Hyperspectral Image Classification Based On Kernel Sparse Representation

With the rapid development of remote sensing technology,and the hyperspectral re-mote sensing technology occupies a very important position in many fields of military applications and civil applications.Because of hyperspectral data have the character-istics of high dimensionality,few samples,redundancy and noise bands,so the joint sparse representation method in the application of hyperspectral data processing has been widely attention.But for hyperspectral image boundary pixels and data sets of not linearly separable still need to improve the effective classification,thus improve the classification of hyperspectral image effect is of great significance in practice.This paper is based on the existing kernel joint sparse representation algorithm,the method of weighting on this basis is discussed,and proposed a weighted joint sparse representation method.Notice the image edge pixel,the neighboring pixels often con-sist of some heterogeneous pixels,such as background noise and heterogeneous pixel,and these heterogeneous pixels affect the sparse representation model.Therefore,in these under the circumstances,they are regarded as having an equal contribution to the classification process of the central test pixels,which is unfavorable for accurate classification.It is worth noting that the weight reflects the correlation between the ad-jacent pixels and the central test pixel.Each neighborhood pixel of the center test pixel is assigned an appropriate weight,which is conducive to accurate classification of the center test pixel.Using the similarity of neighboring pixels in hyperspectral images,the weights of neighboring pixels and the coefficients of sparse representations are op-timized in a regularized sparse model in the classification process of the test pixels,and a kernel sparse representation capability adaptive weight can be obtained.Only the spatial neighborhood of the test pixel is considered in the weighted kernel joint sparse representation classification method.Further discussion of the use of spatial neigh-borhood all the training and test pixels,proposed an optimized nearest regularization weighted kernel joint sparse representation classification algorithm.The proposed algorithm was tested on two hyperspectral remote sensing data set-s of the Indian Pines and Pavia University.The performance of the algorithm was measured using the evaluation criteria:overall accuracy,average accuracy,and kappa coefficient.In the two hyperspectral remote sensing image data,the effects of some pa-rameters of the algorithm on the experimental results are discussed.The experimental results show that the optimized nearest regularized weighted kernel joint sparse rep-resentation makes full use of the contribution of neighboring pixels to the central test pixels.With a certain improvement,the classification performance is better than the existing kernel sparse representation classification algorithm.