationary platforms, as well as top-of-atmosphere reflectance and Ångström exponent. This study also highlights the importance of auxiliary datasets, such as meteorological variables, land-use indicators, demographic, and infrastructural factors, in reducing predictive bias. Methodological developments in data fusion and data preprocessing, specifically regarding data normalization and handling missing AOD values, are identified as main factors affecting the model stability. Comparisons among the model performance reveal that conventional ML models (e.g. Random Forest) remain robust benchmarks, whereas advanced DL architectures like hybrids DL models and transformer-based frameworks achieve better accuracy by capturing complex spatiotemporal (ST) dependencies. Furthermore, this study explores the emerging role of explainable artificial intelligence (XAI) in enhancing model transparency and bridging the gap between data-driven results and physical atmospheric laws. This study concludes by identifying several challenges, such as gaps of satellite data and limited model transferability, and proposes future research directions focusing on ST XAI and physics-guided learning. This review serves as a roadmap for researchers and policymakers to develop interpretable and high-resolution monitoring systems for air quality.