Research On Hyperspectral Image Feature Extraction And Classification Based On Spatial-Spectral Combination

Deep network is widely used in hyperspectral image classification tasks and applications because of its ability of self-learning to obtain deep features that better express the characteristics of hyperspectral data itself.However,in the deep network hyperspectral classification method,there are still problems such as loss of image fine features,information redundancy,and weak discriminability of spatial-spectral fusion features.In response to the above problems,in order to better extract more discriminative spatial and spectral features,and make fully use of spatial-spectral fusion features to classify hyperspectral images,this paper has carried out research work from two aspects of network model structure and feature fusion.The main work of this paper is as following:1.Aiming at the high dimension and strong correlation between spectral bands of hyperspectral image,which leads to over fitting in training and the classification accuracy is reduced.A hyperspectral image classification method based on cross grouping spectral-spatial feature enhancement network is proposed.Firstly,the spatial-spectral cross grouping operation is exploited to group and extract the spectral and spatial features.Secondly,the self-attention module is used to enhance the extracted spectral and spatial features.Finally,the spectral and spatial features are fused to extract the spatial-spectral fusion features for classification.With the experimental results,the network can fully obtain the global and local spectral information,reduce the redundancy of spectral information,and effectively extract the spatial spectral features for classification.2.To acquire fine spectral and spatial features and their interaction information in hyperspectral image classification,a hierarchical spatial-spectral fusion network is proposed.Firstly,the hierarchical feature extraction module was exploited to extract the spectral and spatial features of hyperspectral images respectively.Secondly,the spatial-spectral feature interactive fusion module is designed and employed to fuse the features and produce the joint spatial-spectral features.The proposed network can not only extract and integrate the fine spatial and spectral features of different levels,but also capture the interaction between spectral and spatial features by joint learning.By the experimental results,it is shown that the proposed network performed better than the state-of-the-art deep network-based classification methods.The network is shown the capability of extracting fine features and capturing the spatial-spectral joint features for classification.3.To reduce the loss of high-resolution feature information in the process of feature learning,a spatial-spectral group-convolution dense network is proposed.Firstly,the spatial-spectral 3D group-convolution densenet module is exploited to extract the spectral and spatial features step by step.Secondly,the spectral residual attention module is designed to enhance the regions with abundant spectral information.This method can reduce the inherent information redundancy of hyperspectral data,reuse high-resolution features,and avoid the loss of detailed feature information By the experimental results,it is shown that the proposed network performed better than the state-of-the-art deep network-based classification methods.