Industries / Transport and vehicles / Maritime and shipping / Marine salvage
Surveys
Surveys find a wreck or a ship in trouble and measure where it lies, what shape it is in and what it carries, so that salvors, owners and governments can decide what to do with it.
6
Milestones achieved, from 1960s to 2024.
2
To watch for: announced, not yet confirmed outside.
3
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1960s
Sonar draws the seabed
Side-scan sonar is towed behind a ship and sends pulses of sound out to both sides; the echoes build a picture of the seabed and anything lying on it, much like a photograph made of sound. The MIT Museum says Martin Klein "developed the first commercially viable sidescan sonar system in the 1960s", working at Edgerton, Germeshausen & Grier (EG&G), the firm of his MIT teacher Harold "Doc" Edgerton. Klein left EG&G in 1969 to found Klein Associates, which made a series of side-scan sonars. MIT's alumni magazine credits his sonars with detecting the wreckage of the space shuttle Challenger and of TWA Flight 800.
Credit
Martin Klein; EG&G; Harold Edgerton
1985
Titanic found by its debris
Woods Hole Oceanographic Institution (WHOI) in the US and IFREMER, France's ocean research institute, narrowed the search for the Titanic, lost in 1912, to 100 square miles. From 1 July 1985 the French ship Le Suroit swept it in overlapping lines with SAR, a deep-towed side-scan sonar built by Jean-Louis Michel of IFREMER, and cleared more than 75% of the area without finding the wreck. With 12 days of ship time left, the WHOI team on the research vessel Knorr stopped looking for the hull and looked instead for the trail of debris that currents had spread from it, using Argo, a new towed sonar and video sled. Just after 1:00 a.m. on 1 September 1985 one of the Titanic's boilers came into view under more than 12,400 feet (about 3,800 m) of water. WHOI notes that both Argo and SAR were on sea trials and that finding the Titanic was a secondary aim.
Lessona long trail of debris is easier to cross than a single hull, so a search can aim at what a wreck leaves behind.
Credit
Robert Ballard (WHOI); Jean-Louis Michel (IFREMER)
1988
Probability finds a gold ship
The SS Central America sank in a hurricane on 12 September 1857 with about 425 people and a cargo of gold, and nobody knew where she lay. In 1985 Lawrence D. Stone of Metron, Inc., an operations researcher (operations research uses mathematics to plan decisions), was asked to build a probability map of her position. He combined accounts from 1857, statistics, calculation and expert judgment, and the search plan drawn from his map found the wreck, as he reported in 1992. A US appeals court records that the Columbus-America Discovery Group located her in 1988, 8,000 feet (about 2,400 m) down and 160 miles off South Carolina.
Credit
Lawrence D. Stone (Metron); Columbus-America Discovery Group
2022
Endurance found under ice
Sir Ernest Shackleton's ship Endurance was crushed by pack ice and sank in the Weddell Sea, off Antarctica, in 1915. On 9 March 2022 the Falklands Maritime Heritage Trust said its Endurance22 expedition had found the wreck at a depth of 3,008 m, inside the search area it had set before leaving Cape Town and about four miles south of the position recorded by her captain, Frank Worsley. The team worked from the South African vessel S.A. Agulhas II with Saab's Sabertooth underwater vehicles. The wreck is protected under the Antarctic Treaty, so it was filmed and surveyed but not touched.
Lessonnotes made at the time of a sinking can guide a search a century later; the expedition called Worsley's navigation records "invaluable".
Credit
Falklands Maritime Heritage Trust; John Shears, Mensun Bound and Nico Vincent; Saab
2022
Wrecks named from their shapes
Multibeam sonar sends a fan of sound beams downward and maps depth across a wide strip, so it can measure a wreck's length, width and shape in three dimensions. Innes McCartney of Bournemouth University studied multibeam surveys of all 273 charted wrecks in a study area in the Irish Sea, made from Bangor University's survey ship Prince Madog. Records listed 101 (37%) of the wrecks as unidentified; once each site was assessed the true number was 161 (59%), mainly because of wrong identifications dating from the 1970s. By matching each wreck's measured size and position to archival accounts of losses, he named 129 of the 161, leaving 12% of the area's wrecks unknown, and the study lists 110 amendments to the UK Hydrographic Office's wreck records.
Lessona record that names a wreck wrongly hides it better than no record at all.
Credit
Innes McCartney (Bournemouth University), with Bangor University's School of Ocean Sciences; funded by the Leverhulme Trust
2024
An open wreck sonar set
Teaching software to spot wrecks in sonar needs many labelled examples, and wrecks are rare on the seabed. A team led by Advaith Sethuraman surveyed Thunder Bay National Marine Sanctuary in Lake Huron, US, with an autonomous underwater vehicle (AUV), a robot submarine that runs a survey without a pilot. It released AI4Shipwrecks: 286 side-scan sonar images of 28 wrecks, labelled pixel by pixel with help from marine archaeologists, and benchmark results for segmentation methods (segmentation marks each pixel of an image as wreck or not wreck). The data sit in the University of Michigan's Deep Blue repository.
Credit
Advaith Sethuraman and six co-authors; NOAA's Thunder Bay National Marine Sanctuary
Watch for
A free wreck finder
ShipwreckFinder is an open-source plug-in for QGIS, the free mapping program. It takes multibeam depth data and marks likely wrecks with a deep-learning model trained on wrecks from the Great Lakes and the coasts of Ireland, plus synthetic examples. Its authors report better segmentation than a deep-learning toolkit for ArcGIS and a classical method borrowed from sinkhole detection. The work was a NOAA Ocean Exploration project (July 2023 to June 2025), whose page says the model was tested on known sites such as the wreck of the William P. Rend in Thunder Bay.
Achieved whensomeone outside the team runs it on surveys from a coast it was not trained on and publishes how many charted wrecks it found and how many false alarms it raised.
Claimed by
Anja Sheppard and seven co-authors, September 2025 (accepted to the OCEANS 2025 Great Lakes conference)
Wreck spotting aboard the robot
An AUV can send only small amounts of data to its ship through the water, so a model that spots wrecks on board could report finds or change its route during a dive. Agno Rubim de Assis of the University of Southern Mississippi and colleagues at the University of Mississippi tested six setups of YOLO, an open family of object-detection models from Ultralytics, on public high-resolution depth data from the coast of England that hold more than 630 wrecks. They report precision, recall and F1 above 0.91 for the best model (precision is the share of flagged objects that are real wrecks, recall the share of wrecks that were flagged, F1 a combination of the two), and say it found the wreck in a small deep-water AUV data set from the Gulf of America.
Achieved whenthe model runs on board an AUV during a real survey and its finds and misses are reported against a later check of the same ground.
Claimed by
de Assis, Guilment, D'Emidio and Macelloni, Journal of Marine Science and Engineering, 9 July 2026
Open
01
Name unknown wrecks with AI
A tool that proposes a name for each unidentified wreck by matching its measured length, width and position against historical records of ships lost in the area, scored first on McCartney's Irish Sea study area, where the answers are known.
The briefCredit · unclaimed02
Replay famous wreck searches
An open search planner built on Bayesian search theory, a method that keeps a map of the probability that the target lies in each cell of the sea and lowers it in each cell searched without success, and a public replay of three searches whose details are published: the Titanic (1985), the SS Central America (to 1988) and Endurance (2022). For each, compare how much ground the planner would have covered before reaching the wreck with what the real search covered.
The briefCredit · unclaimed03
Wreck finders on unseen coasts
An outside test of the public wreck-detection models on multibeam surveys from a coast none of them was trained on, scored against the wrecks charted there.
The briefCredit · unclaimed
Open 01 · the brief
Name unknown wrecks with AI
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A tool that proposes a name for each unidentified wreck by matching its measured length, width and position against historical records of ships lost in the area, scored first on McCartney's Irish Sea study area, where the answers are known.
Why it is still open
McCartney's identification of 129 of 161 unknown wrecks was done by hand, wreck by wreck, from 3D models and historical texts. Detection tools such as ShipwreckFinder find a wreck in depth data but do not say which ship it is. This research found no published tool that does the naming step and is scored against known answers.
Built with what exists
McCartney's open-access book as the answer key; the loss records he cites; a large language model such as Claude to turn written loss accounts into a table of ship, date, size and reported position, with each row checked by a person; Python with GeoPandas to score candidates by distance and size. Desk work only: many of these wrecks are war graves and none should be visited.
First step
Type up the measured sizes and the identifications for 30 wrecks from the book, and test whether size and position alone put the right ship in the top three.
Done looks like
For the 161 wrecks that were unknown, the share for which McCartney's identification is among the tool's top three candidates, published with every miss and the code; then a run on one other charted area, with every proposal marked unconfirmed.
Open 02 · the brief
Replay famous wreck searches
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open search planner built on Bayesian search theory, a method that keeps a map of the probability that the target lies in each cell of the sea and lowers it in each cell searched without success, and a public replay of three searches whose details are published: the Titanic (1985), the SS Central America (to 1988) and Endurance (2022). For each, compare how much ground the planner would have covered before reaching the wreck with what the real search covered.
Why it is still open
The method behind the Central America search is described in a 1992 journal paper, and the same approach has been written up for aircraft wreckage: by Stone and colleagues for Air France 447 (Statistical Science, 2014) and by Davey and colleagues of Australia's Defence Science and Technology Group for MH370 (2016). These are accounts by the searchers themselves, not open replays that someone else can run. A search of GitHub's public repositories in November 2026 returned 17 repositories for "bayesian search theory", mostly games and teaching examples such as Battleship players, plus one tool for fishing vessels that switch off their position transmitters, and none that replays a wreck search.
api.crossref.orgapi.crossref.orgapi.crossref.orgapi.github.com
Built with what exists
Python with NumPy; WHOI's account of the 100-square-mile Titanic search box swept in lines about 800 m apart; the Endurance22 account of a search area set before departure and Worsley's recorded position; Stone's paper for the Central America.
First step
Rebuild the 1985 Titanic search box and line spacing from WHOI's account and run a first planner over it.
Done looks like
Code and probability maps published for the three searches; for each, the share of the real search area the planner would have swept before reaching the wreck's true position, with every input and assumption stated.
Open 03 · the brief
Wreck finders on unseen coasts
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An outside test of the public wreck-detection models on multibeam surveys from a coast none of them was trained on, scored against the wrecks charted there.
Why it is still open
ShipwreckFinder was trained on Great Lakes and Irish data; de Assis and colleagues trained on English data and tested on one small Gulf data set; NOAA's project page recommends training on a variety of survey methods, sensors and data quality. No independent comparison of these tools on the same unseen data turned up in this research.
Built with what exists
ShipwreckFinder in QGIS; Ultralytics YOLO; public multibeam surveys from NOAA's National Centers for Environmental Information (NCEI), which NOAA's project also used; the wrecks marked on NOAA's electronic navigational charts as a first answer key, checked by eye.
First step
Download NCEI multibeam for one well-surveyed US harbour approach and list the charted wrecks inside its coverage.
Done looks like
At least two models scored on at least 100 charted wrecks in one region outside their training areas, with recall, false alarms per square kilometre and a list of charted wrecks that do not appear in the data, all published.