SBIR-STTR Award

Machine learning algorithms have demonstrated performance on par with or superior to human analysts on large datasets when sufficient training data is available. For many military applications, sufficient truthed data over relevant operating conditions does not exist and is prohibitively expensive to obtain. A potential solution to thislack of measured datais synthetic data derived from physics simulations to both train and characterize the algorithm.However, to do so we must address the disconnect between the synthetic model and the operating conditions of the real worldmeasured data. This research will develop solutions to this problemutilizing high fidelity physics models, statistical clutter models, radiative transfer models (RTMs), and generative adversarial networks (GANs) to generate training data over a large range of operating conditions (OCs) for which few measured samples exist.

Machine learning algorithms have demonstrated performance on par with or superior to human analysts on large datasets when sufficient training data is available. For military applications obtaining sufficient, truthed data is always a challenge. Three approaches have been used for measured data collections to support sensor exploitation programs: coordinated collections, turntable measurements and scaled model collections. However, these collections have limited availability of relevant targets of interest and the operating conditions over which they are collected are often not meaningfully representative of deployed environments. Our team has delivered scalable synthetic generation pipelines in support of multiple Department of Defense and Intelligence Community programs. In the Phase I proof of concept effort, we demonstrated the benefits of including synthetic foliage models to train machine learning algorithms. In this proposed Phase II effort, we will extend this capability, making it computationally tractable via parametric modeling of the foliage-induced back scatter and forward scatter and incorporating these models into our current synthetic pipeline. This comprehensive tool suite will address the challenge of data availability and provide capability to develop robust inference models.