Hyperspectral Inversion Of Soil Heavy Metal Content In The Three-river Source Region

Hyperspectral remote sensing technology has considerable research value in monitoring and evaluating soil heavy metal pollution.In this study,the Three-River Source Region was taken as the study area.The occurrence relationship of six kinds of heavy metals in soil,such as Mn,Cu,Zn,Pb,Cr,Ni,with soil organic matter,iron-manganese oxides and clay minerals,was studied through the determination and analysis of soil samples and the collection of soil reflectance spectra.A variety of spectral transformation methods were carried out,such as FD,SD,LR,CR,MSC and SD-FD,LR-FD,CR-FD,and MSC-FD obtained by performing first-order differential transformation on the basis of SD,LR,CR,and MSC.Meanwhile,the correlation between soil heavy metals content and soil spectrums were also analyzed to determine the soil characteristic spectrum.And partial least squares(PLS)method,support vector machine(SVM)method and random forest model(RF)were used to model and analyze six soil heavy metal elements and each spectral variable,respectively.The feasibility of using characteristic bands instead of full bands for soil heavy metal content inversion modeling was discussed.And the best regression model of the hyperspectral inversion of each heavy metal element was established.On the basis,the heavy metal contents in the soil of different grassland types were estimated and evaluated.The results showed as follows:(1)The super background rate of Pb element in the Three-River Source Region was76.39%,which was highly variable and seriously enriched.In addition,the results of the Cumulative Index and Index Nemerow of the soil heavy metals indicated that the heavy metal pollution in the Three-River Source Region was mainly Pb,but the overall soil environmental pollution was relatively light.(2)The accuracy of fit of the characteristic band model was mainly achieved by FD,LR-FD and MSC-FD spectral variables,which of the full-band estimation model was mainly achieved by SD,SD-FD and MSC-FD spectral variables.The number of characteristic bands for the extraction of six soil heavy metal elements is 10,17,7,5,8,and 13,respectively.The characteristic bands reflect the spectral characteristics of soil organic matter,clay minerals,and iron-manganese oxide,and can replace the whole band to model the content of soil heavy metals.(3)Based on the three modeling methods,the inversion models of soil heavy metals content in the Three-River Source Region were obtained,which estimation ability showed that RF > SVM > PLS,and the characteristic band model and the full band model of RF are between 0.83-0.90 and 0.89-0.92 respectively,indicating that the RF model had relatively good stability and can effectively invert soil heavy metal content in the Three-River Source Region.It showed that the random forest model has relatively good stability and can effectively retrieve the heavy metal contents of the soil in the Three-River Source Region.(4)Based on the spectral characteristic bands of six heavy metals extracted from the Three-River Source Region and the optimal characteristic-band models of PLS,SVM and RF,the prediction and precision evaluation of heavy metals content in different grassland types were compared.Among them,the hyperspectral prediction results of heavy metal content in meadow grassland,alpine meadow and alpine grassland were similar to the overall inversion rule of heavy metal content in the Three-River Source Region,and there were some differences in the estimation results of other three grassland types.This study can provide technical support for the nondestructive monitoring of soil heavy metal content,large-scale monitoring,and assessment of heavy metal pollution in soil;promote the application of hyperspectral skill in the field of comprehensive monitoring and assessment of ecological environment;and provide theoretical basis for remediation and early warning of soil heavy metal pollution.