Research On Hyperspectral Image Classification Method

Hyperspectral image data is a synthesis of spectral dimension information data and spatial information data,which contains rich and complex information of feature class.The purpose of hyperspectral image classification is to use the classification related technology and theory to fully excavate the information of objects,improve the accuracy of hyperspectral image classification,and provide a solid and reliable information base for the subsequent hyperspectral image application.At present,hyperspectral image classification technology has deeply affected the aspects of modern life,its application in the fields of agriculture,forestry,military,marine and geology has become more and more extensive and mature.In this paper,the traditional hyperspectral image classification algorithm for spectral information and spatial information is not sufficient as the starting point,than the close coupled set of pixels-based adaptive boosting classwise sparse representation classifier for robust hyperspectral image classification algorithm and the block-nearest classifier based boundary constraint algorithm for classification of hyperspectral image algorithm are proposed.The main contents of this paper are as follows:1.Close coupled set of pixels-based adaptive boosting class-wise sparse representation classifier for robust hyperspectral image classification is proposed.In this algorithm,the adaptive-boosting idea is integrated into the sparse representation of the orthogonal matching process.In the iterative optimization of orthogonal matching,the extraction of the decision feature is strengthened,and the close coupled set of pixels is produced in the feature domain created by the local Fisher discriminant analysis.The generated close coupled set of pixels is to smooth the initial feature distribution,and prevent the over-fitting the previous iterative process may produce.Experiments on real hyperspectral images show that the algorithm has better classification performance than similar algorithms.2.The block-nearest classifier based boundary constraint algorithm is proposed.Firstly,the weighting factor is calculated by using the degree of polymerization of the training block and the degree of polymerization of the test sample block.Then,the weighting factor is used to calculate the distance between the blocks,than the label is output according to the distance.At the same time,the algorithm uses the local Fisher discriminant analysis algorithm to reduce the original hyperspectral spectrum to reduce the single band grayscale,then use the local binarization to process the pixel boundary snapshot,and finally use the acquired boundary snapshot information with Level smoothing operation for the label to output the final classification label.The experimental results of the algorithm on three real hyperspectral images show that the classification effect is superior to the similar algorithm.