Autonomous docking and recovery of unmanned underwater vehicles is a key link for their long-term autonomous operation. Accurate perception of the underwater environment and real-time acquisition of depth information constitute the core foundation for accomplishing this task. However, existing lightweight monocular depth estimation networks suffer from insufficient feature representation capability and blurred edge details in complex underwater scenes. To address these issues, this paper proposes an improved self-supervised monocular depth estimation method tailored for underwater scenarios was proposed. Based on the Lite-Mono network, an adaptive decoupled feature enhancement convolution module is designed. By constructing a structured feature enhancement mechanism, this module improves the network's multi-scale feature representation capability and edge detail perception accuracy in underwater scenes. The Depth Anything V2 pre-trained model is introduced to generate high-quality pseudo-labels to assist self-supervised training, so as to make up for the shortage of real labeled data in underwater scenarios. In addition, the loss function is optimized according to the scene characteristics such as underwater light attenuation and color distortion. A depth-supervised loss and a boundary-weighted loss are designed to strengthen spatial constraints and improve edge sharpness. Based on the underwater dataset FLSea, ablation experiments are conducted to verify the effectiveness of each improved module. The results show that, compared with the baseline Lite-Mono, the proposed method reduces the absolute relative error metric for underwater depth estimation accuracy by 29.3% and improves the accuracy metric by 27.9%. In addition, it achieves better recovery of edge details and stronger robustness against complex underwater interferences, validating the effectiveness and superiority of the proposed method for monocular depth estimation in underwater scenarios.