Self-Supervised Learning for SAR Target Recognition with Multi-Task Pretext Training

The scarcity of labeled data and the difference between synthetic and measured imagery impose major obstacles during Synthetic Aperture Radar (SAR) Automatic Target Recognition (ATR). This paper presents a novel self-supervised learning (SSL) framework for SAR target recognition. The model utilizes multi-task pretext training to develop robust feature representations from measured SAR data. Our framework removes reliance on synthetic data by employing complementary pretext tasks to pretrain a deep learning network. The learned features are then evaluated using multiple classifiers, including Support Vector Machine (SVM), XGBoost, and Random Forest, on the Synthetic and Measured Paired and Labeled Experiment (SAMPLE) dataset. The experimental findings demonstrate competitive performance and robust detection capabilities, even with limited training data and low false positive rates. Our framework also outperforms traditional methods and eliminates the dependency on synthetic data. This work highlights the potential of SSL in SAR applications and provides a foundation for future research in domain-specific self-supervised learning. This framework not only advances the state of SAR target recognition but also establishes a foundation for leveraging self-supervised learning in other domain-specific applications with limited labeled data.