SAR ATR Performance Evaluation on Spatially Perturbed Synthetically Generated Signatures

Document Type

Conference Proceeding

Publication Date

5-29-2025

Department

Michigan Tech Research Institute; Great Lakes Research Center; Department of Computer Science

Abstract

Automatic target recognition (ATR) on synthetic aperture radar (SAR) data can be a challenging task due to the limited availability of publicly available measured datasets. Prior work has focused on the Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset and the Synthetic and Measured Paired and Labeled Experiment (SAMPLE/SAMPLE+) datasets. The use of synthetic or modeled data in training AI has increased, both as modeling tools become more effective and also with increasing understanding of how to train AI with synthetic data. In this paper, we investigate the challenge of training an ATR with limited real-world data, focusing on two main problems: i) how best to use training tools to effectively train a robust ATR with combinations of measured and synthetic data, and ii) what is the effect of synthetic model uncertainty when using synthetic data to train ATRs. Model uncertainty could come from, for example, incomplete target data—say, a few images—or uncertain intelligence. Hence, we will show how noise augmentation and dropout affect ATR performance for varying proportions of measured versus synthetic training data. Then we will show how model uncertainty, produced by injecting uniform noise in to the vertices of a CAD model used to simulate SAR imagery, affects ATR performance. The target we will use is SLICY, an MSTAR target that was constructed of radar reflector primitives, which enables interpretability of ATR performance, both in terms of algorithm accuracy but also in saliency of the ATR features. Several experiments are performed that address how noise augmentation, dropout, and model uncertainty affect a canonical ATR used on MSTAR data. Results suggest some rules of thumb in how to train SAR ATRs with limited real-world training data and potential synthetic model uncertainty.

Publication Title

Proceedings of SPIE the International Society for Optical Engineering

ISBN

9781510687158

Link to Full Text

Share

COinS