Hyperspectral Image Classification Based On Dictionary Learning

Nowadays, with the development of remote sensing and computer technology, hy-perspectral image has infiltrated into various fields such as society, economy and so on. At the same time, the number of hyperspectral images is increasing, how to organize the image and classify the hyperspectral images has become an important research topic in the field of remote sensing. Due to the inherent characteristic of hyperspectral image such as high dimensionality of the hyperspectral image. Traditional machine learning methods only consider the spectral characteristics but ignoring spatial characteristics in the image. Moreover, there are "semantic gap" problem between the digital storage of hyperspectral image and human semantic understanding, efficient and reliable method-s for hyperspectral image classification is still full of challenges. Dictionary learning model provides a promising way for the hyperspectral image classification problems. In this thesis, based on the dictionary learning model, we are further mining the spatial information of hyperspectral image and hyperspectral image classification, and try to propose solutions from three aspects.Firstly, for the rich spatial information in hyperspectral images, a novel dictionary learning framework is proposed. We coding the low-level feature full of rich spatial information based on the dictionary learning framework, it can obtain the codes which can maintain the characteristics of spatial information semantic and also eliminates the "semantic gap" between low-level features and high-level image semantics.Secondly, in view of the problem that most of the algorithms in the dictionary learning framework only consider the significant features in the process of coding, a new coding strategy is proposed. We maintain the local spatial information of the dictionary entries in the dictionary generation phase at first. Then in the coding stage, we select feature for coding based on the density adaptively, it can avoid the space information loss of the hyperspectral image in the coding process and can obtain more discriminative features and better classification accuracy.Finally, in view of the problem that typical hyperspectral image sparse dictionary learning study researchers only consider the disadvantages of "nearest neighbor" spa-tial relations in code sample selection stage, we propose a spatial-constrained feature coding method by formulating it as a sparse representation problem and defining a new penalty term to take care of spatial contexts of hyperspectral images, with which the codes generated by feature coding are capable of well preserving their similarities to the spatial neighbors to achieve more discriminative power for classification.