Estimation Of Soil Heavy Metal Arsenic Content Under The Framework Of Hyperspectral Quantitative Remote Sensing

With the continuous development of China’s economy and the rapid improvement of the level of industry and agriculture,the problem of heavy metal pollution in soil is becoming increasingly prominent,which has become a research hotspot in recent years.Heavy metal pollution has seriously endangered human life,which is difficult to degrade,easy to accumulate and toxic,and has an impact on crop growth,yield and quality.Traditional soil heavy metal monitoring methods mainly rely on laboratory tests,which is time-consuming and labor-consuming.With the development of hyperspectral remote sensing analysis technology,it is possible to use remote sensing technology to quantitatively estimate soil parameters.In view of the problems in the current quantitative soil hyperspectral field,the selection of characteristic bands is not accurate,the model accuracy is low,and the generalization ability is weak.The soil of Honghu and Daye in the typical area of Jianghan Plain was selected as the research object.A total of 92 soil surface samples were collected from the farmland area(Honghu)and metal mining area(Daye),The spectral data of Daye area and Honghu area were measured by ASD Field Spec 3 and SVC HR-1024,the wavelength range of the spectrometer is 350～2500 nm,and the content of heavy metal As in the soil is analyzed and determined,Five characteristic band selection algorithms:Spearman correlation analysis(SCA),Stable competitive adaptive reweighted sampling(s CARS),Iteratively retaining informative variables(IRIV),Stable competitive adaptive reweighted sampling Coupling Spearman Feature Enhancement Algorithm(s CARS-SCA),Iteratively retaining informative variables Coupling Spearman Feature Improvement Algorithm(IRIV-SCA)were used to select characteristic bands,the Partial least square regression(PLSR),Support vector machine regression(SVMR),Gradient boost regression tree(GBRT)and XGBoost regression models are established respectively.The accuracy of multiple models in two regions under multiple feature selection modes is comprehensively compared,and the optimal method of soil heavy metal inversion is found out,which is of great significance for the rapid monitoring of soil heavy metal content in this region.The main conclusions are as follows:(1)The average content of heavy metal As in the soil of farmland area(Honghu)and metal mining area(Daye)were 32.86 and 9.28μg/g,respectively.With reference to the national standard soil environmental quality standard of the people’s Republic of China GB15618-2018,the average value of Honghu area exceeds the level three standard(To ensure the critical value of soil for agricultural and forestry production and normal plant growth),which belongs to the polluted area.The mean value of Daye area is lower than the level one standard(the limit value for protecting the natural ecology and maintaining the soil quality of natural background),which belongs to the unpolluted area.(2)The overall correlation between heavy metal As content in soil and original reflectance of spectral band is low,From the perspective of the Honghu region,the overall correlation is low and the volatility is large.The correlation at 400.4 nm is the highest,and the absolute value of the correlation is 0.2232.The correlation tends to stabilize and decrease continuously after700 nm.For Daye area,the absolute value of the correlation coefficient of spectral reflectance and soil heavy metal As content reached the highest correlation at 422 nm,with a maximum value of 0.5317.The correlation tended to stabilize after 731 nm and continued to decrease.After coupling the Spearman feature enhancement method,the correlation between the two is greatly improved.(3)Based on the characteristic band selected by SCA,CARS,IRIV,CARS-SCA,IRIV-SCA variables,a variety of estimation models of heavy metals and spectral data are constructed.For Honghu area,SCA algorithm,CARS algorithm,IRIV algorithm,CARS-SCA algorithm and IRIV-SCA algorithm respectively select4,40,23,14 and 10 feature variables to model,accounting for only 0.44%,4.42%、2.54%,1.55%and 1.1%of the whole band.For Daye area,SCA algorithm,CARS algorithm,IRIV algorithm,CARS-SCA algorithm and IRIV-SCA algorithm respectively select 51,34,48,11 and 14 feature variables to model,accounting for only 2.55%,1.7%,2.4%,0.55%and 0.7%of the whole band.From the modeling results,using the feature variable filtering algorithm to filter the original band can not only ensure the accuracy of the model,but also greatly reduce the complexity of the model.(4)For Honghu area,the highest accuracy is obtained based on the feature variable modeling of CARS-SCA algorithm.R~2,RMSE and MAE of the validation set are 0.983,0.333 and 0.254 respectively;For Daye area,the highest accuracy is obtained based on the feature variable modeling of IRIV-SCA algorithm,R~2,RMSE and MAE of the validation set are 0.984,0.018 and 0.091 respectively.It can predict As content of heavy metals in soil.It is consistent in different regions,which shows that the coupled Spearman Feature Enhancement Algorithm can improve the prediction accuracy of heavy metal As in soil.(5)Compared with four modeling methods:PLSR,SVMR,GBRT,XGBoost.for Honghu region,SVMR model has the best prediction effect,which is better than other models.For Daye region,XGBoost model has the best prediction effect.It shows that SVMR and XGBoost can solve the complex nonlinear relationship between independent variables and dependent variables,so as to build a robust regression model.