Swin Transformer-Based Unsupervised Domain Adaptation for ISAR Satellite Classification

Radar images contain rich information, encompassing motion, dimensionality, and electromagnetic characteristics. Unlike optical images, radar images are resilient to varying weather conditions and lighting, enabling continuous monitoring across diverse applications. However, substantial differences between radar and optical images, including their generation processes, definitions, and noise distributions, pose significant challenges for transfer learning, particularly in zero-shot inference when applying computer vision models to radar images, especially inverse synthetic aperture radar (ISAR) images. In practical applications, access to abundant labeled optical images is common, while labeled ISAR images are often scarce, making it difficult to learn effective representations of ISAR objects in an unsupervised setting. This disparity underscores the need for robust domain adaptation techniques to bridge the gap between labeled optical data and unlabeled ISAR data, enabling accurate ISAR object classification without relying on extensive labeled ISAR datasets. This paper presents a novel approach that integrates state-of-the-art attention-based pre-trained vision models with metric learning to effectively leverage unlabeled ISAR images. By capturing shared features across domains, the proposed method achieves high classification performance on a satellite dataset, demonstrating results comparable to those obtained under supervised settings.