Speaker
Description
Underwater imaging generates large volumes of observational data, but converting these images into reliable, structured scientific observations remains challenging. This is particularly true for benthic organisms that retreat into the sediment when disturbed. The razor clam Ensis provides a representative case, as individuals can rapidly burrow, making direct surveys difficult and potentially biased.
We present BenthicAI, a data-science workflow combining computer vision, photogrammetry, and spatial data processing to detect and spatially reference elusive benthic fauna. ROV imagery is used to develop and evaluate a YOLO-based object detection pipeline that identifies Ensis from persistent surface features such as paired siphon openings, enabling non-invasive detection without triggering a behavioral response.
To preserve the spatial context of these detections, image sequences are processed photogrammetrically to generate georeferenced orthomosaics and local 3D reconstructions of the seabed. AI-based detections can thereby be linked to a common spatial reference system, enabling spatial analysis and relative size estimation.
The workflow addresses key challenges of underwater computer vision, including variable illumination, turbidity, sediment appearance, and image quality. By combining automated detection with spatial reconstruction, BenthicAI demonstrates how complex underwater imagery can be transformed into structured, spatially consistent, and potentially AI-ready observation data. The approach provides a scalable basis for integrating deep learning into marine observation workflows and future autonomous underwater vehicle deployments.