中国东部沿海四省一市PM2.5浓度遥感估算方法研究

卫星遥感反演的气溶胶光学深度（AOD）产品已被广泛应用于近地面PM_2.5浓度的估算。已有研究表明通过构建AOD和PM_2.5之间的高级统计模型—线性混合效应模型（LME）可以有效获取近地面PM_2.5浓度的空间分布，但由于引入了大量的气象和土地利用等因子，使得模型对变量的解译能力有所降低。为此，基于MODIS AOD（空间分辨率：3 km），以我国东部长江三角洲—福建—广东（YRD-FJ-GD）为研究区，构建了两种非参数机器学习模型，即支持向量机（SVM）和随机森林（RF）模型，来估算2018年YRD-FJ-GD地区的近地面PM_2.5浓度，并将其与线性混合效应模型（LME）的估算结果进行对比。研究发现，3种模型估算的PM_2.5浓度与地面实测值之间的R²均高于0.6，其中，RF模型的估算精度最优，模型拟合的R²高达0.91，比SVM模型（R²=0.79）和LME模型（R²=0.64）的估算结果分别提高了13%和30%；且RMSE（~9.07 μg/m³）也远低于LME（~19.09 μg/m³）和SVM模型（~17.29 μg/m³）。此外，由随机森林（RF）模型估算的2018年YRD-FJ-GD地区的PM_2.5空间分布显示，长江三角洲（YRD）地区的年均PM_2.5浓度最高（>46 μg/m³），其次为广东省（GD），福建地区（FJ）的年均PM_2.5浓度最低（<37 μg/m³）；4个季节的平均PM_2.5浓度则呈现冬季（46.32 μg/m³）>春季（38.80 μg/m³）>秋季（36.15 μg/m³）>夏季（30.16 μg/m³）的分布格局。研究结果表明：与高级统计模型（LME）和机器学习（SVM）相比，随机森林（RF）模型能更好地应用于YRD-FJ-GD地区的PM_2.5浓度估算。

Research on Methods of Remotely Sensed PM2.5 Concentrations Estimation in Four Provinces and One City along the East Coast of China

The Aerosol Optical Depth （AOD） derived from remote sensing imageries has been widely used in estimating ground-level PM_2.5 concentrations in large areas. Previous studies that focused on PM_2.5 estimation have reported high predictability of PM_2.5 concentrations when using AOD and the advanced statistical model （i.e.， Linear Mixed Effects model （LME））. However， the interpretation ability of the LME model was lowered， as it introduced many meteorological and land use variables in the model， and the importance of each variable to PM_2.5 concentrations was hard to interpret. Therefore， this study developed two nonparametric machine learning methods， i.e.， Support Vector Machine （SVM） and Random Forest （RF）， to estimate the ground-level PM_2.5 concentrations. The eastern Yangtze River Delta-Fujian-Guangdong （i.e.， YRD-FJ-GD） region in China was employed as our study case， and we also compared the predictability of these two models with the LME model. The results showed that the overall R² between estimated and observed PM_2.5 concentrations exceeded 0.6 for three models， where RF received a R² of 0.9， i.e.， 13% and 30% higher than SVM （R²=0.79） and LME （R²=0.64） model， respectively. The RMSE values were 9.07， 17.29 and 19.09 μg/m³ for RF， SVM and LME model， respectively. In addition， the spatial distribution of PM_2.5 concentrations estimated from the optimal model （i.e.， RF） illustrated high annual PM_2.5 in YRD （>46 μg/m³）， and GD ranked the second. FJ exhibited a relatively low annual PM_2.5 （<37 μg/m³）. The seasonal PM_2.5 concentrations presented a distribution pattern as winter （6.32 μg/m³） > spring （38.80 μg/m³） > autumn （36.15 μg/m³） > summer （30.16 μg/m³）. Our results revealed that the AOD and RF model could be a good proxy for estimating PM_2.5 concentrations in YRD-FJ-GD region.