Research On Spatial-spectral Collaborative Classification For Hyperspectral Image

Classification is an important branch of hyperspectral image processing,its goal is to assign the land-covers classes to pixels contained in the image,the common classification includes three groups i.e.,unsupervised classification,supervised classification and semi-supervised classification.With the rapid development of high-tech,the requirement of the classification accuracy is gradually increasing.In addition,sensors that are able to acquire remote sensing images with both fine spatial and fine spectral resolutions provide a powerful development support for high-accuracy hyperspectral image classification.Nowadays,classification technology develops extremely rapidly.However,the hyperspectral image classification is facing an acid test due to the characteristics of the hyperspectral data.At the same time,the traditional classifiers has obvious shortcomings,that is,the traditional classifiers always purely depend on the spectral information without consideration of the collaborative effect of the spatial information.Therefore,several new classification methods are proposed in this thesis to solve the problem that how to select the representative unlabeled samples,how to obtain the classes of the unlabeled samples and how to use the spatial information to guide the construction of the dictionary,which effectively overcome the problems caused by the small-size training set based on the spatial information.The research works are described as follows:Firstly.spatial-spectral information-based semi-supervised classification(S2ISC)is proposed.The spatial information extracted by a 2-D Gabor filter was stacked with spectral information first.and the spatial neighborhood information of labeled training samples was then combined with active learning algorithm to select the most useful and informative samples.which were used as the unlabeled set to aid the probability model-based supervised support vector machine.This method fully makes use of the spatial information of the hyperspectral image from the unlabeled training samples selection guiding by the spatial information viewpoint.Secondly.spatial-spectral label propagation support vector machine(SS-LPSVM)is proposed for semi-supervised classification of hyperspectral image.2-D Gabor filter was used to extract the spatial information,and the obtained spatial features were stacked with the original spectal features.The spatial graph based on spatial smoothness was constructed,and labels were propagated from labeled samples to unlabeled samples with spatial-spectral graph to update the training set for a basic classifier with the improved classification accuracy.This method fully makes use of the spatial information of the hyperspectral image from the spatial-spectral graph viewpoint.Finally,spatial-dictionary based multiple-mapping-kernel collaborative representation classification(SMCRC)is proposed.The representative and discriminative spatial dictionary was learned via the spatial-spectral graph of the sparse coefficients.The spatial information was further integrated into the CR framework via the multiple-mapping kernel framework to yield higher classification accuracy with much lower computational cost.This method fully makes use of the spatial information of the hyperspectral image from the construction of the spatial dictionary and the multiple-mapping kernel framework viewpoint.