InSAR phase unwrapping method based on fourier domain global hybrid network

Abstract

The application discloses an InSAR phase unwrapping method based on a Fourier domain global hybrid network, and relates to the technical field of remote sensing, and the method comprises the following steps: generating synthetic data of diversified continuous phases through data simulation, superimposing compound noise to obtain a data pair of noisy wrapped phases and real phases; a phase unwrapping model is constructed, a UNet encoder-decoder structure is used as a backbone network, a Fourier mixed residual block FMRB is embedded in a bottleneck layer, an all-space pyramid pooling ASPP module is connected, and a convolution block attention module CBAM is integrated; the phase unwrapping model is trained, a compound loss function is used as an optimization objective, phase unwrapping model parameters are iteratively optimized until the phase unwrapping model converges; and an InSAR interferogram to be unwrapped is input into the phase unwrapping model, and unwrapped phase results are output. The application comprehensively constrains the unwrapping process from multiple physical dimensions through the compound loss function, and the robustness in a complex noise environment and the physical reasonableness of a solution are enhanced.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of remote sensing, and particularly relates to an InSAR phase unwrapping method based on a Fourier domain global hybrid network. BACKGROUND

[0002] Interferometric Synthetic Aperture Radar (InSAR) is an all-weather and all-time active microwave remote sensing technology, which can invert the ground deformation field with centimeter to millimeter level precision, and the core observation quantity is the interference phase. However, due to the periodicity of radar wavelength, the interference phase is wrapped in the interval [-pi, pi), and the true continuous phase needs to be recovered through phase unwrapping, and the accuracy directly determines the quality of the InSAR deformation product, which is the core bottleneck of the technology chain.

[0003] The existing unwrapping technology is a traditional optimization method represented by the least square method, which obtains a globally smooth solution by solving the Poisson equation, and although it can suppress high-frequency noise, excessive smoothing will blur high-gradient structures such as faults and ground fissures, suppress high-frequency deformation amplitudes, and also easily cause low-frequency errors such as plane tilting; and a deep learning method represented by UNet, which learns the nonlinear mapping end to end, but due to the locality of the convolution kernel, it is difficult to model long-distance pixel dependence, and it is insufficient to recover large-scale slowly varying deformation, and the attenuation of deep information easily leads to low-frequency errors, which restricts its reliability in high-precision scenarios. SUMMARY

[0004] The purpose of the present application is to provide an InSAR phase unwrapping method based on a Fourier domain global hybrid network to solve the problems in the prior art.

[0005] To achieve the above purpose, the present application provides the following technical scheme: an InSAR phase unwrapping method based on a Fourier domain global hybrid network, which comprises:

[0006] S100, generating synthetic data of diversified continuous phases through data simulation, superimposing compound noise of atmospheric noise, Gaussian phase noise and isolated residual point noise, and wrapping the phase to a preset interval to obtain a data pair of noisy wrapped phase and true phase, and forming a data set;

[0007] S200, constructing a phase unwrapping model, wherein the phase unwrapping model takes a UNet encoder-decoder structure as a backbone network; a Fourier hybrid residual block FMRB is embedded in the bottleneck layer of the backbone network; an empty space pyramid pooling ASPP module is connected after the Fourier hybrid residual block FMRB; and a convolution block attention module CBAM is integrated in the convolution block of the backbone network;

[0008] S300, training the phase unwrapping model by using the data set, taking a composite loss function as an optimization objective, and iteratively optimizing phase unwrapping model parameters until the phase unwrapping model converges; the loss function comprises a weighted combination of a data fidelity term, a gradient consistency term, a spectral distribution constraint term, and a re-winding consistency term;

[0009] S400, inputting the InSAR interferogram to be unwrapped into the phase unwrapping model, and outputting an unwrapped phase result.

[0010] According to the above scheme, step S100 comprises:

[0011] S110, generating a continuous phase field with diversified spatial distribution characteristics by hybridizing a bicubic interpolation method and a Gaussian function superposition method; the bicubic interpolation method is used to generate a spatial structure of a basic continuous phase, so as to ensure the smoothness and continuity of the spatial distribution of the phase field; and the Gaussian function superposition method is used to simulate phase changes of different amplitudes and different spatial scales, and the two methods are used in cooperation to generate the continuous phase field with diversified spatial distribution characteristics;

[0012] S120, sequentially superimposing simulated atmospheric phase screen noise, random phase noise obeying a Gaussian distribution, and isolated residual point noise caused by phase singular points on the continuous phase field to obtain a noisy phase field; the atmospheric phase screen noise simulates the low-frequency influence of atmospheric disturbance on the phase, the Gaussian random phase noise simulates high-frequency noise caused by temporal or spatial decorrelation, and the isolated residual point noise simulates local mutation type noise caused by the phase singular points; the three kinds of noise are sequentially superimposed according to the noise distribution law of the real InSAR interferogram, so as to ensure that the noisy phase field is close to the real scene;

[0013] S130, performing phase wrapping operation on the noisy phase field, and constraining the value range of the noisy phase field to a preset [-π, π) interval to obtain a noisy wrapped phase;

[0014] S140, constructing the noisy wrapped phase and the corresponding original continuous phase into a data pair to form a data set.

[0015] According to the above scheme, step S200 comprises:

[0016] S210, construct a basic framework with a UNet encoder-decoder structure as a backbone network, the encoder gradually extracts multi-scale features through convolution and down-sampling operations, the decoder gradually restores the spatial resolution through up-sampling and convolution operations, and the encoder and the decoder are connected through a skip connection between corresponding levels to realize feature fusion; the encoder includes consecutive convolution layers and down-sampling layers; the convolution layer is used to extract feature details; the down-sampling layer reduces the spatial resolution of the feature map through convolution operation and expands the receptive field; the decoder includes consecutive up-sampling layers and convolution layers, the up-sampling layer improves the spatial resolution of the feature map through transposed convolution operation; the convolution layer adopts a convolution kernel and is used to restore feature details; the skip connection directly splices the feature map of a certain layer of the encoder and the feature map of the corresponding level of the decoder in the channel dimension, realizes the direct fusion of shallow detail features and deep semantic features, and avoids the loss of detail information;

[0017] S220, embed a Fourier mixed residual block FMRB at the bottleneck layer of the backbone network, the Fourier mixed residual block FMRB reduces the computational complexity through dimension reduction projection, and performs global spatial mixing and global channel mixing operations in the Fourier domain in parallel, and fuses through a residual connection and an input to establish a long-distance pixel dependence relationship;

[0018] S230, connect an atrous spatial pyramid pooling ASPP module at the output end of the Fourier mixed residual block FMRB, capture multi-scale context information by setting a plurality of dilated convolution layers with different expansion rates in parallel;

[0019] S240, integrate a convolution block attention module CBAM in each convolution block of the backbone network, and adaptively refine feature expression through a serially connected channel attention submodule and a spatial attention submodule.

[0020] According to the above scheme, step S220 includes:

[0021] S221, compress the input features in the channel dimension through a 1x1 convolution layer to obtain compressed features;

[0022] S222, perform a fast Fourier transform on the compressed feature to obtain a frequency domain feature; perform a first branch processing and a second branch processing in parallel in the frequency domain, the first branch processing performing a weighting processing on low-frequency components of the frequency domain feature through a learnable complex weight matrix to obtain a spatial mixed feature; the second branch processing performing a cross-channel information fusion on the frequency domain feature through a 1*1 grouped convolution, and applying an adaptive soft threshold operation on the fused frequency domain feature to obtain a channel mixed feature; the fast Fourier transform converts the compressed feature in the spatial domain into the frequency domain feature, separating the low-frequency components and the high-frequency components; in the first branch, the dimension of the learnable complex weight matrix is completely matched with the dimension of the low-frequency components of the frequency domain feature, and the weight value is adaptively adjusted through network training to realize targeted weighting on different low-frequency modes and to strengthen global spatial dependency modeling; in the second branch, the number of groups of the 1*1 grouped convolution is set according to the number of channels of the frequency domain feature to ensure the efficiency of the cross-channel information fusion, and the adaptive soft threshold operation dynamically sets the threshold value based on the amplitude distribution of the frequency domain feature to suppress the noise or irrelevant frequency components below the threshold value and to retain the effective frequency components above the threshold value;

[0023] S223, perform an inverse fast Fourier transform on the spatial mixed feature and the channel mixed feature to obtain corresponding spatial domain features; after superimposing the spatial domain features, perform a channel dimension recovery through a convolution layer to obtain a reconstructed feature; fuse the reconstructed feature and the module input through a residual connection to obtain an output feature with a long-distance pixel dependency relationship in a full image range; the inverse fast Fourier transform converts the spatial mixed feature and the channel mixed feature in the frequency domain back to the spatial domain; the feature superposition adopts an element-by-element addition manner to integrate the information of the two types of mixed features; the number of convolution kernels of the convolution layer is consistent with the number of channels of the FMRB module input feature, so as to ensure that the channel dimension of the reconstructed feature is completely matched with the input feature; the residual connection adds the reconstructed feature and the original input feature of the module element by element to avoid attenuation or loss of feature information in the processing process.

[0024] According to the above scheme, step S230 includes:

[0025] S231, in the ASPP module, N parallel dilated convolution layers are configured, N is a positive integer, and the dilation rates of the dilated convolution layers are different from each other;

[0026] S232, the output feature of the FMRB is convolved by using the dilated convolution layers with different dilation rates to obtain N groups of feature maps with different scales;

[0027] S233, fuse the N groups of feature maps with different scales to obtain an output feature with aggregated multi-scale context information.

[0028] According to the above scheme, step S240 includes:

[0029] S241, in the channel attention submodule, performing global average pooling and global maximum pooling operations on the feature maps output by the backbone network convolution block respectively to obtain two one-dimensional feature vectors;

[0030] S242, inputting the two one-dimensional feature vectors into the shared multi-layer perception, generating a channel attention weight vector after activation function processing, multiplying the weight vector with the feature maps output by the convolution block channel by channel to obtain channel weighted feature maps;

[0031] S243, in the spatial attention submodule, performing average pooling and maximum pooling operations on the channel weighted feature maps along the channel dimension respectively to obtain two two-dimensional feature maps, concatenating the two two-dimensional feature maps and processing them through a convolution layer and an activation function to generate a spatial attention weight map;

[0032] S244, multiplying the spatial attention weight map with the channel weighted feature maps pixel by pixel to output the feature maps refined in channel and space.

[0033] According to the above scheme, step S300 includes:

[0034] S310, dividing the data set into a training set and a validation set, inputting the noisy wrapped phase data of the training set into the phase unwrapping model, and the phase unwrapping model outputs the corresponding predicted unwrapped phase;

[0035] S320, taking a composite loss function as the optimization objective of model training, the composite loss function being composed of a data fidelity term, a gradient consistency term, a spectral distribution constraint term and an unwrapping consistency term by weighting coefficient combination, and comprehensively constraining the phase unwrapping model output, the formula being as follows:

[0036] L total =w1×L data +w2×L grad +w3×L freq +w4×L wrap ;

[0037] Wherein, L total represents the composite loss function; L data represents the data fidelity term; L grad represents the gradient consistency term; L freq represents the spectral distribution constraint term; L wrap represents the unwrapping consistency term; w1, w2, w3 and w4 represent the weight coefficients of each term;

[0038] S330. An optimizer is used to iteratively update the parameters of the phase unwrapping model. After each iteration, the validation set is input into the phase unwrapping model. The predicted unwrapped phase output by the phase unwrapping model is mean aligned. The root mean square error and structural similarity index are used as evaluation indicators to monitor the performance of the phase unwrapping model.

[0039] S340. When the evaluation index of the phase unwrapping model on the validation set remains stable for a continuous preset period and no longer changes significantly, the phase unwrapping model is determined to have converged, and the iteration is stopped.

[0040] According to the above scheme, the data fidelity term is calculated based on the difference between the predicted unwrapped phase and the corresponding true phase; the gradient consistency term is calculated based on the spatial gradient difference between the predicted unwrapped phase and the true phase; the spectral distribution constraint term is calculated based on the frequency domain power spectrum difference between the predicted unwrapped phase and the true phase; and the rewrapping consistency term is calculated based on the consistency between the predicted unwrapped phase and the input noisy wrapping phase in the wrapping gradient domain.

[0041] According to the above scheme, step S330 includes:

[0042] S331. Mean alignment processing includes: calculating the pixel mean of the predicted unwrapped phase and the corresponding true phase, obtaining the mean difference of the phase, subtracting the mean difference from the predicted unwrapped phase, eliminating irrelevant global constant phase bias, and obtaining the aligned predicted phase.

[0043] S332, Root Mean Square Error includes, based on the aligned predicted phase and the true phase, calculating the square of the pixel-level phase difference between the two, and taking the square root of the mean of all pixel square differences. This index is used to quantify the amplitude recovery accuracy of the untangled phase.

[0044] S333. The structural similarity index includes linearly stretching both the aligned predicted phase and the true phase to the [0,1] interval, and calculating the weighted mean, weighted variance, and weighted covariance of the two phase maps within a local sliding window of a preset size, as shown in the following formula:

[0045] SSIM=[(2μ1μ2+C1)(2σ 12 +C2)] / [(μ1 2 +μ2 2 +C1)(σ1 2 +σ2 2 +C2)];

[0046] Where SSIM represents the structural similarity index; μ1 and μ2 are the weighted averages of the two phase maps, respectively; σ1 2 and σ2 2 Let σ be the weighted variance of the two phase diagrams. 12Weighted covariance represented as two-phase maps; C1 and C2 are constants to avoid a zero denominator.

[0047] Compared with the prior art, the present application has the beneficial effects that:

[0048] 1、The present application realizes global information mixing in the frequency domain through the Fourier mixed residual block, effectively establishes long-distance pixel dependency, significantly suppresses low-frequency errors in unwrapping results, and improves the accuracy of geodetic surveying;

[0049] 2、The present application realizes effective aggregation of multi-scale context information and precise enhancement of key features through the synergistic effect of the hollow spatial pyramid pooling module and the convolution block attention module, while maintaining global smoothness, and outstandingly retains local detail features such as faults and cracks;

[0050] 3、The present application comprehensively constrains the unwrapping process from multiple physical dimensions of the pixel domain, the gradient domain, the frequency domain and the wrapping domain through a composite loss function, and combines real noise simulation to greatly enhance the robustness and physical reasonableness of the method in a complex noise environment. BRIEF DESCRIPTION OF DRAWINGS

[0051] Fig. 1 The figure is a step flow chart of the InSAR phase unwrapping method based on the Fourier domain global mixing network of the present application.

[0052] Fig. 2 The figure is a phase unwrapping model structure diagram of the InSAR phase unwrapping method based on the Fourier domain global mixing network of the present application. DETAILED DESCRIPTION

[0053] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0054] Embodiment: As shown in the figure, the present application provides a technical solution, an InSAR phase unwrapping method based on a Fourier domain global mixing network, which comprises the following steps: Figs. 1-2

[0055] S100, generating synthetic data of diversified continuous phases through data simulation, superimposing composite noise of atmospheric noise, Gaussian phase noise and isolated residual point noise, and wrapping the phases to a preset interval to obtain a data pair of noisy wrapped phases and real phases, and forming a data set;

[0056] Specifically, step S100 comprises: ​

[0057] S110, a continuous phase field with diversified spatial distribution characteristics is generated by mixing a bicubic interpolation method and a Gaussian function superposition method; the bicubic interpolation method is used to generate the spatial structure of the basic continuous phase, ensuring the smoothness and continuity of the spatial distribution of the phase field; the Gaussian function superposition method is used to simulate phase changes of different amplitudes and different spatial scales, and the two methods are used to generate a continuous phase field with diversified spatial distribution characteristics;

[0058] For example, the bicubic interpolation method is used to generate a 512*512 pixel basic phase field, the interpolation grid is set to 16*16, and a continuous phase with a smooth change trend is generated. Then, the Gaussian function superposition method is used to randomly generate 5-10 Gaussian kernels, the standard deviation is randomly selected in the range of 20-100 pixels, and the amplitude is randomly set in the range of 0.5π-2π, which is used to simulate deformation characteristics of different spatial scales; this is only an example and is not limited;

[0059] S120, sequentially superimpose simulated atmospheric phase screen noise, Gaussian distributed random phase noise and isolated residual point noise caused by phase singular points on the continuous phase field to obtain a noisy phase field; the atmospheric phase screen noise simulates the low-frequency influence of atmospheric disturbance on the phase, the Gaussian random phase noise simulates the high-frequency noise caused by time or space de-coherence, and the isolated residual point noise simulates the local mutation type noise caused by the phase singular point; the three kinds of noise are sequentially superimposed according to the noise distribution law of the real InSAR interferogram, to ensure that the noisy phase field is close to the real scene;

[0060] For example, the atmospheric phase screen noise is generated by using a fractal noise model, the power spectrum index is set to-2.5, and the amplitude standard deviation is set to 0.3π; the Gaussian random phase noise has a mean of 0 and a standard deviation of 0.2π; the isolated residual point noise randomly generates 50-100 residual points in the phase field, and each residual point causes a phase jump of ±2π; this is only an example and is not limited;

[0061] S130, phase wrapping operation is performed on the noisy phase field to constrain the value range of the noisy phase field to the preset [-π, π) interval, and a noisy wrapped phase is obtained;

[0062] S140, the noisy wrapped phase and the corresponding original continuous phase are constructed into a data pair to form a data set.

[0063] S200, a phase unwrapping model is constructed, and the phase unwrapping model takes a UNet encoder-decoder structure as a backbone network; a Fourier mixed residual block FMRB is embedded in the bottleneck layer of the backbone network; an atrous spatial pyramid pooling ASPP module is connected after the Fourier mixed residual block FMRB; and a convolution block attention module CBAM is integrated in the convolution block of the backbone network;

[0064] Specifically, step S200 comprises:

[0065] S210, a basic framework with a UNet encoder-decoder structure as a backbone network is constructed, the encoder gradually extracts multi-scale features through convolution and down-sampling operations, the decoder gradually restores spatial resolution through up-sampling and convolution operations, and the encoder and the decoder realize feature fusion through a skip connection between corresponding levels; the encoder includes consecutive convolution layers and down-sampling layers; the convolution layers are used to extract feature details; the down-sampling layers reduce the spatial resolution of feature maps through convolution operations and expand the receptive field; the decoder includes consecutive up-sampling layers and convolution layers, the up-sampling layers improve the spatial resolution of feature maps through transpose convolution operations; the convolution layers use convolution kernels to restore feature details; the skip connection directly splices feature maps of a certain layer of the encoder and feature maps of a corresponding level of the decoder in the channel dimension, realizes direct fusion of shallow detail features and deep semantic features, and avoids loss of detail information;

[0066] For example, the encoder includes 4 down-sampling stages, each stage uses a 3x3 convolution kernel, and the number of channels is 64, 128, 256, and 512 in turn; the down-sampling uses a convolution with a step of 2, and the size of the feature map is reduced to 256x256, 128x128, 64x64, and 32x32 in turn; the decoder symmetrically includes 4 up-sampling stages, and uses transpose convolution for 2 times up-sampling; the skip connection splices feature maps of each stage of the encoder and feature maps of a corresponding stage of the decoder in the channel dimension;

[0067] S220, a Fourier mixed residual block FMRB is embedded at a bottleneck layer of the backbone network, the Fourier mixed residual block FMRB reduces the computational complexity through dimension reduction by projection, and performs global spatial mixing and global channel mixing operations in the Fourier domain in parallel, and establishes a long-distance pixel dependency relationship through a residual connection and input fusion;

[0068] Further, step S220 comprises:

[0069] S221, a 1x1 convolution layer is used to compress input features in the channel dimension to obtain compressed features;

[0070] S222, perform a fast Fourier transform on the compressed features to obtain frequency domain features; perform a first branch processing and a second branch processing in parallel in the frequency domain, the first branch processing performing a weighting processing on low frequency components of the frequency domain features through a learnable complex weight matrix to obtain spatial mixed features; the second branch processing performing a cross-channel information fusion on the frequency domain features through a 1x1 grouped convolution, and applying an adaptive soft threshold operation on the fused frequency domain features to obtain channel mixed features; the fast Fourier transform converts the compressed features in the spatial domain into the frequency domain features, separating the low frequency components and the high frequency components; in the first branch, the dimension of the learnable complex weight matrix is completely matched with the dimension of the low frequency components of the frequency domain features, and the weight values are adaptively adjusted through network training to realize targeted weighting on different low frequency modes and to strengthen global spatial dependency modeling; in the second branch, the number of groups of the 1x1 grouped convolution is set according to the number of channels of the frequency domain features to ensure the efficiency of the cross-channel information fusion, and the adaptive soft threshold operation dynamically sets the threshold based on the amplitude distribution of the frequency domain features to suppress the noise or irrelevant frequency components below the threshold and to retain the effective frequency components above the threshold;

[0071] S223, perform an inverse fast Fourier transform on the spatial mixed features and the channel mixed features to obtain corresponding spatial domain features; after superimposing the spatial domain features, perform a channel dimension recovery through a convolution layer to obtain a reconstructed feature; fuse the reconstructed feature and the module input through a residual connection to obtain an output feature with a full image range long distance pixel dependency relationship; the inverse fast Fourier transform converts the spatial mixed features and the channel mixed features in the frequency domain back to the spatial domain respectively; the feature superposition adopts an element-by-element addition manner to integrate the information of the two types of mixed features; the number of convolution kernels of the convolution layer is consistent with the number of channels of the FMRB module input feature, ensuring that the channel dimension of the reconstructed feature is completely matched with the input feature; the residual connection adds the reconstructed feature and the original input feature of the module element by element to avoid attenuation or loss of feature information in the processing process.

[0072] For example: the Fourier mixed residual block FMRB compresses the number of channels of the input feature from 512 to 128 through a 1x1 convolution; the fast Fourier transform converts the feature map of 32x32x128 to the frequency domain; the first branch uses a learnable complex weight matrix with a dimension of 32x32x128 to specially process the 16x16 center component of the low frequency region; the second branch uses a grouped convolution with a group number of 8 to fuse the frequency domain features of 128 channels; in the adaptive soft threshold operation, the threshold is dynamically set according to the 30th percentile of the feature amplitude; the inverse fast Fourier transform returns the processed feature to the spatial domain; after the feature superposition, the number of channels is restored from 128 to 512 through a 1x1 convolution; the residual connection ensures complete information transmission;

[0073] S230, connecting an atrous spatial pyramid pooling (ASPP) module at the output end of the Fourier mixed residual block (FMRB), capturing multi-scale context information by setting different dilated convolution layers in parallel;

[0074] Further, step S230 includes:

[0075] S231, configuring N parallel atrous convolution layers in the atrous spatial pyramid pooling (ASPP) module, N being a positive integer, the dilated rates of the atrous convolution layers being different from each other;

[0076] S232, performing convolution operations on the output features of the Fourier mixed residual block (FMRB) by using atrous convolution layers with different dilated rates respectively, to obtain N groups of feature maps with different scales;

[0077] S233, fusing the N groups of feature maps with different scales to obtain output features with aggregated multi-scale context information.

[0078] For example, the atrous spatial pyramid pooling (ASPP) module is configured with 4 parallel atrous convolution layers, the dilated rates of which are 1, 6, 12 and 18 respectively, and the channel numbers of the output feature maps are all 256, which are finally spliced and fused into 1024 channels;

[0079] S240, integrating a convolution block attention module (CBAM) in each convolution block of the backbone network, to adaptively refine feature expression by concatenating a channel attention sub-module and a spatial attention sub-module;

[0080] Further, step S240 includes:

[0081] S241, in the channel attention sub-module, performing global average pooling and global maximum pooling operations on the feature maps output by the convolution block of the backbone network respectively, to obtain two one-dimensional feature vectors;

[0082] S242, inputting the two one-dimensional feature vectors into a shared multi-layer perceptron, generating a channel attention weight vector after activation function processing, multiplying the weight vector with the feature maps output by the convolution block channel by channel to obtain channel weighted feature maps;

[0083] S243, in the spatial attention sub-module, performing average pooling and maximum pooling operations on the channel weighted feature maps along the channel dimension respectively, to obtain two two-dimensional feature maps, splicing the two two-dimensional feature maps and processing them through a convolution layer and an activation function to generate a spatial attention weight map;

[0084] S244, multiplying the spatial attention weight map with the channel weighted feature maps pixel by pixel to output feature maps refined in both channel and space.

[0085] For example, the multi-layer perceptron of the channel attention sub-module includes two fully connected layers, and the number of neurons is 1 / 16 of the channel number and the original channel number, respectively; the spatial attention sub-module uses a 7*7 convolution kernel to generate a spatial weight map; the two sub-modules are sequentially executed to realize double refinement of the features;

[0086] S300, training the phase unwrapping model using the data set, taking a composite loss function as an optimization objective, iteratively optimizing the phase unwrapping model parameters until the phase unwrapping model converges; the loss function includes a weighted combination of a data fidelity term, a gradient consistency term, a spectral distribution constraint term and a re-wrapping consistency term;

[0087] Further, step S300 includes:

[0088] S310, dividing the data set into a training set and a validation set, inputting the noisy wrapped phase data of the training set into the phase unwrapping model, and the phase unwrapping model outputting the corresponding predicted unwrapped phase;

[0089] S320, taking a composite loss function as the optimization objective of model training, the composite loss function being composed of a data fidelity term, a gradient consistency term, a spectral distribution constraint term and a re-wrapping consistency term by a weighted coefficient combination, and comprehensively constraining the phase unwrapping model output, the formula being as follows:

[0090] L total =w1×L data +w2×L grad +w3×L freq +w4×L wrap ;

[0091] Wherein, L total represents the composite loss function; L data represents the data fidelity term; L grad represents the gradient consistency term; L freq represents the spectral distribution constraint term; L wrap represents the re-wrapping consistency term; w1, w2, w3 and w4 represent the weight coefficients of each term;

[0092] Further, the data fidelity term is calculated based on the difference between the predicted unwrapped phase and the corresponding true phase; the gradient consistency term is calculated based on the spatial gradient difference between the predicted unwrapped phase and the true phase; the spectral distribution constraint term is calculated based on the frequency domain power spectrum difference between the predicted unwrapped phase and the true phase; and the re-wrapping consistency term is calculated based on the consistency of the predicted unwrapped phase and the input noisy wrapped phase in the wrapped gradient domain;

[0093] S330. An optimizer is used to iteratively update the parameters of the phase unwrapping model. After each iteration, the validation set is input into the phase unwrapping model. The predicted unwrapped phase output by the phase unwrapping model is mean aligned. The root mean square error and structural similarity index are used as evaluation indicators to monitor the performance of the phase unwrapping model.

[0094] Furthermore, step S330 includes:

[0095] S331. Mean alignment processing includes: calculating the pixel mean of the predicted unwrapped phase and the corresponding true phase, obtaining the mean difference of the phase, subtracting the mean difference from the predicted unwrapped phase, eliminating irrelevant global constant phase bias, and obtaining the aligned predicted phase.

[0096] S332, Root Mean Square Error includes, based on the aligned predicted phase and the true phase, calculating the square of the pixel-level phase difference between the two, and taking the square root of the mean of all pixel square differences. This index is used to quantify the amplitude recovery accuracy of the untangled phase.

[0097] S333. The structural similarity index includes linearly stretching both the aligned predicted phase and the true phase to the [0,1] interval, and calculating the weighted mean, weighted variance, and weighted covariance of the two phase maps within a local sliding window of a preset size, as shown in the following formula:

[0098] SSIM=[(2μ1μ2+C1)(2σ 12 +C2)] / [(μ1 2 +μ2 2 +C1)(σ1 2 +σ2 2 +C2)];

[0099] Where SSIM represents the structural similarity index; μ1 and μ2 are the weighted averages of the two phase maps, respectively; σ1 2 and σ2 2 Let σ be the weighted variance of the two phase diagrams. 12 It is expressed as the weighted covariance of the two phase diagrams; C1 and C2 are constants to avoid the denominator being zero;

[0100] S340. When the evaluation index of the phase unwrapping model on the validation set remains stable for a continuous preset period and no longer changes significantly, the phase unwrapping model is determined to have converged, and the iteration is stopped.

[0101] S400: Input the InSAR interferogram to be unwrapped into the phase unwrapping model and output the unwrapped phase result.

[0102] This invention provides another technical solution, which is an InSAR phase unwrapping method based on a Fourier domain global hybrid network, for training a phase unwrapping model;

[0103] The dataset is divided into a training set and a validation set in a ratio of 8:2, wherein the training set is used for model parameter learning, and the validation set is used for monitoring the training process and preventing overfitting; the noisy wrapped phase data of the training set is input into the phase unwrapping model, and the phase unwrapping model outputs corresponding predicted unwrapped phases;

[0104] The composite loss function is used as the optimization objective for model training, and the composite loss function is composed of a data fidelity term, a gradient consistency term, a spectral distribution constraint term and a re-wrapping consistency term through weighting coefficient combination, and the composite loss function is used for comprehensive constraint on the output of the phase unwrapping model, and the formula is as follows:

[0105] L total =w1×L data +w2×L grad +w3×L freq +w4×L wrap ;

[0106] Among them, L total represents the composite loss function; L data represents the data fidelity term; L grad represents the gradient consistency term; L freq represents the spectral distribution constraint term; L wrap represents the re-wrapping consistency term; w1, w2, w3 and w4 represent the weight coefficients of each term;

[0107] The data fidelity term is calculated based on the difference between the predicted unwrapped phase and the corresponding true phase; the gradient consistency term is calculated based on the spatial gradient difference between the predicted unwrapped phase and the true phase; the spectral distribution constraint term is calculated based on the frequency domain power spectrum difference between the predicted unwrapped phase and the true phase; and the re-wrapping consistency term is calculated based on the consistency of the predicted unwrapped phase and the input noisy wrapped phase in the wrapped gradient domain;

[0108] The weight coefficients of each loss term are dynamically adjusted according to their convergence in the training process, and the data fidelity term is mainly used in the early stage, and the weight of the physical constraint term is gradually increased in the later stage;

[0109] An optimizer is used to iteratively update the parameters of the phase unwrapping model, and after each iteration, the validation set is input into the phase unwrapping model, and the predicted unwrapped phase output by the phase unwrapping model is processed by mean alignment, and the root mean square error and the structural similarity index are used as evaluation indicators to monitor the performance of the phase unwrapping model;

[0110] For example, using the Adam optimizer, the initial learning rate is set to 0.001, the learning rate is attenuated to 0.5 of the original every 50 epochs, the batch size is set to 8, the total number of training rounds is set to 200, and the performance is evaluated on the validation set every 10 training iterations; this is only an example and is not limited;

[0111] The mean alignment process includes calculating the pixel mean of the predicted unwrapped phase and the corresponding true phase, obtaining the phase mean difference, subtracting this mean difference from the predicted unwrapped phase, eliminating irrelevant global constant phase bias, and obtaining the aligned predicted phase. The root mean square error includes calculating the square of the pixel-level phase difference between the aligned predicted phase and the true phase, taking the square root of the mean of all pixel squared differences, and this index is used to quantify the amplitude recovery accuracy of the unwrapped phase. The structural similarity index includes linearly stretching both the aligned predicted phase and the true phase to the [0,1] interval, and calculating the weighted mean, weighted variance, and weighted covariance of the two phase maps within a local sliding window of a preset size, as shown in the following formulas:

[0112] SSIM=[(2μ1μ2+C1)(2σ 12 +C2)] / [(μ1 2 +μ2 2 +C1)(σ1 2 +σ2 2 +C2)];

[0113] Where SSIM represents the structural similarity index; μ1 and μ2 are the weighted averages of the two phase maps, respectively; σ1 2 and σ2 2 Let σ be the weighted variance of the two phase diagrams. 12 It is expressed as the weighted covariance of the two phase diagrams; C1 and C2 are constants to avoid the denominator being zero;

[0114] When the evaluation metric of the phase unwrapping model on the validation set remains stable for a continuous preset period and no longer changes significantly, the phase unwrapping model is determined to have converged, and the iteration is stopped.

[0115] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. An InSAR phase unwrapping method based on a Fourier domain global hybrid network, characterized in that: The method comprises: S100, generating synthetic data of diversified continuous phases by data simulation, superimposing compound noise of atmospheric noise, Gaussian phase noise and isolated residual point noise, and wrapping the phase to a preset interval to obtain a data pair of noisy wrapped phase and true phase, and forming a data set; S200, constructing a phase unwrapping model, wherein the phase unwrapping model takes a UNet encoder-decoder structure as a backbone network; a Fourier mixed residual block FMRB is embedded in a bottleneck layer of the backbone network; an all-sky pyramid pooling ASPP module is connected after the Fourier mixed residual block FMRB; and a convolution block attention module CBAM is integrated in a convolution block of the backbone network; S300, training the phase unwrapping model by using the data set, taking a compound loss function as an optimization objective, iteratively optimizing phase unwrapping model parameters until the phase unwrapping model converges; and the loss function comprises a weighted combination of a data fidelity term, a gradient consistency term, a spectral distribution constraint term and a re-wrapping consistency term; S400, inputting an InSAR interferogram to be unwrapped into the phase unwrapping model to output an unwrapped phase result.

2. The InSAR phase unwrapping method based on Fourier domain global hybrid network according to claim 1, characterized in that: Step S100 comprises: S110, generating a continuous phase field with diversified spatial distribution characteristics by a bicubic interpolation method and a Gaussian function superposition method; S120, sequentially superimposing simulated atmospheric phase screen noise, random phase noise obeying a Gaussian distribution and isolated residual point noise generated by phase singular points on the continuous phase field to obtain a noisy phase field; S130, performing phase wrapping operation on the noisy phase field to constrain the noisy phase field value range to a preset [-pi, pi) interval to obtain a noisy wrapped phase; S140, constructing the noisy wrapped phase and the corresponding original continuous phase into a data pair to form a data set.

3. The InSAR phase unwrapping method based on Fourier domain global hybrid network according to claim 1, characterized in that: Step S200 comprises: S210, constructing a basic architecture taking a UNet encoder-decoder structure as a backbone network, wherein the encoder gradually extracts multi-scale features through convolution and down-sampling operations, the decoder gradually restores spatial resolution through up-sampling and convolution operations, and the encoder and the decoder realize feature fusion through a skip connection between corresponding levels; S220, embedding a Fourier mixed residual block FMRB in a bottleneck layer of the backbone network, wherein the Fourier mixed residual block FMRB reduces computational complexity through dimension reduction by projection, performs global spatial mixing and global channel mixing operations in the Fourier domain in parallel, establishes a long-distance pixel dependency relationship through residual connection and input fusion, and connects an all-sky pyramid pooling ASPP module at an output end of the Fourier mixed residual block FMRB; S230, connecting an all-sky pyramid pooling ASPP module at an output end of the Fourier mixed residual block FMRB, and capturing multi-scale context information by setting a plurality of dilated convolution layers in parallel with different dilation rates; S240, integrating a convolution block attention module CBAM in each convolution block of the backbone network, and adaptively refining feature expression through a channel attention sub-module and a spatial attention sub-module connected in series.

4. The InSAR phase unwrapping method based on Fourier domain global hybrid network according to claim 3, characterized in that: Step S220 comprises: S221, performing channel dimension compression on input features through a 1x1 convolution layer to obtain compressed features; S222, performing a fast Fourier transform on the compressed feature to obtain a frequency domain feature; performing a first branch processing and a second branch processing in parallel in the frequency domain; the first branch processing performs a weighting processing on a low-frequency component of the frequency domain feature through a learnable complex weight matrix to obtain a spatial mixing feature; the second branch processing performs a cross-channel information fusion on the frequency domain feature through a 1*1 grouped convolution, and applies an adaptive soft threshold operation on the fused frequency domain feature to obtain a channel mixing feature; S223, performing an inverse fast Fourier transform on the spatial mixing feature and the channel mixing feature to obtain a spatial domain feature; performing a channel dimension recovery on the superimposed spatial domain feature through a 1*1 convolution layer to obtain a reconstructed feature; fusing the reconstructed feature and the module input through a residual connection to obtain an output feature with a full image range long distance pixel dependence.

5. The InSAR phase unwrapping method based on Fourier domain global hybrid network according to claim 3, characterized in that: Step S230 includes: S231, in the ASPP module, N parallel dilated convolution layers are configured, N is a positive integer, and the dilation rates of the dilated convolution layers are different from each other; S232, the output feature of the FMRB is respectively convolved by using the dilated convolution layers with different dilation rates to obtain N groups of feature maps with different scales; S233, the N groups of feature maps with different scales are fused to obtain an output feature with aggregated multi-scale context information.

6. The InSAR phase unwrapping method based on Fourier domain global hybrid network according to claim 3, characterized in that: Step S240 includes: S241, in the channel attention submodule, the feature map output by the backbone network convolution block is respectively subjected to a global average pooling and a global maximum pooling operation to obtain two one-dimensional feature vectors; S242, the two one-dimensional feature vectors are input into a shared multilayer perceptron, processed by an activation function to generate a channel attention weight vector, and the channel attention weight vector is fused with the feature map output by the convolution block channel by channel to obtain a channel weighted feature map; S243, in the spatial attention submodule, the channel weighted feature map is respectively subjected to an average pooling and a maximum pooling operation along the channel dimension to obtain two two-dimensional feature maps, and the two two-dimensional feature maps are spliced and processed by a convolution layer and an activation function to generate a spatial attention weight map; S244, the spatial attention weight map and the channel weighted feature map are fused pixel by pixel to output a feature map refined in channel and space.

7. The InSAR phase unwrapping method based on Fourier domain global hybrid network according to claim 1, characterized in that: Step S300 includes: S310, dividing the data set into a training set and a validation set, inputting the noisy wrapped phase data of the training set into a phase unwrapping model, and the phase unwrapping model outputs corresponding predicted unwrapped phase; S320, taking a composite loss function as an optimization target for training the phase unwrapping model, the composite loss function is composed of a data fidelity term, a gradient consistency term, a spectral distribution constraint term and a re-winding consistency term through a weighting coefficient, and the composite loss function is used to comprehensively constrain the output of the phase unwrapping model, and the formula is as follows: L total = w1 x L data + w2 x L grad + w3 x L freq + w4 x L wrap ; wherein L total represents a composite loss function; L data represents a data fidelity term; L grad represents a gradient consistency term; L freq represents a spectral distribution constraint term; L wrap represents a re-wrapping consistency term; w1, w2, w3, and w4 represent weight coefficients for each term; S330, the parameters of the phase unwrapping model are iteratively updated by using an optimizer, after each iteration, the verification set is input into the phase unwrapping model, the predicted unwrapped phase output by the phase unwrapping model is processed by mean alignment, the root mean square error and the structural similarity index are used as evaluation indexes to monitor the performance of the phase unwrapping model; S340, when the evaluation indexes of the phase unwrapping model on the verification set remain stable for a preset period of time and no longer change significantly, it is determined that the phase unwrapping model converges, and the iteration is stopped.

8. The InSAR phase unwrapping method based on the Fourier domain global hybrid network according to claim 7, characterized in that: The data fidelity term is calculated based on the difference between the predicted unwrapped phase and the corresponding real phase; the gradient consistency term is calculated based on the spatial gradient difference between the predicted unwrapped phase and the real phase; the spectral distribution constraint term is calculated based on the frequency domain power spectrum difference between the predicted unwrapped phase and the real phase; and the re-wrapped consistency term is calculated based on the consistency of the predicted unwrapped phase and the input noisy wrapped phase in the wrapped gradient domain.

9. The InSAR phase unwrapping method based on Fourier domain global hybrid network according to claim 7, characterized in that: Step S330 includes: S331, the mean alignment processing includes calculating the pixel mean of the predicted unwrapped phase and the corresponding real phase to obtain a phase mean difference value, subtracting the mean difference value from the predicted unwrapped phase to eliminate irrelevant overall constant phase bias, and obtaining the aligned predicted phase; S332, the root mean square error includes calculating the square of the pixel-level phase difference between the aligned predicted phase and the real phase based on the aligned predicted phase and the real phase, taking the square root of the mean of all pixel square difference values, and this index is used to quantify the amplitude recovery accuracy of the unwrapped phase; S333, the structural similarity index includes linearly stretching the aligned predicted phase and the real phase to the [0, 1] interval, calculating the weighted mean, weighted variance and weighted covariance of the two phase maps in a local sliding window of a preset size, and the formula is as follows: SSIM = [(2μ1μ2 + C1)(2σ 12 + C2)] / [(μ1 2 + μ2 2 + C1)(σ1 2 + σ2 2 + C2)] ; where SSIM is expressed as the structural similarity index; μ1 and μ2 are the weighted mean of the two phase maps, respectively; σ1 2 and σ2 2 are the weighted variance of the two phase maps, respectively, and σ 12 is the weighted covariance of the two phase maps; C1 and C2 are constants to avoid division by zero.

Images