Physics-Based SAR Modeling and Simulation for Large-Scale Data Generation of Multi-platform Vehicles for Deep Learning-Based ATR

Abstract

One critical challenge of Automatic Target Recognition (ATR) systems are that of effective modeling and interpretation of sensory data obtained under constantly changing dynamic environments. In this paper, we address physics-based modeling and simulation of multi-platform vehicles and propose a method for systematic generation of synthetic SAR Imagery for training of Deep Learning (DL) techniques. Starting with computer-aided design (CAD) models of aerial, ground and maritime vehicles, we present a multi-layer method for describing physics-based models of these objects. Next, by considering SAR system constraints and modeling far-field incident and backscattering radiation waves returns, we construct realistic simulated (i.e., synthetic) SAR imagery of the test vehicles and annotate them semantically to aid their DL training. To evaluate and verify the effectiveness of this approach, we compare our synthetically generated SAR imagery against the real SAR images. Several examples of our test scenarios are demonstrated and explained including our post image modulation technique that further enhances realisms of the synthetic SAR images. Finally, we discuss the implication of our technical approach in support of dynamic data-driven applications systems.