2025_programme: Change Detection in Synthetic Aperture Sonar Imagery Using Segment Anything Model

• Day: June 19, Thursday
    Location / Time: B. ERATO at 17:40-18:00
• Last minutes changes: -
• Session: 01. Acoustics in Ocean Observation Systems
  Organiser(s): Jaroslaw Tegowski, Philippe Blondel, Hanne Sagen
  Chairperson(s): Jaroslaw Tegowski, Philippe Blondel
• Lecture: Change Detection in Synthetic Aperture Sonar Imagery Using Segment Anything Model
  Paper ID: 2099
  Author(s): William Hedlund, Per Abrahamsson, Louise Fuchs
  Presenter: William Hedlund
  Abstract: Capabilities for marine research and monitoring of coastal waters have improved with advancements in Autonomous Underwater Vehicles and Synthetic Aperture Sonar (SAS) imagery. These developments have increased the demand for tools to efficiently process information, reducing the workload for operators. In other remote sensing fields, deep learning methods have been applied successfully for similar tasks. However, due to the limited availability of SAS data, deep learning in SAS has not been widely researched. Recent advancements have led to foundation models, which are large general-purpose AI models, with capabilities across unseen tasks and domains. These models could be used to bridge the gap in deep learning for change detection in SAS. \n\nThis paper uses Segment Anything Model (SAM) to investigate foundation models in change detection for SAS imagery. The method applies SAM to segment bi-temporal SAS images, which are then compared to identify changes. Three prompting methods are evaluated: two employ grids of points with varying spacings, and the third employs a traditional log-ratio difference-image to prompt the model. For evaluation, the method is tested on both a real dataset and a synthetic dataset generated through image composition to simulate changes. \n\nComparisons on the synthetic data show that methods using a difference-image and a smaller grid spacing demonstrate similar performance in precision and recall. The larger grid spacing yields more missed detections, with the trade-off of higher precision. When benchmarked against a traditional log-ratio method on the real dataset, the proposed method demonstrates improved precision and recall, reducing false alarms caused by misregistration and environmental noise. \n\nIn conclusion, SAM’s adaptability to new domains and tasks presents a viable alternative for SAS change detection, addressing data limitations and opening new avenues for deep learning applications in marine research and environmental monitoring.
    Download the full paper
  This paper is a candidate for the "Prof. John Papadakis award for the best paper presented by a young acoustician(under 40)"
• Corresponding author: Mr William Hedlund
  Affiliation: SAAB Kockums
  Country: Sweden