Research On Hyperspectral Image Classification Methods Based On Spatial Residual Blocks And Attention Mechanism

Hyperspectral images have numerous spectrums which are featured by high spectral resolution and rich information,and they are suitable for geological surveys,etc.For hyperspectral image classification,previous researchers have proposed a series of methods based on deep learning and scored some achievements.However,hyperspectral images are lacking in spatial variability of spectral features and labeled samples.Meanwhile,deep classification networks are troubled by degraded accuracy and massive calculation.Therefore,how to effectively classify hyperspectral images with fewer labeled samples and less calculation is still one of the main research focuses at present.The main work of this paper is as follows.（1）A hyperspectral image classification network based on spatial residual blocks and parallel networks（SRPNet）is proposed to address the problems of spatial variability of spectral features and degraded accuracy in deep classification networks.Firstly,the proposed network extracts spatial features from the rich spatial context information of hyperspectral images by using spatial residual blocks.For one thing,spatial features can be used as auxiliary information of spectral features to help deal with the spatial variability of spectral features.For another,jump connections in spatial residual blocks are beneficial to the back propagation of gradients,which can alleviate the problem of accuracy degradation in the deep network.Second,a parallel network is designed in the network for extracting spectral features,and the parameters are shared among each parallel branch in the parallel network,which reduces the network parameters that need to be trained and further reduces the demand for network training samples.In addition,the network carries out feature fusion for features of different layers in the spectral feature learning part in order to obtain more features.（2）A hyperspectral image classification method based on attention mechanism and recurrent neural network（AMRNet）is proposed to address the problems of computing redundancy and low classification accuracy of edge pixels,which are common in commonly used hyperspectral image classification methods.To solve the computing redundancy problem,the network uses the original hyperspectral image as input,which greatly reduces the computational complexity.The proposed network consists of four main parts: local feature learning,global feature learning,local and global feature fusion,and classification.First,the local feature learning uses the spectral attention module and the bottleneck attention module to introduce attention mechanisms to spatial and spectral information so as to extract them in a targeted manner and further improve the network’s ability to classify edge pixels.Second,the Re Net module is added to the proposed network to obtain global spatial information.Finally,feature fusion operations are performed on local and global features to obtain multi-scale fused features.This paper has conducted experiments on three representative data sets.Compared with some other advanced methods,the two algorithms proposed can improve the accuracy of hyperspectral image classification,and the comparative experiments show the applicability of the two models proposed.The performance is stronger,which verifies the effectiveness of the proposed algorithm.