Graph Learning-Based Multi-Dimensional Adaptive Scheduling for Multi-QoS Remote Sensing Services in Satellite Networks

Abstract

With the rapid development of inter-satellite laser links, Very low Earth orbit (VLEO) constellations are showing a clear trend of intra-layer networking. As a result, RS transmission is no longer limited to the low Earth orbit (LEO) segment, and the optimization scope should be extended to the full VLEO–LEO process. However, full-process scheduling for multi-QoS RS services faces two major challenges. First, it is highly sensitive to environmental dynamics, since changes in network state may directly cause a group of services to shift from QoS-satisfied to QoS-violated. Second, multiple processes, including inter-layer topology formation, link latency mapping, and scheduling parameter adaptation, form a coupled decision chain that jointly determines the final scheduling result. This imposes strong demands on a system-level adaptive scheduling framework. To address these challenges, this paper proposes a Multi-dimensional Graph representation learning-driven Scheduling and Routing (MGSR) framework. MGSR adopts a hierarchical and collaborative Graph Representation Learning (GRL) architecture to enable system-level adaptive scheduling across multiple dimensions. It includes adaptive uplink node selection, spatiotemporal link latency mapping, and QoS-aware iterative scheduling into a unified decision process. By leveraging graph neural networks to capture topology evolution and traffic dynamics, MGSR can rapidly adjust routing and scheduling strategies under varying service ratios, traffic intensities, and constellation topologies. The simulations show that MGSR consistently outperforms state-of-the-art routing algorithms, achieving significant improvements in QoS satisfaction ratio.