Underwater object detection from sonar and subsea imagery
Automatically find and classify objects on the seabed in sonar, video and survey images.
Darkocean's object detection engine uses deep learning to find items in side-scan sonar, multibeam and subsea imagery that a person would otherwise spend hours scanning for. It speeds up review of large survey datasets and produces a consistent, reviewable list of targets.
Side-scan, multibeam, video, photos
Position, class and confidence
Detections are candidates
Object detection tool from Darkocean
How does AI object detection work on subsea data?
A survey vessel or AUV can collect hundreds of kilometres of side-scan sonar in a single campaign. Reviewing every line by eye is slow, and fatigue means small contacts get missed or logged inconsistently. Object detection models are trained on labelled examples of targets, such as pipelines, debris, anchors, wrecks, boulders, mines and unexploded ordnance, and then scan new data to propose where such objects appear.
Sonar images are not like photographs. Shape is distorted by range and geometry, and acoustic shadow, the dark area behind an object, often tells you more about its height and form than the object itself. Models trained on sonar imagery learn to use both the bright return and the shadow. Optical imagery from ROVs and towed cameras presents a different problem, with turbidity, backscatter and uneven lighting, so models are trained separately for each data type.
Each detection is returned with a position, a bounding box, a class and a confidence score. A reviewer then confirms, rejects or reclassifies it, and those decisions can be fed back to improve the model for your seabed type and sonar. Performance depends on how closely your data matches what the model was trained on, so a short trial on a sample of your own data is the honest way to judge it.
Detection is not identification. A sonar contact flagged as a possible munition, for example, has to be verified by a qualified person with the right data, and often by ROV inspection, before any decision is made.
How it works
Ingest
Load sonar mosaics, waterfall data, video or photographs with their navigation.
Detect
Run the model across the data to propose objects with class and confidence.
Review
A reviewer confirms, rejects or reclassifies each candidate in a single list.
Georeference
Convert confirmed contacts to positions with dimensions where the data allows.
Report
Export the target list and annotated images for the survey report or client.
What you receive
Targets
a. Contact list with position, class and confidence
b. Estimated length, width and height where available
c. GIS and CSV exports
Evidence
a. Annotated sonar and image extracts
b. Shadow-based height estimates for sonar contacts
Quality
a. Detection summary by class and survey line
b. Notes on coverage gaps and poor-quality data
Where it is used
Pipeline and cable inspectionLocate exposure, free spans and debris along subsea assets.
Route and site clearanceFind obstructions, boulders and wrecks before installation.
Munitions screeningFlag candidate contacts for specialist verification in areas with legacy ordnance.
Search and recoveryScan large sonar datasets for a specific object, such as a lost vessel or equipment.
Port and harbour surveysIdentify debris and obstructions in berths and approach channels.
Seabed habitat mappingCount and map objects such as reef structures or mooring scars across a survey area.
AI detection or manual review?
| AI-assisted detection | Manual review only | |
|---|---|---|
| Speed on large datasets | Scans the full dataset and proposes candidates | Limited by analyst hours |
| Consistency | Same criteria applied to every line | Varies between people and with fatigue |
| Unusual targets | May miss classes it was not trained on | A skilled analyst can recognise the unexpected |
| Best practice | Use it to find candidates, then review | Keep it for verification and difficult cases |
Frequently asked questions
What can the model detect?
It depends on the model and the data. Typical classes include pipelines, cables, boulders, debris, wrecks and anchors. For rare or site-specific targets, the model can be retrained on labelled examples from your own surveys.
How accurate is it?
Accuracy depends on sonar frequency, range, seabed type and how similar your data is to the training data. We recommend testing on a sample of your own data and measuring how many real targets are found and how many false alarms are raised.
Does it replace the survey analyst?
No. It reduces the time spent scanning data and makes results more consistent, but every candidate should be reviewed by a person, and safety-critical contacts need verification beyond sonar.
Can I try it myself?
Darkocean offers an online object detection tool at detect.darkocean.ai. Get in touch if you want to discuss using it on your own datasets.
Got sonar data waiting to be reviewed?
Share a sample and we will show you what automated detection finds in it.
