Hyperspectral Image Classification Based On Convolutional Sparse Representation And Composite Kernel Method

Due to the fact that hyperspectral image(HSI)contains rich spectral information of objects in the target area,traditional classification methods can perform relatively accurate classification for different land covers according to the spectral information.With fast development of imaging spectroscopy,HSI not only has high spectral resolution but also has high spatial resolution.Therefore,two novel classification methods based on spectral and spatial information are proposed.Experiments on two real hyperspectral data demonstrate the effectiveness of the proposed methods.The main contents can be concluded as the following two parts:1.By combing the spectral information with the characteristics of spatial neighborhood of pixels,a classification method based on convolutional sparse representation is proposed.Different from the idea of traditional sparse representation that approximates the testing sample through the linear combination of correlative atoms in the dictionary,the proposed method exploits the filter dictionary and the total convolution sum of its corresponding response features in the convolutional sparse representation model to approximate the testing sample with spatial information.Through the convolution,the latent information between two spectral bands can be preserved.Due to the translation invariance of the convolution,the dimensionality of the filter is much less than that of the atoms in the redundant dictionary,thus reducing the redundancy of the dictionary.Besides,the high-dimensional response values are used to replace the original spectral information to enhance the separability,leading to the improvement of the classification performance.2.In order to make full use of the spatial texture information of HSI,a classification method based on composite kernel function is proposed.Different from the traditional kernel classifiers which directly combine the spatial and spectral information of the original data for support vector machine(SVM)classifier,the proposed method adopts Gabor filters to obtain the 3D Gabor features of HSI after performing dimensionality reduction on the raw data by principal component analysis,and then combines spatial and spectral information in the 3D Gabor feature domain through composite kernel technique,followed by SVM classifier.The proposed method can't only improve the flexibility of the exploitation of spatial information,but also successfully apply the kernel technique from a new perspective to strengthen the discriminative ability.