| Over the last decade, hyperspectral remote sensing images have been widely applied for the identification, and classification of the land-cover classes in the agriculture, military affairs and industry. Due to the fact that the hyperspectral sensors capture the radiance of materials in a huge number of contiguous spectral bands, hyperspectral images provide a large amount of space, radiation, and spectral information about a scene, which are very important for land-cover classification with improved accuracy and robustness. In this paper, hyperspectral image(HSI) classification as the application background utilizes spectral information and spatial information of hyperspectral image to classification. The author’s major contributions are outlined as follows:Firstly, aiming at “same material with different spectrum” and “different materials with same spectrum”, this paper proposes a hyperspectral image semi-supervised classification method based on spatial neighbor discriminative graph. Graph based semi-supervised classification has attracted huge attention from various research fields, in which graph construction is the essential part. In this study, we propose to build a novel graph called spatial neighbor discriminative graph for graph-based semi-supervised classification. In which the discriminative information and the spatial layout information are introduced into the graph construction. Then label propagation from the labeled data to the unlabeled data through the graph. More specifically, the discriminative information by enlarging the Euclidean distance between points belonging to different classes consequently obtains more accurate local information. And the spatial layout information via the nonparametric adaptive clustering approach----Mean Shift----to be generated, which make a pixel close link with other pixel in the same patch. By combining the two kinds of graphs, a joint spectral/spatial classifier is presented to deal with hyperspectral image classification problem. Our approach is tested on three hyperspectral images and compare with two supervised methods and two graph-based semi-supervised method. The experiment results demonstrate our approach can get a competitive classification rate.Secondly,Aiming at lots of hyperspectral image classification methods only use one single feature, which only depict the hyperspectral image from one perspective and ignore complementary information from different features, this paper proposes ahyperspectral image classification method based on multi-task low-rank representation. To introduce spatial information, the method uses super pixels as the classification units. We combine two features – spectral feature and spectral gradient feature – to represent a superpixel’s information more fully. Different from low-rank representation, multi-task low-rank representation aims at inferring a unified affinity matrix from two feature spaces, and take advantage of the cross-feature information. Several experiments are done on hyperspectral data, and the result show that this method is superior to other methods.Thirdly,Kernel propagation(KP) is an effective semi-supervised kernel matrix learning(SS-KML) method which improves the comprehensive performance. This paper proposes a hyperspectral images semi-supervised classification based on KP. KP first learns a kernel matrix from labeled and unlabeled samples, and the kernel are combined with the RBF kernel to define a compound kernel. Compound kernel maps the data set into a higher dimensional feature space. Support vector machine(SVM) is chose as the classifier to predict the label of test samples. KP are effectively used for dealing with nonlinear problem. A learned kernel can fit the given data better than a predefined kernel. The effectiveness of our proposed method is evaluated via experiments on hyperspectral dataThis work was supported by the National Natural Science Foundation of China(No. 61272282), and the Program for New Century Excellent Talents in University(NCET-13-0948). |
PDF Full Text Request