- Program
- Master of Applied Ocean Technology, Ocean Mapping — Fisheries and Marine Institute of Memorial University
- Supervisors
- Dr. Sean Mullan and Paul Elliott
- Data
- NOAA side-scan sonar backscatter mosaic from survey H12642 (approaches to New Bedford, Buzzards Bay, Massachusetts)
- Methods
- Computer vision and deep learning: YOLO-based object detection and ResNet
- Presented
- Poster at the Canadian Hydrographic Conference 2024, St. John’s Convention Centre
Background
Boulders are one of the seabed features that surveys are routinely asked to map. They matter for routing cables, siting offshore wind foundations and anchoring, and for describing seabed habitat. On side-scan sonar a boulder is small: a bright return from the face toward the sonar and a dark acoustic shadow behind it, a few pixels across in a mosaic that may cover many square kilometres.
Picking them is usually manual. An interpreter scans the mosaic or the individual lines and marks each contact. It is slow over large areas, and the result can vary from one interpreter to another. That makes boulder picking a natural candidate for object detection, which is built to find many small instances of the same kind of thing in an image.
Data
The work used a side-scan sonar backscatter mosaic from NOAA hydrographic survey H12642, in the approaches to New Bedford in Buzzards Bay, Massachusetts. Jordan Pierce provided the NOAA training dataset and guidance on using it. Boulder samples were picked and labelled on the mosaic to build the training data, with help from Anne Joseph in picking the many samples needed.
Method
The core of the method is a YOLO (“You Only Look Once”) object detector, which predicts bounding boxes and confidence scores for every object in an image in a single pass. That suits sonar mosaics: they are large, the targets are small and numerous, and a detector that is fast at inference can be run over a whole survey area. ResNet, a residual convolutional network widely used for image recognition, was also part of the computer-vision work.
Presentation
I presented the project as a poster at the Canadian Hydrographic Conference 2024 at the St. John’s Convention Centre, under the title “Automated Boulder Detection on Side-Scan Mosaics using YOLO-based Deep Learning Model”.
Where it applies
Automated detection is most useful as a first pass that an interpreter then reviews: on large offshore wind and cable-route surveys, in pre-construction seabed clearance planning, and for repeat surveys where the same area is interpreted more than once. The aim is not to remove the interpreter but to give them a consistent starting set of contacts to confirm, reject or add to.
That is close to the work I later did on offshore wind surveys, where I built side-scan mosaics and ran anomaly and UXO reviews for client reports.
From research to software
Boulder detection carried on into Dolphin Explorer, the survey software I build. Its waterfall has a rule-based boulder detector that models the acoustic shadow directly, and a project can train its own RetinaNet model from a labelled dataset in YOLO or COCO format, so a dataset labelled the same way can be used to train it.
Acknowledgements
Thanks to my supervisors, Dr. Sean Mullan and Paul Elliott; to Jordan Pierce for his guidance and for providing the NOAA training dataset; and to Anne Joseph for her help picking boulder samples.
References
- NOAA National Centers for Environmental Information. Hydrographic survey H12642: Buzzards Bay and Nantucket Sound, approaches to New Bedford.
- Redmon, J., Divvala, S., Girshick, R. and Farhadi, A. (2016). You Only Look Once: Unified, Real-Time Object Detection. CVPR 2016.
- He, K., Zhang, X., Ren, S. and Sun, J. (2016). Deep Residual Learning for Image Recognition. CVPR 2016.