Machine learning-based PM2.5 estimation using NASA’s MERRA-2 and comparison of ML- and SHAP-based meteorological normalization in Thailand

Fine particulate matter (PM 2.5 ) poses severe health risks in Thailand; however, sparse ground monitoring has left long-term trends poorly quantified. This study utilizes NASA’s MERRA-2 reanalysis to estimate historical PM 2.5 levels across Thailand through two approaches: an empirical formulation based on PM 2.5 chemical components (MERRA2_EF) and a machine-learning model (MERRA2_ML) utilizing the Light Gradient Boosting Machine algorithm. Validated against 58 reference air quality stations across Thailand (2012–2022), the MERRA2_ML model significantly outperforms the MERRA2_EF, increasing temporal correlation from 0.48 to 0.84, while reducing MAE from 14.99 to 8.64 µg m⁻³ and RMSE from 24.95 to 13.39 µg m⁻³. Similar improvements in model performance were also observed across the different regions of Thailand, with the MERRA2_ML estimates generally showing stronger temporal agreement and lower errors than the MERRA2_EF. MERRA2_ML substantially improved the representation of observed PM 2.5 variability across Thailand, accurately capturing diurnal peaks, dry-season enhancements, and interannual fluctuations, particularly in biomass-burning- and urban-influenced regions, compared with MERRA2_EF. Long-term analysis revealed significant increases in PM 2.5 across Thailand (1.97 µg m ⁻3 per decade) and most regions from 1981 to 2022, with MERRA2_ML showing consistently lower and more moderate trends than MERRA2_EF, highlighting substantial spatial heterogeneity in long-term PM 2.5 changes. By employing Shapley Additive Explanation (SHAP) for meteorological normalization, this study isolated emission-driven PM 2.5 variations. Overall, decreasing trends in emission-related PM 2.5 were observed across Thailand and several regions; however, statistical significance varied by region and method, and the non-significant national trends do not provide sufficient evidence to attribute the observed decline to a sustained nationwide reduction in emissions or policy interventions. These findings demonstrate that ML-enhanced reanalysis provides robust, spatially explicit data to support evidence-based public health decisions and air quality management throughout Southeast Asia.