MPFNet: A Multiscale Phase Filtering Network for Interferometric SAR

Abstract

Phase filtering is one of the core signal processing steps in interferometric synthetic aperture radar (InSAR). In recent years, InSAR phase filtering algorithms have evolved from traditional solutions to deep learning (DL) methods, significantly improving the processing efficiency. However, most DL-based phase filtering techniques originate from optical filtering methods, and these methods inevitably entail a tradeoff between noise suppression and detail preservation. To resolve this contradiction and fully take into account the characteristics of InSAR phase, a multiscale phase filtering network (MPFNet) based on multilook information fusion is proposed. First, the network adopts the multiscale structure to balance noise suppression and detail preservation, where the multiscale information is obtained through multilook interferograms of varying numbers of looks. Second, drawing on the mechanism of super-resolution, the network incorporates the residual feature distillation blocks (RFDBs) to restore the scale of interferograms. Finally, in response to the demand for complex phase filtering, a loss function based on cosine similarity is constructed, which avoids the discontinuity at $\pm \pi $ affecting the filtering results. Computer simulation and experiments based on real InSAR data verified the effectiveness of the proposed method.