M³FuNet: An Unsupervised Multivariate Feature Fusion Network for Hyperspectral Image Classification

Hyperspectral image (HSI) spectral-spatial joint feature extraction methods generally suffer from low feature retention and weak spatial-spectral dependence, which will lead to single class feature confrontation (SCFC). To solve this problem, an unsupervised multivariate feature fusion network (M3FuNet) is developed in this paper. In M3FuNet, multiscale supervector matrix correction (MSMC) and multiscale random convolution dispersion (MRCD) is used as the spectral and spatial feature extraction method, and the feature retention of spectral and spatial features is improved to achieve feature calibration by feature fusion and decision fusion, called multivariate feature fusion. The MSMC is employed to correct supervector matrix to reduce the intraclass variance in superpixel homogeneous regions, overcome the phenomenon of supervector block drift (SvBD). The MRCD uses random convolution and gaussian smoothing to extract deep spatial features. Due to the similar feature representation ability of the MSMC and MRCD, the obtained spectral-spatial joint features have high feature retention and strong spectral-spatial dependence. Finally, this multivariate feature fusion network is used for realizing classification of HSI. Three commonly HSI datasets are used to validate the effectiveness of the M3FuNet. The experiment results show that the M3FuNet has more superior performance by comparing with several state-of-the-art HSI classification methods. The code of the proposed M3FuNet is available at http://github.com/aichou233/M3FuNet.