Abstract
Accurate water surface elevation estimation is essential for understanding nearshore processes, but it is still challenging due to limitations in measuring water level using in-situ acoustic sensors. This paper presents a vision-based water surface elevation estimation approach using multi-view datasets. Specifically, we propose a visual domain adaptation method to build a water level estimator in spite of a situation in which ocean wave height cannot be measured directly. We also implemented a semi-supervised approach to extract wave height information from long-term sequences of wave height observations with minimal supervision. We performed wave flume experiments in a hydraulic laboratory with two cameras with side and top viewpoints to validate the effectiveness of our approach. The performance of the proposed models were evaluated by comparing the estimated time series of water elevation with the ground-truth wave gauge data at three locations along the wave flume. The estimated time series were in good agreement within the averaged correlation coefficient of 0.98 and 0.90 on the measurement and 0.95 and 0.85 on the estimation for regular and irregular waves, respectively.
Introduction
Understanding nearshore processes, such as how waves interact with beaches and coastal structures, requires precise data on wave elevation. Traditionally, this is achieved using in-situ sensors like acoustic or capacitance gauges. However, these sensors are difficult to maintain in harsh ocean environments and only provide data for a single point. While video-based monitoring is a promising non-contact alternative, it faces two major hurdles: the "perspective problem," where top-down views compress height information, and the "labeling problem," where manually tracing water lines in thousands of frames is labor-intensive and limits scalability.
Experimental Setup
The study utilized a custom dataset generated at the Korea Institute of Ocean Science and Technology (KIOST) using a 50-meter hydraulic wave flume. The facility was equipped with three capacitance wave gauges to provide ground-truth data and two synchronized CCTV cameras providing side-view and top-view perspectives.
The dataset included 57 video sequences capturing both regular waves (consistent periods) and irregular waves (simulating random sea states using the JONSWAP spectrum). Over 100,000 frames were processed at 30 fps to train and validate the proposed models.
Methodology
The researchers developed a two-stage deep learning framework to bridge the gap between laboratory measurements and field-deployable top-view monitoring.
Semi-Supervised Segmentation
To solve the labeling problem, the team used the FloodFill algorithm to automatically generate initial masks for side-view images. These were then used to train a U-Net architecture with a small set of valid automated labels. This allowed the model to learn to distinguish water from the background across long sequences with minimal human intervention, proving more robust than traditional intensity-based thresholding.
Deep Visual Domain Adaptation
The core innovation is the use of domain adaptation for top-view estimation. A Source CNN extracts height features from side-view images where the ground truth is clear. Simultaneously, a Target CNN is trained on top-view images. By minimizing the L2 loss between these two feature sets, the Target CNN learns to map top-view visual patterns—such as foam movement and light reflection—to the height features learned from the side-view.
Results and Discussion
The system's performance was evaluated by comparing its estimates to physical wave gauge data. For side-view measurements, the model achieved an average correlation coefficient (R) of 0.98 for regular waves and 0.90 for irregular waves. The top-view estimation via domain adaptation achieved an R of 0.95 for regular waves and 0.85 for irregular waves.
Root-Mean-Square Error (RMSE) for top-view estimation was approximately 4.13 cm for regular waves and 2.28 cm for irregular waves. The models performed best at locations closest to the cameras, highlighting that image resolution remains critical for physical measurements. While the model tended to underestimate water levels at extreme peaks, it demonstrated significant robustness to "messy" data, such as afterimage effects and transient wave breaking.
Conclusion
This study demonstrates that deep visual domain adaptation can effectively estimate wave elevation from top-down perspectives by leveraging latent visual information. Future work will focus on improving peak estimation and extending the framework to 3D wave flume environments to better simulate complex real-world coastal conditions.
Summary updated