Comparative Study On Feature Selection Methods In Construction Of Forest Biomass Estimation Model By Remote Sensing

In the field of quantitative remote sensing of forest biomass,an increasingly prominent phenomenon is that there are more and more explanatory variables.How to select the explanatory variables effectively has become an important problem.Linear regression model is one of the commonly used remote sensing models.A very important step in the establishment of linear regression model is to select interpretation variables.In this paper,SR(Stepwise Regression Method),BIC criterion(Criterions Based on The Bayes Method),AIC criterion(Criterions Based on Information Theory),Cp criterion(Criterions Based on Prediction Error),LASSO(Least Absolute Shrinkage and Selection Operator),ADALASSO(Adaptive Lasso),SCAD(Smoothly Clipped Absolute Deviation),NNG(Non-negative garrote)are discussed for the selection of variables and the stability of models in the construction of remote sensing estimation model of subtropical forest biomass.NNG and other 8 methods with variable selection ability are studied in this paper.For the purpose of comparison,two methods,OLS,RR,which are generally considered not to have the ability to choose variables,are also compared and discussed.The following factors are considered in comparison:(1)determination coefficient,prediction error,model error,etc.;(2)significant difference in determining coefficient;(3)model parameter stability;(4)variable selection stability;(5)variable selection ability.The test method was a ten-fold cross-test and repeated five times.In some evaluation indexes,the degree of freedom is considered and the degree of freedom is not considered.The results showed that BIC had the best performance and NNG,Cp,AIC had poor overall performance.In other methods,the performance of each index is quite different,while SR has a strong ability in variable selection,although it is poor in common indexes.Shortwave infrared band and its derived texture features are selected by various methods for the maximum number of times,indicating that these variables play an important role in forest biomass estimation.This study provides a new method and reference for the selection and estimation of remote sensing forest biomass remote sensing characteristic variables by comparing the performance of various methods with variable selection ability in the selection of remote sensing forest biomass characteristic variables.The research methods used in this paper are likely to change with the change of the research object,so specific analysis is needed for specific problems.