| Hyperspectral images are abundant with ground object information and are widely used for fine-grained classification.With the advancement of spectral imaging technology,the spatial and spectral information captured by sensors has become richer and more accurate.Deep learning methods have made great progress due to their powerful feature representation capabilities and are widely used in remote sensing,but there are still problems such as insufficient utilization of spatial context information and neglecting the sequence characteristics of hyperspectral data.Therefore,focusing on the problems of how to fully exploit the joint spectral-spatial features with large distinction in hyperspectral images and fully utilize the sequence characteristics of hyperspectral data,this thesis mainly investigates the following two elements.(1)A hyperspectral image classification network with Multi-Scale feature Fusion and Local-Global Representation(MSF-LGR)is proposed to extract richer and more representative spatial and spectral features from multiple scales in spatial and spectral dimensions.First,data dimensionality reduction is performed by Principal Components Analysis(PCA).Then,to effectively extract the joint spectral-spatial features of hyperspectral data and reduce data redundancy,a layer of 3D convolution is used to perform primary joint spectral-spatial feature extraction on the downscaled hyperspectral data.Second,a Multi-Scale Feature Fusion(MSFF)module with channel attention and a Local-Global Representation(LGR)module are designed.MSFF module is able to enhance the feature representation capability at different scales by introducing an attention mechanism in multi-scale branching,LGR module is able to enhance the feature representation capability at different scales.The MSFF module is able to improve the feature representation of features at different scales by introducing an attention mechanism in multi-scale branching,and the LGR module is able to model the input local and global information with fewer parameters.Finally,information fusion of local and global information is performed by a layer of convolutional layers,and classification is performed using a fully connected layer.Experiments on three datasets show that the MSF-LGR method based on multi-scale feature fusion and local-global representation proposed in this thesis can achieve better classification performance.(2)An end-to-end model of Shallow-to-Deep spectral-spatial Feature Enhancement(SDFE)based on Convolutional Neural Network(CNN)and Transformer is designed,which can effectively solve the problem of insufficient discriminative spatial and spectral features extracted in the process of hyperspectral image classification.In this thesis,firstly,the band information with strong spectral characterization ability is extracted by PCA to reduce information redundancy;secondly,a 3D-CNN-based Shallow Spectral-Spectral Feature Enhancement(SSSFE)module is constructed to preserve spatial and spectral correlation while extracting features;then,to avoid the loss of spectral information while enhancing the nonlinear representation capability of the network,a 2D-CNN and attention mechanism-based channel attention residual module(Res-SEConv)is designed to capture deeper spectral-spatial complementary information;Finally,a Vision Transformer-based module is used to extract the spectral-spatial joint features with stronger robustness.The experimental results show that the proposed SDFE method can achieve better classification accuracy,and the classification results are higher than those of MSF-LGR method. |
