基于随机森林的海河流域地表温度降尺度  

Land Surface Temperature Downscaling of Haihe River Basin Based on Random Forest

作　　者：李建明 马燕飞 李仁杰[1,2,4] 商国营 李明明 LI Jianming;MA Yanfei;LI Renjie;SHANG Guoying;LI Mingming(Science College of Resources and Environmental,Hebei Normal University,Shijiazhuang 050024,China;Laboratory of Hebei Environmental Evolution and Ecological Construction,Shijiazhuang 050024,China;Department of Geography,Handan College,Handan,Hebei 056005,China;Hebei Innovation Center for Remote Sensing Identification Technology of Environmental Change,Shijiazhuang 050024,China)

机构地区：[1]河北师范大学资源与环境科学学院,石家庄050024 [2]河北省环境变化遥感识别技术创新中心,石家庄050024 [3]邯郸学院地理系,河北邯郸056005 [4]河北省环境演变与生态建设实验室,石家庄050024

出　　处：《遥感信息》2021年第4期151-158,共8页Remote Sensing Information

基　　金：国家自然科学基金青年科学基金项目(41701426)。

摘　　要：陆地表面温度(land surface temperature,LST)是长波辐射与地气湍流热通量交换的直接驱动力。通过降尺度可以获取高分辨率地表温度。文章以海河流域为研究区,探索了利用植被指数、空气温度、下行短波辐射构建解释向量集对地表温度空间降尺度的方法,并用此方法重建了MODIS MOD11A1影像的缺失像元。对随机森林算法在两个空间尺度(500 m、1000 m)下的训练模型进行分析,并在海河流域划分了五个子研究区来探讨随机森林降尺度模型在不同下垫面性质区域的表现。结果表明,在500 m、1000 m层级下降尺度模型R2分别为0.815、0.896,RSME分别为3.15 K、2.88 K,表明随机森林降尺度模型在1000 m层级的空间异质性更小。地面站点验证中,大兴站、密云站、馆陶站的LST误差分别为3.86 K、2.37 K、1.96 K。随机森林降尺度模型在植被覆盖度高的草地、林地表现更加优秀,LST空间分布合理,在旱地、建设用地降尺度LST略低于MODIS原始LST。最后,基于本研究的方法获取了海河流域无缺失像元的250 m分辨率地表温度数据,为区域地表水热通量研究、干旱监测等提供了数据参考。Land surface temperature(LST)is the direct driving force for the exchange of long-wave radiation and turbulent heat flux between earth and gas.High resolution surface temperature can be obtained by scaling down.Taking Haihe river basin as the research area,the method of constructing the interpretation vector set to the surface temperature spatial scaling by vegetation index,air temperature and downwind short wave radiation is explored,and the missing pixels of MODIS are reconstructed by this method.This paper analyzes the training model of random forest algorithm in two spatial scales(500 m,1000 m),and divides five sub-research areas in Haihe river basin to discuss the performance of random forest downscaling model in different underlying surface properties.The results show that the descending scale model R2 is 0.815 and 0.896 respectively at 500 m and 1000 m,and the RSME is 3.15 K and 2.88 K respectively.The spatial heterogeneity of the descending scale model RF is smaller at 1000 m.In the verification of ground stations,the LST errors of Daxing station,Miyun station and Guantao station are 3.86 K,2.37 K and 1.96 K respectively.The down-scale model RF performs better in grassland and forest land with high vegetation coverage,and the LST spatial distribution is reasonable.The down-scale MODEL LST in dry land and construction land is slightly lower than the MODIS original LST.Finally,the surface temperature data without missing pixels in Haihe river basin is obtained based on the method of this study,which provides data reference for regional surface water and heat flux research and drought monitoring.

关 键 词：海河流域 地表温度 随机森林 降尺度 模型 

分 类 号：TP79[自动化与计算机技术—检测技术与自动化装置]