Collaborative Representation Of Hyperspectral Remote Sensing Image Classification By Fusion Of Spectral And Spatial Information

Hyperspectral image(HSI)has cube form and contains abundant two-dimensional spatial information and three-dimensional spectral information.High-volume spectral bands can effectively distinguish various materials with more details,so HSI is also widely used to detect and characterize land cover types.Classification of HSI is essentially a nonlinear problem,which is mainly challenged by high dimension,lack of labeled samples,spatial variability of spectral information and image quality.A significant improvement in HSI classification can be achieved with appropriate spectral-spatial fusion methods.Spatial context information plays an important role in the classification of hyperspectral images,which can effectively avoid the phenomenon of foreign objects in the same spectrum or different spectra caused by the use of spectral information only.The extraction of spatial features is usually carried out from the neighborhood region by considering the fixed or adaptive size of each pixel window.On this basis,when applying the representation-based method to integrate spectral information and spatial information for classification,there are two main challenges: selecting the appropriate window size for the neighborhood,and the need for a large number of training samples.Therefore,this paper proposed a classifier based on spatial peak sensing probabilistic cooperative representation(SPa Pro CRT)by considering spectral spatial information between superpixel clusters in the framework of probabilistic cooperative representation,and conducted experiments with limited training samples on two data sets to verify the superiority of the model.The specific work of the paper is as follows:(1)Combining the advantages that the probabilistic cooperative representation based classifier(Pro CRC)can effectively utilize training samples from all classes to derive the class labels of test samples,a classifier based on spatial peak sensing probabilistic cooperative representation(SPa Pro CRT)was proposed.The spectral spatial information between superpixel clusters is included in the probabilistic collaborative representation framework,and the probability that the test samples belong to each category is jointly maximized.The key steps are: the original HSI is segmented by superpixel;The spectral distance and spatial correlation of the superpixel cluster corresponding to the training sample and the test pixel are fused.The spectral inducible regularization term,spectral-space feature inducible regularization term and class-specific representation regularization term are included in the cooperative representation.Finally,the classification is based on "representation residual".(2)1% samples and k(k = 5,10,20,50)samples of each type were taken as training samples in two data sets: Indian Pines and Pavia U,respectively.The proposed spatial peak sensing probability classifier SPa Pro CRT was compared with other related classifiers(CRT,Pro CRT,Sa CRT,Sa Pro CRT,SPa CRT)to confirm the advantages of SPa Pro CR classifier in parameter setting and classification effect under the condition of limited samples:Regularized parameter Settings are not sensitive to changes in the data set.In the superpixel segmentation,the compromise coefficient between the average width,space and spectral distance of the superpixel has a large advantage range.Sa Pro CRT has the best performance in both overall accuracy and Kappa coefficient,and the classification map obtained is the closest to the reference map.In addition,due to the high solution efficiency of Pro CRC,the proposed SPa Pro CRT model will not increase the computational burden when class-specific representation regularization terms are added for solving compared with SPa CR model.