A Semi-Supervised Self-Boosting Learning Method for InSAR Phase Denoising

Interferogram denoising is a critical step in interferometric synthetic aperture radar (InSAR) data processing, as it directly affects the accuracy and reliability of final products, such as surface deformation measurements and digital elevation models. Current state-of-the-art data-driven denoising models primarily rely on training datasets with noise simulated using statistical models, such as the complex Gaussian distribution. While effective in controlled scenarios, these simulated noises often fail to capture the intricate variability of real-world interferometric noise, leading to degraded performance in practical applications. In this study, we propose a semi-supervised self-boosting learning method for InSAR phase denoising, marking the first instance of training a model using real-world noise. The method consists of two phases: (1) Model excitation, where the model is initially trained with simulated noise to develop basic denoising capabilities; (2) Refinement boosting, an unsupervised, iterative process where the model gradually refines itself using real-world noise. Specifically, this phase involves four interconnected steps: noise extraction, where noise is extracted from real interferograms; noise purification, which removes residual signal components; data augmentation, where purified noise is updated into the training dataset; and model enhancement, which iteratively refines the model to improve its generalization to real interferograms. We identified the cost-optimal denoising model by conducting experiments across network architectures of varying complexity, using identical training datasets and experimental settings. Experimental results validate the effectiveness of the proposed method on both synthetic interferograms with varying coherence levels and real Sentinel-1 interferograms. On synthetic data, the method demonstrates superior denoising performance, achieving the lowest root mean square errors (RMSE) and highest structural similarity index measures (SSIM) compared to state-of-the-art techniques such as NL-InSAR and InSAR-BM3D, while maintaining comparable inference speeds to simpler methods like BoxCar. On Sentinel-1 interferograms, the approach consistently delivers improved denoising results, as evidenced by fewest phase residues and smoothest phase unwrapping. Our findings also reveal that when training data is comprehensive and well-aligned, increasing model complexity does not necessarily lead to giant improvement; simpler architectures can yield results comparable to those of more sophisticated models. Additionally, by analyzing noises simulated from the coherence-guided statistical model and those extracted from Sentinel-1 interferograms, we observe a significant discrepancy between simulated and real noise distributions, with the former failing to capture the complexities of real-world noise. This underscores the importance of incorporating real-world noise into training datasets for InSAR data-driven models, e.g., denoising, unwrapping, and other applications. Overall, this research introduces a robust methodology for interferogram denoising and enhances our understanding of the complexities of real-world interferometric noise, paving the way for further advancements in noise modeling and interferogram restoration.