Few-Shot SAR Target Classification via Scattering-Aware Interaction and Distribution Alignment

Abstract

Synthetic aperture radar (SAR) automatic target recognition is crucial for detecting high-value targets. However, such targets often exhibit weak scattering and strong concealment, making measured data difficult to obtain and leading to a typical few-shot scenario. In few-shot settings, severe background clutter and high structural similarity among target classes further complicate target-background discrimination and intensify fine-grained interclass confusion. To address these challenges, this article proposes the scattering prior and local interaction joint feature distribution alignment network (SL-JFDA). The core idea is to mitigate recognition confusion in few-shot scenarios through coordinated enhancements in feature representation and distribution alignment. Specifically, to preserve structurally intact and discriminative features under strong clutter, we design a hierarchical backbone consisting of two key modules: the high-frequency scattering prior attention enhances target scattering centers via physical priors, and the local interaction perception captures local context by combining deep semantic and low-level features, compensating for the discriminative deficiency caused by limited samples. To address class-boundary ambiguity in few-shot learning, classification is formulated as an optimal transport problem. The joint feature distribution alignment module uses the Sinkhorn algorithm to compute the optimal transport cost between a query and aggregated class-support distributions, enabling robust distribution-level matching and reducing confusion from outliers and noise. Experiments on MSTAR, OpenSARShip, and FuSARShip demonstrate the effectiveness of SL-JFDA: relative to the strongest baseline using a standard Conv4 backbone, it achieves average accuracy gains of 7.47%, 1.08%, and 2.24% in 1-shot settings; 7.70%, 2.18%, and 2.20% in 5-shot settings, consistently outperforming existing few-shot learning methods in few-shot SAR target recognition.