InSAR Phase Unwrapping via Integrated Multi-Model Deep Learning: Advancing Accuracy in Complex Topographic Hazards

Abstract: Synthetic Aperture Radar Interferometry (InSAR) is a critical tool for monitoring geohazards. However, Phase Unwrapping (PU) remains a significant impediment. Conventional algorithms frequently encounter failure in regions characterised by low coherence or pronounced topographic gradients, resulting in substantial error propagation. In order to address these challenges, the present study proposes an advanced framework that integrates three optimised deep learning (DL) architectures: Attention-enhanced U-Net (UA), Generative Adversarial Network (GAN)-based restoration (GL), and Convolutional Neural Network (CNN) with channel attention (CUA).

The performance of these models was systematically evaluated using a large-scale, diverse InSAR benchmark dataset. Quantitative results demonstrate a substantial leap in accuracy compared to traditional methods. All three proposed DL models achieved a Peak Signal-to-Noise Ratio (PSNR) exceeding 28 dB and a Structural Similarity Index (SSIM) above 0.85. Specifically, the CUA model demonstrated the highest level of precision, achieving a PSNR of 38.24 dB and effectively suppressing noise in complex interferograms. In order to preserve structural integrity in areas of sharp terrain, the UA model (incorporating a 5-layer attention mechanism) achieved an edge SSIM of 0.8888, thereby demonstrating a significant improvement over the Minimum Cost Flow (MCF) algorithm, which frequently encounters difficulties with phase residues in high-gradient regions.

In order to validate the practical applicability of the models, they were tested on real TanDEM-X data from the Weinan region in China. The UA model exhibited a high average SSIM of 0.95, successfully recovering subtle terrain features where traditional MCF demonstrated a mean PSNR of only 18.08 dB. Moreover, a gradient accumulation strategy was introduced with a view to optimising the training process. A thorough efficiency analysis reveals that the GL model (at a 1:1 ratio) can reduce training time by approximately 92% compared to the high-complexity CUA-Accum2 configuration, offering a scalable solution for SAR big data processing.

In conclusion, this research provides a robust, automated, and high-precision methodology for InSAR PU. The present work offers novel insights into the generation of reliable geodetic products for disaster risk reduction in challenging environments by bridging the gap between advanced DL processing and real-world hazard monitoring.

Keywords: U-Net, generative adversarial network (GAN), convolutional neural network (CNN), phase unwrapping(PU), synthetic aperture radar interferometry (InSAR).
(Corresponding author: Ruiqing Niu)