Shipwrecks are valuable sources of archaeological information, but detecting them underwater is not always feasible. Optical cameras are often limited in turbid water; therefore, side-scan sonar (SSS) is an important and widely used tool for surveying the seafloor.
However, sonar images are not without limitations. They can be affected by noise, speckle, and acoustic interference, which may obscure important features. As a result, distinguishing shipwrecks from natural seabed features, such as rocks and sand ripples, can be difficult and time-consuming.
A team from the Harbin Institute of Technology developed a deep learning model called SW-Net, specifically designed to improve shipwreck detection in complex sonar images
SW-Net’s approach
The researchers decided to start with the fact that shipwrecks are artificial objects, so they often have straight edges, sharp corners, and regular geometric shapes that are less common in natural seabed features.
Most AI image models process features in the same way regardless of direction, which can cause the model to treat a clear straight edge from a shipwreck similarly to an irregular patch of noise.
SW-Net uses a different approach. It applies fixed mathematical filters designed to detect edges at specific angles in the sonar image.
Instead of learning these filters during training, the model uses predefined directions, while a directional attention mechanism identifies the most important angles in each part of the seafloor.
This allows the model to focus on meaningful structures and helps distinguish the regular shapes of shipwrecks from noise and random seabed patterns.
Testing SW-Net in field
The researchers evaluated SW-Net against seven established segmentation models using a public dataset containing real shipwreck sonar images. SW-Net achieved the best results in the comparison, obtaining both the highest overlap accuracy and the highest F1 score.
The model also showed strong computational efficiency. With only about 4 million parameters, SW-Net was significantly smaller than the other models tested. It can also process sonar images at approximately 45 frames per second on a single GPU, making real-time use possible.
This could allow an autonomous underwater vehicle to identify and flag potential shipwrecks during a survey, without relying on later processing on a separate computer.
Where the model falls off
The researchers also identified two main limitations of the model. First, SW-Net can sometimes mistake highly reflective rock formations for shipwreck debris, resulting in false positives.
Second, it may fail to detect shipwrecks that are heavily covered by sediment or severely fragmented. In these cases, too little of the original structure may remain for the directional filters to identify a clear pattern, leading to false negatives.
To reduce false positives, the researchers suggest combining sonar data with information from other types of sensors. To reduce false negatives, they suggest using multiple sonar scans from different angles or at different times, which could improve detection, especially for shipwrecks that heavy sediment buries or obscures.
Conclusion
SW-Net offers a more efficient approach to searching for shipwrecks. Rather than manually reviewing large amounts of sonar data after a survey, a system like SW-Net could eventually identify potential wrecks as the data is collected. This would allow survey teams to concentrate on the most promising areas instead of reviewing every sonar image manually.
However, SW-Net is not yet publicly available. Although the paper provides detailed information about the model and its results, the code and trained model have not been released. This means that other researchers cannot currently download and test the system themselves.
Dai, J., & He, J. “SW-Net: A Direction-Aware Deep Learning Model for Shipwreck Segmentation in Side-Scan Sonar Imagery.” Sensors, Vol. 26, Issue 11, Article 3483 (2026). doi.org/10.3390/s26113483
