基于Sentinel-5P卫星反演京津冀地区近地面NO_(2)浓度研究  

Retrieval of Near-Surface NO_(2)Concentration in Beijing-Tianjin-Hebei Region Based on Sentinel-5P Satellite

作　　者：周刚[1,2] 徐彬仁 徐炜达 程文晨 尹文君 王凌 ZHOU Gang;XU Binren;XU Weida;CHENG Wenchen;YIN Wenjun;WANG Ling(Department of Automation,Tsinghua University,Beijing 100084,China;Beijing Insights Value Technology Co.,Ltd.,Beijing 100070,China)

机构地区：[1]清华大学自动化系,北京100084 [2]北京英视睿达科技股份有限公司,北京100070

出　　处：《环境科学研究》2024年第6期1223-1231,共9页Research of Environmental Sciences

基　　金：北京市科技计划项目(No.Z231100003823018)。

摘　　要：为充分利用卫星传感器在探索长时间、大范围区域内近地面二氧化氮(NO2)浓度时空变化规律方面的优势,进一步提高卫星近地面NO_(2)浓度预测的准确性,本研究以Sentinel-5P卫星TROPOMI传感器观测的对流层NO_(2)柱浓度为基础,提出一种融合多源地理要素的“自上而下”近地面NO_(2)浓度遥感估算方法,综合分析随机森林模型(RF)、极致梯度提升树模型(XGBoost)和轻型梯度提升树模型(LightGBM)的性能,基于3种树模型对2019−2020年京津冀地区近地面NO_(2)浓度进行反演,并采用十折交叉验证方法分别对3种模型在近地面NO_(2)浓度预测中的精度差异与稳定性进行了检验比较,利用拟合优度(R2)、均方根误差(RMSE)和平均绝对误差(MAE)对模型进行精度评价,最终选取XGBoost以实现京津冀地区卫星近地面NO_(2)浓度的高效分析预测(R2=0.85,RMSE=6.61μg/m^(3),MAE=2.09μg/m^(3)),在此基础上,从季度、年份等时间尺度对近地面NO_(2)浓度进行空间分析。结果表明:①由于2020年新型冠状病毒感染疫情反弹带来的人类生产活动和出行活动的大量减少,2019年近地面NO_(2)浓度(13.96μg/m^(3))比2020年(13.04μg/m^(3))整体偏高。②近地面NO_(2)浓度具有明显的季节性变化特征,春、夏两季由于大气扩散条件较好,近地面NO_(2)浓度相对较低,在冬季达到全年峰值。③基于SHAP值(沙普利加性解释法)方法对模型特征进行空间分析,并定量研究每个特征对模型的正负贡献程度,其中,对流层NO_(2)柱浓度对预测近地面NO_(2)浓度起到主要促进作用,大气边界层高度对预测近地面NO_(2)浓度起到抑制作用,另外太阳直射辐射、人口密度、地表温度及降水量等指标均对预测近地面NO_(2)浓度有明显影响。研究显示,XGBoost能够更加稳定和准确地预测卫星近地面NO_(2)浓度,为准确识别近地面NO_(2)浓度时空分布特征提供新的手段,可在一定程度上突破现阶�In order to make full use of the advantages of satellite instruments to explore the spatiotemporal variation patterns of nearsurface NO_(2) concentrations in large areas over long periods and further improve the accuracy of near-surface NO_(2) concentration prediction,a‘top-down’remote sensing estimation method was proposed to observe the tropospheric NO_(2) vertical column density data(TROPOMI)for calculating near-surface NO_(2) concentrations by integrating multiple geospatial factors.The study comprehensively analyzed the performance of the Random Forest model(RF),Extreme Gradient Boosting Tree model(XGBoost),and Light Gradient Boosting Tree model(LightGBM)in predicting the near-surface NO_(2) concentrations in the Beijing-Tianjin-Hebei region in 2019 and 2020.The accuracy of the three models was evaluated using the goodness of fit(R2),root mean square error(RMSE)and mean absolute error(MAE).The Extreme Gradient Boosting Tree model was finally selected to efficiently analyze and predict near-surface NO_(2) concentrations in the region(R2=0.85,RMSE=6.61μg/m^(3),MAE=2.09μg/m^(3)).Based on this approach,spatial analysis of near-surface NO_(2) concentrations at different time scales,such as quarterly and yearly was performed.The results showed:(1)Due to the significant reduction in human production activities and travel activities caused by the COVID-19 epidemic,the near-surface NO_(2) concentration in 2019(13.96μg/m^(3))was higher than that in 2020(13.04μg/m^(3)).(2)The near-surface NO_(2) concentration showed obvious seasonal variation characteristics.The NO_(2) concentration was relatively low in spring and summer due to favorable atmospheric diffusion conditions,and reached an annual peak in winter.(3)The SHAP value(Shapley Additive Explanation)method was used to spatially analyze the model features and quantitatively study the positive and negative contributions of each feature to the model.The tropospheric NO_(2) vertical column density plays a positive role in the prediction of near-surface NO_(2) c

关 键 词：NO_(2) TROPOMI 机器学习 XGBoost SHAP值 

分 类 号：X511[环境科学与工程—环境工程]