Feature Extraction And Classification For Hyperspectral Image Based On Capsule Network

Hyperspectral images（HSIs）contain abundant spectral information in hundreds of narrow and contiguous bands,and present rich contextual structure information of imaged scenes.To date,HSIs have been widely applied in many fields,such as agriculture,mineralogy,and en-vironment,whose main processing technologies include feature extraction,unmixing,fusion,target detection,classification,and so on.Among them,feature extraction and classification have gradually become research fronts and research topics in HSI processing.Feature extraction can solve the problem of information redundancy and dimension disas-ter on the premise of ensuring that the useful information in the data is not lost,thus retaining discriminative information for subsequent applications.Classification can obtain the category information of land cover.However,most of the existing HSI feature extraction and classi-fication methods are developed on the basis of traditional machine learning feature extraction methods,most of which are outdated,with limited performance,and lack of design and innova-tion for the characteristics of HSIs.To this end,by combining the latest achievements of deep learning,this thesis has carried out research work on the feature extraction and classification algorithms of HSI based on the capsule network.In addition,the effectiveness of the proposed methods is analyzed and verified from the perspective of theory and computer simulation ex-periments.The main research work is as follows:A robust capsule network for feature extraction and classification of HSIs is proposed.Due to hyperspectral image data is usually interfered by noise,and a large amount of spectral information will have information redundancy,it is very unfavorable for feature extraction and classification applications.Therefore,this thesis introduces the construction of a deep network based on the capsule network for robust feature extraction,and flexibly uses the spectral infor-mation and spatial information of the ground object.First,a capsule network-based framework is proposed for HSI classification,which can model the discriminative features from the per-spective of spectral information and spatial information,respectively.Furthermore,a robust three-dimensional capsule network model is constructed by introducing the maximum corren-tropy criterion,which effectively solves the effects of noise and outliers on HSI classification.Finally,a new multi-source data fusion framework based on maximum correlation entropy is designed,which can overcome the influence of noise and outliers,and effectively fuse spatial-spectral information of HSIs data and elevation information of Li DAR data,thus obtaining more discriminative features for HSI classification.A capsule generative adversarial network method for is developed for HSI classification.It is common knowledge that the labeled samples of HSIs are limited and expensive to ac-quire.However,a large amount of labeled data are required to train and stabilize the most classification methods based on deep learning.The latest research shows that generative ad-versarial networks have been widely used to solve the classification problem under limited samples of HSIs which,however,also suffer from mode collapse and instability problems.As such,a capsule-based generative adversarial network（GAN）for HSI classification is proposed,generating artificial samples for data augmentation to improve the HSI classification perfor-mance with few training samples.In the proposed network,a new discriminator is designed by exploiting Caps Net and convolutional long short-term memory（Conv LSTM）,which extracts low-level features and combines them together with local space sequence information to form high-level contextual features.In addition,a structured sparse_2,1constraint is imposed on sample generation to control the modes of generating data and achieving more stable training.A multi-level capsule network feature extraction and classification method is proposed for HSIs.Deep learning models have recently proven to be effective in addressing the HSI classification problem.However,due to the nature of convolution operations,using all selected pixel in local window may affect the classification performance of the deep learning methods when the local window may have obvious appearance changes.In addition,deep learning method often lack robustness to the perturbation of input,which reduces the generalization ability of these models and undermines their true practicality.To address those problems,a multi-level capsule network is developed for HSI classification.First,spatial-spectral Caps Net structure is applied to explore both spatial context and spectral information simultaneously.Then,a multi-branch LSTM network is designed to fully explore the global properties,which can reduce the impact of other non-supporting pixels.Next,we put forward a global attention fusing module,which fusing robust contextual features with spatial-spectral features to make full use of the complementary advantages between the global and local information.Finally,a spectral norm regularization is imposed to optimize the model improving robustness of the trained model to the minor perturbation.Experiments over three widely HSI datasets have demonstrated the superiority of the proposed method.A capsule domain adaptation adversarial network method for classification of HSIs is de-veloped.HSI domain adaptation is currently one of the effective ways to solve the problem of complicated HSIs scenes and limited samples.However,different HSIs usually have the spec-tral shift between the source domain and target domain due to differences in acquisition and atmospheric conditions or changes in physical properties,which brings great challenges to the existing domain adaptive methods.In order to solve the problem of same body with different spectrum in the domain adaptation of HSIs,a domain adaptation model based on the capsule adversarial network is proposed.First,the capsule network and generative adversarial game criterion are combined to align the feature space of the real domain and the generated domain,and the constraint of wasserstein is introduced to reduce the distribution difference between different data domains,which avoids the problem of pattern collapse.Then,a new capsule generation network model is proposed to learn the global spectral feature representation,which can not only effectively reduce the influence of spectral shift,but also extract the domain invari-ant information and distinguishing information of the data.Finally,the classifier is optimized through generated samples and source domain samples to improve its discriminative ability.In conclusion,this thesis mainly researches on the feature extraction and classification of HSIs.On the basis of capsule network,maximum correntropy criterion,generative adversarial network,multi-level feature fusion,and domain adaptive learning,this thesis constructs plenty of frameworks of feature extraction and classification for HSIs gradually,and demonstrates the effectiveness of the proposed models by conducting the experiments on the some hyperspectral data.