A Multi-path Neural Network for Hyperspectral Image Super-Resolution

The resolution of hyperspectral remote sensing images is largely limited by the cost and commercialization requirements of remote sensing satellites. Existing super-resolution methods for improving the spatial resolution of images cannot well integrate the correlation between spectral segments and the problem of excessive network parameters caused by high-dimensional characteristics. This paper studies a multipath-based residual feature learning method, which simplifies each part of the network into several simple and effective network modules to learn the spatial spectral features between different spectral segments. Through the designed multi-scale feature generation method based on wavelet transform and spatial attention mechanism, the non-linear mapping ability for features is effectively improved. The verification of three general hyperspectral data sets proves the superiority of this method compared with the existing hyperspectral SR methods.