Two-step learning type phase unwrapping method based on residual optimization

Abstract

The invention discloses a two-step learning type phase unwrapping method based on residual optimization. The method comprises the following steps: constructing an absolute phase generation network; dEM data are acquired, and an unwrapped phase data set is constructed based on the DEM data; training an absolute phase generation network based on the unwrapped phase data set to obtain an absolute phase generation model; obtaining an initial unwrapping phase based on the original interferometric phase and an absolute phase generation model; a residual phase is obtained based on the initial unwrapping phase, the residual phase is unwrapped after being filtered, and an unwrapped residual phase is obtained; and obtaining a final unwrapping phase based on the initial unwrapping phase and the unwrapping residual phase. According to the method, a high-precision unwrapping result can be obtained from a large-gradient region, better unwrapping model robustness is achieved, unwrapping errors caused by noise can be weakened by slightly filtering the residual phase, and the precision of a final InSAR product is effectively improved.

Description

Technical Field

[0001] The invention belongs to the technical field of synthetic aperture radar interference data processing, and in particular relates to a two-step learning phase unwrapping method based on residual optimization. Background Art

[0002] Synthetic aperture radar interferometry (InSAR) technology, as an active microwave remote sensing method that integrates cutting-edge radar imaging and interferometry technology, has shown extraordinary application potential in the fields of digital elevation model (DEM) construction and surface deformation monitoring due to its excellent penetration ability and all-weather working ability. Its core lies in the use of at least two SAR images with spatial baseline differences for interferometric processing to extract key information such as ground deformation. However, the interferometric phase generated by the InSAR system is naturally entangled in the [-π,π] interval, so phase unwrapping processing must be performed to restore the continuous and true absolute phase. Phase unwrapping is not only a key link in the InSAR data processing process, but also the main source of error and technical difficulty. Its accuracy directly determines the quality and application value of the final InSAR product.

[0003] At present, phase unwrapping can be mainly divided into three categories according to methods and means. The first category is the path tracking algorithm developed from the branch cutting method. This type of algorithm mainly avoids the area with large errors and error transmission by identifying residual points, reasonably setting branch tangents or guiding according to the quality map. However, in areas with high noise and phase discontinuity, it is easy to form loops due to the dense arrangement of branch tangents, resulting in unwrapping loopholes. Unlike the path tracking method, the minimum norm method is a global optimization algorithm that is independent of the path, but converts the phase unwrapping problem into the optimal solution problem of solving a nonlinear system of equations. This type of algorithm has better stability than the path tracking method and can avoid local error transmission. The minimum norm unwrapping algorithm has the characteristics of small computational complexity and relatively robust numerical calculation. Therefore, this type of algorithm is also the most recognized and widely used algorithm at this stage. However, this type of algorithm still cannot prevent error transmission, the quality of phase unwrapping is difficult to guarantee, and the computational efficiency will drop significantly under high noise conditions. With the development of deep learning technology and multidisciplinary integration, deep learning technology has also been widely used in the field of phase unwrapping. Deep learning technology can be used to learn the nonlinear relationship between the entangled phase and the absolute phase. After training, the network can quickly and accurately obtain high-precision unwrapping results from the entangled phase. However, the unwrapping accuracy in areas with dense fringes and low coherence still needs to be improved. It is prone to problems such as unwrapping phase loss and downsampling. In addition, neural networks are highly dependent on high-quality samples and sample diversity, which greatly limits the performance of deep learning phase unwrapping algorithms. In view of the above challenges, improving the accuracy of phase unwrapping in high-noise and large-gradient areas has become a hot topic and difficulty in current research. Summary of the invention

[0004] In order to solve the above technical problems, the present invention proposes a two-step learning phase unwrapping method based on residual optimization to solve the problems existing in the above prior art.

[0005] To achieve the above object, the present invention provides a two-step learning phase unwrapping method based on residual optimization, comprising:

[0006] Constructing an absolute phase generation network; acquiring DEM data, and constructing an unwrapped phase data set based on the DEM data; training the absolute phase generation network based on the unwrapped phase data set to obtain an absolute phase generation model;

[0007] Obtaining an initial unwrapped phase based on the original interference phase and the absolute phase generation model; obtaining a residual phase based on the initial unwrapped phase, unwrapping the residual phase after filtering, and obtaining an unwrapped residual phase;

[0008] A final unwrapped phase is obtained based on the initial unwrapped phase and the unwrapped residual phase.

[0009] Optionally, a residual network is introduced into the U-Net network structure to obtain a RU-Net network model, and the absolute phase generation network is obtained based on the RU-Net network model.

[0010] Optionally, the U-Net network structure includes an encoder, a decoder and a jump connection, and the encoder and the decoder both include a convolution structure with a convolution kernel size of 4 and a step size of 2.

[0011] Optionally, the process of constructing an unwrapped phase dataset based on DEM data includes:

[0012] Radar encoding is performed on the DEM data to obtain the DEM data in the radar coordinate system. The DEM data in the radar coordinate system is simulated according to the imaging geometric parameters to obtain the absolute phase. The absolute phase is rewound and noise is added to obtain the noisy wrapped phase. The absolute phase and the noisy wrapped phase are cropped to the same size to obtain an unwrapped phase data set.

[0013] Optionally, the process of obtaining the residual phase based on the initial unwrapped phase includes:

[0014] The initial unwrapped phase is rewrapped, and the original interference phase and the rewrap phase image are differentiated by conjugate multiplication to obtain a residual phase.

[0015] Optionally, the residual phase is lightly filtered by a median filtering method with a window size of 5.

[0016] Optionally, an L1 norm unwrapping method is used to unwrap the filtered residual phase to obtain a residual unwrapped phase.

[0017] Optionally, the final unwrapped phase is obtained by summing the initial unwrapped phase and the residual unwrapped phase, and the expression is as follows:

[0018] ψ(s)=ψ1(s)+ψ2(s)

[0019] Among them, ψ(s) represents the final unwrapped phase, ψ1(s) represents the initial unwrapped phase, and ψ2(s) represents the residual unwrapped phase.

[0020] Compared with the prior art, the present invention has the following advantages and technical effects:

[0021] The present invention adopts a two-step learning phase unwrapping method with residual optimization, and utilizes the characteristic of phase loss in the one-step unwrapping method based on deep learning, and divides the phase unwrapping in the large gradient region into two stages. First, the initial unwrapping phase of the first step is obtained by the absolute phase generation network; the residual phase is obtained by conjugate multiplication of the original interference phase and the initial unwrapping phase, but the two-step phase unwrapping method does not process the noise, so the residual phase is slightly filtered; the residual phase is unwrapped in the second step using the L1 norm phase unwrapping method to obtain the residual unwrapped phase; the final unwrapped phase is obtained by summing the initial unwrapping phase obtained by the first step unwrapping and the residual unwrapping phase obtained by the second step unwrapping. The present invention can obtain high-precision unwrapping results from large gradient regions, and has better unwrapping model robustness. Slight filtering of the residual phase can also reduce the unwrapping error caused by noise, effectively improving the accuracy of the final InSAR product. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0023] Figure 1 is a flow chart of a method according to an embodiment of the present invention;

[0024] Figure 2 A schematic diagram of a method according to an embodiment of the present invention;

[0025] Figure 3 This is a schematic diagram of the absolute phase generation network structure according to an embodiment of the present invention;

[0026] Figure 4 Schematic diagram of the residual network structure of an embodiment of the present invention;

[0027] Figure 5 A schematic diagram of disentangled data set preparation according to an embodiment of the present invention;

[0028] Figure 6 A schematic diagram of a disentangled data set according to an embodiment of the present invention;

[0029] Figure 7 It is a schematic diagram of real interference phase data according to an embodiment of the present invention. DETAILED DESCRIPTION

[0030] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0031] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0032] Embodiment 1

[0033] like Figure 1-2 As shown, this embodiment provides a two-step learning phase unwrapping method based on residual optimization, including:

[0034] S1, one-step phase disentanglement network construction based on RU-Net network;

[0035] Using the U-Net network as the basic model, a residual network is added in the encoding stage to construct an absolute phase generator. The network structure is as follows: Figure 3 shown.

[0036] The encoder consists of eight layers of networks. Each layer has a convolution module with a convolution kernel size of 4 and a stride of 2 for feature extraction. The number of convolution kernels in each layer is 32, 64, 128, 256, 256, 256, 256, 256, respectively. Except for the first and last layers, which do not have normalization modules, the remaining layers add normalization modules, activation functions, and residual modules after the convolution modules. The normalization module uses the BatchNorm2d method to achieve batch normalization of images. The activation function uses the ReLU activation function. A residual module is added to each layer to prevent overfitting and gradient explosion. The decoding stage is mainly used for feature extraction. The input image size is 1×256×256. After eight convolutions and downsampling, a 256×1×1 feature information will be obtained.

[0037] The decoder also has eight layers. Each layer consists of a deconvolution module with a convolution kernel size of 4 and a stride of 2 to restore the image resolution. The number of convolution kernels in each layer is 256, 256, 256, 256, 128, 64, 32, and 1 respectively. The first seven layers are composed of deconvolution modules, normalization modules, and ReLU activation functions. Dropout regularization modules are added to each layer, and the dropout rate is set to 0.5 to prevent the model from overfitting. The activation function of the output layer selects the Tanh activation function to compress the output range to [-1,1]. After continuous upsampling, the network can obtain the deep feature information of the image and output a 1×256×256 high-resolution image.

[0038] Residual module: such as Figure 4 As shown in the figure, it consists of two 3×3 convolution modules, a normalization module and a ReLU activation function. A "shortcut" connection is introduced to directly connect the input data to the output to form a residual connection, which helps to alleviate the gradient disappearance and gradient explosion problems that occur in deep neural networks during training.

[0039] S2. Using the existing DEM data, we construct an unwrapped data set that directly maps the entangled phase to the unwrapped phase, train the absolute phase generation network, and obtain the absolute phase generation model. The data set production process is as follows: Figure 5 As shown, specifically:

[0040] S2.1, radar encoding the DEM in the geographic coordinate system to obtain the DEM in the radar coordinate system;

[0041] S2.2, according to the imaging geometric parameters, perform absolute phase simulation on the DEM converted to the radar coordinate system to obtain the real absolute phase;

[0042] S2.3, rewinding the true absolute phase to obtain a noise-free winding phase;

[0043] S2.4, adding noise to the noise-free absolute phase to obtain the noisy winding phase;

[0044] S2.5, crop the true absolute phase and the noisy entangled phase to 256×256 size, and construct the disentangled training data set, such as Figure 6 As shown;

[0045] S2.6. Train the network to obtain the absolute phase generation model.

[0046] S3. Input the interference pattern into the trained absolute phase generation model to generate the initial disentangled phase. The mean absolute error (MAE) is used as the loss function during training. MAE enables the generation network to generate a more realistic absolute phase by minimizing the difference between the generated image and the real image. MAE is robust to outliers and will not cause a significant increase in loss value due to an outlier. It also helps to reduce the complexity of the model and improve generalization ability. The calculation formula of MAE is as follows:

[0047]

[0048] in, To generate the absolute phase, ψ i is the true absolute phase. When training the RU-Net network model, 8 images are input each time, the training rounds are set to 200 rounds, the learning rate is set to 0.0002, MAE is selected as the loss function, and the hyperparameter is set to 100.

[0049] S4, re-wrapping the initial unwrapped phase to obtain a re-wrapped phase, conjugate-multiplying it with the original interference phase to obtain a residual phase, and performing median filtering on the residual phase;

[0050] S4.1. The initial unwrapped phase is rewound, and the original interference pattern and the rewound phase pattern are differentiated by conjugate multiplication, so as to sparse the interference fringes. The expression is as follows:

[0051]

[0052] in, represents the residual phase, ψ represents the original interference phase, and ψ1 represents the initial unwrapped phase.

[0053] S4.2. Use the median filtering method with a window size of 5×5 to slightly filter the residual phase.

[0054] S5. Unwrap the residual phase using the L1 norm phase unwrapping algorithm. Theoretically, the entanglement phase difference between adjacent pixels is equal to the absolute phase difference, but due to the influence of factors such as noise, the entanglement phase difference is often not equal to the absolute phase difference. L1 minimum norm phase unwrapping achieves phase unwrapping by minimizing the difference between the entanglement phase difference and the absolute phase difference of adjacent pixels. The optimization model of the L1 norm is:

[0055]

[0056] Among them, ψ i+1,j and ψ i,j+1are the absolute phases of the (i+1, j) and (i, j+1) pixels on the interference pattern, respectively. The final unwrapped phase is obtained by summing the initial unwrapped phase and the residual unwrapped phase, and its expression is as follows:

[0057] ψ(s)=ψ1(s)+ψ2(s)

[0058] Among them, ψ(s) represents the final unwrapped phase, ψ1(s) represents the initial unwrapped phase, and ψ2(s) represents the residual unwrapped phase.

[0059] In order to verify the technical effect of the present invention, the branch cutting method, the phase unwrapping method based on the U-Net network, the phase unwrapping method based on the VUR-Net network and the phase unwrapping method of the present invention were used to carry out unwrapping processing experiments on the same interference pattern.

[0060] like Figure 7 As shown, the first line is two groups of entangled phase data and their corresponding reference absolute phases, the second line is the unwrapping results of the branch cutting method corresponding to the two groups of data and their corresponding error distribution diagrams, the third line is the unwrapping results of the one-step phase unwrapping algorithm based on the U-Net network corresponding to the two groups of data and their corresponding error distribution diagrams, the fourth line is the unwrapping results of the one-step phase unwrapping algorithm based on the VUR-Net network corresponding to the two groups of data and their corresponding error distribution diagrams, and the fifth line is the unwrapping results of the phase unwrapping algorithm of the present invention corresponding to the two groups of data and their corresponding error distribution diagrams. From the unwrapping results, it can be seen that the unwrapping results of the branch cutting method are significantly different from the reference absolute phase, and from the corresponding error distribution diagrams, it can be found that obvious unwrapping holes appear in high-noise areas. The unwrapping results of the one-step phase unwrapping algorithm based on the U-Net network and the VUR-Net network have better phase continuity, smoother unwrapped phases, and better noise suppression effects. However, due to the weak generalization ability of the one-step phase unwrapping model based on deep learning, its unwrapping results have unwrapping errors and phase losses in areas with denser stripes. The algorithm of the present invention uses a deep neural network to reduce the phase gradient and performs a slight filtering process on the residual phase to be unwrapped in the second step, which effectively suppresses the influence of noise on subsequent unwrapping. It not only improves the stability and generalization ability of the algorithm, but also has higher unwrapping accuracy than other algorithms. The RMSE in the two sets of data is reduced to 1.25rad and 1.30rad respectively.

[0061] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A two-step learning phase unwrapping method based on residual optimization, characterized in that: The following steps are involved: Construct an absolute phase generation network; Acquire DEM data, and construct an unwrapped phase data set based on the DEM data; train the absolute phase generation network based on the unwrapped phase data set to obtain an absolute phase generation model; Obtaining an initial unwrapped phase based on the original interference phase and the absolute phase generation model; Obtaining a residual phase based on an initial unwrapped phase, filtering the residual phase and then unwrapping it to obtain an unwrapped residual phase; A final unwrapped phase is obtained based on the initial unwrapped phase and the unwrapped residual phase.

2. The two-step learning phase unwrapping method based on residual optimization according to claim 1 is characterized in that: A residual network is introduced into the U-Net network structure to obtain a RU-Net network model, and the absolute phase generation network is obtained based on the RU-Net network model.

3. The two-step learning phase unwrapping method based on residual optimization according to claim 2 is characterized in that: The U-Net network structure includes an encoder, a decoder and a skip connection, and both the encoder and the decoder include a convolution structure with a convolution kernel size of 4 and a step size of 2.

4. The two-step learning phase unwrapping method based on residual optimization according to claim 1, characterized in that: The process of constructing an unwrapped phase dataset based on DEM data includes: Radar encoding is performed on the DEM data to obtain the DEM data in the radar coordinate system. The DEM data in the radar coordinate system is simulated according to the imaging geometric parameters to obtain the absolute phase. The absolute phase is rewound and noise is added to obtain the noisy wrapped phase. The absolute phase and the noisy wrapped phase are cropped to the same size to obtain an unwrapped phase data set.

5. The two-step learning phase unwrapping method based on residual optimization according to claim 1, characterized in that: The process of obtaining the residual phase based on the initial unwrapped phase includes: The initial unwrapped phase is rewrapped, and the original interference phase and the rewrap phase image are differentiated by conjugate multiplication to obtain a residual phase.

6. The two-step learning phase unwrapping method based on residual optimization according to claim 1, characterized in that: The residual phase is lightly filtered by a median filter method with a window size of 5.

7. The two-step learning phase unwrapping method based on residual optimization according to claim 1, characterized in that: The L1 norm unwrapping method is used to unwrap the filtered residual phase to obtain the residual unwrapped phase.

8. The two-step learning phase unwrapping method based on residual optimization according to claim 1, characterized in that: The final unwrapped phase is obtained by summing the initial unwrapped phase and the residual unwrapped phase. The expression is as follows: ψ(s)=ψ1(s)+ψ2(s) Among them, ψ(s) represents the final unwrapped phase, ψ1(s) represents the initial unwrapped phase, and ψ2(s) represents the residual unwrapped phase.

Images