基于机器学习和数据融合技术的高分辨率PM2.5浓度估算研究

The method taking advantage of satellite-derived aerosol optical depth (AOD) to estimate PM2.5 concentration has become the primary technique toward a gridded PM2.5 concentration record. Nevertheless, it is recognized that the prior studies have suffered from several drawbacks and limitations in PM2.5 modeling. Firstly, the parameters used in PM2.5 modeling are diversified and always vary among studies. In addition, certain critical parameters such as nitrogen dioxide and sulfur dioxide being highly linked to PM2.5 emissions are rarely considered. Secondly, the AOD products used by most studies suffer from low spatiotemporal resolutions and significant data gaps, which in turn further limit the resolution and spatial coverage of PM2.5 distribution. Thirdly, the regression models established through linear theory can be biased due to the collinearity between parameters. Lastly, the linear methods are less effective in exploring complex interactions between PM2.5 and the related contributing factors as the interactions are always highly nonlinear. .In order to address the aforementioned drawbacks and to improve the resolution and accuracy of the gridded PM2.5 concentration estimations, this study aims at integrating state-of-the-art data fusion and data mining techniques to advance the estimation of satellite-based PM2.5 concentrations. Specifically, this study is to (1) optimize the combination of parameters used for PM2.5 modeling, (2) to develop a data fusion scheme being capable of blending multi-sensor AOD products, as well as (3) to establish robust AOD-PM2.5 models with the optimized parameter combinations and fused AOD products by applying advanced machine learning tools. The PM2.5 concentrations derived from such an integrated scheme would have better spatiotemporal resolutions and higher accuracy with improved spatial coverage, and can then be used as a valuable data source to advance the PM2.5 related studies such as the prevention and control of PM2.5 pollution and the associated health effects.

利用卫星遥感AOD数据估算PM2.5浓度，已成为目前空间大尺度PM2.5浓度格点产品获取的主要手段。针对历史研究在建模过程中存在：(1)变量复杂多样且较少考虑表征污染物排放情况的关键大气参数，(2)AOD产品时空分辨率低且数据缺失严重，以及(3)线性方法不能有效挖掘PM2.5浓度及其关联因子间复杂的非线性时空耦合关系等问题，本研究拟通过开展：(1)PM2.5浓度建模参数组合优化，(2)星地多源AOD数据融合，与(3)基于机器学习算法的高精度PM2.5浓度估算研究，以构建新型高分辨率PM2.5浓度估算技术方法体系，旨在突破现有方法和数据对PM2.5浓度估算精度、时空覆盖范围和分辨率的限制，生成具有高时空分辨率特征的空间大尺度PM2.5浓度格点产品。该方法体系的构建将有望提升和改善PM2.5浓度产品的估算精度和时空适用性，为我国PM2.5污染控制及其健康效应研究提供重要的数据支撑。

为有效利用卫星遥感手段开展时空大尺度 PM2.5 质量浓度估算，生产具有高时空分辨率特征的PM2.5 浓度格点资料，本项目先后围绕PM2.5 浓度建模参数筛选、模型结构优化以及AOD 缺失数据重构等方面开展了一系列研究工作，主要研究内容包括：（1）系统性梳理历史文献并对比不同建模参数或方法应用下的 PM2.5 浓度估算精度，定量评估了不同建模方法和输入变量对PM2.5 浓度估算精度的影响，并遴选出具有较高建模精度的方法以及关键建模参数；（2）针对空间大尺度PM2.5浓度估算中关键变量 AOD 数据因大面积数据缺失造成估算的 PM2.5 浓度产品空间不连续问题，项目组发展了一种顾及时相变化特征的缺失数据重构算法，并将其分别应用于重构站点尺度逐小时PM2.5浓度以及面域AOD 数据产品中的缺失值，进而提升上述产品的时空覆盖度；（3） 针对大尺度 PM2.5 浓度卫星遥感估算，项目组对比分析了不同建模方案生产的 PM2.5 浓度的精度及空间细节质量，进而提出了适合精细化 PM2.5 浓度制图的建模方案。. 具体研究结论如下：（1）引入AOD，气象场和PM2.5时空自相关变量、使用细模态AOD产品、开展AOD产品标高订正和湿度订正都可显著提升PM2.5浓度估算精度；相比之下，提升AOD 产品空间分辨率或PM2.5 浓度产品的时间分辨率对PM2.5浓度估算精度影响不大。其次，相比于其他建模方法，使用混合效应模型，时空地理加权回归、随机森林和深度学习模型可获得较高的估算精度。（2）发展的迭代经验正交函数分解与重构方法可有效恢复PM2.5 和 AOD 数据产品中存在的数据缺失，且具有较高的重构精度，是生产时空连续的 PM2.5 浓度格点资料的重要技术手段。（3）通过对我国地面观测 PM2.5 浓度资料开展均一化检验和订正，生产了一套长时序均质化地基 PM2.5 浓度数据资料。同时，基于优化方案，生产了一套覆盖我国的时空连续的PM2.5 浓度格点资料。. 具体应用实践表明基于本研究提出的优化模型具有较低的模型复杂度但仍具备较高的建模精度，相关研究结果可为后续高精度PM2.5浓度估算及模型优化提供理论参考，发展的缺失数据重构方法可为生产时空连续PM2.5格点资料提供技术支撑。