ASC-SepNet: Enhancing Robust SAR Ground Target Recognition via Attribute Scattering Center and Separability Dual-Driven Learning

作者：Yifan Zhang, Xunzhang Gao, Jingyuan Xia, Weijie Li, Shuanghui Zhang, Li Liu, Xiang Li · 发表于：IEEE Transactions on Aerospace and Electronic Systems · 年份：2025 · DOI：10.1109/taes.2025.3564588 · 被引用次数：14 · 研究领域：Advanced SAR Imaging Techniques、Geophysical Methods and Applications、Underwater Acoustics Research

Deep learning models based on the independent identically distributed (i.i.d.) assumption perform poorly once the assumption cannot be satisfied under two challenges of Synthetic Aperture Radar (SAR) ground target classification: feature distribution shifts and limited training samples. To address these issues, we propose an Attribute Scattering Center and Separability dualdriven Network (ASC-SepNet), which combines the scattering prior and the separability of features to achieve robust classification under various extended conditions. Concisely, ASC-SepNet consists of four components. The Scattering Prior Modeling (SPM) module is responsible for extracting and modeling the physical scattering prior, while the Robust Feature Learning (RFL) module guides the backbone to extract physically consistent features. Together, these two modules enhance the expressiveness of the features by representing and embedding physical priors, which in turn enables the model to maintain robust performance even when faced with feature distribution shifts in test sets. The Feature Separability Measure (FSM) module quantifies the discriminative power of the features and constrains the model to learn features with stronger separability, thereby enabling the model to preserve strong robustness even with limited training samples. Further, to eliminate redundant features and lower parameter count, a Two-Stage Feature Fusion (TSFF) module is proposed to integrate multi-source features. The efficacy of t...