关键词: 合成孔径雷达, 属性散射中心, 深度学习, 复数神经网络

Abstract: Attributed scattering center (ASC) provides crucial physical descriptions for the interpretation of synthetic aperture radar (SAR) targets. However, traditional extraction methods are computationally expensive and exhibit limited performance in complex scenarios. Concurrently, most deep learning approaches are confined to tasks such as target classification, with their internal feature representations showing a weak correlation to physical scattering mechanisms and failing to fully leverage the complex-valued nature of SAR data. To overcome these challenges, this paper proposes a deep network, named the real-complex scattering network (RCS-Net), to achieve end-to-end, efficient, and precise regression from SAR complex-valued data to ASC physical parameters. The core of RCS-Net is a custom-designed real-complex U-shaped architecture, which can directly process and effectively utilize both magnitude and phase information. This study further introduces a training strategy guided by physical model parameters: the orthogonal matching pursuit (OMP) algorithm is used to generate ASC parameters offline as supervisory signals. A physics-constrained multi-task loss function then drives the network to learn the multi-dimensional physical attributes of the scattering centers, including their position, amplitude, and frequency- dependent factors. Experimental results on the MSTAR dataset demonstrate that compared to the conventional OMP method, RCS-Net achieves a nearly three-orders-of-magnitude improvement in inference efficiency. It also shows superior performance in parameter estimation accuracy and the quality of reconstructed images, along with strong generalization capabilities. Ablation studies further validate the effectiveness of the real-complex architecture and the physics-guided training strategy. This research offers an efficient deep learning solution for the inversion of physical parameters from SAR images.

Key words: synthetic aperture radar, attributed scattering center, deep learning, complex-valued neural network