Cross-spectral Domain Adaptation for Hyperspectral Image Super-Resolution

Deep learning methods have seen great success for superresolution tasks over a variety of domains. However, for super-resolution in hyperspectral imaging (HSI) one of the bottlenecks for performance is the availability of high-quality training data sets with high spatial resolution. This issue is further exacerbated in the case of HSI spectral domains for which accessible data is even more scarce, such as those in the infrared range. To address this problem, we propose a novel unsupervised adversarial domain adaptation method that allows the transfer of knowledge from a model trained on hyperspectral images from a range of the electromagnetic spectrum with high-resolution training data to a target spectral range where ground truth high-resolution data is unavailable. We illustrate the performance of our method on a novel hyperspectral data set in the short-wave infrared (SWIR) spectrum. When applying our domain adaptation method, a performance increase can be observed, compared to standard super-resolution methods, in the target domain. The implementation and data set are publicly available on the repository: https://github.com/RMA-4DPL/A-SNLSR.