Research On Two-Dimensional InSAR Phase Unwrapping Method Via Deep Learning

As a widely used geodetic method for 3D reconstruction of terrain and surface physical process analysis,Interferometric Synthetic Aperture Radar(InSAR)technology has gradually matured in the past few decades.As a key and difficult problem in InSAR data processing,phase unwrapping(PU)has garnered a lot of attention since InSAR was proposed,and numerous researchers have carried out a lot of research on it.At present,the efficiency and accuracy of PU in conventional scenes can be guaranteed,but when the data quality is poor,traditional unwrapping methods will inevitably introduce unwrapping errors,which greatly limits the development and progress of the InSAR technique.In essence,phase unwrapping is an ill-posed inverse problem,so we can’t solve it directly using mathematical means.Traditional PU methods belong to the category of model-driven methods and are all constrained under the framework of phase continuity assumption.The phase continuity assumption requires that the wrapped phase of any pixels in the interferogram meets the Nyquist sampling theorem.However,when the phase gradient between adjacent pixels is too large,the phase continuity assumption is no longer tenable.At this time,the traditional method can not obtain the correct unwrapping results.In recent years,deep learning has flourished in many fields.It directly realizes end-to-end learning through data-driven methods without any prior model and shows better effects than traditional methods.Previous research has proved the effectiveness of deep learning in InSAR PU.The basic idea is to transform the unwrapping problem into the fields where deep learning experts in,such as image segmentation,image translation,and so on.In this study,the shortcomings of the existing deep learning-based InSAR phase unwrapping methods are discussed and improved.First,the value of K in the relevant research taking the ambiguity as the ground truth is limited to a fixed range,which affects the generalization performance of the method,and the performance of taking K as the ground truth has not been tested in InSAR field.Secondly,there is a lack of contrast between different Ground Truths under the same external conditions.Moreover,the dataset simulated in existing methods is relatively simple,the samples are not sufficient enough,and the class imbalance problem restricts the unwrapping accuracy when taking the ambiguity gradient as Ground Truth.Given this,the main work and innovations of this paper include:(1)In this study,the noise robustness and performance of several classical 2D InSAR phase unwrapping methods are tested and compared using simulated data and real data,respectively,to provide a reference for the selection of PU methods in different scenes.The failure of phase continuity assumption leads to the bad performance of traditional methods when the data quality is poor,which proves the necessity of developing deep learning-based InSAR phase unwrapping methods.(2)In this study,an InSAR PU method based on cGAN(Conditional Generative Adversarial Networks)is developed,it provides a new means to solve the unwrapping problem in circumstances with low coherence or phase discontinuity.Two kinds of Ground Truth,the absolute phase and the ambiguity,are introduced into the field of InSAR PU.Firstly,the real parameters of two satellite sensors are used to simulate the dataset,and the feasibility of the method is tested using U-Net.Then we adjust U-Net according to the unwrapping task as the generator,Patch GAN is used as the discriminator to build the generative adversarial network.The accuracy of the results can be improved in the zero-sum game.It is worth mentioning that when using ambiguity as Ground Truth,we do not limit it to a specific interval,so the generalization performance of cGAN-PU is greatly improved.When using the absolute phase as Ground Truth,we can choose to correspond the wrapped phase with noise to the absolute phase without noise.In this way,the trained cGAN can realize phase filtering and PU in one step.When the coherence is low or there are too many phase discontinuities,cGAN-PU can achieve better unwrapping accuracy and efficiency.(3)In this study,we propose an InSAR phase unwrapping algorithm via GAUNet(global attention U-Net).To some extent,the class imbalance problem in ambiguity gradient estimation can be solved,and the universality of PU method based on deep learning is enhanced.Firstly,a multi-source multi-noise level dataset is simulated to improve the training effect and generalization ability of the network.Secondly,the global attention mechanism is introduced into the classical U-Net structure to deal with the class imbalance in ambiguity gradient estimation.Furthermore,the quality map is fed into the network together with the wrapped phase for the training process,and it is also used for the construction of the loss function.GAUNet-PU realizes PU in two steps:(1)Estimate the horizontal and vertical ambiguity via GAUNet,respectively;(2)Reconstruct the absolute phase using the estimated ambiguity gradients.GAUNet-PU shows state-of-art unwrapping effects and excellent generalization ability on both simulated data and real data(including topographic interferograms and differential interferograms).The results show that the quality map can provide effective guidance for GAUNet.Compared with traditional methods,GAUNet-PU shows significant advantages in scenes such as surface discontinuous areas or incoherent areas.Compared cGAN-PU and GAUNet-PU,the former has higher unwrapping efficiency and excellent performance in extreme unwrapping scenes,but its accuracy is still not comparable to the traditional methods when the data quality is high,so it is suitable to be used as a supplement to the traditional methods.The latter is more universal and can be competent for most of the unwrapping tasks in extreme scenes such as excessive phase gradient,so it can replace the traditional InSAR PU method.