Research On Dimensionality Reduction Algorithm Of Hyperspectral Remote Sensing Image Based On Subspace

Hyperspectral remote sensing is at the forefront of the development of remote sensing technology today.Compared with traditional remote sensing technology,hyperspectral images have high spectral resolution and rich feature information,which is essential for accurate identification and classification of feature targets.However,the high amount of hyperspectral image data,strong correlation between bands and high information redundancy pose challenges for the high-efficiency recognition and classification of terrain objects.It is particularly important to reduce the dimensionality based on the effective use of hyperspectral data.In this paper,the optimization dimensionality reduction method of hyperspectral image features is proposed.According to the correlation coefficient matrix between the hyperspectral image bands,the initial band group is determined by the threshold priority constraint,which is used as a priori to determine the number of clusters and the initial cluster center.Based on the similarity measurement criteria,the K-means algorithm is used for band clustering,and the intersection of the clustering results under different criteria is used to realize the automatic division of the initial subspace.For the remaining bands that are not covered by the initial subspace,a merging strategy is adopted.By selecting the optimal dividing point to calculate the membership score of the remaining bands,the remaining bands are classified into the corresponding subspaces,and finally several complete and continuous subspaces are formed.Then PCA,MNF and KPCA transform are used to reduce the dimensionality of each subspace,judge the dimensionality reduction according to the contribution rate index and superimpose the dimensionality reduction results of each subspace,so as to obtain the final dimensionality reduction result.The PCA,MNF and KPCA transform based on the whole band of the image are used as the contrast algorithm.When the results of dimensionality reduction are verified,based on the classification separability and correlation evaluation index,two images of Washington D.C.and Pavia university are selected to classify the dimensionality reduction results by SVM and FCM.Through qualitative and quantitative evaluation of classification results and comprehensive judgment of correlation indexes,the feasibility and effectiveness of this algorithm are verified.The experiment shows that the dimensionality reduction effect based on subspace is better than that of all band.Among the segmentation algorithms,the segmentation PCA algorithm has the best dimensionality reduction effect,which can better realize the dimensionality reduction of the image and retain the original image information to a great extent,which provides the possibility for the fast interpretation of the subsequent hyperspectral images.