In-Orbit Lunar Satellite Image Super Resolution for Selective Data Transmission

TL;DR

This paper presents a novel in-orbit satellite image super-resolution system that improves image quality and reduces computational resource usage, enabling selective data transmission and efficient satellite data utilization.

Contribution

The paper introduces a new residual dense non-local attention network and a specialized loss function for in-orbit super-resolution on low-power devices, enhancing image quality and efficiency.

Findings

Outperforms RDN in PSNR by +0.1 dB and +0.19 dB in block-sensitive PSNR.

Consumes 48% less memory and 67% less peak power than standard RDN.

Demonstrates effectiveness on Kaguya DOMs for satellite image super-resolution.

Abstract

Rapid technological advancements have tremendously increased the data acquisition capabilities of remote sensing satellites. However, the data utilization efficiency in satellite missions is very low. This growing data also escalates the cost required for data downlink transmission and post-processing. Selective data transmission based on in-orbit inferences will address these issues to a great extent. Therefore, to decrease the cost of the satellite mission, we propose a novel system design for selective data transmission, based on in-orbit inferences. As the resolution of images plays a critical role in making precise inferences, we also include in-orbit super-resolution (SR) in the system design. We introduce a new image reconstruction technique and a unique loss function to enable the execution of the SR model on low-power devices suitable for satellite environments. We present a residual dense non-local attention network (RDNLA) that provides enhanced super-resolution outputs to improve the SR performance. SR experiments on Kaguya digital ortho maps (DOMs) demonstrate that the proposed SR algorithm outperforms the residual dense network (RDN) in terms of PSNR and block-sensitive PSNR by a margin of +0.1 dB and +0.19 dB, respectively. The proposed SR system consumes 48% less memory and 67% less peak instantaneous power than the standard SR model, RDN, making it more suitable for execution on a low-powered device platform.