Explainable Multi-Source Machine Learning for PM₂.₅ Mapping Across Indonesian Cities: Integrating Satellite, Meteorological, Land-Surface and Fire Information

Abstract

Background: Fine particulate matter (PM₂.₅) is a critical indicator of atmospheric environmental quality, yet its spatial assessment in Indonesia remains constrained by the limited and uneven distribution of ground-monitoring stations. Objectives: This study develops an explainable multi-source machine-learning framework for estimating daily PM₂.₅ concentrations across Indonesian cities by integrating ground observations with satellite-derived aerosol optical depth, meteorological reanalysis, land-surface characteristics, and satellite-detected fire activity. Method: Ground-level PM₂.₅ observations are harmonized with MODIS MAIAC aerosol data, ERA5-Land meteorological variables, VIIRS fire information, and selected geographic covariates. Linear baseline models, Random Forest, and gradient-boosted trees are evaluated using temporal-block validation and leave-one-city-out cross-validation to distinguish temporal predictive performance from geographic transferability. SHapley Additive exPlanations (SHAP) are used to identify predictor contributions, while spatial residual diagnostics are employed to detect systematic geographic prediction errors. Results: In the present simulated analytical scenario, the dataset comprised 3,854 station-days from six Indonesian cities during 2023–2024. The best-performing XGBoost model achieved an out-of-sample R² of 0.79, RMSE of 7.8 µg m⁻³, and MAE of 5.6 µg m⁻³ under temporal validation. Under leave-one-city-out validation, performance decreased to R² = 0.64, RMSE = 10.5 µg m⁻³, and MAE = 7.7 µg m⁻³. Satellite AOD, planetary boundary-layer height, relative humidity, wind speed, and lagged fire radiative power were the dominant predictors. Conclusion: The proposed framework provides a transparent approach for extending PM₂.₅ assessment beyond existing monitoring locations while explicitly evaluating geographic generalizability and model interpretability.