Shipwreck detection in bathymetry data using semi-automated methods: Combining machine learning and topographic inference approaches

This research presents a workflow that integrates emerging machine learning methods with geospatial mapping techniques to improve the identification of shipwrecks in bathymetry data. By first refining the study area into high-potential units, machine learning algorithms can be applied more efficiently. This approach accelerates the process, reduces computational demands, and offers an adaptive method that can eventually be tailored to survey needs and different seabed environments. This paper contributes to the current discourse surrounding the discovery and management of underwater cultural heritage (UCH) in the context of global seabed mapping, developments in autonomous marine survey, and continued offshore development. Shipwrecks constitute a significant proportion of UCH sites that are increasingly likely to be discovered and impacted by these developments, and thus archaeologists need adequate tools for their rapid detection and monitoring to keep pace with the rate of data generation. The proposed workflow uses a raster extraction method as a filtering process to identify areas of seabed with high shipwreck potential, based on their topographic signature in three different visualisations of bathymetry (slope, curvature, and topographic position index). Using these results, several different machine learning algorithms were tested on their ability to identify both intact, visible shipwrecks (‘conspicuous’ wrecks) as well as smaller, possible wreck sites. These methods were tested over an area of 3,131 km 2 from the south coast of England. Results show that the Raster Extraction method was able to filter out 96% of the test data, while still detecting 78% of the test shipwrecks (n=253). Machine learning models trained on different data visualisations (Hillshade, Shaded Relief, Curvature) and algorithms (Single Shot Detector, Faster R-CNN, and Mask R-CNN) had varied performances in terms of recall and precision. • A new semi-automated workflow was developed to detect shipwrecks in bathymetry. • This method combines geospatial mapping tools with machine learning detection. • Raster Extraction filtered 96% of the original data and detected 78% of shipwrecks. • Seven ML models were tested: SSD, FRCNN, MRCNN using different data visualisations. • A Curvature SSD model achieved the best performance (F1 = 0.57) on all shipwrecks.