This project introduces a novel self-supervised learning (SSL) framework for Synthetic Aperture Radar (SAR) Automatic Target Recognition (ATR) that addresses two critical challenges: the scarcity of labeled SAR data, and the persistent domain gap between synthetic and measured SAR imagery. Unlike traditional approaches that rely on synthetic data augmentation, our framework leverages multi-task pretext training to develop robust feature representations directly from measured SAR data, eliminating synthetic data dependencies entirely. Using the SAMPLE (Synthetic and Measured Paired and Labeled Experiment) dataset which contains ten distinct military vehicle classes, we implement nine complementary pretext tasks—including rotations, flips, blurring, denoising, and zoom transformation, to learn discriminative features that generalize well across diverse operating conditions. This approach eliminates synthetic data dependency while achieving new performance benchmarks with strong data and computational efficiency. This approach achieves competitive performance with a diverse range of machine learning, deep learning, and generative adversarial network-based downstream classifiers. In addition, we conduct a systematic layer-wise representation analysis to understand how different network depths influence SAR feature transferability. As a result, the framework is expected to enhance ATR performance, reduce false alarms, and broaden the operational capabilities of SAR systems in complex environments as well as to provide a foundation for future research in domain-specific self-supervised learning.