Hyperspectral Image Clustering Applications Based On Regularization And Collaborative Representation

Hyperspectral images(HSIs)are three-dimensional images composed of space and spectral dimension.They have rich spatial-spectral information and have been widely used in many fields such as precision agriculture,environmental monitoring,and military reconnaissance.HSIs classification is a key technology of image analysis.According to whether label information is used or not,it can be divided into supervised classification and unsupervised classification,among which unsupervised classification is also called clustering.The labeling of groundtruth in HSIs is pixel-level,which is very difficult and expensive to obtain.Therefore,clustering which does no need to use artificial labels,has attracted the attention of many scholars.Due to high dimensional,unbalanced data distribution,and complex spectral-spatial structure,clustering of HSIs is still a challenging task.Focused on the methods of regularization and collaborative representation,this paper researches on the clustering of HSIs.The main idea is to fully exploit the spatial-spectral joint information of HSIs,proposes and improve three clustering algorithms for HSIs.And the effectiveness of the proposed algorithms is demonstrated in three real data sets.The main contributions of this paper are as follows:(1)A new method,locally constrained collaborative representation based Fisher's LDA for clustering(LCR-FLDA),is proposed for the clustering of HSIs.The FLDA model is established by combining the two classical traditional clustering algorithms,K-means and spectral clustering,in the form of maximum Rayleigh entropy to maximize the between-cluster scatter and minimize the within-class scatter.In order to further utilize the local information of HSIs,a similarity matrix of locally constrained collaborative representation is constructed and used in the FLDA model,which improves the representation ability of spatial-spectral information and the discriminativeness of clustering features.The clustering experiments on the hyperspectral standard test data sets prove the effectiveness of this algorithm.(2)We propose a new clustering method called dual graph regularized based collaborative representation clustering for HSIs(DGCRC).Based on the collaborative subspace clustering method(CRC),two graph-regularized model are established which constructs superpixel spatial context structure graph constraint item and spectral neighborhoods graph constraint item.The analysis shows that by adding these two regularization terms,the high possibility that neighbor pixels in the space and spectral domain belong to the same category is constrained.Experiments on three data sets show that the results are better than several traditional clustering algorithms.(3)We propose a deep collaborative subspace clustering method based on non-local neural network for HSIs(DCRC-NLN).On the basis of the autoencoder,the non-local operation module and the collaborative self-representation layer are added to fully integrate the spatial-spectral features of HSIs and increase the discriminability of the features.Design a new joint loss function which is composed of three parts: the reconstruction error loss function,the collaborative self-representation loss function,and a perceptual loss function.It improves the network's reconstruction ability and feature learning ability together.Through experimentations and comparison with other clustering algorithms,this algorithm has excellent clustering performance.