Underwater depth estimation is crucial to understanding underwater three-dimensional scenes, which can support underwater exploration, archaeology, marine mining and other fields. Although many depth estimation methods can work well in structured scenes like indoor scenes, these methods can easily fail in unstructured and turbid underwater scenes. On the other hand, current underwater datasets for depth estimation are quite limited. Insufficient variety and data volume greatly restrict the development of data-driven-based underwater estimation approaches. To address these issues, we propose a novel underwater depth estimation model with a spatial channel attention module, which can improve the feature perception ability of low-level features in unstructured and turbid environments. Furthermore, we release an underwater video dataset with Ultra-High-Definition (UHD) 4K videos to support self-supervised training process. Experimental results prove that the proposed effective network design achieves superior performance.