Research On Classification Methods For Open Set Hyperspectral Remote Sensing Images

Hyperspectral remote sensing images have rich spectral information and are widely used in environmental,military,aerospace systems,and other fields.Moreover,hyperspectral remote sensing image classification is an important step in hyperspectral image analysis,understanding,and application.However,existing research based on closed set classification scenarios neglects the existence of unknown classes in real situations,resulting in a significant reduction in the actual classification accuracy of the model in real-world applications.In response to the shortcomings of current open set hyperspectral image classification,deep learning methods are used to study the open set classification problem of hyperspectral images from three aspects.Existing open set hyperspectral image classification methods mainly focus on first-order statistical features,with little attention paid to second-order and higher-order statistical features.In response to the shortcomings of existing open set hyperspectral image classification algorithms,an open set classification model combining second-order pooling attention neural network and K-Sigma is designed.By using a second-order pooling module with spatial attention,feature extraction is performed on the input image,and an open set hyperspectral image classification method is designed to distinguish between known and unknown classes to improve the accuracy of hyperspectral remote sensing image classification in open set scenarios.In view of the limitations of existing residual networks in extracting the range of receptive fields for each network layer in hyperspectral open set classification research,a multiscale backbone residual feature fusion network open set hyperspectral image classification model is designed.The output of the multiscale backbone residual encoder module includes a variety of sizes,scales,and quantities of receptive fields,allowing the convolutional layer to express multiscale features,At the same time,the connection operation between subsets can obtain a larger receptive field,more effectively process feature map information,and thereby improve the classification accuracy of open set hyperspectral images.In view of the problem that convolutional neural network models often generate a large number of features in open set hyperspectral image classification,resulting in a high degree of redundancy affecting the classification effect,a hybrid link network open set hyperspectral image classification model is designed and reconstructed.The hybrid link network in the encoder module can reduce feature redundancy to a certain extent during the feature extraction process,extracting more discriminative features from hyperspectral images,thereby improving the accuracy of classification of open set hyperspectral images.