Algorithm Research On Semi-Supervised Classification Of Small Samples Hyperspectral Image With Sparse-Based Graph

At present,the related theory and application of machine learning research flourish everywhere.Unsupervised learning and supervised learning are two methods commonly used of machine learning field.However,we should also be aware that,characteristics and advantages of unsupervised learning is that it has no need of training samples,but it is hard to get good learning effect when the spatial distribution of data is rather complex.On the other hand,while the performance of supervised learning is better,supervised learning is rarely to learn the real distribution of the sample when there are little training samples.Therefore,because semi-supervised learning avoids both the drawback of unsupervised learning and supervised learning,combines both the advantages of unsupervised learning and supervised learning,thus can use a large number of unlabeled samples to assist with limited labeled sample,so as to improve the accuracy of learning.Because semi-supervised learning can get higher accuracy with less cost,it is greatly interested by people in the machine learning field both in theory and in practice.Due to the empirical success in practice,graph-based semi-supervised learning has become widely popular in semi-supervised learning community.However,a semi-supervised learning method based on graph has its particularity,composition method of graph will have great influence on the performance of the learning.Facing this problem,we proposal a novel framework to construct the graph structure for semi-supervised classification which is based on a discriminating L1 norm and KNN superposition graph(DL1KNN).We use our method in the classification of small samples hyperspectral image.In this paper,the main research work is as follows:The classification of hyperspectral image with a paucity of labeled samples is a challenging task.For graph-based methods,how to construct a framework to form a graph is the key for successful classifications.A new method of constructing graph isproposed in this paper.Firstly,construct a more discriminating L1 graph called DL1 graph based on L1 graph,then,combining it with KNN graph to solve the classification problem of hyperspectral image in the framework of semi-supervised learning.Our graph construction method contains two steps.Sparse representation is employed firstly to build the probability matrix by estimating if a pairwise pixels belonging to the same class,and this probability matrix is integrated into L1 graph to form a DL1 graph.Then,the DL1 graph is combined with a KNN graph in proportion.Experiments on Indiana Pines hyperspectral data set show that our proposed method outperforms state of the art.This paper superpose probability matrix and weight matrix of L1 graph,thus formed a strong discriminating DL1 graph.Combine local information of space with global information of spectrum by superposition of KNN graph and DL1 graph,thus build a graph-based framework which combines space information and spectral information.Such framework of DL1 KNN graph can reflects more sophisticated structure of hyperspectral image data.Experimental results show that the improvement of classification accuracy is significant when the percentage of labeled samples is 5% by using the label propagation of graph to achieve semi-supervised classification for improving automatic classification accuracy of hyperspectral data with small samples.