Research On Spectral-spatial Feature Extraction And Classification Method For Hyperspectral Image Based On Small Number Of Training Samples

The object classification of hyperspectral remote sensing images(HSI)is one of the most important research topics in the field of HSI processing.There are many characteristics of HSI,such as large amount of data,high degree of redundancy,complex spatial-spectral structures,and the phenomenon about the same objects with different spectral curves and the same spectral curves come from different objects etc.Especially,when the number of training samples is very limited,the existing feature extraction methods of HSI have poorer representation abilities.Therefore,how to efficiently extract and classify the spatial-spectral features of HSI under the limited number of training samples is the main research focus of the thesis.Focusing on the dimensionality reduction,denoising and feature extraction methods,there are some methods have been proposed in this thesis,which can solve the problems of unsatisfactory HSI classification results to a certain extent when the number of training samples is quite limited.The main contents of this paper are summarized as follows:Based on the problem of inaccurate segementation of superpixel segmentation algorithm,a new framework based on two k selection rules is proposed.Specifically,by selecting the most representative training samples and the most representative test samples in each superpixel region,the proposed method can enhance the representation ability of the training data set.Moreover,the results show that the HSI classification algorithm based on the fusion of two-stage k-NN and superpixel representation can effectively choose the most representative samples in the superpixel region,obtain purer spatial context information,and achieve more accurate spatial representation of the test samples.Meanwhile,it can reduce the intra-class differences and overcome the inter-class interferences.Compared with state-of-the-art methods,this method can obtain the highest classification accuracy.Based on the problem of image noise of hyperspectral dataset,which is caused by many factors like sensor material properties,component structure performance,working environment,etc,a new framework based on texture pattern separation(TPS)is proposed.Specifically,it integrates the analysis sparse representation(ASR)and the synthesis sparse representation(SSR)into the joint convolutional analysis-synthesis sparse representation(JCAS).In this way,an HSI can be decomposed into two layers(i.e.,the texture layer with feature information and the background layer with noise information).Finally,only the texture layer is used for recognition and classification of real objects.The experimental results in several real HSI data sets show that the algorithm can effectively separate two images,and the extracted texture layer can greatly improve the classification performance of HSI.Based on the problem of salt and pepper noise in hyperspectral image classification,in order to figure out the process of probability construction and probability optimization of HSI deeply,a dual stage construction of probability method(DSCP)is proposed.Firstly,the initial probability map is constructed as accurately as possible;secondly,the initial probability map is optimized after iteration by using the deeper spatial context information;finally,the final classification result map of HSI can be obtained.In order to verify the classification of the proposed method under the condition of few samples,DSCP uses a training sample set with only six samples per class.The experimental results show that DSCP method can reduce the phenomenon of salt and pepper noise,and improve the classification performance of HSI to a certain extent.