Underwater unexploded ordnance (UXO) are a hazard to the marine environment and human offshore activities; their remediation presents a worldwide challenge. Autonomous underwater vehicles (AUVs) equipped with high-resolution seafloor imaging sonar are suited to this task. AUVs permit the safe and efficient surveying of UXO sites and synthetic aperture sonar (SAS) provides the image resolution necessary to accurately map and monitor these sites. However, the effectiveness of these tools relies on robust autonomous detection and classification algorithms.

While good progress is being made in the application of machine learning to this problem, the scarcity of labelled training data limits the robustness of these solutions under complex and uncertain operating conditions and in novel environments. Data simulation is a cost-effective mechanism for addressing this issue. The over-arching objective of this project is to work toward a simulation capability that provides a good balance between realism and computational speed for this purpose.

A new model is introduced for the rapid simulation of raw coherent SAS data. It is an evolution of the facet diffraction model and it offers an equivalent degree of realism, but at much greater computational speed. Crucially, it captures the coherent wave physics properly and thus enables end-to-end integration within the SAS processing pipeline, including micro-navigation, imaging, interferometry, etc.

Instead of modelling the scattering and propagation from facets (or similar primitives) individually and then using computer graphics (CG) methods (e.g., ray-tracing) to inform the superposition of their contributions to the wavefield, this order is reversed. The aspect-dependent scattering from the scene is rendered directly using CG methods. Then, Fourier-domain methods are used to model propagation of the wavefield and subsequent data acquisition. It is this novel combination and re-ordering of CG and Fourier methods that enables the significant enhancement in speed.

This new approach has the potential to increase the speed of SAS data modelling by orders of magnitude, possibly enabling the simulation of realistic data from realistic scenes in real-time or faster. The outcomes from this project aim to prove the concept and will form the basis for a follow-on CORE proposal to create a fully capable three-dimensional simulator.

This envisaged end-product has value in the augmentation of training data for developing robust machine learning algorithms to detect and classify munitions under complex conditions (e.g., caused by environmental factors, such as: navigation difficulties; seafloor clutter, reverberation, and multipath; target corrosion and burial; etc.). It also offers the opportunity for integration with other SERDP simulation capabilities as a tool for performance estimation and planning.