Abstract
Accurate bathymetric information is essential for coastal engineering, sedimentary studies, and archaeological mapping. While bathymetric LiDAR provides high-quality data, its high cost limits its accessibility for frequent or small-scale surveys. This paper presents a low-cost alternative using Unmanned Aerial Vehicle (UAV) imagery processed with Structure from Motion (SfM) and Multi-View Stereo (MVS) techniques. To overcome the significant depth underestimation caused by water refraction, a machine learning approach using Support Vector Regression (SVR) was developed. By training on known depth observations from LiDAR surveys, the SVR model effectively corrects drone-derived point clouds, achieving vertical accuracies that meet international hydrographic standards.
The Challenge: Water Refraction
Coastal mapping using standard cameras on drones faces a fundamental physical hurdle: refraction. According to Snell’s Law, light rays bend as they transition from water to air. In photogrammetric processing, this bending causes a systematic geometric distortion where the seabed appears shallower than its actual position. This "apparent depth" ($Z_0$) must be corrected to obtain the "real depth" ($Z$).
Traditional correction methods often rely on complex, hand-coded optical algorithms that are difficult to integrate into modern multi-view geometry workflows. This study proposes a data-driven shortcut: using machine learning to learn the systematic relationship between the erroneous apparent depth and the ground truth.
Methodology
The proposed workflow begins with generating an initial 3D point cloud from UAV imagery using standard photogrammetric software (Agisoft Photoscan). This raw point cloud contains the distorted apparent depths.
To correct these errors, the researchers employed Support Vector Regression (SVR) with a linear kernel. The model was trained to map the relationship between the apparent depth ($Z_0$) and the actual depth ($Z$) provided by a Leica HawkEye III bathymetric LiDAR system. SVR was chosen over simple linear regression due to its robustness; its "hinge loss" function allows it to ignore minor noise and outliers—such as sun glint or moving vegetation—that typically plague underwater imagery.
Key Insights and Results
Experimental results from two test sites in Cyprus (Amathouda and Agia Napa) revealed that the relationship between apparent and real depth is remarkably linear. This linearity allows for highly efficient correction models that do not require complex deep-learning architectures.
The SVR model demonstrated impressive performance:
- Accuracy: In the Agia Napa site (depths up to 14.8m), the mean depth error was reduced from 2.23m in the uncorrected data to just -0.10m after SVR correction.
- Standards: The corrected data achieved a vertical accuracy of approximately ±25cm, meeting the International Hydrographic Organization (IHO) standards for shallow water surveys.
- Generalizability: A model trained on the shallower Amathouda site (max depth ~5.5m) successfully corrected data for the deeper Agia Napa site, suggesting the learned refraction physics are transferable across different locations.
Conclusion and Future Work
This research demonstrates that machine learning can bridge the gap between low-cost drone imagery and expensive LiDAR systems for shallow water bathymetry. By treating refraction as a systematic error to be learned rather than a purely optical problem to be modeled, the SVR approach provides a robust and accessible solution for coastal researchers.
Future work will focus on testing the model in more turbid (cloudy) waters and integrating additional variables, such as the angle of incidence and water quality parameters, to further refine accuracy in diverse environmental conditions.
Summary updated