Pulse radio frequency interference (PRFI), which has the potential to corrupt data and lower the quality of synthetic aperture radar (SAR) images, poses a serious threat to the integrity of SAR data. Traditional algorithms often introduce varying degrees of corruption into uncontaminated data and overlook the unique characteristics of SAR signals. This letter proposes a deep neural network with prior knowledge constraints (PKCNet) in the time–frequency domain to address the above issue, specifically designed to suppress PRFI while preserving uncontaminated data. This algorithm incorporates a regional constraint on the time–frequency spectrogram to safeguard uncontaminated data and designs a low-rank reconstruction module (LRM) based on the low-rank characteristic of SAR signals and the sparse characteristic of PRFI. A novel loss function is introduced to guide network training. Compared with frequency-domain notch filtering, time–frequency domain notch filtering, robust principal component analysis, and existing deep-learning-based algorithms, the images of Sentinel-1 recovered by PKCNet exhibit improvements of 19.20%, 2.68%, 7.97%, and 2.05% in terms of the energy of gradient (EOG). The code and dataset will be accessible online ( https://github.com/ZhengFenghao/PKCNet ).

中文翻译：

---

用于合成孔径雷达脉冲射频干扰抑制的先验知识约束网络（PKCNet）

脉冲射频干扰 (PRFI) 有可能损坏数据并降低合成孔径雷达 (SAR) 图像的质量，对 SAR 数据的完整性构成严重威胁。传统算法常常会给未受污染的数据带来不同程度的损坏，并且忽略了 SAR 信号的独特特征。这封信提出了一种在时频域中具有先验知识约束的深度神经网络（PKCNet）来解决上述问题，专门设计用于抑制 PRFI，同时保留未受污染的数据。该算法结合时频谱图的区域约束来保护数据不受污染，并根据SAR信号的低秩特性和PRFI的稀疏特性设计了低秩重建模块（LRM）。引入了一种新颖的损失函数来指导网络训练。与频域陷波滤波、时频域陷波滤波、鲁棒主成分分析以及现有的基于深度学习的算法相比，PKCNet 恢复的 Sentinel-1 图像表现出 19.20%、2.68%、7.97% 的改进，梯度能量 (EOG) 为 2.05%。代码和数据集将可以在线访问（ https://github.com/ZhengFenghao/PKCNet ）。