Research On Hyperspectral Image Ensemble Classification Based On Dominant Set Clustering

For hyperspectral images with less number of samples,a large dimension data lead to poor classification precision and low stability of the algorithm,a hyperspectral image ensemble classification algorithm based on Dominant Set clustering is proposed(Dominant Set Clustering Ensemble Algorithm,referred to as DSCEA).First of all,with the method of random,from the original training samples of all the band selection more groups of the same number of band subsets,extract the corresponding bands of two-dimensional image of multiple image collection.Secondly,the use of DS clustering methods of band selection,image collection of the three dimentional image gradient is calculated,construct local empty spectrum coherence function,selected characteristic subset to represent the original image information and redundancy between informative.Thirdly,with support vector machine as the base classifier,the feature subset after band selection is taken as the new training sample,and a single classifier with difference is trained to classify the target samples.Finally,the simple and efficient majority voting method is used to integrate the classification results of the single classifier,ensemble the final classifier,and dig into the image information deeply while realizing the stable classification of the target samples.The DSCEA algorithm was tested on two classic hyperspectral data sets,the ensemble factor is the band number and number of classifier,the comparison algorithms are the support vector machine(SVM)used separately,the support vector machine(DS_SVM)clustering via DS,the K nearest neighbor(DS_KNN)and the random forest(DS_RT)classification algorithm respectively.the experiment uses the classification accuracy and Kappa coefficient as evaluation index of algorithm,the results show that DSCEA algorithm to complete mission of hyperspectral image classification,and the algorithm has high accuracy and stability.