Three-dimensional inversion method of urban tomographic SAR with multiple types of scatterers

Abstract

The application relates to a kind of urban tomography SAR three-dimensional inversion methods of fusing multiple types of scatterers.The method comprises: selecting PS candidate points based on amplitude deviation index, dividing sub-regions, constructing reference network using the strategy of minimum spanning tree of undirected graph, and obtaining the absolute height of reference network PS through ridge estimation adjustment.Based on this, selecting star network PS candidate points using amplitude threshold, matching adjacent reference network single PS point, and judging and estimating the absolute height of star network PS through sequential self-cancellation generalized likelihood ratio method.Again, selecting statistical homogeneous pixels using confidence interval hypothesis testing, screening out DS candidate points by setting SHP threshold, optimizing its time sequence phase, matching adjacent single PS point for each DS candidate point, selecting the highest connection edge of imaging quality to estimate the absolute height of DS.Finally, fuse three types of height and geocode, and output urban multi-type scatterer height three-dimensional point cloud.Using the method, the three-dimensional information of urban buildings can be accurately obtained.

Description

Technical Field

[0001] This application relates to the field of satellite remote sensing imaging technology, and in particular to a three-dimensional inversion method for urban tomographic SAR that integrates multiple types of scatterers. Background Technology

[0002] In urban scenarios, targets of interest and capable of receiving effective backscattering signals are mainly categorized into permanent scatterers (PS) and distributed scatterers (DS). PS are point-like scattering targets, primarily including buildings, exposed rocks, bridges, power transmission towers, streetlights, and other targets with strong and stable backscattering signals. Within a radar image resolution cell, there is only one primary scatterer (in the absence of overlay). The scattering echoes from other scatterers have minimal impact on the primary scatterer, exhibiting high phase stability. If the scattering echo of a scatterer is represented by a vector in the complex plane, the electromagnetic wave received by the radar can be considered as a composite vector formed by connecting all vectors end-to-end. DS are area-like scattering targets, such as bare soil, grass, and road surfaces, which are non-primary scattering targets with weak backscattering signals. Within a SAR image resolution cell, there is no primary scatterer, but it typically contains a large number of scatterers. Their scattering echo signals interact and exhibit a "random walk" state in complex space. The dashed arrows represent the echo signals of each sub-scatterer, while the solid arrows represent the final superimposed echo signals.

[0003] Different MT-InSAR processing methods have emerged for PS and DS targets. Methods for retrieving PS targets primarily use PSI, employing a single primary image and interfering with other images as secondary images. Methods for retrieving DS targets include SBAS, SqueeSAR, TCP-InSAR, and QPS. Some utilize multi-primary image strategies oriented towards shorter spatiotemporal baselines, while others employ multi-look averaging strategies, all aimed at stably capturing observational information of DS targets. These methods assume that at most one effective scatterer exists within each pixel. However, urban areas are characterized by numerous high-rise buildings, and satellite side-look imaging is prone to overlay phenomena. Multiple scatterers at different heights often exist within the same pixel. Existing methods cannot separate the information of overlaying scatterers, making it difficult to accurately obtain the three-dimensional information of urban buildings and structures, severely limiting the accurate estimation of urban geophysical parameters. Summary of the Invention

[0004] Therefore, it is necessary to provide a three-dimensional inversion method for urban tomographic SAR that can accurately acquire three-dimensional information of urban buildings and structures by fusing multiple types of scatterers.

[0005] A three-dimensional inversion method for urban tomographic SAR that integrates multiple types of scatterers, the method comprising: Acquire single-view composite image dataset and digital elevation model data; use the single-view composite image dataset and digital elevation model data to perform registration, deskewing and radiometric correction operations in sequence to obtain tomographic dataset; Based on the amplitude deviation index, candidate PS points are selected from the tomographic dataset. Sub-regions are divided according to a distance threshold. Within each sub-region, PS connection edges that meet the distance threshold are searched using the undirected graph minimum spanning tree strategy and tomographically imaged. Single PS connection edges with imaging quality reaching the threshold are retained, and duplicate edges are removed. The undirected graph minimum spanning tree reference network with the highest imaging quality is selected globally. Then, the minimum spanning tree reference network is adjusted according to the ridge estimation method to obtain the absolute elevation of the reference network PS. Based on the reference network of the minimum spanning tree of the undirected graph and the absolute elevation of the reference network PS, candidate PS points of the star network are selected from the remaining pixels by using an amplitude threshold. For each candidate point, the single PS point of the reference network with the highest imaging quality in the neighborhood is matched and connected to construct the star network. The sequential self-cancellation generalized likelihood ratio method is used to determine whether it is a single PS point or a double PS point and to estimate the absolute elevation of the star network PS. Based on the undirected graph minimum spanning tree reference network and star network, a hypothesis testing method based on confidence intervals is used to select statistically homogeneous pixels from the remaining pixels; a threshold for the number of SHPs is set to screen out candidate DS points; the interferogram of the candidate DS points is optimized by temporal phase optimization using a phase connection method based on eigenvalue decomposition; for each optimized candidate DS point, a connection edge is established by finding a single PS point in the neighborhood that meets the distance threshold, and the connection edge with the highest imaging quality is selected to estimate the absolute elevation of the DS point; The absolute elevations of the reference network PS, the star network PS, and the DS points are fused and geocoded to output a 3D point cloud of elevations of various types of scatterers in the urban area.

[0006] The aforementioned method for 3D inversion of urban tomographic SAR using multiple scatterer types optimizes the minimum spanning tree reference network construction process by introducing the concept of undirected graphs. After dividing the region into sub-regions based on distance thresholds, it simultaneously searches for PS connection edges and filters effective edges, globally selecting the reference network with the highest imaging quality. This significantly shortens the network construction time for densely packed PS points in urban areas, resulting in a marked efficiency improvement compared to the traditional point-by-point traversal method, meeting the rapid processing needs of large-scale urban scenes. Regarding elevation inversion accuracy, ridge estimation replaces the weighted least squares method to improve the robustness of network adjustment. Addressing signal interference issues caused by the complex urban environment, ridge estimation uses regularization parameters determined by the inflection point of the L-curve to reduce the condition number of the coefficient matrix, minimizing error propagation and making the absolute elevation estimation of the reference network PS more stable. Simultaneously, the star-shaped network employs a sequential self-cancellation generalized likelihood ratio method to accurately identify single / double PS points and invert their elevations separately, breaking the assumption of a single pixel and single scatterer. This significantly reduces elevation inversion errors in overlapping areas, providing high-precision data support for obtaining detailed parameters of buildings and structures. Finally, the fusion of PS and DS multi-type scatterer targets forms a complementary advantage. The first two layers of the network, composed of PS, ensure high-precision inversion of point-like strong scattering targets, while the third layer of the network, composed of DS, covers the area that PS cannot reach through confidence interval hypothesis testing and phase optimization. The output three-dimensional point cloud of elevation of various types of scatterers in urban areas has both high precision and high coverage, comprehensively reflecting the height information of urban buildings and structures. It provides reliable remote sensing data support for scenarios such as urban planning and deformation monitoring, and effectively solves the technical bottlenecks of traditional methods in urban overlay processing and the balance between efficiency and coverage. Attached Figure Description

[0007] Figure 1 This is a flowchart illustrating a three-dimensional inversion method for urban tomographic SAR that integrates multiple types of scatterers, as described in one embodiment. Figure 2 This is a schematic diagram of the framework of a three-dimensional inversion method for urban tomographic SAR that integrates multiple types of scatterers in one embodiment; Figure 3 This is a schematic diagram illustrating the solution of regularization parameters in one embodiment; Figure 3 (a) is a schematic diagram of the L-curve. Figure 3 (b) is a schematic diagram of the slope of the L-curve; Figure 4 This is a schematic diagram showing the location of the experimental test area and the distribution of facilities in one embodiment; Figure 5 This is a schematic diagram illustrating the results of selecting homogeneous pixels in one embodiment; Figure 6 This is a schematic diagram of the three-layer network connection in one embodiment; Figure 7 This is the elevation inversion result of the experimental test area in one embodiment; Figure 7(a) Elevation inversion results for reference network; Figure 7 (b) The elevation inversion results of the double-layer network; Figure 7 (c) is the result of the three-layer network elevation inversion. Detailed Implementation

[0008] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0009] In one embodiment, such as Figure 1 As shown, a three-dimensional inversion method for urban tomographic SAR that integrates multiple types of scatterers is provided, including the following steps: Step 102: Obtain the single-view composite image dataset and digital elevation model data; use the single-view composite image dataset and digital elevation model data to perform registration, deslant removal and radiometric correction operations in sequence to obtain the tomographic dataset.

[0010] like Figure 2 The diagram illustrates the framework of the inversion method. Here, PSC represents PS candidate points, i.e., permanent scatterer candidate points; MSTRN represents a multi-scale temporal regression network, used to construct a temporal model to estimate the relative parameters of the scatterer; SPS represents a stable permanent scatterer; DPS represents a distributed permanent scatterer; and SHP represents statistically homogeneous pixels, referring to pixels with consistent statistical characteristics, often used for phase optimization of distributed scatterers. The inversion process includes: acquiring single-look complex (SLC) images and digital elevation model (DEM) data. The SLC dataset is obtained by multiple observations of the same target at different altitudes with similar incident angles by a multi-pass spaceborne SAR satellite. DEM data can be downloaded from publicly available 30-meter or 90-meter resolution DEM data for the study area. Data preprocessing mainly includes registration, deskewing, and radiometric correction. Registration aims to ensure that the same pixel coordinates in different images correspond to the same target feature; registration accuracy needs to reach the sub-pixel level, and this can be accomplished using different software. Deskewing facilitates height-axis focusing; the angular frequency support domain in the height direction needs to be corrected to near zero frequency. This can be done using slant range deskewing or topographic reference deskewing. Radiometric correction corrects the reflectance amplitude of images observed at different times to the same reference. Completing these two steps yields a tomographic dataset suitable for tomographic processing.

[0011] Step 104: Select PS candidate points from the tomographic dataset based on amplitude deviation index, divide the data into sub-regions according to distance threshold, search for PS connection edges that meet the distance threshold in each sub-region using the undirected graph minimum spanning tree strategy and perform tomographic imaging, retain single PS connection edges with imaging quality reaching the threshold and remove duplicate edges, and globally select the undirected graph minimum spanning tree reference network with the highest imaging quality; then adjust the minimum spanning tree reference network according to the ridge estimation method to obtain the absolute elevation of the reference network PS.

[0012] The first layer (reference network) requires initial identification of PS candidate points using the amplitude deviation index. The global image is then divided into equal-area sub-regions, the specific size of which can be adjusted based on the actual image size and the number of PS candidate points. Each sub-region is then simultaneously searched for and imaged using connecting edges. For each PS candidate point, other PS candidate points are searched within a set distance threshold (500 meters in this embodiment), and connected to them. Beamforming is used for tomography to obtain the image quality. If the quality is less than the image quality threshold (0.7 in this embodiment), the point is discarded; otherwise, it is retained. After imaging of the neighboring connecting edges of all PS candidate points is complete, redundant connecting edges are deleted. Then, based on the image quality weight of each connecting edge, a minimum spanning tree reference network is searched globally and constructed. By dividing the image into sub-regions according to the distance threshold and searching for PS connecting edges in parallel within each sub-region, redundant calculations of global point-by-point traversal are avoided. Combined with the selection rules of retaining high-quality edges and removing duplicate edges, the reference network construction time is significantly shortened, making it particularly suitable for urban scenes with dense PS points, solving the problem of excessively long network construction time in traditional methods. The minimum spanning tree with the highest imaging quality is selected globally as the reference network to ensure that the network structure is based on high-quality connecting edges, reduce the interference of noisy edges on subsequent adjustment, and provide a stable benchmark framework for the entire inversion process.

[0013] Next, the ridge estimation method is used to perform network adjustment on the reference network: ; in, for Absolute estimation of parameter vector, Represents a string consisting of 0, -1, and 1. Transformation matrix, Indicates the connecting edge Relative parameter estimate vector, This represents the image quality composed of connected edges. Weight matrix, This is the regularization parameter. The L-curve can be obtained by solving for the inflection points of the L-curve, which is formed by... and The curve represents the inflection point of the solution and the equilibrium point for minimizing the residuals. The value represents the optimal regularization parameter for this network configuration. To address the interference from scattered signals caused by the complex urban environment, a regularization parameter is introduced to reduce the condition number of the coefficient matrix, avoiding the instability of solutions in ill-conditioned matrices under weighted least squares, reducing the risk of error propagation, and making the absolute elevation estimation of the reference network PS more reliable. The regularization parameter is determined by the inflection point of the L-curve, finding an optimal balance between the solution norm and the residual norm, ensuring both the accuracy of elevation estimation and suppressing the influence of outliers, making it particularly suitable for establishing high-precision benchmarks in complex urban scattering environments. Figure 3 The diagram shown illustrates the solution for the regularization parameter. Figure 3 (a) is a schematic diagram of the L-curve; Figure 3 (b) is a schematic diagram of the slope of the L-curve; the red circles represent the inflection points of the L-curve, and the corresponding regularization parameters are... .

[0014] After network adjustment, the absolute elevation information of all reference networks (PS) relative to a certain elevation reference point can be obtained.

[0015] Step 106: Based on the minimum spanning tree reference network of the undirected graph and the absolute elevation of the reference network PS, select candidate PS points of the star network from the remaining pixels by using the amplitude threshold. For each candidate point, connect the reference network single PS point with the highest imaging quality in the neighborhood to construct the star network. Then, use the sequential self-cancellation generalized likelihood ratio method to determine whether it is a single PS point or a double PS point and estimate the absolute elevation of the star network PS.

[0016] First, candidate PSs are found among the remaining pixels excluding the reference PS using an average amplitude threshold. Then, for each candidate PS, a connecting edge is established by searching for a matching reference PS within a distance threshold. Beamforming is then used for tomographic imaging, and the sequential self-cancellation generalized likelihood ratio method is used for single / double PS verification. The connecting edge with the highest imaging quality is selected, and the absolute elevation estimate of the single / double PS of the star network is calculated based on the reference PS of that connecting edge.

[0017] Based on the reference network, candidate PS points with slightly poor stability are filtered by amplitude thresholding to supplement building details (such as building edges and ancillary structures) not covered by the main reference network, thereby improving the spatial coverage of point scatterers. A sequential self-cancellation generalized likelihood ratio method is employed to directly identify single PS points or double PS points within a single pixel and estimate their elevations separately. This breaks the traditional assumption of a single pixel as a single scatterer, accurately separating multi-height scatterer information caused by overlapping high-rise buildings in urban areas and resolving the elevation confusion problem in overlapping areas.

[0018] Step 108: Based on the undirected graph minimum spanning tree reference network and star network, a hypothesis testing method based on confidence intervals is used to select statistically homogeneous pixels from the remaining pixels; a threshold for the number of SHPs is set to screen out candidate DS points; the interferogram of the candidate DS points is optimized by temporal phase optimization using a phase connection method based on eigenvalue decomposition; for each optimized candidate DS point, a connection edge is established by finding a single PS point in the neighborhood that meets the distance threshold, and the connection edge with the highest imaging quality is selected to estimate the absolute elevation of the DS point.

[0019] Statistical homogeneous pixel identification is performed on the remaining pixels excluding the reference mesh and star mesh PS, and a hypothesis testing method based on confidence intervals is used for discrimination: ; in, Represents probability. Number of samples, sample mean The temporal mean of the neighboring pixels to be estimated replace, Estimated from reference pixel replace, The confidence level is set at a certain threshold. If the temporal average intensity value of a neighboring pixel to be estimated falls within the confidence interval at the given confidence level, it is considered a statistically homogeneous pixel of the reference pixel; otherwise, it is not.

[0020] Based on the statistically homogeneous number of pixels for each pixel, a threshold needs to be set to remove some non-DS points. Here, the SHP threshold is set as the number of SHP points corresponding to 95% of the cumulative distribution function of the existing PS points' SHP, i.e.: ; Where CDF is the cumulative distribution function, and according to the SHP threshold, among the remaining non-PS individuals excluding PS, the SHP is higher than [value missing]. The points that are selected are candidate points for DS.

[0021] Next is DS phase optimization, which involves optimizing the time series observations of a DS target. It is generally assumed to follow a complex circular Gaussian distribution: ; in, Represents the covariance matrix. For determinant operators, This represents the conjugate transpose.

[0022] because The true value is difficult to know, so the sample covariance matrix is ​​generally used as a substitute: ; in, This indicates the number of multiple views, i.e., the number of homogeneous pixels. The average coherence matrix of the samples. Represents the average phase matrix of the sample. This represents the Hadamard product.

[0023] because The SLC dataset contains poor-quality information due to decorrelation, therefore it is necessary to process the phase vectors... The average phase matrix of the samples is then used for further optimization. This can be obtained by solving the maximum likelihood function. ; in, Indicates taking the real part, To solve this nonlinear optimization problem, the eigenvalue decomposition isophase connection algorithm can be used, and the optimized time-series phase obtained is used to replace the original observation phase.

[0024] By using confidence interval hypothesis testing to screen for statistically homogeneous pixels, and determining DS candidate points based on the SHP distribution threshold of PS points, planar scatterers are included in the inversion range, significantly increasing the density of monitoring points in urban areas and covering areas such as roads and grasslands that PS cannot reach. The temporal phase of DS is optimized based on eigenvalue decomposition, eliminating inferior information caused by coherence. Elevation is estimated by combining the connection edges with stable PS points, allowing DS inversion to rely on the high-precision benchmark of PS while leveraging its planar coverage advantage, thus overcoming the coverage deficiency of traditional methods that rely solely on PS.

[0025] Step 110: The absolute elevations of the reference network PS, the star network PS, and the absolute elevations of the DS points are fused and geocoded to output a three-dimensional point cloud of elevations of various types of scatterers in the urban area.

[0026] By fusing elevation data from point targets (PS) and surface targets (DS), the high accuracy of PS on strong scatterers is preserved, while the details of surface areas are supplemented by DS, forming a complete three-dimensional distribution of scatterers in urban areas. Based on a two-layer network oriented towards PS, DS is added as a third layer connecting to the first two layers. This effectively utilizes inversion information based on stable targets from PS, while also progressively densifying the network structure and enriching the information on monitored targets.

[0027] Geocoded 3D point clouds fully reflect the height information of buildings and structures, providing comprehensive data for urban planning, deformation monitoring, disaster assessment, etc., and solving the problem of limited application caused by incomplete coverage or insufficient accuracy of traditional methods.

[0028] The aforementioned method for 3D inversion of urban tomographic SAR using multiple scatterer types optimizes the minimum spanning tree reference network construction process by introducing the concept of undirected graphs. After dividing the region into sub-regions based on distance thresholds, it simultaneously searches for PS connection edges and filters effective edges, globally selecting the reference network with the highest imaging quality. This significantly shortens the network construction time for densely packed PS points in urban areas, resulting in a marked efficiency improvement compared to the traditional point-by-point traversal method, meeting the rapid processing needs of large-scale urban scenes. Regarding elevation inversion accuracy, ridge estimation replaces the weighted least squares method to improve the robustness of network adjustment. Addressing signal interference issues caused by the complex urban environment, ridge estimation uses regularization parameters determined by the inflection point of the L-curve to reduce the condition number of the coefficient matrix, minimizing error propagation and making the absolute elevation estimation of the reference network PS more stable. Simultaneously, the star-shaped network employs a sequential self-cancellation generalized likelihood ratio method to accurately identify single / double PS points and invert their elevations separately, breaking the assumption of a single pixel and single scatterer. This significantly reduces elevation inversion errors in overlapping areas, providing high-precision data support for obtaining detailed parameters of buildings and structures. Finally, the fusion of PS and DS multi-type scatterer targets forms a complementary advantage. The first two layers of the network, composed of PS, ensure high-precision inversion of point-like strong scattering targets, while the third layer of the network, composed of DS, covers the area that PS cannot reach through confidence interval hypothesis testing and phase optimization. The output three-dimensional point cloud of elevation of various types of scatterers in urban areas has both high precision and high coverage, comprehensively reflecting the height information of urban buildings and structures. It provides reliable remote sensing data support for scenarios such as urban planning and deformation monitoring, and effectively solves the technical bottlenecks of traditional methods in urban overlay processing and the balance between efficiency and coverage.

[0029] In one embodiment, adjusting the reference network for the minimum spanning tree of an undirected graph according to the ridge estimation method includes: The process of adjusting the reference network for the minimum spanning tree of an undirected graph using the ridge estimation method is as follows: ; in, for Absolute estimation of parameter vector, Represents a string consisting of 0, -1, and 1. Transformation matrix, Represents the connecting edge Relative parameter estimate vector, This represents the image quality composed of connected edges. Weight matrix, This is the regularization parameter.

[0030] In one embodiment, based on the absolute elevation of the undirected graph minimum spanning tree reference network and the reference network PS, candidate PS points for the star network are selected from the remaining pixels using an amplitude threshold. For each candidate point, a single PS point of the reference network with the highest imaging quality in its neighborhood is connected to construct the star network. The sequential self-cancellation generalized likelihood ratio method is used to determine whether a single PS point or a double PS point is a given, and the absolute elevation of the star network PS is estimated. This includes: Using the average amplitude threshold, candidate PS points are found in the remaining pixels excluding the reference PS. For each candidate PS point, a connection edge is established by searching for a suitable reference PS within the distance threshold. Beamforming is used for tomography and sequential self-cancellation generalized likelihood ratio is used for single / double PS testing. Select the connecting edge with the highest imaging quality, and calculate the absolute elevation estimate of the star network single / double PS based on the reference network PS of the connecting edge.

[0031] In one embodiment, the SHP quantity threshold is set to...

[0032] Where CDF is the cumulative distribution function. This represents the number of homogeneous pixels for existing PS points.

[0033] In one embodiment, a threshold for the number of SHPs is set to filter out candidate DS points, including: Based on the SHP quantity threshold, among the remaining non-PS (excluding PS), the SHP is higher than... The points that are selected are candidate points for DS.

[0034] In one embodiment, the single-view complex image dataset represents a dataset obtained by observing the same urban target from multiple-pass spaceborne synthetic aperture radar satellites at similar incident angles and different altitudes; temporal phase optimization of the interferogram of DS candidate points is performed using a phase connection method based on eigenvalue decomposition, including: The probability density function is calculated based on the time series observations of the DS candidate points; the maximum likelihood optimization objective function is established to optimize the sample covariance matrix in the probability density function; the eigenvalue decomposition algorithm is used to solve the maximum likelihood optimization objective function to obtain the optimized time series phase.

[0035] In one embodiment, the probability density function is calculated based on the temporal observations of the DS candidate points, including: Based on the time-series observations of the DS candidate points The probability density function of the time series observations is ; in, Represents the covariance matrix. For determinant operators, This represents the conjugate transpose.

[0036] In one embodiment, a maximum likelihood optimization objective function is established. ; in, Indicates taking the real part, , Represents conjugate transpose. The phase vector of the single-view complex image dataset, The sample covariance matrix, This is the average coherence matrix of the samples.

[0037] In one embodiment, the sample covariance matrix is ; in, This indicates the number of multiple views, i.e., the number of homogeneous pixels. The average coherence matrix of the samples. The representative sample average phase matrix is ​​composed of the phase vectors from the single-view multi-image dataset. This represents the Hadamard product.

[0038] In a specific embodiment, such as Figure 4 The diagram shown is a schematic representation of the location of the experimental testing area and the distribution of facilities.

[0039] Figure 5 To analyze the results of homogeneous pixel selection, the number of SHPs (Structured Hierarchical Images) for buildings primarily targeting PS (Physical Pixels), including the experimental test area, was relatively small, while the number of SHPs for infrastructure primarily targeting DS (Device Pixels), including the runway in the experimental test area, was relatively large. This indicates that the selection of homogeneous pixels is consistent with reality.

[0040] Figure 6 This is a three-layer network connection, where the red dot is the PS point of the first layer network, the green dot is the PS point of the second layer network, the blue dot is the DS point of the third layer network, and the connecting edge is the yellow line segment.

[0041] Figure 7 The elevation inversion results for the experimental test area are as follows: Figure 7 (a) Reference website; Figure 7 (b) Double-layer network; Figure 7 (c) Three-layer network. Figure 7 (a) shows the elevation estimates of the reference network, and the elevations of the main structures of buildings, including the terminal and satellite hall, are well inverted. Figure 7(b) represents the results of the double-layer network inversion. Based on the first-layer network, the elevation of the PS points on the building with slightly lower stability is estimated, enriching their detailed information. Figure 7 (c) shows the elevation estimation results of the three-layer network. After adding DS points, the coverage of the scattering points is significantly improved. The number of DS points in the third-layer network is 5.4 times that of the number of PS points in the two-layer network. This illustrates the necessity of DS targets in urban monitoring scenarios.

[0042] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0043] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0044] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for three-dimensional inversion of urban tomographic SAR by fusing multiple types of scatterers, characterized in that, The method includes: Acquire single-view composite image dataset and digital elevation model data; use the single-view composite image dataset and digital elevation model data to perform registration, deskewing and radiometric correction operations in sequence to obtain tomographic dataset; Based on the amplitude deviation index, candidate PS points are selected from the tomographic dataset. Sub-regions are divided according to a distance threshold. Within each sub-region, PS connection edges that meet the distance threshold are searched using the undirected graph minimum spanning tree strategy and tomographically imaged. Single PS connection edges that meet the imaging quality threshold are retained, and duplicate edges are removed. The undirected graph minimum spanning tree reference network with the highest imaging quality is selected globally. Then, the minimum spanning tree reference network is adjusted according to the ridge estimation method to obtain the absolute elevation of the reference network PS. Based on the absolute elevation of the undirected graph minimum spanning tree reference network and the reference network PS, candidate PS points of the star network are selected from the remaining pixels by using an amplitude threshold. For each candidate point, a single PS point of the reference network with the highest imaging quality in the neighborhood is matched and connected to construct the star network. The sequential self-cancellation generalized likelihood ratio method is used to determine whether it is a single PS point or a double PS point and to estimate the absolute elevation of the star network PS. Based on the undirected graph minimum spanning tree reference network and star network, a hypothesis testing method based on confidence intervals is used to select statistically homogeneous pixels from the remaining pixels; a threshold for the number of SHPs is set to screen out candidate DS points; the interferogram of the candidate DS points is optimized by temporal phase optimization using a phase connection method based on eigenvalue decomposition; for each optimized candidate DS point, a connection edge is established by finding a single PS point in the neighborhood that meets the distance threshold, and the connection edge with the highest imaging quality is selected to estimate the absolute elevation of the DS point; The absolute elevations of the reference network PS, the star network PS, and the DS points are fused and geocoded to output a 3D point cloud of elevations of various types of scatterers in the urban area.

2. The method according to claim 1, characterized in that, Adjustment of the minimum spanning tree reference network based on the ridge estimation method includes: The process of adjusting the minimum spanning tree reference network according to the ridge estimation method is as follows: ； in, for Absolute estimation of parameter vector, Represents a string consisting of 0, -1, and 1. Transformation matrix, Represents the connecting edge Relative parameter estimate vector, This represents the image quality composed of connected edges. Weight matrix, This is the regularization parameter.

3. The method according to claim 1, characterized in that, Based on the absolute elevation of the undirected graph minimum spanning tree reference network and the reference network PS, candidate PS points for the star network are selected from the remaining pixels using an amplitude threshold. For each candidate point, a single PS point of the reference network with the highest imaging quality in its neighborhood is connected to construct the star network. The sequential self-cancellation generalized likelihood ratio method is used to determine whether a single PS point or a double PS point is used, and the absolute elevation of the star network PS is estimated, including: Using the average amplitude threshold, candidate PS points are found in the remaining pixels excluding the reference PS. For each candidate PS point, a connection edge is established by searching for a suitable reference PS within the distance threshold. Beamforming is used for tomography and sequential self-cancellation generalized likelihood ratio is used for single / double PS testing. Select the connecting edge with the highest imaging quality, and calculate the absolute elevation estimate of the star network single / double PS based on the reference network PS of the connecting edge.

4. The method according to claim 1, characterized in that, A hypothesis testing method based on confidence intervals is used to select statistically homogeneous pixels from the remaining pixels, including: Statistical homogeneous pixel identification is performed on the remaining pixels excluding the reference mesh and star mesh PS, and a hypothesis testing method based on confidence intervals is used for discrimination: ； in, Represents probability. Number of samples, sample mean The temporal mean of the neighboring pixels to be estimated replace, Estimated by reference pixel replace, Confidence level; If the temporal average intensity value of the neighboring pixel to be estimated is within the confidence interval at a given confidence level, it is determined to be a statistically homogeneous pixel of the reference pixel; otherwise, it is not.

5. The method according to claim 1, characterized in that, Set the SHP quantity threshold to Where CDF is the cumulative distribution function. This represents the number of homogeneous pixels for existing PS points.

6. The method according to claim 5, characterized in that, Set a threshold for the number of SHPs to filter out candidate DS points, including: Based on the SHP quantity threshold, among the remaining non-PS (excluding PS), the SHP is higher than... The points that are selected are candidate points for DS.

7. The method according to claim 1, characterized in that, The single-view complex image dataset represents a dataset obtained by observing the same urban target from multiple-pass spaceborne synthetic aperture radar satellites at similar incident angles and different altitudes; temporal phase optimization of the interferograms of DS candidate points is performed using a phase connection method based on eigenvalue decomposition, including: The probability density function is calculated based on the time series observations of the DS candidate points; the maximum likelihood optimization objective function is established to optimize the sample covariance matrix in the probability density function; the eigenvalue decomposition algorithm is used to solve the maximum likelihood optimization objective function to obtain the optimized time series phase.

8. The method according to claim 7, characterized in that, The probability density function is calculated based on the time-series observations of the DS candidate points, including: Based on the time-series observations of the DS candidate points The probability density function of the time series observations is in, Represents the covariance matrix. For determinant operators, This represents the conjugate transpose.

9. The method according to claim 7, characterized in that, Establish the maximum likelihood optimization objective function in, Indicates taking the real part, , Represents conjugate transpose. The phase vector of the single-view complex image dataset, The sample covariance matrix, This is the average coherence matrix of the samples.

10. The method according to claim 8, characterized in that, The sample covariance matrix is: ； in, This indicates the number of views, or the number of homogeneous pixels. The average coherence matrix of the samples. The representative sample average phase matrix is ​​composed of the phase vectors from the single-view multi-image dataset. This represents the Hadamard product.

Images