• The proposed knowledge-embedding-based joint resource allocation method improves average system throughput by 31.19% over MAML-PPO and 80.91% over RL2 in dynamic maritime channel environments.
• An action distribution alignment module prevents unreasonable action-output combinations, significantly enhancing radio resource utilization in complex maritime networks.
• Integrating knowledge embedding with meta-reinforcement learning formulates a physical guidance loss function that reduces required training samples and boosts model generalization capability.
• The approach addresses limitations of traditional heuristic allocation methods, offering flexible, scalable, and real-time resource allocation essential for B5G/6G maritime communications.