Research On Classification Methods Of Hyperspectral Image Based On Spatial-spectral Joint Features

Hyperspectral image is widely used in many fields such as precision agriculture,medical diagnosis and resource exploration because it has the characteristics of high spectral resolution imaging and integration of spatial image and spectral image.The classification performance of hyperspectral image directly affects the overall effect of hyperspectral image application.Therefore,how to extract valuable information from massive data and achieve accurate classification of hyperspectral image is a frontier subject in hyperspectral image processing technology when labeled samples are limited.Some problems exist in hyperspectral image classification such as the longer time consumed in the methods based on deep learning,the poor performance owing to some spatial-spectral join feature methods using insufficient spatial information,and the lower few-shot classification accuracy due to the scarcity of labeled samples.To solve these problems,methods of spatial-spectral joint features classification for hyperspectral image are investigated by using broad learning,sparse representation and collaborative representation.The main contributions are as follows:1.Deep learning methods contribute to improving classification performance by extracting enriched spatial-spectral features.However,deep learning methods require training a large number of network parameters,and at the same time a high-performance computer to work for a long time for getting the better classification performance.In order to ameliorate the above situation,a novel method of spatial-spectral join features classification for hyperspectral image based on local binary pattern and broad learning is proposed.Firstly,the local preserving projection is utilized to reduce the spectral dimensionality of hyperspectral image to extracting the spectral manifold structure features.Secondly,the local binary pattern operation is used to extract local gray-scale and rotation-invariant texture features in the spatial domain of each band of dimensionality-reduction spectral information,so as to achieve the fusion of spatial characteristics and spectral features.Finally,the spatial features fusing with spectral information are input to broad learning system for further sparse processing,and the optimal connecting weights which are used in the final output layer of broad learning are achieved by ridge regression approximation algorithm to complete the classification of hyperspectral image.2.The insufficient use of spatial information in some hyperspectral image classification methods based on spatial-spectral joint features rusults in poor classification performance.In order to fuse deeply spatial features and spectral characteristics for improving the classification performance,a spatial-spectral joint features hyperspectral image classification method based on guided filtering and broad learning system is proposed.Firstly,the Gaussian filter is used to smooth each band of the original spectrum to remove noise based on the spatial information,so as to extract inherent spectral features fusing with spatial information.Secondly,the broad learning network is used to sparse the inputted spatial-spectral joint features,and the test sample labels are computted through the optimal connecting weights in the output layer of broad learning system for constructing the initial probability maps.Finally,based on spatial context information the guided filter is applied to correct the misclassified sample labels to achieve the purpose of further improving the classification accuracy.3.The labeled samples of hyperspectral image are scarce.It will pay great cost if unlabeled samples are labeled.Therefore,more and more researchers begin to pay an attention to the few-shot classification of hyperspectral image.A new method of spatial-spectral joint features classification for hyperspectral image based on sparsity augmented collaborative representation is proposed in the thesis to make the best use of spatial-spectral features for achieving the few-shot classification of hyperspectral image.Firstly,the average filter is applied to filter the spectral information of hyperspectral image to get the spatial-spectral joint features,because there is close correlation among the spectral features of adjacent pixels.Secondly,the augmented representation are achieved by adding the sparse representation to the collaborative representation for extracting enriched discriminative spatial-spectral joint features,and the testing sample labels are computed by utilizing the augmented representation and the train samples for constructing the initial probability maps.Finally,based on spatial context information the guided filter is utilized to correct misclassified samples to improve the classification accuracy.The results of experiments on Indian Pines,Salinas and Pavia University datasets reveal that the classification methods proposed in this thesis can effectively combine spatial characteristics with spectral features of hyperspectral image to obtain better classification performances.There are 34 pictures,20 tables,and 200 references in the thesis.