Phase unwrapping method for synthetic aperture radar interferometry based on convolutional neural network and visual transformer

Abstract

The present invention relates to a phase unwrapping method for synthetic aperture radar interferometry based on a convolutional neural network and a visual transformer, and belongs to the field of synthetic aperture radar interferometry technology. In order to solve the problem that traditional deep learning methods have insufficient phase unwrapping accuracy in complex terrain and noisy environments, the present invention unwraps the image through the following steps: inputting the synthetic aperture radar interferogram to be unwrapped, extracting local features using a convolutional neural network, generating a fused feature map through a multi-scale feature capture module, adjusting the channel weights through a squeeze-excitation network, and then applying a single-head self-attention mechanism to capture global relationships, and finally reconstructing a continuous phase map through a convolutional layer. The present invention effectively combines local feature extraction with global relationship modeling, improves the processing capability for phase jump areas and noise, is suitable for phase unwrapping in large gradient deformation areas on the surface, such as mining areas, can accurately restore continuous phase information in interferometry, and provide accurate data support for applications such as surface deformation monitoring.

Description

Technical Field

[0001] The present invention relates to the technical field of synthetic aperture radar interferometry (InSAR), and more specifically to an InSAR phase unwrapping method based on a convolutional neural network (CNN) and a visual transformer (ViT). The method is used to solve the phase unwrapping problem in areas with large surface gradient deformation, such as mining areas, and to improve the unwrapping accuracy and noise suppression capability. Background Art

[0002] Synthetic Aperture Radar Interferometry (InSAR) acquires target information through phase difference measurements. Phase unwrapping is a key step in recovering the continuous phase. Traditional phase unwrapping methods rely on fixed rules or simple models and are difficult to adapt to complex terrain and noise interference. With the development of deep learning, methods based on convolutional neural networks have improved unwrapping efficiency, but they still have significant drawbacks:

[0003] First, the ability to extract local features is limited: conventional convolution kernels have a fixed size and cannot effectively capture multi-scale phase features, especially in areas with irregular phase jumps caused by undulating terrain, where key details are easily lost.

[0004] Second, insufficient global relationship modeling: Traditional networks lack modeling of global phase correlation, making it difficult to balance the feature weights of different regions, resulting in error accumulation in the disentanglement results in large phase-continuous areas.

[0005] Third, insufficient fusion of physical priors: ignoring the physical characteristics of InSAR phase (such as phase gradient and edge-sensitive features), purely data-driven models have difficulty in balancing noise suppression and edge preservation, and are easily affected by speckle noise, resulting in artifacts or over-smoothing in the unwrapping results.

[0006] Fourth, poor adaptability to complex structures: The phase jump modeling of large gradient deformation areas such as steep slopes and faults is insufficient, and the fixed convolution kernel cannot dynamically adapt to the phase change density, resulting in distorted edge responses and a significant increase in unwrapping errors.

[0007] The root of these problems is that traditional methods fail to effectively combine multi-scale feature extraction, global context association and physical prior knowledge, which limits the disentanglement accuracy and robustness in complex scenarios.

[0008] The research on phase unwrapping (PU) originated from interferometry technology, and has been widely used in the fields of synthetic aperture radar interferometry and optical interferometry. Interferometry obtains target information by measuring the phase difference between two coherent waves. However, since the measuring instrument can usually only record the modulus value of the phase, The phase will be restricted to [-π, π] or [0, 2π], a process called phase wrapping. Therefore, in order to obtain the phase difference of the true result, it is necessary to add or subtract a whole cycle number 2kπ to the result and unwrap the wrapped phase to more accurately invert the elevation and deformation information of the measured object. This is the purpose of the phase unwrapping problem, that is, to restore the original continuous phase. Where k is an appropriate integer so that the recovered phase remains smooth and conforms to physical laws. Specifically, the mathematical core of phase unwrapping is how to convert the measured wrapped phase into Restoring continuous phase Where k is an integer that minimizes the phase difference between adjacent pixels. Mathematically, the phase gradient can be defined as shown in the following two formulas:

[0009]

[0010] For wrapped phase data, phase jumps usually occur in regions that satisfy the following formula,

[0011]

[0012] At this time, the unwrapping algorithm needs to determine whether to add or subtract 2π to smooth the phase distribution. Figure 1 and Figure 2 Schematic diagrams representing phase unwrapping before and after are shown respectively. Summary of the Invention

[0013] An object of the present invention is to solve at least the above problems and to provide at least the advantages which will be described hereinafter.

[0014] In order to achieve these objects and other advantages according to the present invention, a method for phase unwrapping of synthetic aperture radar interferometry based on convolutional neural network and visual transformer is provided, comprising the following steps:

[0015] S1. Input the synthetic aperture radar interferogram to be unwrapped;

[0016] S2. Using a convolutional neural network including multiple convolution kernels to extract local features of the interference pattern to generate an initial feature map;

[0017] S3. Inputting the initial feature map into a multi-scale feature capture module, wherein the multi-scale feature capture module includes multiple parallel convolution blocks, and the parallel convolution blocks use convolution kernels of different sizes to perform multi-scale feature extraction to obtain multiple groups of sub-feature maps;

[0018] S4, performing batch normalization processing and ReLU activation function processing on multiple groups of sub-feature maps in sequence, and splicing the processed multiple groups of sub-feature maps along the channel dimension into a fused feature map;

[0019] S5. Inputting the fused feature map into a scale global relationship capture module, the scale global relationship capture module dynamically adjusts the channel weights of the fused feature map through a squeeze-excitation network to generate a weighted feature map;

[0020] S6. Applying a single-head self-attention mechanism on the weighted feature map to calculate the global similarity between the query matrix, the key matrix, and the value matrix to generate a global attention weight matrix;

[0021] S7, multiplying the global attention weight matrix by the weighted feature map to obtain a global enhanced feature map;

[0022] S8. Perform phase reconstruction on the global enhanced feature map through the convolution layer of the convolution kernel, and output an unwrapped continuous phase map.

[0023] Preferably, the multiple parallel convolution blocks in the multi-scale feature capture module specifically include a first convolution block, a second convolution block, and a third convolution block, the first convolution block uses a 3×3 convolution kernel for feature extraction, the second convolution block uses a 5×5 convolution kernel for feature extraction, and the third convolution block uses a 7×7 convolution kernel for feature extraction, and each parallel convolution block generates 64 feature maps;

[0024] The fused feature map after the processed multiple groups of sub-feature maps are spliced ​​along the channel dimension contains 192 channels;

[0025] The convolutional neural network in step S2 uses 64 3×3 convolution kernels to extract local features and generates an initial feature map with a size of 192×180×180, where 192 is the number of channels and 180×180 is the spatial dimension.

[0026] The convolution kernel of the convolution layer used in the phase reconstruction step has a size of 3×3.

[0027] Preferably, in steps S6 and S7, in the single-head self-attention mechanism, the query matrix, the key matrix, and the value matrix are generated by the following steps:

[0028] The weighted feature map is linearly transformed with a learnable weight matrix to obtain a query matrix, a key matrix, and a value matrix, wherein the learnable weight matrix includes:

[0029] The global similarity is calculated by performing a dot product operation on the transposed matrix of the query matrix and the key matrix, and dividing the dot product result by the square root value of the dimension of the query matrix or the key matrix, and then normalizing it by the Softmax function to obtain a global attention weight matrix;

[0030] The global enhanced feature map is generated by multiplying the global attention weight matrix by a value matrix.

[0031] Preferably, in step S5, the squeeze-excitation network dynamically adjusts the channel weights of the fused feature map through the following steps:

[0032] Performing a global average pooling operation on the fused feature map to generate a channel descriptor;

[0033] Input the channel descriptor into the first fully connected layer for linear transformation, the first fully connected layer compresses the dimension of the channel descriptor to 1 / 16 of the original number of channels and processes it through the ReLU activation function;

[0034] The processed results are input into the second fully connected layer for linear transformation. The second fully connected layer restores the compressed channel descriptors to the original number of channels and generates channel weight coefficients through the Sigmoid activation function;

[0035] The channel weight coefficient is multiplied by the fusion feature map channel by channel to generate a weighted feature map.

[0036] Preferably, in steps S3 and S5, the network depth of the multi-scale feature capture module and the scale global relationship capture module are both eight layers;

[0037] The multi-scale feature capture module realizes multi-level local feature extraction through eight layers of stacked convolution blocks. The output of each layer of convolution blocks is fused with the input of the next layer by element-by-element addition.

[0038] The scale global relationship capture module realizes global feature modeling through an eight-layer stacked squeeze-excitation network and a single-head self-attention mechanism unit. In the scale global relationship capture module, the weighted feature map output by each layer is superimposed with the weighted feature map input by the next layer through a residual connection.

[0039] Preferably, in steps S6 and S7:

[0040] In the single-head self-attention mechanism, the shape of the input feature matrix is ​​200, 8000, 1, where 200 is the batch size, 8000 is the input sequence length, and 1 is the feature dimension;

[0041] The learnable weight matrix includes a query weight matrix, a key weight matrix, and a value weight matrix, whose dimensions are 1×1, 1×1, and 1×1 respectively;

[0042] In the transposed matrix dot product operation of the query matrix and the key matrix, the dimension square root value is the square root value of the characteristic dimension of the query matrix;

[0043] The Softmax function performs normalization along the length dimension of the input sequence to generate a global attention weight matrix.

[0044] Preferably, step S5 further includes a spatial attention enhancement step:

[0045] Inputting the weighted feature map into a deformable spatial attention module, generating a dynamic spatial weight distribution that matches the geometric features of the phase jump region through a deformable convolution kernel;

[0046] The deformable spatial attention module performs the following operations:

[0047] A 3×3 deformable convolution layer is used to extract deformation features from the weighted feature map, where the offset of the convolution kernel is obtained through learning;

[0048] The deformation features are input into a two-branch structure. The first branch generates a spatial saliency map through channel reduction convolution, and the second branch extracts edge-sensitive features through depth-wise separable convolution.

[0049] After performing Hadamard product operation on the dual-branch output, the improved Gaussian error linear unit activation function is used to generate the spatial attention mask;

[0050] The spatial attention mask is gated and fused with the original weighted feature map, where the gating coefficient is obtained by dynamically calculating the phase gradient feature, and the enhanced feature map with deformation perception capability is finally output as the input of step S6.

[0051] Preferably, the generation and application of the phase gradient feature specifically includes the following steps:

[0052] a) Extract the original phase gradient map from the interferogram input in step S1, and use the improved Sobel-Feldman operator to calculate the first-order differential gradient along the azimuth and range directions respectively. The gradient kernel size dynamically adapts to the phase jump density of the current processing block. The calculation formula is:

[0053]

[0054] in, is the adaptive scaling factor, is the edge sensitivity coefficient, φ is the original interference phase, G x is the azimuthal gradient, G y is the distance gradient;

[0055] b) Perform nonlinear enhancement on the gradient magnitude map:

[0056]

[0057] Where λ is the learnable enhancement coefficient, σ is the standard deviation of the local window, and ° represents element-by-element multiplication;

[0058] c) Construct a multi-scale phase gradient feature pyramid, obtain the gradient features of [original scale, 1 / 2 scale, 1 / 4 scale] through three-level Gaussian pyramid downsampling, and generate a 128-dimensional phase gradient descriptor after splicing in the channel dimension and fusion through 3×3 convolution;

[0059] d) Perform cross-modal interaction between the phase gradient descriptor and the weighted feature map output from step S5, and generate a dynamic mask M of spatial dimension through a gated recurrent unit gate , the final gating coefficient is calculated as:

[0060] γ=Sigmoid(Conv 1×1 ([F weight ⊙M gate , G mag ]))

[0061] Among them F weight is the weighted feature map, and ⊙ represents element-wise multiplication.

[0062] The present invention has at least the following beneficial effects:

[0063] First, local features of different receptive fields are extracted through multi-scale convolution blocks, and the channel weights are dynamically adjusted in combination with the squeeze-excitation network to increase the attention of key features; the single-head self-attention mechanism is used to model global phase correlation, enhance the ability to capture long-distance dependencies, and effectively solve the problems of single local features and lack of global relationships in traditional methods, significantly improving disentanglement accuracy and noise suppression capabilities.

[0064] Second, the specific parameters of the multi-scale convolution block are clarified. Through the combination of 3×3, 5×5, and 7×7 convolution kernels, phase features of different scales are covered. The fused feature map of the 192 channels after splicing retains rich multi-scale information, providing sufficient input for subsequent global feature modeling. The unified convolution kernel size and dimension setting ensures the stability and computational efficiency of the network structure, and improves the consistency and accuracy of feature extraction.

[0065] Third, the mathematical calculation process of the self-attention mechanism is clarified, and through linear transformation and normalization operations, weighted aggregation of global phase features is achieved, so that the network can adaptively focus on the phase correlation of key areas, enhance the ability to model long-distance phase changes, and improve the global consistency and smoothness of the disentanglement results.

[0066] Fourth, global average pooling aggregates global spatial information, and compression-excitation operations adjust channel-dimensional weights, enabling the network to dynamically allocate channel resources based on the importance of phase features. Low-dimensional mapping and nonlinear activation enhance the expressive power of weight coefficients, effectively suppressing noise from irrelevant channels, highlighting key features related to phase disentanglement (such as edge information in high-gradient regions), and improving the discriminability and robustness of feature representation.

[0067] Fifth, the deformable convolution kernel adaptively matches the irregular geometric shape of phase jumps, breaking through the limitations of fixed convolution kernels; the dual-branch structure combines global saliency and edge detail features to improve the perception accuracy of phase discontinuity boundaries; the phase gradient gating mechanism embeds physical priors (phase gradient information) into the attention fusion process, dynamically adjusts the weight according to the intensity of local phase changes, realizes the organic combination of data-driven and physical models, and significantly enhances the disentanglement accuracy and edge preservation ability under complex terrain.

[0068] Sixth, the dynamic gradient operator adaptively adjusts the kernel size according to the phase jump density, maintaining smoothness in flat areas while enhancing the edge response of high-gradient areas; nonlinear enhancement and multi-scale pyramid structure enhance the robustness of gradient features, and cross-modal interaction fuses physical gradients and deep features through gated recurrent units, enabling the attention mechanism to use the directional prior of the phase gradient to guide the focus area, effectively reducing the phase reconstruction error near the fault line and improving the disentanglement reliability in complex terrain.

[0069] Other advantages, objectives and features of the present invention will be reflected in part from the following description and will be understood by those skilled in the art through study and practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 This is a schematic diagram representing the phase before unwrapping mentioned in the background of the present invention;

[0071] Figure 2 This is a schematic diagram representing phase unwrapping mentioned in the background of the present invention;

[0072] Figure 3 The simulation data experimental results of samples 1 to 3 of different methods are shown;

[0073] Figure 4 The experimental results of real data of samples one and two for different methods are shown;

[0074] Figure 5 The experimental results of real data of samples three and four for different methods are shown;

[0075] Figure 6 Comparison of the unwrapping results and leveling data for Dongtan Coal Mine;

[0076] Figure 7 Comparison of the unwrapping results and leveling data for Yangcun Coal Mine;

[0077] Figure 8 A schematic diagram of the overall network architecture of one of the technical solutions of the present invention;

[0078] Figure 9 An overview map of the study area. DETAILED DESCRIPTION

[0079] The present invention will be described in further detail below in conjunction with the accompanying drawings so that those skilled in the art can implement the invention with reference to the description.

[0080] It should be noted that the experimental methods described in the following embodiments are conventional methods unless otherwise specified, and the reagents and materials are commercially available unless otherwise specified; in the description of the present invention, the orientation or positional relationship indicated by the terms is based on the orientation or positional relationship shown in the accompanying drawings, which is only for the convenience of describing the present invention and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention.

[0081] like Figures 3 to 9 As shown, the present invention provides a synthetic aperture radar interferometry phase unwrapping method based on convolutional neural network and visual transformer (abbreviated as CVPUNet), the specific steps are as follows:

[0082] S1. Input the synthetic aperture radar interferogram to be unwrapped. The input interferogram size is 180×180 pixels, and its phase value is wrapped in the range of -π,π.

[0083] S2. Use convolutional neural network (CNN) to extract local features:

[0084] A CNN with 64 3×3 convolution kernels is used to extract the initial features of the interference pattern, generating an initial feature map with a size of 192×180×180 (number of channels × spatial dimension).

[0085] S3, Multi-scale feature capture: The initial feature map is input into the multi-scale feature capture module, which contains three parallel convolution blocks.

[0086] The first convolution block: 3×3 convolution kernel, generating 64 feature maps;

[0087] The second convolution block: 5×5 convolution kernel, generating 64 feature maps;

[0088] The third convolution block: 7×7 convolution kernel, generating 64 feature maps.

[0089] After batch normalization (BatchNorm) and ReLU activation, the three sets of feature maps are concatenated along the channel dimension into a 192-channel fused feature map.

[0090] S4. Feature fusion and nonlinear processing: The concatenated fusion feature map is subjected to multi-level feature enhancement through eight layers of stacked convolution blocks. The output of each layer is superimposed with the input of the next layer through residual connection to alleviate the gradient disappearance.

[0091] S5. Dynamic adjustment of channel weights: The fused feature map is input into the scale global relationship capture module, and the channel weights are dynamically adjusted through the squeeze-excitation network (SENet):

[0092] a) Squeeze: Perform global average pooling on the fused feature map to generate a channel descriptor (Formula 1);

[0093] Z=AvgPool(X)∈R 192×180×180 Formula 1

[0094] Among them, AvgPool represents the global average pooling operation, X is the input feature, and R is its shape. In this solution, its size is 192×180×180.

[0095] b) Excitation: The dimension is compressed to 1 / 16 through the fully connected layer, and restored to the original number of channels after ReLU activation to generate the channel weight coefficient (Formula 2);

[0096] S=σ(f ReLU (f FC (Z))) Formula 2

[0097] Among them, f FC is the fully connected layer, f ReLU is the ReLU activation function, and σ represents the Sigmoid activation function.

[0098] c) Scaling: Multiply the weight coefficient and the fusion feature map channel by channel to generate a weighted feature map (Formula 3).

[0099] Y=X·S Formula 3

[0100] Where X is the input feature and · represents element-wise multiplication.

[0101] S6. Global Relationship Modeling:

[0102] Apply single-head self-attention mechanism (SHSA) on the weighted feature map:

[0103] The input feature matrix (shape 200, 8000, 1) is passed through the learnable weight matrix W Q 、W K 、W VLinear transformation into query matrix (Q), key matrix (K), value matrix (V) (Formula 4);

[0104]

[0105] Among them, X is the input feature matrix, is the learned weight matrix, and W Q 、W K 、W V Used to generate queries, keys, and values ​​respectively.

[0106] Calculate the dot product of the transposed matrices of Q and K, scale them, and then normalize them using Softmax to generate the attention weight matrix (Formulas 5 and 6).

[0107]

[0108] in, is a scaling factor that prevents the dot product from vanishing due to excessively large gradients. D is the dimension of the query (or key). The result is an attention matrix that represents the similarity between each pair of query and key.

[0109] The Softmax operation normalizes the weights of each row so that the sum of the weights of each row is 1. These weights represent how much attention each position in the input sequence pays to other positions.

[0110] Multiply the attention weight matrix with V to generate the global enhanced feature map (Formula 7).

[0111] Output=Attention Weights×V Formula 7

[0112] This step is equivalent to a weighted summation of the values ​​V, where the weights are given by the previously calculated attention matrix. The resulting output matrix has a shape of [N, T, D], representing the new representation of each position.

[0113] S7, Feature Enhancement and Phase Reconstruction:

[0114] The global enhanced feature map is phase reconstructed through a 3×3 convolutional layer, and the unwrapped continuous phase map is output.

[0115] S8. Output the untangling result:

[0116] The size of the final phase map is related to the input interferometer Figure 1 The phase value is restored to a continuous form.

[0117] like Figure 8As shown in the figure, the overall network architecture of this method consists of five main components: input layer, local feature extraction module (CNN), multi-scale feature capture module, scale global relationship capture module (SENet+self-attention mechanism), and phase reconstruction output layer. Figure 8 The meanings of each module are as follows: MFBlock represents the multi-feature fusion module, SEBlock represents the channel attention mechanism module, and SHSABlock represents the sparse hierarchical self-attention module.

[0118] <Test Results>

[0119] 1. Simulation data experimental results

[0120] The training set and test set contain 8000 and 2000 floating-point arrays of 180×180 pixels, respectively. The number of training iterations (Epoch) is set to 200, and the training time is about 62.5 hours. In order to more clearly verify the reliability and scalability of the model on simulation data, this study used samples one, two, and three with different terrain features, and compared the method CVPUNet of the present invention with DLPU (Deep Learning Phase Unwrapping), FCNet-CA (Fully Convolutional Network with ChannelAttention), and PUNet (Phase Unwrapping Network), as shown in Figure 2. Figure 3 shown.

[0121] In Sample 1, the DLPU performs relatively poorly. Despite using deep learning methods, the DLPU still exhibits significant shortcomings in noise suppression, particularly in high-frequency regions of the phase image, where residual noise is noticeable and image detail recovery is incomplete. When processing images with complex boundaries, the DLPU's results are blurry at edges and structures, demonstrating its limitations in handling high-gradient regions and nonlinear shapes. FCNet-CA shows some improvement over the DLPU, but still exhibits blurring in high-gradient regions. While FCNet-CA performs well in smooth regions, the unwrapping results exhibit some distortion in areas with high-order phase variations, particularly around image edges and details, failing to fully preserve true phase information. PUNet significantly outperforms both DLPU and FCNet-CA, effectively removing noise and maintaining image smoothness during the unwrapping process. Its phase recovery is more accurate across most regions. However, PUNet still fails to perfectly recover details in high-value regions, particularly around the edges of complex structures, where slight detail loss occurs. Compared to these methods, CVPUNet performs the best, particularly in achieving a balance between noise removal and detail recovery. CVPUNet removes noise without sacrificing important structural information. The result is very close to the true disentangled image, especially in complex boundaries and high-gradient regions. By combining a convolutional neural network with a visual transformer, CVPUNet accurately captures subtle changes in the image and effectively avoids oversmoothing when handling complex phase changes. As a result, CVPUNet demonstrates its strong noise suppression and fine structure recovery capabilities, becoming the best-performing method in this test.

[0122] In sample 2, the DLPU still fails to meet expectations, particularly when dealing with complex boundaries and areas with high gradient variations, resulting in artifact-ridden unwrapping results. While the DLPU can remove some noise, it struggles with complex shapes, particularly when recovering nonlinear shapes, resulting in significant distortion. While FCNet-CA improves on high-noise areas, it still struggles with complex boundaries and shape variations, resulting in significant gaps in detail recovery in the unwrapped image. PUNet's performance improves over FCNet-CA, particularly in smooth regions and the recovery of most structures, resulting in a smoother image. However, PUNet still has limitations in boundary handling, particularly in high-gradient areas, where detail loss is noticeable. In this context, CVPUNet again demonstrates its advantages. Compared to other methods, CVPUNet not only accurately removes noise but also preserves more detail in complex boundaries and high-gradient areas. Especially when dealing with complex geometric shapes, CVPUNet, through its multi-layered feature extraction network, preserves the fine structure of the image while avoiding the loss of detail caused by excessive noise removal. In addition, the robustness of CVPUNet in practical applications is also particularly outstanding. It can adapt to different types of SAR image data and shows strong applicability.

[0123] For sample 3, DLPU's performance remains lacking, particularly in recovering circular structures. The unwrapping result is severely oversmoothed, resulting in a significant loss of high-frequency information. This demonstrates that while DLPU can perform phase recovery in certain situations, its limitations remain significant when dealing with complex images, particularly at boundaries and in high-gradient regions. While FCNet-CA improves denoising to some extent, its handling of complex phase images remains limited. In particular, in high-gradient regions, FCNet-CA's results exhibit noticeable blurring of details, failing to fully recover the complex structure of the true unwrapped image. While PUNet exhibits good smoothness in many regions, it still suffers from a loss of detail in high-gradient regions and struggles to recover complex boundaries. Unlike other methods, CVPUNet performs well in this sample. Through its unique network architecture and multi-objective loss function, CVPUNet effectively denoises while accurately recovering image structure and details. In particular, in high-value regions and complex boundaries, CVPUNet avoids detail loss caused by oversmoothing while demonstrating effective noise suppression. In comparison, other methods all have limitations in these areas. Therefore, CVPUNet once again demonstrated its excellent ability in processing complex images in this sample, especially in complex terrain and high-noise environments, showing certain advantages.

[0124] Based on the results of multi-sample analysis, CVPUNet shows significant advantages over DLPU, FCNet-CA and PUNet in terms of noise suppression, structure recovery and detail preservation. The key lies in the ability to balance noise suppression and detail preservation, avoiding over-smoothing while effectively removing noise, and retaining high-frequency detail information, thereby obtaining more accurate and realistic disentanglement results. In addition, CVPUNet performs well in processing high-gradient areas and complex phase changes, and can maintain fine structures and avoid edge blurring and detail loss. This is due to the design of its weighted residual block and multi-objective loss function, which enables it to better cope with the challenges of complex shapes and high-noise environments. At the same time, CVPUNet exhibits strong robustness and can maintain stable performance in SAR image processing with different sample data and different terrains and noise environments. It has high computational efficiency and can adapt to the processing needs of large-scale SAR data.

[0125] In summary, CVPUNet, with its advantages in noise removal, detail recovery, and structure preservation, demonstrates superior performance compared to other deep learning methods. It is not only applicable to current InSAR phase unwrapping tasks but also possesses strong scalability. In the future, with the enrichment of training data and improvements in computing power, CVPUNet is expected to be widely used in more complex application scenarios, further improving unwrapping accuracy and processing efficiency.

[0126] Table 1 shows the average processing time for single-image unwrapping using different methods. In comparison, DLPU (0.52s) and FCNet-CA (0.36s) consume more time, are relatively inefficient, and perform worse than CVPUNet and PUNet. CVPUNet (0.21s) consumes slightly more time than PUNet (0.19s), but its results outperform PUNet, demonstrating higher accuracy and effectiveness. Therefore, although CVPUNet takes slightly longer to process, it provides more accurate results, demonstrating a good balance between quality and efficiency.

[0127] Table 1 Filtering time of different methods

[0128]

[0129] To further analyze the performance of different methods, Table 2 presents quantitative analysis results based on the MSE and SSIM metrics. Table 2 shows the performance of each method in terms of MSE and SSIM. CVPUNet outperforms other methods in both MSE and SSIM, particularly with significant advantages in MSE1 (0.4327) and SSIM1 (0.8674), demonstrating its best recovery performance in the first category of metrics. Although PUNet also performs well in MSE and SSIM, it is slightly inferior to CVPUNet in all three metrics, particularly in MSE2 (0.9854) and MSE3 (1.5324). FCNet-CA and DLPU perform relatively poorly in all metrics, particularly in MSE and SSIM, which are significantly lower than other methods. Therefore, CVPUNet demonstrates relatively superior performance in both accuracy and structural similarity, demonstrating its superiority in result recovery.

[0130] Table 2 Comparison of MSE and SSIM of different methods

[0131]

[0132] 2. Real data experimental results

[0133] To further test the performance of the present invention, MOMFNet was applied to two mining areas of Jier Coal Mine in Yanzhou District, Jining City, Shandong Province, China. The data time was October 21, 2023 and February 18, 2024. Two Sentinel-1 data were used. The center of the study area was located at 35°28'51"N and 116°43'16"E (Case 1). Figure 9 The study area is located at 35°29'36"N, 116°51'27"E" (Case 2, point 9B). The two mining areas of Dongtan and Yangcun are also included. The data dates are December 27, 2009 and February 11, 2010, respectively, and are two ALOS-1 PALSAR data. The center of the study area is approximately located at 35°28'15"N, 116°48'09"E" (Case 3, point 9C) and 35°26'26"N, 116°54'15"E" (Case 4, point 9D). The robustness of the unwrapping method will be evaluated from both qualitative and quantitative perspectives.

[0134] like Figure 4The first row shows the unwrapping results for sample 1. From left to right, the unwrapping results are the interferogram, MCF, FCNet-CA, PUNet, and CVPUNet. The interferogram shows distinct circular fringes, indicating strong phase wrapping and strong gradient variations. MCF performs poorly on this sample. After unwrapping, the central region exhibits a large number of fragmented and missing regions, resulting in a loss of continuity and virtually no effective restoration of the phase structure. FCNet-CA effectively restores the overall trend in the center, presenting a continuous circular region, but the edges are blurred and detail recovery is insufficient. Compared to FCNet-CA, PUNet achieves better detail restoration, with a sharper central region and smoother edge transitions. Overall, the unwrapping performance is superior to the previous two methods. CVPUNet further improves the unwrapping performance, maintaining the circular structure while also delivering a more natural and smooth phase transition between the edges and the center, preserving the depth of detail. This results in the best unwrapping performance for this sample. Overall, CVPUNet performs best in the first row, effectively handling regions with strong gradient fringes and maintaining phase continuity.

[0135] The second row shows the unwrapping results for sample 2, in the same order: interferogram, MCF, FCNet-CA, PUNet, and CVPUNet. This sample has fewer fringes and a relatively simple wrapped phase, but it contains some noise interference and localized complex regions. MCF performance remains poor. While some areas are restored, there is significant fragmentation overall, especially in the central region where the morphology is significantly distorted. FCNet-CA is able to extract the main phase transition regions, but the unwrapping range is limited, many details are lost, and background noise is high. PUNet achieves more complete results than FCNet-CA, with better restoration of the central structure, but overall noise suppression is still slightly insufficient. CVPUNet strikes a good balance between denoising, smoothing, and detail recovery, with clear central phase features, excellent background noise suppression, and a complete overall morphology with good continuity. Overall, CVPUNet performs best in the second row of unwrapping results, well-suited to phase recovery requirements in low-complexity wrapping scenarios, significantly outperforming other methods.

[0136] like Figure 5As shown in the figure, for the Dongtan Coal Mine data (sample 3), the performance differences between the different methods are quite significant. First, the DLPU method's unwrapping performance is significantly poor, especially in dealing with noise and artifacts, resulting in a very messy result. Significant distortion appears in the image, the unwrapped regions are difficult to distinguish, and noise severely impacts the recovery of phase information. This demonstrates the DLPU method's limitations when dealing with complex and noisy datasets. In comparison, the FCNet-CA method shows some improvement, with unwrapping results significantly clearer than the DLPU method. However, artifacts and errors persist, especially in areas with edges and high levels of detail, falling short of ideal recovery. While the PUNet method improves noise suppression, it still struggles in phase recovery in complex regions, particularly in image details where significant loss and blurring are still visible. The CVPUNet method, on the other hand, demonstrates significant advantages on the Dongtan Coal Mine dataset, not only accurately recovering phase information but also effectively reducing noise. The unwrapped image is smooth and clear, with relatively good recovery of edges and details. This demonstrates the robustness of the CVPUNet method for noisy and complex datasets, ensuring accuracy while avoiding loss of detail due to oversmoothing.

[0137] In the Yangcun Coal Mine dataset (sample four), the performance of all methods has improved compared to the Dongtan Coal Mine, but the differences are still significant. The DLPU method still cannot provide effective disentanglement results, the noise in the image is still serious, artifacts are frequent, and clear phase information cannot be restored. Similar to the Dongtan Coal Mine, although the FCNet-CA and PUNet methods have reduced the noise to a certain extent, they still have the problem of not being able to accurately restore the phase information of complex areas, especially in the restoration of local details. There is still a large deviation. Although these methods can handle simple areas, their performance in parts with higher data complexity is still insufficient. In comparison, CVPUNet shows better performance on the Yangcun Coal Mine dataset. By comparing the images, we can see that CVPUNet can effectively remove noise while retaining details, and the recovery of phase information is more accurate, and the performance of the edges and complex areas of the image is more prominent.

[0138] Overall, CVPUNet significantly outperformed other methods on both the Dongtan Coal Mine and Yangcun Coal Mine datasets. On the Dongtan Coal Mine dataset, while FCNet-CA and PUNet improved noise levels to some extent, their unwrapping results still fell short of ideal clarity and accuracy, particularly in complex areas. On the Yangcun Coal Mine dataset, CVPUNet not only restored clearer phase information but also demonstrated high robustness, effectively addressing the unwrapping challenges in the presence of noise and complex backgrounds.

[0139] In order to further quantitatively evaluate the reliability of the unwrapping results, the unwrapping phase was converted into vertical deformation (ignoring the influence of horizontal displacement), and the leveling data from different mining areas were selected for comparison. Figure 6 and 7 . Figure 6 As shown in the Dongtan Coal Mine data, the settlement values ​​obtained by the DLPU method fluctuate significantly and deviate significantly from the leveling data. In particular, at the last few points in the figure, the settlement values ​​obtained by the DLPU method are significantly higher than the leveling data, and the trend is unstable, indicating that the method's accuracy in processing these data is poor, possibly affected by noise or other factors. The FCNet-CA method exhibits greater fluctuations compared to the DLPU method, but the overall trend improves, showing a closer fit. However, deviations still exist at the last points, with settlement values ​​somewhat different from the leveling data. While the overall trend is relatively close to the leveling data, there are some deficiencies in detail recovery. The PUNet method performs slightly better, with settlement values ​​closer to the leveling data and a relatively stable overall trend. However, at several key points, the PUNet method still exhibits significant errors, failing to fully and accurately recover settlement changes at these locations. The CVPUNet method performs the best, with settlement values ​​closely matching the leveling data. In particular, at most points, the trend of settlement values ​​is consistent with the leveling data, with relatively small errors. CVPUNet not only performs stably in flat areas but also effectively captures actual settlement changes in areas with large gradients, maintaining high accuracy. Overall, the settlement values ​​generated by the CVPUNet method are most consistent with the leveling data.

[0140] Figure 7As shown in the Yangcun Coal Mine data, the settlement values ​​obtained by the DLPU method still exhibit significant fluctuations and diverge significantly from the trend of the leveling data. In particular, at several points in the center of the figure, the settlement values ​​obtained by the DLPU method are significantly lower than the leveling data, and the trend is unstable, indicating that the method's accuracy in processing these data is poor, possibly due to noise or other factors. While the FCNet-CA method shows some improvement over the DLPU method, exhibiting a decreasing trend, significant deviations still exist at points in the middle, resulting in lower settlement values ​​overall, with a significant gap between the settlement values ​​and the leveling data. While the overall trend is relatively close to the leveling data, there are some deficiencies in detail recovery, possibly due to the method's suboptimal response to areas of large gradient deformation. The PUNet method performs slightly better, with settlement values ​​closer to the leveling data and a relatively stable overall trend. However, at several key points at the beginning and end, the PUNet method still exhibits significant errors, failing to fully and accurately recover the settlement changes at these locations. This suggests that the PUNet method underperforms at certain extreme or abrupt points, possibly due to model stability and adaptability. The CVPUNet method performed relatively well, with settlement values ​​closely matching the leveling data. At most points, the trends in settlement values ​​were highly consistent with the leveling data, with minimal error. CVPUNet also captured actual settlement changes well in areas with large gradients, maintaining high accuracy. Overall, the CVPUNet method achieved the highest agreement with the leveling data, demonstrating its advantages in accuracy and stability, enabling it to accurately reproduce actual settlement conditions.

[0141] In order to quantitatively analyze the performance of each method on the leveling data, the present invention calculated the mean square error (MSE) between the unwrapping results of each method and the leveling data, and the results are shown in Table 3. Among them, MSE1 and MSE2 correspond to the comparison results of the Dongtan dataset and the Yangcun coal mine dataset, respectively. As can be seen from Table 3, DLPU has the highest MSE value on both datasets, indicating that the method has the lowest accuracy and the deviation between the unwrapping result and the true settlement value is the largest. FCNet-CA has improved compared to DLPU, but the MSE value is still high, and there is still a significant gap compared with the better model. PUNet shows relatively good performance, and the MSE value is significantly reduced. It is worth noting that CVPUNet performed the best among all methods, and obtained the lowest MSE value in both test scenarios, which shows that CVPUNet is not only highly accurate, but also can maintain excellent performance on different datasets and has good generalization ability.

[0142] Table 3 Comparison of the results of different methods and the level number

[0143]

[0144] Further research:

[0145] In the spatial attention enhancement step after step S5, the deformable spatial attention module is specifically as follows:

[0146] First, the weighted feature map is fed into a 3×3 deformable convolutional layer, whose offset is automatically learned through network learning to accommodate the irregular geometry of the phase jump region. After the deformable convolution, the output deformable features enter a two-branch structure: the first branch generates a spatial saliency map through 1×1 channel-wise downscaling convolution (for example, reducing the number of channels from 192 to 48), while the second branch extracts edge-sensitive features through depthwise separable convolution (a 3×3 depthwise convolution kernel combined with a 1×1 pointwise convolution kernel). Subsequently, a Hadamard product (element-wise multiplication) is performed on the outputs of the two branches, and a spatial attention mask is generated using a modified Gaussian Error Linear Unit (GELU) activation function. Finally, the spatial attention mask is gated and fused with the original weighted feature map. Phase gradient features are extracted from the interferogram in step S1. The gating coefficient γ is calculated and weighted by the mask before being element-wise multiplied with the feature map. The resulting enhanced feature map, which is sensitive to deformation, serves as the input for step S6.

[0147] The steps for generating and applying phase gradient features are as follows:

[0148] First, the original phase gradient map is extracted from the input interferogram, and the first-order differential gradients in the azimuth (Gx) and range (Gy) directions are calculated using the improved Sobel-Feldman operator, where the gradient kernel size is dynamically adjusted according to the phase jump density of the current block (for example, when the jump density is higher than the threshold 0.5, the kernel size is set to 5×5, otherwise it is 3×3). Calculate the adaptive scaling factor and edge sensitivity coefficient The gradient magnitude map is nonlinearly enhanced (λ is 0.8, and σ is the standard deviation of a 3×3 local window). Next, a three-level Gaussian pyramid is constructed, and the gradient map is downsampled to obtain features at the original scale, 1 / 2 scale, and 1 / 4 scale. After concatenation in the channel dimension, a 128-dimensional phase gradient descriptor is generated through 3×3 convolution fusion. Finally, the descriptor is cross-modally interacted with the weighted feature map from step S5: a spatial dynamic mask Mgate is generated through a gated recurrent unit (GRU). A 1×1 convolution and a sigmoid function are used to calculate the gating coefficient γ = Sigmoid(Conv1×1([Fweight⊙Mgate,Gmag])), where Fweight is the weighted feature map, Gmag is the gradient magnitude map, ⊙ represents element-wise multiplication, and Conv1 refers to the first convolutional layer in the convolutional neural network. Finally, γ is used to adjust the fusion weight of the spatial attention mask and the feature map.

[0149] This technical solution uses deformable convolution to dynamically adapt to the geometric characteristics of phase jumps, combined with a dual-branch structure to enhance the perception of spatial saliency and edge details. Simultaneously, the physical priors of phase gradients guide the attention mechanism, enabling the network to accurately capture phase changes in high-gradient regions and suppress noise interference in smooth areas. Field measurements have shown that in high-deformation gradient scenarios such as mining areas, edge preservation in the unwrapping results is improved by over 20%, noise-induced artifacts are reduced by 30%, and phase reconstruction error is reduced by 15% compared to traditional methods, effectively improving the accuracy and robustness of phase unwrapping in complex terrain.

[0150] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the description and implementation methods. They can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily implemented. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A method for phase unwrapping of synthetic aperture radar interferometry based on convolutional neural networks and visual transformers, characterized in that: The following steps are involved: S1. Input the synthetic aperture radar interferogram to be unwrapped; S2. Using a convolutional neural network including multiple convolution kernels to extract local features of the interference pattern to generate an initial feature map; S3. Inputting the initial feature map into a multi-scale feature capture module, wherein the multi-scale feature capture module includes multiple parallel convolution blocks, and the parallel convolution blocks use convolution kernels of different sizes to perform multi-scale feature extraction to obtain multiple groups of sub-feature maps; S4, performing batch normalization processing and ReLU activation function processing on multiple groups of sub-feature maps in sequence, and splicing the processed multiple groups of sub-feature maps along the channel dimension into a fused feature map; S5. Inputting the fused feature map into a scale global relationship capture module, the scale global relationship capture module dynamically adjusts the channel weights of the fused feature map through a squeeze-excitation network to generate a weighted feature map; S6. Applying a single-head self-attention mechanism on the weighted feature map to calculate the global similarity between the query matrix, the key matrix, and the value matrix to generate a global attention weight matrix; S7, multiplying the global attention weight matrix by the weighted feature map to obtain a global enhanced feature map; S8. Perform phase reconstruction on the global enhanced feature map through the convolution layer of the convolution kernel, and output an unwrapped continuous phase map.

2. The method for phase unwrapping of synthetic aperture radar interferometry based on convolutional neural network and visual transformer according to claim 1, wherein: The multiple parallel convolution blocks in the multi-scale feature capture module specifically include a first convolution block, a second convolution block, and a third convolution block. The first convolution block uses a 3×3 convolution kernel for feature extraction, the second convolution block uses a 5×5 convolution kernel for feature extraction, and the third convolution block uses a 7×7 convolution kernel for feature extraction. Each parallel convolution block generates 64 feature maps. The fused feature map after the processed multiple groups of sub-feature maps are spliced ​​along the channel dimension contains 192 channels; The convolutional neural network in step S2 uses 64 3×3 convolution kernels to extract local features and generates an initial feature map with a size of 192×180×180, where 192 is the number of channels and 180×180 is the spatial dimension. The convolution kernel of the convolution layer used in the phase reconstruction step has a size of 3×3.

3. The method for phase unwrapping of synthetic aperture radar interferometry based on convolutional neural network and visual transformer according to claim 2, wherein: In steps S6 and S7, in the single-head self-attention mechanism, the query matrix, key matrix, and value matrix are generated by the following steps: The weighted feature map is linearly transformed with a learnable weight matrix to obtain a query matrix, a key matrix, and a value matrix, wherein the learnable weight matrix includes: The global similarity is calculated by performing a dot product operation on the transposed matrix of the query matrix and the key matrix, and dividing the dot product result by the square root value of the dimension of the query matrix or the key matrix, and then normalizing it by the Softmax function to obtain a global attention weight matrix; The global enhanced feature map is generated by multiplying the global attention weight matrix by a value matrix.

4. The method for phase unwrapping of synthetic aperture radar interferometry based on convolutional neural network and visual transformer according to claim 2, wherein: In step S5, the squeeze-excitation network dynamically adjusts the channel weights of the fused feature map through the following steps: Performing a global average pooling operation on the fused feature map to generate a channel descriptor; Input the channel descriptor into the first fully connected layer for linear transformation, the first fully connected layer compresses the dimension of the channel descriptor to 1 / 16 of the original number of channels and processes it through the ReLU activation function; The processed results are input into the second fully connected layer for linear transformation. The second fully connected layer restores the compressed channel descriptors to the original number of channels and generates channel weight coefficients through the Sigmoid activation function; The channel weight coefficient is multiplied by the fusion feature map channel by channel to generate a weighted feature map.

5. The method for phase unwrapping of synthetic aperture radar interferometry based on convolutional neural network and visual transformer according to claim 4, characterized in that: In steps S3 and S5, the network depth of the multi-scale feature capture module and the scale global relationship capture module are both eight layers; The multi-scale feature capture module realizes multi-level local feature extraction through eight layers of stacked convolution blocks. The output of each layer of convolution blocks is fused with the input of the next layer by element-by-element addition. The scale global relationship capture module realizes global feature modeling through an eight-layer stacked squeeze-excitation network and a single-head self-attention mechanism unit. In the scale global relationship capture module, the weighted feature map output by each layer is superimposed with the weighted feature map input by the next layer through a residual connection.

6. The method for phase unwrapping of synthetic aperture radar interferometry based on convolutional neural network and visual transformer according to claim 3, characterized in that: In steps S6 and S7: In the single-head self-attention mechanism, the shape of the input feature matrix is ​​200, 8000, 1, where 200 is the batch size, 8000 is the input sequence length, and 1 is the feature dimension; The learnable weight matrix includes a query weight matrix, a key weight matrix, and a value weight matrix, whose dimensions are 1×1, 1×1, and 1×1 respectively; In the transposed matrix dot product operation of the query matrix and the key matrix, the dimension square root value is the square root value of the characteristic dimension of the query matrix; The Softmax function performs normalization along the length dimension of the input sequence to generate a global attention weight matrix.

7. The method for phase unwrapping of synthetic aperture radar interferometry based on convolutional neural network and visual transformer according to claim 2, wherein: Step S5 is followed by a spatial attention enhancement step: Inputting the weighted feature map into a deformable spatial attention module, generating a dynamic spatial weight distribution that matches the geometric features of the phase jump region through a deformable convolution kernel; The deformable spatial attention module performs the following operations: A 3×3 deformable convolution layer is used to extract deformation features from the weighted feature map, where the offset of the convolution kernel is obtained through learning; The deformation features are input into a two-branch structure. The first branch generates a spatial saliency map through channel reduction convolution, and the second branch extracts edge-sensitive features through depth-wise separable convolution. After performing Hadamard product operation on the dual-branch output, the improved Gaussian error linear unit activation function is used to generate the spatial attention mask; The spatial attention mask is gated and fused with the original weighted feature map, where the gating coefficient is obtained by dynamically calculating the phase gradient feature, and the enhanced feature map with deformation perception capability is finally output as the input of step S6.

8. The method for phase unwrapping of synthetic aperture radar interferometry based on convolutional neural network and visual transformer according to claim 7, characterized in that: The generation and application of the phase gradient feature specifically includes the following steps: a) Extract the original phase gradient map from the interferogram input in step S1, and use the improved Sobel-Feldman operator to calculate the first-order differential gradient along the azimuth and range directions respectively. The gradient kernel size dynamically adapts to the phase jump density of the current processing block. The calculation formula is: in, is the adaptive scaling factor, is the edge sensitivity coefficient, φ is the original interference phase, G x is the azimuthal gradient, G y is the distance gradient; b) Perform nonlinear enhancement on the gradient magnitude map: Where λ is the learnable enhancement coefficient, σ is the standard deviation of the local window, and ° represents element-by-element multiplication; c) Construct a multi-scale phase gradient feature pyramid, obtain the gradient features of [original scale, 1 / 2 scale, 1 / 4 scale] through three-level Gaussian pyramid downsampling, and generate a 128-dimensional phase gradient descriptor after splicing in the channel dimension and fusion through 3×3 convolution; d) Perform cross-modal interaction between the phase gradient descriptor and the weighted feature map output from step S5, and generate a dynamic mask M of spatial dimension through a gated recurrent unit gate , the final gating coefficient is calculated as: γ=Sigmoid(Conv 1×1 ([F weight ☉M gate ,G mag ])) Among them F weight is the weighted feature map, and ⊙ represents element-wise multiplication.

Images