Research On Feature Selection For Hyperspectral Images

Hyperspectral image(HSI),with hundreds of narrow and adjacent spectral bands,supplies plentiful information to distinguish various land-cover types.However,these spectral bands ordinarily contain a lot of redundant information,leading to the Hughes phenomenon and an increase in computing time.As a popular dimensionality reduction technology,feature selection is to select the optimal band combination from the original band space.This technique tries to eliminate the redundant information between bands,meanwhile maintaining a good classification ability of the selected bands.To deal with the problem of "dimension disaster" and spectral information redundancy of HSI,based on the existing research results,two novel unsupervised band selection methods are proposed to address the problem of the dimensionality reduction.(1)In this dissertation,a band selection method for hyperspectral images is proposed based on gray wolf optimizer(GWO).The method adopts three-step strategy: band subspace decomposition,band selection and band optimization.Based on information gain(IG)and the spectral curve of hyperspectral dataset,the technique of band subspace decomposition is improved.Combined with the optimization strategy of gray wolf optimizer(GWO),the selection of optimal band combination is realized.In order to prove the effectiveness of the proposal,the HSI is classified at the pixel level using support vetcor machine(SVM)classifier and the selected optimal band combination.The proposed IG-GWO method and five competitive band selection methods are tested on three hyperspectral data sets,Indian Pines (AVIRIS 92AV3C),Pavia University(ROSIS-3)and Salinas(AVIRIS 92AV3C),respectively.The comparison results confirm the effectiveness of the band selection method proposed in this dissertation.(2)Based on the improved band subspace decomposition(ISD)and the artificial bee colony(ABC)algorithm,this dissertation develops an unsupervised band selection technique,known as ISD-ABC,to address the problem of dimensionality reduction in HSI classification.Band subspace decomposition is improved by calculating the correlation coefficients between adjacent bands and using the visualization result of the HSI spectral curve.Compared with the existing band subspace decomposition methods,the application of visual technology overcomes the disadvantage that only uses multiple minimum points of correlation coefficient of adjacent band to partition the band space.The Maximum Entropy(ME)criterion and Artificial Bee Colony(ABC)optimization algorithm are used simultaneously to maximize the information of band combination and reduce the redundancy between the selected bands,so that the optimal band combination can be obtained.To evaluate the effectiveness of the proposed method,experiments are conducted on two AVIRIS(3)datasets(Indian Pines and Salinas)and a ROSIS dataset(Pavia University).The experimental results successfully demonstrate that,compared with six other state-of-the-art band selection techniques,the proposed method has good performance in terms of experimental time and classification accuracy.