Toward Intelligent Space-Air-Ground Integrated Network: Architecture, Challenges, and Emerging Directions

Abstract

As a foundational architecture for next generation communication systems, the space-air-ground integrated network (SAGIN) is driving the evolution toward global, resilient, and task-aware connectivity. However, SAGIN is characterized by highly dynamic topologies, heterogeneous nodes, and strong task-driven demands, which pose unprecedented challenges to the real-time performance of scheduling strategies. Artificial intelligence (AI) technologies, particularly reinforcement learning, deep neural networks, and large-scale models, provide promising solutions to these structural bottlenecks. This paper introduces a system-layer-oriented AI capability alignment framework to map model architectures to communication demands, and analyzes key deployment challenges including edge inference, policy consistency, and cross-domain knowledge transfer. We also present a comprehensive review of AI-driven applications in SAGIN, with a focus on critical tasks such as link and routing selection, resource scheduling, traffic offloading, unmanned aerial vehicles (UAV) deployment optimization, semantic communication, and large model integration. Based on this review, the paper outlines future research trends and identifies core technical bottlenecks. The goal is to provide methodological guidance and a development roadmap for building a new generation of intelligent, scalable, and cross-domain adaptive SAGIN architectures.