On Classification Method Of Hyperspectral Images

With the fine spectral resolution at about 10nm, the hyperspectral images could be used to discriminate and detect the land-cover types which could not be discriminated with traditional panchromatic and multi-spectral images. Hyperspectral remote sensing system has been playing an important role in the advanced earth observation remote sensing system in many countries. It provides new ideas and tools for observation of land, ocean, and atmosphere. However, it also brings some new problems, such as how to solve the high dimensionality and small sample problem, how to eliminate the influence of variance between and within different classes, how to realize real-time processing of such huge number of data. Based on the analysis of hyperspectral image properties and state-of-art classification methods, four theory and technique problems of hyperspectral image classification are studied in this paper.First, for the spectral-metric-based methods, a multiresolution spectral angle (MSA) method is proposed to solve the within-class dissimilarity problem of the spectral angel mapper (SAM) method and its revised methods. Based on the pairwise classification framework (PCF), the original multiclass classification problem is decomposed into a set of two-class problems. A class separability criteria is proposed to simultaneously maximize the average between-class spectral angle and minimize the average within-class spectral angle. For each two-class problem, the original set of bands is recursively decomposed into band subsets with large class separability. The band subsets with larger class separability are selected to generate subangles, which will be combined to measure the similarity. The MSA method can effectively eliminate the influence of with-class dissimilarity and obtain significantly better classification accuracy than traditional spectral metric-based methods.Then, to solve the problem of classification using single feature set of hyperspectral images, a multiple feature sets representation and fusion framework is proposed. It uses the complementary information of spectral magnitude and spectral shape and increases the discriminant capability of classes. To evaluate the effectiveness of the framework, fusion methods for SAM and maximum likelihood classification (MLC) are proposed respectively. To improve the capability of the framework more effectively, a novel stacked support vector machine (SSVM) fusion framework is proposed. The experimental results show that it can improve the classification accuracy of hyperspectral images significantly.And then, for the multiresolution analysis of hyperspectral images, discrete wavelet transform (DWT) is adopted to exploit the fine-scale and large-scale information of hyperspectral signals. Based on the analysis of discriminant information of wavelet features at different layers, a fusion method based the probability of support vector m machines (SVMs) is proposed. To better use the complementary diversity of wavelet features, we generalize the SSVM to fuse the wavelet features, as well as band and wavelet features. The experimental results show that the fusion methods are much more effective than direct usage of wavelet features, and can improve the classification accuracy of hyperspectral images significantly.Finally, based on the analysis of current multiclass strategies, an "elimination game" pairwise decision tree method is proposed and generalized to solve the multiclass SVM problem. Experimental results show that it can keep the high accuracy of "round robin" one-against-one (OAO) method, but only needs about 50% test time of OAO. To obtain a much faster multiclass strategy, according to the two factors that directly affect the classification, i.e., the number of two-class problems and the number of support vectors, a fast adaptive binary tree is proposed. Experimental results show that it only needs about 25% test time of OAO, but exhibits less than 0.5% descended classification accuracy.