Deep Attention Network For Hyperspectral Image Classification Research

As a hot issue in the field of remote sensing,hyperspectral image classification is widely used in geological exploration,environmental monitoring,modern military,etc.As one of the representative algorithms of machine learning,deep learning has made breakthrough progress in many research fields.Among them,convolutional neural network(CNN)is widely used in hyperspectral image classification because of its powerful performance.However,most network structures fail to distinguish important features from redundant information,and factors such as image noise and spatial variability of spectral features will seriously degrade the classification performance of the algorithm.To solve the above problems,this paper studies the network architecture based on residual structure and attention mechanism under the comprehensive analysis of hyperspectral image characteristics.The main research contents are as follows:(1)The deep learning theories are deeply studied,and a 3D convolutional network framework is constructed based on the data structure of hyperspectral images.The network can directly process 3D hyperspectral image data,which effectively preserves the spatial structure of image data,and can learn the spectral signatures and spatial signatures of pixels at the same time.(2)In order to distinguish important features from redundant information,residual attention module is embedded into the network.Firstly,the constructed residual attention module alleviates the information loss caused by the deepening of network layers through residual connection.Secondly,the attention module strengthens the spectral signatures and spatial signatures while suppressing unimportant information,which effectively reduces the interference of noise and heterogeneous pixels and enhances the effectiveness of the extracted features.(3)In order to better describe the spatial information of hyperspectral images,this paper further improves the attention mechanism and introduces the coordinate attention module.The coordinate attention module can integrate spectral information and spatial information into the attention map through 1D feature encoding,which further captures the long-distance dependence between pixels and retains more accurate spatial position information,thus improving the performance of the algorithm.(4)In order to demonstrate the effectiveness of the algorithms,the proposed algorithms are tested on three commonly used data sets.Experimental results show that the proposed algorithms not only have stable performance and fast convergence speed,but also significantly improve the classification performance.