Hyperspectral Image Classification Based On Dense Feature Adaptive Fusion Network And Attention Mechanism

With the continuous progress of remote sensing image acquisition technology,hyperspectral images are getting more and more attention.Hyperspectral images have high research and application values and are widely used in many fields in real life,including agricultural informatization,urban and rural construction planning,geological and mineral survey,military defense,etc.Due to the complex spatial structure,difficult sample labeling and redundant spectral information of hyperspectral images,the difficulty of classification is enhanced.Most of the traditional machine learning-based hyperspectral image classification algorithms use manually designed features,which are difficult to extract deep-level feature information.The emergence of deep networks enables models to extract more abstract and deep-level feature information,which are applied to hyperspectral image classification tasks and achieve excellent classification results.In this paper,two deep network models are designed separately based on existing work:(1)Hyperspectral image classification algorithms based on dense feature adaptive fusion networks.Most hyperspectral image classification algorithms based on deep convolutional network architectures use only the last layer of convolutional features as feature outputs.Although the deep convolutional features are rich in semantic information,their spatial information is scarce,while the shallow convolutional features rich in spatial information are easily ignored.Therefore,this paper proposes an algorithm based on a multilayer dense feature adaptive fusion network,which also combines the low-layer,mid-layer,and high-level features of dense convolutional networks to compensate for the defects of deep-layer features.However,different layers of convolutional features carry different information and contribute differently to the output,and how to define the importance of different convolutional layers is still of high research value.Therefore,this paper proposes a weighted adaptive fusion model to automatically assign the respective weights to different convolutional layers to highlight the more discriminative features and suppress the useless ones by weighted fusion to improve the effect of hyperspectral image classification.(2)Hyperspectral image classification algorithm based on residual dual attention network.Hyperspectral images have the problems of spectral redundancy and difficult sample labeling,and how to extract more accurate and discriminative features as much as possible in the absence of training samples is the key to the hyperspectral image classification task.The visual attention mechanism is proposed to focus on the more important local regions in the image while suppressing the noise regions,so as to deeply optimize the feature map.Inspired by this mechanism,this thesis designs two different types of attention modules in the channel dimension and spatial dimension of the residual network to extract more accurate and discriminative features,and finally fuses the results obtained from different attention modules as the final features to enhance the robustness of the features.