SEC-DT: Satellite Edge Computing Enabled Dynamic Data Transmission Based on GNN-Assisted MARL for Earth Observation Missions

Abstract

Recent advancements in low Earth orbit (LEO) satellite technology have facilitated a substantial increase in the number of Earth observation (EO) satellites launched. However, transmitting voluminous imagery generated by these EO satellites to the ground still faces the challenges of limited satellite resources and dynamic satellite networks. To address this problem, we propose SEC-DT, a S atellite E dge C omputing (SEC) enabled computation-aware dynamic D ata T ransmission framework for jointly optimizing the routing selection and in-orbit imagery compression adoption. Specifically, we formulate an online optimization problem for concurrently delivering data from multiple EO satellites in a single EO mission, aiming to minimize the overall transmission and computation latency while ensuring the decent quality of the final downloaded data. Then we cast the problem as a partially observable Markov decision process and adopt an augmented multi-agent reinforcement learning (MARL) algorithm to solve this intractable online decision problem. Considering the natural graph structure of the satellite network, we innovatively integrate the graph neural network (GNN) into the MARL algorithm to form a GNN-assisted MARL framework, wherein GNN can capture the enriched semantic information present in satellite topology to achieve the fusion of diverse environmental states, which is beneficial for improving the decision-making effectiveness of agents. Finally, we conduct extensive experiments and ablation studies in various settings based on real-world datasets of StarLink and SkySat constellations. The experimental results have demonstrated the scalability and excellent performance of our algorithm compared with other baseline schemes.