Dynamic Collaborative Inference for Multi-Type DNN Tasks in Space-Ground Networks: A DRL Approach

The advent of 6G networks is driving a paradigm shift toward integrated “space-air-ground-sea” connectivity, with satellite communication evolving from a supplementary role to a core infrastructure for ubiquitous coverage. Concurrently, the rapid growth in AI-driven real-time processing of massive remote sensing data through deep neural networks (DNNs) has intensified the demand for efficient computational strategies. While onboard DNN inference mitigates data transmission requirements, low Earth orbit (LEO) satellites face significant resource constraints, necessitating collaborative computing between satellites and ground systems. Existing satellite-ground collaborative frameworks, designed for traditional deterministic tasks, fail to address the computational complexity and large-scale data inherent in DNN inference, leading to potential latency violations and energy inefficiency. To overcome these challenges, this paper proposes a Joint Model Partitioning and Resource Allocation (JMP-RA) scheme. By dynamically adapting model partitioning and resource allocation based on real-time channel conditions and task demands, JMP-RA minimizes onboard energy consumption while ensuring reliable task completion. Simulation results show that JMP-RA achieves a 19.3% reduction in task drop rate and a 5.1% improvement in energy efficiency compared to conventional methods. Furthermore, JMP-RA demonstrates strong adaptability across diverse network configurations (bandwidth: 50–300 MHz; computational capacity: 0.1–0.35 TFLOPS), validating its effectiveness in dynamic space-ground edge computing environments.