Method, system and apparatus for multichannel interferogram height reconstruction

Abstract

This invention discloses a method for multi-channel interferometric elevation reconstruction based on a two-stream network, comprising: S1, constructing a two-stream network consisting of an elevation reconstruction stream and a boundary detection stream; S2, constructing a dataset with different terrain features, defining an optimal loss function, and training the two-stream network; S3, inputting the multi-channel interferogram into the trained two-stream network, establishing a dual mapping relationship between the multi-channel interferogram, the elevation map, and the boundary map, and completing the multi-channel interferometric elevation reconstruction. This invention can realize multi-channel interferometric elevation reconstruction based on a two-stream network.

Description

Technical Field

[0001] This invention relates to the field of interferogram reconstruction, and in particular to a method, system, and apparatus for multi-channel interferogram elevation reconstruction. Background Technology

[0002] Interferometric Synthetic Aperture Radar (InSAR) is independent of weather conditions and sunlight. It can obtain elevation information of a terrain surface by applying the known relationship between the phase difference obtained from two SAR images and the terrain height value. It is widely used in various fields, including topographic mapping, surface deformation monitoring, and the detection of special targets. The phase difference obtained from two SAR images is often called the unwound phase. Due to the inverse trigonometric operation of complex InSAR signals, it is mapped to the principal value interval (-π, π) through modulo 2π mapping. Phase unwrapping (PU) is one of the key techniques in interferometric SAR data processing and plays a very important role in the application of InSAR technology. Under strict adherence to the phase continuity assumption (i.e., the absolute value of the phase difference between a pixel and its neighbors is less than π), single-baseline PU methods, represented by the Goldstein branch cut method, quality-guided method, and network flow method, generally work well. However, due to the presence of discontinuities and / or interferometric noise, such as in steep mountains, cliffs, and urban areas, the phase continuity assumption required for a single-baseline PU is not always satisfied. In such cases, the measured wrapped-phase PU becomes an ill-posed problem. Multi-baseline PU eliminates the limitations of the phase continuity assumption that single-baseline PU must follow, and can effectively solve the elevation reconstruction problem of terrain with large slope variations.

[0003] Over the past two decades, multi-channel InSAR elevation reconstruction algorithms based on different strategies (i.e., the multi-baseline PU algorithm in this study) have been proposed, including the Chinese Remainder Theorem (CRT) method [8], Least Squares (LS), Maximum Likelihood (ML), Maximum A posteriori (MAP), Cluster Analysis (CA), Subspace Projection (SPJ), Two-Stage Programming (TSPA), and state estimation methods based on nonlinear filters. The CRT method first constructs the congruence equation of the phase related to the length of the interferometric baseline, and then uses the Chinese Remainder Theorem to solve the equation to obtain the phase corresponding to the observed terrain height value. The CRT method can usually accurately reconstruct the elevation of the observed terrain, provided that its interferometric baseline length strictly meets the relevant conditions required by the method, but its performance is severely limited by the noise present in the interferogram. In order to improve the noise resistance performance of the CRT method, an improved CRT method is reported in this paper. The multi-baseline LS method can be regarded as an extension of the traditional single-baseline LS method. The global solution of the terrain elevation is obtained by averaging the phase weighted by the corresponding baseline length under the LS criterion. Multi-baseline (LS) methods generally outperform CRT methods in terms of noise robustness. However, when extracting elevation gradients from multiple interferograms corresponding to the interferometric baselines, the accuracy of the elevation reconstruction of the observed terrain obtained by this method is significantly reduced, deviating greatly from the true elevation gradient of the observed terrain. Multi-baseline (ML) methods treat the unfolded phase and / or terrain height as parameters within a statistical distribution framework and establish probability density functions for the unfolded phase and terrain height. Optimal parameter estimation is then achieved by maximizing the probability density function constructed according to the ML criterion. ML methods require a large number of observation sources to obtain reliable results, which can significantly increase the cost of implementing a satisfactory multi-channel InSAR system. Similar to ML methods, MAP methods are Markov statistical methods based on Bayesian estimators, where a prior model of the terrain's phase / height is established using Markov random fields. Compared to ML methods, MAP methods reduce the dependence on the number of required interferograms to some extent. Sometimes, even with fewer interferograms than required by ML methods, these methods can obtain reliable elevation estimates of the observed scene, but they require higher computational complexity and more time.

[0004] Methods based on CAP (Charge-Ahead-of-Linear) cluster all pixels in an InSAR interferogram into different groups based on pixel combination information from multiple baselines with varying lengths. Then, they unfold the enclosed pixels group by group using information from each cluster center. These methods are generally effective, but they are often sensitive to measurement bias and noise. Several BM3D noise filtering strategies are applied to improve the robustness and accuracy of CA-based multi-channel elevation reconstruction methods. The SPJ (Single Baseline Noise Reduction) method proposes a joint processing approach to simultaneously perform image registration, phase noise suppression, and power purging (PU) in multi-baseline InSAR systems. A method called TSPA (Transform-Single Baseline Noise Reduction) is proposed to port the single-baseline PU framework to multi-baseline PUs to improve the noise resistance of multi-baseline PUs, providing a perfect approach to improving the elevation reconstruction accuracy of multi-channel InSAR systems. Combining the TSPA method with a tiling strategy based on the multi-baseline LK norm envelope sparsity theorem reduces the computational complexity and memory requirements of PUs on large-scale interferograms, thus effectively handling the PU problem in large-scale interferograms. An improved TSPA-based PU method is proposed, which incorporates an unscented Kalman filter to enhance the algorithm's robustness to noise and measurement bias in observed terrain elevation reconstruction. Another improved TSPA method is proposed to improve the accuracy of interferometric phase or height estimation of observed scenes by jointly processing phase information in local regions to accurately estimate the ambiguity number gradient. State estimation algorithms based on Extended Kalman Filter (EKF), Unscented Kalman Filter (UKF), and Particle Filter (PF) are subsequently proposed. These methods typically use interferometric phase or terrain height as the state variables to be estimated. By establishing derived state-space equations and observation equations representing the phase or height of the terrain, the estimation of PU or terrain height is transformed into a state variable estimation problem. Then, the phase or height estimate of the terrain is obtained by effectively fusing multi-baseline InSAR data. These methods can generally reconstruct elevation from noisy interferograms and have strong noise robustness, but their time consumption is usually high. Furthermore, the PUMA method is applied to PU from a single interferogram, extending it to multi-channel InSAR elevation reconstruction to reduce the impact of noise present in the interferogram on the accuracy of terrain elevation reconstruction in the InSAR system.

[0005] Deep learning (DL) is a machine learning technique that has been successfully applied in speech, automatic control, and image processing, demonstrating significant advantages over traditional techniques. With the rapid development of related hardware such as CPUs and GPUs, DL has gradually been applied to SAR target detection, SAR image filtering, and interferogram processing (PU). In 2019, Teng Zhang transformed the interferogram PU problem into an interferogram semantic segmentation problem. In this problem, the integer phase wrapper counter (PWC) value of a pixel is first obtained using DeepLabv3+, and then the unfolded phase of the pixel is obtained by adding the PWC and its corresponding wrapper phase. In 2021, Pu constructed a DL network called PGNET to improve the estimation accuracy of interferogram phase gradients, and then combined it with the traditional Pu algorithm to obtain highly dependent unfolded phases. Gao proposed a DL network called D-Link to improve the accuracy of interferogram phase gradient estimation, thereby enhancing the robustness of the unscented Kalman filter (PU) algorithm in PU to interferograms with severe noise and large phase gradients. A conditional generative adversarial network, PU-GAN, was proposed to perform PU operations on interferograms (i.e., PU is achieved by using the network to perform the mapping between the wrapped phase and the unfolded phase of the interferogram), and acceptable results were obtained from noisy interferograms. In 2022, Zhou proposed an unsupervised deep convolutional neural network called CANET, which transformed the PU problem of multi-channel InSAR systems into an unsupervised clustering analysis problem, and obtained a popular solution. The above work has promoted the development of DL technology in the field of interferogram PU and provided a perfect approach for the further application of DL technology in the fields of multi-channel InSAR system PU or elevation reconstruction. Summary of the Invention

[0006] The purpose of this invention is to provide a method, system, and apparatus for multi-channel interferogram elevation reconstruction, aiming to solve the problem of multi-channel interferogram elevation reconstruction.

[0007] This invention provides a multi-channel interferogram elevation reconstruction method based on a two-stream network, comprising:

[0008] S1. Construct a two-stream network consisting of elevation reconstruction stream and boundary detection stream;

[0009] S2. Construct a dataset with different terrain features, define the optimal loss function, and train a two-stream network;

[0010] S3. Input the multi-channel interferogram into the trained dual-stream network, establish a dual mapping relationship between the multi-channel interferogram, elevation map, and boundary map, and complete the elevation reconstruction of the multi-channel interferogram.

[0011] This invention provides a system for multi-channel interferometric elevation reconstruction based on a two-stream network, comprising:

[0012] The dual-stream network module is used to construct a dual-stream network consisting of an elevation reconstruction stream and a boundary detection stream.

[0013] The training module is used to construct datasets with different terrain features, define the optimal loss function, and train a two-stream network.

[0014] The reconstruction module is used to input the multi-channel interferogram into the trained dual-stream network, establish a dual mapping relationship between the multi-channel interferogram, elevation map, and boundary map, and complete the elevation reconstruction of the multi-channel interferogram.

[0015] This invention also provides a system for multi-channel interferometric elevation reconstruction based on a dual-stream network, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the above method.

[0016] This invention also provides a computer-readable storage medium storing an information transmission implementation program, which, when executed by a processor, implements the steps of the above-described method.

[0017] Using the embodiments of the present invention, multi-channel interferogram elevation reconstruction based on a dual-stream network can be achieved.

[0018] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0019] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 This is a flowchart of the method for multi-channel interferogram elevation reconstruction based on a dual-stream network according to an embodiment of the present invention;

[0021] Figure 2 This is a training prediction diagram of the method for multi-channel interferogram elevation reconstruction based on a dual-stream network according to an embodiment of the present invention;

[0022] Figure 3 This is a schematic diagram of a dual-stream network for a method of multi-channel interferogram elevation reconstruction based on a dual-stream network according to an embodiment of the present invention.

[0023] Figure 4 This is a schematic diagram of the bneck model of the method for multi-channel interferogram elevation reconstruction based on a dual-stream network according to an embodiment of the present invention;

[0024] Figure 5 This is a schematic diagram of the spatial attention module of the multi-channel interferogram elevation reconstruction method based on a dual-stream network according to an embodiment of the present invention;

[0025] Figure 6 This is a schematic diagram of the residual block of the multi-channel interferogram elevation reconstruction method based on a dual-stream network according to an embodiment of the present invention;

[0026] Figure 7 This is a schematic diagram of a system for multi-channel interferogram elevation reconstruction based on a two-stream network, according to an embodiment of the present invention.

[0027] Figure 8 This is a schematic diagram of a device for multi-channel interferogram elevation reconstruction based on a dual-stream network according to an embodiment of the present invention. Detailed Implementation

[0028] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] Method Implementation Examples

[0030] According to embodiments of the present invention, a system for multi-channel interferometric elevation reconstruction based on a two-stream network is provided. Figure 1 This is a flowchart of the method for multi-channel interferogram elevation reconstruction based on a two-stream network according to an embodiment of the present invention, as follows: Figure 1 As shown, it specifically includes:

[0031] This study proposes a multi-channel InSAR elevation reconstruction method based on deep learning. A dual-stream network, ERSBDS, consisting of an elevation reconstruction stream and a boundary detection stream, is used to reconstruct elevation maps of the observed terrain (i.e., to map the elevations of multiple interferograms obtained from different baselines to the elevations of the observed terrain using the constructed network). First, elevation reconstruction and terrain boundary detection tasks are performed using the elevation reconstruction stream network and the boundary detection stream network, respectively. Second, the elevation reconstruction stream network employs an improved DeepLabv3+, following a classic encoder-decoder structure. To reduce model complexity and improve elevation reconstruction accuracy, the Xception network in the encoder is replaced by a lightweight network based on MobileNetV3, which helps to capture phase information more accurately while reducing the number of network parameters. Spatial attention (SA) modules are added to the encoding and decoding paths to enhance attention to the spatial information of the feature maps. Third, the boundary detection stream network mainly consists of simple and efficient residual modules, capable of performing terrain boundary detection tasks. Finally, the boundary detection stream network converges boundary information into the elevation reconstruction stream, enhancing the network's robustness and effectively improving the elevation reconstruction accuracy of the observed scene. The trained network can directly process multiple interferograms from different baselines to achieve elevation reconstruction without any post-processing procedures.

[0032] 2 Recommended Methods

[0033] The ERSBDS network constructs a suitable network comprising elevation reconstruction stream and boundary detection stream, builds datasets with different terrain features, defines an optimal loss function, selects an appropriate training strategy, and establishes a dual mapping relationship between multi-channel interferograms, elevation maps, and boundary maps. Figure 2 This is a training prediction diagram of the method for multi-channel interferogram elevation reconstruction based on a dual-stream network according to an embodiment of the present invention;

[0034] like Figure 2 As shown in (a), three interferograms of size 256×256 corresponding to three different baseline lengths are combined into a three-channel feature map of size (256×256)×3, which serves as the input to the ERSBDS network. By processing the input interferogram using the trained network, an elevation map of the observed terrain [i.e., a three-channel input feature map of size (256×256)×3, corresponding to three different baseline lengths] can be obtained, as shown in (a). Figure 2 (b) shows the training network. The elevation reconstruction stream reconstructs the elevation map of the observed terrain, while the boundary detection stream detects the boundaries of the observed terrain and updates the network using the optimal loss function described later.

[0035] Figure 3 This is a schematic diagram of a dual-stream network for a method of multi-channel interferogram elevation reconstruction based on a dual-stream network according to an embodiment of the present invention.

[0036] DeepLabV3+ networks, as a classic semantic segmentation architecture, follow an encoder-decoder structure and have been applied in many fields due to their excellent performance. Therefore, this study uses an improved version of DeepLabV3+ as the backbone of the elevation reconstruction flow, such as... Figure 3 As shown. The improved Deep LabV3+ network consists of an encoder and a decoder.

[0037] (1) The encoder mainly consists of four blocks: a deep convolutional neural network (DCNN), a shrinking spatial pyramid pooling (ASPP) module, a spatial attention module, and convolutional layers. The DCNN uses an optical network called MobileNetV3 to replace the modified aligned Xception network used in DeepLabV3+, which can extract more refined feature maps with higher semantic information. The ASSP module fuses feature maps of different scales, and the spatial attention module is used to enhance the attention to the spatial information of the feature maps output by the ASSP module. Then, the number of channels of the output feature map is adjusted using the "1×1 Conv (convolution) + BN (batch normalization) + ReLU (rectified linear unit)" operation. The encoder outputs a 256-channel feature map with high semantic information and a resolution of 16×16.

[0038] (2) In the decoding stage, the low-level feature map output by DCNN (i.e., the feature map is output by the sixth bneck unit in MobileNetV3) is adjusted by "1×1 convolution + BN + ReLU" operation to obtain a 128-channel feature map of size 64×64 (represented as DCNN feature map). At the same time, the feature map with high semantic information output by the encoder is subjected to "4+1×1 convolution upsampling" operation (i.e. 4×4 bilinear interpolation upsampling convolution operation) to obtain a 256-channel feature map of size 64×64, which is concatenated with the DCNN feature map and further processed by the spatial attention module and "1×1 Conv + BN + ReLU" unit, as well as "4+1×1 convolution upsampling" unit to obtain a 256-channel feature map of size 256×256 (represented as ERS feature map). Then, the 32-channel feature map containing phase details and edge information output by the Res block unit in the boundary detection stream is concatenated with the ERS feature map and further processed by the “3×3Conv+BN+ReLU” unit to obtain an elevation map of the observed terrain corresponding to the wrapped phase image input from the network.

[0039] A. Mobile Network V3

[0040] As a lightweight network, MobileNetV3 combines depthwise separable convolutions, inverse residual structures, and compression and activation (SE) units, significantly reducing the number of parameters while maintaining the network's ability to fully extract feature information from multi-channel interferograms. The DCNN module uses the MobileNetV3 network, where "Input" is the dimension of the input feature map, "exp-size" is the number of channels in the feature map output by the first layer of the bneck unit's 1×1 convolution, "out" is the number of channels in the output feature map, "SE" indicates whether an SE unit is used, "NL" indicates which activation function is used, "HS" represents the h-swish activation function, "RE" represents the ReLU activation function, and "s" represents the stride of the convolution.

[0041] Figure 4 This is a schematic diagram of the bneck model of the method for multi-channel interferogram elevation reconstruction based on a dual-stream network according to an embodiment of the present invention;

[0042] Figure 4 The bneck model in MobileNetV3 is shown.

[0043] B. Spatial Attention Module

[0044] Figure 5 This is a schematic diagram of the spatial attention module of the multi-channel interferogram elevation reconstruction method based on a dual-stream network according to an embodiment of the present invention;

[0045] Spatial attention module such as Figure 5 As shown, the input feature map (Fin) is processed through parallel average pooling and max pooling, then concatenated for further processing via a "3×3 Conv" operation and a sigmoid function unit to obtain the weights for each feature channel. Finally, the weights are multiplied by the input feature map (Fin) to obtain the output feature map. The spatial attention module allows the network to focus more on important spatial information in the interferogram and helps generate more effective feature maps.

[0046] Figure 6 This is a schematic diagram of the residual block of the multi-channel interferogram elevation reconstruction method based on a dual-stream network according to an embodiment of the present invention;

[0047] Boundary detection flow such as Figure 3 As shown, it contains three residual blocks (Res blocks, such as...) Figure 6(As shown) and four "Conv+BN+ReLU" units. The boundary information output from the third Res block is merged into the elevation reconstruction stream to enhance its sensitivity to complex terrain boundaries, thereby improving the accuracy of elevation reconstruction. The final convolutional unit reduces the number of channels in the feature map and outputs the estimated boundary map. The structural parameters of the Res block and "Conv+BN+ReLU" units in the boundary detection stream are shown, where "input" represents the size of the input feature map and "channels" represents the number of output channels for each convolutional layer.

[0048] The ERSBDS network establishes a dual mapping between multi-channel interferograms, elevation maps, and boundary maps, essentially functioning as a dual-output network. Therefore, two suitable loss functions are considered: the mean square error (MSE) of the elevation reconstruction stream and the binary cross-entropy (BCE) of the boundary detection stream. The MSE and BCE are calculated as follows:

[0049]

[0050] Where Loss1 is the loss of the elevation reconstruction flow and Loss2 is the loss of the boundary detection flow. N is the number of image pixels, and Yi is the height value at pixel i in the ground truth elevation map (i.e., the label image). is the height value at pixel i of the estimated elevation map obtained by ERSBDS Net; yi represents the binary value at pixel i of the binary label image representing the boundary of the true elevation map, which can be 0 or 1; p(yi) is the probability of label yi obtained by ERSBDS Net; loss refers to the sum of the net losses of ERSBDS.

[0051] This invention employs an improved DeepLabV3+ architecture for elevation reconstruction. The Xception network is replaced in the encoder with a lightweight network called MobileNetV3, reducing network parameters while maintaining performance. Spatial attention modules are then added to the encoding and decoding paths to enhance the network's focus on spatial information in the feature maps. Secondly, the boundary detection stream primarily consists of residual blocks, which detect boundary information of the observed terrain and incorporate it into the elevation reconstruction stream to improve the accuracy of elevation estimation for the observed scene. Finally, a suitable dataset is constructed to enable the trained network to accurately reconstruct the elevation map of the observed scene. Experiments on multi-channel InSAR elevation reconstruction of the observed scene demonstrate the effectiveness of the proposed method and show its advantages in accuracy and efficiency compared to some of the most commonly used methods.

[0052] System Implementation Examples

[0053] According to embodiments of the present invention, a system for multi-channel interferometric elevation reconstruction based on a two-stream network is provided. Figure 7 This is a schematic diagram of a system for multi-channel interferogram elevation reconstruction based on a dual-stream network according to an embodiment of the present invention, as shown below. Figure 7 As shown, it specifically includes:

[0054] The dual-stream network module is used to construct a dual-stream network consisting of an elevation reconstruction stream and a boundary detection stream.

[0055] The training module is used to construct datasets with different terrain features, define the optimal loss function, and train a two-stream network.

[0056] The reconstruction module is used to input the multi-channel interferogram into the trained dual-stream network, establish a dual mapping relationship between the multi-channel interferogram, elevation map, and boundary map, and complete the elevation reconstruction of the multi-channel interferogram.

[0057] The dual-stream network module is specifically used to construct a dual-stream network consisting of an elevation reconstruction stream and a boundary detection stream, wherein the dual-stream network includes an encoder and a decoder.

[0058] The encoder specifically comprises four blocks: a deep convolutional neural network, a shrinking spatial pyramid pooling module, a spatial attention module, and a convolutional layer connected in sequence. The deep convolutional neural network is used to extract feature maps of different scales that provide high-level semantic information from the multi-channel interferogram. The shrinking spatial pyramid pooling module is used to fuse feature maps of different scales. The spatial attention module is used to enhance the attention to the spatial information of the fused feature map output by the shrinking spatial pyramid pooling module. A convolutional layer with 1×1 convolution, batch normalization, and rectified linear units is used to adjust the number of channels in the output feature map. The encoder outputs a 256-channel feature map with high semantic information and a resolution of 16×16.

[0059] The decoder specifically includes:

[0060] In the decoding stage, the deep convolutional neural network outputs a low-level feature map. This low-level feature map is then adjusted using a 1×1 convolution + BN + ReLU operation to obtain a 128-channel feature map of size 64×64. The feature map output by the encoder, which has high semantic information, is then upsampled using a 4+1×1 convolution to obtain a 256-channel feature map of size 64×64. This channel feature map is concatenated with the output feature map and further processed by the spatial attention module and a 1×1 Conv + BN + ReLU unit. After processing, it is upsampled by a 4+1×1 convolution to obtain a 256-channel feature map of size 256×256. Then, the 32-channel feature map containing phase details and edge information output by the residual block unit in the boundary detection stream is concatenated with the residual block feature map and further processed by a 3×3 Conv + BN + ReLU unit to obtain an elevation map of the observed terrain corresponding to the wrapped phase image input from the network, thus completing the reconstruction.

[0061] The embodiments of the present invention are system embodiments corresponding to the above method embodiments. The specific operation of each module can be understood by referring to the description of the method embodiments, and will not be repeated here.

[0062] Device Example 1

[0063] This invention provides a schematic diagram of a device for multi-channel interferogram elevation reconstruction based on a dual-stream network, as shown in the embodiment of the invention. Figure 8 As shown, it includes: a memory 80, a processor 82, and a computer program stored on the memory 80 and executable on the processor 82. When the computer program is executed by the processor, it implements the steps in the above method embodiments.

[0064] Device Example 2

[0065] This invention provides a computer-readable storage medium storing an information transmission implementation program, which, when executed by a processor 82, implements the steps described in the above method embodiments.

[0066] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions to the technical solutions of the embodiments of the present invention do not cause the essence of the corresponding technical solutions to deviate from the scope of the present solution.

Claims

1. A method for multi-channel interferometric elevation reconstruction based on a two-stream network, characterized in that, include: S1. Construct a two-stream network consisting of elevation reconstruction stream and boundary detection stream; S2. Construct a dataset with different terrain features, define the optimal loss function, and train a two-stream network; S3. Input the multi-channel interferogram into the trained dual-stream network, establish a dual mapping relationship between the multi-channel interferogram, elevation map, and boundary map, and complete the elevation reconstruction of the multi-channel interferogram. The boundary detection stream mainly consists of residual blocks, which detect the boundary information of the observed terrain and merge it into the elevation reconstruction stream.

2. The method according to claim 1, characterized in that, S1 specifically includes: constructing a two-stream network consisting of an elevation reconstruction stream and a boundary detection stream, wherein the two-stream network includes an encoder and a decoder.

3. The method according to claim 2, characterized in that, The encoder specifically comprises four blocks: a deep convolutional neural network, a shrinking spatial pyramid pooling module, a spatial attention module, and a convolutional layer connected in sequence. The deep convolutional neural network is used to extract feature maps of different scales containing high-level semantic information from multi-channel interferograms. The shrinking spatial pyramid pooling module is used to fuse feature maps of different scales. The spatial attention module is used to enhance the attention to the spatial information of the fused feature map output by the shrinking spatial pyramid pooling module. A convolutional layer consisting of 1×1 convolutions, batch normalization, and rectified linear units is used to adjust the number of channels in the output feature map. The encoder outputs a 256-channel feature map with high semantic information and a resolution of 16×16.

4. The method according to claim 3, characterized in that, The decoder specifically includes: In the decoding stage, the deep convolutional neural network outputs a low-level feature map. This low-level feature map is then adjusted using a 1×1 convolution + BN + ReLU operation to obtain a 128-channel feature map of size 64×64. The feature map output by the encoder, which has high semantic information, is then upsampled using a 4+1×1 convolution to obtain a 256-channel feature map of size 64×64. This channel feature map is concatenated with the output feature map and further processed by the spatial attention module and a 1×1 Conv + BN + ReLU unit. After processing, it is upsampled by a 4+1×1 convolution to obtain a 256-channel feature map of size 256×256. Then, the 32-channel feature map containing phase details and edge information output by the residual block unit in the boundary detection stream is concatenated with the residual block feature map and further processed by a 3×3 Conv + BN + ReLU unit to obtain an elevation map of the observed terrain corresponding to the wrapped phase image input from the network, thus completing the reconstruction.

5. A system for multi-channel interferometric elevation reconstruction based on a two-stream network, characterized in that, include: The dual-stream network module is used to construct a dual-stream network consisting of an elevation reconstruction stream and a boundary detection stream. The training module is used to construct datasets with different terrain features, define the optimal loss function, and train a two-stream network. The reconstruction module is used to input the multi-channel interferogram into the trained dual-stream network, establish a dual mapping relationship between the multi-channel interferogram, elevation map, and boundary map, and complete the elevation reconstruction of the multi-channel interferogram. The boundary detection stream mainly consists of residual blocks, which detect the boundary information of the observed terrain and merge it into the elevation reconstruction stream.

6. The system according to claim 5, characterized in that, The dual-stream network module is specifically used to: construct a dual-stream network consisting of an elevation reconstruction stream and a boundary detection stream, wherein the dual-stream network includes an encoder and a decoder.

7. The system according to claim 6, characterized in that, The encoder specifically comprises four blocks: a deep convolutional neural network, a shrinking spatial pyramid pooling module, a spatial attention module, and a convolutional layer connected in sequence. The deep convolutional neural network is used to extract feature maps of different scales containing high-level semantic information from multi-channel interferograms. The shrinking spatial pyramid pooling module is used to fuse feature maps of different scales. The spatial attention module is used to enhance the attention to the spatial information of the fused feature map output by the shrinking spatial pyramid pooling module. A convolutional layer consisting of 1×1 convolutions, batch normalization, and rectified linear units is used to adjust the number of channels in the output feature map. The encoder outputs a 256-channel feature map with high semantic information and a resolution of 16×16.

8. The system according to claim 7, characterized in that, The decoder specifically includes: In the decoding stage, the deep convolutional neural network outputs a low-level feature map. This low-level feature map is then adjusted using a 1×1 convolution + BN + ReLU operation to obtain a 128-channel feature map of size 64×64. The feature map output by the encoder, which has high semantic information, is then upsampled using a 4+1×1 convolution to obtain a 256-channel feature map of size 64×64. This channel feature map is concatenated with the output feature map and further processed by the spatial attention module and a 1×1 Conv + BN + ReLU unit. After processing, it is upsampled by a 4+1×1 convolution to obtain a 256-channel feature map of size 256×256. Then, the 32-channel feature map containing phase details and edge information output by the residual block unit in the boundary detection stream is concatenated with the residual block feature map and further processed by a 3×3 Conv + BN + ReLU unit to obtain an elevation map of the observed terrain corresponding to the wrapped phase image input from the network, thus completing the reconstruction.

9. A device for multi-channel interferometric elevation reconstruction based on a two-stream network, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the method for multi-channel interferogram elevation reconstruction based on a dual-stream network as described in any one of claims 1 to 4.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an implementation program for information transmission, which, when executed by a processor, implements the steps of the method for multi-channel interferogram elevation reconstruction based on a dual-stream network as described in any one of claims 1 to 4.

Images