PASTE: Physics-Aware Scattering Topology Embedding Framework for SAR Object Detection

Jiacheng Chen

doi:10.1109/tgrs.2026.3736050

PASTE: Physics-Aware Scattering Topology Embedding Framework for SAR Object Detection

2026-09-22

0

PRE

AI

DOI：10.1109/tgrs.2026.3736050

Original

Original request for help

Save

Abstract

En 中文

Deep SAR object detectors often inherit optical-image paradigms and consequently underuse the sparse, discrete electromagnetic scattering structure of targets. Existing scattering-enhancement methods commonly rely on amplitude statistics or costly frequency-domain processing, making it difficult to incorporate physically meaningful structure into modern detectors. We propose Physics-Aware Scattering Topology Embedding (PASTE), a closed-loop framework that integrates scattering topology priors into SAR object detection. Scattering Keypoint Automatic Annotation (SKAA) automatically generates ASC-based keypoint annotations, Scattering Topology Injection Module (STIM) injects the learned topology into multi-scale features, and Scattering Prior Supervision Strategy (SPSS) uses Gaussian soft targets to supervise the scattering-map branch. Across seven CNN-, Transformer- and Mamba-based detectors, PASTE yields relative mAP@0.5 gains of 1.3% to 11.3%. Visualizations further show that the learned scattering representations distinguish target and background regions, improving interpretability together with detection accuracy. In addition, parameter ablation experiments and robustness perturbation experiments are also conducted to verify the practicality of PASTE.

Keywords:

Synthetic Aperture Radar (SAR)

Physics-Aware SAR Object Detection

Attributed Scattering Center Model

Scattering Topology Prior

Physics Prior Fusion

PASTE: Physics-Aware Scattering Topology Embedding Framework for SAR Object Detection

Abstract

Journal

Organization

Cited Papers