| Hyperspectral image classification is an important task in hyperspectral image processing and applica-tions,which utilizes the combined spatial and spectral information of hyperspectral data to classify each pixel in an image,with the aim of achieving high-precision classification and automated recognition of ground tar-gets.With the continuous development of remote sensing technology,hyperspectral images have higher spec-tral and spatial resolution.The complex spatial structure and high-dimensional spectral information of such data pose challenges for accurate classification using traditional machine learning methods.Deep learning,as a powerful feature extraction tool,can effectively solve nonlinear problems and has become the main-stream algorithm in hyperspectral image classification.However,faced with the limitations of convolutional neural networks in long-range dependency and key information perception,efficient attention mechanisms have become a favorable approach for improving the performance of deep models.This paper proposes two deep learning frameworks based on attention mechanisms,aiming to extract more discriminative spatial and spectral features using lightweight deep models and achieve precise ground object recognition by taking into account the special properties of hyperspectral remote sensing images.1)Facing the three-dimensional data characteristics of hyperspectral images containing spectral informa-tion,a three-dimensional attention mechanism is more advantageous for spectral key information perception and representation.However,three-dimensional multi-head attention mechanisms face serious parameter optimization,computation,and storage burdens.Therefore,this paper proposes a hyperspectral image clas-sification method based on a multi-head context self-attention network model.This method decouples the spectral attention perception into two parallel sub-modules of spatial attention and spectral attention,con-structs a plug-and-play decoupling context self-attention mechanism(DMu CA),and dynamically calibrates the attention of spatial and spectral features.Furthermore,this method further utilizes equidistant sampling to perform grouping representation on each dimension of the data,groups the query set and the context in-teraction keys and places them in groups,and obtains the multi-head attention weights to ensure the model’s perception performance while effectively reducing the model’s parameter quantity and computational burden.A large number of experiments show that the decoupling sub-modules and dynamic and static weighting op-erations in the proposed method demonstrate the expected effects in hyperspectral data important information perception and ground object recognition,especially for key spectral information.2)To deal with the difficulty in learning ground object representation caused by the uneven size of object targets in hyperspectral remote sensing images,this paper proposes a channel interaction attention model with multi-scale spatial neighborhood perception capability(MCIAN),which is used to improve the recognition accuracy and location precision of multi-scale ground objects.Firstly,different-scale filters are used to extract the spatial multi-scale information of hyperspectral data in each channel group.Then,a first-order-second-order joint channel attention mechanism is introduced on different scale groups to measure the importance of each channel from multiple perspectives and improve the reliability of feature calibration.The channel grouping leads to a lack of information interaction between groups in attention calibration.Therefore,a multi-scale feature progressive recombination attention mechanism is introduced after the grouping atten-tion to achieve multi-scale information interaction and realize local and global feature calibration.Finally,the proposed attention modules are embedded into the residual backbone network to achieve accurate clas-sification of hyperspectral images.Parameter analysis,ablation experiments,and comparative analysis with existing deep learning methods on two public experimental datasets are conducted to verify the effectiveness of the proposed method.In summary,for the high-dimensional nonlinear distribution structure of hyperspectral data,this paper constructs deep network models from the perspectives of spatial spectrum decoupling and multi-scale atten-tion perception.At the same time,a large number of experimental analyses have been conducted to verify the feasibility and advantages of the proposed method compared to the existing state-of-the-art methods. |
