土壤质地对灌区土壤盐分高光谱反演精度影响研究  

Effect of Soil Texture on the Accuracy of Hyperspectral Inversion of Soil Salinity in Irrigated Area

作　　者：周世勋 尹娟[1] 杨莹攀 杨震 常布辉[2] ZHOU Shi-xun;YIN Juan;YANG Ying-pan;YANG Zhen;CHANG Bu-hui(School of Civil and Hydraulic Enigineering,Ningxia University,Yinchuan 750021,China;Yellow River Institute of Hydraulic Research,Zhengzhou 450045,China)

机构地区：[1]宁夏大学土木与水利工程学院,银川750021 [2]黄河水利科学研究院,郑州450045

出　　处：《节水灌溉》2024年第6期1-10,共10页Water Saving Irrigation

基　　金：国家重点研发计划重大项目课题5专题4(2021YFD1900605-04);宁夏高等学校一流学科建设(水利工程)资助项目(NXYLXK2021A03)。

摘　　要：针对土壤质地对高光谱反演土壤盐分精度的影响不明确问题,于2023年4月1-10日在内蒙古河套灌区沈乌灌域共采集了132个不同盐渍化程度的土壤样品,并同步采集了对应的光谱信息,研究了不同盐渍化程度下土壤光谱反射率的变化特征以及不同土壤质地光谱特征与土壤盐分的相关性,探讨了土壤样本适宜的数学变换方法,并筛选敏感波段,建立了基于全部样本以及不同土壤质地下的土壤盐分含量的高光谱反演模型。结果表明:随着土壤盐分含量的增加,高光谱反射率逐渐增大;随着土壤粒度的减小,不同土壤质地下土壤盐分与不同波段的反射率及其相关系数呈先增加后下降的变化趋势。通过对光谱数据进行数学变换后,发现以倒数对数微分、对数微分、平方根微分3种变换效果最佳。通过建立多元逐步线性回归(BPNN)、偏最小二乘回归(PLSR)、支持向量机回归(SVM)以及BP神经网络(BPNN)4种模型,对光谱变换下的盐分含量进行了估算,4种模型的估算精度由高到低表现为:BPNN>SVM>PLSR>MLSR。相较于全部样本的土壤盐分估算结果,考虑不同土壤质地的盐分估算精度均有所提升,其中砂粒质地估算精度R2由0.918提升到0.962,RPD由3.493提升到4.313;粉粒质地估算精度R2由0.866提升到0.902,RPD由2.613提升到3.310;黏粒质地估算精度R2由0.876提升到0.926,RPD由2.651提升到3.953,且在3种土壤质地背景下建立的模型均达到了出色模型的标准。说明在考虑土壤质地的前提下进行含盐量的高光谱反演,有利于提升反演精度。This study addresses the unclear influence of soil texture on the accuracy of hyperspectral inversion for soil salinity.From April 1 to 10,2023,a total of 132 soil samples with different salinization degrees were collected in Shenwu irrigation area of Hetao Irrigation District,Inner Mongolia.Corresponding spectral information was collected simultaneously.The variation characteristics of spectral reflectance of soil under different salinization degree and the correlation between spectral characteristics of soil texture and soil salinity were studied.The appropriate mathematical transformation method of soil samples was discussed,and the sensitive bands were selected.A hyperspectral inversion model based on all samples and different soil textures was established.The results showed that the hyperspectral reflectance increased with the increase of soil salt content.As the soil grain size decreased,soil salinity and reflectance of different bands and their correlation coefficients exhibited an increasing trend followed by a decreasing trend.After mathematical transformation of spectral data,it is found that reciprocal logarithmic differentiation,logarithmic differentiation and square root differentiation have the best effect.Four models of multiple stepwise linear regression(BPNN),partial least squares regression(PLSR),support vector machine regression(SVM)and BP neural network(BPNN)were established to estimate the salt content under spectral transformation.The estimation accuracy of the four models was as follows:BPNN>SVM>PLSR>MLSR.Compared with the soil salt estimation results of all samples,the salt estimation accuracy of different soil textures was improved.For sand texture,the estimation accuracy R2 was increased from 0.918 to 0.962,and the RPD was increased from 3.493 to 4.313.The grain texture estimation accuracy R2 increased from 0.866 to 0.902,and the RPD increased from 2.613 to 3.310.The accuracy of clay texture estimation R2 increased from 0.876 to 0.926,and the RPD increased from 2.651 to 3.953,and the mod

关 键 词：土壤盐分 高光谱 土壤质地 光谱变换 反演模型 

分 类 号：S27[农业科学—农业水土工程]