A method for fusing ground-based multi-source deformation monitoring data based on virtual target point coding

Abstract

This invention relates to the field of geographic mapping and discloses a method for fusing space-based multi-source deformation monitoring data based on virtual target coding. The method includes: acquiring raw complex domain echo data from space-based SAR and point cloud data from local sensors; extracting the complex scattering coefficients, phase center positions, and phase center change rates of coherent scattering points from the SAR complex domain data; extracting Gaussian curvature extrema and their local normal vector change rates from the point cloud data; generating virtual target codes independent of physical reflectors through complex wavelet transform coding; establishing a constraint to maximize mutual information between the complex domain and spatial domain; using the virtual target codes as a fixed benchmark, iteratively updating the spatial domain transformation matrix through gradient descent to achieve cross-scale registration; and generating a unified deformation field based on the registration results. This invention eliminates the need for physical targets and utilizes the SAR phase center change rate to provide subwavelength-level constraints, achieving high-precision cross-modal fusion of space-based macroscopic deformation and local microscopic deformation.

Description

Technical Field

[0001] This invention relates to the field of geographic surveying and mapping, and in particular to a method for fusing multi-source deformation monitoring data of the sky and ground based on virtual target point coding. Background Technology

[0002] Synthetic Aperture Radar (SAR) and multi-source sensors such as ground-based lidar (LiDAR) or millimeter-wave radar each have their advantages in deformation monitoring: Space-based SAR, using satellite platforms as carriers, can acquire large-scale, all-weather surface deformation information through interferometric (InSAR) technology, with phase measurement accuracy reaching millimeter or even sub-millimeter levels. However, limited by side-view imaging geometry, it can only acquire one-dimensional deformation components along the radar line of sight (LOS), and its spatial resolution is relatively low (typically meter- to ten-meter levels). Local sensors (such as ground-based lidar, millimeter-wave radar, or visual sensors) can acquire high-resolution three-dimensional point cloud data with spatial positioning accuracy down to the centimeter level, capable of finely characterizing local geometric deformations. However, their coverage is limited, and they are insensitive to minute deformations at the subwavelength level. Fusion of space-based SAR data with local point cloud data to achieve unified deformation monitoring across scales and modes is an important research direction in the fields of surveying and mapping remote sensing and geological disaster monitoring. However, due to the fundamental differences in the imaging mechanisms, spatial scales, coordinate references, and information dimensions of the two types of data, how to achieve high-precision registration and fusion without increasing hardware costs (such as deploying physical corner reflectors or ground control points) remains a technical challenge that urgently needs to be solved in this field.

[0003] To address the aforementioned issues, several deformation monitoring methods based on multi-source data fusion have been proposed in the prior art. For example, Chinese patent application CN121170189A discloses a method for updating the terrain of metal mining areas based on lidar. This method constructs a probabilistic grid model of terrain change by fusing satellite SAR remote sensing imagery, UAV orthophotos, and mining operation plans, thereby guiding lidar to perform targeted scanning of high-probability change areas. While this method achieves adaptive detection of change areas, it relies on the mining operation plan as prior information, and the SAR data is only used for change detection rather than for direct registration with point cloud data. Chinese patent application CN110058237A discloses an InSAR point cloud fusion and 3D deformation monitoring method for high-resolution SAR imagery. This method generates point clouds from multi-platform SAR data using distributed scatterer InSAR technology and then uses an iterative nearest-point algorithm for registration and fusion. However, this method handles the same-modal point cloud registration between SAR data (i.e., point clouds derived from one SAR and point clouds derived from another SAR), not cross-modal registration between SAR complex domain data and locally measured point cloud data. Furthermore, its registration accuracy depends on external DEM data and ground control points. In addition, Chinese patent application CN120259131A discloses a high-precision lidar mapping method based on multi-source data fusion. This method uses a deep learning network to fuse point cloud, image, and inertial navigation data, but it still relies on ground control points for geometric correction and does not involve the processing of SAR complex domain data. The common shortcomings of existing technologies are as follows: First, most rely on physical reflectors, ground control points, or external DEM data as registration benchmarks, which increases deployment costs and system complexity, making them particularly difficult to implement in vast mining areas or complex terrain regions. Second, the registration of SAR data and point cloud data is usually performed independently in the spatial domain, failing to fully utilize the subwavelength-level physical information such as the phase center position and phase center change rate contained in the SAR complex domain echo signal, resulting in registration accuracy being limited by the spatial resolution of the point cloud data itself. Third, there is a lack of a unified cross-scale registration framework between the macroscopic deformation of space-based SAR and the local microscopic deformation, leading to a prominent problem of error accumulation. Summary of the Invention

[0004] The technical problem to be solved by this invention is to achieve high-precision cross-modal and cross-scale registration and fusion of space-based SAR complex domain data and local sensor point cloud data without relying on physical reflectors, ground control points or external DEMs and other hardware facilities.

[0005] Specifically, this invention provides a method for fusing multi-source deformation monitoring data of the space-ground plane based on virtual target point encoding, the method comprising the following steps: S1. Acquire raw complex domain echo data of space-based SAR imagery and point cloud data from local sensors; wherein, the raw complex domain echo data is an SLC complex image obtained by focusing raw SAR echo data, which contains both real and imaginary part information; the point cloud data does not contain physical quantities corresponding to the phase information of the complex domain echo data. S2. Extract the complex scattering coefficients, phase center positions, and phase center change rates of coherent scattering points from the original complex domain echo data of the space-based SAR image to construct a first complex feature set; extract the Gaussian curvature extrema points and their local normal vector change rates corresponding to the spatial positions from the point cloud data of the local sensor to construct a second complex feature set; encode the first and second complex feature sets into the same complex domain feature space through complex wavelet transform to generate virtual target point codes that do not depend on physical reflector hardware; S3. Establish constraints for maximizing mutual information between the complex domain and the spatial domain, and define cross-modal mutual information. ,in Represents information entropy. The set of complex-domain feature vectors encoding virtual target points. The set of spatial coordinates of the registered point cloud; with the goal of maximizing the mutual information, the parameters of the spatial domain transformation matrix are iteratively updated by gradient descent, wherein the complex domain virtual target encoding serves as a fixed reference to guide the optimization direction of the spatial domain parameters, thereby realizing cross-scale iterative registration between space-based SAR imagery and local point cloud data; S4. Based on the registration results, the macroscopic deformation data corresponding to the space-based SAR image and the microscopic deformation data corresponding to the local point cloud data are mapped to the same physical coordinate reference to generate unified deformation field data and output it.

[0006] A multi-source heterogeneous monitoring data fusion system, the system comprising: The data acquisition module is used to acquire raw complex domain echo data of space-based SAR images and point cloud data of local sensors; wherein, the raw complex domain echo data is a complex format baseband signal without imaging processing, and contains both real and imaginary part information; the point cloud data does not contain physical quantities corresponding to the phase information of the complex domain echo data; A virtual target generation module, connected to the data acquisition module, is used to extract the complex scattering coefficients, phase center positions, and phase center change rates of coherent scattering points from the original complex domain echo data of the space-based SAR image, constructing a first complex feature set; extracting the Gaussian curvature extrema points and their local normal vector change rates from the point cloud data of the local sensor, constructing a second complex feature set; and encoding the first and second complex feature sets into the same complex domain feature space through complex wavelet transform to generate virtual target code that does not depend on physical reflector hardware. The cross-scale registration module, connected to the virtual target generation module, is used to establish a constraint for maximizing mutual information between the complex domain and the spatial domain, and to define cross-modal mutual information. ,in Information entropy is represented; with maximizing the mutual information as the optimization objective, the spatial domain transformation matrix parameters are iteratively updated using the gradient descent method, wherein the complex domain virtual target point encoding serves as a fixed reference benchmark to guide the optimization direction of the spatial domain parameters, thereby achieving cross-scale iterative registration between space-based SAR images and local point cloud data; The unified deformation field output module is connected to the cross-scale registration module. It is used to map the macroscopic deformation data corresponding to the space-based SAR image and the microscopic deformation data corresponding to the local point cloud data to the same physical coordinate reference based on the registration results, generate unified deformation field data and output it.

[0007] The beneficial effects of this invention are as follows: First, by extracting the phase center position and rate of change of natural coherent scatterers from the raw complex domain echo data of SAR, a virtual target encoding is constructed, completely replacing the function of traditional physical corner reflectors or ground control points, achieving cross-modal benchmark unification with zero hardware cost. Second, by utilizing the constraint of maximizing mutual information between the complex domain and the spatial domain, the sensitivity of SAR phase information to subwavelength-level deformation is introduced into the point cloud registration process, enabling the registration accuracy to break through the limitations of the spatial resolution of the point cloud itself, reaching the sub-pixel level (i.e., subwavelength level). Third, through multi-scale decomposition and layer-by-layer registration strategies, the cross-scale registration problem between the macroscopic scale of space-based SAR and the microscopic scale of local sensors is effectively solved, avoiding error accumulation. Finally, through a closed-loop feedback mechanism, the generated unified deformation field is used to guide the data acquisition and virtual target encryption of the next cycle, forming an adaptive intelligent monitoring system. In summary, this invention significantly improves the registration accuracy and fusion efficiency of multi-source heterogeneous deformation monitoring data, and is particularly suitable for complex scenarios such as vast mining areas without ground control points, steep slopes, or geological hazard points where physical reflectors cannot be deployed. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0009] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0010] Before formally describing the present invention, a general description of the solution of the present invention will be given first to facilitate understanding.

[0011] Example 1: Please refer to Figure 1 The present invention provides a method for fusing space-ground multi-source deformation monitoring data based on virtual target point encoding, comprising the following steps: This invention provides a method for fusing space-ground multi-source deformation monitoring data based on virtual target point encoding. The method includes the following steps: S1. Acquire raw complex domain echo data of space-based SAR imagery and point cloud data from local sensors; wherein, the raw complex domain echo data is an SLC complex image obtained by focusing raw SAR echo data, which contains both real and imaginary part information; the point cloud data does not contain physical quantities corresponding to the phase information of the complex domain echo data. As one embodiment, step S1 of the present invention specifically involves: establishing a data link between a ground receiving station and a commercial SAR satellite (such as TerraSAR-X, Sentinel-1, or the domestically produced Gaofen-3 satellite) to receive raw complex domain echo data transmitted from the satellite in real time. This raw complex domain echo data is an unprocessed complex format baseband signal (raw data), with each sampling point containing real (I-channel) and imaginary (Q-channel) information. The data format is typically a signed 16-bit integer or a single-precision floating-point complex number. For example, for a metal mining area (geographical coordinates 118°15′E, 40°30′N, area approximately 50 square kilometers), receiving raw data from a single TerraSAR-X satellite strip mode with a 12-day reception period, a center frequency of 9.65GHz (X-band), a wavelength of approximately 3.1cm, a range sampling frequency of 150MHz, an azimuth pulse repetition frequency of 3000Hz, and a raw data file size of approximately 2GB. Simultaneously, local sensors deployed on the mine surface (such as a 16-line mechanical rotating lidar or a 77GHz millimeter-wave radar) collect local point cloud data at a fixed scanning frequency (e.g., 10Hz). The lidar point cloud data format includes three-dimensional coordinates (x, y, z, in meters) for each point, with a coordinate accuracy of ±2cm, and does not include any physical quantities (such as velocity, angular velocity, or phase value) corresponding to SAR phase information. To ensure spatiotemporal synchronization, a GPS timing module adds a timestamp accurate to the microsecond level to the local sensor data.

[0012] It should be noted that a preprocessing step is included between step S1 and step S2: T1. The original complex domain echo data of the space-based SAR image is encoded with range and azimuth compression, and the real and imaginary parts of the complex signal are completely preserved to obtain a complex domain single-view complex image; wherein, amplitude extraction is not performed during the encoding stage. Specifically, for the received raw SAR echo data (e.g., a 2GB binary file), range-directed matched filtering is first performed: a complex conjugate of the transmitted signal is constructed as a matched filter, and convolution is performed on each range gate to obtain range-compressed data. Then, azimuth-directed synthetic aperture processing is performed: a sub-aperture algorithm or the Omega-K algorithm is used to complete azimuth compression, resulting in a complex-domain single-look complex image (SLC). The value of each pixel in this SLC image is a complex number, for example, pixel... Corresponding complex values Throughout this process, the real part is always retained. and the virtual part No Amplitude extraction is not performed. Phase unwrapping ensures that subsequent virtual target encoding can utilize complete phase information.

[0013] T2. Perform voxel downsampling and statistical outlier filtering on the point cloud data of the local sensor to obtain a sparse point cloud, and calculate the local normal vector change rate of each point. Specifically, the raw LiDAR point cloud data is large (e.g., 1 million points per frame). Voxel filtering is used: the voxel side length is set. Divide the space into A grid was created, retaining the centroid coordinates of all points within each grid while discarding others, resulting in approximately 200,000 points. Then, a statistical outlier filter was applied: the average distance from each point to its k nearest neighbors (k=50) was calculated, assuming the distance distribution followed a Gaussian distribution, and points with an average distance greater than the global mean plus three standard deviations were removed. Finally, for each point, a quadratic surface was fitted using its neighborhood points. Calculate the normal vector Then, the rate of change of the normal vector is calculated using the variance of the normal vector in the neighborhood. .

[0014] T3. Based on satellite orbit parameters and local sensor timestamps, perform coarse temporal alignment between the complex domain single-view complex image and the sparse point cloud.

[0015] Specifically, the SAR imagery is acquired with the time accurate to the second (e.g., 06:23:45 UTC on March 1, 2025), while the local point cloud data includes a GPS timestamp (e.g., 06:23:45.123 UTC on March 1, 2025). Both are unified to the same time base (e.g., UTC). If the time difference is less than a threshold (e.g., 1 second), it is considered coarsely aligned. If the time difference exceeds the threshold, the SAR imagery is time-interpolated based on satellite orbital parameters (e.g., satellite velocity, position), or misaligned data frames are discarded.

[0016] S2. Extract the complex scattering coefficients, phase center positions, and phase center change rates of coherent scattering points from the original complex domain echo data of the space-based SAR image to construct a first complex feature set; extract the Gaussian curvature extrema points and their local normal vector change rates corresponding to the spatial positions from the point cloud data of the local sensor to construct a second complex feature set; encode the first and second complex feature sets into the same complex domain feature space through complex wavelet transform to generate virtual target point codes that do not depend on physical reflector hardware; As one embodiment, step S2 of the present invention specifically includes the following sub-steps: First, range compression (matched filtering) and azimuth compression (synthetic aperture processing) are performed on the raw SAR complex domain echo data to obtain a complex domain single-look complex image (SLC). In this SLC image, each pixel corresponds to a complex scattering coefficient. ,in For amplitude, For phase, pixels with a radar cross section (RCS) exceeding the background noise threshold by more than 15 dB are identified as coherent scattering points using a constant false alarm rate (CFAR) detection algorithm. For each coherent scattering point, its phase center position (sub-pixel level coordinates) is extracted through local peak interpolation, and the phase center position is determined using two consecutive image periods (time baseline). Calculation of the phase center change rate of the interference phase difference (days): ,in , These represent the interference phase values ​​(in radians) for the two phases, respectively. For example, in a region of exposed rock, the phase center change rate of the extracted coherent scattering points is... The corresponding surface deformation rate is It reaches the subwavelength level (better than) ).

[0017] Secondly, the local sensor point cloud data is processed: voxel grid filtering (voxel size 0.2m×0.2m×0.2m) is used for downsampling, and then the Gaussian curvature of each point is calculated. (in , (Main curvature), extracting Gaussian curvature whose absolute value is greater than a threshold The points are taken as extreme points; at the same time, the normal vector of each point is calculated through local surface fitting. And calculate the rate of change of the normal vector. (Rate of change along spatial distance).

[0018] Finally, complex wavelet transform (e.g., using the Morlet complex wavelet as the mother wavelet) is employed to encode the first complex feature set (a complex vector composed of complex scattering coefficients, phase center position, and phase center change rate) and the second complex feature set (a real vector composed of the coordinates of Gaussian curvature extrema points and their local normal change rates, mapped to a complex vector through a complex form) into the same complex domain feature space. Specifically, for each feature point, its complex wavelet coefficients are calculated. ,in For scale parameters, For translation parameters, The wavelet coefficients are combined into a high-dimensional complex feature vector, which serves as the encoding for virtual target points. This encoding relies entirely on natural scatterers (such as exposed rocks, building corners, etc.) and requires no physical corner reflectors or ground control points.

[0019] It should be noted that the calculation method for the phase center change rate in step S2 is as follows: First, identify the interferometric coherence coefficients in the raw complex domain echo data of space-based SAR. Higher than the preset threshold The scattering points are taken as effective coherent scattering points, where the interference coherence coefficient is defined as:

[0020] in This represents the complex scattering coefficient vector of the corresponding pixel in the two phases of complex domain echo data. The number of pixels within the window. The range of values ​​is

[0021] For coherent scattering points that satisfy the aforementioned interference coherence coefficient threshold, the rate of change of their phase center is calculated as follows:

[0022] in, and These represent the interference phase values ​​of the same coherent scattering point in the complex domain echo data from the two consecutive periods. The time baseline for the two periods of data is used; after the interferometric coherence coefficient is greater than the preset threshold and atmospheric correction is performed, the subwavelength level constraint is obtained based on the phase change rate to provide for the spatial domain point cloud transformation during the cross-scale iterative registration process.

[0023] As one embodiment, the specific calculation process of the phase center change rate in step S2 of the present invention is as follows: Taking a slope in a mining area as an example, two raw data sets were selected from TerraSAR-X satellites collected on March 1, 2025 and March 13, 2025, respectively, with the time baseline as an example. The two datasets were registered and interferometrically processed to obtain an interferometric phase map. A coherent scattering point formed by exposed rock on the slope was identified on the interferometric phase map, with the interferometric phase values ​​of the preceding and following periods being... , The rate of change of the phase center is then calculated as follows: Corresponding surface deformation rate , where wavelength ,have to This value achieves subwavelength level measurement accuracy. In the subsequent S3 registration process, each virtual target point is encoded with this subwavelength level deformation information, which improves the constraint accuracy of point cloud transformation from the centimeter level to the sub-millimeter level, the so-called "sub-pixel level constraint".

[0024] For example, considering another stationary region (such as the metal structure on the roof of a mining office building), where the phase difference between two consecutive periods is almost zero, the calculated rate of change of the phase center... This indicates that the point can be used as a stable virtual control point.

[0025] It should be noted that step S2, which generates virtual target code that does not depend on physical reflector hardware, specifically includes: S21. Identify natural coherent scatterers with radar cross sections exceeding a preset background noise threshold from the raw complex domain echo data of space-based SAR as candidate virtual control points. Perform interferometry on the same candidate virtual control point using SAR images at at least two different incident angles, and solve the following system of equations to inversely solve for its initial geographic coordinates or candidate spatial coordinates. :

[0026] in , These are the slant ranges of the scattering point in the two image phases, respectively. , For the corresponding satellite antenna phase center coordinates, , For atmospheric delay correction, scattering points with residuals less than a preset distance threshold and coherence coefficients higher than a preset threshold are selected as effective virtual control points, forming a spatially uniform virtual control point network. Specifically, taking a certain metal mining area as an example, two TerraSAR-X images with different incident angles were acquired: Image 1 is an ascending orbit with an incident angle of... Image 2 is a descending orbit, with an incident angle of... First, natural coherent scatterers with a radar cross section (RCS) exceeding the background noise by more than 20 dB were identified using CFAR detection. For example, an abandoned concrete mixing plant on the east side of the mining area had an RCS of 12.5 dB and a mean background noise of -9.8 dB, a difference of 22.3 dB, meeting the threshold condition and was selected as a candidate virtual control point. The slant range of this scattering point was read from the header files of the two image sets. , The phase center coordinates of the satellite antenna are respectively and All units are meters. The atmospheric delay correction term is estimated using external water vapor data or a GPS-based interpolation model, for example... , Substitute the values ​​into the system of equations and solve for the initial geographic coordinates or candidate spatial coordinates of the point. The solution is obtained using Newton's iterative method, with the initial value being the midpoint between the positions of the two satellites. The method converges after 5 iterations, yielding the solution. , , Calculate the residual: Substitute the solution into the left side of the equation and subtract the right side to obtain the result. less than The actual residual of 0.008 meters is greater than 0.0031 meters, which does not meet the requirements. Therefore, this point is eliminated. Another exposed rock mass is selected, with a residual of 0.0025 meters, which meets the requirements and is used as an effective virtual control point. Finally, about 50 effective virtual control points with even spatial distribution (interval greater than 100 meters) are selected in the mining area to form a virtual control point network.

[0027] S22. Extract the Gaussian curvature extremum points that match the spatial location of the virtual control point network from the point cloud data of the local sensor, and calculate the initial transformation parameters between the two. Specifically, using local LiDAR point cloud data, the Gaussian curvature of each point is calculated, and points with an absolute curvature value greater than 0.01 are extracted as candidate points. Then, the nearest Gaussian curvature extremum is searched within a 0.5-meter radius around the virtual control point network to form matching point pairs. For example, for virtual control point A (coordinates as shown above), an extremum point is found at a distance of 0.23 meters in the point cloud, with a Gaussian curvature of... normal vector Using all matching point pairs (at least 3 pairs, 15 pairs in this embodiment), the initial rotation matrix is ​​calculated using SVD decomposition. Translation vector This achieves coarse registration.

[0028] S23. Encode the virtual control point network and the Gaussian curvature extremum points into the same complex domain feature space through complex wavelet transform to generate virtual target point codes.

[0029] Specifically, the three-dimensional coordinates of the coarsely registered virtual control point network Coordinates of the corresponding point cloud extreme points Construct complex vectors respectively: For example, for the th Construct complex vectors for points Then, Morlet complex wavelets are used for multi-scale decomposition, yielding a set of wavelet coefficients for each point. These coefficients are concatenated into a 64-dimensional complex feature vector, which serves as the encoding for that virtual target point. The set of encodings for all virtual target points constitutes the virtual target point encoding library.

[0030] It should be noted that the specific process of extracting the Gaussian curvature extrema points and their local normal vector change rates from the point cloud data of the local sensor in step S2 to construct the second complex feature set includes: S221. Perform local surface fitting on the point cloud data: For each point cloud data point... Utilizing its neighborhood point set Fitting the quadratic surface using the least squares method And calculate the first fundamental form coefficients of the surface. Second fundamental form coefficient ; S222, Calculate the Gaussian curvature of each point cloud data point. And extract the values ​​of Gaussian curvature whose absolute values ​​are greater than a preset threshold. The point is taken as the extreme point of Gaussian curvature, denoted as . ; S223, For each Gaussian curvature extreme point To calculate the rate of change of its local normal vector: first, the normal vector is obtained by fitting the plane through the neighborhood point set. Then calculate the relationship between this point and all points in its neighborhood. The rate of change of the normal vector difference along the spatial distance:

[0031] in For point The set of neighborhood point indices; S224. The three-dimensional coordinates of the Gaussian curvature extremum point With the corresponding rate of change of the local normal vector Combination, by constructing complex forms Form features in the complex field and construct a second set of complex features.

[0032] S3. Establish constraints for maximizing mutual information between the complex domain and the spatial domain, and define cross-modal mutual information. ,in Represents information entropy. The set of complex-domain feature vectors encoding virtual target points. The set of spatial coordinates of the registered point cloud; with the goal of maximizing the mutual information, the parameters of the spatial domain transformation matrix are iteratively updated by gradient descent, wherein the complex domain virtual target encoding serves as a fixed reference to guide the optimization direction of the spatial domain parameters, thereby realizing cross-scale iterative registration between space-based SAR imagery and local point cloud data; As one embodiment, step S3 of the present invention is specifically implemented as follows: First, let the set of complex domain feature vectors in the virtual target encoding generated by S2 be denoted as... Each for 3D complex vectors (e.g.) The set of spatial coordinates of the local point cloud data after initial transformation is denoted as... each Mutual information is calculated using kernel density estimation: [Estimation is performed separately...] , and joint distribution For example, decomposing a complex eigenvector into its real and imaginary parts and concatenating them... The vector is a 3D real vector, estimated using a Gaussian kernel density, with the kernel width determined by the Silverman criterion. Initial transformation matrix. The results are given by the coarse registration of the virtual control point network (see claim 3), including the rotation matrix. (Three Euler angles) Translation vector .

[0033] The optimization objective is to maximize , equivalent to minimizing The gradient descent method is used, and the gradient of mutual information with respect to the transformation parameters is calculated in each iteration: ,in The update rule is as follows: Learning rate Initially set to The mutual information decreases to 0.9 of its original value every 100 iterations. The iteration terminates when the mutual information changes by less than 0.9% for 10 consecutive iterations. After approximately 500 iterations, the optimal transformation matrix is ​​obtained. The registration is then completed. During this registration process, the complex domain virtual target point encoding remains unchanged as a "fixed reference benchmark," while the spatial domain point cloud coordinates are continuously adjusted as the transformation matrix is ​​updated, thus enabling the complex domain information to guide the spatial domain registration.

[0034] It should be noted that in step S3, the mutual information is calculated. Specifically, it includes: S311, Encode the set of virtual target vectors in the complex field. Each complex vector in the vector is represented as a two-dimensional real vector. The probability density function is calculated using kernel density estimation. ; Specifically, assuming that it was generated by S2 Each virtual target point is encoded as a 64-dimensional complex vector. Separating the real and imaginary parts of each complex vector yields a 128-dimensional real vector. Gaussian kernel density estimation is used: ,in nuclear width Select via Scott's rules: We approximate it to be 1.0.

[0035] S312, Set the coordinates of the spatial domain point cloud After normalizing the three-dimensional coordinates, the probability density function is calculated using multivariate kernel density estimation. ; Specifically, after registration, the coordinate range of point cloud data may be very large (e.g., , , Normalization is required: first, subtract the minimum value from each coordinate, divide by the range, and map to... The interval is then determined. Multivariate Gaussian kernel density estimation is then used, with similar kernel width selection. This yields... .

[0036] S313. Establish candidate correspondences using initial geographic coordinates / projected coordinates, pair each complex domain vector with its nearest neighbor point cloud coordinates in the spatial domain, and then calculate the joint probability density function of the corresponding samples. ; Specifically, for each complex field eigenvector In the spatial domain point cloud, the nearest neighbor point is searched using a KD-tree. .form For paired samples Then, a two-dimensional joint kernel density estimation is used: ,in It is a Gaussian kernel.

[0037] S314, Substitute into the mutual information formula

[0038] Calculate mutual information; Specifically, the feature space of the complex field is discretized into The grid (each dimension) is discretized into spatial domain coordinates. While a grid could be used, the computational cost is high. Therefore, a sampling method is employed: 5000 points are randomly sampled in both the feature space and the spatial domain, and a histogram is used to approximate the probability distribution. These points are then substituted into the formula to calculate the mutual information. For example, the initial registration yields the following results: Bit.

[0039] S315. Using the mutual information as the optimization objective, calculate its partial derivatives with respect to the parameters of the spatial domain transformation matrix, and iteratively update the transformation matrix parameters using the gradient descent method.

[0040] Specifically, to maximize Calculate the gradient for the target. For the transformation parameters... The gradient is approximately equal to The Adam optimizer is used, with a learning rate of... First-order moment attenuation coefficient Second-order moment attenuation coefficient After 500 iterations, the mutual information increased to Bits, the transformation matrix converges.

[0041] in: The spatial domain transformation matrix Use one of the following forms: Rigid body transformation matrix: contains rotation parameters Translation parameters The total degrees of freedom are 6; Affine transformation matrix: contains linear transformation matrices Translation vector The total degrees of freedom are 12. Non-rigid transformation matrix: parameterized using thin plate spline interpolation function, with degrees of freedom three times the number of control points; Set of complex-domain feature vectors for virtual target encoding Dimensions ,in For the number of virtual target points, The original complex dimension of the complex eigenvectors; the set of point cloud coordinates in the spatial domain. Dimensions ,in The number of point cloud data points; In step S313, the sample pairing method uses nearest neighbor search: for each complex domain feature vector In the spatial domain point cloud, the point with the closest Euclidean distance is searched using a KD-tree. ,form For paired samples Meanwhile, for each spatial domain point Search for the nearest neighbor eigenvector in the set of eigenvectors in the complex field to form a bidirectional pairing to enhance the robustness of mutual information estimation; The iteration stopping condition in step S315 is that any of the following conditions are met: continuous The absolute value of the change in mutual information in each iteration is less than a preset threshold; In the gradient descent method, the gradient magnitude of all transformation parameters is less than a preset threshold. Reaching the maximum number of iterations .

[0042] It should be noted that before performing cross-scale iterative registration in step S3, a multi-scale decomposition step based on SAR imaging geometric priors is also included: S321. Obtain the spatial resolution of space-based SAR imagery. Satellite incident angle Spatial resolution of digital elevation model (DEM) data and local sensor point cloud data ; For example, in a certain mining area, the space-based SAR imagery is in TerraSAR-X high-resolution spotlight mode, with a spatial resolution of... The resolution of the local lidar point cloud after downsampling is [resolution value missing]. Satellite incident angle The DEM data is the result of pre-acquired drone photogrammetry, with a grid spacing of 0.5 meters, covering the entire mining area.

[0043] S322, Calculate the number of basic decomposition layers And extract the terrain slope angle of each pixel from the DEM data. ,when Add extra layers Total number of decomposition layers This DEM data is only used for slope priors in multi-scale decomposition and is not used as a registration control benchmark. Specifically, this involves calculating the number of basic decomposition layers. Extract the terrain slope angle of each pixel from the DEM: For example, there is a steep slope on the north slope of the mining area, and the slope angle calculated from the DEM... greater than ,therefore No additional layers are added; while another cliff area ,but Total number of decomposition layers For flat areas, No layers added.

[0044] S323, Decompose the space-based SAR imagery and local point cloud data into... The layer-scale spatial representation uses standard Gaussian pyramid downsampling for the azimuth direction of the SAR image and non-uniform downsampling for the range direction based on the projection relationship between slant range and ground distance, so that the registration error at different scales remains consistent under the slant range projection geometry. Registration is performed layer by layer from the coarsest scale to the finest scale, and the registration result of each layer is used as the initial transformation matrix for the next layer.

[0045] Specifically, for SAR images, construction A multi-layer Gaussian pyramid. In the azimuth direction (along the track), standard downsampling (image size halved per layer) is used. In the range direction (perpendicular to the track), due to the slant distance... Distance from the ground The relationship is ( (where is the angle of incidence), the projection distortion varies with the angle of incidence. Therefore, when downsampling at distance, the non-uniform sampling factor and This is inversely proportional to ensure consistent ground distance resolution at every scale. For example, if the original distance sampling interval is 0.25 meters, the ground distance resolution is... Meters; after downsampling at the first layer, the interval is 0.5 meters, with a ground distance resolution of approximately 0.94 meters, and so on. For point cloud data, voxel grid downsampling is used, with the voxel side length doubling layer by layer. From the coarsest scale ( Starting with the lowest resolution layer, the transformation matrix at the coarse scale is obtained using the S3 registration method, and then used as the next layer (i.e., the layer with the lowest resolution). The initial value of the registration is set at each layer and then refined layer by layer until the finest scale (layer 0). This "coarse-to-fine" strategy effectively avoids local minima and improves the robustness of the registration.

[0046] S4. Based on the registration results, the macroscopic deformation data corresponding to the space-based SAR image and the microscopic deformation data corresponding to the local point cloud data are mapped to the same physical coordinate reference to generate unified deformation field data and output it.

[0047] As one embodiment, step S4 of the present invention specifically involves: using the optimal transformation matrix obtained in S3. The deformation field of the space-based SAR imagery (e.g., a line-of-sight deformation rate map obtained through time-series InSAR technology, with a spatial resolution of 10m × 10m, covering the entire mining area) is resampled to the spatial resolution of the local point cloud data (e.g., 0.2m × 0.2m). Simultaneously, the deformation field of the local point cloud data (e.g., local deformation obtained through multi-period point cloud differential analysis, with an accuracy of ±2cm) is... The inverse transformation is mapped to the SAR coordinate system. Then, under a unified coordinate datum, adaptive weighted fusion is used: for each spatial grid, the fused deformation value... The weight , , and The standard deviations of deformation measurements for the two types of data are respectively (e.g. , The final result is a unified deformation field data, output in GeoTIFF format, with each pixel storing a three-dimensional deformation vector. The unit is millimeters. This deformation field is sent to the mining area monitoring terminal for slope stability assessment and early warning.

[0048] It should be noted that a closed-loop feedback step is included after step S4: S5. Input the unified deformation field generated in the current cycle into the time series prediction model to predict the deformation region distribution and deformation rate in the next monitoring cycle. For example, the current cycle (the first) The uniform deformation field generated (periodically) is a three-dimensional deformation vector field with a spatial resolution of 0.5 meters. A Long Short-Term Memory (LSTM) network is used as the temporal prediction model. The input is the deformation field sequence of the past 5 periods (each period is 12 days), and the output is the predicted deformation field for the next period. The model is trained using monitoring data from the past 12 months. For example, the predicted deformation field for the next period is... The deformation rate of the southeast corner slope of the periodic mining area will increase to (Currently 0.2 mm / day), while the deformation rate in the stable region of the northwest corner is still close to 0.

[0049] S6. Based on the prediction results, dynamically adjust the extraction strategy of virtual target point encoding in step S2 of the next cycle: increase the spatial density of virtual target points in areas where the predicted deformation rate is higher than the preset threshold, and reduce the density of virtual target points in areas where the predicted deformation rate is lower than the preset threshold. In this invention, a preset threshold is set. For areas with predicted deformation rates higher than 0.3 mm / day (such as the southeast corner slope), in step S2, the extraction density of virtual target point encoding is increased from the default of one per 100 square meters to one per 20 square meters; for areas with a rate lower than 0.1 mm / day, it is reduced to one per 500 square meters. Specifically, a sampling probability density function is generated based on the predicted deformation rate map, and then a Poisson disk sampling method is used to generate new virtual target point locations.

[0050] S7. Feed the adjusted virtual target encoding back to the space-based SAR data acquisition strategy to select alternative SAR image data sources with better incident angles or heading angles for the next cycle, and increase the data acquisition frequency in high-risk deformation areas; at the same time, feed the prediction results back to the local sensor to optimize the scanning path planning and sampling density distribution of the local sensor for the next cycle.

[0051] Finally, the adjusted virtual target point coding distribution map is fed back to the space-based SAR data acquisition strategy. For the high-risk area in the southeast corner, a request is made to the satellite operator to acquire spotlight imagery (0.5-meter resolution) of the area in the next cycle, replacing the original stripe pattern (3-meter resolution), and the revisit cycle is shortened from 12 days to 6 days. Simultaneously, the prediction results are sent to the local UAV lidar platform to replan the scanning path: the UAV will focus its flight path on the southeast corner slope, increasing the scanning density from 100 points / m² to 500 points / m²; for stable areas, the scanning frequency is reduced or skipped. Ultimately, in the [missing information - likely a specific phase or event]... The periodic collection of more accurate data has further improved the timeliness and accuracy of deformation monitoring.

[0052] It should be noted that the local sensor includes at least one of millimeter-wave radar, visual sensor, or lidar; the original complex domain echo data of the space-based SAR image is a complex format baseband signal directly transmitted by the SAR satellite; in the constraint of maximizing mutual information between the complex domain and the spatial domain, the complementarity of information between the two domains is guaranteed by the following physical relationship: the phase information of the SAR complex domain echo data is sensitive to subwavelength level deformation, and the spatial coordinates of the point cloud data are sensitive to geometric deformation at the centimeter level and above. The complementary characteristics of the two physical quantities at the deformation scale help to coordinately constrain the registration error at the two scales.

[0053] Specifically, in actual deployments, different types of local sensors can be selected based on monitoring needs and costs. For example: For high-precision monitoring of open-pit mine benches and slopes, a 16-line lidar (such as the Velodyne VLP-16) is used, acquiring 300,000 points per second with an accuracy of ±2cm. For large-scale, all-weather monitoring, a 77GHz millimeter-wave radar (such as the Texas Instruments AWR1843) is used, capable of penetrating rain and fog, but the point cloud is relatively sparse (approximately 1000 points per frame). For auxiliary texture information, an RGB-D vision sensor (such as the Intel RealSense D455) is used to acquire color images and depth maps. The method of this invention is compatible with point cloud data from any one or more of the above-mentioned sensors.

[0054] Regarding space-based SAR data, the TerraSAR-X, Sentinel-1A / B, COSMO-SkyMed, and Gaofen-3 satellites all provide raw complex domain echo data (Level 0 products). For example, the "raw data" product from the Gaofen-3 satellite, ordered through the China Center for Resources Satellite Data and Application, is in binary file format with auxiliary annotations, where each pixel represents the I and Q components as 16-bit signed integers.

[0055] Taking a mining slope as an example, space-based SAR phase information can detect slow deformations (such as creep) at the millimeter or even sub-millimeter level, but its spatial resolution (approximately 3m) is insufficient to locate specific cracks. Local lidar point clouds can provide three-dimensional geometric information with centimeter-level accuracy, identifying cracks as small as 5cm in width, but are insensitive to minute deformations (<1mm). The mutual information maximization constraint of this invention essentially seeks a spatial transformation that maximizes the statistical dependence between the spatial coordinate distribution of the point cloud and the SAR phase deformation distribution. For example, on a slope, the SAR phase shows that a 50-meter diameter area is sinking at a rate of 0.1mm / day, while the point cloud shows that the geometry of this area also undergoes corresponding micro-changes (although the noise of a single measurement is relatively large). By maximizing mutual information, the transformation parameters of the point cloud are adjusted to best match the trend of SAR phase changes, thereby simultaneously optimizing the registration accuracy at both scales.

[0056] Example 2: A multi-source heterogeneous monitoring data fusion system, the system comprising: The data acquisition module is used to acquire raw complex domain echo data of space-based SAR images and point cloud data of local sensors; wherein, the raw complex domain echo data is an SLC complex image obtained by focusing the raw complex echo data of SAR, and contains both real and imaginary part information; the point cloud data does not contain physical quantities corresponding to the phase information of the complex domain echo data. A virtual target generation module, connected to the data acquisition module, is used to extract the complex scattering coefficients, phase center positions, and phase center change rates of coherent scattering points from the original complex domain echo data of the space-based SAR image, constructing a first complex feature set; extracting the Gaussian curvature extrema points and their local normal vector change rates from the point cloud data of the local sensor, constructing a second complex feature set; and encoding the first and second complex feature sets into the same complex domain feature space through complex wavelet transform to generate virtual target code that does not depend on physical reflector hardware. The cross-scale registration module, connected to the virtual target generation module, is used to establish a constraint for maximizing mutual information between the complex domain and the spatial domain, and to define cross-modal mutual information. ,in Information entropy is represented; with maximizing the mutual information as the optimization objective, the spatial domain transformation matrix parameters are iteratively updated using the gradient descent method, wherein the complex domain virtual target point encoding serves as a fixed reference benchmark to guide the optimization direction of the spatial domain parameters, thereby achieving cross-scale iterative registration between space-based SAR images and local point cloud data; The unified deformation field output module is connected to the cross-scale registration module. It is used to map the macroscopic deformation data corresponding to the space-based SAR image and the microscopic deformation data corresponding to the local point cloud data to the same physical coordinate reference based on the registration results, generate unified deformation field data and output it.

[0057] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for data fusion of ground-based multi-source deformation monitoring based on virtual target point coding, characterized in that: Includes the following steps: S1. Acquire raw complex domain echo data of space-based SAR imagery and point cloud data from local sensors; wherein, the raw complex domain echo data is an SLC complex image obtained by focusing raw SAR echo data, which contains both real and imaginary part information; the point cloud data does not contain physical quantities corresponding to the phase information of the complex domain echo data. S2. Extract the complex scattering coefficients, phase center positions, and phase center change rates of coherent scattering points from the original complex domain echo data of the space-based SAR image to construct a first complex feature set; extract the Gaussian curvature extrema points and their local normal vector change rates corresponding to the spatial positions from the point cloud data of the local sensor to construct a second complex feature set; encode the first and second complex feature sets into the same complex domain feature space through complex wavelet transform to generate virtual target point codes that do not depend on physical reflector hardware; S3, establishing a complex domain-space domain mutual information maximization constraint, defining a cross-modal mutual information quantity wherein represents information entropy, is a complex domain feature vector set coded for a virtual target point, is a set of spatial coordinates of the registered point cloud; taking the maximization of the mutual information quantity as an optimization objective, iteratively updating the spatial domain transformation matrix parameters by a gradient descent method, wherein the complex domain virtual target point coding serves as a fixed reference benchmark to guide the optimization direction of the spatial domain parameters, thereby realizing the cross-scale iterative registration of the space-based SAR image and the ground point cloud data. S4. Based on the registration results, the macroscopic deformation data corresponding to the space-based SAR image and the microscopic deformation data corresponding to the local point cloud data are mapped to the same physical coordinate reference to generate unified deformation field data and output it.

2. The method of claim 1, wherein the method is a virtual target point coding based ground-based multi-source deformation monitoring data fusion method. The calculation method for the phase center change rate in step S2 is as follows: Identifying an interferometric coherence coefficient in space-based SAR raw complex domain echo data above a preset threshold as an effective coherent scattering point, wherein the interferometric coherence coefficient is defined as: in This represents the complex scattering coefficient vector of the corresponding pixel in the two phases of complex domain echo data. The number of pixels within the window. The range of values ​​is For coherent scattering points that satisfy the aforementioned interference coherence coefficient threshold, the rate of change of their phase center is calculated as follows: in, and These represent the interference phase values ​​of the same coherent scattering point in the complex domain echo data from the two consecutive periods. The time baseline for the two periods of data is used; after the interferometric coherence coefficient is greater than the preset threshold and atmospheric correction is performed, the subwavelength level constraint is obtained based on the phase change rate to provide for the spatial domain point cloud transformation during the cross-scale iterative registration process.

3. The method for fusing space-ground multi-source deformation monitoring data based on virtual target coding as described in claim 1, characterized in that: Step S2, which generates virtual target encoding that does not depend on physical reflector hardware, specifically includes: S21. Identify natural coherent scatterers with radar cross sections exceeding a preset background noise threshold from the raw complex domain echo data of space-based SAR as candidate virtual control points. Perform interferometry on the same candidate virtual control point using SAR images at at least two different incident angles, and solve the following system of equations to inversely solve for its initial geographic coordinates or candidate spatial coordinates. : in , These are the slant ranges of the scattering point in the two image phases, respectively. , For the corresponding satellite antenna phase center coordinates, , For atmospheric delay correction, scattering points with residuals less than a preset distance threshold and coherence coefficients higher than a preset threshold are selected as effective virtual control points, forming a spatially uniform virtual control point network. S22. Extract the Gaussian curvature extremum points that match the spatial location of the virtual control point network from the point cloud data of the local sensor, and calculate the initial transformation parameters between the two. S23. Encode the virtual control point network and the Gaussian curvature extremum points into the same complex domain feature space through complex wavelet transform to generate virtual target point codes.

4. The method for fusing space-ground multi-source deformation monitoring data based on virtual target coding as described in claim 1, characterized in that: In step S3, the mutual information is calculated. Specifically, it includes: S311, Encode the set of virtual target vectors in the complex field. Each complex vector in the vector is represented as a two-dimensional real vector. The probability density function is calculated using kernel density estimation. ; S312, Set the coordinates of the spatial domain point cloud After normalizing the three-dimensional coordinates, the probability density function is calculated using multivariate kernel density estimation. ; S313. Establish candidate correspondences using initial geographic coordinates / projected coordinates, pair each complex domain vector with its nearest neighbor point cloud coordinates in the spatial domain, and then calculate the joint probability density function of the corresponding samples. ; S314, Substitute into the mutual information formula Calculate mutual information; S315. Using the mutual information as the optimization objective, calculate its partial derivatives with respect to the parameters of the spatial domain transformation matrix, and iteratively update the transformation matrix parameters using the gradient descent method. in: The spatial domain transformation matrix Use one of the following forms: Rigid body transformation matrix: contains rotation parameters Translation parameters The total degrees of freedom are 6; Affine transformation matrix: contains linear transformation matrices Translation vector The total degrees of freedom are 12. Non-rigid transformation matrix: parameterized using thin plate spline interpolation function, with degrees of freedom three times the number of control points; Set of complex-domain feature vectors for virtual target encoding Dimensions ,in For the number of virtual target points, The original complex dimension of the complex eigenvectors; the set of point cloud coordinates in the spatial domain. Dimensions ,in The number of point cloud data points; In step S313, the sample pairing method uses nearest neighbor search: for each complex domain feature vector In the spatial domain point cloud, the point with the closest Euclidean distance is searched using a KD-tree. ,form For paired samples Meanwhile, for each spatial domain point Search for the nearest neighbor eigenvector in the set of eigenvectors in the complex field to form a bidirectional pairing to enhance the robustness of mutual information estimation; The iteration stopping condition in step S315 is that any of the following conditions are met: continuous The absolute value of the change in mutual information in each iteration is less than a preset threshold; In the gradient descent method, the gradient magnitude of all transformation parameters is less than a preset threshold. Reaching the maximum number of iterations .

5. The one according to claim 1, characterized in that: The specific process of extracting the Gaussian curvature extrema points and their local normal vector change rates from the point cloud data of the local sensor in step S2 to construct the second complex feature set includes: S221. Perform local surface fitting on the point cloud data: For each point cloud data point... Utilizing its neighborhood point set Fitting the quadratic surface using the least squares method And calculate the first fundamental form coefficients of the surface. Second fundamental form coefficient ; S222, Calculate the Gaussian curvature of each point cloud data point. And extract the values ​​of Gaussian curvature whose absolute values ​​are greater than a preset threshold. The point is taken as the extreme point of Gaussian curvature, denoted as . ; S223, For each Gaussian curvature extreme point To calculate the rate of change of its local normal vector: first, the normal vector is obtained by fitting the plane through the neighborhood point set. Then calculate the relationship between this point and all points in its neighborhood. The rate of change of the normal vector difference along the spatial distance: in For point The set of neighborhood point indices; S224. The three-dimensional coordinates of the Gaussian curvature extremum point With the corresponding rate of change of the local normal vector Combination, by constructing complex forms Form features in the complex field and construct a second set of complex features.

6. The method for fusing space-ground multi-source deformation monitoring data based on virtual target coding as described in claim 1, characterized in that, Before performing cross-scale iterative registration in step S3, a multi-scale decomposition step based on SAR imaging geometric priors is also included: S321. Obtain the spatial resolution of space-based SAR imagery. Satellite incident angle Spatial resolution of digital elevation model (DEM) data and local sensor point cloud data ; S322, Calculate the number of basic decomposition layers And extract the terrain slope angle of each pixel from the DEM data. ,when Add extra layers Total number of decomposition layers This DEM data is only used for slope priors in multi-scale decomposition and is not used as a registration control benchmark. S323, Decompose the space-based SAR imagery and local point cloud data into... The layer-scale spatial representation uses standard Gaussian pyramid downsampling for the azimuth direction of the SAR image and non-uniform downsampling for the range direction based on the projection relationship between slant range and ground distance, so that the registration error at different scales remains consistent under the slant range projection geometry. Registration is performed layer by layer from the coarsest scale to the finest scale, and the registration result of each layer is used as the initial transformation matrix for the next layer.

7. The method for fusing space-ground multi-source deformation monitoring data based on virtual target coding as described in claim 1, characterized in that: A preprocessing step is also included between step S1 and step S2: T1. The original complex domain echo data of the space-based SAR image is encoded with range and azimuth compression, and the real and imaginary parts of the complex signal are completely preserved to obtain a complex domain single-view complex image; wherein, amplitude extraction is not performed during the encoding stage. T2. Perform voxel downsampling and statistical outlier filtering on the point cloud data of the local sensor to obtain a sparse point cloud, and calculate the local normal vector change rate of each point. T3. Based on satellite orbit parameters and local sensor timestamps, perform coarse temporal alignment between the complex domain single-view complex image and the sparse point cloud.

8. The method for fusing space-ground multi-source deformation monitoring data based on virtual target coding as described in claim 1, characterized in that: Step S4 is followed by a closed-loop feedback step: S5. Input the unified deformation field generated in the current cycle into the time series prediction model to predict the deformation region distribution and deformation rate in the next monitoring cycle. S6. Based on the prediction results, dynamically adjust the extraction strategy of virtual target point encoding in step S2 of the next cycle: increase the spatial density of virtual target points in areas where the predicted deformation rate is higher than the preset threshold, and reduce the density of virtual target points in areas where the predicted deformation rate is lower than the preset threshold. S7. Feed the adjusted virtual target encoding back to the space-based SAR data acquisition strategy to select alternative SAR image data sources with better incident angles or heading angles for the next cycle, and increase the data acquisition frequency in high-risk deformation areas; at the same time, feed the prediction results back to the local sensor to optimize the scanning path planning and sampling density distribution of the local sensor for the next cycle.

9. The method for fusing space-ground multi-source deformation monitoring data based on virtual target coding as described in claim 1, characterized in that: The local sensor includes at least one of millimeter-wave radar, visual sensor, or lidar; the original complex domain echo data of the space-based SAR image is a complex format baseband signal directly transmitted from the SAR satellite; in the constraint of maximizing mutual information between the complex domain and the spatial domain, the complementarity of information between the two domains is guaranteed by the following physical relationship: the phase information of the SAR complex domain echo data is sensitive to subwavelength level deformation, and the spatial coordinates of the point cloud data are sensitive to geometric deformation at the centimeter level and above. The complementary characteristics of the two physical quantities at the deformation scale help to coordinately constrain the registration error at the two scales.

10. A multi-source heterogeneous monitoring data fusion system, characterized in that, The system includes: The data acquisition module is used to acquire raw complex domain echo data of space-based SAR images and point cloud data of local sensors; wherein, the raw complex domain echo data is an SLC complex image obtained by focusing the raw complex echo data of SAR, and contains both real and imaginary part information; the point cloud data does not contain physical quantities corresponding to the phase information of the complex domain echo data. A virtual target generation module, connected to the data acquisition module, is used to extract the complex scattering coefficients, phase center positions, and phase center change rates of coherent scattering points from the original complex domain echo data of the space-based SAR image, constructing a first complex feature set; extracting the Gaussian curvature extrema points and their local normal vector change rates from the point cloud data of the local sensor, constructing a second complex feature set; and encoding the first and second complex feature sets into the same complex domain feature space through complex wavelet transform to generate virtual target code that does not depend on physical reflector hardware. The cross-scale registration module, connected to the virtual target generation module, is used to establish a constraint for maximizing mutual information between the complex domain and the spatial domain, and to define cross-modal mutual information. ,in Information entropy is represented; with maximizing the mutual information as the optimization objective, the spatial domain transformation matrix parameters are iteratively updated using the gradient descent method, wherein the complex domain virtual target point encoding serves as a fixed reference benchmark to guide the optimization direction of the spatial domain parameters, thereby achieving cross-scale iterative registration between space-based SAR images and local point cloud data; The unified deformation field output module is connected to the cross-scale registration module. It is used to map the macroscopic deformation data corresponding to the space-based SAR image and the microscopic deformation data corresponding to the local point cloud data to the same physical coordinate reference based on the registration results, generate unified deformation field data and output it.

Images