Hyperspectral image classification based on twin spectral attention consistency

Abstract

The invention relates to a hyperspectral image classification method based on twinborn spectrum attention consistency. The hyperspectral image classification method comprises the following steps: 1, performing data generation: performing segmentation operation on sample data x = {x1, x2, x3} extracted from a hyperspectral image to obtain data set features of two space blocks with different sizes respectively; 2, refining spectral characteristics: sending the obtained data set features into a Dense network, and refining all highly correlated spectral features again to obtain information; 3, performing spectral feature enhancement: by applying a channel attention mechanism, spectral characteristics are enhanced, and key channels are highlighted; 4, performing channel consistency regularization: establishing a special link for connecting an upper branch and a lower branch by assuming and modeling channel consistency; 5, integrating the classification loss and the channel consistency loss of the two branches into a unified network, performing end-to-end training, enabling the verification set to pass through the trained model to obtain an optimal training model, and then obtaining a classification result of the hyperspectral image through testing. According to the invention, the spectral attention consistency model is applied to hyperspectral image classification, and the consistency loss function is applied to surface feature classification of the hyperspectral image, so that the accuracy of hyperspectral image classification is improved.

Description

technical field [0001] The present invention relates to the technical field of image processing, in particular to the technical field of machine learning and hyperspectral image classification, and more specifically, to a hyperspectral image classification method based on twin spectral attention consistency, using Dense network, twin network, channel attention Force module and channel attention consistency regularization for hyperspectral image classification. Background technique [0002] Hyperspectral imaging is the continuous imaging of a certain area from the ultraviolet spectrum to the infrared spectrum with a nanoscale spectral resolution using an imaging spectrometer. As a special type of remote sensing image, hyperspectral image has rich spectral information, which creates good conditions for object recognition. Hyperspectral image classification is used in vegetation monitoring, atmospheric environment research, ocean remote sensing, urban planning and other fields ...

AI Technical Summary

Problems solved by technology

This method has shortcomings: it only pays attention to the spectral band information, ignores the spatial information, and cannot effectively extract the hyperspectral image features, thus failing to achieve the best classification effect

This method has the following deficiencies: Experiments have proved that the spatial information extracted by the spatial attention module used is not significant, and has no effect on improving the classification model, and the model ignores the consistency of attention under the spatial transformation of hyperspectral images.

Images