A mixed-supervision spectral super-resolution methodology for multispectral image

Existing convolutional neural work-based spectral super-resolution (SSR) for multispectral images are supervised by high-resolution hyperspectral images ($\mathbf{H}^{\mathbf{2}} \mathbf{S I s}$). However, $\mathbf{H}^{\mathbf{2}} \mathbf{S I s}$ are actually unknown. In practical, pairs of low-resolution hyperspectral and high-resolution multispectral image are accessible. Therefore, we propose an SSR methodology combining self-supervision and supervision, which is based on the linear degradation relationship between the accessible image pairs. Under our methodology, the existing SSR network could be driven by the actual image pairs rather than the simulated $\mathbf{H}^{\mathbf{2}}$ SIs. The effectiveness and superiority of our methodology are demonstrated with the state-of-the-art SSR networks.