A spatially robust and explainable machine learning framework for 100 m grid-level estimation of urban PM2.5 and PM10 using multi-source earth observation data

Accurate spatial estimation of urban particulate matter (PM) is important for exposure assessment, environmental monitoring, and air-quality management, particularly in data-scarce urban regions. Conventional monitoring networks are spatially sparse, and satellite-based PM downscaling studies that rely on random validation can substantially overestimate predictive performance because of spatial dependence. This study developed an explainable machine learning framework for generating 100 m grid-level estimates of summer PM 2 . 5 and PM 10 across Tehran by integrating multi-source Earth observation data, meteorological reanalysis, remotely sensed surface indices, and GIS-derived predictors. Two complementary feature-selection strategies—Random Forest (RF) importance and SHAP-based explainable artificial intelligence—were evaluated with Random Forest, XGBoost, and LightGBM. Model performance and spatial generalization were assessed using regular K-fold cross-validation, Leave-One-Station-Out cross-validation, and Spatial Block Cross-Validation (SBCV). Feature selection reduced the predictor space by 76–79% while maintaining comparable or improved performance under spatially independent validation. RF-selected predictors provided more consistent spatial generalization for PM 2 . 5 , whereas RF- and SHAP-selected configurations performed comparably for PM 10 . Regular K-fold cross-validation produced 17–28% lower error estimates than spatially independent validation, highlighting the importance of spatial validation when assessing performance at unmonitored locations. SHAP-based analysis identified elevation as the most influential model-attributed predictor for PM 2 . 5 , while temperature, surface pressure, and boundary-layer height were among the most influential predictors for PM 10 . These attributions were relatively stable across spatial partitions and random seeds but are interpreted as associations within the fitted models rather than causal effects. The resulting 100 m estimates revealed a pronounced north–south gradient, with the highest concentrations in southern industrial and high-traffic districts. Overall, the study demonstrates the value of jointly evaluating predictive performance, feature selection, model-based interpretability, and spatial generalization when developing urban PM estimation models. The framework provides a methodological basis for grid-level air-quality estimation in data-scarce urban environments, while its broader applicability requires external validation and local recalibration.