Autonomous underwater vehicles and remotely operated robotic systems increasingly depend on vision systems to perceive complex seabed topography and navigate surrounding hazards, yet acquiring reliable geometric information underwater remains difficult. Traditional underwater depth sensing has often relied on dedicated hardware such as acoustic sonar arrays, synchronized stereo cameras, or specialized LiDAR units. However, deploying these instruments on compact underwater robots introduces significant engineering hurdles related to physical payload dimensions, energy consumption, and total integration expense, alongside sensor-specific performance limits. In response to these operational constraints, estimating continuous three-dimensional scene depth from a single onboard color camera has emerged as a promising research alternative. Translating standard two-dimensional images into metric distances requires learning a mapping function to predict the separation between each image pixel and the camera optical center. In self-supervised learning frameworks, where ground truth depth annotations are absent, models are typically trained through view synthesis, requiring the network to jointly estimate dense depth and inter-frame camera motion between successive video frames.
Despite considerable progress in terrestrial monocular depth estimation, adapting these computational methods to marine environments poses acute challenges due to the optical physics of water. Submerged cameras operate under severe optical degradation caused by wavelength-dependent light attenuation, medium scattering, and pervasive color distortion. Earlier underwater depth estimation strategies attempted to adapt terrestrial models via transfer learning or used synthetic training datasets, but synthetic datasets incur heavy computational overhead while terrestrial models struggle to bridge the extensive optical divergence between atmospheric and aquatic environments. Other traditional techniques exploited hand-crafted optical priors, such as blurriness or the differential attenuation across color channels where red light attenuates much faster than blue and green wavelengths. While methods based on the dark channel prior, underwater dark channel prior, and generalized dark channel prior demonstrate that medium transmission maps correlate inversely with scene depth, these formulations often degrade under fluctuating water turbidity, dynamic illumination, and structural complexity, rarely undergoing direct verification against ground truth depth.
To systematically address these optical constraints without requiring ground truth supervision, a research team led by Peng Yao, Lang Xiang, Yalu Wang, and Bin Fan developed PGU-Depth, a self-supervised monocular depth estimation framework guided directly by underwater physical imaging principles. Rather than relying solely on standard RGB feature extraction or unconstrained neural representations, the system incorporates physical imaging models to provide structured guidance during depth learning. In standard underwater optical formulations, such as the Jaffe-McGlamery model, captured image intensity is governed by the scene radiance multiplied by medium transmittance plus ambient background light. Because the medium transmittance decreases exponentially with scene distance according to the medium attenuation coefficient, the inverted transmission map serves as a natural geometric cue that correlates positively with real scene depth. PGU-Depth extracts this physical transmission map directly from raw underwater frames using the generalized dark channel prior, providing spatial consistency and clear structural boundaries to anchor the learning process.
The overall PGU-Depth architecture processes raw RGB inputs and inverted transmission maps through a coordinated dual-branch encoder coupled with an information-preserving decoder. When an RGB frame enters the network, a joint feature extraction module fuses local spatial details with global image context to produce shallow feature maps. In parallel, the single-channel transmission map derived from the generalized dark channel prior is converted into an inverted transmission representation. Because single-channel physical priors differ markedly in representational capacity compared to multi-channel color representations, the architecture introduces a specialized underwater physical information extraction module. This extraction module expands the single-channel inverted transmission input into a rich multi-channel feature space matching the channel dimensions of the RGB network. It then applies progressive downsampling across multiple spatial scales, ensuring that high-level physical representations align structurally with the hierarchical representations generated across the primary image encoding pipeline.
Within the underwater physical information extraction module, feature optimization is achieved through a dual-branch processing strategy designed to capture fine structural contours and prevent feature degradation. An auxiliary path utilizes depth-wise separable convolutions to extract detailed boundary and texture cues with minimal computational overhead, specifically targeting abrupt depth transitions and object outlines where transmission values shift rapidly. The resulting features are concatenated with preliminary features from the primary path and passed into a channel attention mechanism that dynamically amplifies the most informative physical feature channels. An optimized spatial attention mechanism then refines spatial focus by combining average pooling and max pooling operations along the channel dimension, generating complementary spatial description maps that simultaneously reflect global medium distribution and local salient regions. To prevent the attention weights from suppressing original physical properties, a final residual connection preserves the underlying optical prior.
To integrate these refined physical representations with primary visual features, the researchers designed an underwater physical information aggregation module deployed at multiple encoder resolution stages, specifically across scales corresponding to one-fourth, one-eighth, and one-sixteenth of the input resolution. Because underwater degradation is spatially non-uniform, relying unconditionally on physical priors can inject noise if the transmission estimation is locally distorted. The aggregation module mitigates this vulnerability through a three-stage fusion pipeline comprising an adaptive fusion mechanism, a multiscale texture enhancement block, and a gated fusion network. The adaptive fusion stage computes dynamic channel weights using global average pooling and a Softmax activation, balancing RGB features against physical transmission features while employing an interaction branch for hybrid non-linear modeling. A gated regulation step then adjusts how strongly physical features influence the composite representation before spatial texture refinement.
The multiscale texture enhancement stage inside the aggregation module subsequently applies squeeze-and-excitation channel attention to emphasize depth-discriminative representations before performing spatial pooling. Average and max pooling operations are concatenated and routed through parallel convolutional filters with varying receptive field sizes of three, five, and seven pixels. This multiscale receptive design captures localized boundary transitions alongside broad environmental gradients, generating an enriched spatial attention map that guides depth consistency across blurred regions. Finally, the gated fusion network computes a dynamic gating signal from the concatenated enhanced features and original RGB features using a convolutional layer and a Sigmoid function. This final gating mechanism regulates the balance between texture-enhanced features and initial color features, preventing the depth estimation model from over-relying on imperfect physical priors while maintaining structural coherence across the reconstructed scene geometry.
At the decoding stage, the network progressively reconstructs spatial resolution starting from the lowest resolution level using an information fusion decoder. To prevent the loss of geometric information during successive upsampling operations, the decoder re-incorporates physical prior representations. In comparative ablations of decoder fusion strategies, direct pixel-wise addition of transmission maps yielded modest improvements due to rigid linear superposition, whereas a gated mechanism permitted dynamic weighting. The highest performance was achieved using a shared feature fusion configuration, where inverted transmission features previously extracted and refined by the encoder are shared directly with the decoder through a gating mechanism. To supervise training without ground truth labels, the framework employs view synthesis with a lightweight PoseCNN network estimating relative camera transformations between adjacent frames, optimizing an objective function combining an edge-aware smoothness loss, an underwater light attenuation loss, and a photometric reprojection loss that weights structural similarity at ninety percent and pixel-wise L1 differences at ten percent.
Quantitative evaluations conducted on the FLSea-VI marine benchmark—comprising 22,451 underwater images across twelve distinct scenes from Red Sea areas and complex canyon terrains—demonstrated that PGU-Depth outperformed both pure physical prior models and standard self-supervised deep learning baselines across all error metrics, including absolute relative error, squared relative error, root mean squared error, and logarithmic root mean squared error. Qualitative comparisons revealed that while traditional methods suffered from blurred object boundaries and pure deep learning baselines lost fine structural details, PGU-Depth accurately recovered sharp peaks, slender cable lines, and small swimming fish. In targeted evaluations across four challenging environments—tiny canyon, U canyon, landward path, and pier path—the model exhibited higher overall prediction accuracy than specialized distillation frameworks like WaterMono. Ablation experiments verified that while the baseline model with dual-branch physical extraction reduced prediction errors, maximum accuracy required the unified deployment of the extraction module, the multi-stage aggregation module, and the shared-feature decoder.
Cross-dataset evaluations on the USOD10K dataset, involving 1,026 test samples spanning diverse marine objects and complex backgrounds, confirmed that PGU-Depth generalizes robustly to unseen underwater environments without retraining. In complex scenes featuring human statues and divers, the model produced coherent depth distributions that clearly separated foreground subjects from ambient aquatic backgrounds. Despite these demonstrated advantages, the researchers noted specific operational limitations in environments with extensive optical caustics, where dynamic water surface wave action produces non-physical illumination patterns that distort transmission map estimation and degrade local depth accuracy. Additionally, training and evaluation were carried out using a single fixed random seed at an input resolution of 640 by 192 pixels on a single graphics processing unit, balancing detail retention against the computational limits of underwater vehicles. The authors identified multi-run evaluations across diverse random seeds as an essential focus for ongoing research to further establish statistical robustness across operational marine environments.
Subject of Research: Self-supervised monocular depth estimation for underwater robotics using optical physical guidance
Article Title: PGU-Depth: an underwater self-supervised monocular depth estimation method based on physical guidance
Article References: Yao, P., Xiang, L., Wang, Y., & Fan, B. (2026). PGU-Depth: an underwater self-supervised monocular depth estimation method based on physical guidance. Intelligent Marine Technology and Systems, 4(1), Article 30. https://doi.org/10.1007/s44295-026-00120-8
Image Credits: AI Generated
DOI: 10.1007/s44295-026-00120-8
Keywords: Underwater Robotics, Monocular Depth Estimation, Self-Supervised Learning, Computer Vision, Physical Guidance, Transmission Map, Optical Attenuation, Deep Learning, View Synthesis, Encoder Decoder, FLSea Dataset, USOD10K Dataset, Marine Technology, Autonomous Underwater Vehicles, Image Restoration, Photometric Reprojection
Cite Scienmag News
Blake Davidson. (November 17, 2026). Physical Guidance Framework Enhances Self-Supervised Monocular Underwater Depth Estimation Accuracy. Scienmag. https://scienmag.com/physical-guidance-framework-enhances-self-supervised-monocular-underwater-depth-estimation-accuracy/
Blake Davidson. "Physical Guidance Framework Enhances Self-Supervised Monocular Underwater Depth Estimation Accuracy." Scienmag, 17 November 2026, https://scienmag.com/physical-guidance-framework-enhances-self-supervised-monocular-underwater-depth-estimation-accuracy/. Accessed 17 November 2026.
Blake Davidson. "Physical Guidance Framework Enhances Self-Supervised Monocular Underwater Depth Estimation Accuracy." Scienmag. November 17, 2026. https://scienmag.com/physical-guidance-framework-enhances-self-supervised-monocular-underwater-depth-estimation-accuracy/








