Dimension Reduction Using Spatial And Spectral Features For Hyperspectral Image Classification

Hyperspectral images are more suitable to the classification of surface feature than multispectral images,because the abundant spectral resolution of it can reflect different material characteristic of the surface.Thus,the research on hyperspectral images has been gained more and more attention.Although the accuracy of a classifier to detect and identify various types of surface features can be improved because of the feature of high dimensional data,there are large cost of determining the feature category,lack of training samples,high expectation of the algorithm's running time and questions caused by a large number of spectral bands such as "Huges" phenomenon and so on.The above problems can be effectively solved via sparse representation.From the basic sparse theory,the spectral and spatial information are integrated in the sparse representation in this paper.Meanwhile,the principal component analysis(PCA)is adopted to reduce the dimension of data.Finally,the based on the MKSOMP algorithm framework combining the spatial and spectral characteristics is proposed to classify the hyperspectral image.The main contents are as follows:Firstly,it is insufficient that using only the spectral information to measure the similarity between pixels and reduce the dimension of data due to the characteristics of typical high dimension and less training samples.Numerous have confirmed that the classification result of hyperspectral images can be greatly improved by the integration of spatial information.Therefore,based on the weighted mean filter(WMF)method combining spatial information,the spatial and spectral features are fused to give new hybrid pixels.Then,the PCA is adopted to reduce the dimension of hybrid pixels and the based on spatial and spectral WMF-PCA algorithm is proposed to mitigate the "Huges"phenomenon.Secondly,the KSOMP algorithm using spatial feature is adopted for classification because of the importance of spatial information for hyperspectral image classification.A fixed-size sliding window is used in KSOMP algorithm.However,it is a time-consuming to determine the optimal size of the window.Thus,a based multi-sliding window MKSOMP algorithmis proposed to avoid the problem in this paper.Finally,a system related to the algorithms involved in the paper is designed and implemented in order to facilitate the classification of hyperspectral images and parameter setting.In this section,the modular design is adopted to effectively extend the system and increase the flexibility of it.In this paper,the experiments comparison is conducted on the public dataset Indian Pines.The experimental results show that the proposed hyperspectral classification framework can improve the accuracy of classification.