Ground-based air quality monitoring in Hanoi, Vietnam, remains limited by sparse spatial coverage, hindering the characterization of fine-scale PM2.5 variability. This study develops a machine learning framework for hourly PM2.5 estimation by integrating Himawari-8/9 satellite observations with ERA5 meteorological reanalysis data. Ground-based PM2.5 measurements from three monitoring stations in Hanoi were used as reference observations, and three ensemble learning algorithms—LightGBM, XGBoost, and Random Forest—were trained using cloud-filtered daytime observations. Among the evaluated models, XGBoost achieved the best performance, with an R² of 0.56, an RMSE of 22.10 μg/m³, and an MAE of 15.54 μg/m³. The model captured temporal variations in PM2.5 across a broad concentration range, from approximately 10 to over 100 μg/m³, including major pollution episodes. The remaining unexplained variance likely reflects sub-grid-scale variability associated with localized emission sources, including traffic, industrial activities, and open burning, as well as land-use heterogeneity within the approximately 10 km × 10 km satellite retrieval grid. These results demonstrate the feasibility of combining geostationary satellite observations and meteorological reanalysis data with machine learning for hourly PM2.5 estimation in Hanoi. Future work will incorporate near-real-time NOAA GFS forecasts and low-cost sensor networks to improve temporal continuity and capture hyper-local spatial variability.