Machine Learning for Shipwreck Detection: A Computer Vision Approach in Maritime Archaeology
0Cal T. Pols
The efficacy of machine learning in archaeology has been well established by many terrestrial studies; however, these methods remain under-utilised in maritime archaeological research, especially in relation to underwater site detection. The majority of previous machine learning studies in maritime contexts have primarily used either side-scan sonar data (SSS) or aerial/satellite imagery as the basis for object detection. This research aimed to assess the performance and potential of machine learning methods for the detection of shipwreck sites in open-access bathymetry data from the United Kingdom (UK). Manual identification was completed to act as a baseline detection level and used to evaluate the performance of a pre-trained shipwreck detection model (esri) and two custom detection models. The custom models were trained to detect shipwrecks using high-resolution (1m) bathymetry data and using hillshade visualisation, based on a Single Shot Detector (SSD) model type with a ResNet50 backbone. Both detection models achieved a high average precision score (0.77) from a training dataset of known shipwreck instances, which used data augmentation to effectively increase the number of training samples.
The shipwreck detection models created for this research highlight the feasibility of implementing machine learning workflows into maritime archaeological research. As the challenges of Big Data begin applying to marine survey, manual analysis of these datasets seems to be increasingly untenable. For the soon-to-be reality of global seabed mapping, machine learning methods offer effective, systematic, and efficient ways to analyse vast datasets for a variety of purposes including archaeological potential. Going forward, these methods can also help develop underwater cultural heritage management strategies and aid field-based investigations through site identification, classification, and initial assessment.