Resolving phase unwrapping in interferometric synthetic aperture radar using deep recurrent residual U-Net

Abstract

Phase unwrapping (PhU) is a primary task in interferometric synthetic aperture radar (InSAR) with the ability to resolve the ambiguity of modulo 2π and obtain the absolute change of phase. This paper addresses the PhU problem by introducing a novel network configuration inspired by recurrent residual U-Net. There current residual convolutional blocks improve feature representation, while the U-Net shape architecture maintains the fusion of a low level with high spatial features. This architecture fused the advantage of the of residual learning, recurrent and U-Net connection. The residual connection boosts the flow of information, recalibrate the data and the efficiency of the parameter. Finally, ample experiments were conducted on Synthetic dataset, which contains of simulated InSAR images with 2000 image pairs (256 × 256). To further assess the proposed network using the Sentinel-1A dataset for the Sukari gold mine in Egypt. Experiments on simulated and real Sentinel-1A datasets proved that the proposed network outperforms traditional SNAPHU and other deep learning-based approaches on the basis of qualitative and quantitative comparisons, and that the proposed network architecture outperforms traditional SNAPHU and other deep learning-based approaches.

Keywords

• Phase unwrapping
• Convolutional neural networks
• U-Net architecture
• Recurrent residual blocks