Satellite-based retrieval has become a popular PM_2.5 monitoring method. To improve the retrieval performance, multiple variables are usually introduced as auxiliary variables, in addition to aerosol optical depth (AOD). The different kinds of variables are usually at different resolutions, varying from sub-kilometer to dozens of kilometers. Generally speaking, when undertaking the retrieval, the variables at different resolutions are resampled to the same resolution as the AOD product to ensure scale consistency (single-scale retrieval). However, a drawback of doing this is that the information contained within the different resolutions (scales) is discarded. To fully utilize the information contained in the different scales, a dual-scale retrieval approach is proposed in this paper. In the first stage, the variables which influence PM_2.5 concentration at a large scale are used for PM_2.5 retrieval at a coarse resolution. Then, in the second stage, the variables which affect PM_2.5 distribution at a finer scale are used for the further PM_2.5 retrieval at a high resolution (sub-km level resolution), with the retrieved low-resolution PM_2.5 from the first stage also acting as input. In this study, four different regression models were adopted to test the performance of the dual-scale retrieval approach at both daily and annual scales: multiple linear regression (MLR), geographically weighted regression (GWR), random forest (RF), and the generalized regression neural network (GRNN). Compared with the traditional single-scale retrieval approach, the proposed dual-scale retrieval approach can achieve PM_2.5 mapping at a finer resolution and with a higher accuracy. Dual-scale retrieval can utilize the information contained in different scales, thus achieving an improvement in both resolution and retrieval accuracy. The proposed approach has the potential to be used for the generation of quantitative remote sensing products in various fields, and will promote the quality improvement of these quantitative remote sensing products.

基于卫星的检索已成为流行的PM _2.5监控方法。为了提高检索性能，除了气溶胶光学深度（AOD）之外，通常还引入多个变量作为辅助变量。不同种类的变量通常具有不同的分辨率，从亚公里到几十公里不等。一般而言，进行检索时，将不同分辨率的变量重新采样为与AOD产品相同的分辨率，以确保刻度一致性（单刻度检索）。但是，这样做的缺点是丢弃了包含在不同分辨率（比例）内的信息。为了充分利用不同尺度下所包含的信息，本文提出了一种双尺度检索方法。在第一阶段，影响PM _2.5的变量大规模浓缩可用于粗略分离PM _2.5。然后，在第二阶段中，将影响较小尺度PM _2.5分布的变量用于高分辨率（亚千米级分辨率）的进一步PM _2.5检索，而检索到的低分辨率PM _2.5从第一阶段起也充当输入。在这项研究中，采用了四种不同的回归模型来测试双尺度检索方法在日尺度和年尺度上的性能：多元线性回归（MLR），地理加权回归（GWR），随机森林（RF）和广义回归神经网络（GRNN）。与传统的单尺度检索方法相比，该双尺度检索方法可以达到PM _2.5以更高的分辨率和更高的精度进行贴图。双尺度检索可以利用不同尺度中包含的信息，从而提高分辨率和检索精度。所提出的方法有可能用于各个领域的定量遥感产品的产生，并将促进这些定量遥感产品的质量改进。