Loss Function Design for Wrapped Phase Fitting in InSAR Deep Learning Network Training

作者：Chunjing Chen, Chisheng Wang, Haidong Yu · 发表于：IEEE Geoscience and Remote Sensing Letters · 年份：2024 · DOI：10.1109/lgrs.2024.3433007 · 被引用次数：4 · 研究领域：Synthetic Aperture Radar (SAR) Applications and Techniques、Advanced SAR Imaging Techniques、Geophysical Methods and Applications

In recent years, the application of deep learning in the field of interferometric synthetic aperture radar (InSAR) measurement has gradually increased. Unlike the real-value data typically handled by conventional deep learning methodologies, InSAR technology grapples with complex-value data. Specifically, the phase values of this complex data are constrained within the interval ($-\pi,+\pi $]. However, in deep learning frameworks, the calculation of loss functions often neglects the mathematical constraint that phase differences should also adhere to the [$-\pi,\pi $] range. This oversight impedes the model’s ability to accurately fit the target data, thereby diminishing training efficiency. To address this challenge, we introduce three foundational loss functions tailored for fitting wrapped phase values: cosine similarity-based (La), mean square error (Lb) for wrapped phase, and ensemble coherence-based (Lc). By combining these approaches, we formulate a total of seven distinct loss functions. To identify the most effective one, we evaluated their performance using a temporal convolutional network (TCN) as a test scenario. Our findings reveal that the optimal combination function, Labc, achieves a 72.78% accuracy rate. Compared to other function with over 72% accuracy, Labc reduces the required iterations for model convergence by at least 30% and effectively delineates terrain feature boundaries.