Abstract
In the large-scale Marine Internet of Things, UAV-assisted task offloading as a promising solution alleviates the computational burden encountered by fixed maritime surface nodes in multi-task ocean scenarios. However, most existing works focus on UAV trajectory design and resource allocation, while overlooking the dynamic demands and optimization potential of maritime surface nodes. Moreover, static resource allocation strategies and single optimization methods often limit global search capability and adaptability in dynamic ocean environments. We propose SA-TD3, a hybrid decision-making framework. We design a UAV-assisted computation offloading and resource optimization mechanism from the perspective of maritime surface nodes to better capture dynamic task demands. Furthermore, we develop an enhanced TD3 algorithm that integrates simulated annealing with an environment-aware dual-channel advantage function, improving global search capability and policy robustness. Finally, we construct a graph neural network-based dynamic prioritized replay mechanism to capture inter-node correlations and improve training efficiency. Extensive experiments demonstrate that SA-TD3 reduces average latency by 19.7% and improves overall performance by 13.2% across diverse ocean environments, effectively reducing the computational load and communication latency of surface nodes while enhancing energy efficiency.