Multibranch Dilated Residual Network for InSAR Phase Denoising

Abstract

Interferometric phase noise governs the accuracies of both subsequent interferometric synthetic aperture radar (InSAR) data processing and the final measurements. Filtering is the main method for reducing the phase noise of InSAR. However, the traditional filtering algorithms operating in the spatial or transform domains often fail to reconcile effective noise suppression with the preservation of fine fringe details. Besides, the existing deep learning-based denoising methods still exhibit limitations, such as excessive smoothing, generation of spurious fringes, and redundant computations. To address these issues, this letter proposes a novel denoising network model, i.e., multibranch dilated residual network (MDRNet). It integrates dilated convolutions with a channel attention mechanism, achieving adaptive modulation of the receptive field, thereby effectively preserving fringe edges and authentic deformation signals while suppressing spurious fringes in decorrelated regions. Both simulation and real experiment results indicate that MDRNet attains superior performance under conditions of strong noise and low coherence. It achieves a more favorable balance between denoising efficacy and phase fidelity compared to existing methods. The corresponding test code is available at https://github.com/xiangHQ/MDRNet.git