This study used four individual machine learning (ML) models (random forest, adaptive boosting, gradient boosting, and extreme gradient boosting), and a stacked ensemble model (SEM) for PM_2.5 estimation over Greater Bangkok (GBK) during the dry season for 2018–2022. Aerosol optical depth (AOD) from Fengyun-4A satellite was used as the main predictor variable. The other predictor variables include meteorological variables, fire hotspots, vegetation index, terrain elevation, and population density. Surface PM_2.5 from 17 air quality monitoring stations was used for model development and evaluation. Satellite AOD aligns reasonably well with AOD from two AERONET stations in the study area in terms of correlation coefficient (r), mean bias (MB), mean error (ME), and root mean square error (RMSE). Among the individual models, adaptive boosting performed the best with r = 0.75, MB = 0.55 µg m⁻³, ME = 9.1 µg m⁻³, and RMSE = 12.9 µg m⁻³. As for SEM which comprises all individual models, it outperformed every individual model, with r = 0.84, zero MB, ME = 7.2 µg m⁻³, and RMSE = 10.4 µg m⁻³. In two additional cases of haze hours and clean hours, SEM is best overall while adaptive boosting is superior to the other individual ML models. The case of haze hours has lower model predictability, suggesting elevated PM_2.5 is difficult to predict. SEM was thus chosen to map PM_2.5 as well as exposure intensity over GBK. Good agreement between the observed and predicted diurnal and monthly patterns is achieved by every model. PM_2.5 tends to be relatively high at 08–10 LT and declines in later hours, corresponding to higher traffic emissions in the morning and daytime meteorological conditions more favorable to dilute air pollutants, respectively. PM_2.5 intensifies in the winter but decreases in March and April. During these two months, the areas outside Bangkok tend to have higher PM_2.5 than within Bangkok, possibly linked to active summertime biomass burning in those areas that are less urbanized with more agricultural lands. Relatively high exposure intensity is constrained to Bangkok due likely to its much denser population. The findings indicate a significant potential for leveraging the Fengyun-4A satellite data and ML to advance space-based air quality monitoring for Thailand and other data-scare regions in Southeast Asia. A satellite-based PM_2.5 dataset could support the formulation of effective air quality management strategies in GBK.

本研究使用四种单独的机器学习 (ML) 模型（随机森林、自适应提升、梯度提升和极限梯度提升）和堆叠集成模型 (SEM) 来估计大曼谷 (GBK) 旱季期间的PM _2.5 2018–2022。风云四号A卫星的气溶胶光学深度（AOD）被用作主要预测变量。其他预测变量包括气象变量、火灾热点、植被指数、地形海拔和人口密度。来自 17 个空气质量监测站的地表 PM _2.5用于模型开发和评估。卫星 AOD 在相关系数 ( r )、平均偏差 (MB)、平均误差 (ME) 和均方根误差 (RMSE) 方面与研究区域两个 AERONET 站的 AOD 相当吻合。在各个模型中，自适应增强表现最好，r  = 0.75，MB = 0.55 µg m ⁻³，ME = 9.1 µg m ⁻³，RMSE = 12.9 µg m ⁻³。对于包含所有单独模型的SEM，它优于每个单独模型，r  = 0.84，MB为零，ME = 7.2 µg m ^-3，RMSE = 10.4 µg m ^-3。在雾霾时间和清洁时间的另外两种情况下，SEM 总体上是最好的，而自适应增强则优于其他单独的 ML 模型。雾霾时段的模型预测能力较低，表明 PM _2.5升高难以预测。因此选择 SEM 来绘制 PM _2.5以及 GBK 上的暴露强度。每个模型都实现了观察到的和预测的每日和每月模式之间的良好一致性。 PM _2.5在 08-10 LT 时往往相对较高，并在晚些时候下降，分别对应于早晨较高的交通排放和白天的气象条件更有利于稀释空气污染物。 PM _2.5在冬季增强，但在三月和四月减弱。在这两个月中，曼谷以外地区的 PM _2.5浓度往往高于曼谷境内，这可能与城市化程度较低、农业用地较多的地区夏季生物质燃烧活跃有关。曼谷的暴露强度相对较高，可能是因为曼谷人口密集得多。研究结果表明，利用风云四号A卫星数据和机器学习来推进泰国和东南亚其他数据匮乏地区的天基空气质量监测具有巨大潜力。基于卫星的 PM _2.5数据集可以支持 GBK 制定有效的空气质量管理策略。