Hyperspectral Image Spatial-Spectral Feature Extraction And Classification Based On Manifold Learning

Hyperspectral Image(HSI)has detailed spectral information and spatial information,and it provides a new opportunity for the fine recognition and classification of the ground objects.However,the high spatial-spectral resolution also bring the problems of large amount of data,high redundancy,many bands and strong correlation between spectral bands,it is very easy to produce the "curse of dimensionality" if it is classified directly.Therefore,it is an urgent problem to reveal the useful discriminant features from HSI,to improve the classification accuracy of the ground objects,and to reduce the computational complexity.Based on the nonlinear data structure of HSI,this paper mainly studied from two aspects of manifold learning and spatial-spectral feature fusion for feature extraction and classification methods of HSI.The main research works are as follows:(1)Several typical feature extraction algorithms of HSI are summarized from the aspects of global linear,nonlinear manifold learning and linear manifold learning.And it also introduced the commonly used classification methods,evaluation criteria of classification results and some hyperspectral data sets that were used in this paper.(2)A new feature extraction algorithm called Spatial-Spectral Coordination Embedding(SSCE)and a new classifier called Spatial-Spectral Coordination Nearest Neighbor(SSCNN)are proposed in this paper.Aiming at the problem that the traditional classification methods just apply the spectral information while they often ignore the spatial information,based on the theory of graph embedding in manifold learning,the new algorithm fully integrates the spatial-spectral information in HSI.Firstly,the proposed algorithm defines a new similarity measurement method called Spatial-Spectral Coordination Distance(SSCD),and it can reduce the probability that heterogeneous objects are selected as nearest neighbors;Then,it constructs a spatial-spectral neighborhood graph and enhances the aggregation of data through raising weight of the spatial-spectral neighbor points to extract the discriminant features;At last,it uses the SSCNN to classify the reduced dimensional data.The classification experimental results on PaviaU and Salinas data sets show that SSCE can get better classification results than other methods,and can effectively improve the classification accuracy.(3)A new feature extraction algorithm combining Weighted Mean Filter(WMF)and Manifold Reconstruction Preserving Embedding(MRPE)is proposed in this paper.According to the spatial consistency property of HSI,firstly,the proposed algorithm applies WMF to the all pixels in the image which can reduce the spectral difference of data points from the same class;Then,the discriminant features are extracted through enhancing the weights of the spatial neighbor points in the manifold reconstruction.The classification experimental results on PaviaU and Urban data sets show that in the same experimental conditions,MRPE has better classification results.In summary,this paper mainly studied the hyperspectral image spatial-spectral feature extraction and classification based on manifold learning,and proposed two new algorithms.The effectiveness of the proposed methods are verified by the hyperspectral data set.