Estimating The Concentration Of Soil Heavy Metal Based On HyMAP-C Airborne Hyperspectral Image

The problem of soil heavy metal pollution in China is serious,but the traditiona l methods of soil heavy metal monitoring and assessment cannot meet the needs of large-scale continuous space.In addition,although there are many successful cases in the study of estimating surface soil heavy metals based on ground spectral features,few studies have focused on inversion of surface soil heavy metal concentrations by imaging hyperspectral remote sensing,which means that the great advantage of estimation surface soil heavy metals based on imaging spectral has not been brought into full application.The monitoring of surface soil heavy metals by imaging hyperspectral remote sensing still faces complex practical problems,theoretica l construction and technical difficulties etc.,and there are still a great deal of research problems to explore.Therefore,it has been implemented Hy MAP-C aerial hyperspectral imagery of the Yitong County area in Jilin Province as an example to explore the estimation of surface heavy metal concentration in black soil in the thesis.The main work and conclusions of this paper are as follows:(1)With the selection of heavy metal spectral features,the competitive adaptive reweighted sampling(CARS)method is simple and reliable structure.Two feature selection methods,CARS and Pearson correlation,are used to select the characterist ic spectra of four heavy metals(As,Cr,Pb,and Zn),and the validity of their spectral characteristic is verified.The evaluation results established by the traditional method show that the CARS method can select effective spectral features for all four heavy metals,while the Pearson method can only work for three heavy metals(As,Pb,Zn).At the same time,the CARS feature selection method is just only operated on the reflectance data,while the Pearson selection method is operated on four types of preprocessed data;and the Pearson correlation selection is usual y subject to human influence.(2)The spectral characteristics of the four heavy metals have commonalit y,moreover,there is a common characteristic of multiple heavy metals in the spectral wavelength range of 2.0-2.3 ?m.With the verification of the spectral features in above research,the spectral characteristics of the four heavy metals are analyzed and summarized.Although Pearson and CARS have different strategies and perspectives for the selection of feature methods,the commonality of the analysis results for the heavy metals(including the effective features of As,Pb,Zn,and CARS for Cr)are very high.In addition to the common features of the spectral features,each heavy metal has different spectral characteristics from other heavy metals.(3)In this thesis,Extra Trees method is applied to the heavy metal invers io n estimation.It is found that the Extra Trees model is better than the traditional classical method for the linear spectral features selected by the Pearson method.The fitting ability of the nonlinear combined spectral features selected by the CARS method is lower than the SVM model of the corresponding data.(4)In order to overcome the problems of overfitting and model instability,we propose a method for estimating soil heavy metals based on Stacking model.Comparing the accuracy evaluation indexes of all the models based on the linear features selected by the Pearson correlation method or the combined features selected by the CARS,the accuracy,stabilityand resistance of the Stacking method perform as wel;that is,they can overcome the problems caused by unbalanced data and small sample dataset.Moreover,even for heavy metal As with high spatial heterogeneity,the heavy metal concentration distribution of model estimation with the hyperspectral image is the consistent with the actual verification analysis.The reliability of the Stacking estimation model is higher.