The growing availability of Synthetic Aperture Radar (SAR) target datasets allows for the consolidation of different SAR Automatic Target Recognition (ATR) tasks using a foundational model powered by Self-Supervised Learning (SSL). SSL aims to derive supervision signals directly from the data, thereby minimizing the need for costly expert labeling and maximizing the use of the expanding sample pool in constructing a foundational model. This study investigates an effective SSL method for SAR ATR, which can pave the way for building the foundation model. The primary obstacles faced in SSL for SAR ATR are the scale problem of the remote sensing images and speckle noise in SAR images. To overcome these challenges, we present a novel approach called Knowledge-Guided Predictive Architecture (SAR-KPGA), which leverages local masked patches to predict the multi-scale SAR feature representations of unseen context. The key aspect of SAR-KPGA is integrating SAR domain features to ensure high-quality target features for SSL. Furthermore, we employ local masks and multi-scale features to accommodate the large image scale and target scale variations in remote sensing scenarios. By evaluating our framework on three target recognition datasets (vehicle, ship, and aircraft), we demonstrate its outperformance over other SSL methods and its effectiveness with increasing SAR data. This study showcases the potential of SSL for SAR target recognition across diverse targets, scenes, and sensors.

翻译：随着合成孔径雷达（SAR）目标数据集的日益丰富，基于自监督学习（SSL）构建基础模型有望统一不同的SAR自动目标识别（ATR）任务。自监督学习直接从数据中提取监督信号，从而减少对昂贵专家标注的依赖，并最大化利用不断扩大的样本池来构建基础模型。本研究探索了一种适用于SAR ATR的高效自监督学习方法，为构建基础模型奠定基础。当前SSL在SAR ATR中面临的主要挑战包括遥感图像的大尺度问题以及SAR图像中的散斑噪声。为克服这些障碍，我们提出了一种名为知识引导预测架构（SAR-KPGA）的新方法，该方法利用局部掩码图像块来预测未见上下文的多尺度SAR特征表示。SAR-KPGA的核心在于融合SAR领域特征，以确保用于自监督学习的高质量目标特征。此外，我们采用局部掩码与多尺度特征，以适应遥感场景中图像大尺度与目标尺度变化的问题。通过在三个目标识别数据集（车辆、船舶与飞机）上评估我们的框架，证明了其优于其他自监督学习方法，并展示了随SAR数据增加而增强的有效性。本研究展示了自监督学习在跨目标、跨场景及跨传感器SAR目标识别中的潜力。