Machine learning-driven regional prediction of PM2.5 concentrations in the eastern mediterranean bridging spatial data gaps in air quality monitoring

Fine particulate matter (PM_2.5) posing significant risks due to its ability to penetrate deep into the respiratory system. This study introduces the Regional PM_2.5 Predictor (RPP), a machine learning-based framework designed to estimate PM_2.5 concentrations across Turkiye, especially in regions with limited PM_2.5 monitoring infrastructure. Leveraging satellite-derived Aerosol Optical Thickness (AOT) data, meteorological variables from ERA-5, and ground-based air quality measurements, the model integrates diverse datasets spanning 2018 to 2023, the RPP employs XGBoost algorithms to address spatial monitoring gaps. The model demonstrates strong predictive performance across multiple evaluation scenarios: the seasonal analysis yielded RMSE values of 4.39–10.01 μg/m³ and R² values of 0.66–0.84; temporal evaluations achieved an average RMSE of 8.28 μg/m³ and R² of 0.76; spatial (station-blinded) cross-validation maintained reliable predictions with average RMSE of 9.21 μg/m³ and R² of 0.71; while random sampling achieved RMSE of 6.82 μg/m³ and R² of 0.85 with an 80-20 % split. The framework successfully captured Turkiye's air quality trend, with PM_2.5 levels decreasing from 25.52 μg/m³ (2018) to 18.88 μg/m³ (2023), while identifying performance variations across diverse topographical regions. The model demonstrated remarkable stability during the COVID-19 pandemic period, achieving its best performance in 2020 (RMSE: 7.54 μg/m³, R²: 0.80). This approach demonstrates how machine learning can complement traditional monitoring networks, providing cost-effective air quality assessments for public health interventions and environmental policy evaluation.