Nested U-Net with Separable 3D Convolution for Hyperspectral Image Super-Resolution

Hyperspectral visible-band image acquisition currently encounters challenges of low spatial resolution, which arises from the necessary trade-offs between achieving higher spectral resolution and ensuring rapid data capture. Deep learning-based single-modality image super-resolution methods have proven effective in recovering the resolution of hyperspectral image sequences and garnered considerable attention. The main challenge in hyperspectral image super-resolution lies in leveraging spectral relevance across successive bands and spatial details in a three-dimensional (3D) image cube. To address this issue, this paper introduces a Nested U-Net (NUNet) for single-modality hyperspectral image super-resolution. NUNet utilizes interval grouping to minimize information redundancy and incorporates a nested U-shaped structure with skip and residual connections to deepen the network and enhance multi-level information utilization. We also develop a separable residual 3D unit featuring parallel separable 3D convolutions to effectively extract and integrate spatial and spectral features. The method includes a feature context fusion module for merging feature information across groups and employs pixel shuffle for high-resolution image reconstruction. Experimental evidence demonstrates NUNet’s strengths over existing state-of-the-art techniques.