Research On Random-Subspace-Ensemble-Based Classification Algorithms For Hyperspectral Image

Classification is one of the most important technique for analyzing hyperspectral remote sensing image.Nonetheless,there are many challenging problems arising in this research area.Two common issues are the curse of the imbalanced training sample set and high feature-toinstance ratio.In this paper,we present a new general framework to train series of effective classifiers with enhanced spatial feature subspace.This framework can achieve good classification performance effectively.First,a feature-optimization method is proposed to extract enhanced features possess relatively low dimensionality and distinctive information.We use the feature-extraction method to reduce the dimension of the original spectral data.Then,the extracted feature is projected onto a higher dimensional space to generate the enhanced feature.The extracted enhanced feature can improve the classification accuracy and efficiency of classifiers.Second,a similar-neighboring-sample-search method is proposed to address the issue of imbalanced training samples.According to an important fact: around a labeled sample,there will be more than one samples with the same label in the hyperspectral images.Therefore,in this method,the spatial adjacent samples of the real training samples are taken as the optional sample set.Then,the neighboring sample are most likely to be labeled with the actual training samples are selected as the new training samples.Similar-neighboring-sample-search method not only calculate the similarity between the real training pixel and its spatial neighbors,but also the internal consistency of the optional samples.Therefore,this method can extract the new training samples accurately,so as to balance the distribution of training samples.Then,a random subspace ensemble based enhanced classification algorithm is proposed.In this method,the similar-neighboring-sample-search method is used to balance the distribution of training sample set.Then,the enhanced spatial feature subspaces are utilized to train individual classifiers.This method can overcome the curse of high computational complexity and unstable classification performance of the original random subspace ensemble method.In addition,in order to verify the effectiveness of the feature-optimization method,we used feature-optimization method to generate enhanced spectral feature to improve the classification performance of spectral classifier.Finally,the excellent classification performance of the proposed method is proved by a variety of experiments.The relevant coefficients of the proposed method are also analyzed in experiment.By using the enhanced spatial features and enhanced spectral features,the classifiers can achieve higher classification accuracy more effectively,which further proves the effectiveness of the proposed feature-optimization method.