<![CDATA[Air pollution due to fine particulate matter (PM2.5) has emerged as a critical public health concern in tropical megacities, where rapid urbanization often outpaces environmental regulation. Indonesia, characterized by limited air quality monitoring networks and complex meteorological conditions, remains particularly vulnerable to extreme pollution episodes. This study investigates the spatio-temporal dynamics of extreme PM2.5 concentrations in three Indonesian cities—Kemayoran, Semarang, and Malang—selected for their contrasting urban morphologies, topographic settings, and meteorological regimes. Using hourly PM2.5 observations from 2022 to 2024 combined with ERA5 reanalysis data, this study examines the role of atmospheric conditions in governing pollution variability across space and time. A hybrid modeling framework was implemented by integrating multiple linear regression with nonlinear machine learning algorithms, namely Random Forest and XGBoost, to predict hourly PM2.5 concentrations based on meteorological variables including temperature, dew point, wind speed, precipitation, and surface pressure. The results indicate that machine learning models outperform linear methods, with Random Forest providing a strong balance between predictive accuracy and interpretability. Wind speed emerged as the most consistent predictor, followed by dew point and precipitation, exhibiting notable spatial and seasonal variability. Extreme PM2.5 episodes, defined as hourly concentrations exceeding the 95th percentile, were most frequent during dry and transitional seasons under stagnant, humid, and low-rainfall conditions. Kemayoran recorded the highest concentrations, while Malang, despite lower emission levels, demonstrated vulnerability due to weak atmospheric ventilation. Overall, the findings highlight the dominant role of meteorological stagnation in driving PM2.5 extremes and demonstrate the potential of hybrid modeling approaches for developing localized early-warning systems in tropical urban environments.]]>
