Research On A Deep Learning-based Algorithm For Hyperspectral Image Classification

Hyperspectral imaging is an image that combines imaging and spectral techniques to simultaneously obtain high-dimensional spatial and spectral information of objects.Due to the different features of objects in different dimensions,the dense spectral dimension provides good conditions for accurate classification of objects,making hyperspectral imaging widely used in agriculture,environmental and climate monitoring,urban development,military security,and other fields.However,due to the three-dimensional,redundant,same-spectrum-different-object,and low spatial resolution characteristics of hyperspectral data,hyperspectral image classification remains a challenging problem.This thesis proposes three more efficient hyperspectral image classification methods based on spatial-spectral attention feature extraction mechanism,Transformer network structure,and self-attention mechanism.The main contributions of this thesis are as follows:(1)Considering the issue of insufficient extraction of spatial and spectral features of hyperspectral imagery(HSI)in current convolutional neural network(CNN)-based HSI classification methods,we propose a multi-scale residual network-based HSI classification method with attention mechanism.This method processes the original HSI data using patches of different sizes to handle hyperspectral data at different scales.The hyperspectral data from different scales are used to extract the spatial-spectral features of HSI by adding a spatialspectral attention mechanism to the residual network,and further aggregate the feature information of HSI by embedding a lightweight attention module in the CNN.Finally,the features from different paths are fused and output.Experimental results show that this method achieves better classification performance compared to other mainstream methods on multiple public datasets.(2)An end-to-end Transformer model is proposed to address the problem of convolutional operations being inadequate for capturing long-range dependent interaction features.The model utilizes spectral attention mechanism and self-attention mechanism to extract spatial-spectral features of hyperspectral images(HSI).Firstly,the original HSI data is preprocessed and transformed into multiple vectors via a spectral attention module.Linear compression is then performed on each vector to obtain a sequence of vector lengths.Positional encoding vectors and learnable embedding vectors are introduced to address the issue of capturing continuous spectral relationships in HSI over long distances.Subsequently,an encoder module with multihead self-attention is employed to extract image features,and a residual network structure is utilized to address gradient vanishing and overfitting issues.Experimental results demonstrate that the proposed method outperforms other mainstream HSI classification methods.(3)Given the characteristics of hyperspectral data,including threedimensionality,redundancy,and noise,it is challenging to represent hyperspectral data adequately using current methods.To address this issue,a multiscale feature fusion network that integrates 3D selfattention is proposed.The network first employs multiple parallel convolutional kernels for multiscale feature extraction,sampling different granularities of the feature maps,and effectively fusing the spatial and spectral features of the feature maps.Then,an improved 3D self-attention mechanism is proposed to provide local feature details for the self-attention branch and fully utilize the contextual relationship of the input matrix.Subsequently,an improved 3D multi-head self-attention mechanism is used to enhance convolution,and a3Dcov＿attention module is proposed to combine the convolution mapping that extracts local features with the self-attention feature mapping that models global dependencies,enhancing the local receptive field while capturing long-distance interactions.Experimental results on multiple public datasets show that the proposed method outperforms other mainstream HSI classification methods.