干旱区土壤盐渍化信息遥感建模

Remote sensing modeling of soil salinization information in arid areas

以新疆塔里木盆地北缘的渭干河-库车河三角洲绿洲为研究区,利用GF^(-1)与Landsat8 OLI影像数据作为基本数据源,从影像上提取15个盐分指数和5个光谱植被指数,通过灰度关联分析法,对0~10 cm表层土壤含盐量与影像光谱指数进行分析和筛选,确定出与土壤含盐量相关性较高的综合光谱指数,采用多元线性回归,偏最小二乘法回归,支持向量机回归三种方法分别对GF^(-1)与Landsat8 OLI影像构建基于实测数据和影像数据的综合指数土壤含盐量估算模型,并选出最优模型。结果表明:(1)在20个光谱指数中,相关性较好的光谱指数是SR、CSRI、SI、BI、S6、ARVI、SAVI、NDSI,关联系数均达到0.7以上,并基于这8个光谱指数构建综合光谱指数。(2)3种估算模型:基于GF^(-1)多元线性回归模型决定系数R^2为0.6856,高于决定系数R2为0.5142的Landsat8 OLI;偏最小二乘回归模型1~8个主成分,GF^(-1)决定系数2个>3个>1个,其中2个主成分最高可达0.6104,Landsat决定系数4个>3个>2个,其中4个主成分最高可达0.549;支持向量机模型3种函数,GF^(-1)决定系数RBF>Polynomial>Linear,其中RBF函数最高可达0.7969,Landsat决定系数Polynomial>RBF>Linear,其中Polynomial函数最高可达0.7154。对比3种模型可知,支持向量机回归模型的R2最高,因此该模型相对于多元线性回归和偏最小二乘回归更适于土壤盐渍化估算。

Taking Weigan-Kuqa River Delta Oasis located in the northern margin of the Tarim Basin in Xinjiang as study area. Found out 15 salt indexes and 5 spectral vegetation indexes from the images in GF^(-1) and Landsat8 OLI data. By means of gray relational analysis,the soil salinity and image spectral index of 0 ~ 10 cm soil layer were analyzed and screened comprehensively,and the comprehensive spectral index with high correlation with soil salinity was further determined. Using multivariate linear regression,partial least-squares regression and support vector machine regression to construct the soil salinity comprehensive index estimation model based on the measured data and the image data respectively for GF^(-1) and Landsat8 OLI images,and then select the optimal model.The results:( 1) SR,CSRI,SI,BI,S6,ARVI,SAVI,NDSI were the higher of all 20 spectral indices and the correlation coefficients were above 0. 7,and build a composite spectral index based on the eight spectral indices.( 2)Three kinds of estimation models: the coefficient R2 of multiple linear regression model based on GF^(-1) is 0.6856,which is higher than that of the model constructed based on Landsat8 OLI R20.5142; 1-8 partial least squares regression principal components,GF^(-1) coefficient of determination: 2 > 3 > 1,2 principal components reached 0.6104. Landsat,the coefficient of determination: 4>3>2,4 principal components up to 0.549; support vector machine model of the 3 kinds of function,the GF^(-1) coefficient ofdetermination RBF>Polynomial>Linear,RBF function is up to 0.7969. Landsat,the coefficient of determination Polynomial>RBF>Linear,Polynomial function is upto 0.7154. Thethree models show that the R^2 of support vector machine regression model is the highest,so the model is better than multiple linear regression and partial least squares regression for the estimation of soil salinization.