The Internet of Underwater Things (IoUT) enables distributed sensing and analytics in marine environments, yet suffers from fundamental constraints in communication capabilities, energy consumption, and data heterogeneity. We propose a hierarchical federated transfer learning (HFTL) framework under a cloud–edge–end paradigm, leveraging unmanned aerial vehicle (UAV) enabled edge computing and model partitioning for efficient distributed training. By decoupling models into frozen and trainable components, IoUT devices perform lightweight feature extraction, offloading intermediate features to UAVs for computation-intensive updates. A hybrid acoustic–electromagnetic communication scheme bridges underwater–aerial domains. Extensive evaluations demonstrate that our framework achieves superior trade-offs in payload, energy consumption, and training time, consistently outperforming baseline methods in both accuracy and system cost.