Task-Oriented Semantic Delivery in Large-Scale Heterogeneous Satellite Networks: A Local Topological Information-Dependable Deep Learning Approach

In the Large-Scale Heterogeneous Satellite Networks (LSHSNs) integrating Low Earth Orbit (LEO) and Medium Earth Orbit (MEO) satellites, data delivery faces complex topology and dynamic connectivity, which poses a significant challenge to current graph-dependable transmission strategies requiring global topological information, incurring huge computational cost and interactive overhead. To address this issue, in this paper, we propose a Local Information-dependable Semantic Delivery Mechanism (LISDM) by exploiting topological features at the semantic level, in which we develop a Task-oriented Semantic-aware Topology Compression Network (TSTCN) to condense the global topology according to specific task demands. Besides, we develop a Deep Q-Network enabled Semantic Coded Routing (DQNSCR) algorithm for the semantic delivery in the LISDM by designing a novel metric called Semantic Delivery Efficiency (SDE). The simulation results indicate that the proposed mechanism performs better in improving the required topology scale and throughput compared with typical data delivery mechanisms such as the conventional Open Shortest Path First (OSPF) routing algorithm, the DQN-based Intelligent Routing (DQN-IR) algorithm, and Real-Time Hop-by-Hop Routing (RTHop) algorithm with Space-Time Graph (STG) model, respectively.