BB-PIP-U²Net: a physics-constrained optimization framework for robust multibaseline InSAR phase unwrapping.

Area of Science:

• Geophysics and Remote Sensing
• Artificial Intelligence and Machine Learning
• Signal Processing

Background:

• Phase unwrapping (PU) is crucial for InSAR but is severely limited by noise.
• Existing methods struggle with accuracy and computational efficiency in noisy environments.

Purpose of the Study:

• To develop a robust and accurate phase unwrapping method for multi-baseline InSAR (MB-InSAR) that overcomes noise limitations.
• To integrate deep learning with mathematical optimization for improved interpretability and performance.

Main Methods:

• A physics-constrained hybrid framework, BB-PIP-U²Net, combining deep learning (U²Net with depthwise separable convolutions and hybrid attention) and branch-and-bound pure integer programming (BB-PIP).
• Construction of a large-scale, physics-realistic dual-simulation integrated (PRDS) dataset using Zernike polynomials for noise modeling.
• Utilizing bottleneck latent features as a warm start for BB-PIP to solve a constrained global integer optimization problem.

Main Results:

• BB-PIP-U²Net demonstrates high noise robustness, achieving an RMSE of 1.18 rad under 3 dB SNR, outperforming PIPNet (1.58 rad) and TSPA (3.66 rad).
• The method is 5.8% faster than PIPNet.
• Validation on real SAR data confirms a balance between computational efficiency and unwrapping accuracy.

Conclusions:

• BB-PIP-U²Net offers a significant advancement in MB-InSAR phase unwrapping, providing mathematically optimal and physically interpretable results.
• The framework enhances measurement accuracy and deformation monitoring capabilities in InSAR technology.