Towards Deep Line-Based Architectures for Onboard Hyperspectral Image Super-Resolution

Hyperspectral images can provide the fine spectral resolution needed for several tasks involving material identification, but they suffer from limited spatial resolution due to instrument constraints. The emerging paradigm of onboard computing seeks to address several detection problems of interest directly onboard of a satellite, in order to minimize latency due to the downlink and processing chain. In this paper, we design a novel lightweight architecture for onboard spatial super-resolution of hyperspectral images with the goal of improving onboard detection methods that rely on such images. The core of our contribution is a novel neural network architecture that works in a line-by-line fashion thanks to a recurrent-attentive mechanism in the along-track direction. This design greatly limits the memory requirements and computational complexity of the model, making it suitable for onboard usage. We show that this novel architecture is competitive with respect to the state of the art in term of super-resolution quality, while providing significant savings in complexity and memory.