Correctly labeled data is essential for training AI/ML-based automatic target recognition (ATR). The training process is all the more complicated in synthetic aperture radar (SAR) images because of their unique phenomenology, such as orientation-sensitive target signatures, layover, cross-range smearing, and radio frequency interference. New automated technology must reduce the cost and accelerate the timeline of manual labeling. We propose Auto Label SAR (AL-SAR) by extending our DELPHI Active Learning (AL) framework for single-target SAR image classification to a framework for multi-target classification. In Phase I, we will leverage open-source Python and state-of-the-art AL with Core-set selection. Our Phase 1 feasibility study will simulate human-in-the-loop labeling using publicly available or government provided pre-labeled training data. Resources will be focused on critical technical questions and experiments to illuminate and prioritize key enabling capabilities for AL-SAR development during Phases II-III. AL-SAR will be an automated, cost-effective, and adaptable target labeling tool which could also be used to label targets in other mission areas with different data sources.
