2023_programme: Application of a Convolutional Neural Network trained with simulated low-frequency Synthetic Aperture Sonar data for classification of buried UXO

• Session: 15. Towards Automatic Target Recognition. Detection, Classification and Modelling
  Organiser(s): Johannes Groen, Yan Pailhas, Roy Edgar Hansen, Jessica Topple and Narada Warakagoda
• Lecture: Application of a Convolutional Neural Network trained with simulated low-frequency Synthetic Aperture Sonar data for classification of buried UXO
  Paper ID: 1892
  Author(s): van de Sande Jeroen, Pereboom-Huizinga Wyke, den Hollander Richard, van der Burg Dennis, Mulders Ingrid, van Vossen Robbert
  Presenter: van de Sande Jeroen
  Abstract: The application of Artificial Intelligence to target classification problems typically requires substantial amounts of training data. Applications in the underwater domain often suffer from a lack of experimental target data. In addition, operational conditions that affect the target response may be uncertain or unknown. In this paper, a method is described and evaluated that explores the use of Target-In-Environment-Response (TIER) simulations to generate sufficient amounts of training data for the discrimination of proud or (partially) buried Unexploded Ordnance and clutter, using low-frequency Synthetic Aperture Sonar (SAS). As a use case the attempt is to discriminate two types of cylindrical dummy targets from clutter. The TIERs of the dummy targets are simulated for a variety of burial depths, orientations and sediment types. Together with field recordings of clutter, they are used for training a Convolutional Neural Network. The trained network is evaluated against discriminating synthetic target data and clutter as well as real target data and clutter, the latter both recorded in field experiments using TNO’s MUD system. The use of two SAS data representations has been analyzed: conventional time-domain SAS images and Multi-Aspect Acoustic Color (MAAC) images. The Receiver Operating Curves (ROC) of the CNN for evaluation against an independent set of synthetic data shows an average Area-Under-Curve of 0.99 for both data types. Application on real data shows a performance reduction for SAS images to an AUC of 0.86, which is considered reasonable. For MAAC images, however, the performance breaks down to an AUC of 0.64. A mismatch between target model and actual target, in conjunction with CNN overfitting, is suspected to be the main reason for this performance drop. Suggestions are made to extend the training set with simulations that include several variations in the target models, in order to reduce the sensitivity to the precise target properties.
    Download the full paper
• Corresponding author: Mr Jeroen van de Sande
  Affiliation: TNO
  Country: Netherlands
  e-mail: