Energy-Efficient Computation Framework for Remote Sensing Data Processing in LEO Satellite Network

The exponential increase in Low Earth Orbit (LEO) satellite deployments has significantly boosted Earth observation capabilities, necessitating advanced data-processing solutions to handle the surging volume of remote sensing data. However, the “bent-pipe” architecture between satellites and ground station can lead to high latency and data loss issues. Thus, processing data on the LEO satellite becomes a feasible approach. Given the unique constraints of space, deciding how to offload tasks and process them within limited energy remains one of the key challenges in on-orbit satellite processing. Traditional Deep Reinforcement Learning (DRL) algorithms have the capability for autonomous decision-making, but their high time and energy consumption for inference limit their application in space scenarios. This paper explores the Broad Reinforcement Learning (BRL) with an Actor-Critic architecture for Satellite Edge Computing (SEC), aiming to enhance on-orbit decision-making and optimizing computational resource utilization in LEO satellites. We propose a computational offloading decision framework that reduces on-board computing energy consumption through model compression and the low decision-cost advantages of BRL, enhancing the efficiency of on-orbit computing. This approach can maintain system performance while reducing the dependency on satellite system resources, thus improving the efficiency of remote sensing data processing. The simulation results show that, compared to the DRL decision framework, the BRL decision framework can reduce energy consumption by 15%. Furthermore, the computation framework that incorporates knowledge distillation can reduce the overall energy consumption to 39% of the original baseline.