A phase unwrapping method and apparatus, system, storage medium
By combining the construction of U-NetHD network and L1 norm unwrapping, the problem of insufficient phase unwrapping accuracy in low coherence and large gradient regions in InSAR data processing was solved, achieving high-precision unwrapping effect and improving the quality of InSAR products.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA UNIV OF MINING & TECH
- Filing Date
- 2024-06-19
- Publication Date
- 2026-06-26
AI Technical Summary
Existing technologies have insufficient phase unwrapping accuracy in InSAR data processing in low coherence and high gradient regions, making it difficult to obtain high-precision unwrapping results, especially in regions with dense fringes and high noise, where the unwrapping effect is poor.
A phase unwrapping method based on a deep learning network model is adopted. By constructing a U-NetHD absolute phase generation network, the initial unwrapped phase is generated through training. The residual phase is then unwrapped by combining the L1 norm unwrapping method. Finally, a high-precision unwrapping result is obtained by summing the phases.
It improves the accuracy of phase unwrapping, especially in low coherence and high gradient regions, achieving high-precision unwrapping results and enhancing the quality of InSAR products.
Smart Images
Figure CN118795476B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of synthetic aperture radar interferometric data processing technology, and particularly relates to a phase unwrapping method, device, system, and storage medium based on a deep learning network model. Background Technology
[0002] InSAR technology is an active microwave remote sensing technology that combines radar imaging and interferometry. It mainly uses two (or more) SAR images to interferometrically obtain ground information. Thanks to its strong penetration ability and all-weather operation capability, it is now widely used in tasks such as DEM acquisition and surface deformation monitoring.
[0003] InSAR systems use two SAR images with a certain baseline length to perform differential analysis to obtain interferograms. Due to the limitations of InSAR systems, the interferometric phase is a wrapped phase in the range [-π, π]. The process of adding ambiguity numbers to the wrapped phase to recover the absolute phase is called phase unwrapping. Phase unwrapping is one of the key steps in InSAR data processing, and it is not only a major source of error but also a difficult aspect of InSAR data processing. Therefore, the accuracy of phase unwrapping determines the quality of InSAR products.
[0004] Currently, phase unwrapping can be mainly divided into three categories based on methods and techniques. The first category consists of path-tracking algorithms developed from the branch-cutting method. These algorithms primarily identify residual points and rationally set branch-cutting lines or use quality maps as guidance to ensure the unwrapping path avoids areas with large errors, thus preventing error propagation. However, in areas with high noise and phase discontinuities, dense branch-cutting line placement can easily form loops, leading to unwrapping vulnerabilities. Unlike the path-tracking method, the minimum norm method is a path-independent global optimization algorithm that transforms the phase unwrapping problem into solving the optimal solution to a system of nonlinear equations. This type of algorithm has better stability than the path-tracking method and can avoid local error propagation. The minimum norm unwrapping algorithm has the characteristics of low computational cost and robust numerical computation, making it the most widely accepted and applied algorithm at present. However, this type of algorithm still cannot prevent error propagation; when interference fringes are dense or coherence is low, the quality of phase unwrapping is difficult to guarantee. With the development of deep learning technology and multidisciplinary integration, deep learning technology is also widely used in the field of phase unwrapping. Deep learning techniques can quickly and accurately obtain unwrapping results, but their accuracy in densely striped regions and low-coherence regions still needs improvement. They are prone to phase loss and downsampling issues, and neural networks are highly dependent on high-quality and diverse samples, which significantly limits the performance of deep learning phase unwrapping algorithms. Therefore, obtaining high-precision phase unwrapping results in high-gradient regions has become an important research task. Summary of the Invention
[0005] The technical problem to be solved by the present invention is to provide a phase unwrapping method, apparatus, system, and storage medium based on a deep learning network model, which can solve the problem of difficulty in obtaining terrain in low coherence and large gradient regions, and effectively improve the phase unwrapping accuracy.
[0006] To achieve the above objectives, the present invention adopts the following technical solution:
[0007] A phase unwrapping method, comprising:
[0008] Step S1: Construct a U-NetHD absolute phase generation network that directly maps the entangled phase to the unentangled phase;
[0009] Step S2: Construct an unwrapped dataset that directly maps the wrapped phase to the unwrapped phase, train the absolute phase generation network, and obtain the absolute phase generation model;
[0010] Step S3: Input the interferogram into the trained absolute phase generation model to generate the initial unwrapped phase;
[0011] Step S4: Rewrap the initial unwrapped phase to obtain the rewrapped phase, and multiply it by the original interference phase to obtain the residual phase;
[0012] Step S5: Unwrap the residual phase to obtain the unwrapped residual phase. Summate the initial unwrapped phase and the unwrapped residual phase to obtain the final unwrapped phase.
[0013] Preferably, in step S5, the residual phase is unwrapped using the L1 norm unwrapping method to obtain the residual unwrapped phase, and the initial unwrapped phase and the residual unwrapped phase are summed to obtain the final unwrapped phase.
[0014] Preferably, step S2 includes:
[0015] S2.1. Radar encoding is performed on the DEM in the geographic coordinate system to obtain the DEM in the radar coordinate system;
[0016] S2.2. Based on the imaging geometric parameters, perform absolute phase simulation on the DEM transformed into the radar coordinate system to obtain the true absolute phase;
[0017] S2.3 Rewound the true absolute phase to obtain a noise-free wound phase;
[0018] S2.4 Add noise to the noise-free absolute phase to obtain a noisy wrapped phase;
[0019] S2.5. Clip the true absolute phase and the noisy entangled phase to a size of 256×256 to construct the untangling training dataset;
[0020] S2.6. Train the absolute phase generation network to obtain the absolute phase generation model.
[0021] The present invention also provides a phase unwrapping device, comprising:
[0022] The first processing module is used to construct the U-NetHD absolute phase generation network for the direct mapping from the wrapped phase to the unwrapped phase;
[0023] The second processing module is used to construct an unwrapped dataset that directly maps the wrapped phase to the unwrapped phase, and to train the absolute phase generation network to obtain the absolute phase generation model.
[0024] The third processing module is used to input the interferogram into the trained absolute phase generation model to generate the initial unwrapped phase.
[0025] The fourth processing module is used to rewrap the initial unwrapped phase to obtain the rewrapped phase, and then multiply it by the original interference phase to obtain the residual phase.
[0026] The fifth processing module is used to unwrap the residual phase to obtain the unwrapped residual phase, and to sum the initial unwrapped phase and the unwrapped residual phase to obtain the final unwrapped phase.
[0027] Preferably, the fifth processing module unwraps the residual phase using the L1 norm unwrapping method to obtain the residual unwrapped phase, and sums the initial unwrapped phase and the residual unwrapped phase to obtain the final unwrapped phase.
[0028] Preferably, the second processing module simulates the absolute phase using DEM data, constructs an absolute phase dataset, and then performs de-wrapping and noise addition operations on the absolute phase to construct a wrapped phase dataset; the unwrapped dataset is used to train the U-NetHD network to obtain an absolute phase generation model.
[0029] The present invention also provides a phase unwrapping system, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program executes a phase unwrapping method when executed by the processor.
[0030] The present invention also provides a storage medium storing a computer program, which executes a phase unwrapping method when running.
[0031] This invention utilizes deep learning to divide the unwrapping of large gradient regions into two stages. First, an initial unwrapped phase is obtained through an absolute phase generation network. Then, the residual phase is obtained by conjugate multiplication of the original interference phase and the initial unwrapped phase. Finally, L... 1The norm-based phase unwrapping method performs a second unwrapping step on the residual phase to obtain the residual unwrapped phase. The final unwrapped phase is obtained by summing the initial unwrapped phase obtained in the first unwrapping step and the residual unwrapped phase obtained in the second unwrapping step. Compared with other existing conventional phase unwrapping methods, this invention can obtain high-precision unwrapping results from regions with large gradients, and the method has better robustness of the unwrapping model, effectively improving the accuracy of the final InSAR product. Attached Figure Description
[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0033] Figure 1 This is a flowchart of the phase unwrapping method according to an embodiment of the present invention;
[0034] Figure 2 This is a schematic diagram of the U-NetHD network structure according to an embodiment of the present invention;
[0035] Figure 3 This is a schematic diagram of the residual network structure according to an embodiment of the present invention;
[0036] Figure 4 A schematic diagram illustrating the fabrication of the untangled dataset in an embodiment of the present invention;
[0037] Figure 5 This is a schematic diagram of the untangled dataset according to an embodiment of the present invention;
[0038] Figure 6 This is a schematic diagram of the actual interference phase data used in this embodiment of the invention;
[0039] Figure 7 This is a schematic diagram of the actual interference phase data used in this embodiment of the invention. Detailed Implementation
[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0041] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0042] Example 1:
[0043] like Figure 1 As shown, this embodiment of the invention provides a phase unwrapping method based on a deep learning network, comprising the following steps:
[0044] S1. Construct a U-NetHD absolute phase generation network that directly maps the entangled phase to the unentangled phase.
[0045] This invention uses the U-Net network as the basic model and constructs a high-resolution absolute phase generator using two residual networks. The generator consists of a global generation network (module two) and local enhancement networks (module one and module three). The network structure is as follows: Figure 2 As shown.
[0046] Module 1 consists of an adaptive average pooling layer to reduce the data dimensionality before it is passed to Module 2, and two convolutional modules with different kernel sizes, mainly used for feature extraction and downsampling, and the feature maps are passed to Module 3.
[0047] Module 2: This is a U-shaped structure without skip connections. The encoder has five layers. The first layer is a convolutional module with a kernel size of 7. The remaining layers consist of convolutional modules with a kernel size of 3, a normalization module, and a ReLU activation function. The encoder downsamples the 1×128×128 wrapped phase output from Module 1 into a 1024×8×8 feature map. After nine residual modules, the decoder also has five layers. The first four layers consist of deconvolutional modules, a normalization module, and a ReLU activation function. To avoid overfitting, regularization modules are added to the second to fourth layers. The last layer is a convolutional module with a kernel size of 7. The decoder upsamples the 1024×8×8 feature map into a low-resolution 1×128×128 absolute phase.
[0048] Module 3 consists of three residual modules, one deconvolution module, and one convolution module. The final output activation function uses the Tanh activation function, which compresses the range of the high-resolution absolute phase values to [-1, 1].
[0049] Residual module: such as Figure 3 As shown, it consists of two 3×3 convolutional modules, a normalization module, and a ReLU activation function. The introduction of "shortcut" connections directly connects the input data to the output, forming residual connections. This helps to alleviate the gradient vanishing and gradient exploding problems that occur in deep neural networks during training.
[0050] S2. Using DEM data, construct an unwrapped dataset that directly maps entangled phases to unwrapped phases. Train the absolute phase generation network to obtain the absolute phase generation model. The unwrapped dataset creation process is as follows: Figure 4 As shown in the diagram, the untangled dataset is illustrated below. Figure 5 As shown.
[0051] Furthermore, step S2 specifically includes:
[0052] S2.1. Radar encoding is performed on the DEM in the geographic coordinate system to obtain the DEM in the radar coordinate system;
[0053] S2.2. Based on the imaging geometric parameters, perform absolute phase simulation on the DEM transformed into the radar coordinate system to obtain the true absolute phase;
[0054] S2.3 Rewound the true absolute phase to obtain a noise-free wound phase;
[0055] S2.4 Add noise to the noise-free absolute phase to obtain a noisy wrapped phase;
[0056] S2.5. Clip the true absolute phase and the noisy entangled phase to a size of 256×256 to construct the untangling training dataset;
[0057] S2.6. Train the absolute phase generation network to obtain the absolute phase generation model.
[0058] During network training, the input is a noisy, tangled phase, and the output is the true absolute phase; the loss function combining mean square error and mean absolute error is used as the final objective during training.
[0059] Mean Absolute Error (MAE) minimizes the difference between the generated image and the real image, enabling the generator network to produce a more realistic absolute phase. The formula for calculating MAE is as follows:
[0060]
[0061] in, To generate absolute phase, ψ i It is a true absolute phase.
[0062] Mean Square Error (MSE) is a loss function commonly used in regression problems. It primarily measures the average squared error between the predicted and true values. Using it as a loss function to optimize network parameters minimizes the pixel difference between the generated absolute phase and the true absolute phase. The formula for calculating MSE is as follows:
[0063]
[0064] in, To generate absolute phase, ψ iIt is a true absolute phase.
[0065] This invention combines MAE and MSE by weighted averaging to form a combined loss function. MSE is sensitive to outliers, while MAE is not. Combining them can mitigate the excessive penalty imposed by MSE on outliers to some extent, helping to stabilize the training process and improve model convergence. The total loss is defined as follows:
[0066] L Total =λL MAE +(1-λ)L MSE
[0067] Where λ is a weighting parameter used to balance the proportion of the two loss functions in the total loss.
[0068] S3. Input the interferogram into the trained absolute phase generation model to generate the initial unwrapped phase.
[0069] S4. Rewrap the initially unwrapped phase, and then perform a difference between the original interferogram and the rewrapped phase map by conjugate multiplication, thereby sparsifying the interference fringes. The expression is as follows:
[0070]
[0071] in, Let ψ represent the residual phase, ψ represent the original interference phase, and ψ1 represent the initial unwrapping phase.
[0072] S5, via L 1 The norm-based phase unwrapping algorithm unwrappes the residual phase. Theoretically, the wrapped phase difference between adjacent pixels should be equal to the absolute phase difference. However, due to factors such as noise, the wrapped phase difference and the absolute phase difference are often not equal. 1 Minimum norm phase unwrapping is achieved by minimizing the difference between the wrapped phase difference and the absolute phase difference between adjacent pixels. 1 The optimization model for the norm is:
[0073]
[0074] Where, ψ i+1,j and ψ i,j+1 These represent the absolute phases of pixels (i+1,j) and (i,j+1) on the interferogram, respectively. The final unwrapped phase is obtained by summing the initial unwrapped phase and the residual unwrapped phase, and its expression is as follows:
[0075] ψ(s) = ψ1(s) + ψ2(s)
[0076] Where ψ(s) represents the final unwrapped phase, ψ1(s) represents the initial unwrapped phase, and ψ2(s) represents the residual unwrapped phase.
[0077] To verify the technical effects of the embodiments of the present invention, unwrapping experiments were conducted on the same interferogram using Branch-cut, L1 norm phase unwrapping method, MCF, SNAPHU, phase unwrapping method based on U-Net network, phase unwrapping method based on U-NetHD network, and the phase unwrapping method of the present invention. The data used in the experiments were real data from the LuTan-1 satellite. Figure 6 (a) is the actual interference pattern in size 256×256. Figure 6 (b) represents the corresponding absolute phase. Figure 6 (c) and Figure 6 (j) shows the untangling results and corresponding error diagrams of the branch cutting method. Figure 6 (d) and Figure 6 (k) shows the unwrapping results and corresponding error diagrams of the L1 norm phase unwrapping method. Figure 6 (e) and Figure 6 (l) shows the untangling results and corresponding error diagrams for the minimum cost flow method. Figure 6 (f) and Figure 6 (m) represents the untangling results and corresponding error diagrams of the statistical cost flow method. Figure 6 (g) and Figure 6 (n) represent the unwrapping results and corresponding error maps based on the U-Net network. Figure 6 (h) and Figure 6 (o) shows the unwrapping results and corresponding error diagrams based on the U-NetHD network. Figure 6 (i) and Figure 6(p) shows the unwrapping results and corresponding error maps of the TSPU algorithm. The unwrapping results show that due to discontinuous regions and significant noise in the interferogram, the branch tangents are densely arranged, resulting in severe fragmentation of the unwrapping results and regional unwrapping errors. For the L1 norm phase unwrapping algorithm and the mcf algorithm, although both are path-independent optimization algorithms and suppress noise to some extent, phase jumps occur in dense fringe regions and phase aliasing regions. While the statistical cost flow algorithm does not show severe phase jumps, it still exhibits significant unwrapping errors in regions with phase aliasing where clear fringes cannot be formed. The one-step phase unwrapping algorithm based on the U-Net network can overcome the phase continuity assumption and directly obtain high-precision unwrapping results from the interferogram. The absolute phase generated by the one-step phase unwrapping algorithm based on the U-NeHDt network has a significantly improved resolution compared to the U-Net network, producing a more refined absolute phase. This invention utilizes a deep neural network to reduce the phase gradient and extract residual phase information by post-processing the unwrapping results. This method achieves higher unwrapping accuracy than other algorithms, with an RMSE of only 1.8714 rad.
[0078] Figure 7 (a) A real interference pattern of the same size (256×256). Figure 7 (b) represents the corresponding absolute phase. Figure 7 (c) and Figure 7 (j) shows the untangling results and corresponding error diagrams of the branch cutting method. Figure 7 (d) and Figure 7 (k) shows the unwrapping results and corresponding error diagrams of the L1 norm phase unwrapping method. Figure 7 (e) and Figure 7 (l) shows the untangling results and corresponding error diagrams for the minimum cost flow method. Figure 7 (f) and Figure 7 (m) represents the untangling results and corresponding error diagrams of the statistical cost flow method. Figure 7 (g) and Figure 7 (n) represent the unwrapping results and corresponding error maps based on the U-Net network. Figure 7 (h) and Figure 7 (o) shows the unwrapping results and corresponding error diagrams based on the U-NetHD network. Figure 7 (h) and Figure 7(p) shows the unwrapping results and corresponding error maps of the TSPU algorithm. The unwrapping results show that the branch-cutting method is severely affected by noise, resulting in numerous unwrapped islands. While the L1 norm phase unwrapping algorithm, minimum cost flow method, and statistical cost flow algorithm have good spatial continuity, they still exhibit unwrapping jump regions of varying areas. Although data 1 and data 2 are different regions of the same interferogram, the performance of the one-step phase unwrapping algorithm based on U-Net and U-NetHD networks is significantly lower in data 2 than in data 1, with a noticeable phase loss in the unwrapping results. In this embodiment, the L1 norm phase unwrapping algorithm is combined with the one-step phase unwrapping algorithm based on U-NetHD. Although the performance of the neural network model decreases and its generalization ability is insufficient, the traditional phase unwrapping algorithm effectively compensates for this deficiency, maintaining good spatial continuity and reducing the RMSE from 10.8599 rad to 2.0830 rad. These two experiments demonstrate that this embodiment not only improves the unwrapping accuracy of phase unwrapping in low coherence and high gradient regions but also enhances the generalization ability and robustness of the neural network.
[0079] Example 2:
[0080] The present invention also provides a phase unwrapping device based on a deep learning network, comprising:
[0081] The first processing module is used to construct the U-NetHD absolute phase generation network for the direct mapping from the wrapped phase to the unwrapped phase;
[0082] The second processing module is used to construct an unwrapped dataset that directly maps the wrapped phase to the unwrapped phase, and to train the absolute phase generation network to obtain the absolute phase generation model.
[0083] The third processing module is used to input the interferogram into the trained absolute phase generation model to generate the initial unwrapped phase.
[0084] The fourth processing module is used to rewrap the initial unwrapped phase to obtain the rewrapped phase, and then multiply it by the original interference phase to obtain the residual phase.
[0085] The fifth processing module is used to unwrap the residual phase to obtain the unwrapped residual phase, and to sum the initial unwrapped phase and the unwrapped residual phase to obtain the final unwrapped phase.
[0086] As one embodiment of the present invention, the fifth processing module unwraps the residual phase using the L1 norm unwrapping method to obtain the residual unwrapped phase, and sums the initial unwrapped phase and the residual unwrapped phase to obtain the final unwrapped phase.
[0087] As one embodiment of the present invention, the second processing module simulates the absolute phase using DEM data, constructs an absolute phase dataset, and then performs de-wrapping and noise addition operations on the absolute phase to construct a wrapped phase dataset; the unwrapped dataset is used to train the U-NetHD network to obtain an absolute phase generation model.
[0088] Example 3:
[0089] This invention also provides a phase unwrapping system based on a deep learning network, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program executes a phase unwrapping method based on a deep learning network when executed by the processor.
[0090] Example 4:
[0091] This invention also provides a storage medium storing a computer program that executes a phase unwrapping method based on a deep learning network during runtime.
[0092] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.
Claims
1. A phase unwrapping method, characterized in that, include: Step S1: Construct a U-NetHD absolute phase generation network that directly maps the entangled phase to the unentangled phase; Step S2: Construct an unwrapped dataset that directly maps the wrapped phase to the unwrapped phase, train the absolute phase generation network, and obtain the absolute phase generation model; Step S3: Input the interferogram into the trained absolute phase generation model to generate the initial unwrapped phase; Step S4: Rewrap the initial unwrapped phase to obtain the rewrapped phase, and multiply it by the original interference phase to obtain the residual phase; Step S5: Unwrap the residual phase to obtain the unwrapped residual phase. Summate the initial unwrapped phase and the unwrapped residual phase to obtain the final unwrapped phase. The U-NetHD absolute phase generation network uses the U-Net network as the basic model and constructs a high-resolution absolute phase generator using two residual networks. It consists of three modules: module 1, module 2, and module 3. Module 2 is the global generation network, and modules 1 and 3 form a local enhancement network. Module 1 consists of an adaptive average pooling layer to reduce the data dimensionality before feeding it into Module 2, and two convolutional modules with different kernel sizes for feature extraction and downsampling, and the feature map is fed into Module 3. Module 2: This is a U-shaped structure without skip connections. The encoder has five layers. The first layer is a convolutional module with a kernel size of 7. The remaining layers consist of convolutional modules with a kernel size of 3, a normalization module, and a ReLU activation function. The encoder downsamples the 1×128×128 wrapped phase output from Module 1 into a 1024×8×8 feature map. After nine residual modules, the decoder also has five layers. The first four layers consist of deconvolutional modules, a normalization module, and a ReLU activation function. Regularization modules are added to the second to fourth layers. The last layer is a convolutional module with a kernel size of 7. The decoder upsamples the 1024×8×8 feature map into a low-resolution 1×128×128 absolute phase. Module 3 consists of three residual modules, one deconvolution module, and one convolution module. The final output activation function uses the Tanh activation function to compress the range of the high-resolution absolute phase values to [-1, 1]. The residual module consists of two 3×3 convolutional modules, a normalization module, and a ReLU activation function. It introduces a shortcut connection to directly connect the input data to the output, forming a residual connection.
2. The phase unwrapping method as described in claim 1, characterized in that, In step S5, the residual phase is unwrapped using the L1 norm unwrapping method to obtain the residual unwrapped phase. The initial unwrapped phase and the residual unwrapped phase are summed to obtain the final unwrapped phase.
3. The phase unwrapping method as described in claim 2, characterized in that, Step S2 includes: S2.
1. Radar encoding is performed on the DEM in the geographic coordinate system to obtain the DEM in the radar coordinate system; S2.
2. Based on the imaging geometric parameters, perform absolute phase simulation on the DEM transformed into the radar coordinate system to obtain the true absolute phase; S2.3 Rewound the true absolute phase to obtain a noise-free wound phase; S2.4 Add noise to the noise-free absolute phase to obtain a noisy wrapped phase; S2.
5. Clip the true absolute phase and the noisy entangled phase to a size of 256×256 to construct the untangling training dataset; S2.
6. Train the absolute phase generation network to obtain the absolute phase generation model.
4. A phase unwrapping apparatus for implementing the phase unwrapping method of claim 1, characterized in that, include: The first processing module is used to construct the U-NetHD absolute phase generation network for the direct mapping from the wrapped phase to the unwrapped phase; The second processing module is used to construct an unwrapped dataset that directly maps the wrapped phase to the unwrapped phase, and to train the absolute phase generation network to obtain the absolute phase generation model. The third processing module is used to input the interferogram into the trained absolute phase generation model to generate the initial unwrapped phase. The fourth processing module is used to rewrap the initial unwrapped phase to obtain the rewrapped phase, and then multiply it by the original interference phase to obtain the residual phase. The fifth processing module is used to unwrap the residual phase to obtain the unwrapped residual phase, and to sum the initial unwrapped phase and the unwrapped residual phase to obtain the final unwrapped phase.
5. The phase unwrapping device as described in claim 4, characterized in that, The fifth processing module unwraps the residual phase using the L1 norm unwrapping method to obtain the residual unwrapped phase. The initial unwrapped phase and the residual unwrapped phase are then summed to obtain the final unwrapped phase.
6. The phase unwrapping apparatus as described in claim 5, characterized in that, The second processing module simulates the absolute phase using DEM data, constructs an absolute phase dataset, and then performs unwrapping and noise addition operations on the absolute phase to construct a wrapped phase dataset. The unwrapped dataset is used to train the U-NetHD network to obtain an absolute phase generation model.
7. A phase unwrapping system based on a deep learning network, comprising: A memory and a processor, wherein the memory stores a computer program executed by the processor, the computer program performing the phase unwrapping method as described in any one of claims 1-3 when executed by the processor.
8. A storage medium storing a computer program, which, when executed, performs the phase unwrapping method as described in any one of claims 1-3.