InSAR phase unwrapping method based on multi-scale feature fusion denoising CNN network
By using a multi-scale feature fusion denoising CNN network, the problem of inaccurate unwrapping caused by noise interference in InSAR phase maps by traditional methods is solved, and efficient and accurate phase unwrapping is achieved in complex noise environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- TIANJIN UNIV
- Filing Date
- 2023-04-28
- Publication Date
- 2026-06-05
AI Technical Summary
Traditional phase unwrapping methods are not effective in InSAR phase images with high noise interference. In particular, the branch cutting method is prone to phase islands, the least squares method causes error propagation, the network flow method makes mistakes in unwrapping in low-quality areas, and the deep learning method has insufficient unwrapping accuracy and efficiency in complex noise environments.
A multi-scale feature fusion denoising CNN network is adopted, which utilizes the DnCNN framework to combine dilated convolution, deformable convolution and residual modules. It is trained on a simulated InSAR dataset to extract multi-scale feature information and perform fusion recovery to solve the noise interference problem.
It improves the accuracy and efficiency of InSAR phase unwrapping, reduces the impact of noise interference on the unwrapping results, and ensures accuracy and stability in complex environments.
Smart Images
Figure CN116664419B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of phase unwrapping technology, and more particularly to InSAR single-baseline phase unwrapping, and especially to an InSAR phase unwrapping method for a multi-scale feature fusion denoising CNN network. Background Technology
[0002] Traditional phase unwrapping methods are generally classified into three categories: (1) Path-tracking-based phase unwrapping algorithms, which integrate the phase gradients of adjacent pixels by selecting an appropriate integration path to achieve phase unwrapping; (2) Minimum norm-based phase unwrapping methods, which achieve phase unwrapping by minimizing the difference between the wrapped phase gradient and the true phase gradient; and (3) Network flow methods, which transform the phase unwrapping problem into a network flow problem with minimum computational cost, and limit the propagation of phase errors in low-quality regions by minimizing the difference between the discrete partial derivatives of the unwrapped phase and the wrapped phase, thereby solving for the global optimal solution. All three types of traditional algorithms can achieve good unwrapping results in scenarios with low noise interference and good phase continuity. However, when the quality of the InSAR interferometric phase map is poor, the path-tracking unwrapping method, represented by the branch-cutting method, is prone to "phase islands," resulting in unwrapping gaps in low-quality regions, and the computation time is long. The minimum norm unwrapping method, represented by the least squares method, causes unwrapping errors in local low-quality regions to spread globally, resulting in fast unwrapping speed but low unwrapping quality. Network flow methods, such as minimum network cost flow, can balance unwrapping accuracy and unwrapping efficiency to some extent, but they still make unwrapping errors on phase edge information in low-quality regions and cannot fully recover phase information.
[0003] Deep learning methods employ various strategies to achieve phase unwrapping through supervised optimization of neural networks on specific datasets. These methods can be broadly categorized into two types: deep learning regression analysis phase unwrapping methods and deep learning-based entanglement number estimation methods. Deep learning regression analysis phase unwrapping methods treat unwrapping as a regression problem, with the neural network directly learning the mapping relationship between the wrapped phase and the absolute phase. Deep learning-based entanglement number estimation methods first transform the problem into a semantic segmentation problem. The input is the wrapped phase, and the trained network outputs the entanglement number, followed by post-processing to obtain the final unwrapping result. Both one-step unwrapping networks based on regression problems and two-step unwrapping networks based on semantic segmentation perform well in their respective application scenarios. However, they still cannot fully meet the requirements when dealing with complex InSAR phase images with complex noise distributions and interference factors such as atmospheric effects. Furthermore, the network structure often determines the performance of deep learning phase unwrapping methods. Many unwrapping networks choose the U-Net framework, but frequent downsampling operations inevitably lead to the loss of interferometric fringe information. Semantic segmentation-based unwrapping networks primarily learn the number of entanglements through the network. They have good unwrapping accuracy when the fringes are clear. However, the complex fluctuations in the InSAR phase image can easily affect the classification of the number of entanglements, causing the network to need to reclassify them globally. This sacrifices unwrapping efficiency and cannot avoid the generation of errors. Summary of the Invention
[0004] Therefore, the purpose of this invention is to provide an InSAR phase unwrapping method and system for multi-scale feature fusion denoising CNN networks, which solves the problem that traditional phase unwrapping cannot achieve good unwrapping results due to noise interference in InSAR.
[0005] To achieve the above objectives, the present invention provides an InSAR phase unwrapping method for a multi-scale feature fusion and denoising CNN network, characterized by comprising the following steps:
[0006] S1. Obtain the simulated InSAR interferometric phase map and construct the InSAR simulation dataset;
[0007] S2. Input the InSAR simulated dataset into the InSAR phase unwrapping model of the multi-scale feature fusion and denoising CNN network for phase unwrapping training;
[0008] The InSAR phase unwrapping model of the multi-scale feature fusion and denoising CNN network adopts the denoising network DnCNN as the framework. It extracts multi-scale features by setting dilated convolution and deformable convolution, fuses the extracted multi-scale feature information, and uses the residual module to perform phase unwrapping to restore the feature information.
[0009] S3. Input the real InSAR interferometric phase map into the InSAR phase unwrapping model of the trained multi-scale feature fusion and denoising CNN network, and output the unwrapped phase map.
[0010] More preferably, in S1, acquiring the simulated InSAR interferometric phase map includes:
[0011] S101. Adjust the parameters of the two-dimensional Gaussian surface to generate Gaussian surfaces of different sizes and patterns; add a random matrix to the generated Gaussian surface to produce distortions of different directions and sizes, forming an interference phase map simulating terrain phase and deformation phase.
[0012] S102. Perlin noise of different frequencies and amplitudes is superimposed to obtain fractal Perlin noise, which simulates the local atmospheric phase; the interferometric phase map is superimposed with the local atmospheric phase to serve as the real phase map for training.
[0013] S103. The real phase diagram is wound to form a noise-free wound interference phase diagram, which is used as a simulated interference phase diagram.
[0014] S104. Using Gaussian noise to simulate uncorrelated noise, generate a complex noise matrix with the same noise level from the real and imaginary parts of the interference phase diagram in S101. Multiply the complex noise matrix with the simulated interference diagram obtained in S103 to obtain a simulated interference phase diagram containing uncorrelated noise.
[0015] Further preferably, step S105 is used to filter the simulated interferometric phase map containing uncorrelated noise generated in step S104 using the Goldstein filtering algorithm, and the filtered interferometric phase map is used as the final InSAR simulation dataset.
[0016] More preferably, the phase unwrapping training of the InSAR phase unwrapping model of the multi-scale feature fusion and denoising CNN network includes the following steps:
[0017] S201. Input the InSAR phase map of the InSAR simulation dataset into the input layer of the InSAR phase unwrapping model of the multi-scale feature fusion denoising CNN network;
[0018] S202. Use 64 convolutional kernels of size 3×3 to extract features from the input InSAR phase map to obtain 64 feature maps;
[0019] S203. Using two dilated convolutions with different sampling rates and a deformable convolution, the extracted initial feature map is abstracted at multiple levels, and 192 different scales of interferogram noise and fringe information are extracted.
[0020] S204. Batch normalize the noise and fringe information of the extracted interferograms at different scales, use the ReLU activation function for adaptive learning, and fuse the processed feature maps; use residual convolution to recover the feature information of the fused feature maps; repeat S203-S204 multiple times until the feature map has completed the recovery of feature information.
[0021] S205. The feature map with recovered feature information is passed through the output convolutional layer and output in a single channel to obtain the unwrapped phase map that meets the requirements.
[0022] More preferably, in S203, the extraction of the initial feature map is subjected to multi-level abstraction using two dilated convolutions with different sampling rates and one deformable convolution, including:
[0023] The extracted initial feature map is abstracted in parallel at multiple levels using a first dilated convolution with a sampling rate of 5*5, a second dilated convolution with a sampling rate of 7*7, and a deformable convolutional layer.
[0024] The first and second dilated convolutions are set with different dilation rates to increase the receptive field without changing the resolution of the image output feature map.
[0025] The deformable convolution introduces a learnable offset into the receptive field, so that the receptive field is no longer a regular square, but an irregular shape that fits the features of the target object.
[0026] The first dilation convolution, the second dilation convolution, and the deformable convolution were used to obtain three sets of feature information at different scales, with 64 features in each set; finally, noise and fringe information of 192 interferograms at different scales were extracted.
[0027] More preferably, in S204, the step of recovering feature information from the fused feature map using residual convolution includes the following process:
[0028] The noise and fringe information of the interferograms at 192 different scales were extracted and spliced and fused using 192 3*3 convolutional kernels. The feature information was restored using a 3*3 residual network convolutional kernel, and the number of channels was restored to 64.
[0029] The recovered 64 feature maps are then processed through a final 3×3 convolution to become a single-channel output.
[0030] The present invention also provides an InSAR phase unwrapping system for a multi-scale feature fusion and denoising CNN network, comprising: a data acquisition unit and an InSAR phase unwrapping model for a multi-scale feature fusion and denoising CNN network;
[0031] The data acquisition unit is used to acquire InSAR interferometric phase maps;
[0032] The InSAR phase unwrapping model of the multi-scale feature fusion denoising CNN network is used to unwrap the input InSAR interferometric phase map and output an unwrapped phase map.
[0033] The InSAR phase unwrapping model of the multi-scale feature fusion denoising CNN network adopts the denoising network DnCNN as the framework. It extracts multi-scale features by setting up a multi-scale feature fusion module with dilated convolution and deformable convolution, fuses the extracted multi-scale feature information, and uses the residual module to perform phase unwrapping to restore the feature information.
[0034] More preferably, the InSAR phase unwrapping model of the multi-scale feature fusion denoising CNN network is trained using an InSAR simulated dataset, which is obtained through the following process:
[0035] The interferometric phase map generation module is used to adjust the parameters of the two-dimensional Gaussian surface to generate Gaussian surfaces of different sizes and patterns. A random matrix is added to the generated Gaussian surface to produce distortions of different directions and sizes, forming an interferometric phase map that simulates the phase of terrain and deformation.
[0036] The real phase map generation module is used to superimpose Perlin noise of different frequencies and amplitudes to obtain fractal Perlin noise, simulating the local atmospheric phase; the interferometric phase map is superimposed with the local atmospheric phase; and used as the real phase map for training.
[0037] The simulated interference phase diagram generation module is used to wrap the real phase diagram to form a noise-free wrapped interference phase diagram, which serves as the simulated interference phase diagram.
[0038] The simulated interferometric phase map generation module containing uncorrelated noise uses Gaussian noise to simulate uncorrelated noise, generates a complex noise matrix with the same noise level from the real and imaginary parts of the interferometric phase map, and multiplies the complex noise matrix with the obtained simulated interferometric map to obtain the simulated interferometric phase map containing uncorrelated noise.
[0039] Further preferably, the system also includes a filtering module, which uses the Goldstein filtering algorithm to filter the generated simulated interferometric phase map containing uncorrelated noise, and uses the filtered interferometric phase map as the final InSAR simulation dataset.
[0040] More preferably, the dilated convolution is provided in two parts, with the sampling rate of the first dilated convolution being 5*5 and the sampling rate of the second dilated convolution being 7*7.
[0041] This application discloses an InSAR phase unwrapping method and system based on a multi-scale feature fusion denoising CNN network. Using a denoising CNN network as a framework, during data training of the multi-scale feature fusion InSAR denoising CNN phase unwrapping model, multiple fittings are performed on the simulated data based on real InSAR data, including terrain phase, deformation phase, atmospheric phase, and phase loss of correlation noise, to form the final interferometric phase map. The phase unwrapping model incorporates two dilated convolutions and deformation convolutions with different sampling rates to extract multi-scale data from the input InSAR data. This solves the problem of inaccurate phase unwrapping results caused by the filtering of loss of correlation noise in traditional InSAR data phase unwrapping.
[0042] This application uses the DnCNN network as the basic framework. In order to make the network more suitable for InSAR phase unwrapping applications, a dataset that conforms to the characteristics of InSAR phase is constructed by simulating each component in the SAR interferometric phase.
[0043] To ensure the network suppresses noise in noisy interferograms while maximizing unwrapping accuracy, a multi-scale feature extraction module is constructed in parallel using dilated convolutions with different dilation rates and deformable convolutions to extract and fuse multi-scale information from the feature map. Furthermore, the inclusion of a residual module improves network speed, avoids network degradation, and ensures network robustness. Attached Figure Description
[0044] Figure 1 The flowchart shows the InSAR phase unwrapping method of the multi-scale feature fusion denoising CNN network of the present invention.
[0045] Figure 2 This is a flowchart for creating simulated SAR interferometric phase data in an example of the present invention;
[0046] Figure 3(a) shows the simulated original phase diagram;
[0047] Figure 3(b) shows the phase diagram after the original phase diagram has been wrapped.
[0048] Figure 3(c) shows the phase diagram of the entangled interference containing noise;
[0049] Figure 3(d) shows the coherence coefficient calculated from the phase diagram;
[0050] Figure 4 This is a schematic diagram of the phase unwrapping network of the present invention;
[0051] Figure 5 This is an architecture diagram of the multi-scale feature extraction module of the present invention. Detailed Implementation
[0052] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0053] like Figure 1 As shown, one embodiment of the present invention provides an InSAR phase unwrapping method for multi-scale feature fusion and denoising CNN networks, which includes the following steps:
[0054] S1. Obtain the simulated InSAR interferometric phase map and construct the InSAR simulation dataset;
[0055] S2. Input the InSAR simulated dataset into the InSAR phase unwrapping model of the multi-scale feature fusion and denoising CNN network for phase unwrapping training;
[0056] The InSAR phase unwrapping model of the multi-scale feature fusion and denoising CNN network adopts the denoising network DnCNN as the framework. It extracts multi-scale features by setting dilated convolution and deformable convolution, fuses the extracted multi-scale feature information, and uses the residual module to perform phase unwrapping to restore the feature information.
[0057] S3. Input the real InSAR interferometric phase map into the InSAR phase unwrapping model of the trained multi-scale feature fusion and denoising CNN network, and output the unwrapped phase map.
[0058] Among them, in such Figure 2 In the illustrated embodiment, obtaining a simulated InSAR interferometric phase map and constructing an InSAR simulation dataset includes the following steps:
[0059] S101: Adjust the parameters of the two-dimensional Gaussian surface to generate Gaussian surfaces of different sizes and patterns. Add a random matrix to the Gaussian surface to control the distribution of the midpoints of the surface, causing the Gaussian surface to produce distortions of different directions and magnitudes, forming an interferometric phase map simulating the phase of terrain and deformation.
[0060] S102: Perlin noise of different frequencies and amplitudes is superimposed to obtain fractal Perlin noise, simulating the local atmospheric phase. The interferometric phase map generated in S1-1 is superimposed with the atmospheric phase as the real phase map for subsequent training.
[0061] S103: Wrap the real phase generated in S102 to form a noise-free wrapped interference phase map. The phase map size in the example is 186 pixels × 186 pixels.
[0062] S104: Use Gaussian noise to simulate uncorrelated noise. Set the noise level according to the deformation phase gradient obtained in S101. Generate a complex noise matrix with the same noise level from the real and imaginary parts of the interference phase diagram in S101. Multiply the complex noise matrix with the simulated interference diagram obtained in S103 to obtain a simulated interference phase diagram containing uncorrelated noise.
[0063] S105: Apply the Goldstein filtering algorithm to the interferometric phase generated in S104, set the filtering window to 32×32, and the filtering coefficient to 0.5. The filtered interferometric phase map forms the final InSAR simulation dataset. The filtered interferometric phase map is used as the input value for network training.
[0064] Figure 2 The simulation process of InSAR simulation data is shown. Figure 3 shows the corresponding 186pix×185pix phase map generated. Figure 3(a) shows the original phase containing surface features and atmospheric phase, which will be used as the true value for network training. Figure 3(b) shows the wrapped phase corresponding to the original phase. Figure 3(c) shows the wrapped phase after adding noise that varies with the deformation gradient. Figure 3(d) shows the coherence coefficient of the simulated interferogram.
[0065] The simulated InSAR dataset was generated in Matlab R2020a and used to form the dataset. The training dataset contains 14,000 pairs of floating-point arrays of size 186 pixels × 186 pixels, and the validation dataset contains 3,000 pairs of floating-point arrays of the same size.
[0066] Furthermore, S2 includes the following steps:
[0067] S201: Input a single-channel InSAR wrapped phase map of size 186pix×186pix into the first layer of the unwrapping model;
[0068] S202 uses 64 3×3 convolutional kernels to extract 64 feature maps, and uses the ReLU activation function to model the non-linearity of the data.
[0069] S203: The multi-scale feature fusion module employs two dilated convolutions with different sampling rates (5×5, 7×7) and a deformable convolutional layer to extract noise and fringe information from interferograms at different scales in parallel (64 per group, 3*64 in total), achieving multi-level abstraction of features, such as... Figure 5 As shown.
[0070] S204: By cascading batch normalization (BN) and ReLU activation functions, 192 3×3 convolutional kernels are used to splice and fuse the extracted multi-scale information and complete the feature information recovery under the action of a 3×3 convolutional kernel. At this time, the number of channels becomes 64.
[0071] Repeat steps 203 and 204 until the feature map completes feature information recovery in the final residual module.
[0072] S205: The recovered feature information is finally processed by a 3×3 convolution kernel to turn the feature map into a single-channel output, resulting in the desired unwrapped phase map.
[0073] Figure 4 The diagram shows the overall structure of the unwrapped network of this invention. As can be seen from the diagram, the network uses a DnCNN network as its basic framework and includes an input layer, batch normalization (BN), ReLU activation function, multi-scale feature extraction module, residual module, and output layer. The functions of the main components are described below:
[0074] Batch normalization: By transporting the system parameters in the search space, the robustness of the system is increased, thereby accelerating the convergence speed of the network, ensuring gradients, and alleviating overfitting.
[0075] ReLU activation function: As an activation function, ReLU is simple to implement, fast to compute, has strong nonlinear fitting ability, and can effectively avoid the gradient vanishing problem.
[0076] Dilated convolution: By setting different dilation rates, dilated convolution increases the receptive field without changing the resolution of the image output feature map, and at the same time avoids information loss caused by downsampling to a certain extent during feature extraction.
[0077] Deformable convolution: Deformable convolution introduces learnable offsets into the receptive field, making the receptive field no longer a regular square, but an irregular shape that fits the features of the target object. Therefore, the convolutional region can always cover the target, and the learned offsets can adapt regardless of how the target stripes are deformed. The addition of deformable convolution allows the network to extract more complex edge information.
[0078] Residual networks: Adding residual network structures to deep neural networks causes the deep neural network to degenerate into a shallow network, solving the degradation problem that easily occurs when the network depth increases and improving network performance.
[0079] Furthermore, the InSAR phase unwrapping of the multi-scale feature fusion and denoising CNN network described in S2 uses the DnCNN network as its main framework, combining the characteristics of dilated convolution, deformable convolution, batch normalization, ReLU activation function, and residual network to construct a multi-scale feature extraction module and a residual module. In the multi-scale feature extraction module, the 64 feature maps extracted from the previous layer are subjected to dilated convolution and deformable convolution to capture phase map information. Without affecting the resolution, dilated convolution achieves regular expansion of the receptive field by setting different dilation rates. Deformable convolution learns an additional offset compared to conventional convolution, making the receptive field more closely fit the complex shape of the interference phase, thereby obtaining more detailed information globally. The three sets of feature information containing different scales are fused using 192 3×3 convolution kernels through a cascade of batch normalization (BN) and ReLU activation function. The multi-scale information is then restored by the residual module, ensuring that the network depth is increased while avoiding network degradation.
[0080] Furthermore, S3 includes the following steps:
[0081] A single-channel INSAR wrapped phase map is input into a trained unwrapped network. After passing through the first convolutional layer, a 64-channel feature map is output. This map then passes through a multi-scale feature extraction module in the intermediate layers, outputting a feature map fused with features from different scales. Finally, a residual module reconstructs the feature map. For example... Figure 4 The multi-scale feature extraction module shown has 8 modules and the residual module has 10 modules. Finally, the fused 64-channel feature map is converted into a single-channel output by a 3×3 convolution kernel to obtain the final unwrapped phase map that meets the expectations.
[0082] The network in this example is developed based on the Tensorflow 2.9.0 deep learning framework using Python 3.8. The main computer parameters for network training and experimental testing are as follows: Tesla T4 GPU + 8vCPU, Intel Xeon Processor (Skylake, IBRS) CPU + 64G RAM.
[0083] The present invention also provides an InSAR phase unwrapping system for a multi-scale feature fusion and denoising CNN network, for implementing the above-mentioned phase unwrapping method, comprising: a data acquisition unit and an InSAR phase unwrapping model for a multi-scale feature fusion and denoising CNN network;
[0084] The data acquisition unit is used to acquire InSAR interferometric phase maps;
[0085] The InSAR phase unwrapping model of the multi-scale feature fusion denoising CNN network is used to unwrap the input InSAR interferometric phase map and output an unwrapped phase map.
[0086] The InSAR phase unwrapping model of the multi-scale feature fusion denoising CNN network adopts the denoising network DnCNN as the framework. It extracts multi-scale features by setting up a multi-scale feature fusion module with dilated convolution and deformable convolution, fuses the extracted multi-scale feature information, and uses the residual module to perform phase unwrapping to restore the feature information.
[0087] The InSAR phase unwrapping model of the multi-scale feature fusion and denoising CNN network is trained using an InSAR simulation dataset, which is obtained through the following process:
[0088] The interferometric phase map generation module is used to adjust the parameters of the two-dimensional Gaussian surface to generate Gaussian surfaces of different sizes and patterns. A random matrix is added to the generated Gaussian surface to produce distortions of different directions and sizes, forming an interferometric phase map that simulates the phase of terrain and deformation.
[0089] The real phase map generation module is used to superimpose Perlin noise of different frequencies and amplitudes to obtain fractal Perlin noise, simulating the local atmospheric phase; the interferometric phase map is superimposed with the local atmospheric phase; and used as the real phase map for training.
[0090] The simulated interference phase diagram generation module is used to wrap the real phase diagram to form a noise-free wrapped interference phase diagram, which serves as the simulated interference phase diagram.
[0091] The simulated interferometric phase map generation module containing uncorrelated noise uses Gaussian noise to simulate uncorrelated noise, generates a complex noise matrix with the same noise level from the real and imaginary parts of the interferometric phase map, and multiplies the complex noise matrix with the obtained simulated interferometric map to obtain the simulated interferometric phase map containing uncorrelated noise.
[0092] It also includes a filtering module, which uses the Goldstein filtering algorithm to filter the generated simulated interferometric phase map containing uncorrelated noise, and uses the filtered interferometric phase map as the final InSAR simulation dataset.
[0093] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.
Claims
1. A method for InSAR phase unwrapping of a multi-scale feature fusion and denoising CNN network, characterized in that, Includes the following steps: S1. Obtain the simulated InSAR interferometric phase map and construct the InSAR simulation dataset; S2. Input the InSAR simulated dataset into the InSAR phase unwrapping model of the multi-scale feature fusion and denoising CNN network for phase unwrapping training; The InSAR phase unwrapping model of the multi-scale feature fusion and denoising CNN network adopts the denoising network DnCNN as the framework. It extracts multi-scale features by setting dilated convolution and deformable convolution, fuses the extracted multi-scale feature information, and uses the residual module to perform phase unwrapping to restore the feature information. S3. Input the noisy real InSAR interferometric phase map into the trained multi-scale feature fusion denoising CNN network's InSAR phase unwrapping model, and output the unwrapped phase map; the phase unwrapping training of the multi-scale feature fusion denoising CNN network's InSAR phase unwrapping model includes the following steps: S201. Input the InSAR phase map of the InSAR simulation dataset into the input layer of the InSAR phase unwrapping model of the multi-scale feature fusion denoising CNN network; S202. Use 64 convolutional kernels of size 3×3 to extract features from the input InSAR phase map to obtain 64 feature maps; S203. Using two dilated convolutions with different sampling rates and a deformable convolution, the extracted initial feature map is abstracted at multiple levels, and 192 different scales of interferogram noise and fringe information are extracted. S204. Batch normalization is performed on the noise and fringe information of the extracted interferograms at different scales, and adaptive learning is performed using the ReLU activation function. The processed feature maps are then fused. The feature map after fusion is used to recover feature information using residual convolution; Repeat steps S203-S204 multiple times until the feature map has completed feature information recovery; S205. The feature map with recovered feature information is passed through the output convolutional layer and output in a single channel to obtain the unwrapped phase map that meets the requirements.
2. The InSAR phase unwrapping method for multi-scale feature fusion and denoising CNN networks according to claim 1, characterized in that, In S1, obtaining the simulated InSAR interferometric phase map includes: S101. Adjust the parameters of the two-dimensional Gaussian surface to generate Gaussian surfaces of different sizes and patterns; add a random matrix to the generated Gaussian surface to produce distortions of different directions and sizes, forming an interference phase map simulating terrain phase and deformation phase. S102. Perlin noise of different frequencies and amplitudes is superimposed to obtain fractal Perlin noise, which simulates the local atmospheric phase; the interferometric phase map is superimposed with the local atmospheric phase to serve as the real phase map for training. S103. The real phase diagram is wound to form a noise-free wound interference phase diagram, which is used as a simulated interference phase diagram. S104. Using Gaussian noise to simulate uncorrelated noise, generate a complex noise matrix with the same noise level from the real and imaginary parts of the interference phase diagram in S101. Multiply the complex noise matrix with the simulated interference diagram obtained in S103 to obtain a simulated interference phase diagram containing uncorrelated noise.
3. The InSAR phase unwrapping method for multi-scale feature fusion and denoising CNN networks according to claim 2, characterized in that, It also includes S105, which uses the Goldstein filtering algorithm to filter the simulated interferometric phase map containing uncorrelated noise generated by S104, and uses the filtered interferometric phase map as the final InSAR simulation dataset.
4. The InSAR phase unwrapping method for multi-scale feature fusion and denoising CNN networks according to claim 1, characterized in that, In S203, the extraction of the initial feature map is abstracted at multiple levels using two dilated convolutions with different sampling rates and one deformable convolution, including: The extracted initial feature map is abstracted in parallel at multiple levels using a first dilated convolution with a sampling rate of 5*5, a second dilated convolution with a sampling rate of 7*7, and a deformable convolutional layer. The first and second dilated convolutions are set with different dilation rates to increase the receptive field without changing the resolution of the image output feature map. The deformable convolution introduces a learnable offset into the receptive field, so that the receptive field is no longer a regular square, but an irregular shape that fits the features of the target object. The first dilation convolution, the second dilation convolution, and the deformable convolution were used to obtain three sets of feature information at different scales, with 64 features in each set; finally, noise and fringe information of 192 interferograms at different scales were extracted.
5. The InSAR phase unwrapping method for multi-scale feature fusion and denoising CNN networks according to claim 4, characterized in that, In S204, the process of restoring feature information from the fused feature map using residual convolution includes the following steps: The noise and fringe information of the interferograms at 192 different scales were extracted and spliced and fused using 192 3*3 convolution kernels. The feature information was restored using a 3*3 residual network convolution kernel, and the number of channels was restored to 64. The recovered 64 feature maps are then processed through a final 3×3 convolution to become a single-channel output.
6. An InSAR phase unwrapping system for a multi-scale feature fusion and denoising CNN network, characterized in that, include: InSAR phase unwrapping model of data acquisition unit and multi-scale feature fusion denoising CNN network; The data acquisition unit is used to acquire InSAR interferometric phase maps; The InSAR phase unwrapping model of the multi-scale feature fusion denoising CNN network is used to unwrap the input InSAR interferometric phase map and output an unwrapped phase map. The InSAR phase unwrapping model of the multi-scale feature fusion denoising CNN network adopts the denoising network DnCNN as the framework. It extracts multi-scale features by setting up a multi-scale feature fusion module with dilated convolution and deformable convolution, fuses the extracted multi-scale feature information, and uses the residual module to perform phase unwrapping to restore the feature information. The phase unwrapping training of the InSAR phase unwrapping model of the multi-scale feature fusion and denoising CNN network includes the following steps: S201. Input the InSAR phase map of the InSAR simulation dataset into the input layer of the InSAR phase unwrapping model of the multi-scale feature fusion denoising CNN network; S202. Use 64 convolutional kernels of size 3×3 to extract features from the input InSAR phase map to obtain 64 feature maps; S203. Using two dilated convolutions with different sampling rates and a deformable convolution, the extracted initial feature map is abstracted at multiple levels, and 192 different scales of interferogram noise and fringe information are extracted. S204. Batch normalize the noise and fringe information of the extracted interferograms at different scales, use the ReLU activation function for adaptive learning, and fuse the processed feature maps. The feature maps are then recovered using residual convolution. Repeat steps S203-S204 multiple times until the feature map has completed feature information recovery; S205. The feature map with recovered feature information is passed through the output convolutional layer and output in a single channel to obtain the unwrapped phase map that meets the requirements.
7. The InSAR phase unwrapping system of the multi-scale feature fusion and denoising CNN network according to claim 6, characterized in that, The InSAR phase unwrapping model of the multi-scale feature fusion and denoising CNN network is trained using an InSAR simulation dataset, which is obtained through the following process: The interferometric phase map generation module is used to adjust the parameters of the two-dimensional Gaussian surface to generate Gaussian surfaces of different sizes and patterns. A random matrix is added to the generated Gaussian surface to produce distortions of different directions and sizes, forming an interferometric phase map that simulates the phase of terrain and deformation. The real phase map generation module is used to superimpose Perlin noise of different frequencies and amplitudes to obtain fractal Perlin noise, simulating the local atmospheric phase; and to superimpose the interferometric phase map with the local atmospheric phase. As the true phase map for training; The simulated interference phase diagram generation module is used to wrap the real phase diagram to form a noise-free wrapped interference phase diagram, which serves as the simulated interference phase diagram. The simulated interferometric phase map generation module containing uncorrelated noise uses Gaussian noise to simulate uncorrelated noise, generates a complex noise matrix with the same noise level from the real and imaginary parts of the interferometric phase map, and multiplies the complex noise matrix with the obtained simulated interferometric map to obtain the simulated interferometric phase map containing uncorrelated noise.
8. The InSAR phase unwrapping system of the multi-scale feature fusion and denoising CNN network according to claim 7, characterized in that, It also includes a filtering module, which uses the Goldstein filtering algorithm to filter the generated simulated interferometric phase map containing uncorrelated noise, and uses the filtered interferometric phase map as the final InSAR simulation dataset.
9. The InSAR phase unwrapping system of the multi-scale feature fusion and denoising CNN network according to claim 7, characterized in that, The dilated convolution is provided in two parts: the sampling rate of the first dilated convolution is 5*5, and the sampling rate of the second dilated convolution is 7*7.