Inversion Of Soil Heavy Metal Content Based On GF-5 Hyperspectral Image

Soil heavy metal pollution is becoming more and more serious.It is particularly important to timely and accurately grasp the content and distribution of heavy metals in soil.With the development of remote sensing technology,it has been proved effective to use remote sensing image to carry out large-scale monitoring and evaluation of soil heavy metal content.However,when selecting the feature band of hyperspectral image data,correlation coefficient method or single band selection algorithm is used in most studies,and the accuracy of the model is relatively limited.On the other hand,the research idea of laboratory combined spectral image data modeling is relatively simple,how to effectively combine two kinds of data to invert heavy metal content is also a problem that needs to be solved.Therefore,this paper takes Tongguan mining area in Shaanxi Province as the research object and carries out the following experiments:（1）In view of the low accuracy of the traditional feature band selection algorithm in the inversion model,this paper coupled the single band selection algorithm according to the strategy of rough selection before selection.Taking GF-5（High Score 5）image as data source,the accuracy of partial least squares model（PLSR）of Cd and Hg content constructed based on different band selection algorithms was compared under two kinds of first-order differential（FD）and standard normal transformation（SNV）spectra,and the superiority of the coupled band selection algorithm was verified;The optimal model was selected to analyze the importance of spectral features,and the spatial mapping of Cd and Hg content was carried out.（2）Aiming at the problem of the single idea of modeling the current laboratory combined spectral image data,this paper proposed a new idea based on simulating the characteristic spectrum of GF-5 data.Taking Cd element as an example,the laboratory spectrum was used as auxiliary data,and the joint inversion of Cd content between the laboratory spectrum and GF-5 image data was realized by constructing the simulation of GF-5 data and spectral correction.The main conclusions of this paper are as follows:（1）The coupled band selection algorithm has obvious advantages in improving the accuracy of inversion model.By comparing three single algorithms based on competitive adaptive reweighting algorithm（CARS）,Monte Carlo uninformative variable elimination method（MC-UVE）and iterative retention of informative variable method（IRIV）,and the modeling accuracy of heavy metal content constructed by competitive adaptive reweighting algorithm-Genetic algorithm（CARS-GA）and CARS-IRIV coupling algorithms.It was found that under the two transform spectrum,the accuracy of Cd and Hg element content models constructed by CARS-GA coupling algorithm is the best,which indicates the superiority of coupling algorithm.（2）When retrieving heavy metal content based on GF-5 image,the best inversion model for Cd and Hg elements was CARS-GA-PLSR model under FD transform spectrum.The R²and RMSE of Cd element verification set were 0.636 and 0.094mg·kg^-1 respectively.The R²and RMSE of Hg verification set were 0.617 and 1.387mg·kg^-1 respectively;An optimal inversion model for Cd and Hg elements was used to evaluate the importance of the characteristic bands,and it was found that the two heavy metals were highly likely to be affected by the adsorption of Fe₂O₃ and clay minerals;By analyzing the spatial distribution map of heavy metal content obtained from inversion,it was found that Cd element is mainly distributed in the lower mountain of Xiaoqinling Mountain in the south of Tongguan County and the eastern region of Tongguan County,while Hg element showed a small regional radioactive distribution,which is consistent with the research results of other scholars.（3）Using the new idea of combining laboratory spectrum and image data proposed in this paper to model Cd content,the inversion accuracy has been improved to a certain extent.The R² of the verification set was 0.649 and the RMSE was 0.095mg·kg^-1,which showed the effectiveness of the proposed method.