Cross-City Transfer Learning for Sentinel-5P-Driven NO₂ Prediction in Data-Sparse Urban Environments

Abstract. Traditional forecasting methods of air pollutants show intrinsic limitations due to the complexity of atmospheric interactions. Recent research has moved toward the employment of artificial intelligence (AI)-based approaches and satellite data processing. The framework proposed in this study is a transfer learning (TL) model to estimate surface-level NO₂ concentrations across multiple locations by using satellite and environmental data. The approach integrates Sentinel-5P TROPOMI-derived tropospheric NO₂ columns, meteorological variables (temperature, precipitation, wind speed and direction), spatial coordinates and temporal features. A CatBoost regression model is implemented, leveraging a Leave-One-City-Out (LOCO) TL framework across five cities (Berlin, London, Madrid, Paris and Toronto) in the world. This enables the model transfer from multiple source domains to a new target city with minimal ground-based data. Experimental results are outperforming city-specific baseline models, by showing a reduced Root Mean Square Error (RMSE) by approximately 7% and a Coefficient of Determination (R²) higher by 2.7%. Toronto, which represents an environment with a low monitoring density, benefits most from TL, with R² improving from 0.58 (baseline) to 0.66 (transfer) and RMSE dropping from 6.44 μg/m³ to 5.84 μg/m³. A detailed Leave-One-Block-Out (LOBO) ablation study shows how each group of features contributes to the performance of the model. Spatial coordinates and meteorological features are the most influential predictors of NO₂ concentration, while the satellite NO₂ data increase model generalization. These results highlight the potential of cross-city TL and remote sensing synergy for scalable urban air pollution monitoring, especially in limited ground-based monitoring scenarios.