Physics-based data augmentation for high frequency 3D radar systems

The detection of side-attack explosive hazards remains challenging due to the significant variation in size, shape, construction materials, and placement on or above the surface. Some of the most challenging-to-detect side-attack explosive hazards are those placed inside of naturally occurring clutter such as vegetation. High-frequency radar systems with 3D resolution have been observed to be an effective technology for detecting and discriminating surface-laid sideattack explosive hazards from both natural and manmade clutter. Automated target recognition on the 3D voxel radar data is a complex problem that is well suited for deep convolutional neural networks. The main drawback of such approaches is the requirement for a large amount of training data, which is expensive and time-consuming to collect. Ad hoc and generative models have been used to augment data for deep learning with some degree of success; however, these methods often generate examples closely resembling instances from the training data, and any deviations are potentially not physically realistic for the sensing phenomenology. More accurate and effective augmentation can be accomplished by leveraging sensor physics along with large amounts of readily available background data. Observations of target signatures under clutter-free conditions can be inserted into a cluttered scene in a way consistent with the physics governing the sensor. We show that our physics-based data augmentation technique yields realistic synthetic data that is useful for augmenting the available training data and leads to improved discrimination performance.