基于机器学习的星载短波红外CO_(2)柱浓度估算  

Estimation of the column concentration of carbon dioxide using spaceborne shortwave infrared spectrometer

作　　者：李静波 张莹 盖荣丽[1] LI Jing-bo;ZHANG Ying;GAI Rong-li(School of Information Engineering,Dalian University,Dalian 116622,China;State Key Laboratory of Remote Sensing Science,Aerospace Information Innovation Research Institute,Chinese Academy of Sciences,Beijing 100101,China)

机构地区：[1]大连大学信息工程学院,辽宁大连116622 [2]中国科学院空天信息创新研究院遥感科学国家重点实验室,北京100101

出　　处：《中国环境科学》2023年第4期1499-1509,共11页China Environmental Science

基　　金：国家重点研发计划(2019YFC1509203)。

摘　　要：利用OCO-2卫星遥感数据、全球碳柱总量观测网(TCCON)站观测数据、NDVI归一化植被指数数据、ERA5大气数据,采用决策树和集成学习(XGBoost、普通随机森林、极端随机森林、梯度提升)对二CO_(2)平均柱浓度进行预测.通过相关性分析、特征选择与特征提取,建立模型预测CO_(2)平均柱浓度,再与TCCON站点的地基观测数据进行比对.通过分析不同模型(决策树、XGBoost、普通随机森林、极端随机森林、梯度提升)预测的结果,发现使用极端随机森林回归模型预测CO_(2)平均柱浓度的精度最高,R^(2)、均方根误差(RMSE)、平均绝对误差(MAE)、平均相对误差(MRE)分别为:0.953、0.492×10^(-6)、0.260×10^(-6)、0.063%,其余模型次之,因此对极端随机森林回归模型的预测性能随自身参数的影响进行了分析,结果表明,在误差允许的范围内(±2×10^(-6)),极端随机森林回归模型和梯度提升回归模型预测的准确率一样,都为98.10%.由于CO_(2)的背景浓度较高,而边界层内CO_(2)浓度的空间差异相对较小,因此需要进一步缩小误差的范围,在±1×10^(-6)误差范围内,极端随机森林回归模型和梯度提升预测的准确率分别为91.82%和90.51%.所以采用极端随机森林算法预测CO_(2)柱浓度的结果更好,精度更高,符合CO_(2)预测的精度要求.In this paper we used OCO-2 satellite observations,TCCON(total carbon column observing network,TCCON)observations,NDVI(Normalized Difference Vegetation Index)records,and atmospheric parameters from ERA5,and apply various machine learning models including decision tree and ensemble learning(e.g.XGBoost,random forest,extreme random forest,gradient lifting)to predict the carbon dioxide column concentration.Models are established to predict the concentration of carbon dioxide column through correlation analysis,feature selection and extraction.Model prediction results are then compared with the data from XCO_(2)observations from TCCON.By comparing the prediction results from different models,it has been found that the extreme random forest regression model has the best performance in predicting CO_(2)column concentration.R2,RMSE(root mean square error),MAE(mean absolute error)and MRE(mean relative error)of Extra Trees Regressor(extreme random forest regression model)are 0.953,0.492×10^(-6),0.260×10^(-6)and 0.063%respectively.Thereafter,the prediction performance of ExtraTreesRegressor is analysed by changing the input parameters.The results show that within the acceptable error range(±2×10^(-6)),the prediction accuracy of extreme random forest regression model and gradient lifting regression model is the same,which is 98.10%.Because the variation range of CO_(2)column concentration is small,it is necessary to further narrow the error range.Within the error range of±1×10^(-6),the prediction accuracy of extreme random forest regression model and gradient lifting regression model are 91.82%and 90.51%respectively.Therefore,the extreme random forest algorithm results in better accuracy in predicting CO_(2)column concentration,which meets the accuracy requirements of CO_(2)prediction.

关 键 词：二氧化碳平均柱浓度 集成学习 极端随机森林 梯度提升 决策树 XGBoost 短波红外 

分 类 号：X87[环境科学与工程—环境工程] TP181[自动化与计算机技术—控制理论与控制工程]