Reseach On Hyperspectral Target Detection And Image Classification Based On Attention Mechanism

Hyperspectral image(HSI)is finely divided in terms of spectral dimensions,which is different from the color image containing three spectral bands of Red,Green and Blue.It consists of dozens or even hundreds of spectral bands.The rich spectral information makes HSI have a strong ability to distinguish different matters,and has a wide range of applications in urban planning,environmental monitoring and crop yield estimation,which involves hyperspectral target detection and image classification technology.However,as for hyperspectral target detection,the traditional methods often have limited abilities to nonlinearly express the spectral information and are difficult to effectively distinguish the different types of hyperspectral targets.In the HSI classification task,the problems that insufficient utilization of spatial and spectral features and insufficient number of sample labels result in a great impact on the improvement of HSI classification accuracy.This thesis proposes the corresponding improved methods to mitigate the problems of hyperspectral target detection and image classification tasks.The main works of this thesis can be summarized as follows:(i)For the problem that the traditional hyperspectral target detection method has limited nonlinear spectral expression ability and is difficult to distinguish different types of hyperspectral targets,a hyperspectral target detection method integrating attention mechanism and multi-objective constrained energy minimization(E-IMTCEM)is proposed.Specifically,the E-IMTCEM method combines multiple improved detectors of constraint energy minimization(CEM)with ensemble learning;an internal parallel CEM detector is constructed,and the spectral weight is redistributed by introducing the attention mechanism.The multi-layer detector is connected by external cascade to enhance the feature expression ability concerning the target to be detected.The effectiveness of the proposed E-IMTCEM method is verified on synthetic and real hyperspectral datasets.The results show that EIMTCEM obtains high accuracy for multi-class hyperspectral target detection.(ii)Aiming at the problem of insufficient utilization of spatial and spectral information and insufficient number of sample labels in hyperspectral remote sensing image classification,a HSI classification method combining multi-scale three-dimensional convolutional neural network(3D-CNN)and Convolutional Block Attention Mechanism(CBAM)is proposed.Firstly,the feature mapping method is utilized to fully mine and fuse the spatial and spectral features of hyperspectral images from different receptive fields,and the fused spatial-spectral features are further refined by CBAM.Then the deep network is constructed by the residual idea,and the Dropout method is adopted to deal with the overfitting problem.Finally,the Softmax classifier is used for classification.A large number of experiments on three commonly used hyperspectral datasets show that the proposed method achieves better classification performance than other classical methods.(iii)To make full use of the spatial-spectral features of hyperspectral images,an end-toend dual-channel HSI classification method is designed by combing multi-scale 3D-CNN and an improved residual attention mechanism.Firstly,the spatial features and spectral features of hyperspectral images are extracted by two branches.Then,the residual idea is adopted to improve the spatial and spectral attention modules,which is used to refine the extracted spatial and spectral features respectively.Finally,both refined features of HSI are fused and are classified by Softmax classifier.The classification experiments are carried out on the commonly used hyperspectral datasets,and the results show that the proposed improved method obtains better classification results than other classical methods.