| Hyperspectral images(HSIs)record the responses of objects using hundreds of continuous narrow spectral bands ranging from visible light to near infrared.HSIs carry plentiful spectral information,which have been widely applied in various classification tasks.Hyperspectral classification is the task of allocating a label to each pixel.HSIs contain not only spectral information,but also spatial information such as texture and morphology.Thus,extracting features from the spectral-spatial domain can perform better than only utilizing spectral information.Based on the above content,this thesis studies the hyperspectral classification algorithm using the spatial-spectral joint model.The main research contents are as follows:1.In the practical situations,only a few training data can be acquired.Besides,the phenomena of inter-and intra-class spectral variability pose a challenge to hyperspectral classification tasks.To solve the problem,a novel two-branch spatial-spectral network based on the pixel-pair and spatial patch(PPSP)model is proposed in this paper.The PPSP model is designed to construct the PPSP combinations,significantly expanding the amount of training samples.Then,the features of PPSP combinations are extracted via the two-branch spectral-spatial network from the spectral-spatial domain to overcome the problem of inter-and intra-class spectral variability.During the test procedure,the predicted label of the test pixel is determined via the joint classification with the voting strategy.Experiments have been conducted on three hyperspectral datasets,showing that our method has superior classification performance than its counterparts.2.In order to fully extract the spatial-spectral features and prevent the interference of useless features on classification,a novel attention mechanism-aided spatial-spectral residual network is proposed in this thesis.The spatial-spectral features are extracted from the spatial domain and spectral domain by the 3-D convolutional neural network.Besides,the residual blocks are utilized to prevent the network from vanishing gradient.Then,the channel attention module and spatial attention module are used to reduce the interference of useless features on classification.The whole network consists of three parts.The first part is the attention mechanism-aided 3-D residual network,which is used to extract the spatial-spectral features effectively.The second part is the attention mechanism-aided 2-D residual network,which is utilized to finely extract the spatial features.Finally,the third part is the fully-connected module,which is employed to complete the classification task.Experimental results over two hyperspectral datasets indicate that the proposed method achieves superior classification performance over its counterparts. |
