Study On The Hyperspectral Remote Sensing Inversion Of Soil Heavy Metal Concentrations Based On Random Forest Model

With the exploitation of coal resources,soil heavy metal pollution problems increasingly aggravate.So the research on soil heavy metal pollution in coal mining area is very important to human health and social resources sustainable development.The traditional heavy metal detection method requires much time and effort.The emergence of hyperspectral imager,with many bands,high spectral resolusion,and large amount of information,offers a new way for the studies of the soil heavy metal inversion.There are several problems in traditional linear or nonlinear inversion method,such as low generalization ability and low operation efficiency.As random forest model is the fusion of plenty of decision trees,its prediction performance greatly improved,and become the hotplot of research in recent years.The study area is Liuxin mining area,located in Xuzhou,Jiangsu Province.ASD hyperspectral data and Hyspex imaging hyperspectral data are used to estimate soil heavy metal content in the study area.The main conclusions are as follows:(1)In the research of soil heavy metal quantitative estimation based on laboratory ASD hyperspectral data,Local Correction Maximization denoising method and a variety of variable selection and modeling method are used.Preprocessing results show that Local Correction Maximization denoising method can maximize denoising and leave enough information correlation with soil heavy metal at the same time,correlation is higher and correlation bands are more.Partial Least Square(PLS),Support Vector Machine(SVM),Support Vector Machine based on Recursive Feature Elimination(RFE-SVM),Least Absolute Shrinkage and Selection Operator(LASSO),Ridge,Elastic Net(EN),Ridge_c,Random Forest(RF),Regularized Random Forest(RRF),Guided Random Forest(GRF),Guided Regularized Random Forest(GRRF)models are compares in estimation.The inversion results indicate that regularization variable selection method based on Random Forest can extract the optimal band with highest efficiency.In the optimal results,RRF is best in the estimation of Cr,with Rp2,RMSEp,MREp are 0.8708,3.0687,0.0467 separately;RRF is best in the estimation of Cu,with estimation indexs are 0.8215,3.7583,0.1657 separately;GRRF is best in the estimation of Pb,with estimation indexs are 0.8363,2.4325,0.1191 separately.(2)Using indoor and field ASD hyperspectral data to estimate heavy metal on Hyspex airborne hyperspectral data after atmospheric correction.Indoor results indicate that Random Forest is the optimal model.GRF is best in the estimation of Cr,with Rp2,RMSEp,MREp are 0.7179,15.8920,0.2530 separately;RRF is best in the estimation of Cu,with Rp2,RMSEp,MREp are 0.6388,11.7921,0.8174 separately;GRF is best in the estimation of Pb,with Rp2,RMSEp,MREp are 0.3661,5.3969,0.3617 separately.Inversion results of field ASD spectral are better than indoor.RRF is best in the estimation of Cr,with Rp2,RMSEp,MREp are 0.4113,15.0728,0.2473 separately;RRF is best in the estimation of Cu,with estimation indexs are 0.4146,13.1818,0.9933 separately;RRF is best in the estimation of Pb,with estimation indexs are 0.5710,4.9364,0.2719 separately.(3)Research on unmixing is carried out on Hyspex image data,and heavy metal estimation is implemented based on original and unmixed airborne hyperspectral data.Results indicate that,the accuracy of inversion improved significantly after unmixing.In the estimation results of unmixed spectal data,Random Forest is still the optimal model.RRF is best in the estimation of Cr,with Rp2,RMSEp,MREp are 0.7358,6.0063,0.0965 separately;GRRF is best in the estimation of Cu,with estimation indexs are 0.7017,7.2088,0.3734 separately;GRRF is best in the estimation of Pb,with estimation indexs are 0.7424,2.8030,0.1385 separately.Applying the characteristic bands extracted by four main estimation models to entire Hyspex image,then mapping the heavy metal content.Results show that Random Forest model is the most stable model,the heavy metal content estimation values distribute centralized,and estimation average value is close to measured average value,verify the effectiveness of the method further more.