Research On Semi-supervised Hyperspectral Image Classification Method Based On Sparse Representation Graph

Hyperspectral image(HSI)classification refers to the task of determining the category of each pixel by analyzing the spectral characteristics of the objects,thereby dividing the objects into different types of regions.It is an important way for people to obtain useful information from remote sensing images.HSI classification often faces the issue of limited number of labeled data,which are often costly,effortful and time-consuming to label.On the other hand,we can obtain a large number of unlabeled data effortlessly.Semi-supervised learning(SSL),which can utilize both small amount of labeled instances and abundant yet unlabeled samples,has recently been proposed to deal with this issue.Among current SSL methods,graph-based SSL methods are particularly appealing since they have elegant mathematical formulation and can obtain a close-form solution.The key of the graph-based semi-supervised HSI classification is to construct a graph on the hyperspectral image data that can truly reflect the similarity relationship between samples.For high-dimensional data such as hyperspectral images,due to the high dimensionality and non-linearity of the data,the traditional distance metrics can not represent the similarity relation between them well,thus cannot construct a good graph structure.In order to solve this problem,this dissertation focuses on how to learn an informative and discriminative graph structure under the framework of sparse represntation model,the leared graph was hoped to be able to fully exploit the spectral and spatial information of hyperspectral data,thereby truly reflecting the internal structure of hyperspectral data.Firstly,in this dissertation,we use a small amount of label information to estimate the probabilistic class structure of data,which can denote the probabilistic relationship between each sample and each category.Then,we incorporate this probabilistic class structure information into the sparse representation model,and propose a novel graph construction method called probabilistic class structure regularized sparse representation(PCSSR)graph for semi-supervised HSI classification.The proposed graph structure penalizes samples with different class distributions,i.e.,assigning smaller weights to them,which truly reflects the internal structure of the data.The experimental results show that the proposed PCSSR graph obtains superior performance aompared with the state of art graph-based SSLmethods.Secondly,since the conventional graph construction method does not consider the rich spatial information in hyperspectral images,we improves the sparse representation graph from the following two aspects:(1)assigning similar weight coefficients to the spatially neighboring pixels;(2)assigning smaller weight to the samples of different class distribution,and propose a more informative and discriminative graph called spatial and class structure constrained(SCSSR)graph for semi-supervised HSI classification This graph structure not only makes the weights among the samples with different class distribution as small as possible,but also guarantees the representation coefficients of the neighboring pixels to be similar,thus improving the discriminability of the graph significantly.Experimental results on serveral hyperspectral image datasets show that the proposed method can effectively improve the performance of semi-supervised HSI classification.Finally,for the existing graph construction method is either obtained directly from the original data or is derived from the self-expression coefficient of the data,in this dissertation,we attempt to construct a graph structure from representation space,and propose a discriminative graph construction method based on the representation space,which can learn the representations of samples and the similarity matrix of representations simultaneously.The proposed graph structure is constructed according to the distance between the representations of the samples,the latent assumption is that samples should have a lager probability to share the same label if their representations have a smaller distance.Moreover,we explicitly incorporate the probabilistic class relationship between sample and class into the model.Such a priori information can guarantee that the samples with smaller distance of class distribution can have bigger weights.After taking into account the distance between the representations and the class distribution of samples,we can obtain a more informative and discriminative graph.The experimental results show the validity of the proposed graph construction method.