Underwater shipwreck covered in algae with a scuba diver in the background.

Software Tools to Enable Automated Detection of Submerged Archaeological Sites From Multibeam Sonar

Past Expedition

Project Summary

Using artificial intelligence, roboticists have developed a new way to virtually detect and discover submerged archaeological sites — like shipwrecks. Their approach relies on a machine learning model to automate the search for such sites in multibeam sonar data collected by surface vessels. It limits the need for human involvement, reducing the time, effort, and cost required to detect submerged archaeological sites and enables discoveries to be made and shared with the maritime heritage community and the public.

The model is specifically designed to detect and segment shipwreck sites in multibeam sonar images:

Detect: Draw a bounding box label around a region of interest (e.g., a shipwreck site) in a sonar image.
Segment: Label each pixel in the sonar image as “shipwreck” or “not shipwreck” to output pixelwise segmentation masks.

Sonar and Segmentation — Multibeam sonar image (left) and segmentation (right) of the *William P. Rend* shipwreck site in Thunder Bay National Marine Sanctuary used to test the model created as part of the Software Tools to Enable Automated Detection of Submerged Archaeological Sites From Multibeam Sonar project. The large circles are zoomed-in insets of the site. *Image courtesy of Software Tools to Enable Automated Detection of Submerged Archaeological Sites From Multibeam Sonar (using data from the NOAA National Centers for Environmental Information).*
Download largest version (png, 201 KB).

The model is the core of a QGIS toolkit (QGIS is an open-source library) designed to increase the efficiency of processing ocean exploration data. The toolkit will be easily accessible for practical use by marine archaeologists and other users interested in shipwreck detection.

Model Flowchart — Flowchart of the model developed to detect shipwrecks in multibeam data, part of a QGIS toolkit designed during the Software Tools to Enable Automated Detection of Submerged Archaeological Sites From Multibeam Sonar project. *Image courtesy of Software Tools to Enable Automated Detection of Submerged Archaeological Sites From Multibeam Sonar (using data from the NOAA National Centers for Environmental Information).*
Download largest version (jpg, 1.13 MB).

Training and Validation

To train and validate the model, the project team used open-source sonar images from Thunder Bay National Marine Sanctuary (TBNMS) in Michigan, part of Lake Huron and the resting place for hundreds of known and yet to be located shipwrecks. TBNMS is an ideal testbed for developing tools like this one. Data included multibeam sonar images collected during three ship-based hydrographic surveys conducted by NOAA’s Office of Coast Survey and bathymetry data derived from side-scan sonar images collected using an IVER 3 autonomous underwater vehicle as part of a previous NOAA Ocean Exploration supported project. To supplement these data, they created synthetic multibeam images and segmentation masks of shipwreck sites by superimposing shipwreck sites onto real terrain images.

SS Monrovia — Side-scan sonar image of the wreck of SS *Monrovia*, part of the AI4Shipwrecks Dataset used to train and validate the model developed for the Software Tools to Enable Automated Detection of Submerged Archaeological Sites From Multibeam Sonar project. *Image courtesy of Machine Learning for Automated Detection of Shipwreck Sites from Large Area Robotic Surveys.*
Download largest version (png, 3 MB).

Background

Recent advances in sensor technology and marine robotic systems have resulted in massive amounts of multibeam sonar data. However, finding submerged objects, like shipwrecks, within this data is time consuming and typically requires expert knowledge. Concurrent advances in machine learning have demonstrated its utility for visual tasks, like object detection in large datasets, offering potential solutions for processing increasing amounts of data.

Takeaways and Recommendations

By testing their model with sonar images from known shipwreck sites, the team demonstrated that a model trained with real and simulated data can return accurate results. But it wasn’t all smooth sailing, and they identified a number of recommendations for future users of their model, including:

Incorporate a variety of data (i.e., variations in survey methodologies and sensors used, data quality, the presence/absence of shipwrecks, and the shape and structure of shipwrecks).
Preprocess the data for machine learning input (e.g., cropping, smoothing).
Augment the data (e.g., random scaling, translation, rotation).
Train the model to learn how to handle gaps in data coverage.

What’s Next

In the Great Lakes, this toolkit will be demonstrated on data from Thunder Bay National Marine Sanctuary to search for shipwrecks and other archaeological sites in its waters. But the approach is also applicable in the deep ocean.

At the OCEANS 2025 Great Lakes conference, the project team will present a tutorial on their work. During “AI4Shipwrecks: Artificial Intelligence and Machine Learning for Shipwreck Detection,” they will provide an overview of the AI4Shipwrecks dataset, a new side-scan sonar dataset featuring 28 distinct shipwreck sites in Thunder Bay National Marine Sanctuary, and demonstrate their approach for automated detection of submerged archaeological sites from multibeam sonar data.

Maritime Heritage

Seafloor Mapping

Partners

University of Michigan Field Robotics Group (FROG) (Principal Investigator: Katie Skinner)
University of Michigan Robotics
University of Michigan Naval Architecture & Marine Engineering

Media Contacts

- Emily Crum
  Communication Specialist
  NOAA Ocean Exploration
  ocean-explore-comms@noaa.gov