Synthetic aperture radar interferometry phase unwrapping method based on convolutional neural network and visual converter

Abstract

The invention relates to a synthetic aperture radar interferometry phase unwrapping method based on a convolutional neural network and a visual converter, and belongs to the technical field of synthetic aperture radar interferometry. Aiming at the problem that a traditional deep learning method is insufficient in phase unwrapping precision in a complex terrain and noise environment, the method comprises the following steps of inputting a synthetic aperture radar interferogram to be unwrapped, extracting local features by using a convolutional neural network, generating a fusion feature map through a multi-scale feature capture module, and performing phase unwrapping on the fusion feature map; and after channel weights are adjusted through an extrusion-excitation network, a single-head self-attention mechanism is applied to capture a global relationship, and finally a continuous phase diagram is reconstructed through a convolutional layer. According to the method, local feature extraction and global relation modeling are effectively combined, the processing capacity for a phase jump area and noise is improved, the method is suitable for phase unwrapping of large-gradient surface deformation areas such as mining areas, continuous phase information in interference measurement can be accurately recovered, and accurate data support is provided for application such as surface deformation monitoring.

Description

Technical Field

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

[0002] Synthetic aperture radar interferometry (InSAR) obtains target information through phase difference measurement, and phase unwrapping is the key link to restore the continuous phase. Traditional phase unwrapping methods rely on fixed rules or simple models, and it is difficult to adapt to complex terrains and noise interference. With the development of deep learning, although the method based on convolutional neural network has improved the unwrapping efficiency, there are still significant defects:

[0003] First, the local feature extraction ability is limited: the size of the conventional convolution kernel is fixed, and it is unable to effectively capture multi-scale phase features. Especially in the region of irregular phase jumps caused by terrain undulation, key details are easily lost.

[0004] Second, the global relationship modeling is insufficient: traditional networks lack the modeling of the global correlation of the phase, and it is difficult to balance the feature weights of different regions, resulting in error accumulation in the unwrapping results in large-area phase continuous regions.

[0005] Third, the physical prior fusion is insufficient: ignoring the physical properties of the InSAR phase (such as phase gradient, edge-sensitive features), the pure data-driven model is difficult to balance noise suppression and edge preservation, and is easily affected by speckle noise, resulting in artifacts or over-smoothing in the unwrapping results.

[0006] Fourth, the adaptability to complex structures is poor: the modeling of phase jumps in large gradient deformation regions 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 cause 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 unwrapping accuracy and robustness in complex scenarios.

[0008] The research on phase unwrapping (PU) originated from interferometry technology and has been widely applied especially 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 usually can only record the modulus of the phase, The phase is restricted to the interval [-π, π] or [0, 2π], and this process is called phase wrapping. Therefore, in order to obtain the phase difference of the true result, an integer multiple of a full cycle, 2kπ, must be added to or subtracted from the result to unwrap the wrapped phase, so as to more accurately invert the elevation and deformation information of the object under measurement. This is the purpose of the phase unwrapping problem, that is, to restore the original continuous phase. where k is an appropriate integer such that the restored phase remains smooth and conforms to physical laws. Specifically, the mathematical core of phase unwrapping is how to restore the continuous phase from the measured wrapped phase Restore the continuous phase where k is the 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 respectively show schematic diagrams representing before and after phase unwrapping. Summary of the Invention

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

[0014] To achieve these objects and other advantages of the present invention, a synthetic aperture radar interferometry phase unwrapping method based on a convolutional neural network and a vision transformer is provided, including the following steps:

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

[0016] S2. Use a convolutional neural network including multiple convolutional kernels to perform local feature extraction on the interferogram to generate an initial feature map;

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

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

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

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

[0021] S7. Multiply the global attention weight matrix by the weighted feature map to obtain a globally enhanced feature map.

[0022] S8. Perform phase reconstruction on the globally enhanced feature map through a convolutional layer with a convolutional kernel to output an unwrapped continuous phase map.

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

[0024] The fused feature map obtained by concatenating the processed multiple groups of sub-feature maps along the channel dimension contains 192 channels.

[0025] In step S2, the convolutional neural network uses 64 3×3 convolutional kernels for local feature extraction to generate 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 convolutional kernel of the convolutional layer used in the phase reconstruction step is 3×3 in size.

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

[0028] Perform linear transformation on the weighted feature map with learnable weight matrices respectively to obtain the query matrix, key matrix, and value matrix. The learnable weight matrices include;

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

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

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

[0032] Perform 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] Input the processed result into the second fully connected layer for linear transformation. The second fully connected layer restores the compressed channel descriptor to the original number of channels and generates a channel weight coefficient through the Sigmoid activation function;

[0035] Multiply the channel weight coefficient with the fused feature map channel by channel to generate a weighted feature map.

[0036] Preferably, in steps S3 and S5, the network depths 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 stacked convolutional blocks. The output of each layer of convolutional block is fused with the input of the next layer through element-wise addition;

[0038] The scale global relationship capture module realizes global feature modeling through eight stacked squeeze-and-excitation networks 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 of the next layer through 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, and their dimensions are 1×1, 1×1, and 1×1 respectively;

[0042] In the dot product operation between the query matrix and the transposed matrix of the key matrix, the square root value of the dimension is the square root value of the feature 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-wise multiplication;

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

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

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

[0061] where 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, extract local features with different receptive fields through multi-scale convolutional blocks, combine with the squeeze-and-excitation network to dynamically adjust the channel weights, and enhance the attention to key features; use the single-head self-attention mechanism to model the global phase correlation, enhance the ability to capture long-range dependencies, effectively solve the problems of single local features and lack of global relationships in traditional methods, and significantly improve the disentanglement accuracy and noise suppression ability.

[0064] Second, clarify the specific parameters of the multi-scale convolutional block. Through the combination of 3×3, 5×5, and 7×7 convolutional kernels, cover phase features at different scales. The 19-co-channel fusion feature map after concatenation retains rich multi-scale information, providing sufficient input for subsequent global feature modeling; the unified convolutional kernel size and dimension settings ensure the stability and computational efficiency of the network structure, and improve the consistency and accuracy of feature extraction.

[0065] Third, clarify the mathematical calculation process of the self-attention mechanism. Through linear transformation and normalization operations, realize the weighted aggregation of global phase features, enable the network to adaptively focus on the phase correlation of key regions, enhance the ability to model long-range phase changes, and improve the global consistency and smoothness of the disentanglement results.

[0066] Fourth, aggregate global spatial information through global average pooling, and the squeeze-and-excitation operation realizes weight adjustment in the channel dimension, enabling the network to dynamically allocate channel resources according to the importance of phase features. Low-dimensional mapping and non-linear activation enhance the expressive ability of weight coefficients, effectively suppressing noise in irrelevant channels and highlighting key features related to phase unwrapping (such as edge information in high-gradient regions), improving the discriminability and robustness of feature representation.

[0067] Fifth, the deformable convolution kernel adaptively matches the irregular geometric morphology 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 discontinuous boundaries; the phase gradient gating mechanism embeds physical priors (phase gradient information) into the attention fusion process, dynamically adjusting weights according to the severity of local phase changes, realizing the organic combination of data-driven and physical models, and significantly enhancing the unwrapping accuracy and edge preservation ability under complex terrains.

[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 in high-gradient regions; non-linear enhancement and multi-scale pyramid structures strengthen the robustness of gradient features, and cross-modal interaction fuses physical gradients and depth features through gated recurrent units, enabling the attention mechanism to use the direction prior of phase gradients to guide the focusing area, effectively reducing the phase reconstruction error near the fault line, and improving the unwrapping reliability under complex terrains.

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

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

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

[0072] Figure 3 are the simulation data experimental results of Samples 1-3 of different methods;

[0073] Figure 4 are the real data experimental results of Samples 1 and 2 of different methods;

[0074] Figure 5 are the real data experimental results of Samples 3 and 4 of different methods;

[0075] Figure 6 is the comparison between the unwrapping result of Dongtan Coal Mine and the leveling data;

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

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

[0078] Figure 9 General situation map of the study area. Specific implementation manners

[0079] The following further describes the present invention in detail with reference to the accompanying drawings, so that those skilled in the art can implement it according to the description in the specification.

[0080] It should be noted that, unless otherwise specified, the experimental methods described in the following implementation manners are all conventional methods, and the reagents and materials, unless otherwise specified, can all be obtained from commercial channels; 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 should not be construed as a limitation of the present invention.

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

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

[0083] S2. Use the convolutional neural network CNN for local feature extraction:

[0084] Use a CNN with 64 3×3 convolutional kernels to perform initial feature extraction on the interferogram, and generate an initial feature map with a size of 192×180×180 (number of channels × spatial dimension).

[0085] S3. Multi-scale feature capture: Input the initial feature map into the multi-scale feature capture module, which includes three parallel convolutional blocks,

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

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

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

[0089] After the three groups of feature maps are batch-normalized (BatchNorm) and activated by ReLU, they are concatenated along the channel dimension into a fused feature map with 192 channels.

[0090] S4. Feature fusion and non-linear processing: The concatenated fused feature map is subjected to multi-level feature enhancement through eight stacked convolutional blocks. The output of each layer is superimposed on the input of the next layer through a residual connection to alleviate the vanishing gradient.

[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-and-excitation network (SENet):

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

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

[0094] where AvgPool represents the global average pooling operation, X is the input feature, and R is its shape, which is 192×180×180 in this solution.

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

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

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

[0098] c) Scaling: The weight coefficients are multiplied element-wise with the fused feature map 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 the single-head self-attention mechanism (SHSA) to the weighted feature map:

[0103] The input feature matrix (with a shape of 200, 8000, 1) passes through the learnable weight matrices W Q 、W K 、W VThe linear transformation is the query matrix (Q), the key matrix (K), and the value matrix (V) (Equation 4);

[0104]

[0105] where X is the input feature matrix, and W is the learned weight matrix Q W K W V are used to generate the query, key, and value respectively.

[0106] Calculate the dot product of Q and the transpose of K, and after scaling, generate the attention weight matrix through Softmax normalization (Equations 5, 6);

[0107]

[0108] where is a scaling factor to prevent the dot product value from being too large and causing the gradient to vanish. D is the dimension of the query (or key). The calculated result is an attention matrix, representing the similarity between each pair of queries and keys.

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

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

[0111] Output = Attention Weights × V Equation 7

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

[0113] S7. Feature Enhancement and Phase Reconstruction:

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

[0115] S8. Output the Unwrapping Result:

[0116] The size of the final phase map is consistent with the input interference Figure 1 and the phase values are restored to the continuous form.

[0117] Such as Figure 8As shown, the overall network architecture of this method consists of five main components: an input layer, a local feature extraction module (CNN), a multi-scale feature capture module, a scale global relationship capture module (SENet + self-attention mechanism), and a phase reconstruction output layer. Figure 8 The interpretations 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] <Experimental results>

[0119] 1. Experimental results of simulation data

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

[0121] In Sample 1, the performance of the DLPU is relatively weak. Despite the use of deep learning methods, there are still obvious deficiencies in the noise suppression of the DLPU, especially in the high-frequency region of the phase diagram, where the noise residue is relatively significant and the image detail recovery is incomplete. When processing images with complex boundaries, the results of the DLPU are relatively blurred at the edges and structures, indicating its limitations in dealing with high-gradient regions and non-linear shapes. The FCNet-CA has improved compared to the DLPU, but there is still blurring when processing high-gradient regions. Although the FCNet-CA performs well in smooth regions, in the high-order phase change region, the unwrapping results show a certain degree of distortion, especially at the image edges and details, and the true phase information is not fully retained. The performance of the PUNet is significantly better than that of the DLPU and the FCNet-CA. It can better remove noise during the unwrapping process and maintain the smoothness of the image, and the phase recovery effect is more accurate in most regions. However, the PUNet still fails to perfectly recover the details in the high-value regions, especially at the edges of complex structures, where there is a slight loss of details. Compared with these methods, the CVPUNet performs the best, especially in the balance between noise removal and detail recovery. The CVPUNet can remove noise without sacrificing important structural information. Its results are very close to the true unwrapping map, especially in the performance in complex boundaries and high-gradient regions. By combining convolutional neural networks with vision transformers, the CVPUNet can accurately capture the subtle changes in the image and effectively avoid the over-smoothing problem when dealing with complex phase changes. Therefore, the CVPUNet demonstrates its strong noise suppression ability and fine structure recovery ability, becoming the best-performing method in this test.

[0122] In Sample 2, DLPU still fails to meet the requirements. Especially when dealing with complex boundaries and regions with high gradient changes, its unwrapping results are full of artifacts. Although DLPU can remove some noise, it still seems inadequate in dealing with complex shapes, especially in the recovery of non-linear shapes, where there are significant distortions. Although FCNet-CA has improved in high-noise regions, it still fails to solve the problems of complex boundaries and shape changes well, resulting in a large gap in the detail recovery of the unwrapped image. The performance of PUNet is improved compared to FCNet-CA. Especially in the recovery of smooth regions and most structures, the image shows a relatively smooth effect. However, PUNet still has limitations in dealing with boundaries, especially in high-gradient regions, where obvious detail loss occurs. In this case, CVPUNet once again shows certain advantages. Compared with other methods, CVPUNet can not only accurately remove noise but also retain more details in complex boundaries and high-gradient regions. Especially when dealing with complex geometric shapes, CVPUNet can maintain the fine structure of the image through a multi-level feature extraction network, while avoiding the problem of detail loss caused by excessive noise removal. In addition, the robustness of CVPUNet in practical applications is also particularly prominent, and it can adapt to different types of SAR image data, showing strong applicability.

[0123] For Sample 3, the performance of DLPU is still lacking, especially in the recovery of circular structures, where the unwrapping results are severely over-smoothed, losing a large amount of high-frequency information. This shows that although DLPU can perform phase recovery in some cases, its model limitations are still very obvious when facing complex images, especially when dealing with boundaries and high-gradient regions. Although FCNet-CA has improved the denoising effect to a certain extent, its processing of complex phase images is still not delicate enough. Especially when dealing with high-gradient regions, the results of FCNet-CA have obvious detail blurring and cannot fully recover the complex structures in the true unwrapped image. Although PUNet shows good smoothness in many regions, there is still a lack of details in high-gradient parts, and its recovery effect for complex boundaries is not good. Different from other methods, CVPUNet still performs well in this sample. Through its unique network architecture and multi-objective loss function, CVPUNet can accurately recover the structure and details of the image while effectively removing noise. Especially when dealing with high-value regions and complex boundaries, CVPUNet can avoid detail loss caused by over-smoothing and show efficient noise suppression ability. In contrast, other methods all have certain problems in these aspects. Therefore, CVPUNet once again proves its excellent ability in dealing with complex images in this sample, especially showing certain advantages in complex terrain and high-noise environments.

[0124] Combining the analysis results of multiple samples, CVPUNet demonstrates significant advantages over DLPU, FCNet-CA, and PUNet in aspects such as noise suppression, structure restoration, and detail retention. The key lies in its ability to balance noise suppression and detail retention, effectively removing noise while avoiding over-smoothing and retaining high-frequency detail information, thereby obtaining more accurate and realistic unwrapping results. In addition, CVPUNet performs excellently in handling high-gradient regions and complex phase changes, capable of maintaining fine structures and avoiding edge blurring and detail loss. This benefits from the design of its weighted residual blocks and multi-objective loss function, enabling it to better cope with the challenges of complex shapes and high-noise environments. At the same time, CVPUNet exhibits strong robustness, capable of maintaining stable performance in SAR image processing with different sample data, different terrains, and noise environments, and has high computational efficiency, able to adapt to the processing requirements of large-scale SAR data.

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

[0126] Table 1 shows the average processing time for unwrapping a single image using different methods. In comparison, DLPU (0.52 s) and FCNet-CA (0.36 s) have higher time consumption, relatively lower efficiency, and their results are inferior to those of CVPUNet and PUNet. While CVPUNet (0.21 s) has slightly higher time consumption than PUNet (0.19 s), its result performance is better than that of PUNet, demonstrating higher accuracy and effectiveness. Therefore, although CVPUNet has a slightly longer processing time, it can provide more accurate results, reflecting 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 shows the quantitative analysis results based on the MSE and SSIM metrics. It can be seen from Table 2 that the performance of each method in terms of MSE and SSIM. CVPUNet outperforms other methods in both MSE and SSIM, especially with significant advantages in MSE1 (0.4327) and SSIM1 (0.8674), indicating that it has the best recovery effect in the first type of metrics. Although PUNet also performs well in MSE and SSIM, it is slightly inferior to CVPUNet in all three metrics, especially in MSE2 (0.9854) and MSE3 (1.5324). FCNet-CA and DLPU perform relatively poorly in each metric, especially in terms of MSE and SSIM, which are much lower than other methods. Therefore, CVPUNet shows relatively superior performance both in terms of accuracy and structural similarity, demonstrating its advantage in result recovery.

[0130] Table 2 Comparison of MSE and SSIM of Different Methods

[0131]

[0132] 2. Experimental Results of Real Data

[0133] To further test the performance of the present invention, MOMFNet is applied to two mining areas in the Second Jining Coal Mine, Yanzhou District, Jining City, Shandong Province, China. The data time is October 21, 2023 and February 18, 2024, two Sentinel-1 data. The center of the study area is located at 35°28'51” N, 116°43'16” E (Case 1, Figure 9 Point A) and 35°29'36” N, 116°51'27” E (Case 2, Point B). And two mining areas of Dongtan and Yangcun. The data time is December 27, 2009 and February 11, 2010, 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 C) and 35°26'26” N, 116°54'15” E (Case 4, Point D). The robustness of the unwrapping method will be evaluated from both qualitative and quantitative perspectives.

[0134] As Figure 4As shown, the first row shows the unwrapping effect of Sample 1. From left to right are the unwrapping results of the interferogram, MCF, FCNet-CA, PUNet, and CVPUNet. Observing the interferogram, obvious circular fringes can be seen, indicating that the phase wrapping is very obvious and there is a strong gradient change. MCF performs poorly on this sample. After unwrapping, a large number of broken and missing areas appear in the central region, losing continuity and almost unable to effectively restore the phase structure. FCNet-CA can better restore the overall trend in the center, presenting a continuous circular region, but there is a blur at the edge and insufficient detail recovery. Compared with FCNet-CA, PUNet has better detail recovery, a clearer central region, a smoother edge transition, and the overall unwrapping effect is better than the previous two. The unwrapping effect of CVPUNet is further improved. While maintaining the circular structure, the phase transition between the edge and the central region is more natural and smooth, and the detail level is also retained. It is the best group in this sample. Generally speaking, CVPUNet performs the best in the first row, can better handle the strong gradient fringe region, and maintain phase continuity.

[0135] The second row shows the unwrapping effect of Sample 2. The same order is the interferogram, MCF, FCNet-CA, PUNet, and CVPUNet. This sample has fewer fringes, the wrapped phase is relatively simple, but it contains certain noise interference and local complex regions. MCF still performs poorly. Although some regions are restored, there are obvious fragmentation phenomena as a whole, especially the morphological distortion in the central region is obvious. FCNet-CA can extract the main phase mutation regions, but the unwrapping range is small, many details are lost, and there is more background noise. Compared with FCNet-CA, PUNet has a more complete result, better restoration of the central structure, but the overall suppression of noise is still slightly insufficient. CVPUNet achieves a good balance between denoising, smoothing, and detail recovery. The central phase characteristics are clear, the background noise suppression is excellent, the overall morphology is complete, and the continuity is also good. Generally speaking, CVPUNet also has the best effect in the second row of unwrapping results, can well meet the phase recovery requirements in the low-complexity wrapping scenario, and is significantly better than other methods.

[0136] As Figure 5As shown, for the data of Dongtan Coal Mine (Sample 3), the performance differences of different methods are very obvious. First of all, the unwrapping effect of the DLPU method is significantly poor, especially in dealing with noise and artifacts, and the results appear very messy. There are obvious distortions in the image, the unwrapped area cannot be clearly distinguished, and the noise seriously affects the recovery of phase information. This indicates that the DLPU method has great limitations when facing complex and noisy datasets. In contrast, the FCNet-CA method has been relatively improved. The unwrapping result is significantly clearer than that of DLPU, but there are still certain artifacts and errors, especially in the parts with more edges and details, and the effect has not fully achieved the ideal recovery effect. Although PUNet has made improvements in noise suppression, it is still insufficient in the phase recovery of complex regions. Especially in the detail parts of the image, obvious missing and blurring phenomena can still be seen. The CVPUNet method shows significant advantages on the Dongtan Coal Mine dataset. It not only accurately recovers the phase information, but also effectively reduces the noise. The unwrapped image is smooth and clear, and the recovery of the edge and detail parts is relatively good. This proves that CVPUNet has a certain robustness on noisy and complex datasets, and can avoid the loss of details caused by over-smoothing while ensuring accuracy.

[0137] In the dataset of Yangcun Coal Mine (Sample 4), the performance of all methods has improved compared with that of Dongtan Coal Mine, but the differences are still significant. The DLPU method still cannot provide effective unwrapping results. The noise in the image is still serious, artifacts appear frequently, and clear phase information cannot be recovered. Similar to Dongtan Coal Mine, although the FCNet-CA and PUNet methods have reduced the noise to a certain extent, they still have problems in accurately recovering the phase information of complex regions, especially there are still large deviations in the reduction of local details. Although these methods can handle simple regions, their performance is still insufficient in the parts with higher data complexity. In contrast, CVPUNet shows better performance on the Yangcun Coal Mine dataset. By comparing the images, it can be seen that CVPUNet can effectively remove the noise while retaining the details, and the recovery of the phase information is more accurate, and the performance of the edge and complex regions of the image is more prominent.

[0138] Generally speaking, the performance of CVPUNet on both the Dongtan Coal Mine and Yangcun Coal Mine datasets is significantly better than other methods. On the Dongtan Coal Mine dataset, although the FCNet-CA and PUNet have improved the noise problem to a certain extent, their unwrapping results still cannot reach the ideal clarity and accuracy, especially in the processing of complex regions, the effect is not as good as CVPUNet. On the Yangcun Coal Mine dataset, CVPUNet not only recovers clearer phase information, but also shows high robustness, and can effectively cope with the unwrapping challenges under noise and complex backgrounds.

[0139] To further quantitatively evaluate the reliability of the unwrapping results, the unwrapped phase is converted into vertical deformation (ignoring the influence of horizontal displacement), and the leveling data of different mining areas are selected for comparison as Figure 6 and 7 . Figure 6 (Dongtan Coal Mine) As shown, the settlement values of the DLPU method fluctuate greatly and deviate significantly from the leveling data. Especially in the last few points in the figure, the settlement values of the DLPU method are significantly higher than the leveling data, and the change trend is unstable, indicating that the accuracy of this method in processing these data is poor and may be affected by noise or other factors. The FCNet-CA method fluctuates more than the DLPU, but the overall trend is improved compared to the DLPU, showing a more consistent trend, but there are still deviations at the last points, and there is a certain gap between the settlement values and the leveling data. Although the overall trend is relatively close to the leveling data, there are still some deficiencies in the detailed recovery. The performance of the PUNet method is slightly better, the settlement values are relatively close to the leveling data, and the overall trend is also relatively stable. However, at several key points, the PUNet method still has relatively large errors and fails to accurately recover the settlement changes at these positions. The performance of the CVPUNet method is the best, the settlement values are relatively close to the leveling data, especially at most points, the change trend of the settlement values is consistent with the leveling data, and the error is small. The CVPUNet not only performs stably in the flat area but also can better capture the actual settlement changes in the area with large gradient changes and maintain a high accuracy. Overall, the settlement values of the CVPUNet method have the highest degree of agreement with the leveling data.

[0140] Figure 7As shown in (Yangcun Coal Mine), the settlement values of the DLPU method still have large fluctuations and are quite different from the trend of the leveling data. Especially at several points in the middle of the figure, the settlement values of the DLPU method are significantly lower than the leveling data, and the change trend is unstable, indicating that the accuracy of this method in processing these data is poor and may be affected by noise or other factors. Although the FCNet-CA method has improved compared to DLPU and shows a downward trend, there are still large deviations at the points in the middle, and the overall value is still relatively low, with a certain gap between the settlement value and the leveling data. Although the overall trend is relatively close to the leveling data, there are still certain deficiencies in the detailed recovery, which may be due to the fact that the response of this method in the large-gradient deformation area is not as expected. The performance of the PUNet method is slightly better, the settlement values are relatively close to the leveling data, and the overall trend is also relatively stable. However, at several key points at the beginning and end, the PUNet method still has large errors and fails to accurately recover the settlement changes at these positions completely. This indicates that the performance of PUNet at some extreme or mutation points is not as expected, which may be related to the stability and adaptability of the model. The performance of the CVPUNet method is relatively good, the settlement values are relatively close to the leveling data, especially at most points, the change trend of the settlement values is highly consistent with the leveling data, and the error is small. CVPUNet can also better capture the actual settlement changes in the area of large gradient changes and maintain a high accuracy. Overall, the settlement values of the CVPUNet method have the highest degree of coincidence with the leveling data, indicating that it has certain advantages in terms of accuracy and stability and can more accurately restore the true settlement situation.

[0141] In order to quantitatively analyze the performance of each method on the leveling data, the present invention calculates the mean square error (MSE) between the unwrapping results of each method and the leveling data. 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, the MSE values of DLPU are the highest in both datasets, indicating that this method has the lowest accuracy and the largest deviation between the unwrapping result and the true settlement value. FCNet-CA has improved compared to DLPU, but the MSE value is still relatively high, and there is still an obvious gap compared with the better models. PUNet shows relatively good performance, and the MSE value is significantly reduced. It is worth noting that CVPUNet performs the best among all methods and obtains the lowest MSE values in both test scenarios. This shows that CVPUNet not only has high accuracy but also can maintain excellent performance on different datasets and has good generalization ability.

[0142] Table 3 Comparison of results of different methods with leveling data

[0143]

[0144] Subsequent further research:

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

[0146] First, the weighted feature map is input into a 3×3 deformable convolutional layer, and the offsets of this convolutional layer are automatically obtained through network learning to adapt to the irregular geometric shape of the phase jump region. After being processed by the deformable convolution, the output deformed features enter a two-branch structure: the first branch generates a spatial saliency map through a 1×1 channel reduction convolution (for example, reducing the number of channels from 192 to 48), and the second branch extracts edge-sensitive features through a depthwise separable convolution (a 3×3 depth convolution kernel combined with a 1×1 point convolution kernel). Subsequently, a Hadamard product operation (element-wise multiplication) is performed on the outputs of the two branches, and a spatial attention mask is generated through an improved Gaussian error linear unit (GELU) activation function. Finally, the spatial attention mask is fused with the original weighted feature map in a gated manner: phase gradient features are extracted from the interferogram in step S1, the gating coefficient γ is calculated, and after weighting γ with the mask, it is multiplied element-wise with the feature map to output an enhanced feature map with deformation perception ability, which is used as the input for step S6.

[0147] The generation and application steps of the phase gradient features are as follows:

[0148] First, the original phase gradient map is extracted from the input interferogram, and the first-order difference gradients in the azimuthal direction (Gx) and range direction (Gy) are calculated using an 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 the edge-sensitive coefficient Perform non-linear enhancement processing on the gradient magnitude map (λ is taken as 0.8, and σ is the standard deviation of a 3×3 local window). Then, 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 concatenating in the channel dimension, they are fused through a 3×3 convolution to generate a 128-dimensional phase gradient descriptor. Finally, cross-modal interaction is performed between the descriptor and the weighted feature map in step S5: a spatial dynamic mask Mgate is generated through a gated recurrent unit (GRU), and the gating coefficient γ = Sigmoid(Conv1×1([Fweight⊙Mgate,Gmag])) is calculated through a 1×1 convolution and a Sigmoid function, 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, the fusion weight between the spatial attention mask and the feature map is adjusted through γ.

[0149] In this technical solution, deformable convolutions are used to dynamically adapt to the geometric features of phase jumps, and a dual-branch structure is combined to enhance the perception of spatial saliency and edge details. At the same time, the physical prior of the phase gradient is used to guide the attention mechanism, enabling the network to accurately capture the phase changes in high-gradient regions and suppressing the interference of noise on smooth regions. Through actual measurement, in high-deformation gradient scenarios such as mining areas, the edge retention of the unwrapping results is improved by more than 20%, the artifacts caused by noise are reduced by 30%, and the phase reconstruction error is reduced by 15% compared with traditional methods, effectively improving the accuracy and robustness of phase unwrapping under complex terrains.

[0150] Although the embodiments of the present invention have been disclosed as above, they are not limited to the applications listed in the specification and embodiments. It can be fully applied to various fields suitable for the present invention. For those familiar with the field, additional modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and the illustrated and described examples here.

Claims

1. A synthetic aperture radar interferometry phase unwrapping method based on a convolutional neural network and a vision transformer, characterized in that It includes the following steps: S1. Input the synthetic aperture radar interferogram to be unwrapped; S2. Use a convolutional neural network including multiple convolutional kernels to perform local feature extraction on the interferogram to generate an initial feature map; S3. Input the initial feature map into a multi-scale feature capture module. The multi-scale feature capture module includes multiple parallel convolutional blocks. The parallel convolutional blocks use convolutional kernels of different sizes to perform multi-scale feature extraction to obtain multiple groups of sub-feature maps; S4. Perform batch normalization processing and ReLU activation function processing on multiple groups of sub-feature maps in sequence, and splice the processed multiple groups of sub-feature maps along the channel dimension into a fused feature map; S5. Input 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-and-excitation network to generate a weighted feature map; S6. Apply a single-head self-attention mechanism to the weighted feature map, calculate the global similarity between the query matrix, key matrix, and value matrix, and generate a global attention weight matrix; S7. Multiply the global attention weight matrix by the weighted feature map to obtain a globally enhanced feature map; S8. Perform phase reconstruction on the globally enhanced feature map through a convolutional layer with a convolutional kernel to output an unwrapped continuous phase map.

2. The synthetic aperture radar interferometric phase unwrapping method based on a convolutional neural network and a vision transformer according to claim 1, characterized in that, The multiple parallel convolutional blocks in the multi-scale feature capture module specifically include a first convolutional block, a second convolutional block, and a third convolutional block. The first convolutional block uses a 3×3-sized convolutional kernel for feature extraction, the second convolutional block uses a 5×5-sized convolutional kernel for feature extraction, the third convolutional block uses a 7×7-sized convolutional kernel for feature extraction, and each parallel convolutional block generates 64 feature maps; The fused feature map obtained by splicing the processed multiple groups of sub-feature maps along the channel dimension contains 192 channels; In step S2, the convolutional neural network uses 64 3×3-sized convolutional kernels for local feature extraction to generate 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 convolutional kernel of the convolutional layer used in the phase reconstruction step is 3×3 in size.

3. The method for synthetic aperture radar interferometry phase unwrapping based on convolutional neural network and vision 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 through the following steps: Linearly transform the weighted feature map with learnable weight matrices respectively to obtain the query matrix, key matrix, and value matrix. The learnable weight matrices include; The calculation of the global similarity is performed by taking the dot product of the query matrix and the transposed matrix of the key matrix, dividing the dot product result by the square root of the dimension of the query matrix or the key matrix, and then normalizing it through the Softmax function to obtain the global attention weight matrix; The globally enhanced feature map is generated by multiplying the global attention weight matrix by the value matrix.

4. The synthetic aperture radar interferometry phase unwrapping method based on convolutional neural network and vision transformer according to claim 2, wherein In step S5, the squeeze-and-excitation network dynamically adjusts the channel weights of the fused feature map through the following steps: Perform 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; Input the processed result into the second fully-connected layer for linear transformation. The second fully-connected layer restores the compressed channel descriptor to the original number of channels and generates channel weight coefficients through the Sigmoid activation function; Multiply the channel weight coefficients with the fused feature map channel by channel to generate a weighted feature map.

5. The synthetic aperture radar interferometry phase unwrapping method based on a convolutional neural network and a vision transformer according to claim 4, characterized in that, In steps S3 and S5, the network depths 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 stacked convolutional blocks, and the output of each layer of convolutional blocks is fused with the input of the next layer through element-wise addition; The scale global relationship capture module realizes global feature modeling through eight stacked squeeze-and-excitation networks and single-head self-attention mechanism units. In the scale global relationship capture module, the weighted feature map output by each layer is superimposed on the weighted feature map input of the next layer through residual connection.

6. The synthetic aperture radar interferometric phase unwrapping method based on convolutional neural network and vision 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, and their dimensions are 1×1, 1×1, and 1×1 respectively; In the dot product operation between the query matrix and the transposed matrix of the key matrix, the square root value of the dimension is the square root value of the feature dimension of the query matrix; The Softmax function performs normalization processing along the input sequence length dimension to generate a global attention weight matrix.

7. The synthetic aperture radar interferometry phase unwrapping method based on a convolutional neural network and a vision transformer according to claim 2, characterized in that, After step S5, there is also a spatial attention enhancement step: Input the weighted feature map into the deformable spatial attention module to generate a dynamic spatial weight distribution that matches the geometric features of the phase jump region through a deformable convolution kernel; Among them, the deformable spatial attention module performs the following operations: Adopt a 3×3 deformable convolutional layer to extract deformation features from the weighted feature map, where the offset of the convolution kernel is obtained through learning; Input the deformation features 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 depthwise separable convolution; After performing the Hadamard product operation on the outputs of the two branches, use an improved Gaussian error linear unit activation function to generate a spatial attention mask; Perform gated fusion on the spatial attention mask and the original weighted feature map, where the gating coefficient is dynamically calculated through the phase gradient feature, and finally output an enhanced feature map with deformation perception ability as the input of step S6.

8. The synthetic aperture radar interferometric phase unwrapping method based on a convolutional neural network and a vision transformer according to claim 7, characterized in that The generation and application of the phase gradient feature specifically include the following steps: a), Extract the original phase gradient map from the interferogram input in step S1. Use an improved Sobel-Feldman operator to calculate the first-order difference gradients along the azimuth and range directions respectively, where the gradient kernel size dynamically adapts to the phase jump density of the current processing block. The calculation formula is as follows: Among them, is the adaptive scaling factor, is the edge sensitivity coefficient, φ is the original interference phase, G x is the azimuth gradient, G y is the range gradient; b), Perform non-linear enhancement processing on the gradient magnitude map: where λ is the learnable enhancement coefficient, σ is the standard deviation of the local window, and ° represents element-wise multiplication; c), Construct a multi-scale phase gradient feature pyramid. Obtain the gradient features at [original scale, 1 / 2 scale, 1 / 4 scale] through downsampling of a three-level Gaussian pyramid, and after concatenating in the channel dimension, generate a 128-dimensional phase gradient descriptor through 3×3 convolution fusion; d), perform cross-modal interaction between the phase gradient descriptor and the weighted feature map output in step S5, and generate a dynamic mask M in the spatial dimension through a gated recurrent unit gate , and finally the 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