Research On Hyperspectral Image Classification Based On Space-spectrum Features

The classification technology of hyperspectral images is one of the important research directions in the field of remote sensing.However,the high-dimensional characteristics of hyperspectral data cause certain difficulties in data processing and analysis.Traditional classification methods do not make full use of spatial and spectral information,and in the selection of training samples,most of them use traditional random sampling strategies,which causes the overlap of training samples and test samples,leading to problems such as inaccurate evaluation of classification algorithms.In response to the above problems,this article has done an in-depth study on the spatial spectrum classification method of hyperspectral images based on previous studies,and paid attention to the sampling method,applied the control sampling strategy method to select training samples,and strived for a more realistic and objective evaluation system.The specific research content of this paper is as follows:(1)Within the scope of the supervised classification,the traditional empty hyperspectral image spatial-spectrum feature extraction algorithms are often adopt the method of random sampling,This method will enhance the overlap between the training samples and the test samples,which in turn increases the classification accuracy,resulting in the deviation of the improper evaluation of the classification algorithm.To objective evaluation of the accuracy of hyperspectral image empty spectral classification algorithm,this paper introduces a new method for control of sampling,and applies it respectively in spectral classification and spatialspectrum classification,the experiments show that the spatial-spectrum classification algorithms which can reduce the overlaps between training samples and testing samples,provide more accurate and objective evaluation.(2)This paper presents a hyperspectral image classification algorithm based on band selection and 3D-Gabor filtering.Compared with other algorithms,the classification accuracy of hyperspectral images can Improve classification accuracy when there are fewer label samples in hyperspectral images.Firstly,the algorithm uses an extensible self-representation learning band selection method,and selects the bands with a large amount of information through cache vector.Then,the 3D-Gabor filtering is applied to respond to the selected band,obtain the feature vectors of all pixels,and use these feature vectors for classification.Finally,experiments on three data sets show that the algorithm can improve the classification accuracy and shorten the experimental time.(3)In this paper,a hyperspectral image classification algorithm based on band restoration and multi-scale spatial spectrum feature fusion is proposed.In the case of sufficient label samples for hyperspectral images,compared with other classification algorithms,it can fully demonstrate the advantages of deep learning,fully extract the space spectrum characteristics of hyperspectral images,and improve the classification accuracy.The algorithm firstly uses a twolayer hierarchical Dirichlet process to model the noise structure of hyperspectral images and remove the noise of different bands of hyperspectral images.Then,the VGG16 convolution module is used to extract deep discriminating spatial features and obtain spectral features on the corresponding scale.Then,Cooperative Sparse Autoencoder is used to integrate deep space and spectral features and perform multi-scale fusion representation.Experiments on three data sets show that this method can improve classification accuracy.