Research On Hyperspectral Image Classification Methods Based On Broad Network

As one of the basic and key technologies in the field of remote sensing,hyperspectral image(HSI)classification is an important means for human beings to recognize their environment.However,the existence of problems such as difficulty in obtaining labeled samples and difficulty in fully characterizing complex spectral-spatial features have severely restricted the further application and development of HSI classification.In response to the above problems,broad learning,deep learning,and domain adaptation technology are used to study the HSI classification in this thesis.The main work includes:1.To solve the problem that it is difficult to classify HSIs at a low labeling cost,a semisupervised HSI classification method using data augmentation broad network(DABN)is proposed.First,the graph convolution operation is applied to extract deep and nonlinear spectral-spatial features from the original HSI as input.Second,the combinatorial average method(CAM)is proposed to use valuable paired samples to generate sample expansion set for DABN model training.Third,the broad learning system(BLS)is used to perform broad expansion on spectral-spatial features extracted by graph convolution network and extended by CAM,which further enhances the feature representation ability.Finally,the output weights can be calculated by the ridge regression theory.2.To solve the problem that using linear sparse autoencoder to map features in the original BLS will lead to the decline of feature discrimination,a HSI classification method using broad graph convolutional network(BGCN)is proposed.In the proposed method,the graph convolution operation is first used to capture the nonlinear spectralspatial features,instead of only using the linear sparse autoencoder in BLS.Then,the discriminative spectral-spatial features are expanded with the graph convolution operation,which helps to further enhance the feature representation ability,thereby improving the classification ability of the BGCN.Finally,the output weights are calculated with the ridge regression theory.3.In order to solve the problem that the original BLS cannot complete the crossdomain HSI classification,a HSI classification method using domain adaptation broad learning(DABL)is proposed.First,according to the importance of the marginal and conditional distributions,the maximum mean discrepancy(MMD)is used in mapped features to adapt these distributions between source and target domains.Meanwhile the manifold regularization is added to maintain the manifold structure of the input HSI data.Second,to further reduce the distribution difference and maintain manifold structure,the domain adaptation and manifold regularization are added to the construction process of the output layer objective function.Finally,the ridge regression theory is exploited to get the output weights.4.To solve the the problem that DABL is difficult to fully characterize the complex target-domain spectral information and cannot reduce the class weight deviation during the adaptation process,a HSI classification method using convolutional broad domain adaptation network(CBDAN)is proposed.Firstly,the convolutional domain adaptation network(CDAN)is proposed.A domain adaptation layer is added into the convolutional neural network to adapt the marginal distribution and second-order statistics information of both domains.Secondly,the weighted conditional maximum mean discrepancy(WCMMD)is proposed to calculate the conditional distribution divergences and changes of class prior distributions.Then the WCMMD based domain adaptation term is added during input mapping to the mapped features in the BLS to obtain the weighted conditional broad network(WCBN).On the one hand,WCBN can reduce the difference of conditional distribution and class weight bias of the domaininvariant features extracted by CDAN.On the other hand,the representation ability of the domain-invariant features expanded by CDAN can be further enhanced.5.To solve the the problem that the distribution of cross-domain hyperspectral data is not easy to adapt sufficiently,with the help of transfer learning and adversarial learning ideas,by transferring the labeled sample information of the source domain to the unlabeled target domain,a HSI classification method using domain adversarial broad adaptation network(DABAN)is proposed.Firstly,the bottleneck adaptation module is added into the domain adversarial neural network,by simultaneously performing domain adversarial learning,reducing both the marginal distribution difference and second-order statistic difference between two domains,the distributions of the source and target domains are aligned,and thus the domain adversarial adaptation network(DAAN)is designed.Then the conditional distribution adaptation regularization term based on MMD is embedded into a BLS to obtain the conditional adaptation broad network(CABN).On the one hand,CABN can perform the classlevel distribution adaptation on the domain-invariant features extracted by DAAN.On the other hand,the representation ability of the domain-invariant features expanded by CABN can be further enhanced.Experiments on the Indian Pines,Pavia University,Kennedy Space Center,Botswana,Pavia City,and Houston Bright datasets are conducted.The experimental results show that the proposed methods can acquire high accuracy on single-domain or cross-domain HSI classification tasks.There are 36 figures,25 tables,and 190 references in this dissertation.