| Hyperspectral Image(HSI)is mainly composed of spatial,spectral and radiation information,and can describe the objective world in multiple scales,angles and dimensions.Therefore,HSI classification tasks are widely used in agricultural monitoring,disaster early warning,and medical detection.Significance and broad application prospects.Nowadays,HSI algorithms can be mainly classified into several types: 1)In early work,classification methods based on 2D convolution operations were proposed for hyperspectral image classification,such as 2D-CNN,which mainly uses HSIs are regarded as common twodimensional optical images to extract spatial information;2)According to the characteristic that HSIs contain more spectral information,a classification method based on 3D convolution operation is proposed for HSI classification,such as 3D-CNN,which mainly uses the spectral information in HSIs;3)Compared with the first two methods that use a single convolution operation to achieve HSI classification,some scholars try to fuse different convolutions together.For 3D convolutional operation to achieve HSI classification;4)HSI is generally in the atmosphere,it will be affected by non-human factors,such as water vapor,fog,light,etc.,so a series of related algorithms to remove these natural factors have emerged one after another,that high-quality hyperspectral images are obtained through denoising and dehazing for improving HSI classification accuracy.In the case of sufficient training samples,the above work has achieved certain results in the field of hyperspectral image classification.However,in actual situations,sometimes it is impossible to obtain enough data.With the reduction of training sample data,traditional methods cannot extract enough effective feature map information.The resulting prediction map will have more misclassified regions.In view of the defects and deficiencies of the above work,the following work is proposed to ensure the accuracy of the network model in the case of fewer training samples: 1)Based on the residual structure,the method of 3D and2 D convolution cross-mixing is proposed to obtain the spectral information and spatial information in the HSI,and complete the HSI classification;2)in the same Based on the residual structure,the performance of the multi-attention module on HSI classification tasks is explored,and the multi-attention network model MA-Net is constructed to realize HSI classification.(1)In the first work,principal component analysis(PCA)was used in the data processing stage to reduce the redundant spectrum to reduce the calculation,and a threedimensional(3D)residual structure is designed to extract spectral features,a twodimensional(2D)residual structure is designed to extract spatial information in the feature map,and an auxiliary feature extraction(AFE)structure is designed to connect the first two structures.By fusing 3D and 2D residual structures with AFE,an end-to-end convolutional neural network named CMR-CNN is proposed for HSI classification.(2)The second work proposes to explore the influence of multi-attention and HSI classification based on residual structure: SE-Block is used to capture the local discriminant region of the feature map;GCP module is used to ensure the accuracy of the model and speed up the convergence of the model.Finally,the feature maps obtained by the two attention modules are fused together by means of addition and sent to the Vi T classifier for classification.Vi T obtains the position information of adjacent pixels in the feature map by means of encoding and decoding,and finally gets the classification result.(3)The two proposed experiments of HSI classification were conducted in five different public HSI datasets: Indian Pines,University of Pavia,Salinas Scene,KSC and Xuzhou.The results show that compared with some of the most advanced methods,the proposed two HSI classification network models achieve better classification results in three evaluation indexes:overall accuracy(OA),average accuracy(AA)and Kappa value. |
PDF Full Text Request