InSAR phase unwrapping method, device and storage medium based on deep convolutional neural network optimization

Abstract

The present disclosure relates to an InSAR phase unwrapping method, device and storage medium based on deep convolutional neural network optimization, the method comprising constructing a training data set, including randomly generated images and randomly generated Gaussian images; training a ResNet network based on the obtained training data set; correcting the results predicted by the ResNet network through an optimizer and a loss function; and inputting the InSAR measured data interference pattern into a tested neural network model to obtain the unwrapped real phase. Through the various embodiments of the present disclosure, higher accuracy and robustness, and better noise resistance are achieved in phase unwrapping.

Description

Technical Field

[0001] The present disclosure relates to the field of synthetic aperture radar interferometry technology, and in particular to an InSAR phase unwrapping method based on deep convolutional neural network optimization, an InSAR phase unwrapping device based on deep convolutional neural network optimization, and a computer-readable storage medium. Background Art

[0002] InSAR (Interferometric Radar) measurement technology is a remote sensing technology that uses synthetic aperture radar (SAR) data to measure surface deformation and elevation. It obtains information about surface targets by comparing the phase difference between two or more radar images. The problem of estimating the "true phase" of the ground from the "wrapped phase" of the interference pattern is called "phase unwrapping". Existing phase unwrapping methods have a high dependence on the phase continuity assumption, poor noise resistance, and low robustness. Summary of the invention

[0003] The present disclosure intends to provide an InSAR phase unwrapping method based on deep convolutional neural network optimization, an InSAR phase unwrapping device based on deep convolutional neural network optimization, and a computer-readable storage medium, which have higher accuracy and robustness in phase unwrapping and better anti-noise performance.

[0004] According to one of the solutions of the present disclosure, an InSAR phase unwrapping method based on deep convolutional neural network optimization is provided, comprising:

[0005] Construct a training data set, including randomly generated images and randomly generated Gaussian images;

[0006] Train the ResNet network based on the obtained training data set;

[0007] Correct the results predicted by the ResNet network through the optimizer and loss function;

[0008] The interferogram of InSAR measured data is input into the tested neural network model to obtain the untangled true phase.

[0009] In some embodiments, constructing a training data set includes:

[0010] By randomly generating matrices, pixel blocks of different sizes and intensities are obtained through scaling and intensity transformation; or by generating a certain number of random Gaussian distributions, randomly selecting channel positions and combining them into an image;

[0011] Generate a dataset consisting of these images.

[0012] In some embodiments, wherein randomly generating an image comprises:

[0013] Generate a random 2D array of size base;

[0014] Scale it to an array of size final dimensions;

[0015] Multiply by a random coefficient and set its minimum value to zero;

[0016] Randomly generate Gaussian images, including:

[0017] Generate a two-dimensional array of all zeros;

[0018] Create a one-dimensional array with a preset range and a preset step size;

[0019] Generate a grid matrix to represent the horizontal and vertical coordinates of the image, use the coordinate grid to calculate the value of the Gaussian distribution and accumulate it to a two-dimensional all-zero array;

[0020] Repeat the loop an appropriate number of times, generating a Gaussian distribution each time, and then adding multiple Gaussian distributions to the array, each of which consists of a random variance, a random translation, and a random amplitude, until the minimum value of the array is reset to zero;

[0021] Generate a dataset consisting of these images, including:

[0022] Random array-based images and images consisting of multiple Gaussian distributions are generated by cyclic interleaving to construct the training dataset.

[0023] In some embodiments, the ResNet network is trained based on the obtained training data set, including:

[0024] The input features are passed through a series of convolutional layers, batch normalization layers, and activation functions to obtain the residual function;

[0025] Add the residual function input features element by element to get the desired output.

[0026] In some embodiments, the configuration of the optimizer includes:

[0027] Introduce momentum correction term based on Adam optimizer;

[0028] The optimizer is obtained by updating the equation after momentum correction.

[0029] In some embodiments, testing the neural network model includes:

[0030] Input the generated test set into the trained neural network model to obtain the untangled true phase;

[0031] Set the range of the linear mapping of the wrapped phase image;

[0032] Different levels of Gaussian noise, salt-pepper noise, and multiplicative noise are added to the wrapped phase image, and linearly mapped back to the same range as the input;

[0033] The noisy phase image is unwrapped using the ResNet network and an adaptive moment estimation model that introduces a momentum correction term.

[0034] In some embodiments, the InSAR measured data interferogram is input into a tested neural network model to obtain the untangled real phase, including:

[0035] The ResNet network and the adaptive moment estimation model with the introduction of momentum correction term are used to unwrap the InSAR measured data interferogram to obtain the unwrapping result of the interferogram.

[0036] In some embodiments, the ResNet network is configured as at least a ResNET-18 network, and a corresponding ResNet model is obtained based on different numbers of channels and residual blocks in the module.

[0037] According to one of the solutions disclosed herein, an InSAR phase unwrapping device based on deep convolutional neural network optimization is provided, comprising:

[0038] A construction module configured to construct a training dataset;

[0039] A training module configured to train the ResNet network based on the obtained training data set;

[0040] A correction module, which is configured to correct the results predicted by the ResNet network through an optimizer and a loss function;

[0041] The phase unwrapping module is configured to input the InSAR measured data interference pattern into the tested neural network model to obtain the unwrapped real phase.

[0042] According to one of the solutions of the present disclosure, a computer-readable storage medium is provided, on which computer-executable instructions are stored, and when the computer-executable instructions are executed by a processor, the following are implemented:

[0043] According to the above-mentioned InSAR phase unwrapping method based on deep convolutional neural network optimization.

[0044] The InSAR phase unwrapping method based on deep convolutional neural network optimization, the InSAR phase unwrapping device based on deep convolutional neural network optimization and the computer-readable storage medium of various embodiments of the present disclosure at least construct a training data set, including randomly generated images and randomly generated Gaussian images; train the ResNet network based on the obtained training data set; correct the results predicted by the ResNet network through the optimizer and the loss function; input the InSAR measured data interference map into the tested neural network model to obtain the unwrapped real phase, aiming to obtain efficient and accurate unwrapping results, restore the correct phase, and thus obtain accurate surface deformation inversion results, and propose an InSAR phase unwrapping method based on deep convolutional neural network optimization, which replaces the traditional phase unwrapping method, realizes accurate prediction of the real phase of the input data, and then uses the improved adaptive moment estimation algorithm and loss function to correct the neural network calculation results, and finally obtains accurate real phase data. The method proposed in each embodiment of the present disclosure has higher accuracy and robustness than the traditional algorithm, and better anti-noise performance.

[0045] It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosure, as claimed. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In the drawings, which are not necessarily drawn to scale, similar reference numerals in different views may represent similar components. Similar reference numerals with letter suffixes or similar reference numerals with different letter suffixes may represent different instances of similar components. The drawings generally illustrate various embodiments by way of example and not limitation, and together with the description and claims, serve to explain the disclosed embodiments.

[0047] Figure 1 A flow chart of an InSAR phase unwrapping method based on deep convolutional neural network optimization according to an embodiment of the present disclosure is shown;

[0048] Figure 2 A schematic diagram of an implementation step of an InSAR phase unwrapping method based on deep convolutional neural network optimization according to an embodiment of the present disclosure is shown;

[0049] Figure 3 A schematic diagram showing two types of residual blocks in a ResNet network according to an embodiment of the present disclosure is shown;

[0050] Figure 4 A schematic diagram showing the ResNET-18 network structure of an embodiment of the present disclosure is shown;

[0051] Figure 5 The phase unwrapping result under ideal conditions is shown;

[0052] Figure 6 The phase unwrapping result obtained by using the embodiment of the present disclosure after adding noise is shown;

[0053] Figure 7 A phase unwrapping result diagram obtained by using the measured data interference diagram of an embodiment of the present disclosure is shown;

[0054] Figure 8 An architectural diagram of an InSAR phase unwrapping device based on deep convolutional neural network optimization according to an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solution and advantages of the embodiments of the present disclosure clearer, the technical solution of the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings of the embodiments of the present disclosure. Obviously, the described embodiments are part of the embodiments of the present disclosure, not all of the embodiments. Based on the described embodiments of the present disclosure, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present disclosure.

[0056] In combination with the above-mentioned background technology, the present disclosure records the corresponding solutions in the form of embodiments to solve the defects in the prior art, but it is not intended to limit the scope of protection of the patent right claimed by the present disclosure.

[0057] InSAR measurement technology is a remote sensing technology that uses synthetic aperture radar (SAR) data to measure surface deformation and elevation. It obtains information about surface targets by comparing the phase difference between two or more radar images, and has a great advantage among many ground observation methods. InSAR data processing flow, the step of forming interference is an indispensable part of InSAR data processing and is also the most important part related to the accuracy of the entire experimental results. Since the phase map can only be measured as its value modulo 2π, that is, the phase imaging technology can only measure the remainder of the real phase minus several integer multiples of 2π, so the purpose of phase unwrapping is to find the integer multiples of 2π lost in the interference map and restore the physical continuity of the phase mapping. Therefore, improving the phase unwrapping method of the phase interferogram can help obtain more accurate deformation results.

[0058] At present, the InSAR two-dimensional phase unwrapping algorithms can be roughly divided into the following categories: the first category is based on path tracking; the second category is based on the minimum norm; the third category is based on the optimal estimation. Representative algorithms include Goldstein branch cutting method, mask cutting algorithm (Mask-cut), quality-guided algorithm, least squares method, and minimum cost flow (MCF) algorithm.

[0059] The above InSAR 2D phase unwrapping method has some technical defects, including the following:

[0060] 1. Non-uniqueness problem: There is non-uniqueness in the phase unwrapping process, that is, there are multiple possible solutions. This is because in the interference pattern, the phase value range is wrapped in an interval, and the true phase offset cannot be determined during the unwrapping process. This non-uniqueness makes the result of phase unwrapping possibly not unique, and needs to be further processed in combination with other information or constraints;

[0061] 2. Noise and interference: There are often noise and interference in the interference pattern, such as interference caused by light scattering, errors in the interferometer, environmental vibration, etc. These noise and interference will affect the precision and accuracy of phase unwrapping, and may lead to error accumulation and erroneous unwrapping results during the unwrapping process;

[0062] 3. Regional limitation: Phase unwrapping is usually performed based on local regions, that is, the interference pattern is divided into small blocks for unwrapping. This local limitation may lead to boundary effects and discontinuities, especially in the presence of large phase changes or complex phase structures.

[0063] 4. Computational complexity: Phase unwrapping usually involves a lot of calculations and algorithms, especially when dealing with large-scale interferograms or performing complex phase analysis. This requires efficient algorithms and computing resources to process large-scale data, and may require long computing times.

[0064] From the above, we can see that phase unwrapping technology still has some challenges and limitations in practical applications. It is necessary to comprehensively consider factors such as non-uniqueness, noise interference, regional restrictions and computational complexity to obtain accurate and reliable unwrapping results. In practical applications, it may be necessary to combine other technologies and methods, such as multi-sensor fusion, prior knowledge constraints and model optimization, to improve the accuracy and reliability of phase unwrapping.

[0065] In order to obtain efficient and accurate unwrapping results, restore the correct phase, and thus obtain accurate surface deformation inversion results, as one of the solutions, such as Figure 1 As shown, an embodiment of the present disclosure provides an InSAR phase unwrapping method based on deep convolutional neural network optimization, comprising:

[0066] S100: construct a training data set, including randomly generated images and randomly generated Gaussian images;

[0067] S200: training the ResNet network based on the obtained training data set;

[0068] S300: Correct the predicted results of the ResNet network through the optimizer and loss function;

[0069] S400: Input the InSAR measured data interference pattern into the tested neural network model to obtain the untangled real phase.

[0070] In response to the problems raised in the previous article, the embodiments of the present invention aim to propose an InSAR phase unwrapping method based on deep convolutional neural network optimization, which replaces the traditional phase unwrapping method, realizes the accurate prediction of the true phase of the input data, and then uses the optimizer and loss function to correct the neural network calculation results, and finally obtains accurate true phase data, which has higher accuracy and robustness and better noise resistance than traditional algorithms.

[0071] In some embodiments, the method of the embodiment of the present disclosure, constructing a training data set, includes:

[0072] By randomly generating matrices, pixel blocks of different sizes and intensities are obtained through scaling and intensity transformation; or by generating a certain number of random Gaussian distributions, randomly selecting channel positions and combining them into an image;

[0073] Generate a dataset consisting of these images.

[0074] like Figure 1 and Figure 2 As shown, in this embodiment, a data set containing multiple elements can be generated. There are two types of elements, one is a randomly generated small image, and the other is a randomly generated Gaussian image. By randomly generating a matrix, scaling and intensity transformation are performed to obtain pixel blocks of different sizes and intensities; or by generating a certain number of random Gaussian distributions, randomly selecting channel positions and combining them into an image. Finally, a data set consisting of these two images is generated.

[0075] In some embodiments, the method of the disclosed embodiment may be: randomly generating an image, including: generating a random two-dimensional array of a base size; scaling it to an array of a final size; multiplying it by a random coefficient and returning its minimum value to zero;

[0076] Randomly generate a Gaussian image, including: generating a two-dimensional array of all zeros; creating a one-dimensional array with a preset range and a preset step size; generating a grid matrix for representing the horizontal and vertical coordinates of the image, using the coordinate grid to calculate the value of the Gaussian distribution and adding it to the two-dimensional array of all zeros; looping an appropriate number of times, each time generating a Gaussian distribution, and then adding multiple Gaussian distributions to the array, each Gaussian distribution consisting of a random variance, a random translation, and a random amplitude, until the minimum value of the array is reset to zero;

[0077] Generating a dataset consisting of these images includes generating random array-based images and images consisting of multiple Gaussian distributions by cyclic crossover to construct a training dataset.

[0078] For example, a specific method of constructing a training data set can be implemented as a technical solution represented by the following steps.

[0079] Generate a random small image: first generate a random 2D array of the base size, then scale it to an array of the final size, then multiply it by a random coefficient and set its minimum value to zero.

[0080] Generate a random Gaussian image: First, generate a 256×256 two-dimensional array of all zeros. Then create a one-dimensional array x and y with a range of -π to π and a step size of 0.0246. Then generate grid matrices xx and yy to represent the horizontal and vertical coordinates of the image, use the coordinate grid to calculate the value of the Gaussian distribution and add it to the two-dimensional array of all zeros. Then loop an appropriate number of times, generate a Gaussian distribution each time, and then add multiple Gaussian distributions to this array. Each Gaussian distribution consists of a random variance, a random translation, and a random amplitude. Finally, the minimum value of the array is reset to zero.

[0081] Taking a data set size of n and a loop variable of i as an example, this embodiment can generate a data set of size n through a loop. When i is an even number, a random array-based image is generated. When i is an odd number, an image composed of multiple Gaussian distributions is generated.

[0082] Through the above implementation steps, this embodiment can generate real phase images with different characteristics and changes for model training and testing. For example, 5,000 real phase images are generated and these images are intertwined to serve as training sets and test sets. Figure 6 As shown in the figure, for example, to be closer to the interference phase map measured in actual conditions, random noise can be added to the image in the data set.

[0083] In some embodiments, the method of the embodiment of the present disclosure may be: training the ResNet network based on the obtained training data set, including:

[0084] The input features are passed through a series of convolutional layers, batch normalization layers, and activation functions to obtain the residual function;

[0085] Add the residual function input features element by element to get the desired output.

[0086] Continue to combine Figure 3 and Figure 4As exemplified, the present embodiment trains the ResNet network based on the obtained training data set. The ResNet network can be trained using the training data set obtained in the aforementioned steps, and is implemented as a technical solution represented by the following steps.

[0087] Assuming that in a certain layer, the original input is x, and the optimal function is recorded as H(x), then the fitted objective function F(x) is defined as:

[0088] F(x)=H(x)-x

[0089] The function F(x) is called the "residual function" and its function is to learn the residual information in the input feature x, that is, the difference between the learning input feature and the expected output H(x).

[0090] The residual block can contain multiple convolutional layers, batch normalization layers and activation functions. The specific process is: first, the input feature x is calculated through a series of convolutional layers, batch normalization layers and activation functions to obtain F(x), and then F(x) is added to the input feature x element by element to obtain the expected output H(x).

[0091] In some embodiments, the embodiments of the present disclosure may be: the ResNet network is at least configured as a ResNET-18 network, and a corresponding ResNet model is obtained based on different numbers of channels and numbers of residual blocks in the module.

[0092] Exemplarily, the present embodiment can use the ResNET-18 network as a network model for training. The first two layers of ResNet-18 are a convolutional layer with an output channel number of 64 and a stride of 2, followed by a maximum pooling layer with a stride of 2, and a batch normalization layer is added after each convolutional layer. In the meantime, 4 modules composed of residual blocks are used, and each module uses several residual blocks with the same number of output channels. The number of channels of the first module is the same as the number of input channels. Since the maximum pooling layer with a stride of 2 has been used before, there is no need to reduce the height and width. Each subsequent module doubles the number of channels of the previous module in the first residual block and halves the height and width. Then all residual blocks are added to ResNet-18, where each module uses 2 residual blocks. Finally, a global average pooling layer and a fully connected layer output are added to ResNet-18.

[0093] Each module has 4 convolutional layers, excluding the 1×1 convolutional layer for identity mapping, plus the first 7×7 convolutional layer and the last fully connected layer, for a total of 18 layers. Different ResNet models can be obtained by configuring different numbers of channels and residual blocks in the module.

[0094] In some implementation schemes, the method of the embodiment of the present disclosure may be: a configuration method of an optimizer, including:

[0095] Introduce momentum correction term based on Adam optimizer;

[0096] The optimizer is obtained by updating the equation after momentum correction.

[0097] For example, the specific configuration of the optimizer can be implemented as a technical solution represented by the following steps.

[0098] The optimizer can optimize the parameters of the network model by updating the parameters. The Adam optimization algorithm is an effective stochastic optimization method that calculates the individual adaptive learning rates of different parameters by estimating the first and second order moments of the gradient. The name Adam comes from the adaptive moment estimation. Given a learning rate of α and a momentum parameter of β 1 and β 2 , the model parameter is θ, and the gradient of the loss function is g t .

[0099] m t =β 1 m t-1 +(1-β 1 ) t

[0100]

[0101] m t represents the first-order moment estimator, V t represents the second-order moment estimator, t represents the number of iterations, and the initial values ​​of all three are 0. In each iteration step, the gradient g is calculated t And update the first-order moment analyzer and the second-order moment estimator, and correct the deviation of the first-order and second-order moment estimates:

[0102]

[0103]

[0104] Update model parameters:

[0105]

[0106] where ε is a constant used for numerical stability. To calculate the bias-corrected first-order moment estimator, Calculates the bias-corrected second moment estimator.

[0107] First-order moment estimator m t Used to estimate the first-order moment (mean) of the gradient, the second-order moment estimator V tUsed to estimate the second moment (variance) of the gradient. By correcting the bias of these estimators, the first and second moments of the gradient can be estimated more accurately and used to update the model parameters.

[0108] The Nadam algorithm introduces the Nesterov momentum correction term on the basis of the Adam algorithm, and obtains the Nadam update equation:

[0109]

[0110] m t =β 1 m t-1 +(1-β 1 ) t

[0111]

[0112] The introduction of the Nesterov momentum correction term can help the algorithm better handle the local minima of the parameters and provide faster convergence, better generalization performance and a more stable learning process, which is more advantageous than the Adam optimizer.

[0113] The loss function used in this embodiment is L1 loss. L1 loss is also called absolute value loss. It is a regression loss function used to measure the difference between the predicted value and the target value. It calculates the absolute value of the difference between the predicted value and the target value and takes the average value as the loss value.

[0114] Specifically, for a given sample, assume that the predicted value is y p , the target value is y t , the calculation formula of L1 loss function is as follows:

[0115]

[0116] Where N is the number of samples, Σ represents the summation symbol, and || represents the absolute value operation.

[0117] During the model training process, the model parameters are optimized by minimizing the L1 loss function so that the difference between the predicted value and the target value is as small as possible. This can encourage the model to learn prediction results that are closer to the target value.

[0118] In some embodiments, the disclosed embodiments may be: testing a neural network model, including:

[0119] Input the generated test set into the trained neural network model to obtain the untangled true phase;

[0120] Set the range of the linear mapping of the wrapped phase image;

[0121] Different levels of Gaussian noise, salt-pepper noise, and multiplicative noise are added to the wrapped phase image, and linearly mapped back to the same range as the input;

[0122] The noisy phase image is unwrapped using the ResNet network and an adaptive moment estimation model that introduces a momentum correction term.

[0123] For example, the specific testing process can be implemented as a technical solution represented by the following steps.

[0124] Test the ResNet network and the Nesterov accelerated adaptive moment estimation model, and input the test set generated in the above steps into the trained neural network model to obtain the unwrapped real phase. In order to test the anti-noise performance of the ResNet network and the Nesterov accelerated adaptive moment estimation model, the range of the entangled phase image is linearly mapped to [0,1], and then different levels of Gaussian noise, salt and pepper noise, and multiplicative noise are added to the entangled phase image, and then linearly mapped back to the same range as the input, and the noisy phase image is unwrapped using the ResNet network and the Nesterov accelerated adaptive moment estimation model. In some embodiments, the standard deviation of Gaussian noise and multiplicative noise can be 0.01 to 0.40, respectively, so that the density of salt and pepper noise ranges from 0.01 to 0.40.

[0125] In some embodiments, the disclosed embodiment may be: inputting the InSAR measured data interferogram into a tested neural network model to obtain the untangled real phase, including:

[0126] The ResNet network and the adaptive moment estimation model with the introduction of momentum correction term are used to unwrap the InSAR measured data interferogram to obtain the unwrapping result of the interferogram.

[0127] For example, see Figures 5 to 7 , Figure 5 The phase unwrapping result under ideal conditions is shown. Part (a) shows the wrapped phase, part (b) shows the real phase, and part (c) shows the three-dimensional display of the real phase. Figure 6 The figure shows the phase unwrapping results obtained by the model using the ResNet network and Nesterov accelerated adaptive moment estimation after adding noise. Part (a) of the figure shows the unwrapping result of the ResNET model, and part (b) is a three-dimensional display of the unwrapping result of the ResNET model. Figure 7The phase unwrapping result diagram obtained by processing the InSAR measured data interferogram using the ResNet network and the Nesterov accelerated adaptive moment estimation model is shown. In the figure, parts (a), (d), and (g) are three InSAR interferograms respectively; parts (b), (e), and (h) are the unwrapping results obtained by the ResNet network and the Nesterov accelerated adaptive moment estimation model after unwrapping the three InSAR interferograms; parts (c), (f), and (i) are three-dimensional displays of the unwrapping results obtained by the ResNet network and the Nesterov accelerated adaptive moment estimation model after unwrapping the three InSAR interferograms. Combined with the previous content, this embodiment uses the ResNet network and the Nesterov accelerated adaptive moment estimation model to unwrap the InSAR measured data interferogram to obtain the unwrapping result of the interferogram.

[0128] Compared with the current phase unwrapping method, the InSAR phase unwrapping method based on deep convolutional neural network optimization in each embodiment of the present disclosure has the following beneficial effects:

[0129] 1) The model using ResNet network and Nesterov accelerated adaptive moment estimation is not restricted by the phase continuity assumption, which improves the accuracy of phase unwrapping;

[0130] 2) Nesterov accelerated adaptive moment estimation and loss function can effectively solve the problem of multiple possible solutions in the phase unwrapping process;

[0131] 3) In the presence of noise, the model using ResNet network and Nesterov accelerated adaptive moment estimation has better noise resistance performance.

[0132] As one of the solutions, Figure 8 As shown, an embodiment of the present disclosure provides an InSAR phase unwrapping device based on deep convolutional neural network optimization, comprising:

[0133] A construction module configured to construct a training dataset;

[0134] A training module configured to train the ResNet network based on the obtained training data set;

[0135] A correction module, which is configured to correct the results predicted by the ResNet network through an optimizer and a loss function;

[0136] The phase unwrapping module is configured to input the InSAR measured data interference pattern into the tested neural network model to obtain the unwrapped real phase.

[0137] As an implementation mode, the InSAR phase unwrapping device based on deep convolutional neural network optimization disclosed in the present invention can be combined with the steps described above, and the building module is further configured as follows:

[0138] By randomly generating matrices, pixel blocks of different sizes and intensities are obtained through scaling and intensity transformation; or by generating a certain number of random Gaussian distributions, randomly selecting channel positions and combining them into an image;

[0139] Generate a dataset consisting of these images.

[0140] As an implementation mode, the InSAR phase unwrapping device based on deep convolutional neural network optimization disclosed in the present invention can be combined with the steps described above, and the building module is further configured as follows:

[0141] Randomly generate images, including:

[0142] Generate a random 2D array of size base;

[0143] Scale it to an array of size final dimensions;

[0144] Multiply by a random coefficient and set its minimum value to zero;

[0145] Randomly generate Gaussian images, including:

[0146] Generate a two-dimensional array of all zeros;

[0147] Create a one-dimensional array with a preset range and a preset step size;

[0148] Generate a grid matrix to represent the horizontal and vertical coordinates of the image, use the coordinate grid to calculate the value of the Gaussian distribution and accumulate it to a two-dimensional all-zero array;

[0149] Repeat the loop an appropriate number of times, generating a Gaussian distribution each time, and then adding multiple Gaussian distributions to the array, each of which consists of a random variance, a random translation, and a random amplitude, until the minimum value of the array is reset to zero;

[0150] Generate a dataset consisting of these images, including:

[0151] Random array-based images and images consisting of multiple Gaussian distributions are generated by cyclic interleaving to construct the training dataset.

[0152] As an implementation mode, the InSAR phase unwrapping device based on deep convolutional neural network optimization disclosed in the present invention can be combined with the steps described above, and the training module is further configured as follows:

[0153] The input features are passed through a series of convolutional layers, batch normalization layers, and activation functions to obtain the residual function;

[0154] Add the residual function input features element by element to get the desired output.

[0155] As an implementation mode, the InSAR phase unwrapping device based on deep convolutional neural network optimization disclosed in the present invention can be combined with the steps described above, and the correction module is further configured as follows:

[0156] Introduce momentum correction term based on Adam optimizer;

[0157] The optimizer is obtained by updating the equation after momentum correction.

[0158] For example, the optimizer can optimize the parameters of the network model by updating the parameters. The Adam optimization algorithm is an effective stochastic optimization method that calculates the individual adaptive learning rates of different parameters by estimating the first and second order moments of the gradient. The name Adam comes from the adaptive moment estimation. Given a learning rate of α and a momentum parameter of β 1 and β 2 , the model parameter is θ, and the gradient of the loss function is g t .

[0159] m t =β 1 m t-1 +(1-β 1 ) t

[0160]

[0161] m t represents the first-order moment estimator, V t represents the second-order moment estimator, t represents the number of iterations, and the initial values ​​of all three are 0. In each iteration step, the gradient g is calculated t And update the first-order moment analyzer and the second-order moment estimator, and correct the deviation of the first-order and second-order moment estimates:

[0162]

[0163]

[0164] Update model parameters:

[0165]

[0166] where ε is a constant used for numerical stability. To calculate the bias-corrected first-order moment estimator, Calculates the bias-corrected second moment estimator.

[0167] First-order moment estimator mt Used to estimate the first-order moment (mean) of the gradient, the second-order moment estimator V t Used to estimate the second moment (variance) of the gradient. By correcting the deviation of these estimators, the first and second moments of the gradient can be estimated more accurately and used to update the model parameters. The Nadam algorithm introduces the Nesterov momentum correction term on the basis of the Adam algorithm, and obtains the Nadam update equation:

[0168]

[0169] m t =β 1 m t-1 +(1-β 1 ) t

[0170]

[0171] The introduction of the Nesterov momentum correction term can help the algorithm better handle the local minima of the parameters and provide faster convergence, better generalization performance and a more stable learning process, which is more advantageous than the Adam optimizer.

[0172] As an implementation mode, the InSAR phase unwrapping device based on deep convolutional neural network optimization disclosed in the present invention can be further configured as follows in combination with the steps described above:

[0173] Input the generated test set into the trained neural network model to obtain the untangled true phase;

[0174] Set the range of the linear mapping of the wrapped phase image;

[0175] Different levels of Gaussian noise, salt-pepper noise, and multiplicative noise are added to the wrapped phase image, and linearly mapped back to the same range as the input;

[0176] The noisy phase image is unwrapped using the ResNet network and an adaptive moment estimation model that introduces a momentum correction term.

[0177] Exemplarily, the ResNet network and the Nesterov accelerated adaptive moment estimation model are tested, and the test set generated by the above steps is input into the trained neural network model to obtain the unwrapped real phase. In order to test the anti-noise performance of the ResNet network and the Nesterov accelerated adaptive moment estimation model, the range of the entangled phase image is linearly mapped to [0,1], and then different levels of Gaussian noise, salt and pepper noise, and multiplicative noise are added to the entangled phase image, and then linearly mapped back to the same range as the input, and the noisy phase image is unwrapped using the ResNet network and the Nesterov accelerated adaptive moment estimation model. In some embodiments, the standard deviation of Gaussian noise and multiplicative noise can be 0.01 to 0.40, respectively, so that the density of salt and pepper noise is from 0.01 to 0.40.

[0178] As an implementation mode, the InSAR phase unwrapping device based on deep convolutional neural network optimization disclosed in the present invention can be combined with the steps described above, and the phase unwrapping module is further configured as follows:

[0179] The ResNet network and the adaptive moment estimation model with the introduction of momentum correction term are used to unwrap the InSAR measured data interferogram to obtain the unwrapping result of the interferogram.

[0180] As an implementation mode, the InSAR phase unwrapping device based on deep convolutional neural network optimization disclosed in the present invention can be combined with the steps described above, the ResNet network is configured as at least a ResNET-18 network, and the corresponding ResNet model is obtained based on different numbers of channels and residual blocks in the module.

[0181] Specifically, one of the inventive concepts of the present disclosure is to construct at least a training data set, including randomly generated images and randomly generated Gaussian images, through the above-mentioned InSAR phase unwrapping method based on deep convolutional neural network optimization, the InSAR phase unwrapping device based on deep convolutional neural network optimization and the computer-readable storage medium; train the ResNet network based on the obtained training data set; correct the results predicted by the ResNet network through the optimizer and the loss function; input the InSAR measured data interference map into the tested neural network model to obtain the unwrapped real phase, aiming to obtain efficient and accurate unwrapping results, restore the correct phase, and thus obtain accurate surface deformation inversion results, and propose an InSAR phase unwrapping method based on deep convolutional neural network optimization, which replaces the traditional phase unwrapping method, realizes accurate prediction of the real phase of the input data, and then uses the improved adaptive moment estimation algorithm and loss function to correct the neural network calculation results, and finally obtains accurate real phase data. The method proposed in each embodiment of the present disclosure has higher accuracy and robustness than the traditional algorithm, and better noise resistance.

[0182] The present disclosure also provides a computer-readable storage medium having computer executable instructions stored thereon. When the computer executable instructions are executed by a processor, the InSAR phase unwrapping method based on deep convolutional neural network optimization is mainly implemented, including:

[0183] Construct a training data set, including randomly generated images and randomly generated Gaussian images;

[0184] Train the ResNet network based on the obtained training data set;

[0185] Correct the results predicted by the ResNet network through the optimizer and loss function;

[0186] The interferogram of InSAR measured data is input into the tested neural network model to obtain the untangled true phase.

[0187] The above embodiments are only exemplary embodiments of the present disclosure and are not intended to limit the present disclosure. The protection scope of the present disclosure is defined by the claims. Those skilled in the art may make various modifications or equivalent substitutions to the present disclosure within the essence and protection scope of the present disclosure, and such modifications or equivalent substitutions shall also be deemed to fall within the protection scope of the present disclosure.

Claims

1. InSAR phase unwrapping method based on deep convolutional neural network optimization, including: Construct a training data set, including randomly generated images and randomly generated Gaussian images; Train the ResNet network based on the obtained training data set; Correct the results predicted by the ResNet network through the optimizer and loss function; The interferogram of InSAR measured data is input into the tested neural network model to obtain the untangled true phase; in, Construct a training dataset, including: By randomly generating matrices, pixel blocks of different sizes and intensities are obtained through scaling and intensity transformation; by generating a certain number of random Gaussian distributions, channel positions are randomly selected and combined into an image; Generate a dataset consisting of these images; Randomly generate images, including: Generate a random 2D array of size base; Scale it to an array of size final dimensions; Multiply by a random coefficient and set its minimum value to zero; Randomly generate Gaussian images, including: Generate a two-dimensional array of all zeros; Create a one-dimensional array with a preset range and a preset step size; Generate a grid matrix to represent the horizontal and vertical coordinates of the image, use the coordinate grid to calculate the value of the Gaussian distribution and accumulate it to a two-dimensional all-zero array; Repeat the loop an appropriate number of times, generating a Gaussian distribution each time, and then adding multiple Gaussian distributions to the array, each of which consists of a random variance, a random translation, and a random amplitude, until the minimum value of the array is reset to zero; Generate a dataset consisting of these images, including: Random array-based images and images consisting of multiple Gaussian distributions are generated by cyclic interleaving to construct the training dataset.

2. The method according to claim 1, wherein: Based on the obtained training data set, the ResNet network is trained, including: The input features are passed through a series of convolutional layers, batch normalization layers, and activation functions to obtain the residual function; Add the residual function input features element by element to get the desired output.

3. The method according to claim 2, wherein: The optimizer configuration includes: Introduce momentum correction term based on Adam optimizer; The optimizer is obtained by updating the equation after momentum correction.

4. The method according to claim 3, wherein: Testing neural network models, including: Input the generated test set into the trained neural network model to obtain the untangled true phase; Set the range of the linear mapping of the wrapped phase image; Different levels of Gaussian noise, salt-pepper noise, and multiplicative noise are added to the wrapped phase image, and linearly mapped back to the same range as the input; The noisy phase image is unwrapped using the ResNet network and an adaptive moment estimation model that introduces a momentum correction term.

5. The method according to claim 4, wherein: The InSAR measured data interferogram is input into the tested neural network model to obtain the untangled true phase, including: The ResNet network and the adaptive moment estimation model with the introduction of momentum correction term are used to unwrap the InSAR measured data interferogram to obtain the unwrapping result of the interferogram.

6. The method according to any one of claims 1 to 5, wherein: The ResNet network is configured as at least a ResNET-18 network, and a corresponding ResNet model is obtained based on different numbers of channels and residual blocks in the module.

7. InSAR phase unwrapping device based on deep convolutional neural network optimization, including: A construction module configured to construct a training dataset; A training module configured to train the ResNet network based on the obtained training data set; A correction module, which is configured to correct the results predicted by the ResNet network through an optimizer and a loss function; A phase unwrapping module configured to input the InSAR measured data interferogram into a tested neural network model to obtain an unwrapped real phase; in, Construct a training dataset, including: By randomly generating matrices, pixel blocks of different sizes and intensities are obtained through scaling and intensity transformation; by generating a certain number of random Gaussian distributions, channel positions are randomly selected and combined into an image; Generate a dataset consisting of these images; Randomly generate images, including: Generate a random 2D array of size base; Scale it to an array of size final dimensions; Multiply by a random coefficient and set its minimum value to zero; Randomly generate Gaussian images, including: Generate a two-dimensional array of all zeros; Create a one-dimensional array with a preset range and a preset step size; Generate a grid matrix to represent the horizontal and vertical coordinates of the image, use the coordinate grid to calculate the value of the Gaussian distribution and accumulate it to a two-dimensional all-zero array; Repeat the loop an appropriate number of times, generating a Gaussian distribution each time, and then adding multiple Gaussian distributions to the array, each of which consists of a random variance, a random translation, and a random amplitude, until the minimum value of the array is reset to zero; Generate a dataset consisting of these images, including: Random array-based images and images consisting of multiple Gaussian distributions are generated by cyclic interleaving to construct the training dataset.

8. A computer-readable storage medium having computer-executable instructions stored thereon, wherein when the computer-executable instructions are executed by a processor, the computer-executable instructions implement: The InSAR phase unwrapping method based on deep convolutional neural network optimization according to any one of claims 1 to 6.

Images