Unexploded ordnance (UXO) in cluttered, occluded, and buried environments present a major environmental and security risk for munitions response. Synthetic aperture sonar technology presents an opportunity to mitigate these risks but suffer from reduced fidelity and lowered recognition for potential targets in these challenging environmental conditions. This project developed a suite of algorithms to enhance synthetic aperture sonar beamforming for automated target recognition for buried UXO detection. The primary objectives for this project were to merge new advances in physics-based machine learning with three-dimension synthetic aperture sonar to achieve higher resolution, improved imagery, and better task performance in evaluation.

This project introduces two primary innovations to advance the project’s objectives: (1) a new differentiable beamforming method based on neural rendering for volumetric synthetic aperture sonar reconstruction, and (2) a machine learning based tone mapper for improving automated target recognition for buried UXO detection. Both methods were developed using state-of-the-art advances in machine learning and computational imaging and were evaluated on exemplar data from a sub-bottom synthetic aperture sonar of various man-made objects distributed on a lakebed.

Differentiable beamforming resulted in improved three-dimensional synthetic aperture sonar volumetric reconstructions including better visualizations of targets relative to clutter for a circular in-air system as well as a bistatic water-based system. Further, machine learning based tone mapping was shown to boost performance in both precision and recall for automated target recognition for partially and fully buried objects.

Benefits to the Department of Defense and the scientific community include new knowledge of how to merge the physics of underwater acoustics and beamforming with machine learning. New algorithms and methodologies have been developed and shown via experimental demonstrations as proof-of-concept. As shown in the Figure below, these algorithms can help synthetic aperture sonar systems better visualize and recognize three-dimensional targets on the seafloor.

Figure: Differentiable beamforming for 3D synthetic aperture sonar - The proposed algorithms enable better target visualization as compared to background/clutter in shallow water.