基于极端梯度提升树和深度学习方法估算中国地表NO_(2)浓度  

Estimating ground-level NO_(2)over China by using extreme gradient boosting and deep learning method

作　　者：王思晨 霍彦峰[1,3,4] 穆溪 江鹏 朱立[2] 荀尚培 何彬方[1,3,4] 吴文玉 WANG Sichen;HUO Yanfeng;MU Xi;JIANG Peng;ZHU Li;XUN Shangpei;HE Binfang;WU Wenyu(Key Laboratory for Atmospheric Sciences&Remote Sensing of Anhui Province,Anhui Institute of Meteorological Sciences,Hefei 230031;School of Resources and Environmental Engineering,Anhui University,Hefei 230601;Shouxian National Climate Observatory,Huainan 232200;Huaihe River Basin Typical Farmland Ecological Meteorological Field Science Experiment Base of Meteorological Administration,Huainan 232200)

机构地区：[1]安徽省气象科学研究所,安徽省大气科学与卫星遥感重点实验室,合肥230031 [2]安徽大学,资源与环境工程学院,合肥230601 [3]寿县国家气候观象台,淮南232200 [4]中国气象局淮河流域典型农田生态气象野外科学试验基地,淮南232200

出　　处：《环境科学学报》2023年第10期298-308,共11页Acta Scientiae Circumstantiae

基　　金：风云卫星应用先行计划(No.FY-APP-2022.0603);国家自然科学基金区域创新发展联合基金(No.U21A2028);安徽省气象局创新发展专项(No.YJG202203);国家高技术研究发展计划无场地定标关键技术项目(No.8-060011)。

摘　　要：二氧化氮(NO_(2))是备受关注的重要大气污染物之一,与人体呼吸系统和心血管疾病有紧密关系.卫星遥感是获得大尺度NO_(2)分布情况的有效方法,搭载于Aura卫星上的臭氧监测仪(Ozone Monitoring Instrument,OMI)可以反演全球尺度的对流层NO_(2)柱浓度.然而,由于观测条件(如云覆盖)和传感器物理异常影响,OMI在中国地区存在1/2以上的缺失数据,严重限制了数据的应用价值.本文首先基于深度学习方法重建OMI对流层NO_(2)柱浓度缺失数据,然后结合气象资料和地面信息(如道路密度)等数据,利用梯度提升树模型估算了2018—2020年中国近地面NO_(2)浓度日均值,最后使用机器学习解释性算法评估了OMI数据对近地面NO_(2)估算的适用性和敏感性.结果表明:OMI数据缺失值的重建效果和近地面NO_(2)估算精度良好,OMI缺失数据重建值与原始数据的交叉验证R^(2)为0.81,近地面NO_(2)浓度估算值与中国环境总站监测值交叉验证R^(2)为0.84;气象要素对近地面NO_(2)的敏感性最高,特征重要度为36.7%,OMI对流层NO_(2)柱浓度的特征重要度约为8%.Nitrogen dioxide(NO_(2))is a significant atmospheric pollutant closely associated with respiratory and cardiovascular diseases.Satellite remote sensing has proven to be an effective approach for acquiring the large-scale spatial distribution of NO_(2).The ozone monitoring instrument(OMI),mounted on the Aura satellite,enables global retrieval of tropospheric NO_(2)column concentrations.However,the application of OMI data in China is severely limited due to the presence of missing values,which account for over half of the dataset,primarily caused by observation conditions such as cloud cover and sensor anomalies.This study addresses this challenge by firstly reconstructing the missing values of OMI tropospheric NO_(2)columns using a deep learning method.Subsequently,by incorporating meteorological data and ground information(e.g.,road density),a gradient boosting tree model is employed to estimate the daily mean near-surface NO_(2)concentrations over China from 2018 to 2020.Finally,the applicability and sensitivity of the OMI data in estimating near-surface NO_(2)are evaluated using machine learning interpretability algorithms.The results demonstrate a favorable performance in reconstructing the missing OMI data and estimating near-surface NO_(2)concentrations.The cross-validated R^(2)between the reconstructed missing values of OMI and the original data is 0.81,while the cross-validated R^(2)between the estimated near-surface NO_(2)concentrations and the measurements from the China National Environmental Monitoring Station is 0.84.Meteorological factors exhibit the highest sensitivity to near-surface NO_(2)concentrations,accounting for 36.7%of the feature importance,while the feature importance of OMI tropospheric NO_(2)column concentrations is approximately 8%.

关 键 词：二氧化氮 臭氧监测仪(OMI) 深度学习 缺失值重建 NO_(2)浓度反演 

分 类 号：X831[环境科学与工程—环境工程] X87