Hyperspectral Image Classification Method Based On ERS Preprocessing And Dual-Attention Mechanism Res-CNN

Hyperspectral imaging is a remote sensing technique aimed at capturing and analyzing the reflected or emitted spectra of a scene across a wide range of wavelengths,typically containing hundreds to thousands of spectral bands.Due to the unique spectral features of each material and object,hyperspectral images contain abundant information about the composition,structure,and properties of objects and materials in a scene.The significance of hyperspectral image classification lies in its various application scenarios,such as land cover mapping,mineral exploration,crop monitoring and prediction,environmental assessment,and so on.In recent years,deep convolutional neural networks have been widely used in image processing and have achieved good classification results.In the task of hyperspectral image classification,although convolutional neural networks have mitigated the problem of network degradation by increasing network depth,classification performance still needs improvement due to the inability to extract global features and the inability to effectively suppress noise.In addition,simple end-to-end models cannot model long-range dependencies well and ignore the correlation between bands.Traditional principal component analysis for dimensionality reduction preprocessing ignores the correlation between different homogeneous regions and cannot fully utilize the spatial correlation between bands.Therefore,to address the above problems,this thesis adopts the spectral-spatial dual-dimensional attention mechanism residual network and ERS segmentation homogeneous subregion preprocessing method,respectively.Based on the spectral attention module,the spectral features are weighted,making the network focus more on discriminative features.The spatial attention module weights different regions in the image on the spatial dimension to improve feature resolution and capture local image features better.Based on the dual attention mechanism module,the model’s classification accuracy and generalization ability can be improved,which is effective in feature focus,feature position focus,and noise suppression.The complexification of the residual network structure and the increase of depth can effectively ensure its ability to extract feature information.Embedding the attention mechanism in the residual structure can not only enhance the learning features of the final task and weaken the features with less information by weighting the learned features but also reduce the number of training parameters by changing the network depth.The superpixel segmentation preprocessing module generates dense and compact adjacent homogeneous subregions to refine and distinguish the features corresponding to different pixels,extract the low-dimensional spatial features of the image itself,fuse them with high-dimensional spectral features,and ultimately obtain the hyperspectral image classification results,further improving the classification accuracy.The work of this thesis includes:(1)This thesis proposes a residual convolutional neural network based on the dual attention mechanism for hyperspectral image classification.The residual structure can solve the gradient problem,that is,as the gradient decreases in the network layer during the training process,it becomes difficult to train an effective feature representation.The dual attention mechanism is more fully utilized in feature extraction in different dimensions.This method uses four publicly available hyperspectral image classification datasets,namely Indian Pines,KSC,Pavia U,and Xuzhou datasets,and evaluates the results based on three indicators: Overall Accuracy(OA),Average Accuracy(AA),and Kappa coefficient.The average OA score on the four test sets is0.99508,and the average AA and Kappa coefficient scores are 0.99452 and 0.99453,respectively,which demonstrate the accuracy of the method used in this thesis compared with current hyperspectral image classification methods SVM,3D-CNN,SSRN,Hybrid SN,ENL-FCN,DCCN,SSAtt.(2)This thesis improves the DARCNN architecture by introducing a superpixel segmentation preprocessing algorithm,namely Super DARN,for image classification.The superpixel segmentation feature generates superpixels of similar size,homogeneity,and compactness,adaptively segments the hyperspectral image’s boundary,distinguishes the homogeneous features of adjacent subregions,and fuses low-dimensional and high-dimensional features.This method was tested on two complex land-cover datasets,namely Indian Pines and Pavia U datasets.Through visualization and quantitative analysis of the classification results,this method achieved superior OA,AA,and Kappa coefficient of 0.99635,0.99625 and 0.99750,respectively.(3)In this thesis,different numbers of superpixels are formulated for different datasets aiming to obtain the highest classification accuracy through the effect of different numbers of superpixels on OA.The robustness and feasibility of the proposed model are corroborated by the ablation experimental results in this thesis.In addition,network complexity calculations are performed and the experimental data show that the method achieves a good balance between classification results and complexity calculations.The experimental visualisation results and quantitative data show that it has been further improved for hyperspectral image classification.