DP-Net: A U-Shaped Multidimensional Multiscale Fusion Neural Network for InSAR Phase Filtering

Abstract

Synthetic aperture radar interferometry (InSAR) is an essential tool for observing the Earth’s surface, widely employed in geohazard and ground subsidence monitoring. Enhancing interferogram quality through phase filtering is particularly significant. Traditional filtering methods are often ineffective, while emerging deep learning approaches still face challenges in noise removal and stripe edge preservation. This article proposes a novel InSAR phase filtering method, the dilated phase network (DP-Net), based on a U-shaped multidimensional and multiscale fusion neural network. The proposed method employs a U-shaped network architecture to achieve effective fusion and fine processing of interferogram features across multiple dimensions. By incorporating a Dilated module with embedded cavity convolution, the network enhances its capability to capture features at various scales. Furthermore, the method integrates features at different levels during the encoding-decoding process, enabling effective noise reduction while preserving interferogram details and improving filtering quality. Additionally, a simulated dataset is generated and trained using digital elevation model (DEM) inversion with hierarchical noise addition. The efficacy of the method is validated through filtering experiments on both simulated and real data.