We investigate how a self-assessment capability can be obtained and exploited considering a typical use case in which limited training data is available, namely target recognition for buried objects in a marine environment in low-frequency sonar imagery.

Problem Context

To resolve the problem that limited target data is available for target recognition for buried objects in a marine environment, the usage of augmented simulated targets will be investigated. This project can be considered as a feasibility study aiming to address challenges encountered in many military use cases. In general, the industry does not provide these solutions, and TNO aims to fill this niche by providing tailored solutions to the Netherlands Ministry of Defence once concept solutions have been demonstrated.


The project uses data collected by TNO’s Mine Underground Detection (MUD) system, which has been developed to detect and localize buried objects in the seabed in inshore environments, such as harbors. It uses an interferometric low-frequency-synthetic aperture sonar (SAS) as primary sensor to detect, localize, and classify buried objects. To aid the training, relevant targets will also be simulated using a high-fidelity target simulation capability. A data augmentation strategy will subsequently be developed to augment simulated targets in measured data.


The project provided new knowledge on the application of state-of-the-art machine learning techniques to problems for which limited training data are available; The project also provided new knowledge on the application of these techniques to sonar data (time series data, low-frequency synthetic aperture sonar (LF-SAS)), which poses new challenges compared to optical camera imagery.


  • Robbert van Vossen, Senior Research Scientist, TNO, e-mail: