As the maritime industry evolves, the need for efficient network applications, such as marine navigation and ocean monitoring, along with reliable communication and computing services, has become critical. However, traditional terrestrial networks struggle with sparse base station deployment and poor signal coverage, resulting in unreliable connectivity. Prior research shows that low Earth orbit (LEO) satellite networks can improve coverage but face challenges like long communication distances and high operational costs, increasing latency and limiting support for high-demand applications. To address these issues, this paper proposes a novel space-air-ground-sea multi-layer architecture for computation offloading, integrating high-altitude platform stations as relay nodes. Using a multi-agent proximal policy optimization approach, modeled as a partially observable Markov decision process, the framework optimizes task offloading and resource allocation to minimize delays, energy consumption, and operational costs. Numerical results demonstrate that the proposed method outperforms baseline approaches in terms of overall performance metrics, achieving reductions of 14.4% and 10.6%, respectively.