光谱分辨率对土壤组分建模影响分析  

Effects of Different Spectral Resolutions on Modeling Soil Components

作　　者：陈玉[1] 魏永明[1] 王钦军[1,2] LI Lin 雷少华 路春燕 CHEN Yu;WEI Yong-ming;WANG Qin-jun;LI Lin;LEI Shao-hua;LU Chun-yan(CAS Key Laboratory of Digital Earth Science,Aerospace Information Research Institute,Chinese Academy of Sciences,Beijing100094,China;Sanya Institute of Remote Sensing,Sanya 572029,China;Department of Earth Sciences,Indiana University-Purdue University Indianapolis(IUPUI),IN 46202,USA;Nanjing Normal University,Nanjing 210023,China;Fujian Agriculture and Forestry University,Fuzhou 350002,China)

机构地区：[1]中国科学院数字地球重点实验室,中国科学院空天信息创新研究院,北京100094 [2]三亚中科遥感研究所,海南三亚572029 [3]Department of Earth Sciences,Indiana University-Purdue University Indianapolis(IUPUI),IN 46202,USA [4]南京师范大学,江苏南京210023 [5]福建农林大学,福建福州350002

出　　处：《光谱学与光谱分析》2021年第3期865-870,共6页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金项目(41601383);国家重点研发计划项目(2017YFC1500902);兵团科技攻关项目(2017DB005-01);海南省百人专项联合资助。

摘　　要：实验室可见-近红外高光谱数据(VIS-NIR)具有快速、高效、无损等技术优势,被越来越多应用于土壤组分反演中。光谱分辨率越高所能表达的土壤信息越丰富,但也带来了数据冗余。目前,对于不同光谱分辨率对土壤组分建模影响效应分析的研究相对较少。以欧洲土壤中心数据集19036个土壤样本为数据源,以土壤总氮(N)、有机碳(OC)、碳酸钙(CaCO_(3))、粘土(Clay)为例,基于偏最小二乘回归方法(PLS)并选择30%的随机样本独立验证的方式开展相关研究。首先将所有样本原始0.5 nm分辨率4200个波段的高光谱数据采用等间距取均值方法分别重采样到2,4,8,…,1024 nm开展分析。结果表明:随着光谱分辨率的降低,土壤各类组分反演精度均呈下降趋势,光谱分辨率在64 nm以上,4类土壤组分普遍具有较高的模型验证精度(R^(2)>0.65,RPD>1.7),光谱分辨率在128 nm以下CaCO 3和Clay组分精度显著变差;4类组分中,CaCO_(3)对光谱分辨率敏感性最强,在高光谱分辨率下反演精度较高(R 2>0.86,RPD>2.72),但随光谱分辨率降低精度下降最快。此外,基于光谱响应函数将样本光谱重采样到GF2,S3A,L8,Aster,Modis和S3OLCI六种常见卫星传感器的光谱分辨率展开评价。结果表明:土壤N、OC在各传感器中均可获得较高的精度,甚至在GF2传感器仅有4个波段情况下,也具有不错的验证精度(R^(2)=0.56;RPD=1.51),而土壤CaCO_(3)及Clay反演精度普遍较差;除传感器光谱波段数量外,波段位置对土壤组分的反演能力的影响也很显著,拥有近红外长波(1100~2500 nm)光谱范围的传感器对土壤组分的反演能力优于缺少该光谱波段的传感器,特别是粘土矿物的吸收峰多位于近红外长波段,S3A,L8,Aster和Modis传感器的Clay反演能力均优于光谱波段数更多的S3OLCI。该研究成果对土壤组分高光谱数据预处理、卫星数据源的选择及未来传感器光谱通道的设计具有指导意义�The laboratory visible-near infrared(VIS-NIR)spectroscopy has been frequently used in quantifying soil components because it is effective,fast and nondestructive etc.The higher spectral resolution is the richer soil information we could obtain.However,hyperspectral data are red undant and should be preprocessed.The study of the effects of different spectral resolutions on the modeling of soil components is relatively inadequate.Taking advantage of the European Land Use/Cover Area Frame Statistical Survey(LUCAS)dataset having 19036 soil samples,we investigate the effects of different spectral resolutions on modeling soil components:total soil nitrogen(N),organic carbon(OC),calcium carbonate(CaCO3),and clay.To achieve this,we took the partial least squares regression(PLS)method as the evaluation model and randomly chose 30%samples for independent verification.Firstly,the spectral data which have 4200 bands with 0.5 nm spectral resolution were resampled to 2,4,8,…,1024 nm respectively using average reflection value by of uniform interval sampling.The results are as follows:(1)when the spectral resolution was decreased,the inversion accuracy of soil components showed a downward trend;(2)when the spectral resolution was higher than 64 nm,higher model validation accuracies were obtained for estimating the four selected soil components(R^(2)>0.65,RPD>1.7);(3)the accuracy for CaCO3 and clay components was significantly reduced when the spectral resolution was lower than 128 nm;(4)of the four soil components,CaCO3 was the most sensitive to spectral resolution.It has higher accuracy(R^(2)>0.86,RPD>2.72)at high spectral resolutions,but the accuracy reduced most rapidly as the spectral resolution decreases.Secondly,based on the spectral response functions for a group of common satellite sensors,the inversion performances of using GF2,S3 A,L8,Aster,S3 OLCI,and Modis spectral bands are summarized as follows:(1)all sensors achieved higher accuracy for soil N and OC even if GF2 has 4 different bands(R^(2)=0.56;RPD=1.51);(2)a l

关 键 词：土壤组分 实验室可见近红外光谱 卫星传感器 光谱分辨率 偏最小二乘法 

分 类 号：O657.3[理学—分析化学]