基于分布式感知深度神经网络的高分辨率PM2.5值估算

细颗粒物（PM_2.5）个体暴露水平是健康效应研究中的关键问题,然而历史数据缺失和地面监测点覆盖范围小阻碍了相关研究。基于美国国家航空航天局遥感数据提供的气溶胶光学厚度（AOD）,融合地面监测、气象等多源数据进行建模来估算近地面PM_2.5浓度,所得结果的空间覆盖范围广、时间连续性强、方法成本低。本文基于2018年京津冀鲁地区,引入气象、NDVI、时间节点、空间标识等50个特征分析AOD-PM_2.5关系。鉴于传统插补方法单一所造成的信息损失,运用时空多视图插补方法来提高插补的精度和广度。考虑到特征的滞后作用、特征间相关性与偏相关性所导致的复杂关系,运用分布式感知深度神经网络模型来分别捕捉多源特征间的高阶特性。结果表明：① 时空多视图插补方法的相对误差为27.5%,数据平均缺失52.1%降至4.84%。② 分布式感知深度神经网络模型在时间预测上平均绝对误差、相对误差、均方误差、均方根误差分别为17.7 μg/m ³、46.8%、766.2 μg ²/m ⁶、26.9 μg/m ³,空间上,为16.6 μg/m ³、41.8%、691.5 μg ²/m ⁶、26.6 μg/m ³,从精度、稳健性、泛化能力和耗时方面综合来看,结果优于线性统计模型和常见深度学习架构。

High resolution PM2.5 estimation based on the distributed perception deep neural network model

The individual exposure level of fine particulate matter(PM2.5) is a key issue in the study of health effects. However, the lack of historical data and the small coverage of ground monitoring stations have hindered the development of related research. Based on the aerosol optical depth(AOD) provided by NASA remote sensing data, multi-source data such as ground monitoring and meteorological data were integrated for modeling to estimate near-ground PM2.5concentration. The results have wide spatial coverage, strong time continuity and low method cost. Based on the Beijing-Tianjin-Hebei-Shandong region in 2018, this paper introduces 50 features such as meteorological elements, NDVI, time nodes and spatial markers to analyze the relationship between AOD and PM2.5. In view of the information loss caused by the single traditional interpolation method, the spatiotemporal multi-view interpolation method is used to improve the accuracy and coverage of interpolation. Considering the complex relationship caused by hysteresis of features, and the correlation and partial correlation between features,this paper uses a distributed perception deep neural network model(DP-DNN) to separately capture higher-order features between multiple-source features. The results show that:(1) The relative error of the spatiotemporal multi-view interpolation method is 27.5%, and the average proportion of missing data decreases from 52.1% to 4.84%.(2) In terms of time prediction,mean absolute error, relative error, mean square error and root mean square error of DP-DNN are 7.7 μg/m³, 46.8%, 766.2 μg²/m⁶ and 26.9 μg/m³, respectively. In terms of space prediction,they are 16.6 μg/m³, 41.8%, 691.5 μg²/m⁶ and 26.6 μg/m³, respectively. In aspects of accuracy,robustness, generalization ability and time consuming, the results are superior to linear statistical models and common deep learning architecture.