@article{Zhuang2026, 
author = {Peixian Zhuang and Qi Tian and Yijian Wang and Jiaheng Han and Yuchun Wang},
title = {Underwater Monocular Depth Estimation: A Survey of Theory, Methods, and Applications},
year = {2026},
journal = {CAAI Artificial Intelligence Research},
keywords = {Underwater scene, monocular depth, formation model, deep learning, multimodal fusion},
url = {https://www.sciopen.com/article/10.26599/AIR.2026.9150016},
doi = {10.26599/AIR.2026.9150016},
abstract = {Underwater monocular depth estimation aims to infer the distances between scene objects from the camera from a single underwater image. Nevertheless, the propagation of visible light in water is subject to wavelength-dependent absorption and scattering, which introduce color distortion, low contrast, blur, and texture degradation, thereby hindering the reliable extraction of depth cues. The degradation of underwater images varies with the propagation distance of visible light, providing a physical basis for scene distance estimation. Leveraging this property, underwater monocular depth estimation has evolved from inferring distances using single physical cues to comprehensive estimation approaches that integrate deep learning with multi-source observations. Starting from the theoretical foundations of underwater imaging and light propagation, Bayesian estimation, and prior constraints, our survey classifies existing methods into three categories: traditional physical-prior methods, deep-learning-driven methods, and multimodal fusion methods. Moreover, we review the utilization of depth cues across a range of underwater tasks, including underwater image enhancement and restoration, underwater three-dimensional (3D) reconstruction, underwater object detection, autonomous underwater vehicle (AUV) navigation and localization, and underwater robotic operations. On this basis, our survey analyzes current challenges, including complex optical degradation, insufficient cross-domain generalization, the scarcity of high-quality real-world data, and the difficulty of balancing accuracy with computational efficiency. Further, we outline promising avenues for future research, including the integration of underwater imaging priors and deep learning, the improvement of multimodal fusion reliability, the development and benchmarking of real-world underwater depth datasets, and the advancement of real-time depth estimation tailored for underwater intelligent devices.}
}