Feature Extraction Of Hyperspectral Image Based On Spatial-Spectral Manifold Learning

The hyperspectral remote sensing images realize the “integration of maps and spectra”,which provides the possibility for accurately realizing the recognition and classification of the ground objects.However,the higher spectral resolution result in a large amount of bands and complex data,which leads to a "dimensionality disaster".Therefore,how to keep the effective information in the original data as much as possible while reducing the data redundancy is the key to the classification of hyperspectral remote sensing images.Feature extraction is an effective way to solve the "dimensionality disaster".Through feature extraction,data redundancy can be reduced,more explanatory features can be obtained,and the classification effect can be improved.In view of the facts that most traditional methods assume that hyperspectral data is distributed in a single manifold structure and the complex multi-manifold characteristic is not considered,and only relying on spectral information while ignoring spatial information.In the paper,taking the internal structure of hyperspectral images data as a viewpoint,and combining the graph embedding theory in manifold learning,carrying out the research on feature extraction methods for hyperspectral remote sensing images.The main work is as follows:(1)Introducing research status of hyperspectral imaging technology and feature extraction,analyzing the existing feature extraction methods,and finding that most methods fail to fully discover the multi-manifold property of hyperspectral data,and also ignore the spatial consistency,as a result,the classification performance of ground objects is limited,which provides the theoretical foundation for the new methods.(2)A supervised multi-manifold discriminant embedding(SMMDE)feature extraction method is proposed.According to the inherent multi-manifold structure distribution of hyperspectral data.First,the data is divided into sub-manifolds to obtain corresponding submanifold structures of different categories of data;secondly,the intra-manifold and intermanifold graphs are constructed to represent the manifold structur.And defining the multimanifold distance;finally,by maximizing intra-manifold distance to disperse the data on and minimizing inter-manifold distance to gather the data,to extract low-dimensional features on each manifold.The experiments on the PaviaU and KSC hyperspectral data sets prove that the SMMDE can effectively characterize the multi-manifold structure of hyperspectral data and improve the accuracy of classification.(3)A supervised spatial regularization manifold discriminat analysis(SSRMDA)feature extraction method is proposed.In the light of the shortcomings of traditional feature extraction methods that only use spectral information without considering spatial features,based on graph embedding theory in manifold learning,the paper fully integrates spectral-spatial information and studys a new spatial-spectral method.The method extracts spatial features through entropy rate superpixels,and then constructs the neighboring graphs in the spectral domain and spatial domain according to the graph embedding model,and using the label information to enhance the data aggregation in the class for extracting more discriminat features.Experiments on Indian Pines and Washington DC Mall hyperspectral data sets show that the SSRMDA has a better classification effect than related methods.To sum up,for the problem of "dimensionality disaster" in the hyperspectral classification,two new feature extraction methods are proposed based on manifold learning.The validity of the methods are verified on the hyperspectral data sets.