Research On Hyperspectral Imagery Classification And Endmember Extraction Methods

Hyperspectral remote sensing is the multi-dimensional information obtaining technology,which combines target detection and spectral imaging technology together. That is, it couldobtain the two-dimensional object distribution information and one-dimensional spectralfeature characteristic information at the same time. With the increasing of spectral resolution,regarding the spectral characteristic of the objects, people’s recognition ability goes deeperthat many characteristics originally hidden in the narrow spectral bands could be discovered."Endmember" is defined as the ideal pure data, which could represent the characteristic of theobjects. Since endmember is generally used as the pre-knowledge of hyperspectral imageryprocessing methods, it plays an important role for the following processing steps， whichdepends on whether the obtained endmembers could represent the characteristic of the objects.Compare with multi-spectral remote sensing, hyperspectral remote sensing provides fruitfulspectral information, which could highlight the tiny spectral characteristic of the objects. Thewide spectral range make it possible for the user to select the special bands to highlight thecharacteristic of the objects, which could provide more original data for the hyperspectralimagery processing methods and make the precise processing of spectral information possible.However, due to the high dimension of the hyperspectral imagery, it has a huge limit whentraditional imagery processing methods are used in hyperspectral imagery. Based on the basichyperspectral remote sensing imagery processing theories and relevant subjects, this studymainly focuses on the feature reduction, endmember extraction, and the classification ofhyperspectral remote sensing imagery.Regarding feature reduction, due to the high dimension of hyperspectral data, theclassification accuracy is severely affected when there are few training samples. Featurereduction is a common method to deal with this phenomenon. However, most of the featurereduction methods can’t provide optimal feature reduction number. So this study proposes toutilize the statistic estimation characteristic of Monte Carlo random experiments to calculateoptimal feature reduction number and conduct hyperspectral imagery classification withrelevance vector machine. Experiment results show the reliability of the feature reduction number calculated by Monte Carlo method. Compare with the classification of original data,it has a significant improvement on the classification accuracy with the feature reduction data.Regarding hyperspectral endmember extraction, through the analysis of variousalgorithms, we mainly focus on the N-FINDR endmember extraction algorithm which hasbetter performance. However, the order of the samples has a certain effect on the endmemberextraction, and traditional N-FINDR algorithm also needs to reduce the dimensionality basedon the number of the endmembers, which will limit its application. In the actual hyperspectraldata, the incompact clustering of the same species presented in the high dimensional spacealso increases the difficulty of endmember extraction. So this study proposed an improvedstop rule and the pretreatment of the features, and utilizing Support Vector Machine (SVM) toconduct the second endmember extraction. Experiments show that the improved stop rulefurther increased the volume of the convex polyhedron composed of the endmembers. Thepretreatment of the features and the second SVM based endmember extraction increase theseparability of the data and the precision of the extracted endmembers respectively.Regarding hyperspectral imagery classification, fuzzy C-means clustering algorithm iswidely utilized for its simpleness and fast convergence rate. Due to the high dimensionality ofhyperspectral data, the nonlinear characteristic of the spectral bands makes it difficult for thetraditional fuzzy C-means clustering algorithm to have good clustering result in the originalspace. Moreover, fuzzy C-means clustering algorithm just uses the membership degree tocalculate the clustering center, which omits the space distribution that intrinsic exists amongthe samples. So this study proposes fuzzy kernel weighted C-means clustering algorithm.When it is used to calculate the fuzzy kernel clustering center, different weights will beassigned to each sample according to the space distribution. As a result, different sampleshave different set of kernel clustering centers. Compare with traditional fuzzy C-meansclustering algorithm, experiments results of both standard data and actual hyperspectral dataprove that the proposed fuzzy kernel weighted C-means clustering algorithm has asignificance improvement on the overall classification accuracy.Support vector machine is widely used in the classification of hyperspectral reflectancedata. In traditional SVM, features are generated from all or subsets of spectral bands witheach feature contributing equally to the classification. In the classification of smallhyperspectral reflectance data sets, a common challenge is Hughes phenomenon, which is caused by many redundant features and resulting in subsequent poor classification accuracy.In this study, we examined two approaches to assigning weights to SVM features to increaseclassification accuracy and reduce adverse effects of Hughes phenomenon:1)“RSVM” refersto support vector machine with ReliefF feature weighting algorithm, and2)“FRSVM” refersto support vector machine with fuzzy ReliefF feature weighting algorithm. Analyses wereconducted on a reflectance data set of individual maize kernels from three inbred lines and apublic data set with three selected land-cover classes. Both weighting methods increasedclassification accuracy of traditional SVM and therefore reduced adverse effects of Hughesphenomenon.