Research On Learning Algorithms For Spectral-spatial Classification Of Hyperspectral Images

Classification is one of the most important research content in hyperspectral image analysis. Hyperspectral images can provide very rich spatial and spectral information. The rich information has brought opportunities and challenges for hyperspectral image classifi-cation. There mainly exist several critical issues need to be addressed:1) the high dimen-sionality of hyperspectral images; 2) the small amount of available labeled samples; 3) the spatial variability of spectral information; 4) the quality of hyperspectral images. Here, the high dimensionality of hyperspectral images and small amount of available labeled samples may pose the Hughes phenomenon, while the spatial variability of spectral information and the quality of hyperspectral images can lead to another phenomenon, that is, the same ground objects have different spectral information, and different ground objects have very similar spectral information. According to the characteristics of hyperspectral images, this thesis is to conduct in-depth research on spectral-spatial classification of hyperspectral images by exploiting supervised, semi-supervised and unsupervised learning algorithms.Spectral-spatial classification is an important research topic of hyperspectral image classification. Some classical hyperspectral image classification algorithms generally only exploit the spectral information. In order to obtain a good classification effect, they gen-erally need sufficient labeled samples. However, in practice true samples labeling is very difficult. Introducing the spatial information cannot only make up the lack of samples la-beling information, but also effectively overcome the phenomenon that the same ground objects have different spectral information, and different ground objects have very similar spectral information. The Gaussian weighted local mean operator is first proposed, which can extract the spatial information of each sample by exploiting the similarity of the spec-tral information in its neighborhood. Due to the high dimensionality of the spectral infor-mation, the dimensionality of the integrated information with the spatial information can become higher, which leads to the serious information redundancy and curse of dimension-ality. To solve the problems, composite kernels discriminant analysis is proposed for feature extraction. This algorithm implements the different combinations of the spatial and spectral information by the stacked kernel, series kernel and parallel kernel, and then extracts the corresponding features by the nonlinear discriminant analysis for supervised learning clas-sification. In the case of very small amounts of labeled samples, compared with the popular composite kernels and generalized composite kernels algorithms, the proposed supervised learning spectral-spatial classification algorithm can extract the more effective discriminant features, and obtain better classification accuracies.The lack of labeled samples is one of the most important factors that directly affects hyperspectral image classification. Thus, for supervised learning algorithms, achieving the high-accuracy classification of hyperspectral image is still a very difficult task. In hyperspec-tral image classification, although true labeled samples are difficult to collect, the unlabeled samples are easy to obtain. Thus, a semi-supervised learning spectral-spatial classification framework is proposed. This framework can learn new labeled samples by adopting the semi-supervised learning algorithm for making up the lack of labeled samples. Compared with other popular semi-supervised learning classification algorithms, the proposed frame-work has a significant advantage in terms of the classification accuracy.Although the proposed framework has achieved a good classification accuracy, there are several aspects that need to be improved. Thus, this thesis puts forward a spectral-spatial classification model integrating the unsupervised, semi-supervised and supervised learning algorithms. In this model, the principal component analysis is first used for dimensionality reduction of hyperspectral images and extracts the few first principal components. Secondly, a semi-supervised learning algorithm is applied to the separation of easily separable samples and difficultly separable samples. Next, the highly confident set is selected from the easily separable samples by utilizing the spatial neighborhood information and the dilation operator in mathematical morphology. Then the proper number of high confident samples are added into the training set for the classification task of the test set, which consists of difficultly separable samples. Finally, a supervised learning spectral-spatial classification algorithm is designed for the classification task of difficultly separable samples. The experiment results show that the proposed model can achieve the better classification effect with less time than several popular semi-supervised learning classification algorithms in hyperspectral image processing.