Spectral-spatial Classification Methods Of Hyperspectral Image With Deep Residual Structure

Hyperspectral image(HSI)classification is a key problem in remote sensing image processing.Its main task is to fine classify ground objects by learning rich feature information of HSI.However,the problems such as high dimension,spatial heterogeneity of spectral information,limited training samples and information redundancy greatly limit the HSI classification performance.Based on the problems mentioned above,this paper discusses and studies HSI classification methods with deep residual structures,which mainly include HSI classification method based on deep hybrid dilated residual networks,HSI classification method based on multi-scale feature fusion residual network,HSI classification method based on cascaded dual-scale crossover network,and feature-grouped network with spectral-spatial connected attentions for HSI classification method.The details are summarized as follows:1.For the problem on the insufficient feature information of HSI,a novel hybrid dilated residual network method is designed.This method aims to embed the network into the process of continuously extracting spectral-spatial information,which not only can increase the size of the receptive field,but also can solve the grid effect problem caused by dilated convolution.Meanwhile,the residual network is used to alleviate the phenomenon of gradient vanishing and gradient explosion.The use of batch normalization and activation function could speed up the training and improve the training accuracy of the model.Experimental results show that the proposed method has better classification performance.2.For the case of the limited label HSI,a multi-scale features learning method is proposed.The key idea is to improve the performance of the network by increasing the width of the network.Specifically,features are extracted by the cascaded dual-scale cross network and multi-scale feature fusion residual network,and then integrated by adding or multiplying features at different scales to increase the information flow between the networks and enhance the contrast between pixels,which is conducive to feature extraction.Experimental results indicate that the proposed method can alleviate the problem of small samples and significantly improve the classification performance of HSI.3.For the difficulties on selecting the spectral band and spatial position due to the high dimension of HSI,this paper constructs a feature-grouped network with spectral-spatial connected attentions.First,use the attention module to continuously focus on the initial HSI from spectral to spatial,which could select important bands and mark the spatial location of interest.Then,construct a features grouped network to divide the coarsely extracted features into several groups,and extract the features from each group through a continuous residual network.The extracted features are fused to capture the high-level feature information of HSI.Experimental results illustrate that the proposed method can produce better classification results.