Spatiotemporal modeling of long-term PM2.5 concentrations and population exposure in Greece, using machine learning and statistical methods

Abstract

The lack of high-resolution, long-term PM_2.5 observations in Greece and the Eastern Mediterranean hampers the development of spatial models that are crucial for providing representative exposure estimates to health studies. This work presents a spatial modeling approach to address this gap and assess PM_2.5 spatial variability for the first time on a national level in Greece, by integrating in situ observations, meteorology, emissions and satellite AOD data among others. A high-resolution (1 km²) gridded dataset of PM_2.5 concentrations across Greece from 2015 to 2022 was developed, and seven statistical, machine learning, and hybrid models were evaluated under different prediction scenarios. Random Forest (RF) models demonstrated superior performance, (R² = 0.73, MAE = 2.2 μg m^-3), validated against ground-based measurements. Winter months consistently showed the highest PM_2.5 levels, averaging 16.8 μg m^-3, over the domain, due to residential biomass burning (BB) and limited atmospheric dispersion. Summer months had the lowest concentrations, averaging 10.3 μg m^-3, while substantial decreases nationwide were observed during the 2020 COVID-19 lockdown. Population exposure analysis indicated that the entire Greek population was exposed to long-term PM_2.5 concentrations exceeding the WHO air quality guideline (AQG) of 5 μg m^-3. Moreover, the dataset revealed elevated PM_2.5 levels across several regions of mainland Greece. Notably, 70 % to 90 % of the population experience levels exceeding 10 μg m^-3 in Central and Northern regions of continental Greece like Thessaly, Central Macedonia, and Ioannina. The Ioannina region, which is severely impacted by residential BB, recorded pollution levels up to five times the WHO AQG highlighting the urgent need for targeted interventions. The high-resolution RF model's superior performance for monthly average concentrations, compared to the Copernicus Atmosphere Monitoring Service (CAMS) dataset, renders it a reliable tool for long-term PM_2.5 assessment in Greece that can support air quality management and health studies.

Keywords: Air quality; Artificial intelligence; Particulate matter - PM(2.5); Population exposure; Random Forest; Spatial model.