Performance Enhancement of Deep-learning-based InSAR Phase Unwrapping by Optimizing Training Data and Model Structure

Abstract

Phase unwrapping is an essential processing step in SAR interferometry, which estimates the absolute phase from the wrapped phase within (- 𝜋, 𝜋]. Phase unwrapping is an essential data processing procedure for synthetic aperture radar interferometry. Accordingly, a lot of traditional unwrapping algorithms have been developed. Phase unwrapping is still a challenging problem in the presence of steep phase gradients and a noisy area. Recently, deep-learning-based phase unwrapping approaches have been proposed, and they show superior performance than conventional phase unwrapping algorithms. However, recent studies have not considered 1) the locally different noise, and 2) the data balance of phase gradient and noise. In addition, although, the unwrapped phase is estimated by accumulating relative phase differences between adjacent pixels from the reference point on the entire wrapped phase image, conventional model structures for semantic segmentation were adopted as it is without consideration of the phase unwrapping process. Therefore 3) the models have difficulty exploiting the phase information of the entire image together. In this study, training data and model structure were optimized for the performance enhancement of deep-learning-based phase unwrapping. For that, the training data was simulated with simple and local noise. And data augmentation was applied for balancing the phase gradient and noise level. Besides, the multi-encoder U-Net regression model structures are suggested, which have different kernels of 3X3, 5X5, and diliated 3X3. Also, the best model structure was determined by comparing the unwrapping performance according to the numbers of pooling layers and encoders. Finally, we found that optimizations of training data and model structure are a valid approach for enhancing deep-learning-based phase unwrapping. The mean absolute errors for applying suggested models, which were trained by simple and local noise, to real synthetic aperture interferograms were 0.592 and 0.445 respectively. Single-kernel model trained by local noise showed only a mean absolute error of 0.542. For the same phase data, mean absolute errors of minimum cost flow and statistical-cost, network-flow algorithm for phase unwrapping were 0.953 and 0.861 respectively. We expect that this study will contribute to designing the model structure and training data simulation approaches for the phase unwrapping, and also help to clarify earth internal processes and mechanisms.