Hyperspectral Inversion Model Study Of Soil Hg And As Contents

Heavy metals Hg and As are heavy metal pollutants with global mobility,high bioconcentration,and high biotoxicity.The concentration acquisition is crucial for soil monitoring and remediation.In this paper,90 soil samples were collected for the gold mining area and the suburban area.Firstly,the soil samples were analyzed by indoor spectrometry and chemical tests to obtain hyperspectral data and soil heavy metal content data.Then,five kinds of spectral pre-processing(FD,SD,SG,MSC,SNV)and four kinds of spectral indices(DI,RI,NDVI,BI)were combined with the raw data,followed by correlation analysis with the Hg and As content data,respectively.Finally,partial least squares regression(PLSR),support vector machine(SVR),and gray wolf algorithm optimized back propagation neural network(GWO-BPNN)were used to establish the soil heavy metal content inversion model.The main research results are as follows.(1)All five spectral pre-processing methods of soil heavy metal hyperspectral data can achieve the amplification of information of soil Hg and As-related variables.Based on the spectral curves,it is concluded that FD and SD transformations can obtain more spectral feature peaks,and the trends of spectral features are the same after MSC and SNV transformations.(2)The correlation analysis of the spectral data with the measured values after the cross-combination of five pretreatments and four spectral index methods is a method to explore the effective extraction of the sensitive feature bands of soil heavy metals Hg and As.The correlation analysis between the soil heavy metal Hg and As content data and the spectral data after each combination of treatments all showed that the four crosscombinations of DI+SNV,RI+SNV,NDVI+SNV,and BI+R had better results.(3)In the soil Hg content inversion model,the GWO-BPNN nonlinear model constructed with 1265 variables screened by the combination of pretreatment and spectral index method(DI+SNV)+(RI+SNV)+(NDVI+SNV)+(BI+R)had the best prediction accuracy.the Hg sensitive bands ranged from 400-463 nm,615-690 nm,766-The GWO-BPNN nonlinear model constructed by screening 1455 variables with the combination of pretreatment and spectral index DI+SNV had the best prediction accuracy in the soil As content inversion model.The extracted As sensitive bands ranged from 400-571 nm,824-1316 nm,1357-1507 nm,1658-1897 nm,1980-2054 nm,and 2144-2368 nm.(4)The transferability results of the prediction models for soil Hg and As contents under the 2 types of land use in the suburban and mining areas showed that the best soil Hg and As migration ability were both based on the mixed soil data from the mining and suburban areas as the modeling set prediction model.However,the Hg prediction model had poorer migration ability in the suburban data compared to the validation set of mine data.The As prediction model has poorer migration ability in the mine data compared to the validation set of suburban data.