A distributed geosynchronous orbit synthetic aperture radar three-dimensional deformation intelligent inversion method

Abstract

This invention belongs to the field of synthetic aperture radar (SAR) technology, and particularly relates to a distributed intelligent inversion method for three-dimensional deformation of SAR in geostationary orbit. In this method, the generator uses a VUR-net unwrapped network with good phase-preserving properties, and the discriminator uses a Markov discriminator model (PatchGAN), which can solve the problem that traditional deep learning networks do not care about accuracy preservation. To prevent network structure collapse, a pseudo-cyclic consistency loss function is added. Since the data obtained directly is usually entangled, the designed deep learning network has the nonlinear mapping capability of unwrapping; to enhance the network's phase unwrapping capability, a phase unwrapping loss function is introduced.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of synthetic aperture radar, and particularly relates to a three-dimensional deformation intelligent inversion method for a distributed geosynchronous orbit synthetic aperture radar. BACKGROUND

[0002] Differential Interferometry SAR (D-InSAR) of synthetic aperture radar extracts phase information to obtain ground deformation in the interval between two or more observations by differentially processing two or more synthetic aperture radar (SAR) images covering the same area before and after deformation. There are generally two methods for measuring and processing deformation: two-track method and three-track method. The two-track method needs to use external DEM to obtain terrain phase. With the completion and popularization of global high-precision DEM surveying, the two-track method has gradually become the mainstream of D-InSAR technology. Distributed spaceborne SAR systems use formation satellites as platforms to jointly complete multi-angle imaging or interferometric imaging, improve the performance of earth observation through flexible configurations, and can realize high-precision InSAR height measurement, D-InSAR deformation measurement, ground moving target indication (GMTI) and other tasks.

[0003] Spaceborne SAR can be divided into low earth orbit SAR (LEO SAR), medium earth orbit SAR (MEO SAR) and geosynchronous SAR (GEO SAR) according to the orbit height. Current spaceborne SARs in orbit are almost LEO SARs, but LEO SARs have long revisit periods (several days to tens of days), which cannot meet the timeliness requirements of disaster area observation. High-orbit SAR, also known as geosynchronous orbit SAR, is a SAR satellite operating in a geosynchronous orbit at an altitude of about 36000 km. Compared with LEO SAR, GEO SAR system has the advantages of rapid revisit (24 hours) to the target area, longer observation time (several hours) and larger beam coverage area (up to thousands of kilometers of mapping width). At the same time, the distributed configuration can provide multi-angle information at the same time, and the three-dimensional deformation of the scene can be inverted in time through multi-orbit differential interferometry.

[0004] However, the scattering characteristics of the target and the imaging quality of the image are closely related to the observation angle, the image angle of the distributed GEO SAR changes greatly, the images obtained by different observation angles have different texture characteristics and forms, and the image characteristics differ greatly. In addition, the geometric configuration of the distributed GEO SAR is complex, the image angle changes greatly, the imaging geometry is seriously affected by the change of the terrain and the oblique imaging, and the image has different degrees of geometric distortion, which is very unfavorable for subsequent processing. Therefore, for the same scene, the images obtained by different observation angles of the distributed GEO SAR will have significant differences in resolution, geometric distortion and scattering characteristics. The precision of three-dimensional deformation inversion of multi-angle distributed GEO SAR data is low. SUMMARY

[0005] The technical problem of the present application is: in order to solve the above problems, the present application provides a three-dimensional deformation inversion method for intelligent processing of distributed GEO SAR differential interference data at different angles, which is based on a neural network intelligent processing framework, can remove the differences in resolution, geometric distortion and scattering characteristics caused by different observation angles, improve the comprehensive cognition of the observed scene information, and improve the performance of three-dimensional deformation inversion processing.

[0006] The technical solution of the present application is:

[0007] A three-dimensional deformation intelligent inversion method for distributed geosynchronous orbit synthetic aperture radar, the steps of the method comprise:

[0008] Firstly, the trajectory information of the distributed geosynchronous orbit synthetic aperture radar is obtained;

[0009] The trajectory information includes velocity and coordinates;

[0010] The trajectory information is obtained by Satellite Tool Kit software;

[0011] Secondly, the distributed geosynchronous orbit synthetic aperture radar observes the scene center according to the trajectory information obtained in the first step at every set time interval, and obtains three SAR images at different angles for each observation, the three SAR images are a first image, a second image and a third image, wherein the first image pair is provided by the monostatic data obtained by the main star in self-receiving, the second image pair is provided by the bistatic data obtained by the main star transmitting and the satellite GEO-2 receiving, and the third image pair is provided by the bistatic data obtained by the main star transmitting and the satellite GEO-3 receiving;

[0012] Thirdly, the three SAR image pairs obtained by twice observation of the same area in the second step are subjected to differential interference processing, and the differential interference phase images of the three images are obtained, and the real phase images of the three images are obtained after removing the 2π wrapping.

[0013] The fourth step is to train the deep learning network by using the differential interference phase image obtained in the third step and the real phase image as a training set;

[0014] During the training, the image with the highest accuracy is selected as the migration target data set, and the other two images are used as the original images for migration and unwrapping;

[0015] The fifth step is to use the generator trained in the fourth step to migrate and unwrap the features of the three images;

[0016] The sixth step is to obtain one-dimensional deformation data in the line-of-sight angle direction of the distributed geosynchronous orbit synthetic aperture radar according to the differential interference phase image obtained in the fifth step after migration and unwrapping, and to obtain three-dimensional deformation inversion of the entire scene using the least squares method.

[0017] In the first step, the trajectory information is obtained according to the orbital elements of the three formation satellites, including semi-major axis a c , eccentricity e c , inclination i c , right ascension of the ascending node Omega c , argument of perigee omega c , and mean anomaly M c .

[0018] In the second step, the scene includes a scene center point, a scene size, and a scene DEM. When the scene center is observed, the deformation variable within the observation interval is included.

[0019] In the fourth step, the deep learning network is a pseudo-cycle-consistent image fusion deep learning network (Pseudo-CycleGAN). The deep learning network includes a generator and a discriminator. The generator uses an unwrapping network VUR-net with good phase-preserving properties, and the discriminator uses a Markov discriminator model (PatchGAN). The two sets of adversarial loss functions used in the deep learning network are represented as follows:

[0020]

[0021]

[0022] Where L GAN (·) represents the adversarial loss, E[·] represents the mathematical expectation of the mean value, v~Pdata(v) represents that the variable v belongs to the data domain Pdata(v), D V and D U represent the discriminators V and U, respectively, and F and G represent the generators F and G, respectively.

[0023] The pseudo-cycle consistency loss function used in the deep learning network is represented as:

[0024]

[0025] Wherein, L fcyc (·) represents the cycle consistency loss, and ||·||1 represents the matrix 1 norm.

[0026] The phase unwrapping loss function used in the deep learning network is represented as:

[0027]

[0028] Wherein, L wrap (·) represents the unwrapping loss.

[0029] Advantages

[0030] (1) In the method of the present application, the generator uses the unwrapping network VUR-net with good phase preserving properties, and the discriminator uses the Markov discriminator model (PatchGAN), which can solve the problem that the traditional deep learning network does not care about the precision preservation problem;

[0031] (2) In the method of the present application, in order to prevent the network structure from collapsing, a pseudo-cycle consistency loss function is added;

[0032] (3) In the method of the present application, since the data obtained directly is usually wrapped, the designed deep learning network has the ability of unwrapping this nonlinear mapping, and in order to enhance the phase unwrapping ability of the network, a phase unwrapping loss function is introduced. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 The flowchart of the method proposed in the present application is shown in the figure;

[0034] Figure 2 The network structure diagram of Pseudo-CycleGAN proposed in the present application is shown in the figure;

[0035] Figure 3 The basic diagram of distributed GEO SAR three-dimensional deformation inversion is shown in the figure;

[0036] Figure 4 Scene DEM and deformation;

[0037] Figure 5 The generated three-direction differential interferograms are shown in the figure;

[0038] Figure 6 The output result after using the Pseudo-CycleGAN proposed in the present application is shown in the figure;

[0039] Figure 7The three-dimensional deformation inversion result output after using the method. DETAILED DESCRIPTION

[0040] The application will be further described below in conjunction with the accompanying drawings and examples.

[0041] The overall processing flow of the application is shown in Figure 1 The three-dimensional deformation inversion accuracy is improved by unifying SAR image features through a multi-angle fusion deep learning network. The method completes the migration of image feature differences under different viewing angles to the highest image by designing the structure of the deep learning network. Since the obtained image is usually wrapped, and the unwrapped image is needed for subsequent processing, an unwrapping loss is added in the traditional cycle generative adversarial network (CycleGan) to complete the nonlinear mapping while completing the image feature migration. Essentially, image feature migration and phase unwrapping belong to two different nonlinear mappings, and the essence of the Pseudo-CycleGAN proposed in the method is to complete these two nonlinear mappings with high phase preservation.

[0042] The application will be further described below in conjunction with the accompanying drawings and examples.

[0043] The training simulation input of the method of the application is the orbital elements (semi-major axis a c , eccentricity e c , inclination i c , ascending node right ascension Omega c , argument of perigee omega c , mean anomaly M c ) of three formation satellites, the target scene position (scene center point, scene size, scene DEM), the observation interval, and the deformation variable in the observation interval; and the output is the three-dimensional deformation inversion result after fusion of three angle data. The specific steps are as follows:

[0044] First, according to the input orbital elements of the satellite formation, the velocity and coordinate trajectory information of the distributed GEO SAR system is obtained through the Satellite Tool Kit software;

[0045] Second, the input scene DEM is as shown in Figure 4 (a), and the height direction deformation variable is as shown in Figure 4(b) shown. At every set time interval, the distributed geosynchronous orbit synthetic aperture radar observes the scene center according to the track information obtained in the first step, and each observation obtains three SAR images of different perspectives, which are the first image, the second image and the third image, wherein the first image pair is provided by the monostatic data obtained by the main star in self-transmission and self-reception, the second image pair is provided by the bistatic data obtained by the main star in transmission and the satellite GEO-2 in reception, and the third image pair is provided by the bistatic data obtained by the main star in transmission and the satellite GEO-3 in reception. The data registration of the multiple distributed satellite SAR image pairs is realized by using the cross-correlation method or the interference data registration method based on external DEM assistance. Then, the conjugate multiplication of the main and auxiliary images of each registered distributed SAR image pair is performed to generate an interferogram, as shown in Figure 5 ;

[0046] In the third step, differential interference processing is performed. According to the satellite track information and the scene DEM data, the reference terrain and flat phase interference phase information is calculated through the range-doppler equation, the satellite interferogram is conjugate multiplied with the respective reference terrain and flat phase to remove the terrain and flat phase, and the differential interference phase information of each satellite is obtained. The differential interference phase images of the three images are obtained, and the real phase images of the three images are obtained after removing the 2π wrapping, and the wrapped phase image is as shown in Figure 5 ;

[0047] In the fourth step, the network is trained according to the data set generated in the third step. The deep learning network structure is as shown in Figure 2 , reasonable parameters are selected for network training and verification, and the network training parameters need to be adjusted according to the characteristics of the data set;

[0048] In the fifth step, the trained deep learning network is used for image feature migration and phase unwrapping. The training result contains all the generators and discriminators in the network structure, and the required result is the mapping network from the original wrapped image to the target unwrapped image, that is, one of the generators. The generator can complete the image feature unification and the unwrapping non-linear mapping at the same time, and the result after the deep learning network mapping is as shown in Figure 6 ;

[0049] In the sixth step, the one-dimensional deformation data of the radar line-of-sight angle can be obtained according to the migrated and unwrapped differential interference phase image obtained in the fifth step, and then the least square method is used to solve and obtain the three-dimensional deformation information of the entire scene. The basic schematic diagram of the distributed GEO SAR three-dimensional deformation inversion is as shown in Figure 3 , and the deformation inversion result is as shown in Figure 7 .

[0050] d = (d N , d E , dU ) T d N is the north-south direction deformation, d E is the east-west direction deformation, d U is the vertical direction deformation. are the unit vectors of the north, east and vertical directions, respectively. According to the least square method, the estimation of the three-dimensional deformation d

[0051]

[0052] where Θ is the three-dimensional deformation model coefficient matrix, which can be expressed as

[0053]

[0054] C Φ is the covariance matrix of the phase error, which can be expressed as

[0055]

[0056] Step 7, three-dimensional deformation inversion accuracy analysis.

[0057] The concept of positioning accuracy coefficient (Position Dilution of Precision, PDOP) is introduced to evaluate the comprehensive performance of the target three-dimensional deformation inversion accuracy under different sub-aperture data combinations. According to the classical definition of PDOP, the accuracy coefficient PDOP d of the three-dimensional deformation inversion is defined as

[0058]

[0059] It is assumed that the phase errors of the three sub-apertures are consistent, and the above formula can be simplified as

[0060]

[0061] At this point, all steps are completed.

[0062] Next, an example of implementation is given combined with specific parameters.

[0063] In this example, the parameters of the distributed satellite formation are shown in Table 1, considering an L-band system with an imaging resolution of 9.738m*13.762m. Taking east longitude 120.74 and north latitude 31.64 as the center of the scene, it is assumed that the middle of the scene is a mountainous area with terrain fluctuations, and the scene size is about 680m*680m. The scene DEM is as follows: Figure 4As shown in (a), distributed GEO SAR observes scene deformation at daily intervals. During this period, complex height-direction deformations ranging from a few centimeters to tens of centimeters occurred in the mountainous area. The height-direction deformations are as follows: Figure 4 As shown in (b), there is no variation in the east-west and north-south directions. Multi-angle observations were conducted at three different observation angles, corresponding to the latitude arguments of the aperture center positions: 1°, 146°, and 50°. The scene is now complete.

[0064] Table 1 GEO SAR Formation Parameters

[0065]

[0066]

[0067] The training parameters are set as shown in Table 2.

[0068] To compare with traditional multi-angle-based 3D deformation methods, a dataset of differential interferometric phase maps without entanglement was simulated. Using traditional multi-angle-based 3D deformation methods, the deformation inversion errors in the east-west, north-south, and height directions were 1.81 cm, 1.82 cm, and 2.42 cm, respectively. Using the algorithm proposed in this patent, the deformation inversion errors in the east-west, north-south, and height directions were 0.48 cm, 0.65 cm, and 1.97 cm, respectively, representing improvements of 73%, 64%, and 18% in the three directions compared to traditional methods. Simulation results demonstrate that the multi-angle image fusion-based 3D deformation inversion method can improve the accuracy of distributed GEO SAR 3D deformation inversion.

[0069] Table 2 Training Parameter Settings

[0070]

[0071] It is particularly important to point out that this model essentially adds a nonlinear mapping of untangling to the solution of feature transfer problem. Theoretically, this nonlinear mapping can be replaced with different mapping functions in different scenarios. Simply replace the generator in Pseudo-CycleGAN with a network structure with special functions to achieve style transfer and nonlinear mapping in different scenarios.

[0072] Of course, the present invention may have other various embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art can make various corresponding changes and modifications according to the present invention, but these corresponding changes and modifications should all fall within the protection scope of the appended claims.

Claims

1. A distributed geosynchronous orbit synthetic aperture radar three-dimensional deformation intelligent inversion method, characterized in that... The steps of this method include: The first step is to acquire trajectory information from a distributed geosynchronous orbit synthetic aperture radar. The second step involves the distributed geosynchronous orbit synthetic aperture radar observing the center of the scene based on the trajectory information obtained in the first step at set time intervals. Each observation yields three SAR images from different perspectives. The third step involves performing differential interferometry on the three SAR image pairs obtained from two observations of the same area in the second step to obtain the differential interferometric phase map of the three images. After removing the 2π winding, the true phase map of the three images is obtained. The fourth step involves using the differential interferometric phase maps of the three images obtained in the third step and the true phase maps of the three images as the training set to train the deep learning network. The fifth step involves using the generator trained in the fourth step to transfer and untangle the features of the three images. The sixth step involves obtaining one-dimensional deformation data in the line-of-sight angle direction of the distributed geosynchronous orbit synthetic aperture radar based on the differential interferometric phase map obtained in the fifth step after migration and unwrapping. The three-dimensional deformation inversion of the entire scene is then obtained using the least squares method.

2. The intelligent inversion method for three-dimensional deformation of distributed geosynchronous orbit synthetic aperture radar according to claim 1, characterized in that: In the first step, the trajectory information includes speed and coordinates, and the trajectory information is obtained through Satellite Tool Kit software.

3. The intelligent inversion method for three-dimensional deformation of distributed geosynchronous orbit synthetic aperture radar according to claim 2, characterized in that: When acquiring trajectory information, it is obtained based on the orbital root numbers of the three formation satellites, including the semi-major axis 'a'. c eccentricity e c Inclination angle i c Right ascension of ascending node Ω c Argument of perigee ω c , the angle of near point M c .

4. The intelligent inversion method for three-dimensional deformation of distributed geosynchronous orbit synthetic aperture radar according to any one of claims 1-3, characterized in that: In the second step, the three SAR images are the first image, the second image, and the third image, respectively. The first image provides monostatic data obtained by the primary satellite through self-transmission and self-reception, the second image provides bistatic data transmitted by the primary satellite and received from satellite GEO-2, and the third image provides bistatic data transmitted by the primary satellite and received from satellite GEO-3.

5. The intelligent inversion method for three-dimensional deformation of distributed geosynchronous orbit synthetic aperture radar according to claim 1, characterized in that: In the second step, the scene includes the scene center point, scene size, and scene DEM. When observing the scene center, the deformation within the observation interval is included.

6. The intelligent inversion method for three-dimensional deformation of distributed geosynchronous orbit synthetic aperture radar according to claim 1, characterized in that: In the fourth step, during training, the image with the highest accuracy is selected as the target dataset for transfer, and the other two images are used as the original images for transfer and unwrapping.

7. The intelligent inversion method for three-dimensional deformation of distributed geosynchronous orbit synthetic aperture radar according to claim 6, characterized in that: In the fourth step, the deep learning network is a pseudo-cyclic consistency image fusion deep learning network, Pseudo-CycleGAN. The deep learning network includes a generator and a discriminator. The generator uses a phase-preserving untangled network, VUR-net, and the discriminator uses a Markov discriminator model, PatchGAN.

8. The intelligent inversion method for three-dimensional deformation of distributed geosynchronous orbit synthetic aperture radar according to claim 7, characterized in that: The two sets of adversarial loss functions used in deep learning networks are expressed as follows: Among them, L GAN (·) represents the adversarial loss, E[·] represents the mathematical expectation of the mean, v~Pdata(v) means that the variable v belongs to the data domain Pdata(v), D V and D U Let V and U represent discriminators respectively, and F and G represent generators F and G respectively.

9. The intelligent inversion method for three-dimensional deformation of distributed geosynchronous orbit synthetic aperture radar according to claim 8, characterized in that: The pseudo-cyclic consistency loss function used in deep learning networks is expressed as: Where L fcyc (·) represents the cycle consistency loss, and ||·||1 represents the matrix 1-norm.

10. The intelligent inversion method for three-dimensional deformation of distributed geosynchronous orbit synthetic aperture radar according to claim 9, characterized in that: The phase unwrapping loss function used in deep learning networks is expressed as: Where L wrap (·) indicates the loss from untangling.

Images