Multiscale Residual Denoising Network for Robust InSAR Phase Reconstruction

Interferometric Synthetic Aperture Radar (InSAR) phase denoising is essential for deformation monitoring and topographic mapping, yet it remains challenging due to the inherent trade-off between denoising strength, structural preservation, and computational efficiency. Classical filters maintain phase integrity but oversmooth fine details, non-local methods deliver effective noise suppression at high computational cost, and existing deep learning approaches often alter low-SNR regions, introducing artifacts. To overcome these limitations, we propose the Multi-Scale Residual Denoising Network (MSRD-Net), a hierarchical phase reconstruction framework that integrates the nonlinear recovery capability of deep learning with the robustness of frequency-domain filtering. MSRD-Net decomposes interferometric phase into low-frequency and corresponding residual components, preserving dominant structures while progressively restoring fine-scale details. A self-boosting training strategy using real-world noise extracted from Sentinel-1 and TanDEM-X data further improves the network's robustness and generalization. Evaluations across four representative scenarios—earthquakes, landslides, rainforests, and urban areas—show that MSRD-Net achieves a balanced improvement in noise suppression, structural fidelity, and efficiency. Furthermore, its scalable design allows adaptive adjustment of denoising strength, while its iterative nature ensures reliable performance under extreme noise. These characteristics establish MSRD-Net as a practical and versatile solution for both large-scale operational monitoring and high-precision geoscience studies with InSAR.