Interferometric synthetic aperture radar baseline error estimation method based on mechanism training

Abstract

The invention discloses an interferometric synthetic aperture radar baseline error estimation method based on mechanism training, which comprises the following steps: firstly, initializing interferometric synthetic aperture radar data, defining two interferometric baseline error models, calculating the distance from a baseline to a scenic spot to obtain two baseline interferometric phases, and generating a baseline error interferometric phase diagram; then inputting the interferometric phase diagram into a deep learning network, reconstructing the interferometric phase diagram, training the deep learning network by minimizing loss between the interferometric phase diagram and the reconstructed interferometric phase diagram, and finally inputting an interferometric phase diagram test data set containing errors into the trained deep learning network, and obtaining two baseline error parameters and a baseline trajectory result after motion error compensation. The method provided by the invention solves the problems of strong external data dependence, insufficient time-varying baseline adaptability, low phase sensitivity and scene limitation in the prior art, and improves the correctness of the interferometric phase and the precision of the finally obtained ground elevation or deformation.

Description

Technical Field

[0001] The invention belongs to the technical field of remote sensing detection, and in particular relates to an interferometric synthetic aperture radar baseline error estimation method based on mechanism training. Background Art

[0002] Interferometric synthetic aperture radar baseline error estimation technology is one of the important issues in the field of remote sensing detection. The baseline error is an important parameter in interferometric synthetic aperture radar processing and is of great significance for improving the correctness of the interferometric phase and the accuracy of the ultimately obtained ground elevation or deformation.

[0003] Existing interferometric synthetic aperture radar baseline error estimation methods are mainly divided into two categories: spatial domain methods and transform domain methods. In the spatial domain method, the document "DEM-assisted InSAR zero-IF processing baseline estimation method. Journal of Huazhong University of Science and Technology (Natural Science Edition), 2017, 45(03):1-6" proposes the use of joint optimization of baseline parameters. This method uses external terrain data to constrain the baseline error, significantly improving the estimation accuracy in flat areas. However, it relies on high-precision digital elevation models and ground control point layout, and requires high-precision external terrain data. The document "Spaceborne Squint SAR Maneuvering Target Imaging and Motion Parameter Estimation Method. Proceedings of the 8th Annual Conference on High-Resolution Earth Observation, 2022:13-22" improves imaging quality through motion parameter estimation. However, this method does not incorporate the physical relationship between the platform dynamic parameters and baseline error into the model, making dynamic baseline error difficult to correct. In the transform domain method, the paper “A new method for baseline estimation of spaceborne InSAR based on interference fringe frequency. Journal of Electronics, 2011, 39(6):1289-1295” inverts the baseline parameters by analyzing the interference phase spectrum characteristics. This method assumes that the baseline is a fixed value and the phase is globally continuous. However, in actual scenarios, sudden changes in terrain or low coherence areas can lead to spectral aliasing, which in turn causes baseline estimation errors.

[0004] In summary, existing methods are limited by factors such as external control points and terrain conditions, and estimate the baseline as a constant value. These methods are not applicable to situations such as complex terrain, lack of external auxiliary information, and time-varying baseline errors. Summary of the Invention

[0005] To solve the above technical problems, the present invention provides an interferometric synthetic aperture radar baseline error estimation method based on mechanism training, which improves the correctness of the interferometric phase and the accuracy of the ultimately acquired ground elevation or deformation.

[0006] The technical solution adopted by the present invention is: an interferometric synthetic aperture radar baseline error estimation method based on mechanism training, the specific steps are as follows:

[0007] S1. Initialize interferometric synthetic aperture radar parameters;

[0008] Initialize the digital elevation model (DEM) of the target area, set the interferogram resolution in the X direction to dx, the interferogram resolution in the Y direction to dy, and the interferogram image size to N. Then initialize the radar parameters, including baseline length B, baseline errors E1 and E2, synthetic aperture radar carrier frequency fc, synthetic aperture radar wavelength λ, and wave number K.

[0009] S2. Based on step S1, a distributed interferometric SAR single-view complex image is selected, and a pairwise interferometric baseline error model is defined to calculate the distance from the baseline to the scene point, thereby obtaining the interferometric phase of the two baselines and generating an interferometric phase map including the baseline error.

[0010] S3, inputting the interference phase image obtained in step S2 into the deep learning network, outputting the error baseline vector, reconstructing the interference phase image through the mechanism generation module, and training the deep learning network through backpropagation by minimizing the loss between the interference phase image and the reconstructed interference phase image, thereby obtaining a trained deep learning network;

[0011] Based on step S2, different error parameter combinations are randomly generated and imported into the baseline error model to obtain the interferometric phase image dataset. The interferometric phase image dataset containing the error is represented as X, and the interferometric phase image dataset X is divided into the training set X according to the set ratio. train and the test set X test , training set X train As the training input of the deep learning network, the test set X test As input to the trained deep learning network.

[0012] S4. Based on step S3, the interferometric phase image test data set containing errors is input into the trained deep learning network to obtain two baseline error parameters and the baseline trajectory results after motion error compensation, thereby realizing distributed synthetic aperture radar imaging.

[0013] Furthermore, the step S2 is specifically as follows:

[0014] Define the ideal interferometric baseline pair B ideal1 (n), B ideal2 (n), construct a pairwise interference baseline error model, and define the two baseline errors as a form including multiple parameters: E1(n; x1, x2, ..., x m )、E2(n;y1,y2,...,y m ), then the actual baseline model expression with errors is as follows:

[0015] B1(n)=B ideal1 (n)+E1(n;x1,x2,...,x m )

[0016] B2(n)=B ideal2 (n)+E2(n;y1,y2,...,y m )

[0017] Where n represents a discrete time series, E1(n; x1, x2, ..., x m )、E2(n;y1,y2,...,y m ) represent the baseline error vector containing m parameters.

[0018] Then, use the digital elevation model of the target area to assist or select the flat area of ​​the target area to calculate the shortest slant range difference between the main and secondary antennas to a certain scene point Q. The expression is as follows:

[0019] R1(n;Q)=||B1(n)-Q||2

[0020] R2(n;Q)=||B2(n)-Q||2

[0021] Where Q = [a, b, h] T , represents the scene point Q space position vector, [·] T Indicates transpose.

[0022] Finally, the interference phase is calculated to obtain the interference phase diagram. The expression is as follows:

[0023]

[0024] in, Indicates the phase difference between the primary and secondary antennas and scene point Q.

[0025] Furthermore, the step S3 is specifically as follows:

[0026] The output of the deep learning network is an error baseline vector including baseline errors E1 and E2 Then input the error baseline vector Y into the mechanism generation module and output the reconstructed interference phase map

[0027] Assume that the deep learning network is represented as f(X train ;ω)=ωX train +b, where ω represents the deep learning network parameter weight and b represents the network bias term, which can be ignored. The network is represented as f(X train ;ω)=ωX train , the mechanism generation module is expressed as

[0028] Then the loss between the interferometric phase image with error and the reconstructed interferometric phase image is calculated and expressed as During training, X trainis the training input, then the loss is a function of Y, denoted as L(X train ,g(Y)), minimize the loss to get the optimal network weight When the network gradient is back-propagated, the gradient of the loss with respect to the network weight is calculated, and the expression is as follows:

[0029]

[0030] in, Obtained by automatic back propagation of the deep learning network, Represents the gradient of the mechanism generation module. It is calculated by hybrid gradient derivation. The expression is as follows:

[0031]

[0032] Where ΔY is a vector of the same dimension as Y, representing a small increment of the error baseline vector Y, which is used to approximate the gradient. Then, the gradient descent method is used to update the parameters in the deep learning network and update the weight ω of the k+1th step. k+1 , the expression is as follows:

[0033]

[0034] Here, α represents the learning rate.

[0035] Finally, train until the network converges. When the loss function is less than the convergence threshold σ, the training has converged and the training is stopped, and a trained deep learning network is obtained.

[0036] Furthermore, the step S4 is specifically as follows:

[0037] The interferometric phase image test data set X test Input the trained deep learning network and output the error vector estimate According to the error vector estimate Get the estimated value of the two baseline error vectors Compensate the baseline error model and calculate the correct trajectory of the two baselines. The expression is as follows:

[0038]

[0039] Finally, the error vector estimate is substituted into the distributed interferometric SAR antenna phase center trajectory, the baseline error model is compensated, and the correct trajectory of the two baselines is calculated. It is then combined with the distributed interferometric SAR tomography data to realize distributed synthetic aperture radar imaging.

[0040] The beneficial effects of the present invention are as follows: the method of the present invention first initializes the interferometric synthetic aperture radar data, defines two interferometric baseline error models, calculates the distance from the baseline to the scene point to obtain the two baseline interferometric phases, generates an interferometric phase map including the baseline error, then inputs the interferometric phase map into a deep learning network, reconstructs the interferometric phase map, trains the deep learning network by minimizing the loss between the interferometric phase map and the reconstructed interferometric phase map, and finally inputs the interferometric phase map test data set containing the error into the trained deep learning network to obtain the two baseline error parameters and the baseline trajectory result after motion error compensation. Compared with the existing radar baseline error estimation method, the method of the present invention fully utilizes the feature extraction and pattern recognition capabilities of mechanism training, analyzes the interferometric synthetic aperture radar data through the training model, obtains the baseline error parameter estimation that conforms to the actual situation and corrects the satellite trajectory, and can effectively suppress the error propagation under complex terrain conditions, improve the elevation inversion accuracy, and achieve high-precision interferometric synthetic aperture radar data elevation inversion. It can be used for various data analysis and processing tasks under complex terrain conditions, solves the problems of strong external data dependence, insufficient adaptability to time-varying baselines, low phase sensitivity and scene limitations in the existing technology, and improves the accuracy of the interferometric phase and the accuracy of the ultimately obtained ground elevation or deformation. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 The present invention is a flow chart of an interferometric synthetic aperture radar baseline error estimation method based on mechanism training.

[0042] Figure 2 Schematic diagram of mechanism training in an embodiment of the present invention.

[0043] Figure 3 This is a diagram showing the simulation results of the interferometric synthetic aperture radar baseline error estimation in an embodiment of the present invention. DETAILED DESCRIPTION

[0044] The method of the present invention is further described below with reference to the accompanying drawings and embodiments.

[0045] In order to facilitate the description of the content of the present invention, the following terms are first defined:

[0046] Definition 1: Elevation digital terrain (DEM);

[0047] A digital elevation model (DEM) is a digital model that discretizes the spatial distribution of surface elevation using a regular grid. It is acquired through techniques such as aerial photogrammetry, radar interferometry, and lidar scanning. Each grid cell records the elevation value corresponding to a geographic coordinate. The accuracy of this data directly affects the terrain inversion results from radar interferometry.

[0048] Definition 2: Interferometric synthetic aperture radar (InSAR);

[0049] Interferometric synthetic aperture radar (InSAR) is a microwave remote sensing technology based on synthetic aperture radar (SAR) image pairs. Using radar antennas to observe the same surface feature at different spatial locations, InSAR calculates the phase difference between the two images to invert the surface shape or three-dimensional terrain structure. Its measurement accuracy is affected by parameters such as baseline length, radar wavelength, and platform orbit stability.

[0050] Definition 3: Interference phase pattern;

[0051] An interferometric phase map is a two-dimensional phase distribution generated by interferometry of complex InSAR image pairs. Its pixel values ​​represent the phase variations caused by differences in radar signal propagation paths. This phase map incorporates the coupled effects of surface elevation, deformation, and systematic errors, requiring phase unwrapping and error compensation to separate the effective information.

[0052] Definition 4: Deep learning network;

[0053] Deep learning networks are machine learning models with multi-layered nonlinear transformation structures that achieve end-to-end task solving by automatically learning the intrinsic characteristics of data. Typical architectures include convolutional neural networks (CNNs) and generative adversarial networks (GANs), which are commonly used in remote sensing to solve complex nonlinear mapping problems such as image segmentation, noise suppression, and parameter inversion. Based on a data-driven, high-dimensional nonlinear mapping model, end-to-end training establishes a correlation mapping between phase distortion and error parameters, outputting error estimates to drive system parameter correction, ultimately improving interferometric measurement accuracy.

[0054] like Figure 1 As shown in FIG, a flow chart of an interferometric synthetic aperture radar baseline error estimation method based on mechanism training of the present invention is shown, and the specific steps are as follows:

[0055] S1. Initialize interferometric synthetic aperture radar parameters;

[0056] To estimate the interferometric synthetic aperture radar baseline error, a digital elevation model (DEM) is required, or a flat area in the target area is selected. Initialize the DEM (digital elevation model) of the target area, setting the interferogram resolution in the X direction to dx, the interferogram resolution in the Y direction to dy, and the interferogram phase image size to N. Then initialize the radar parameters, including the baseline length B (the spacing between the primary and secondary antennas), baseline errors E1 and E2, synthetic aperture radar carrier frequency fc, synthetic aperture radar wavelength λ, and wave number K.

[0057] In this embodiment, the baseline error includes: baseline amplitude error A, baseline frequency error f, and baseline phase error P. Specifically, E1(n; A1, f1, P1) and E2(n; A2, f2, P2) represent baseline error vectors containing the baseline amplitude error A, frequency error f, and phase error P, respectively.

[0058] S2. Based on step S1, a distributed interferometric SAR single-view complex image is selected, and a pairwise interferometric baseline error model is defined to calculate the distance from the baseline to the scene point, thereby obtaining the interferometric phase of the two baselines and generating an interferometric phase map including the baseline error.

[0059] S3, inputting the interference phase image obtained in step S2 into the deep learning network, outputting the error baseline vector, reconstructing the interference phase image through the mechanism generation module, and training the deep learning network through backpropagation by minimizing the loss between the interference phase image and the reconstructed interference phase image, thereby obtaining a trained deep learning network;

[0060] Based on step S2, different error parameter combinations are randomly generated and imported into the baseline error model to obtain an interferometric phase image dataset. The interferometric phase image dataset containing errors is represented as X, and the interferometric phase image dataset X is divided into a training set X according to a set ratio (8:2 in this embodiment) train and the test set X test , training set X train As the training input of the deep learning network, the test set X test As input to the trained deep learning network.

[0061] S4. Based on step S3, the interferometric phase image test data set containing errors is input into the trained deep learning network to obtain two baseline error parameters and the baseline trajectory results after motion error compensation, thereby realizing distributed synthetic aperture radar imaging.

[0062] In this embodiment, step S2 is specifically as follows:

[0063] Define the ideal interferometric baseline pair B ideal1 (n), B ideal2 (n), construct a pairwise interference baseline error model, and define the two baseline errors as a form including multiple parameters: E1(n; x1, x2, ..., x m )、E2(n;y1,y2,...,y m ), in this embodiment, it is E1(n; A1, f1, P1) and E2(n; A2, f2, P2), then the actual baseline model expression with error is as follows:

[0064] B1(n)=B ideal1 (n)+E1(n;A1,f1,P1)

[0065] B2(n)=B ideal2 (n)+E2(n;A2,f2,P2)

[0066] Where n represents a discrete time series, E1(n; x1, x2, ..., x m)、E2(n;y1,y2,...,y m ) represent the baseline error vectors containing m parameters. The two baseline error vector expressions are defined as follows:

[0067] E1(n;A1,f1,P1)=0

[0068]

[0069] Then, use the digital elevation model of the target area to assist or select the flat area of ​​the target area to calculate the shortest slant range difference between the main and secondary antennas to a certain scene point Q. The expression is as follows:

[0070] R1(n;Q)=||B1(n)-Q||2

[0071] R2(n;Q)=||B2(n)-Q||2

[0072] Where Q = [a, b, h] T , represents the scene point Q space position vector, [·] T Indicates transpose.

[0073] Finally, the interference phase is calculated to obtain the interference phase diagram. The expression is as follows:

[0074]

[0075] in, Indicates the phase difference between the primary and secondary antennas and scene point Q.

[0076] In view of the fact that radar baseline error is sensitive to interference phase in interferometric synthetic aperture radar data processing and analysis technology, the SAR interference pattern characteristics under different error conditions are integrated into the deep learning network training process, so that the network has the ability to recognize the interference pattern characteristics under specific interferometric synthetic aperture radar parameters and ground environmental conditions.

[0077] In this embodiment, step S3 is specifically as follows:

[0078] The specific mechanism training diagram of the mechanism generation module in this embodiment is as follows Figure 2 As shown in the figure, the system consists of four key components: parameter representation graph input, network estimation module, mechanism generation module, and decision module. The parameter representation graph is input into the network estimation module for parameter estimation and numerical gradient generation. The physical mapping model and mechanism generation module generate a representation graph corresponding to the estimated parameters. Finally, the results are input into the decision module to evaluate the representation effect and calculate the loss, which is used for network training and ultimately optimize model performance.

[0079] The output of the deep learning network is an error baseline vector including baseline errors E1 and E2 In this embodiment Then input the error baseline vector Y into the mechanism generation module and output the reconstructed interference phase map

[0080] Assume that the deep learning network is represented as f(X train ;ω)=ωX train +b, where ω represents the deep learning network parameter weight and b represents the network bias term, whose effect on the network output is independent of the input and can be ignored. The network is represented as f(X train ;ω)=ωX train , the mechanism generation module is expressed as

[0081] Then the loss between the interferometric phase image with error and the reconstructed interferometric phase image is calculated and expressed as During training, X train is the training input, then the loss is a function of Y, denoted as L(X train ,g(Y)), minimize the loss to get the optimal network weight When the network gradient is back-propagated, the gradient of the loss with respect to the network weight is calculated, and the expression is as follows:

[0082]

[0083] in, Obtained by automatic back propagation of the deep learning network, Indicates the gradient of the mechanism generation module, which cannot be directly derived. It is calculated using mixed gradient derivation. The expression is as follows:

[0084]

[0085] Where ΔY is a vector of the same dimension as Y, representing a small increment of the error baseline vector Y, which is used to approximate the gradient. Then, the gradient descent method is used to update the parameters in the deep learning network and update the weight ω of the k+1th step. k+1 , the expression is as follows:

[0086]

[0087] Here, α represents the learning rate.

[0088] Finally, train until the network converges. When the loss function is less than the convergence threshold σ, the training has converged and the training is stopped, and a trained deep learning network is obtained.

[0089] In this embodiment, step S4 is specifically as follows:

[0090] The interferometric phase image test data set Xtest Input the trained deep learning network and output the error vector estimate In this embodiment According to the error vector estimate Get the estimated value of the two baseline error vectors Compensate the baseline error model and calculate the correct trajectory of the two baselines. The expression is as follows:

[0091]

[0092] Finally, the error vector estimate is substituted into the distributed interferometric SAR antenna phase center trajectory, the baseline error model is compensated, and the correct trajectory of the two baselines is calculated. It is then combined with the distributed interferometric SAR tomography data to realize distributed synthetic aperture radar imaging.

[0093] The simulation results of the interferometric synthetic aperture radar baseline error estimation in this embodiment are as follows: Figure 3 As shown in the figure, it can be seen that by minimizing the loss between the interference phase image dataset and the reconstructed interference phase image, the amplitude error, frequency error and phase error can be trained.

[0094] In summary, the method of the present invention applies the mechanism learning method to the interferometric synthetic aperture radar baseline error estimation method, generates interference fringes containing dynamic baseline errors by interfering pairwise single-view complex SAR images, constructs an optimization search problem by using the SAR interferometric fringes mechanism, solves the optimization parameters through network training, obtains the estimation of the dynamic baseline error corresponding to the interferometric baseline, substitutes the estimated dynamic baseline error into the distributed interferometric SAR antenna phase center trajectory, obtains the antenna phase center trajectory after motion error compensation, and combines it with the distributed interferometric SAR tomography data for subsequent sparse three-dimensional imaging and interferometric and tomographic data processing. This improves the correctness of the interferometric phase and the accuracy of the ultimately acquired ground elevation or deformation. The mechanism modeling defines dynamic error parameters, mechanism-deep learning network fusion training, and dynamic compensation correction trajectory, which systematically solves the problems of the existing technology such as strong external data dependence, insufficient adaptability to time-varying baselines, low phase sensitivity, and scene limitations.

[0095] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and it should be understood that the scope of protection of the present invention is not limited to such specific descriptions and embodiments. Those skilled in the art can make various other specific variations and combinations based on the technical teachings disclosed in the present invention without departing from the essence of the present invention, and such variations and combinations are still within the scope of protection of the present invention.

Claims

1. A method for estimating baseline error of interferometric synthetic aperture radar based on mechanism training, the specific steps are as follows: S1. Initialize interferometric synthetic aperture radar parameters; Initialize the digital elevation model (DEM) of the target area, set the X-direction interference pattern resolution to dx, the Y-direction interference pattern resolution to dy, and the interference phase pattern image size to N; Reinitialize the radar parameters, including baseline length B, baseline errors E1, E2, synthetic aperture radar carrier frequency fc, synthetic aperture radar wavelength λ, and wave number K; S2. Based on step S1, a distributed interferometric SAR single-view complex image is selected, and a pairwise interferometric baseline error model is defined to calculate the distance from the baseline to the scene point, thereby obtaining the interferometric phase of the two baselines and generating an interferometric phase map including the baseline error. S3, inputting the interference phase image obtained in step S2 into the deep learning network, outputting the error baseline vector, reconstructing the interference phase image through the mechanism generation module, and training the deep learning network through backpropagation by minimizing the loss between the interference phase image and the reconstructed interference phase image, thereby obtaining a trained deep learning network; Based on step S2, different error parameter combinations are randomly generated and the baseline error model is imported to obtain an interferometric phase image dataset. The interferometric phase image dataset containing the error is represented as X, and the interferometric phase image dataset X is divided into a training set X according to a set ratio. train and the test set X test , training set X train As the training input of the deep learning network, the test set X test As input to a trained deep learning network; S4. Based on step S3, the interferometric phase image test data set containing errors is input into the trained deep learning network to obtain two baseline error parameters and the baseline trajectory results after motion error compensation, thereby realizing distributed synthetic aperture radar imaging.

2. The method for estimating baseline error of interferometric synthetic aperture radar based on mechanism training according to claim 1, characterized in that: The step S2 is specifically as follows: Define the ideal interferometric baseline pair B ideal1 (n), B ideal2 (n), construct a pairwise interference baseline error model, and define the two baseline errors as a form including multiple parameters: E1(n; x1, x2, ..., x m )、E2(n;y1,y2,...,y m ), then the actual baseline model expression with errors is as follows: B1(n)=B ideal1 (n)+E1(n;x1,x2,…,x m ) <h2 style=";text-align:left;direction:ltr">B2(n)=B<h2 style=";text-align:left;direction:ltr"> ideal2 <h2 style=";text-align:left;direction:ltr"> (n)+E2(n;y1,y2,...,y<h2 style=";text-align:left;direction:ltr"> m <h2 style=";text-align:left;direction:ltr"> ) Where n represents a discrete time series, E1(n; x1, x2, ..., x m )、E2(n;y1,y2,...,y m ) represent the baseline error vector containing m parameters respectively; Then, use the digital elevation model of the target area to assist or select the flat area of ​​the target area to calculate the shortest slant range difference between the main and secondary antennas to a certain scene point Q. The expression is as follows: R1(n;Q)=||B1(n)-Q||2 R2(n;Q)=||B2(n)-Q||2 Where Q = [a, b, h] T , represents the scene point Q space position vector, [·] T represents transpose; Finally, the interference phase is calculated to obtain the interference phase diagram. The expression is as follows: in, Indicates the phase difference between the primary and secondary antennas and scene point Q.

3. The interferometric synthetic aperture radar baseline error estimation method based on mechanism training according to claim 1, characterized in that: The step S3 is specifically as follows: The output of the deep learning network is an error baseline vector including baseline errors E1 and E2 Then input the error baseline vector Y into the mechanism generation module and output the reconstructed interference phase map Assume that the deep learning network is represented as f(X train ;ω)=ωX train +b, where ω represents the deep learning network parameter weight and b represents the network bias term, which can be ignored. The network is represented as f(X train ;ω)=ωX train , the mechanism generation module is expressed as Then the loss between the interferometric phase image with error and the reconstructed interferometric phase image is calculated and expressed as During training, X train is the training input, then the loss is a function of Y, denoted as L(X train ,g(Y)), minimize the loss to get the optimal network weight When the network gradient is back-propagated, the gradient of the loss with respect to the network weight is calculated, and the expression is as follows: in, Obtained by automatic back propagation of the deep learning network, Represents the mechanism generation module gradient; uses mixed gradient derivation calculation The expression is as follows: Where ΔY is a vector of the same dimension as Y, representing a small increment of the error baseline vector Y, which is used to approximate the gradient; the gradient descent method is then used to update the parameters in the deep learning network and the weight ω of the k+1th step is updated. k+1 , the expression is as follows: Among them, α represents the learning rate; Finally, train until the network converges. When the loss function is less than the convergence threshold σ, the training has converged and the training is stopped, and a trained deep learning network is obtained.

4. The method for estimating baseline error of interferometric synthetic aperture radar based on mechanism training according to claim 1, characterized in that: The step S4 is specifically as follows: The interferometric phase image test data set X test Input the trained deep learning network and output the error vector estimate According to the error vector estimate Get the estimated value of the two baseline error vectors Compensate the baseline error model and calculate the correct trajectory of the two baselines. The expression is as follows: Finally, the error vector estimate is substituted into the distributed interferometric SAR antenna phase center trajectory, the baseline error model is compensated, and the correct trajectory of the two baselines is calculated. It is then combined with the distributed interferometric SAR tomography data to realize distributed synthetic aperture radar imaging.

Images