Towards Satellite-Based PM₂.₅ Estimation: Exploring AOD–PM₂.₅ Relationships and Key Drivers through Statistical and Machine Learning Approaches

Fine particulate matter (PM2.5) remains one of the most critical pollutants affecting air quality, public health, ecosystem protection and climate from a One Health perspective. Although ground-based monitoring networks represent the reference framework for air quality assessment, their heterogeneous spatial distribution limits the characterization of PM2.5 variability across multiple scales.
Satellite-derived Aerosol Optical Depth (AOD) is widely used as a proxy for atmospheric aerosol loading; however, the relationship between columnar AOD and surface PM2.5 is affected by meteorological conditions, aerosol vertical distribution, emission regimes and local characteristics. This study investigates the AOD–PM2.5 relationship over Italy during 2019 using MODIS Terra and Aqua MAIAC AOD products, ERA5 meteorological variables from the Copernicus reanalysis, and hourly in situ PM2.5 measurements from the European Environment Agency (EEA) monitoring stations.
A multiple linear regression model (MLR) was initially developed as a baseline and compared with Random Forest (RF), XGBoost and Support Vector Regression (SVR) approaches. Models based on conventional EEA station classifications showed limited predictive capability (R² < 0.15), whereas seasonal and station-specific MLR formulations including ERA5 variables substantially improved performance, reaching seasonal R² values of 0.30–0.62. Machine learning models provided improvements. Globally, RF slightly outperformed MLR (R² = 0.107 vs. 0.105), while SVR and XGBoost showed lower performances. However, RF achieved better results in suburban areas and during winter, whereas SVR performed best during summer and autumn. At the local scale, nonlinear approaches showed advantages, with XGBoost improving winter predictions at the Museo Nazionale station in Naples (R² = 0.333 vs. 0.165 for MLR).
These findings highlight that machine learning approaches can provide localized and season-dependent benefits for satellite-based PM2.5 estimation, particularly where nonlinear AOD–PM2.5 relationships occur. This initial assessment will be further developed by increasing the dataset to evaluate model performance under different environmental conditions.