Underwater visual perception is a fundamental capability for intelligent unmanned systems. Among various approaches, monocular depth estimation has emerged as a key research topic due to its low cost and broad applicability. To address the dual challenges of underwater optical distortion and real-time inference constraints, this paper proposes a lightweight depth estimation model that integrates an RMI input space transformation with an enhanced Adaptive Bins framework. Specifically, we first transform RGB inputs into the RMI space, leveraging the attenuation characteristics of red light in underwater environments to alleviate visual information loss. Then, a U-Net-based architecture is designed, incorporating a lightweight encoder and a modified AdaBins module. This combination enables dynamic depth interval partitioning and multi-scale feature fusion, enhancing model robustness. Finally, a hybrid loss function is employed to optimize the training process. Experiments conducted on the USOD10K dataset demonstrate that our method outperforms existing models in both threshold accuracy (δ₁/δ₂/δ₃ = 0.460/0.691/0.813) and error metrics (AbsRel = 0.687). The proposed model significantly improves depth differentiation for fine-grained objects and complex scenes, while maintaining real-time performance, offering an efficient perception solution for underwater unmanned systems.