Semi-supervised Hyperspectral Image Classification Algorithm Based On Sparse Representation

With the development of hyperspectral remote sensing technology,hyperspectral remote sensing technology has attracted widely concern and high attention in geological prospecting,fine agriculture and interplanetary exploration.The classification of hyperspectral remote sensing image is also the prerequisite for image comprehension and interpretation,which is of great significance.However,the hyperspectral remote sensing image who has wide band width and the complicated spatial information has brought great challenges to hyperspectral image classification.So,how to make use of the inherent attributes of hyperspectral image to improve its classification accuracy is very important.In machine learning area,the classification methods are divided into three different categories: supervised learning,semisupervised learning and unsupervised learning in accordance with the use of labeled samples in the training classifier process.These three different methods have achieved remarkable results in the application of image classification.Therefore,this thesis proposes a variety of hyperspectral image classification methods based on sparse representation and semisupervised learning,then tests these methods on the real hyperspectral image.The main work of this thesis is summarized as follows:1.In the case of a large number of marked samples in hyperspectral image classification,the traditional classifier based on sparse representation can get good results.However,when the number of marked samples is little,the effect is unsatisfactory.On the basis of this situation,a new method of semi-supervised hyperspectral image classification based on sparse representation is proposed,which makes use of the label information of the marked samples and the structure information of the unmarked samples to improve the classification accuracy.2.The k-means clustering method simply calculating the euclidean distance of each unlabeled sample to the clustering center,does not use a small number of labeled samples' information and take the proximity between the unmarked samples and the different categories into account.Therefore,this thesis proposes a hyperspectral image classification method based on similarity and linear representation.This method makes fully use of the label information of labeled samples and the total number of the categories to predict unknown samples,so that the result of the proposed method is better than the traditional classifiers in handling small sample size problem.3.The traditional classifier based on sparse representation predicts each unmarked sample by minimizing the reconstruction error,completely ignoring the relationship between unmarked samples and their neighborhood information.So a new method of k-nearest neighbor and sparse representation in hyperspectral image classification is introduced.The algorithm judges whether the marked samples which have been abstracted from the dictionary are the neighborhood of the unmarked sample.If it is,use k-nearest neighbor to predict the unmarked samples.Otherwise,use the label propagation algorithm to predict the unmarked samples.