Attention-based Multi-scale Hyperspectral Image Classification Network

In recent years,hyperspectral imaging technology has developed rapidly and is widely used in many fields such as agriculture and military affairs.Hyperspectral images provide rich information for the recognition and classification of ground objects with their high resolution,large amount of data and correlation between adjacent bands.However,there are few labeled samples,high cost of labeled samples,high spectral dimensions,“same objects with different spectral” and “different objects with same spectrum” have also brought great challenges to hyperspectral image classification.Hyperspectral image feature extraction and classification is the basis of hyperspectral data processing,and has always been the focus of image processing technology researches.The existing classification methods based on two-dimensional convolutional neural network extract spatial features and spectral features at the same time,and the interaction between the two features leads to unsatisfactory classification results.How to make full use of spectral and spatial features to improve classification accuracy is one of the difficulties in hyperspectral image classification.On the basis of summarizing the existing feature extraction and classification methods,this thesis makes an in-depth analysis and research on the characteristics and existing problems of hyperspectral data.The main work includes:(1)In order to make full use of the local and global features of spectrum and space,this thesis designs a multiscale fusion strategy.Based on this strategy,we proposed a attention-based multi-scale hyperspectral image classification network(AMSN).(2)Two branches of spectrum and space are designed in the network,and spectral features and spatial features are extracted respectively by three-dimensional convolution.The two branches take the densely connected convolution network as the basic framework,and each layer of the densely connected convolution network leads out an output channel to obtain local features,the final output are the global features.The multi-scale fusion features are obtained by combining the local features and the global features.(3)The channel attention blocks are introduced after the spectral multi-scale fusion features,and the spatial attention blocks are introduced after the spatial multi-scale fusion features.The channel attention blocks and the spatial attention blocks can adaptively focus on the regions with large differences,and greatly improve the ability of feature extraction of the network.(4)Finally,spatial and spectral features are fused to obtain spatial and spectral joint features,and softmax classifier is used to classify hyperspectral images.This thesis uses three data sets to verify the proposed method.The overall accuracy,average accuracy and κ coefficient are used to evaluate the proposed method and other related methods.The overall accuracy of the three data sets is improved by 0.32%,0.39% and 0.32% respectively,indicating the feasibility of the proposed method.In addition,the impact of multi-scale fusion strategy on the experimental results is verified by stopping some branch work,all the methods using multi-scale strategy have higher classification accuracy than the existing methods,which shows the effectiveness of multi-scale strategy.