Research On InSAR Phase Unwrapping Via Deep Learning Based Region Segmentation

Synthetic Aperture Radar Interferometry(InS AR)is a new technology that combines S AR imaging and interferometry,and has made great achievements in three-dimensional reconstruction and deformation monitoring.Since the InSAR system cannot obtain accurate geophysical quantities directly through the interferometric phase,phase unwrapping has been an important p art of the InSAR workflow,which can restore the absolute phase by wrapped phase,so that the entire image can provide meaningful information.On the one hand,deep learning technology is profoundly affecting technological changes in all walks of life;on the other hand,phase unwrapping technology urgently needs some new algorithms with higher unwrapping efficiency and unwrapping accuracy.Combining deep learning to innovate this technology has become an inevitable trend.Due to the "phase continuous distribution" characteristic of the wrapped phase,the wrapped phase always exists in the form of interference fringes,which undoubtedly brings the possibility of the "deep learning region segmentation" method.Therefore,this article first carried out the "single baseline phase unwrapping technology based on deep learning region segmentation",through the convolutional neural network to divide the wrapped phase into a number of "continuous phase regions",and establish the region as the object of the unwrapping model.Simulation and measured data will be used to verify the effectiveness of the algorithm.Experiments show that the region segmentation method can significantly improve the stability of the unwrapping result and improve the accuracy.Based on the obvious results of "single baseline phase unwrapping of deep learning region segmentation method",this article further extends the deep learning region segmentation method to multi-baseline phase unwrapping.By analyzing the characteristics of the multi-baseline wrapped phase,and combining the characteristics of the ambiguity number,linear overdetermined equations with the ambiguity number as the unknown quantity have been built,and then complete the unwrapping of the multi-baseline wrapped phase diagram.This method has achieved obvious results in multi-baseline unwrapping,and has a good unwrapping effect on complex research objects with sudden changes in absolute phase.The above studies have shown that the phase unwrapping technology introduced by the deep learning region segmentation method shows strong vitality,and the innovation of the phase unwrapping technology by deep learning is far more than the application in region segmentation.Based on the study of region segmentation,this paper also proposes an improved least squares unwrapping method that combines deep learning to predict the horizontal and vertical phase gradients,which achieves "fast and good" unwrapping result as a whole.It can be said that the research of phase unwrapping technology combined with deep learning is a key to the development of InSAR phase unwrapping technology and determines the development direction of future unwrapping technology.