Hyperspectral Imagery Analysis Based On Broad Learning

Hyperspectral imagery analysis is a key step in the analysis and application of remote sensing technology to the Earth’s surface,and it is also one of the important means for humans to recognize the Earth.In order to meet the requirements of classification accuracy,efficiency and generalization,aiming at the characteristics of hyperspectral imagery with complex spatial-spectral characteristics,nonlinear separability and limited label samples,several hyperspectral imagery analysis methods are studied based on deep learning and broad learning techniques.The main work includes:1.Due to the limited ability of conventional graph to represent high-order manifold relationship,it is difficult to adequately characterize the complex spectral information of hyperspectral imagery.Besides,in the case of limited label samples,the capability of deep learning to represent the spatial information of hyperspectral imagery is limited.Regarding the issues above,a supervised hypergraph and samples expanded convolutional neural network based hyperspectral imagery feature extraction method is proposed.Specifically,we first construct the supervised within/between-class hypergraph to extract the spectral features of HSI.Then,a samples expanded convolutional neural network is designed to extract the spatial features of HSI.Finally,the spectral-spatial features used for hyperspectral imagery classification can be obtained by intergrating the spectral and spatial features.2.In the case of no labeled samples,a hyperspectral image clustering algorithm based on broad learning is proposed.Firstly,the graph regularized sparse autoencoder is used to fine-tune the weights of different parts in the broad learning,which will not only maintain the manifold structure of the input,but also enhance the stability of the broad learning.Then,by combining the l₂ norm of the output layer weights term and graph regularization term,the objective function of unsupervised broad learning is constructed.Finally,the weights of the output layer can be obtained by solving the generalized eigenvalue decomposition problem,and the clustering result can be got by performing spectral clustering on the output vector.3.Aiming at the problem that the conventional broad learning cannot utilize the large amount of unlabeled samples,a semi-supervised broad learning based hyperspectral imagery classification method is proposed.Firstly,to make full use of abundant spectral and spatial information of hyperspectral imagery,hierarchical guidance filtering is performed on the original hyperspectral imagery to get its spectral-spatial representation.Then,the class-probability structure is incorporated into the broad learning to calculate the relationship matrix between the labeled and unlabeled samples,and the pseudo-labels of the unlabeled samples are obtained.Finally,the connecting weights of broad structure can be easily computed through the ridge regression theory.4.Due to the use of linear features in broad learning,the complicated spatial-spectral information of hyperspectral imagery cannot be fully expressed.To address this issue,two kinds of convolutional neural networks are proposed,and the single-stage or multi-stage features extracted with convolutional neural network are expanded in wide sense with broad learning.Firstly,the convolutional neural network is pre-trained with limited labeled samples,and the features of multiple stages are extracted.Then,the features of single-stage or multi-stage are expanded with BLS.Finally,the features extracted by convolutional neural network and BLS-based expanded features are all connected to the output layer,and the connecting weights can be easily obtained through the ridge regression theory.5.Aiming at the problem that it is difficult for the conventional broad learning to use the knowledge of relevant domains to improve the accuracy of the target domain classification task,a broad domain adaptive network based hyperspectral imagery classification method is proposed.Due to the fact that the adversarial learning can learn the features with smaller discrepancy in domain distribution through the inter-domain adversarial mechanism,the adversarial learning is utilized to design a broad domain adaptation network to realize cross-domain hyperspectral imagey classification.Furthermore,in order to align the multimodel structure of the source domain and the target domain to improve the accuracy of the domain discrimination,the classifier prediction vectors containing the discriminant information is used as a condition for the domain discrimination of the domain discriminator,and the conditional broad domain adaptive network is proposed.In addition,for the problem that the conventional broad learning algorithm can not update the weights of middle layer in the broad domain adaptive network,a gradient descent increment learning algorithm for the broad domain adaptive network is designed,so that the domain invariant features can be learned.The research results can not only provide new ideas and technical reserves for a variety of hyperspectral image analysis problems,but also further deepen and enrich the existing theoretical theories of machine learning,remote sensing science,etc.,which is of great theoretical significance and practical value.