多源遥感对宁夏干湿季土壤含盐量敏感性分析与定量反演

Sensitivity analysis and quantitative inversion of multi-source remote sensing to soil salt content in dry and wet seasons in Ningxia

快速准确反演干旱地区土壤盐渍化程度是有效防止盐渍化扩张的前提。为探讨地面高光谱和Landsat 8 OLI影像数据针对干湿季土壤含盐量敏感性分析与定量反演的问题,本研究以宁夏银北平罗县干季(4月)和湿季(10月)表层土壤(0~20 cm)含盐量、地面高光谱和Landsat 8 OLI影像为数据源,利用线性和非线性函数检验2种光谱数据及其对应盐分指数对研究区土壤含盐量的敏感性。采用偏最小二乘回归(PLSR)、支持向量机(SVM)和反向传播神经网络模型(BPNN)构建地面高光谱和Landsat 8 OLI影像的盐分估算模型。结果表明:干、湿季土壤含盐量平均值分别为6.17和4.28 g·kg^(-1),土壤盐渍化较为严重;不同季节地面高光谱和影像光谱对土壤含盐量敏感的波段和盐分指数不同,地面高光谱经重采样波段和盐分指数与土壤含盐量敏感性均表现为极显著;干、湿季土壤含盐量的BPNN估算模型稳定性和预测能力均优于PLSR和SVM模型;干、湿季均以高光谱-BPNN模型效果最佳,其决定系数R2分别为0.739和0.819,RPD分别为1.49和1.95;经重采样地面高光谱模型校正后的干季影像模型精度R2从0.685提升到0.844;湿季影像模型精度R2从0.654提升到0.788,有效提高了较大尺度下的土壤含盐量估算精度。本研究实现了遥感监测土壤含盐量由点向面的空间转换,为宁夏银北地区土壤盐渍化的识别和防治提供了科学参考。

Rapid and accurate inversion of soil salinity in arid areas is the premise to effectively prevent the expansion of salinization.To address sensitivity analysis and quantitative inversion of soil salt content in dry and wet seasons by means of ground hyperspectra and Landsat 8 OLI images,we obtained data of topsoil(0-20 cm)salt content,ground hyperspectra and Landsat 8 OLI image in the dry season(April)and wet season(October)in Pingluo County,Ningxia.Linear and nonlinear functions were used to test the sensitivity of the spectral data and corresponding salinity index.We established the models to estimate soil salt content based on ground hyperspectral and image data using partial least squares regression(PLSR),support vector machine(SVM),and back propagation neural network model(BPNN).The average soil salt content in dry and wet seasons was 6.17 and 4.28 g·kg^(-1),respectively,indicating serious soil salinization.The sensitive bands and salinity index of ground hyperspectra and image spectra to soil salinity differed among seasons.The resampling band and salinity index of ground hyperspectra showed extremely high sensitivity to soil salinity.The stability and prediction ability of BPNN estimation model of soil salt content were better than those of PLSR and SVM.Hyperspectral-BPNN in dry and wet seasons was the best estimation model,with the prediction accuracy of 0.739 and 0.819,and the relative analytical errors of 1.49 and 1.95,respectively.After calibrated by the resampled measured spectrum model,the estimation accuracy of the image-spectra-based model increased from 0.685 to 0.844 in the dry season and from 0.654 to 0.788 in the wet season,which effectively enhanced the accuracy in estimating soil salt content at large scale.We successfully made the spatial transformation of soil salt content from small to large scale.The results provided a scientific reference for identification and prevention of soil salinization in Yinbei of Ningxia.