One answer is to generate simulated, or synthetic, SAR data to train the machine learning (ML) models. In June 2021, an Air Force client tasked Booz Allen with developing simulated SAR images of specific ground vehicle targets and training ML models to recognize and identify those targets.
Simulating data from remote sensors, such as SAR, is much more difficult than simulating data from optical sensors, such as a camera. Optical sensors collect data in the visible, near-infrared, and short-wave infrared portions of the electromagnetic spectrum. In contrast, remote sensors, such as those used by SAR, utilize longer wavelengths of the spectrum—often referred to as bands—which give them special properties, such as the ability to see through clouds or forests.
Generating simulated data of this type and then training machine learning models to analyze the data require highly specialized expertise, tools, and resources that are not widely held. However, Booz Allen has advanced capabilities in all of those areas.
Within 5 months, the Booz Allen team developed a solution that generated the needed quantities of simulated images and used that data to train algorithms that could be used for an end-to-end automated target recognition system.
Moreover, the Booz Allen team did this using only 120 simulated SAR images for each target class—one image for every three degrees of a 360-degree view—which is a small fraction of the hundreds or thousands of images that are typically employed to train an automated target-recognition model.
To develop this SAR simulation solution, Booz Allen brought together its industry-leading expertise in artificial intelligence (AI) and experience in Air Force mission areas. Using computer-aided design models of the relevant ground targets, the Booz Allen team simulated radiating those ground vehicles with SAR radar waves from many angles. This produced the simulated SAR images of the targets.
The team then supplemented those simulated images with real SAR images of the ground targets taken from an unclassified internet-accessible repository known as MSTAR. Finally, the team tested various combinations of real and synthetic SAR data on deep learning neural networks developed for radar data by the University of California at Berkeley.