Graph Based Semi-supervised Learning For Hyperspectral Remote Sensing Data Classification

Hyperspectral remote sensing data has a very high spectral resolution, which can obtain the spectral and spatial information of earth surface at the same time. Unidentifiable features in the mutil-band remote sensing can be identified in the hyperspectral remote sensing. Classification for hyperspectral remote sensing plays an important role in understanding object distribution. Hyperspectral remote sensing data has a high dimension, that traditional supervised classifiers need a large numbers of training samples for hyperspectral remote sensing data classification. However, the high spectrum image acquisition categories labeled data is a time-consuming and costly work. Graph based semi-supervised classification method can solve the small sample size problem, by using a large numbers of unlabeld samples, which can improve the classification accuracy of hyperspectral remote sensing data.The main work of graph based semi-supervised method, is constructing the graph. Graph reflect the inner geometry information. Manifold learning is a nonlinear dimensionality reduction method that can grasp the nonlinear characteristics of Hypersectral image. Each Manifold learning method corresponds to an unique graph structure. Hence, manifold learning methods can be used to construct the graph structure within the graph based semi-supervised classification algorithm. Dissimilar information exists among data, but it is very difficult to obtain. Traditional method can only use the similar information. Graph based semi-supervised algorithm combine with dissimilar and similar information of data, will be helpful for hyperspectral remote sensing data classification. This paper focus on the graph based semi-supervised classification method, and the researchs of hyperspectral remote sensing data classification mainly from the following several aspects:1With the small sample size situation, K Nearest Neighbor algorithm is not very well for hyperspectral remote sensing data classification. Using semi-supervsied classification method to automatically obtain the label information of unlabeled samples, can also effectively overcome the problem of insufficient samples. There also exist some risks we must to face. Once there have a wrong classification result, such errors will be continued and expanded in subsequent use. In order to reduce the risk of error caused by semi-supervised classification, we using the probability belonging to each category as the label information. We use two groups of hyperspectral remote sensing data as the experimental data. Experiments show that using classification method to obtain the labele information of unlabeled data, is an effective and simple way to improve the classification of hyperspectral remote sensing data.2The graph based semi-supervised algorithms can also overcome the small sample size problem in hyperspectral remote sensing data classification. Comtrast to transductive graph based semi-supervised method, Laplacian Support Vector Machine algorithm can be directly used for classification after train. Howevere, LapSVM is lack of sufficient mining for those nonlinear characteristics of hyperspectral remote sensing data. Manifold learning method can be used to mining the nonlinear characteristics of data. Manifold learning algorithms can be used to construct the graph structure, in order to have a better describe of hyperspectral data and improve the classification accuracy. In this paper, we choosing three kinds of hyperspectral remote sensing data as experimental data, and focus on analyze the effects of two kinds of manifold learning algorithm on hyperspectral data classification. We found that the classification accuracy of manifold graph based LapSVM algorithm for hyperspectral remote sensing data improved significantly than original one.3There exist a lot of dissimilar information among hyperspectral remote sensing. Traditional graph based semi-supervised classification method only using the similarity information of data. In this paper, we proposed a dissimilar probability based method to measure the dissimilarity information, and construct the dissimilar propability based graph. In the dissimilar propability based graph, graph edges connect among those dissimilar points, and graph weight caculate by the heat kernel function. This method combine the dissimilar graph with similar graph by using graph Laplacian, that embedding in the graph based semi-supervised method for hyperspectral image classification. Experimental show that dissimilar information can improve the hyperspectral remote sensing data classification accuracy, especially lack of train samples.