Quantitative Inversion Of Soil Heavy Metals Based On Visible-Near-Infrared Spectroscopy

The traditional geochemical methods of soil heavy metal monitoring have some limitations,while the visible near infrared spectroscopy technology has the advantages of large range and high efficiency,and is becoming a means of soil quality monitoring.Guizhou Province is a typical province with high geochemical background in Southwest China.The upper reaches of the Duliu river basin is a famous antimony producing area in Guizhou Province.The soil safety of the basin is very important to the environmental safety of the region and its downstream areas.Taking the upper reaches of the Duliu River as the research area,soil samples were collected on the spot.Based on the content of heavy metals as and Sb in soil samples and the visible near infrared hyperspectral data of soil samples,the characteristic bands of heavy metals as and Sb in soil were screened through a variety of spectral preprocessing combined with genetic algorithm(GA),and the partial least squares regression(PLSR)and random sampling were used The forest regression(RF) method was used to construct the visible near infrared hyperspectral quantitative inversion model of soil heavy metals as and Sb,and compared with the widely used correlation coefficient screening band method.In order to solve the problem of incomplete manual extraction of spectral features,a convolution neural network quantitative inversion model of soil heavy metals as and Sb in the whole spectral region was established.The main conclusions are as follows(1)The results showed that the minimum value of As was 0.49 mg/kg,the maximum value was 111.68 mg/kg,the average value was 20.52 mg/kg,which was slightly higher than the background value of 20 mg/kg in Guizhou Province,the exceeding standard rate was 30.11%;the minimum value of Sb was 1.19 mg/kg,the maximum value was 137.07 mg/kg,the average value was 10.75 mg/kg,which was much higher than the background value of 2.2 mg/kg in Guizhou Province The over standard rate of samples was 80%.(2)The results showed that the spectral characteristics of soil spectra could be highlighted and the correlation between soil spectra and heavy metals As and Sb could be enhanced by nine kinds of spectral pretreatment;the results of multiple scattering correction(MSC)and standard normal transformation correction(SNV)were similar,and there was no significant difference in the range of correlation coefficient and the number of bands passing the significance test;the combination of MSC and SNV with differential treatment could improve the correlation between soil spectra and heavy metals The inversion model constructed by characteristic band screening shows that the effect of multiple scattering correction+first-order differential preprocessing(MSC+FD)is the best,and the model has the best prediction effect for heavy metals As and sb.(3)The R_P~2of the optimal partial least squares regression model and random forest regression model of soil heavy metals As and Sb were 0.354,-0.042 and 0.242,-0.015,respectively;the R_P~2of the optimal partial least squares regression model and random forest regression model of soil heavy metals As and Sb were 0.170,-0.003and 0.123,0.041,respectively;the R_P~2of the optimal partial least squares regression model and random forest regression model of soil heavy metals as and Sb were 0.170,-0.003 and 0.123,0.041,respectively The partial least squares regression model and random forest regression model of soil heavy metals As and Sb can not meet the minimum accuracy requirements.(4)PLSR and RF based on GA characteristic band selection can greatly improve the prediction accuracy of soil heavy metals,and the optimal model of heavy metals As is MSC+FD＿GA＿PLSR,R~2_cis 1,RPD_cis 89.569,R~2_Pis 0.986,RPD_cis 8.890;the optimal model of heavy metal sb is MSC+FD＿GA＿RF,R_c~2is 1,RPD_cis 3.8538e+15,R~2_Pis 0.942,RPD_cis 4.387。(5)the optimal model of soil heavy metal As shows that there are 105 characteristic bands of soil heavy metal as,which are distributed in 446.1～468.4nm,534.7～561.2nm,685～728.3nm,735.7～766.6nm,817～839.7nm and 852.2～1077nm;the optimal model of soil heavy metal sb shows that there are 259 characteristic bands of soil heavy metal Sb,which are relatively wide The sensitive bands are 343.6,346.8,351.6,359.6,671.5,686.5,746,753.4,921.2,931.7,936.6,957.4,959.9～961.2,967.5,973.8,978.8,985～986.3,994.9,998.6,1024.3,1031.6and 1035.3～1037.7nm.(6)For As,R_c~2,RPD_c,R_P~2and RPD_care 0.805,2.297,0.581 and 1.638, respectively;for Sb,R_c~2,18.946,0.670 and 1.847,respectively;the convolution neural network models of the two kinds of heavy metals are better than PLSR and RF screened by full spectrum region and correlation coefficient method,but the convolution neural network model is more sensitive to As and RF due to the small number of samples The prediction effect of Sb is not as good as PLSR and RF screened by GA band.