Machine-learning-based estimation of ground-level NO₂ concentrations across Southeastern Europe using multi-source...

Reliable estimation of ground-level nitrogen dioxide (NO₂) remains challenging due to limited monitoring coverage and complex interactions among emissions, meteorology, and land-use factors. This study develops a machine-learning framework to estimate surface NO₂ concentrations across Southeastern Europe, covering thirteen countries with diverse emission sources and sparse monitoring networks. Daily NO₂ observations from 317 monitoring stations were integrated with multiple satellite and ancillary datasets, including tropospheric NO₂ from Sentinel-5P TROPOMI, vegetation indices from MODIS, and nighttime light data from VIIRS. Additional inputs included meteorological variables from ERA5, anthropogenic emissions from EDGAR v8.1, land-use data from Copernicus Corine Land Cover, elevation from the Copernicus Digital Elevation Model, and population density from the Global Human Settlement Layer. All datasets were processed and harmonized using Google Earth Engine. Three Gradient Boosting Decision Tree models (XGBoost, LGBM, and CatBoost) were trained and optimized with Optuna. Model performance was assessed using 10-fold cross-validation and independent test sets with R², MAE, and RMSE metrics. All models performed consistently, with LGBM achieving the best results (R² = 0.86; MAE = 5.34 μg/m³; RMSE = 8.32 μg/m³). SHapley Additive exPlanations (SHAP) identified key predictors and improved interpretability. Seasonal analysis showed higher accuracy in summer and autumn than in winter and spring. These results demonstrate that integrating multi-source data with advanced machine learning enables reliable surface NO₂ estimation in regions with limited monitoring, supporting air quality assessment, exposure analysis, and evidence-based policymaking in Southeastern Europe.