As the occurrence of hyperspectral images has brought a new way for the observation of the earth, it has also introduced new topics for remote sensing image processing. Among all the topics related to hyperspectral images, classification is a perennial one. As a matter of fact, the research on classification of hyperspectral images has a history of years. However, due to the limitation of hardware and processing techniques, there are still lots of problems unsolved. Besides, with the emergence of new techniques and their development in the field of machine learning and image processing, the processing ways for classification of hyperspectral images have been gradually enriched and propelled to a higher stage. This thesis starts with an overview of literature for classification of hyperspectral images, presenting some research focuses as well as pointing out some issues that are difficult to solve. Practical works are then carried out and concentrated on several selected topics. Having the principles in the field of machine learning and image processing as theoretical basis, this thesis have actively integrated these principles with the characteristics of hyperspectral images. As a result, considerable innovative ideas are contributed and several new algorithms are proposed, which are practical for classification of hyperspectral images. The main contributions and novelties can be summarized as follows. 1. A semi-supervised classification method based on graph cut is proposed. In this method, the graph cut algorithm is introduced into classification tasks of hyperspectral images and the fuzzy version of support vector machine (SVM) is adopted to overcome the problem of insufficient training samples. In real hyperspectral classification applications, the number of training samples for each class is often limited. As a result, the graph cut algorithm can not be directly applied in these classification tasks. To overcome this difficulty, a two-step framework is proposed. In this framework, a fuzzy support vector machine is employed with these limited training samples, as a preliminary classification stage. For these preliminary results, we reserve those samples with higher posterior probabilities belonging to certain classes as reliable ones, while others are rejected for further classification. In the second stage, the reliable samples are fed as training samples to the input of the graph cut algorithm. With the establishment of a more subtle mode...
修改评论