基于随机森林模型的中国气溶胶光学厚度反演及变化分析  

Random Forest Model–based Aerosol Optical Depth Inversion and Variation Analysis in China

作　　者：赵乐鋆 李夫星[3,4,5] 王丽梅 蒋镕致 孔韵宁 王金 吴佳仪[3,4,5] 赵子渝 王诗瑶 ZHAO Lejun;LI Fuxing;WANG Limei;JIANG Rongzhi;KONG Yunning;WANG Jin;WU Jiayi;ZHAO Ziyu;WANG Shiyao(Nanjing Institute of Geography and Limnology,Chinese Academy of Sciences,Nanjing 211135,China;University of Chinese Academy of Sciences,Beijing 100049,China;School of Geographical Sciences,Hebei Normal University,Shijiazhuang 050024,China;Hebei Technology Innovation Center for Remote Sensing Identification of Environmental Change,Shijiazhuang 050024,China;Hebei Key Laboratory of Environmental Change and Ecological Construction,Shijiazhuang 050024,China)

机构地区：[1]中国科学院南京地理与湖泊研究所,南京211135 [2]中国科学院大学,北京100049 [3]河北师范大学地理科学学院,石家庄050024 [4]河北省环境变化遥感识别技术创新中心,石家庄050024 [5]河北省环境演变与生态建设实验室,石家庄050024

出　　处：《地球与环境》2025年第1期74-88,共15页Earth and Environment

基　　金：河北省自然科学青年基金项目(D2019205027);大学生创新创业训练计划项目(202310094010)。

摘　　要：以地面气象站点监测数据与MAIAC AOD数据构建随机森林模型,对2018—2019年中国气溶胶光学厚度(AOD)反演,并分析时空变化。结论如下:(1)将数据集划为春、夏、秋、冬4组,分组训练模型优化参数,最终确定决策树个数分别为250、250、200、200,决策树最大深度分别为40、90、40、40。经检验在该参数下模型的学习精度较高,参数具有较强的可靠性,决定系数(R^(2))分别为0.925、0.922、0.935、0.930,均方根误差(RMSE)分别为0.091、0.082、0.055、0.065;(2)基于确定的参数完成AOD反演后,利用AERONET AOD数据检验模型的估值精度,R^(2)为0.891,RMSE为0.129,整体精度较高。同时,各地面气象站点反演后的AOD年均值与原始AOD年均值相比变化趋势相近,且略高于原始年均值;(3)相较于2018年,2019年中国AOD值整体略低,空气质量有所改善;不同季节的AOD大小关系为:春季>夏季>冬季>秋季;(4)中国AOD值空间分布呈现出“东南高西北低”的特征,与胡焕庸线揭示的我国人口分布格局相一致。华北平原、华中地区、四川盆地等AOD偏高;南疆地区由于气候干旱,植被稀疏,沙尘天气频繁,AOD值也较高;东北地区、青藏高原、云贵高原与内蒙古高原,受地形地势、地理位置等影响,经济实力偏弱,人口分布稀疏,AOD偏低。Using ground meteorological station monitoring and MAIAC AOD data,a random forest model was constructed to invert the aerosol optical depth(AOD)in China from 2018 to 2019 and to analyze its spatiotemporal variations.The conclusions are as follows:(1)The dataset was divided into four groups corresponding to spring,summer,autumn,and winter.The models were trained with optimized parameters for each group.The final number of decision trees was set to 250,250,200,and 200,with maximum depths of 40,90,40,and 40,respectively.The model demonstrated high learning accuracy with these parameters,showing strong reliability.The coefficients of determination(R^(2))were 0.925,0.922,0.935,and 0.930,and the root mean square errors(RMSE)were 0.091,0.082,0.055,and 0.065.(2)After completing the AOD inversion based on the determined parameters,the model′s estimation accuracy was verified using AERONET AOD data,yielding an R²of 0.891 and an RMSE of 0.129,indicating high overall accuracy.Additionally,the annual average AOD values derived from the inversion at various ground meteorological stations showed a trend similar to the original AOD annual averages,albeit slightly higher.(3)Compared to 2018,the overall AOD values in China in 2019 were slightly lower,indicating an improvement in air quality.The AOD values for different seasons followed the order of spring>summer>winter>autumn.(4)The spatial distribution of AOD values in China exhibited a pattern of″high in the southeast and low in the northwest,″consistent with the population distribution revealed by the Hu Huanyong Line.High AOD values were observed in the North China Plain,central China,and the Sichuan Basin.The southern Xinjiang region,characterized by arid climate,sparse vegetation,and frequent dust storms,also had high AOD values.Conversely,the Northeast region,the Tibetan Plateau,the Yunnan-Guizhou Plateau,and the Inner Mongolia Plateau,influenced by topography,geographical location,weaker economic strength,and sparse population distribution,exhibited lower AOD values

关 键 词：随机森林模型 MAIAC AOD AERONET AOD 时空分布特征 变化趋势 

分 类 号：P407[天文地球—大气科学及气象学]