| As a key step of interferometry synthetic aperture radar(InSAR)data processing,phase unwrapping plays a decisive role in the digital elevation reconstruction.The traditional algorithm has the problem that the accuracy and efficiency of phase unwrapping cannot be taken into account.Deep learning has the strong ability of feature representation and the ability to solve nonlinear inverse problems,and the trained network model is usually very efficient in efficient in performing the specified tasks,which makes it have potential advantages in the field of phase unwrapping for interferograms.In this paper,phase unwrapping for interferograms is studied under the framework of deep learning theory and technology.The main research contents are as follows:(1)The principle of InSAR technology and classical phase unwrapping algorithms,including branch-cut method,the quality-guided phase unwrapping method,iterative least square phase unwrapping method,minimum cost flow method,unscented Kalman filtering phase unwrapping method,are studied,and their performances are demonstrated through the experiments of phase unwrapping for simulated and experimental interferograms.(2)A deep learning phase unwrapping method based on adaptive noise-level evaluation of interferograms,is proposed,where a network using U-Net3+ as the skeleton is built to establish the nonlinear mapping relationship between the wrapped phase of interferograms and its unwrapped phase by combining full-scale skip connection and residual network.Firstly,a noise level evaluation system of interferograms is designed by fusing the quality maps and the phase residues of the interferograms,to divide noisy interferograms into different groups based on their estimated noise levels.Secondly,the network is trained using training sets with different noise levels,to obtain the trained networks suitable for phase unwrapping of the interferograms with different noise levels.Finally,the noise levels of the interferograms to be unwrapped are estimated,and then are unwrapped by the trained networks corresponding to the same noise levels as those of the interferograms to be unwrapped.The results obtained with different types of interferograms,including synthetic interferograms and experimental measured interferograms,demonstrate the effectiveness of the proposed algorithm,and show this method can obtain robust solutions from noisy wrapped phase images,with very popular time consumption.(3)A phase unwrapping method based on deep learning semantic segmentation is proposed.Firstly,the initial unwrapped phase of the interferograms is obtained using the multi-scale skip connection network to extract the phase-wrap-numbers of the interferograms.Secondly,the phase discontinuity is corrected by using the "minimum discontinuity theory" to improve the accuracy of phase unwrapping for interferograms;Finally,the phase unwrapping for large-scale interferograms is explored,and the phase unwrapping of large scale interferograms with noise is demonstrated by combining Kalman filtering phase unwrapping algorithm. |
