Research On The Hyperspectral Image Classification Algorithm Based On Kernel Method And Dictionary Learning

Hyperspectral image classification is an important research content in its processing and analysis process,and has been widely used in many fields,such as environmental protection,land monitoring,military reconnaissance and others.Therefore,it has received much attention from researchers.Due the influence of unfavorable factors,such as the insufficient labeled samples,"the different objects with same spectrum and the same objects with different spectrum",etc.,the traditional classification model using only spectral feature has low accuracy.The study found that the spatial information characterizing the relationship between pixel and their neighbors can assist in the classification task.Therefore,the research of constructing a classification model using both spatial and spectral information has made gratifying progress.However,the simple artificial features used in this research have limited expressive capabilities and are susceptible to noise pixels in hyperspectral image.The model based on sparse representation and dictionary learning can effectively extract the robust feature representation of pixel,so it has been widely concerned in the field of hyperspectral image classification.However,it is necessary to further study that how to effectively use the multiple features of pixel in hyperspectral image.Therefore,this thesis focuses on the research of sparse representation and dictionary learning based on multi-feature,so as to improve the classification accuracy of hyperspectral image.The main research work of the thesis is as follows:1.A multi-feature kernel hyperspectral image classification algorithm based on adaptive sparse representation(MFK-ASR)is proposed.The algorithm extracts multiple features of hyperspectral image and divides the image into several spatial groups by using the watershed segmentation algorithm.By applying adaptive sparse itemladoptive,0 to multiple features of all pixels in each spatial group and embedding the coefficient reconstruction errors of multi-feature,each feature can adaptively select appropriate dictionary atoms,thereby effectively utilizing spatial neighbor information and correlated yet complementary information of multiple features to improve the discriminability of coding coefficients.Furthermore,a multi-feature nonlinear adaptive sparse representation model is constructed by using a kernel method.The experimental results on Indian Pines image with large spatial structure and spectral confusion classes and University of Pavia image containing small and complex spatial structures show that the proposed algorithm can effectively improve the classification accuracy of hyperspectral image,and has high classification accuracy for classes with a small number of samples.2.A multi-feature kernel hyperspectral image classification algorithm based on spatial-aware dictionary learning(MFK-SADL)is proposed.The algorithm uses the same method as MFK-ASR model to extract multiple features and divide the spatial groups of hyperspectral image.By applying a joint sparse constraint term lrow,0 to multiple features of all pixels in each spatial group and embedding coefficient reconstruction errors of multi-feature,the spatial neighborhood information and multi-feature information can be used for dictionary learning.Moreover,a multi-feature nonlinear dictionary learning model is proposed by using the kernel method.The experimental results show that the proposed method has better classification accuracy for images with different scenes(Indian Pines and University of Pavia),and especially has significant classification accuracy for hyperspectral image containing small and complex spatial structures.3.A multi-feature kernel hyperspectral image classification algorithm based on class sub-dictionary learning(MFK-CSDL)is proposed.The algorithm uses the same feature extraction method and image segmentation method as MFK-SADL algorithm for multi-feature extraction and spatial groups division of hyperspectral image.By applying a joint sparse constraint lrow,0 to multiple features of each class of training samples and embedding coefficient reconstruction errors of multi-feature,the correlated yet complementary information of multiple features can be effectively combined.Thus,a discriminative dictionary is obtained for each feature of each class of samples.The joint sparse coding for multiple features of all pixels in each spatial group based on a pre-leaned dictionary improves the distinguishability of the coding coefficients.Furthermore,a multi-feature nuclear sub-dictionary learning model is proposed by using the kernel method to extend linear model.This model allows the training phase to run for a short time by using only each type of training samples to learn the corresponding sub-dictionary.The experimental results on the public data set of Indian Pines and University of Pavia verify the effectiveness of the proposed algorithm.