Are Deep Learning Models Trained on Synthetic Data Actually Reliable in InSAR? A Self-Supervised Learning Approach for InSAR Phase Denoising

Deep learning is revolutionizing the way InSAR data are processed. Supervised deep learning models trained on synthetic datasets have demonstrated strong performance according to conventional evaluation metrics. However, in practice, many supervised InSAR phase denoising filters tend to oversmooth fringe patterns and generate ghost fringes in decorrelated regions, raising concerns about the reliability of models trained solely on synthetic data in real-world applications. In this work, we introduce Noise2Fringe, an unsupervised learning framework that enables the direct training of denoising filters on noisy interferograms, eliminating the need for clean references, while achieving performance comparable to that of supervised models. This framework allows a fair comparison between models trained exclusively on real versus synthetic data, independent of the training scheme. On synthetic benchmarks, models trained on real data appear to underperform compared to those trained on synthetic data. However, in real-world interferograms, the opposite holds: synthetic-trained models oversmooth and generate ghost fringes, whereas real-trained models avoid these artifacts. These findings reveal (i) a much larger gap between synthetic and real data than previously assumed, (ii) an overestimation of the effectiveness of synthetic-trained models, and (iii) deficiencies in current evaluation methods that rely heavily on simulated benchmarks. We further find that incorporating auxiliary information, such as the Amplitude Dispersion Index, into the input can enhance denoising performance. Overall, this study demonstrates that models trained solely on synthetic data may be far less reliable in real-world applications than suggested by their benchmark results, and emphasizes that generating more realistic training datasets should take precedence over designing increasingly complex model architectures.