A 3D reconstruction method and system for sparse array tomographic SAR based on SVD-DCS

Abstract

The present invention provides a sparse array tomographic SAR three-dimensional reconstruction method and system based on SVD-DCS. This method addresses the difficulty of accurate reconstruction and the high number of false targets in tomographic SAR under irregular and sparse sampling conditions. The method preprocesses the sparse array SAR image data, including complex image registration, phase de-skewing, and amplitude and phase error correction. The method also reduces the DCS data processing dimensionality by identifying adjacent pixels of equal height and performing SVD decomposition. The system then reconstructs the reference pixel scattering intensity profile through noise energy estimation and SVD-DCS spectrum estimation. Furthermore, the method selects the model order, estimates target parameters, and geocodes the data to output the spatial position and scattering intensity data of the tomographic SAR point cloud. This method improves the density and reliability of sparse array SAR three-dimensional reconstruction. The method achieves high point cloud reconstruction density, better sidelobe noise suppression, and improved reliability of tomographic SAR point cloud reconstruction.

Description

Technical Field

[0001] The present invention belongs to the technical field of high-dimensional synthetic aperture radar imaging data processing, and in particular relates to a sparse array tomographic SAR three-dimensional reconstruction method and system based on SVD-DCS. Background Art

[0002] Synthetic Aperture Radar (SAR) is an all-day, all-weather, active microwave imaging radar with a certain degree of surface penetration. Developed based on SAR, tomoSAR is a new technology for time-efficient, true stereoscopic, and high-precision 3D reconstruction. It has broad application prospects in areas such as urban 3D construction, field exploration and rescue, environmental reconnaissance in non-cooperative areas, and forest biomass and ice reserve estimation. TomoSAR achieves third-dimensional resolution by adding multiple antennas or performing multiple passes in the normal direction of the azimuth-slant range plane and performing a third-dimensional elevation synthetic aperture, enabling the recovery of information from overlapping areas. With the development of new-generation spaceborne SAR constellations, several satellites are now able to acquire multiple high-resolution SAR images of the same area within a relatively short time interval, providing reliable data support for tomoSAR 3D reconstruction.

[0003] However, when tomographic SAR uses a third-dimensional synthetic aperture along the normal to the azimuth-slant-range plane, multiple flybys or multi-antenna sampling creates an uneven, sparse array. Traditional spectral analysis methods such as DFT, BF, and Capon require uniform elevation signal sampling and a sampling frequency that satisfies the Nyquist sampling theorem for elevation inversion in the overlaid region, resulting in tomographic resolution unable to surpass the limitations of Rayleigh resolution. De-overlapping methods such as MUSIC and singular value decomposition (SVD) can surpass Rayleigh resolution in elevation, but at the expense of spatial resolution. Although tomographic SAR systems have elevation resolution, the limited number of antennas results in a relatively short synthetic aperture in elevation, resulting in elevation resolution far less than both azimuth and slant-range resolution. To achieve third-dimensional super-resolution imaging under conditions of sparse and irregular baseline distribution, the SL1MMER method, based on compressed sensing (CS), has emerged. This method exploits the sparse nature of the target signal structure and performs non-correlated measurements on low-dimensional, low-resolution, sub-Nyquist-sampled data to achieve accurate signal reconstruction. However, when there are few and unevenly distributed baselines, the SL1MMER method struggles to suppress third-dimensional sidelobe noise, leading to the introduction of a large number of false targets.

[0004] As an extension of CS theory, in 2005, Baron et al. proposed distributed compressed sensing (DCS) to address the joint sparsity of multiple signals. Based on the information theory principle that the joint entropy of two sources is less than the sum of their individual entropies, DCS defines the concept of joint sparsity to fully exploit the correlations between multiple signals and perform joint CS. Its performance is superior to that of individual CS signals, thereby reducing the system's sampling rate requirements. Given DCS's superiority over CS, research on DCS-based 3D reconstruction methods for tomographic SAR is crucial for reducing system requirements and improving the timeliness of 3D reconstruction. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a sparse array tomographic SAR three-dimensional reconstruction method and system based on SVD-DCS, so as to improve the density and reliability of sparse array SAR three-dimensional reconstruction.

[0006] The technical solution adopted by the present invention to solve the above technical problems is: a sparse array tomographic SAR three-dimensional reconstruction method based on SVD-DCS, comprising the following steps:

[0007] S1: Acquire sparse array SAR images and preprocess them to obtain preprocessed tomographic SAR stack data;

[0008] S2: With the reference pixel as the center, within the defined window range, the average logarithmic coherence threshold and average amplitude constraint criteria are used to search and identify adjacent contour pixels pixel by pixel in the preprocessed tomographic SAR stack data;

[0009] S3: Adaptive weight design is performed based on the coherence coefficient index weight, and SVD decomposition is performed on the tomographic SAR data stack composed of adjacent equal-height pixels to reduce the DCS data processing dimension;

[0010] S4: Observe the SVD projection through MMV and estimate the noise energy at the reference pixel of the reduced-dimensional tomographic SAR data stack as the parameter of the SVD-DCS spectrum estimation model;

[0011] S5: Substitute the reduced-dimensional tomographic SAR data stack, estimate the spectrum matrix using the SVD-DCS spectrum estimation model, and reconstruct the reference pixel scattering intensity profile;

[0012] S6: Constructing a model order solution model with information theory constraints through the spectrum matrix to determine the order of the tomographic SAR point cloud model;

[0013] S7: Construct a target three-dimensional complex scattering coefficient optimization solution model through the spectrum matrix to estimate the scattering intensity value of the tomographic SAR point cloud;

[0014] S8: Geocode the tomographic SAR point cloud data in the range-Doppler-elevation coordinate system and output the spatial position and scattering intensity data of the tomographic SAR point cloud.

[0015] According to the above scheme, in step S1, the specific steps are:

[0016] S11: performing complex image registration on the input sparse array SAR image to obtain a SAR complex image;

[0017] S12: inputting reference terrain data, performing phase de-skewing processing on the SAR complex image to obtain a de-skewing SAR complex image;

[0018] S13: performing amplitude and phase error correction on the de-slanted SAR complex image to obtain pre-processed tomographic SAR stack data.

[0019] According to the above scheme, in step S2, the specific steps are:

[0020] S21: Let M be the number of frames of sparse array SAR image, x i (m) represents the complex value of the mth pre-processed tomographic SAR stack data at the reference pixel i, y i (m) represents the complex value of the pixel adjacent to the reference pixel i in the mth pre-processed tomographic SAR stack data. * Represents the conjugate operation of complex numbers; constructs a similarity measure index between the reference pixel x and the adjacent pixel y The distribution range is between (0,1]:

[0021]

[0022] S22: Take the negative logarithm of the similarity measure index between the reference pixel x and the adjacent pixel y, and the distribution range is between [0,∞):

[0023]

[0024] S23: Let N be the total number of pixels in the SAR image block; within the coherence calculation window, calculate the similarity measurement index between the reference pixel image block and the corresponding pixels of the adjacent pixel image block one by one, calculate the average similarity measurement index and use it as the similarity measurement index between the reference pixel image block and the adjacent pixel image block:

[0025]

[0026] S24: Calculate the average amplitude A of the tomographic SAR data at the reference pixel and the adjacent pixels x 、A y ;

[0027] S25: average amplitude A at the reference pixel xGreater than 0.5 times the average amplitude A of adjacent pixels y , average amplitude A at the reference pixel x Less than 2 times the average amplitude A of adjacent pixels y , and the similarity measure index D(x,y) with the reference pixel image block is less than -lnC thresh The central pixel of the adjacent pixel image block is identified as the adjacent equal-height pixel, where the coherence threshold C is set thresh =0.5;

[0028] S26: Traverse all pixels as reference pixels and find the adjacent pixels of the same height for all pixels according to steps S41 to S45.

[0029] Furthermore, in step S3, the specific steps are:

[0030] S31: Let K be the rank of the matrix, L be the number of adjacent pixels of equal height plus 1, S be the diagonal matrix composed of singular value elements arranged from large to small, K < L, K < M, U M×K 、V L×K The eigenvectors corresponding to the first K singular values ​​constitute the signal subspace, and the eigenvectors corresponding to the remaining singular values ​​constitute the noise subspace. The tomographic SAR data stack Y composed of adjacent equal-height pixels is decomposed into subspaces corresponding to the signal and noise by SVD, and only the signal subspace is retained:

[0031] Y=USV H =[U M×K ,U M×(M-K) ]S M×L [V L×K ,V L×(L-K) ] H (4);

[0032] S32: Order Among them I K×K is the K×K unit matrix, 0 (L-K)×K is a zero matrix of (LK)×L; the tomographic SAR data stack after dimensionality reduction is calculated as

[0033] Furthermore, in step S4, the specific steps are:

[0034] S41: Set represents the diagonal matrix of singular values, Represents the matrix composed of singular value vectors; the singular value decomposition of the compressed sensing matrix A is as follows:

[0035]

[0036] S42: Set represents the nth singular value vector, Represents the tomographic SAR data stack after dimensionality reduction The i-th column vector of is used as the energy of the noise at the reference pixel, which is used for the noise energy constraint of SVD-DCS spectrum estimation and the order determination of the tomographic SAR point cloud model. The formula for calculating the noise energy is as follows:

[0037]

[0038] Furthermore, in step S5, the specific steps are:

[0039] S51: Assume || || 2,1 、|| || F They represent the L1-L2 mixed norm and F norm of the matrix respectively; substitute the reduced-dimensional tomographic SAR data stack to construct the SVD-DCS spectrum estimation model:

[0040]

[0041] S52: Estimating the spectral matrix using the spectral gradient projection method

[0042] S53: Set Represents the estimated spectrum matrix The i-th row vector of the final reference pixel gives the scattering intensity result

[0043] Furthermore, in step S6, the specific steps are:

[0044] S61: Set is the number of valid targets at the reference pixel, z is the front The position of the maximum spectral peak, Before The three-dimensional scattering coefficient corresponding to the maximum spectrum peak position; construct the model order solution model constrained by information theory:

[0045]

[0046] The information constraint terms use the BIC and MDL models:

[0047]

[0048] S62: Based on the spectrum peaks from large to small, the cost function is substituted and calculated by exhaustive method, and the number of effective scattering targets and the elevation position corresponding to the target distribution with the minimum cost function are used as the effective target number at the reference pixel. and elevation position information.

[0049] Furthermore, in step S7, the specific steps are:

[0050] S71: Construct the target three-dimensional complex scattering coefficient optimization solution model:

[0051]

[0052] S72: Set express The i-th column vector of is solved by the least square method, and the scattering intensity value of the tomographic SAR point cloud is estimated as follows:

[0053]

[0054] A sparse array tomographic SAR 3D reconstruction system based on SVD-DCS includes an SLC image registration module, a phase de-skew module, an amplitude and phase error correction module, an adjacent contour pixel recognition module, an adaptive weight design module, a noise energy estimation module, an SVD-DCS spectrum estimation module, a model order selection module, a target parameter estimation module, and a geocoding module;

[0055] The SLC image registration module is used to acquire sparse array SAR images and perform complex image registration to obtain SAR complex images;

[0056] The phase de-skewing module is used to perform phase de-skewing on the SAR complex image according to the reference terrain data to obtain a de-skewing SAR complex image;

[0057] The amplitude and phase error correction module is used to perform amplitude and phase error correction on the de-slanted SAR complex image to obtain pre-processed tomographic SAR stack data;

[0058] The adjacent contour pixel recognition module is used to search and identify adjacent contour pixels pixel by pixel in the pre-processed tomographic SAR stack data within a defined window range, with the reference pixel as the center, and adopts the average logarithmic coherence threshold and average amplitude constraint method.

[0059] The adaptive weight design module is used to perform adaptive weight design based on the coherence coefficient index weight, and perform SVD decomposition on the tomographic SAR data stack composed of adjacent equal-height pixels to reduce the DCS data processing dimension;

[0060] The noise energy estimation module is used to estimate the noise energy at the reference pixel of the tomographic SAR data stack after dimensionality reduction by observing the SVD projection through MMV as the parameter of the SVD-DCS spectrum estimation model;

[0061] The SVD-DCS spectrum estimation module is used to estimate the spectrum matrix after substituting the reduced-dimensional tomographic SAR data stack and reconstruct the reference pixel scattering intensity profile;

[0062] The model order selection module is used to construct a model order solution model with information theory constraints through the spectrum matrix to determine the order of the tomographic SAR point cloud model;

[0063] The target parameter estimation module is used to construct a target three-dimensional complex scattering coefficient optimization solution model through the spectrum matrix and estimate the scattering intensity value of the tomographic SAR point cloud;

[0064] The geocoding module is used to geocode the tomographic SAR point cloud data in the range-Doppler-elevation coordinate system and output the spatial position and scattering intensity data of the tomographic SAR point cloud.

[0065] A computer memory stores a computer program executable by a computer processor, wherein the computer program executes a sparse array tomographic SAR three-dimensional reconstruction method based on SVD-DCS.

[0066] The beneficial effects of the present invention are:

[0067] 1. The present invention provides a sparse array tomographic SAR 3D reconstruction method and system based on SVD-DCS. This method addresses the problems of difficult to accurately reconstruct tomographic SAR and the high number of false targets under irregular and sparse sampling conditions. The method preprocesses the sparse array SAR image data, including complex image registration, phase deskew processing, and amplitude and phase error correction, to obtain preprocessed tomographic SAR stack data. The method reduces the DCS data processing dimension by identifying adjacent pixels of equal height and performing SVD decomposition. The scattering intensity profile of the reference pixel is reconstructed through noise energy estimation and SVD-DCS spectrum estimation. The method and system select the model order, estimate the target parameters, and geocode the data to output the spatial position and scattering intensity data of the tomographic SAR point cloud, thereby improving the density and reliability of sparse array SAR 3D reconstruction.

[0068] 2. Compared with traditional beamforming, Capon, and Music multi-view methods, as well as SVD-Wiener and SL1MMER single-view methods, the proposed method has the highest number of point cloud reconstructions and high 3D reconstruction density.

[0069] 3. Compared with the existing SVD-Wiener and SL1MMER super-resolution single-view methods, the method of the present invention suppresses sidelobe noise better and improves the reliability of tomographic SAR point cloud reconstruction through precise reconstruction. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 is a flow chart of an embodiment of the present invention.

[0071] Figure 2 This is a comparison of the sparse array tomographic SAR 3D reconstruction results using various methods;

[0072] in:

[0073] Figure 2 (a) is the 3D reconstruction result of sparse array tomographic SAR using the beamforming method;

[0074] Figure 2 (b) is the result of sparse array tomographic SAR 3D reconstruction using the Capon method;

[0075] Figure 2 (c) is the result of sparse array tomographic SAR 3D reconstruction using the Music method;

[0076] Figure 2 (d) is the result of 3D reconstruction of sparse array tomographic SAR using the single-view SVD-Wiener method of Dehang Aerospace;

[0077] Figure 2 (e) is the 3D reconstruction result of sparse array tomographic SAR using the SL1MMER method;

[0078] Figure 2 (f) is a diagram showing the sparse array tomographic SAR 3D reconstruction results using the SVD-DCS method of this embodiment.

[0079] Figure 3 This is a locally enlarged image comparing the tomographic SAR 3D reconstruction results of various methods.

[0080] in:

[0081] Figure 3 (a) is a partial enlarged view of the tomographic SAR 3D reconstruction result of the German Aerospace single-view SVD-Wiener method;

[0082] Figure 3 (b) is a local enlarged image of the tomographic SAR 3D reconstruction result using the SL1MMER method; Figure 3 (c) is a partially enlarged view of the tomographic SAR 3D reconstruction result of the SVD-DCS method of this embodiment. DETAILED DESCRIPTION

[0083] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0084] Example 1

[0085] See also Figure 1 The embodiment of the present invention includes three major data processing processes: data preprocessing, spectrum estimation, and tomographic 3D point cloud generation, which specifically include the following steps:

[0086] Step 1: Acquire sparse array SAR images and preprocess them to obtain preprocessed tomographic SAR stack data;

[0087] Step 1.1: Acquire sparse array SAR image data and perform complex image registration on the sparse array SAR image data;

[0088] Step 1.2: Input the reference terrain data and perform phase de-skew processing on the registered SAR complex image;

[0089] Step 1.3: Perform amplitude and phase error correction on the de-slanted SAR complex image to obtain pre-processed tomographic SAR stack data;

[0090] Step 2: With the reference pixel as the center, search and identify adjacent pixels of equal height within the defined window range pixel by pixel. The method used to identify adjacent pixels of equal height is the method of average logarithmic coherence threshold and average amplitude constraint criterion. The specific steps are as follows:

[0091] Step 2.1: Construct a similarity measure index between the reference pixel x and the adjacent pixel y

[0092]

[0093] Its distribution range is between (0,1], where M is the number of frames of sparse array SAR images, x is i (m) represents the complex value of the mth pre-processed tomographic SAR stack data at the reference pixel i, y i (m) represents the complex value of the pixel adjacent to the reference pixel i in the mth pre-processed tomographic SAR stack data. * Represents the conjugation operation of a complex number.

[0094] Step 2.2: Take the negative logarithm of the similarity measure index between the reference pixel x and the adjacent pixel y:

[0095]

[0096] Its distribution range is between [0,∞).

[0097] Step 2.3: Within the coherence calculation window, calculate the similarity measurement index between the reference pixel image block and the corresponding pixels of the adjacent pixel image block one by one, calculate the average similarity measurement index, and use it as the similarity measurement index between the reference pixel image block and its adjacent pixel image blocks:

[0098]

[0099] Where N is the total number of pixels in the SAR image block.

[0100] Step 2.4: Calculate the average amplitude A of the tomographic SAR data at the reference pixel and its adjacent pixels x 、A y .

[0101] Step 2.5: Average the amplitude A at the reference pixel x Greater than 0.5 times the average amplitude A of adjacent pixels y , and the average amplitude A at the reference pixel x Less than 2 times the average amplitude A of adjacent pixels y , and the similarity measurement index D(x,y) with the reference pixel image block is less than -lnC thresh The central pixel of the adjacent pixel image block is identified as the adjacent equal-height pixel, where the coherence threshold C is set thresh =0.5.

[0102] Step 2.6: Traverse all pixels as reference pixels and find the adjacent pixels of the same height according to steps 4.1 to 4.5.

[0103] Step 3: Perform SVD decomposition on the tomographic SAR data stack composed of adjacent equal-height pixels to reduce the DCS data processing dimension;

[0104] Step 3.1: Decompose the tomographic SAR data stack Y consisting of adjacent equal-height pixels into subspaces corresponding to signals and noises by SVD and only retain the signal subspace, i.e.

[0105] Y=USV H =[U M×K ,U M×(M-K) ]S M×L [V L×K ,V L×(L-K) ] H (4)

[0106] Where K is the rank of the matrix, L is the number of adjacent pixels of equal height plus 1, S is a diagonal matrix composed of singular value elements arranged from large to small, K<L, K<M, U M×K 、V L×K It means that the eigenvectors corresponding to the first K singular values ​​constitute the signal subspace, and the eigenvectors corresponding to the remaining singular values ​​constitute the noise subspace.

[0107] Step 3.2: Make Among them I K×K is the K×K unit matrix, 0 (L-K)×K is a zero matrix of (LK)×L; the tomographic SAR data stack after dimensionality reduction is calculated as

[0108] Step 4: Estimate the noise energy of the SVD-DCS spectrum estimation model;

[0109] Step 4.1: Perform singular value decomposition on the compressed sensing matrix A as follows:

[0110]

[0111] in represents the diagonal matrix of singular values, Represents the matrix of singular value vectors.

[0112] Step 4.2: The formula for calculating noise energy is as follows:

[0113]

[0114] in represents the nth singular value vector, Represents the tomographic SAR data stack after dimensionality reduction The i-th column vector of is used as the energy of the noise at the reference pixel, which is used for the noise energy constraint of SVD-DCS spectrum estimation and the order determination of the tomographic SAR point cloud model.

[0115] Step 5: Substitute the reduced-dimensional tomographic SAR data stack and use the SVD-DCS spectrum estimation model to reconstruct the reference pixel scattering intensity profile;

[0116] Step 5.1: Substitute the reduced-dimensional tomographic SAR data stack and construct the SVD-DCS spectrum estimation model:

[0117]

[0118] where || || 2,1 、|| || F They represent the L1-L2 mixed norm and F norm of the matrix respectively.

[0119] Step 5.2: Use the spectral gradient projection method to estimate the spectrum matrix

[0120] Step 5.3: Give the final scattering intensity result at the reference pixel in Represents the estimated spectrum matrix The i-th row vector of .

[0121] Step 6: Determine the order of the tomographic SAR point cloud model;

[0122] Step 6.1: Construct the model order solution model with information theory constraints:

[0123]

[0124] in is the number of valid targets at the reference pixel, z is the front The position of the maximum spectral peak, Before The three-dimensional scattering coefficient corresponding to the maximum spectrum peak position, and the information constraint term uses the BIC and MDL models:

[0125]

[0126] Step 6.2: Based on the spectrum peaks from large to small, use the exhaustive method to substitute and calculate the cost function, and use the number of effective scattering targets and elevation positions corresponding to the target distribution with the minimum cost function as the effective number of targets at the reference pixel. and elevation position information.

[0127] Step 7: Estimate the scattering intensity value of the tomographic SAR point cloud;

[0128] Step 7.1: Construct the target three-dimensional complex scattering coefficient optimization solution model:

[0129]

[0130] Step 7.2: Solve the above equation by the least squares method and estimate the scattering intensity value of the tomographic SAR point cloud using the following equation:

[0131]

[0132] in express The i-th column vector of .

[0133] Step 8: Geocode the tomographic SAR point cloud data in the range-Doppler-elevation coordinate system and output the spatial position and scattering intensity data of the tomographic SAR point cloud.

[0134] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0135] Example 2

[0136] This embodiment uses the method of embodiment 1 to process 12-channel X-band SAR data from an airborne array in an urban area of ​​a certain city. The data baseline is horizontally distributed along the cross-track direction, and the bandwidth is 810 MHz.

[0137] Figure 2The 3D reconstruction results of several classic tomographic SAR methods were compared, including traditional multi-look beamforming, Capon, and Music; De Aerospace's single-look SVD-Wiener and SL1MMER; and the SVD-DCS method of this example. The results show that the single-look SVD-Wiener and SL1MMER methods have better point cloud reconstruction coverage, but suffer from greater sidelobe noise. The multi-look beamforming, Capon, and Music methods have better sidelobe noise suppression capabilities, but suffer from poor point cloud reconstruction coverage.

[0138] Figure 3 for Figure 2 From the enlarged view of the local area of ​​the results of the last three methods, it can be seen that in the case of sparse array, the SVD-DCS method of this embodiment has better monitoring sidelobe noise suppression capability, point cloud coverage and density.

[0139] Table 1 shows the statistical results of the number of point cloud recovery rates of several methods. The benchmark number of point cloud recovery is 1860755. It can be seen that the method of this embodiment has the highest recovery rate, even better than the benchmark level.

[0140] Table 1 Number of point clouds reconstructed by tomographic SAR using different methods

[0141] method Number of point clouds Beamforming 929,267 Capon 863,326 Music 900,179 SVD-Wiener 1,569,377 SL1MMER 1,652,245 DCS 2,153,021

[0142] The above embodiments are intended only to illustrate the design concepts and features of the present invention. Their purpose is to enable those skilled in the art to understand the contents of the present invention and implement them accordingly. The scope of protection of the present invention is not limited to the above embodiments. Therefore, any equivalent changes or modifications made based on the principles and design concepts disclosed in the present invention are within the scope of protection of the present invention.

Claims

1. A sparse array tomographic SAR 3D reconstruction method based on SVD-DCS, characterized by: The following steps are involved: S1: Acquire sparse array SAR images and preprocess them to obtain preprocessed tomographic SAR stack data; S2: With the reference pixel as the center, within the defined window range, the average logarithmic coherence threshold and average amplitude constraint criteria are used to search and identify adjacent contour pixels pixel by pixel in the preprocessed tomographic SAR stack data; S3: Adaptive weight design is performed based on the coherence coefficient index weight, and SVD decomposition is performed on the tomographic SAR data stack composed of adjacent equal-height pixels to reduce the DCS data processing dimension; S4: Observe the SVD projection through MMV and estimate the noise energy at the reference pixel of the reduced-dimensional tomographic SAR data stack as the parameter of the SVD-DCS spectrum estimation model; the specific steps are: S41: Set represents the diagonal matrix of singular values, 、 Represents the matrix composed of singular value vectors; for compressed sensing matrix Perform singular value decomposition as follows: （5）； S42: Set is the number of frames of sparse array SAR image, Indicates the singular value vectors, Represents the tomographic SAR data stack after dimensionality reduction No. Column vector, calculate the noise energy at the reference pixel: （6）； S5: Substitute the reduced-dimensional tomographic SAR data stack, estimate the spectrum matrix using the SVD-DCS spectrum estimation model, and reconstruct the reference pixel scattering intensity profile; the specific steps are: S51: Set 、 They represent the L1-L2 mixed norm and F norm of the matrix respectively; substitute the reduced-dimensional tomographic SAR data stack to construct the SVD-DCS spectrum estimation model: , s.t. (7); S52: Estimating the spectral matrix using the spectral gradient projection method ; S53: Set Represents the estimated spectrum matrix No. Row vector; gives the final scattering intensity result at the reference pixel ; S6: Construct a model order solution model constrained by information theory through the spectrum matrix to determine the order of the tomographic SAR point cloud model; the specific steps are: S61: Set is the number of valid targets at the reference pixel, For the front The position of the maximum spectral peak, Before The three-dimensional scattering coefficient corresponding to the maximum spectrum peak position; construct the model order solution model constrained by information theory: （8）； The information constraint items adopt BIC and MDL models: （9）； S62: Based on the spectrum peaks from large to small, the cost function is substituted and calculated by exhaustive method. The number of effective scattering targets and the elevation position corresponding to the target distribution with the minimum cost function are used as the effective target number at the reference pixel. and elevation position information; S7: Construct a target three-dimensional complex scattering coefficient optimization solution model through the spectrum matrix to estimate the scattering intensity value of the tomographic SAR point cloud; the specific steps are: S71: Construct the target three-dimensional complex scattering coefficient optimization solution model: （10）； S72: Set express No. Column vector; solve formula (10) by the least squares method and estimate the scattering intensity value of the tomographic SAR point cloud as: （11）； S8: Geocode the tomographic SAR point cloud data in the range-Doppler-elevation coordinate system and output the spatial position and scattering intensity data of the tomographic SAR point cloud.

2. The sparse array tomographic SAR 3D reconstruction method based on SVD-DCS according to claim 1, characterized in that: In the step S1, the specific steps are: S11: Acquire a sparse array SAR image and perform complex image registration to obtain a SAR complex image; S12: performing phase de-skewing processing on the SAR complex image according to the reference terrain data to obtain a de-skewing SAR complex image; S13: performing amplitude and phase error correction on the de-slanted SAR complex image to obtain pre-processed tomographic SAR stack data.

3. The sparse array tomographic SAR 3D reconstruction method based on SVD-DCS according to claim 1, characterized in that: In the step S2, the specific steps are: S21: Set Indicates the The pre-processed tomographic SAR stack data is placed in the reference pixel The complex value of Indicates the The pre-processed tomographic SAR stack data and the reference pixel The complex value of the adjacent pixel, Represents the conjugate operation of complex numbers; constructs reference pixels With adjacent pixels Similarity measure index , the distribution range is between: （1）； S22: For reference pixels With adjacent pixels The similarity metric index takes the negative logarithm operation and the distribution range is between: （2）； S23: Set is the total number of pixels in the SAR image block; In the coherence calculation window, the similarity measurement index between the corresponding pixels of the reference pixel image block and the adjacent pixel image block is calculated one by one, and the average similarity measurement index is calculated and used as the similarity measurement index between the reference pixel image block and the adjacent pixel image block: （3）； S24: Calculate the average amplitude of the tomographic SAR data at the reference pixel and adjacent pixels 、 ; S25: average amplitude at the reference pixel Greater than 0.5 times the average amplitude of adjacent pixels , average amplitude at the reference pixel Less than 2 times the average amplitude of adjacent pixels , and the similarity measure index with the reference pixel image block Less than The central pixel of the adjacent pixel image block is identified as the adjacent equal-height pixel, where the coherence threshold is set ; S26: Traverse all pixels as reference pixels and find the adjacent pixels of the same height for all pixels according to steps S41 to S45.

4. The sparse array tomographic SAR 3D reconstruction method based on SVD-DCS according to claim 3, characterized in that: In the step S3, the specific steps are: S31: Set is the rank of the matrix, The number of adjacent pixels of the same height plus 1, is a diagonal matrix consisting of singular value elements arranged from large to small, 、 , 、 Before The signal subspace is composed of the eigenvectors corresponding to the singular values, and the eigenvectors corresponding to the other singular values ​​are the noise subspace. The tomographic SAR data stack composed of adjacent equal-height pixels is constructed by SVD. Decompose into subspaces corresponding to signal and noise, and only retain the signal subspace: （4）； S32: Order ,in for The unit array, for The zero matrix of the tomographic SAR data stack after dimensionality reduction is calculated as .

5. A sparse array tomographic SAR 3D reconstruction system based on SVD-DCS, characterized by: The system is used to implement the method described in any one of claims 1 to 4; It includes SLC image registration module, phase de-skew module, amplitude and phase error correction module, adjacent contour pixel recognition module, adaptive weight design module, noise energy estimation module, SVD-DCS spectrum estimation module, model order selection module, target parameter estimation module and geocoding module; The SLC image registration module is used to acquire sparse array SAR images and perform complex image registration to obtain SAR complex images; The phase de-skewing module is used to perform phase de-skewing on the SAR complex image according to the reference terrain data to obtain a de-skewing SAR complex image; The amplitude and phase error correction module is used to perform amplitude and phase error correction on the de-slanted SAR complex image to obtain pre-processed tomographic SAR stack data; The adjacent contour pixel recognition module is used to search and identify adjacent contour pixels pixel by pixel in the pre-processed tomographic SAR stack data within a defined window range, with the reference pixel as the center, and adopts the average logarithmic coherence threshold and average amplitude constraint method. The adaptive weight design module is used to perform adaptive weight design based on the coherence coefficient index weight, and perform SVD decomposition on the tomographic SAR data stack composed of adjacent equal-height pixels to reduce the DCS data processing dimension; The noise energy estimation module is used to estimate the noise energy at the reference pixel of the tomographic SAR data stack after dimensionality reduction by observing the SVD projection through MMV as the parameter of the SVD-DCS spectrum estimation model; The SVD-DCS spectrum estimation module is used to estimate the spectrum matrix after substituting the reduced-dimensional tomographic SAR data stack and reconstruct the reference pixel scattering intensity profile; The model order selection module is used to construct a model order solution model with information theory constraints through the spectrum matrix to determine the order of the tomographic SAR point cloud model; The target parameter estimation module is used to construct a target three-dimensional complex scattering coefficient optimization solution model through the spectrum matrix and estimate the scattering intensity value of the tomographic SAR point cloud; The geocoding module is used to geocode the tomographic SAR point cloud data in the range-Doppler-elevation coordinate system and output the spatial position and scattering intensity data of the tomographic SAR point cloud.

6. A computer memory, characterized in that: A computer program executable by a computer processor is stored therein, and the computer program executes the SVD-DCS-based sparse array tomography SAR three-dimensional reconstruction method according to any one of claims 1 to 4.

Images