Research On Deep Learning Algorithms For InSAR Phase Unwrapping

Interferometric Synthetic Aperture Radar(InSAR)is an important technology for obtaining geophysical information about the Earth,with phase unwrapping being a crucial step in the process.In recent years,with the continuous development of deep learning technologies and the unique characteristics of phase unwrapping tasks,deep learning has been increasingly applied to phase unwrapping,giving rise to various high-performance algorithms.However,each algorithm has its own advantages and disadvantages,and a balance must be struck between unwrapping accuracy,efficiency,and fringe consistency.In recent years,research on deep learning-based phase unwrapping algorithms has become a hotspot.Based on a summary of the advantages and disadvantages of various algorithms,this paper conducts research on InSAR phase unwrapping deep learning algorithms,and the main work and innovations can be divided into two parts:In the first part,an interpretable neural network for large-scale phase unwrapping is proposed.Current end-to-end one-step phase unwrapping algorithms struggle to ensure the preservation of unwrapped details,resulting in very blurry mountain textures and unclear results.Additionally,the insufficient receptive field of convolutional neural networks has a significant negative impact on large-scale phase unwrapping,and the lack of global features leads to non-uniform errors in the unwrapping results.To address these two issues,this paper designs a hybrid convolutional block that retains multi-scale features while obtaining an adequate receptive field.Furthermore,a multi-scale structural similarity loss function is employed to ensure the accuracy of unwrapping results.The performance of various methods was validated based on simulation data,and the generalization ability of the proposed method was verified on real-world data.Finally,this paper attempts to analyze the unwrapping principle of the proposed neural networkbased method by combining the basic theory of phase unwrapping,illustrating that unwrapping is a gradual transition process from fringe lines and phase gradients to the distribution patterns of fringe orders.This provides a basis for future improvements and design of neural networks.Experiments show that the proposed method has superior performance,effectively addressing the two major issues of preserving phase details and large errors in end-to-end networks during large-scale phase unwrapping,and possesses a certain degree of interpretability.In the second part,a phase unwrapping method based on fringe line detection is proposed.By combining deep learning with the basic theory of phase unwrapping,a fringe line detection algorithm based on deep learning is designed.The fringe lines are then repaired using an undirected graph and the principle of cyclic integration,and unwrapping is performed according to the repaired fringe lines.Finally,a neural network is used to identify and correct error points in the unwrapping results.The accuracy of each step is validated using simulation data.Ultimately,the effectiveness of the algorithm is verified using real-world data.The advantages of the proposed method lie in its ability to ensure fringe consistency and simplify the learning content by having a relatively simple learning objective for the neural network,resulting in strong generalization capabilities.