1. Introduction
With the rapid advancement and extensive deployment of Internet of Things (IoT) technologies, the Marine Internet of Things (M-IoT) has become a critical enabler of a broad range of maritime applications, including oceanographic monitoring, marine resource exploration, environmental surveillance and intelligent maritime traffic management [1,2,3]. By interconnecting diverse marine sensors, unmanned platforms, and communication terminals, the M-IoT facilitates enhanced situational awareness and supports intelligent decision-making in complex ocean environments [4,5,6,7].
However, the inherently harsh and heterogeneous nature of marine settings poses significant challenges to the computational efficiency and operational sustainability of M-IoT systems. These challenges are exacerbated by the widespread reliance on resource-constrained edge devices, intermittent and high-latency satellite communication links, and limited onboard energy supplies [8]. Collectively, these constraints impede real-time, efficient task execution, diminish system responsiveness, and restrict the scalability and adaptability of distributed computing architectures in maritime contexts.
Meanwhile, in real-world deployment, the strict payload and energy restrictions of UAV platforms impose significant limits on onboard neural inference and edge server integration, as continuous policy execution and GPU-based acceleration can substantially reduce flight endurance. Maritime communication links are subject to intermittent obstruction, long propagation delays and fluctuating bandwidth, requiring robust channel estimation, adaptive modulation and coding schemes and predefined fallback policies during connectivity outages. Moreover, unanticipated environmental variations demand real-time policy adaptation through lightweight online fine-tuning or federated meta-reinforcement learning, implemented with careful management of computational and communication budgets. Finally, when extended to multi-UAV operations, decentralized or hierarchical training frameworks with periodic aggregation of local experiences can preserve the collective performance without overwhelming the limited high-capacity links available at sea.
2. Related Work
Resource management and computation offloading optimization problems usually involve non-convex optimization with multi-objective coupling, such as power control, user association and arithmetic allocation. To address this problem, several classes of traditional algorithms are applied by researchers. For instance, the decomposition optimization algorithm is used to split the large-scale non-convex programming problem into easily solvable subproblems. In [9], the authors consider a binary offloading policy in a wireless-powered multi-user MEC system and propose a joint ADMM-based decomposition algorithm to tackle the combinatorial coupling between offloading decisions and time allocation. In order to overcome the doubly near-far effect, Hu et al. [10] address the “doubly near-far” problem in two-device...
