Interferometric synthetic aperture radar (InSAR) is an advanced radar remote sensing technology widely used in geodesy, mainly used for terrain reconstruction and surface deformation estimation. The technology in the past 30 years has developed rapidly, one of the fastest growing research direction is wrapped phase solution (PU), it is a kind of recovering from observations of wrapped phase that can directly reflect the physical characteristics of target absolute phase (that is, the elimination of two PI fuzzy degrees) technology, interferometric synthetic aperture radar signal processing is one of the most critical step in, Determines the final quality of the topographic reconstruction and surface deformation products. Artificial intelligence is a branch of computer science, since the 1970s has been known as one of the world's three cutting-edge technologies, in the past three decades has also obtained rapid development, has been widely used in many disciplines, and has achieved many fruitful results. Therefore, how to introduce artificial intelligence technology into phase unwrapping algorithm has become a hot research direction in the field of interferometric synthetic aperture radar in recent years.

Traditional single baseline phase unwrapping (SBPU, also known as traditional two dimensional PU) is an ill-posed inverse problem. This means that no matter how skilled a PU algorithm designer is, it is impossible to design a PU algorithm that can correctly handle every case, i.e., for any given SBPU method, we can always generate its unsolvable input entanglement interference phase in polynomial time. Therefore, we can only design the best SBPU algorithm in the statistical sense, but we cannot make the SBPU algorithm designed to be the best for every interferogram. In this case, it is of great significance to accumulate SBPU processing experience from different research cases for algorithm design. Artificial intelligence technology (especially deep learning technology) provides a potential framework for us to accumulate data processing experience. A large number of valuable data from different InSAR sensors enable the phase unwrapping method based on deep learning technology to exceed the traditional model-based phase unwrapping method.

AI based SBPU approaches can be roughly divided into the following two categories: path tracking PU using AI and global PU using AI, which have completely different PU mechanisms and strategies. The PATH tracking PU technology based on AI regards SBPU as a local integral problem, which shows very promising results in many applications. Based on two mature concepts in path tracking PU (residual handicap and branch tangent), the path tracking PU methods using AI can be divided into two categories: residual handicap estimation based on AI and branch tangent placement based on AI. The first attempt to use AI for residual estimation was in a neural network for detecting phase gradient information proposed by Schwartzkopf et al in early 2000. Since then, however, there has been no research on phase gradient estimation using neural networks for nearly 20 years. Until 2020, another researcher finally proposed a deep convolutional neural network (DCNNs) -phase gradient neural network (PGNet)[1] for phase gradient estimation, as shown in FIG. 1, which transformed the estimation problem of phase gradient into a root

Figure 1. PGNet architecture

Where, the input is the winding interference phase data, the output is the vertical phase gradient and horizontal phase gradient data, Conv is the convolution layer, BN is the batch normalization layer, ReLu is the linear correction unit.

Two DCNNs are used to predict the vertical and horizontal phase gradients for the image segmentation problem of classifying interferogram according to the polarity of the residual handicap (i.e. 0, +1 and -1). Some researchers propose a similar phase gradient recognition method, and the coherence coefficient map is also input into the neural network, so as to provide the neural network with a local indication of whether there is a residual handicap. The first step is preprocessing. In this stage, the system firstly extracts the clustering features of the interferogram according to the distribution of residual handicaps, and then obtains the clustered sub-interferogram by tiling strategy, and finally processes them independently of each other. In the second step, traditional artificial intelligence techniques such as genetic algorithm, simulated annealing algorithm and ant colony algorithm were used to solve the placement problem of branch tangents. Some of these methods are robust, but they carry heavy computational burdens. Recently, a deep learning based branch tangent placement method was proposed, which transformed the residual handicap balance problem into a semantic segmentation problem, namely the branch tangent network (BCNet) problem. One limitation of these AI path tracking PU approaches is that they only apply AI techniques to assist one major step in the phase unwrapping process chain. For example, PGNet is only used for phase gradient estimation and BCNet is only used for branch tangent placement. In view of this, a second intelligent SBPU method is proposed: AI based global PU method. The development of the method has also gone through two stages. The first stage of the method is based on global optimization, by using AI techniques similar to the traditional minimum norm PU method to minimize the energy function, so that the unwrapped phase image can be fitted with the wound phase image. In the second stage, deep learning framework is directly used to transform the input winding interference phase diagram into unwinding phase. From the number of processing steps, the global PU methods using deep learning technology can be roughly divided into two groups: AI based one-step PU method and two-step PU method. The basic idea of AI based one-step PU method is to design a discrete cosine neural network, which can directly derive the unwrapped phase from the entangled interferogram through regression, but the disadvantage of this method is that the phase fringes of the input may be excessively filtered destructively. The two-step PU method converts the phase unwrapping problem into the semantic segmentation problem in the first step and the correction problem of the phase jump pixels in the second step. Compared with the one-step PU method based on AI, the main advantage of this method is that the fringe consistency can be guaranteed accurately through post-processing. However, this method also has disadvantages, that is, it can not use the quality map in the traditional PU method, and is not suitable for directly processing the large size interferogram.

Multi-baseline phase unentangling (MBPU) is an advanced technique that can recover the only absolute phase from multiple interferogram of the same scene, which can refine the traditional SBPU problem from ill-posed problem to well-posed problem. However, because the mathematical theory of MBPU method is based on Chinese remainder theorem, and the neighborhood information of pixel is not fully utilized, the noise robustness is not strong, and it is often difficult to get ideal PU results in the practical application process. Artificial intelligence and machine learning algorithms can simultaneously obtain millions of interferences with and without noise, and can learn what noise is and how best to eliminate it through autonomous learning, so as to solve the problem of low noise robustness encountered in MBPU. In this case, some researchers have begun to use artificial intelligence techniques and machine learning methods to improve the noise robustness of MBPU.

Early MBPU methods using machine learning technology can be roughly divided into three categories: maximum likelihood (ML), maximum posteriori (MAP) and cluster analysis (CA). Later, the AI-based unsupervised strategy was applied to the MBPU model, which requires neither any training images nor corresponding ground real data in advance. In this case, an AI-based MBPU can effectively eliminate the hassle of building complex training data sets and producing near-real-time prediction results without any off-line training process. However, in general, the learning accuracy and reliability of AI-based machine learning strategies in interference graphs without the help of training data sets are lower than those of AI-based machine learning strategies with the help of training data. In addition, the performance of the AI-based MBPU method is directly related to the selection of the standard baseline length in the interferometric SAR system, which is similar to the traditional model-based SBPU method. Therefore, in order to maintain optimal unwrapping results, the design problem of optimal baseline still needs to be considered in THE AI-based MBPU method.

To sum up, the current research results of AI phase unwrapping are encouraging and promising. However, some research areas of AI-based PU are relatively immature and are expected to achieve rapid development in the next few years.

[1] Zhou L , Yu H ,  Lan Y . Deep Convolutional Neural Network-Based Robust Phase Gradient Estimation for Two-Dimensional Phase Unwrapping  IEEE Transactions on Geoscience and Remote Sensing, 2020, 58(7): 4653 -- 4665.