Research On Hyperspectral Image Classification Methods Based On Spectral And Spatial Joint Feature Learning

Hyperspectral imaging is one of the import directions of remote sensing application,which has been widely used in environmental monitoring,precision agriculture,and meteorological observation.Hyperspectral images classification is an effective way to distinguish and recognize complex objects,which has important research significance and application value.In this dissertation,the hyperspectral image classification method based on neural networks is systematically studied.The main contents and innovations of this dissertation are as follows:Aiming at the problem that the accuracy of classification methods is limited due to the high similarity of spectral information and the diversity of spatial information among samples of hyperspectral image datasets,a cascade residual capsule network model via combining residual network and capsule network is proposed in this dissertation.The model combines residual network and capsule network to learn highlevel spectral features and spatial context orientation features,which can effectively distinguish highly similar spectral information and diverse spatial information.Experimental results show that the proposed method improves the overall classification accuracy of hyperspectral images by 0.14%-5.88% compared with state-of-the-art hyperspectral image classification methods.Aiming at the problem that the feature extraction from objects of hyperspectral image dataset samples is interfered by other hyperspectral pixels,an attention asymmetric autoencoder model is proposed in this dissertation.The model firstly senses objects in the hyperspectral image through the attention module,and then the attention module is embedded into the asymmetric autoencoder module to learn the highly discrimitive spatial spectral features.Experimental results show that the proposed method can effectively suppress the interference of non-goal to objects in hyperspectral images,and improves the overall classification accuracy of hyperspectral images by0.02%-5.85% compared with state-of-the-art hyperspectral image classification methods.In this dissertation,an attention residual capsule network is proposed by combining the core network structure of the attention asymmetric autoencoder and the cascade residual capsule network.The model perceives the features of objects from the spatial and spectral dimensions of hyperspectral images through the attention module,respectively.And the residual capsule module is connected in a serial way to learn high level spectral features and spatial context orientation features from spectral and spatial dimensions,respectively.It can effectively suppress the interference of non-goal and distinguish highly similar spectral information as well as diverse spatial information.Experimental results show that,the proposed method further improves the overall classification accuracy of hyperspectral images by 0.04%～1.29%.In this dissertation,a hyperspectral image classification method based on spectral space joint feature self-supervised learning is proposed via extending the supervised spectral spatial joint feature learning network to self-supervised hyperspectral image classification method.The method combines residual network and capsule network to learn high-level spectral features and spatial context orientation features in spectral and spatial dimensions.And it uses the L2 norm of the output feature vectors of the network model as the pseudo-labels of samples in the hyperspectral image dataset to realize the self-supervised learning of feature.Experimental results show that the proposed method improves the overall classification accuracy of hyperspectral images by 0.79% to 4.24%compared with state-of-the-art hyperspectral image clustering methods.