A study in the journal Remote Sensing shows the implementation of a deep learning model that uses digital elevation data to detect shipwrecks automatically and rapidly over a large geographic area.
The paper, “Archaeologic Machine Learning for Shipwreck Detection Using Lidar and Sonar,” is authored by Leila Character, Agustin Ortiz Jr, Tim Beach, and Sheryl Luzzadder-Beach. It is intended to apply a new methodology to the field of underwater archaeology.
Shipwrecks represent a major resource to understand maritime human activity over millennia, but underwater archaeology is expensive, misappropriated, and hazardous.
An automated tool to rapidly detect and map shipwrecks can therefore be used to create more accurate maps of natural and archaeological features to aid management objectives, study patterns across the landscape, and find new features. Additionally, more comprehensive and accurate shipwreck maps can help to prioritize site selection and plan excavation.
The model is based on open source topo-bathymetric data and shipwreck data for the United States available from NOAA. The model uses transfer learning to compensate for a relatively small sample size and addresses a recurring problem that associated work has had with false positives by training the model both on shipwrecks and background topography.
Results of statistical analyses conducted—ANOVAs and box and whisker plots—indicate that there are substantial differences between the morphologic characteristics that define shipwrecks vs. background topography, supporting this approach to addressing false positives. The model uses a YOLOv3 architecture and produced an F1 score of 0.92 and a precision score of 0.90, indicating that the approach taken herein to address false positives was successful.
Deep learning and open-source high-resolution bathymetric data used
This project’s objective was to determine whether deep learning and open-source high-resolution bathymetric data could be used to accurately predict the locations of shipwrecks over a large geographic area. This work is intended to introduce a new methodology to the field of underwater archaeology that can be used in conjunction with existing methodologies.
A methodology to automatically map all potential shipwrecks over a large geographic area accurately and rapidly can help archaeologists prioritize site selection and plan excavation. The model implementation presented accurately detects shipwrecks in remotely sensed imagery collected from large areas of the coast of the continental United States, Alaska, and Puerto Rico. The methodological approach can easily be replicated in other locations around the world.
Machine learning applications in archaeology have made significant progress in predictive accuracy and utility in the last decade, with work over the last four years focused on the application of a specific type of machine learning algorithm called deep learning.
Deep learning is a type of machine learning that can identify features of interest in remotely sensed imagery by recognizing the unique visual patterns by which they are represented. It is particularly powerful for detecting archaeological features because of its ability to identify many different morphologies and orientations of the same features.
Applications of deep learning to underwater archaeology are very limited because of the paucity of high-resolution bathymetric data as compared to land-based elevation data. The new model implementation presented here uses high-resolution open-source bathymetric data accessed through the NOAA Office of Coastal Management Data Access Viewer, as well as the GPS locations of confirmed shipwrecks from NOAA’s Office of Coast Survey’s Automated Wreck and Obstruction Information System (AWOIS) database.
Part of NOAA’s mission is to document the seafloor to ensure natural resources are protected and waterways are navigable for mariners. This includes a database of more than 10,000 shipwrecks and underwater obstructions. This model will be used by the Navy’s Underwater Archaeology Branch to find unmapped or unknown naval shipwrecks to aid management objectives by creating more accurate and complete maps of shipwreck locations and by studying shipwreck patterns across the underwater landscape.
This work seeks to make machine learning methods applicable and relevant to archaeologists and others interested in studying, managing, and conserving the maritime landscape.
Read the study in its entirety here.
