Advancing SAR Target Recognition Through Hierarchical Self-Supervised Learning with Multi-Task Pretext Training

Area of Science:

• Remote Sensing
• Artificial Intelligence
• Computer Vision

Background:

• Synthetic Aperture Radar (SAR) Automatic Target Recognition (ATR) systems struggle with limited labeled data and domain gaps between synthetic and real-world imagery.
• Current methods often rely on synthetic data, which can introduce inaccuracies and fail to generalize well to measured SAR data.

Purpose of the Study:

• To develop a comprehensive self-supervised learning (SSL) framework for SAR ATR that eliminates the need for synthetic data.
• To achieve state-of-the-art performance in SAR ATR by utilizing multi-task pretext training and evaluating diverse downstream classifiers.
• To establish a new paradigm for SAR ATR that leverages the inherent structure of radar data.

Main Methods:

• A multi-task SSL framework with nine pretext tasks (geometric invariance, signal robustness, multi-scale analysis) was developed.
• The framework was evaluated across various downstream classifiers including Support Vector Machines (SVM), Random Forest, XGBoost, Gradient Boosting, ResNet, U-Net, MobileNet, EfficientNet, and Generative Adversarial Networks.
• Extensive experiments were conducted using the SAMPLE dataset with rigorous cross-validation and comparison against the SimCLR baseline.

Main Results:

• The SSL framework significantly improved SAR ATR performance across multiple classifiers.
• SVM achieved 99.63% accuracy, Random Forest reached 99.26% accuracy, and ResNet18 attained 97.40% accuracy.
• Task-based SSL demonstrated superior performance compared to contrastive learning (SimCLR) for SAR ATR.

Conclusions:

• The proposed SSL framework offers a robust and effective approach to SAR ATR without synthetic data augmentation.
• This work provides practical guidelines for deploying SSL-based SAR ATR systems and sets a foundation for future research in domain-specific SSL for remote sensing.
• The findings highlight the potential of leveraging inherent radar data structures for improved ATR performance.