Data-Efficient Deep Learning for SAR Automatic Target Recognition: A Comprehensive Survey
Year: 2026
Type: article
Abstract: Synthetic aperture radar (SAR) automatic target recognition (ATR) is a critical capability in remote sensing with unique challenges stemming from high intraclass variation, speckle noise, and limited labeled datasets. Deep learning has shown promise in enhancing SAR ATR performance, yet conventional supervised approaches remain data-hungry and computationally intensive. This survey presents the first comprehensive review of modern data-efficient learning paradigms tailored for SAR ATR, encompassing self-supervised learning, semisupervised learning, knowledge distillation, and incremental learning. For each paradigm, we define foundational concepts, architectural patterns, and representative methods, emphasizing adaptations required for the SAR domain. We analyze model performance, generalization, and robustness, particularly in few-shot and class-incremental settings. Unlike prior surveys, this work unifies fragmented literature, introduces a new taxonomy based on learning strategy rather than architecture, and highlights emergent methods such as transformer backbones and self-distillation. We identify research gaps in domain adaptation, pseudolabel reliability, and long-term forgetting, establishing a roadmap for advancing SAR ATR under real-world constraints.
Institution: Duke University
Language: English
Subfield: Aerospace Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $1,800
OpenAlex ID: W7167279566