Hyperspectral Image Classification Based on Capsule Network

• 作 者: MA Qiaoyu ZHANG Xin ZHANG Chunlei ZHOU Heng
• 机 构: School of Science,China University of Geosciences(Beijing),Beijing 100083,China;School of Statistics
• 出 处: 电子学报（英文版）
• 发表日期: 2022年1期
• 关键词: Capsule network Space-spectrum fu-sion Hyperspectral Classification Deep learnin

The conventional convolutional neural network performs not well enough in the ground objects classification because of its insufficient ability in maintain-ing sensitive spectral information and characterizing the covariance of spatial structure,resulting from the narrow sensitive frequency band and complex spatial structure with diversity of hyperspectral remote sensing data which caused more serious phenomena of“same material,differ-ent spectra”and“different material,same spectra”.Therefore,an improved capsule network is proposed and introduced into hyperspectral image target recognition.A convolution structure combining shallow features and multi-scale depth features is put forward to reduce the phenomena of“different material,same spectra”firstly,and then the diversity of the spatial structure is ex-pressed by the capsule vector and sub-capsule division in channel wise,so that the averaging effect of the convolu-tion process is weakened in the spectral domain and the spatial domain to reduce the phenomena of“same materi-al,different spectra”.By comparing the experimental res-ults on the hyperspectral data sets such as Indian Pines,Salinas,Tea Tree and Xiongan,the capsule network shows strong spatial structure expression ability,flexible deep and shallow feature fusion ability in multi-scale,and its accuracy in target recognition is better than that of conventional convolutional neural networks,so it is suit-able for the recognition of complex targets in hyperspec-tral images.