Marine oil spills cause severe damage to ocean ecosystems and coastal economies. Timely scheduling of oil spill emergency resources is crucial for effective emergency response. Addressing the challenge of efficient resource scheduling under dynamically evolving oil spill conditions in marine environments, this paper proposes a hierarchical hybrid intelligent architecture. Firstly, it integrates an improved A*-PSO path planning module and an attention-based Bayesian Proximal Policy Optimization (SA-BPPO) reinforcement learning module to construct a 'path navigation-decision optimization' model for oil spill emergency resource scheduling. The improved A*-PSO path planning module generates initial paths and performs smoothing optimization, enabling rapid planning of efficient paths from start to end points, enhancing their feasibility and safety. Secondly, the self-attention Bayesian PPO (SA-BPPO) reinforcement learning module focuses on optimizing ship cleaning strategies. By incorporating an attention mechanism, the model’s ability to focus on key information is enhanced. The Bayesian method estimates the uncertainty of strategy outputs, allowing ships to dynamically adjust cleaning strategies according to environmental changes. Experimental results demonstrate that the proposed architecture significantly outperforms algorithms like Deep Q-Learning (DQN), Soft Actor-Critic (SAC), Ant Colony Optimization (ACO), and Dijkstra in key indicators such as cleaning time, efficiency, and path optimization. Compared to DQN, the improved A*-PSO-SA-BPPO algorithm improves average cleaning efficiency by 36\%. Compared to SAC, it improves average cleaning efficiency by 9.17\%. Compared to ACO and Dijkstra algorithms, it improves average cleaning efficiency by 36.78\% and 75\%, respectively. Ablation experiments further validate the effectiveness of each module that adding the improved A*-PSO module reduces cleaning time by 17.24\% and increases cleaning efficiency by 39.45\%.