Ensemble machine learning-based PM2.5 modeling using hotspot counts (0-1000 km) reflecting Chiang Mai, Thailand’s extreme pollution

Abstract

Persistent local and transboundary smog has critically elevated PM2.5 levels in Northern Thailand over the past decade, resulting in significant health risks. The spatial distribution of hotspot counts, indicative of biomass burning and smoke dispersion, demonstrates a strong correlation with PM2.5 concentration patterns, underscoring the importance of incorporating such data into air quality analyses. This study integrates hotspot data to capture both temporal dynamics and external influences in PM2.5 prediction models. The importance of lagged hotspot counts within 100–1000 km of Chiang Mai—ranked as the world’s most polluted city during the study period—and lagged ground-level PM2.5 is assessed using Lasso regularization. The analysis reveals that the cumulative effects of hotspots extend their influence on air quality in Chiang Mai up to approximately 700 km. Advanced tree-based ensemble machine learning methods, including Random Forest (RF), Gradient Boosting (GB), and Extreme Gradient Boosting (XGBoost), are implemented alongside the Long Short-Term Memory (LSTM) deep learning model to evaluate their predictive performance. This approach provides a novel framework for PM2.5 modeling in Northern Thailand. Five key features with specific day lags were identified for modeling. These include PM2.5 at lag 1, short-range hotspots within 100 km at lags 1 to 3, mid-range hotspots at 200 and 400 km at lags 2 to 4, and long-range hotspots beyond 700 km at lag 5. Incorporating hotspot data improved model performance by approximately 20%, as evidenced by error metrics and residual analysis. Among the models tested, GB outperformed XGBoost, RF, and LSTM, achieving the highest R² (0.97), lowest RMSE (5.49), MAE (2.08), and MAPE (5.8%), along with near-zero MBE and minimal MdAE (0.48). Statistical validation confirmed the model’s reliability with no significant bias.