An Experimental Methodology for Automated Detection of Surface Turbulence Features in Tidal Stream Environments

Area of Science:

• Marine biology
• Renewable energy engineering
• Computer vision

Background:

• Tidal stream environments are crucial for marine renewable energy (MRE) and as foraging grounds for marine megafauna.
• Hydrodynamic features in these areas impact prey availability and megafauna behavior, posing risks for MRE device interaction.
• Uncrewed aerial vehicles (UAVs) provide high-resolution data on surface turbulence and animal presence, essential for environmental impact assessments.

Purpose of the Study:

• To develop and demonstrate an automated method for detecting surface turbulence features in UAV imagery.
• To assess the feasibility of using deep learning for analyzing large UAV datasets in tidal stream environments.

Main Methods:

• A Faster R-CNN deep learning model was employed for autonomous detection of kolk-boils (turbulence features).
• The model was trained on pre-labeled UAV images of kolk-boils, enhanced using environmental condition-specific techniques.
• Model performance was evaluated, with a 75-epoch variant showing optimal recall and precision.

Main Results:

• The study successfully demonstrated the automated detection of kolk-boils using a deep learning approach.
• The 75-epoch Faster R-CNN model achieved high average recall and precision.
• Limitations included a tendency for false positive detections, indicating areas for further refinement.

Conclusions:

• Deep learning offers a viable solution for automating the detection of surface turbulence in UAV imagery from tidal stream environments.
• Further research is needed to standardize data, benchmark models, and refine pre-processing techniques for improved accuracy.
• This automated approach can significantly aid in understanding and mitigating the environmental impacts of MRE development.