Empirical Model Selection And Feature Extraction For Retrieving Soil Heavy Metal Concentration From Airborne Hyperspectral Imagery

In this paper,soil heavy metals retrieval from unmanned aerial vehicles(UAVs)near-infrared hyperspectral is the research goal.Three aspects of heavy metal spatial correlation,model selection and feature selection were used to enhance the generalization ability of inversion model and improve the accuracy of soil heavy metal inversion.(1)The spatial correlation and spatial heterogeneity between five heavy metals(Cr,Fe,Cu,Zn,Cd)in the three study areas were analyzed.The spatial distribution of soil heavy metals in the study areas at different spatial scales are not affect-ted by the scope,and also existed on large spatial scales,which provides feasibility for large-scale inversion of soil heavy metal concentration;(2)In the model selection,this paper used the experimental design of the system to verify the inversion accuracy of the three models on five heavy metals.Partial estimation showed that partial least squares(PLSR)was superior to support vector regression(SVR)and artificial neural network(ANN).Experiments show that the inversion of heavy metal concentration on different scales requires model selection and for the optimal model;(3)In the feature selection,the two study areas of Malanzhuang were used to evaluate the improvement of the accuracy of heavy metal inversion by principal component analysis(PCA)and minimum noise fraction(MNF).Experiments show that feature extraction has different degrees of improvement on the accuracy of heavy metal inversion.It is necessary to compare the sensitivity of different heavy metals to the feature selection through experiments.This paper constructed the optimal model of five heavy metals in three study areas,and generated thematic maps with the best model,which provides technical support for the selection of models and features of soil heavy metal retrieval with hyperspectral imagery in different spatial scales.