InSAR Deformation Inversion Method Based on Ground Feature Deformation Fingerprint and Kinematic Phase Coupling

Abstract

This invention discloses an InSAR deformation inversion method based on ground feature deformation fingerprint and kinematic phase coupling. The method includes the following steps: S1. Acquiring multi-source remote sensing data of the area to be monitored during a continuous monitoring period and performing preprocessing; S2. For the preprocessed data at each monitoring time, calculating the normalized vegetation index, thermal response rate, microwave scattering coefficient, and texture features of each pixel, and constructing the ground feature deformation fingerprint features of the pixel using the calculated parameters; S3. Calculating the similarity between the ground feature deformation fingerprint features and those of each pixel in each sample in the sample library; S4. Determining the pixel category based on the similarity; S5. Selecting the corresponding dynamic phase coupling model according to the pixel category to calculate the phase of each pixel, and then converting the phase into its deformation amount; S6. Repeating S2-S5 to obtain the deformation amount of each pixel at each time point in the continuous monitoring period, and filtering the deformation amount to obtain the InSAR deformation time series of the area to be monitored during the continuous monitoring period.

Description

Technical Field

[0001] This invention relates to the field of InSAR deformation monitoring technology, specifically to an InSAR deformation inversion method based on the coupling of ground feature deformation fingerprints and kinematic phase. Background Technology

[0002] With its advantages of wide coverage and high precision, InSAR technology has become one of the core technologies for deformation monitoring. Traditional InSAR inversion methods use a "one-size-fits-all" mathematical model (such as linear velocity + seasonal period) to process all pixels, ignoring the differences in deformation physical mechanisms among different land features. For example, rigid infrastructure (bridges, railways) is mainly subject to thermoelastic deformation, and there are essential differences in core driving factors, response laws (linear / nonlinear), and key influencing parameters. This leads to semantic classification errors such as "land feature category ≠ deformation mechanism," crosstalk between different mechanisms caused by irrelevant parameters in the calculation, and the model failing to fit the actual deformation physical laws. Ultimately, this results in the core contradiction of "deformation mechanism and model mismatch," which leads to noise interference, frequent misjudgments, and low inversion accuracy. Soil and vegetation areas are mainly subject to hydraulic-mechanical deformation, but traditional models do not include "humidity-rainfall coupling terms" and only consider one factor, resulting in soil moisture change signals interfering with infrastructure monitoring, or the amplifying effect of rainfall on deformation being ignored. Load-sensitive areas (warehouses, heavy-load roads) are mainly subject to load-driven deformation, but traditional models lack a "load-deformation" physical correlation term and use a general periodic model for fitting, leading to deformation caused by load changes being misjudged as natural settlement. The limitation of this single model leads to interference between thermal noise and hydrological signals, frequently resulting in problems such as "misjudging thermal expansion and contraction as subsidence" and "soil moisture changes interfering with infrastructure monitoring," which seriously restricts the accuracy of inversion.

[0003] Existing improvements mainly focus on data preprocessing stages such as atmospheric correction and phase unwrapping, failing to fundamentally address the core contradiction of "mismatch between deformation mechanisms and models." While traditional semantic classification methods can distinguish land cover categories, they rely solely on semantic information from optical images and cannot directly reflect the deformation mechanisms of land covers. For example, some concrete roads may be misclassified due to the influence of surrounding vegetation, and land covers of the same category may exhibit different deformation mechanisms due to environmental differences.

[0004] Therefore, there is an urgent need for a feature extraction method that can directly correlate with the deformation mechanism of ground features, so as to achieve precise coupling between the deformation mechanism and the kinematic model and break through the existing technical bottlenecks. Summary of the Invention

[0005] To address the aforementioned shortcomings in existing technologies, the InSAR deformation inversion method based on ground feature deformation fingerprint and kinematic phase coupling provided by this invention solves the problem of low InSAR deformation inversion accuracy caused by inaccurate determination of deformation type in existing technologies.

[0006] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: A method for InSAR deformation inversion based on the coupling of ground feature deformation fingerprints and kinematic phase is provided, which includes the following steps: S1. Collect multi-source remote sensing data of the area to be monitored during the continuous monitoring period, as well as rainfall, soil moisture and load changes, and preprocess the multi-source remote sensing data. S2. For the preprocessed data at each monitoring time point, calculate the normalized vegetation index, thermal response rate, microwave scattering coefficient, and texture features for each pixel, and construct the ground feature deformation fingerprint of the pixel using the calculated parameters. ; S3, Calculation Ground feature deformation fingerprint features of each pixel in each sample in the sample library The similarity is such that the sample library includes the ground feature deformation fingerprint features of each pixel of several samples; S4. Select the K pixels with the highest similarity and count the categories of the K pixels. Select the category with the highest proportion as the most similar pixel. The corresponding pixel categories include thermoelastic deformation, hydraulic-mechanical deformation, and load-driven deformation. S5. Select the corresponding dynamic phase coupling model according to the pixel category, and calculate the phase of each pixel by combining rainfall, soil moisture and load change. Then convert the phase into its deformation. S6. Repeat steps S2 to S5 to obtain the deformation amount of each pixel at each moment in the continuous time period, and filter the deformation amount of all pixels to obtain the InSAR deformation time series of the area to be monitored in the continuous monitoring period.

[0007] Furthermore, the expressions for the dynamic phase coupling models corresponding to the thermoelastic deformation, hydraulic-mechanical deformation, and load-driven deformation are respectively: in, It contributes phase to thermoelastic deformation; The phase coefficient is the linear thermal expansion coefficient; To monitor the temperature change over time; This is the initial phase of thermal expansion; This is a correction factor for the nonlinear thermal response; Contributes phase to hydraulic-mechanical deformation; The soil moisture-sensitive phase coefficient; This represents the change in soil moisture. This refers to the phase coefficient of the rainfall response. To monitor rainfall at any given time; This represents the initial phase of hydrological expansion; This is the humidity-rainfall coupling correction factor; Contributes phase to load-driven deformation; The load-sensitive phase coefficient; To monitor the change in load over time; This is a correction factor for nonlinear loads; The initial phase is driven by the load.

[0008] Furthermore, the parameters in the dynamic phase coupling model , , , , , , , , and Inversion methods include: A1. Construct a parameter design vector that satisfies the parameter constraints. : A2. Obtain the InSAR time series corresponding to multiple known observations of the area to be monitored, and construct a design matrix where each row corresponds to one InSAR time series. : in, , and This represents the temperature change corresponding to the 1st, 2nd, and Nth observations; , and These represent the changes in soil moisture at the 1st, 2nd, and Nth observations, respectively. , and These represent the rainfall amounts corresponding to the 1st, 2nd, and Nth observations, respectively. , and These represent the load changes corresponding to the 1st, 2nd, and Nth observations, respectively. A3. Design vectors based on parameters. and design matrix Constructing a simulated phase vector Then, the parameters are solved using the constrained weighted least squares algorithm. , , , , , , , , and .

[0009] Furthermore, the parameter constraint condition is: when the pixel is a thermoelastic deformation type, , , , and Set to 0; when the pixel is of the hydraulic-mechanical deformation type, , , , Set to 0; when the pixel is a load-driven deformation type, , , , , Set to 0; The objective function of the constrained weighted least squares algorithm is: in, For the first Weighted residuals of each observation; For the first InSAR coherence coefficient of the second observation; For the first Pixel-level simulated phase of the next observation; For designing a matrix The OK.

[0010] Furthermore, the InSAR deformation inversion method also includes parameter... , and Make corrections: , in, For the updated , and ; The solution obtained using the constrained weighted least squares algorithm , and ; The calibration step size is set to 0.005~0.02. To and The simulated phase calculated by the dynamic phase coupling model corresponding to the number of observations; For the first The observation weight of each observation.

[0011] Furthermore, the phase of a pixel is converted into its deformation using a phase-deformation conversion formula, which is as follows: in, The phase corresponding to the pixel; The radar wavelength; This represents the deformation amount of a pixel.

[0012] Furthermore, the texture features include contrast, entropy value, and correlation, and the land feature deformation fingerprint features The expression is: in, , , , , and These are the normalized vegetation index, thermal response rate, microwave scattering coefficient, contrast, entropy, and correlation, respectively. They are respectively , , , , and The corresponding weighting coefficients.

[0013] Furthermore, the multi-source remote sensing data includes high-resolution optical images, thermal infrared images, and radar remote sensing images. The preprocessing of the three types of images is as follows: first, the three types of images are spatiotemporally registered, and then radiometric correction, geometric correction, and noise removal are performed in sequence. The spatiotemporal registration is based on the optical image, and the thermal infrared image and radar remote sensing image are aligned with it. Near-infrared band using optical imaging and the red band of thermal infrared images Calculate the normalized vegetation index for each pixel. ; The thermal response rate of each pixel is calculated using thermal infrared images from a continuous monitoring period. , and These are the temperatures at pixel l corresponding to time m+1 and time m, respectively; The radar remote sensing image is Sentinel-1 data VV / VH dual-polarization data, and the microwave scattering coefficient is read from Sentinel-1 data VV / VH dual-polarization data; The texture features include contrast, entropy, and correlation, all of which are calculated from the optical image using a gray-level co-occurrence matrix.

[0014] The beneficial effects of this invention are as follows: This scheme fuses multi-source remote sensing data (high-resolution optical images, thermal infrared images, and radar remote sensing images) with auxiliary data (rainfall, soil moisture, and load changes) to extract stable feature combinations directly related to the physical mechanisms of land cover deformation, constructing pixel-by-pixel land cover deformation fingerprint feature vectors, and thereby accurately attributing (classifying) the deformation mechanism. Unlike traditional methods that directly construct deformation models based solely on semantic or empirical assumptions, the land cover deformation fingerprint features obtained by this scheme... It does not directly participate in InSAR phase inversion calculations, but instead provides pixel-level physical prior information for subsequent inversion processes. This physical prior is output in the form of "deformation mechanism category labels," which are essentially a mechanism constraint used to limit the types of physical models that can participate in subsequent inversions and their parameter space ranges.

[0015] This solution replaces traditional semantic classification by using deformation fingerprints and mechanism attribution. By fusing multi-source data to extract unique deformation association features (deformation fingerprints) of ground features, it accurately matches the corresponding physical deformation model and achieves dynamic coupling between deformation mechanism and kinematic equation. This effectively eliminates cross-mechanism noise interference and significantly improves the deformation inversion accuracy of scenarios such as infrastructure and soil vegetation. It is suitable for high-precision scenarios such as engineering monitoring and geological disaster early warning. Attached Figure Description

[0016] Figure 1 This is a block diagram illustrating the principle of the InSAR deformation inversion method based on the coupling of ground feature deformation fingerprints and kinematic phase. Detailed Implementation

[0017] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0018] refer to Figure 1 , Figure 1 A flowchart of an InSAR deformation inversion method based on the coupling of ground feature deformation fingerprints and kinematic phase is shown, as follows: Figure 1 As shown, the method includes steps S1 to S6.

[0019] In step S1, high-resolution optical images, thermal infrared images, and radar remote sensing images of the area to be monitored during the continuous monitoring period, as well as rainfall, soil moisture, and load changes, are collected, and the three types of images are preprocessed. High-resolution optical images can be selected from Sentinel-2, Gaofen series satellite images, or aerial imagery, with a spatial resolution ≥10 meters, including visible and near-infrared bands. Level-1C and higher level data can be downloaded from the satellite data distribution platform. Thermal infrared images can be selected from Sentinel-3 and Landsat 8 satellite thermal infrared data, with a temperature inversion accuracy ≥±0.5℃. Level-2 data (with radiometric calibration and atmospheric correction completed) can be downloaded. Microwave scattering coefficient data can be selected from Sentinel-1 VV / VH dual-polarization data, with a polarization ratio calculation accuracy ≥0.01. Level-1 SLC or Level-2 GRD data can be downloaded.

[0020] The three images are preprocessed as follows: Spatiotemporal registration uses the optical image as a reference, aligning it with thermal infrared and microwave data, with a registration error ≤ 1 pixel, and uniformly adopting the WGS84 coordinate system. Radiometric correction eliminates the influence of atmospheric scattering and sensor errors on the optical / thermal infrared images; radiometric calibration is performed on the microwave data, converting the echo intensity into dB-level scattering coefficients (…). Geometric correction is optimized using ground control points to ensure that the positional error of ground features is ≤5 meters; optical images are cropped to retain the effective area. Noise removal: median filtering is used to eliminate cloud residue in thermal infrared images, and low-quality pixels with a coherence coefficient <0.3 are removed from radar remote sensing images.

[0021] In step S2, for the three types of preprocessed data at each monitoring time, the normalized vegetation index, thermal response rate, microwave scattering coefficient, and texture features of each pixel are calculated, and the calculated parameters are used to construct the ground feature deformation fingerprint of the pixel. .

[0022] In one embodiment of the present invention, the texture features include contrast, entropy value, and correlation, and the ground feature deformation fingerprint features The expression is: in, , , , , and These are the normalized vegetation index, thermal response rate, microwave scattering coefficient, contrast, entropy, and correlation, respectively. They are respectively , , , , and The corresponding weighting coefficients, specifically, (Spectral feature weights): 0.2~0.3 (Thermal response weight): 0.25~0.35 (Microwave feature weight): 0.15~0.25 (Texture feature weights) are 0.05~0.1 each.

[0023] Near-infrared band using optical imaging and the red band of thermal infrared images Calculate the normalized vegetation index for each pixel. ; The thermal response rate of each pixel is calculated using thermal infrared images from a continuous monitoring period. , and These are the temperatures at pixel l corresponding to time m and time m-1, respectively; The radar remote sensing image is Sentinel-1 data VV / VH dual-polarization data, and the microwave scattering coefficient is read from Sentinel-1 data VV / VH dual-polarization data; The texture features include contrast, entropy, and correlation, all three of which are calculated from the optical image using the gray-level co-occurrence matrix (GLCM). Contrast reflects texture sharpness. The formula for Contrast is... Entropy reflects texture complexity; the formula is Entropy. Correlation (Corr) reflects texture correlation; the formula Corr is... .

[0024] In the above formula The gray level row index represents the pixel gray level value (i.e., the gray level of the center pixel / reference pixel) in the row direction of the Gray-Level Co-occurrence Matrix (GLCM). The column index for grayscale level represents the pixel grayscale value in the column direction of GLCM (i.e., the grayscale level of neighboring pixels / target pixels); The number of gray levels in an image (e.g., 8-bit grayscale image) (i.e., the grayscale value range is 0~255); The gray-level co-occurrence matrix (GLCM) is the first... Line number Column elements represent the grayscale levels under a specified direction and step size. With gray level The normalized probability of the co-occurrence of pixel pairs (satisfying) );in grayscale The weighted mean: (by For weighted pairs (Calculate the average); grayscale The weighted mean: (by For weighted pairs (Calculate the average); grayscale Weighted standard deviation: ; grayscale Weighted standard deviation: .

[0025] The contrast formula in this scheme measures the "depth difference" of texture. Textures of man-made structures (such as the grid of a bridge or the parallel lines of a railway track) are regular and clear, resulting in high contrast values; textures of natural areas (such as soil and grassland) are messy, resulting in low contrast. It refines the morphological differences of features, avoiding misclassification of "vegetated concrete roads" as pure soil areas. The higher the entropy value calculated by the formula, the more "chaotic" the texture. For example, the distribution of crops in farmland and the vegetation cover on hillsides have high entropy values; while the textures of warehouse roofs and bridge surfaces are regular, resulting in low entropy values. Combined with contrast, it can more comprehensively describe the morphology of features, making the fingerprint vector more discriminative. The correlation formula examines the "regularity" of texture. The closer the correlation is to 1, the more directional the texture is (such as the direction of railway tracks or the arrangement of beams in a bridge); textures of natural features do not have a fixed direction, resulting in low correlation. It further strengthens the distinction between man-made structures and natural features, providing more detailed evidence for mechanism attribution.

[0026] In step S3, calculate Ground feature deformation fingerprint features of each pixel in each sample in the sample library The similarity is calculated using the following formula: The sample library includes the ground feature deformation fingerprint features of each pixel in several samples. in, To ensure the accuracy of the formula, which is a very small number, and to avoid extreme cases (such as a feature value of 0 leading to calculation errors), making the matching results more stable; this formula is the core of determining "which sample database's mechanism is most similar to the feature to be classified" -- It is a "deformed identity card" for the features to be identified. It serves as the "identity card" for known mechanisms in the sample library. By calculating the similarity between the two, the most matching sample is found.

[0027] In step S4, the K pixels with the highest similarity are selected, and the categories of the K pixels are counted. The category with the highest proportion is taken as the most similar category. For each pixel, the sample categories include thermoelastic deformation, hydraulic-mechanical deformation, and load-driven deformation. In this scheme, thermoelastic deformation samples include bridge and railway samples; hydraulic-mechanical deformation samples include soil and farmland samples; and load-driven deformation samples include warehouse and heavy-duty road samples. Each category has ≥500 samples, covering different regions and environmental conditions to ensure sample diversity.

[0028] In step S5, a corresponding dynamic phase coupling model is selected based on the pixel category, and the phase of each pixel is calculated by combining rainfall, soil moisture, and load change. The phase is then converted into its deformation. The expressions for the dynamic phase coupling models corresponding to the thermoelastic deformation, hydraulic-mechanical deformation, and load-driven deformation are as follows: , and : in, Contributes phase to thermoelastic deformation (unit: ); The linear thermal expansion phase coefficient (unit: / ); To monitor the temperature change over time (unit: ); The initial phase of thermal expansion (unit: ); Nonlinear thermal response correction coefficient (unit: ), steel price .

[0029] This model is tailor-made for the "thermoelastic deformation label" pixel, specifically describing deformation caused by temperature changes. Linear terms. It is the main contribution of deformation, not the nonlinear term. This is to correct for deformation deviations under extreme temperatures. For example, when steel is at high or low temperatures, the deformation is no longer a purely linear change. Adding this term makes the model more closely match the actual situation and avoids the fitting error of traditional linear models.

[0030] in, Contributing phase to hydraulic-mechanical deformation (unit: ); Soil moisture-sensitive phase coefficient (unit: / ); Changes in soil moisture (unit: ); Rainfall response phase coefficient (unit: / ); Rainfall at the time of monitoring (unit: ); Initial phase of hydrological expansion (unit: ); Humidity-rainfall coupling correction factor (unit: ), values ​​for clay soil .

[0031] This model corresponds to the "hydraulic-mechanical deformation label," specifically describing the deformation of soil and vegetation areas—increased soil moisture causes expansion, and rainfall also triggers deformation, with the two influencing each other (for example, the effect of rainfall is more pronounced in high humidity). Coupling terms This is to capture this synergistic effect and solve the problem that traditional models only consider humidity or rainfall individually, leading to inaccurate simulations.

[0032] in, Contributes phase to load-driven deformation (unit: ); Load-sensitive phase coefficient (unit: ), value of reinforced concrete ; To monitor the load change at any given time (unit: ); Nonlinear load correction factor (unit: ), value ; Initial phase driven by load (unit: ).

[0033] This model is a dedicated model for the "load-driven label," filling the gap in traditional models that lack a "load-deformation" relationship. For example, in heavily loaded roads and warehouses, increased load triggers deformation, and this deformation deviates from linearity under heavy loads (e.g., reinforced concrete exhibits slight nonlinear deformation under heavy loads). The nonlinear term... This is to correct this deviation and avoid misjudging load deformation as natural settlement.

[0034] In this plan, , , and The data sources and acquisition methods can be found in Table 1.

[0035] Table 1 This scheme calculates the phase of each pixel through a dynamic phase coupling model. This process essentially establishes the mapping relationship between "deformation fingerprint - physical model - parameter space", transforming the subsequent inversion problem from a traditional unconstrained solution with full parameters into a physical inversion problem with consistent mechanism and constrained parameters, which significantly improves the model's identifiability and physical rationality.

[0036] In implementation, this scheme preferably uses a phase-deformation conversion formula to convert the phase of a pixel into its deformation. The phase-deformation conversion formula is as follows: in, The phase corresponding to the pixel; The radar wavelength; This represents the deformation amount of a pixel.

[0037] In step S6, steps S2 to S5 are repeated to obtain the deformation amount of each pixel at each moment in the continuous time period, and the deformation amount of all pixels is filtered to obtain the InSAR deformation time series of the area to be monitored in the continuous monitoring period; this scheme preferably uses adaptive Kalman filtering noise reduction to filter the deformation amount of all pixels.

[0038] In one embodiment of the present invention, the parameters in the dynamic phase coupling model , , , , , , , , and Inversion methods include: A1. Construct a parameter design vector that satisfies the parameter constraints. : The parameter constraint is: when the pixel is a thermoelastic deformation type, , , , and Set to 0; when the pixel is of the hydraulic-mechanical deformation type, , , , Set to 0; when the pixel is a load-driven deformation type, , , , , Set it to 0.

[0039] A2. Obtain the InSAR time series corresponding to multiple known observations of the area to be monitored, and construct a design matrix where each row corresponds to one InSAR time series. : in, , and This represents the temperature change corresponding to the 1st, 2nd, and Nth observations; , and These represent the changes in soil moisture at the 1st, 2nd, and Nth observations, respectively. , and These represent the rainfall amounts corresponding to the 1st, 2nd, and Nth observations, respectively. , and These represent the load changes corresponding to the 1st, 2nd, and Nth observations, respectively. In this scheme, the G-matrix construction based on deformation mechanism constraints ensures that parameter solving in the InSAR inversion process is performed only within a physically consistent parameter subspace. Its technical advantages include at least the following: (1) The mathematical coupling between parameters of different deformation mechanisms is eliminated at the design matrix structure level, avoiding mutual interference between temperature change, hydrological change and load change signals; (2) The identifiability and physical rationality of model parameters are improved, and the inversion error caused by model mismatch is reduced; (3) Clear physical prior conditions are provided for subsequent parameter solving using the constrained weighted least squares algorithm, which significantly improves the stability and accuracy of deformation inversion results.

[0040] Therefore, the G matrix in this invention is not a fixed, unified design matrix, but a constrained physical design matrix dynamically constructed based on a pixel-by-pixel deformation mechanism. It is one of the key technical links to realize the "coupling inversion of ground feature deformation fingerprint and kinematic model".

[0041] In implementation, the objective function of the preferred constrained weighted least squares algorithm in this scheme is: in, For the first Weighted residuals of each observation; For the first InSAR coherence coefficient of the second observation; For the first Pixel-level simulated phase of the next observation; For designing a matrix The OK.

[0042] The objective function designed in this scheme aims to find a set of optimal parameters. The goal is to make the "observed phase" and the "model-calculated phase" as close as possible. The weighted residual is the difference between the two, multiplied by a weight, summed, and then the minimum value is found. The parameters obtained in this way not only closely match actual observations but also highlight the role of high-quality data and avoid low-quality data dragging down the accuracy.

[0043] A3. Design vectors based on parameters. and design matrix Constructing a simulated phase vector Then, the parameters are solved using the constrained weighted least squares algorithm. , , , , , , , , and .

[0044] In implementation, the preferred InSAR deformation inversion method in this scheme also includes parameter... , and Make corrections: , in, For the updated , and ; The solution obtained using the constrained weighted least squares algorithm , and ; The calibration step size is set to 0.005~0.02. To and The simulated phase calculated by the dynamic phase coupling model corresponding to the number of observations; For the first The observation weight of each observation.

[0045] The effectiveness of the InSAR deformation inversion method provided in this scheme will be illustrated below with specific examples: 1. Data Acquisition and Preprocessing This embodiment selects a steel bridge spanning a river as the monitoring object, and collects Sentinel-1 InSAR time series data covering the area, with radar wavelength... cm, time span 12 months, revisit period 12 days; synchronous acquisition -2 optical image (10m resolution) -3 thermal infrared data, Sentinel-1 VV / VH dual-polarization scattering coefficient data, hourly temperature, rainfall, ground soil moisture, local road load monitoring data from meteorological stations, and GNSS measured deformation data. All data underwent spatiotemporal registration, radiometric correction, geometric correction, and noise removal, with a registration error of <1 pixel, and were unified to the WGS84 coordinate system.

[0046] 2. Extraction of fingerprint features of ground feature deformation Pixel-by-pixel calculation of 6-dimensional deformation fingerprint features in the study area: Normalized Difference Vegetation Index (NDVI): 0.1~0.2 for the main body of the steel bridge, 0.3~0.4 for the edge slope and soil area; Thermal Response Rate (... ): Steel bridge main body 2.5~3.0℃ / h, edge soil area 0.5~1.0℃ / h; microwave scattering coefficient ( ): The main body of the steel bridge has a contrast of -10 to -5dB, while the loose edge area has a contrast of -20 to -15dB; Texture features: the main body has a contrast of 80 to 100, low entropy, and high correlation; the edge area has low contrast, high entropy, and weak correlation.

[0047] The above features are normalized according to their weights to construct a pixel-by-pixel deformable fingerprint feature vector: in ,satisfy .

[0048] 3. Attribution of pixel-by-pixel deformation mechanism Each pixel-wise deformed fingerprint is matched pixel-by-pixel with the sample database using an improved cosine similarity method. Smoothing factor Nearest neighbor number .

[0049] Attribution results: ① Approximately 85% of the main steel bridge pixels: have the highest similarity to thermoelastic deformation samples, and are determined to be thermoelastic deformation mechanism; ② Approximately 10% of the edge slope protection / soil pixels: have the highest similarity to hydraulic-mechanical deformation samples, and are determined to be hydraulic-mechanical deformation mechanism; ③ Approximately 5% of the locally heavy-load pavement pixels: have the highest similarity to load-driven deformation samples, and are determined to be load-driven deformation mechanism.

[0050] Finally, a pixel-by-pixel deformation mechanism attribution mask is generated for subsequent dynamic matching of the model.

[0051] 4. Three types of kinematic phase models and parameters 4.1 Thermal expansion phase model (thermoelastic deformation): , ; ; ; 4.2 Hydrological Expansion Phase Model (Hydraulic-Mechanical Deformation): , , , , ; 4.3 Load-driven phase model (load-driven deformation): , , , .

[0052] 5. Phase transition distortion: Then, adaptive Kalman filtering is performed for noise reduction.

[0053] 6. Results and Accuracy: ① 85% of the main body of the steel bridge: Deformation is highly correlated with temperature changes, and there is no misjudgment of settlement due to thermal expansion and contraction; ② 10% of the edges: Deformation is consistent with the coupling changes of humidity and rainfall; ③ 5% of the local area: Deformation matches the changes of heavy load; ④ The RMSE of the global inversion results is ≤0.3mm, which has a high degree of agreement with the GNSS measurement and no cross-mechanism noise crosstalk. It can be directly used for bridge safety monitoring and early warning.

[0054] In summary, this solution utilizes a self-developed "land feature deformation fingerprint" feature extraction method, integrating multi-source data to directly correlate deformation mechanisms. Compared to traditional semantic classification, it avoids the misjudgment problem of "land feature category ≠ deformation mechanism," improving attribution accuracy by over 20% and ensuring the accuracy of model matching from the source. The kinematic model introduces nonlinear correction terms and coupling correction terms, which better reflect the actual deformation physics of land features, resulting in significantly better model fitting accuracy than traditional linear models. The optimization strategy combining weighted least squares algorithm and adaptive Kalman filtering highlights the role of high-quality observation data while effectively suppressing noise interference, achieving a root mean square error (RMSE) of ≤0.3mm for deformation inversion. The system integrates full-process functions such as multi-source data acquisition, fingerprint extraction, mechanism attribution, model coupling, and solution optimization, exhibiting a high degree of automation and flexible adaptability to different monitoring scenarios, demonstrating broad engineering application value and market competitiveness.

Claims

1. An InSAR deformation inversion method based on the coupling of ground feature deformation fingerprints and kinematic phase, characterized in that, Including the following steps: S1. Collect multi-source remote sensing data of the area to be monitored during the continuous monitoring period, as well as rainfall, soil moisture and load changes, and preprocess the multi-source remote sensing data. S2. For the preprocessed data at each monitoring time point, calculate the normalized vegetation index, thermal response rate, microwave scattering coefficient, and texture features for each pixel, and construct the ground feature deformation fingerprint of the pixel using the calculated parameters. ; S3, Calculation Ground feature deformation fingerprint features of each pixel in each sample in the sample library The similarity is such that the sample library includes the ground feature deformation fingerprint features of each pixel of several samples; S4. Select the K pixels with the highest similarity and count the categories of the K pixels. Select the category with the highest proportion as the most similar pixel. The corresponding pixel categories include thermoelastic deformation, hydraulic-mechanical deformation, and load-driven deformation. S5. Select the corresponding dynamic phase coupling model according to the category of the pixel, and calculate the phase of each pixel by combining rainfall, soil moisture and load change, and then convert the phase into its deformation. S6. Repeat steps S2 to S5 to obtain the deformation amount of each pixel at each moment in the continuous time period, and filter the deformation amount of all pixels to obtain the InSAR deformation time series of the area to be monitored in the continuous monitoring period.

2. The InSAR deformation inversion method according to claim 1, characterized in that, The expressions for the dynamic phase coupling models corresponding to the thermoelastic deformation, hydraulic-mechanical deformation, and load-driven deformation are as follows: in, It contributes phase to thermoelastic deformation; The phase coefficient is the linear thermal expansion coefficient; To monitor the temperature change over time; This is the initial phase of thermal expansion; This is a correction factor for the nonlinear thermal response; Contributes phase to hydraulic-mechanical deformation; The soil moisture-sensitive phase coefficient; This represents the change in soil moisture. This refers to the phase coefficient of the rainfall response. To monitor rainfall at any given time; This represents the initial phase of hydrological expansion; This is the humidity-rainfall coupling correction factor; Contributes phase to load-driven deformation; The load-sensitive phase coefficient; To monitor the change in load over time; This is a correction factor for nonlinear loads; The initial phase is driven by the load.

3. The InSAR deformation inversion method according to claim 2, characterized in that, Parameters in the dynamic phase coupling model , , , , , , , , and Inversion methods include: A1. Construct a parameter design vector that satisfies the parameter constraints. : A2. Obtain the InSAR time series corresponding to multiple known observations of the area to be monitored, and construct a design matrix where each row corresponds to one InSAR time series. : in, , and This represents the temperature change corresponding to the 1st, 2nd, and Nth observations; , and These represent the changes in soil moisture at the 1st, 2nd, and Nth observations, respectively. , and These represent the rainfall amounts corresponding to the 1st, 2nd, and Nth observations, respectively. , and These represent the load changes corresponding to the 1st, 2nd, and Nth observations, respectively. A3. Design vectors based on parameters. and design matrix Constructing a simulated phase vector Then, the parameters are solved using the constrained weighted least squares algorithm. , , , , , , , , and .

4. The InSAR deformation inversion method according to claim 3, characterized in that, The parameter constraint is: when the pixel is a thermoelastic deformation type, , , , and Set to 0; when the pixel is of the hydraulic-mechanical deformation type, , , , Set to 0; when the pixel is a load-driven deformation type, , , , , Set to 0; The objective function of the constrained weighted least squares algorithm is: in, For the first Weighted residuals of each observation; For the first InSAR coherence coefficient of the second observation; For the first Pixel-level simulated phase for each observation; For designing a matrix The OK.

5. The InSAR deformation inversion method according to claim 3, characterized in that, It also includes parameters , and Make corrections: ， in, For the updated , and ; The solution obtained using the constrained weighted least squares algorithm , and ; The calibration step size is set to 0.005~0.

02. To and The simulated phase calculated by the dynamic phase coupling model corresponding to the number of observations; For the first The observation weight of each observation.

6. The InSAR deformation inversion method according to claim 1, characterized in that, The phase-deformation conversion formula is used to convert the phase of a pixel into its deformation. The phase-deformation conversion formula is as follows: in, The phase corresponding to the pixel; The radar wavelength; This represents the deformation amount of a pixel.

7. The InSAR deformation inversion method according to claim 1, characterized in that, The texture features include contrast, entropy, and correlation; the ground feature deformation fingerprint features The expression is: in, , , , , and These are the normalized vegetation index, thermal response rate, microwave scattering coefficient, contrast, entropy, and correlation, respectively. They are respectively , , , , and The corresponding weighting coefficients.

8. The InSAR deformation inversion method according to claim 1, characterized in that, The multi-source remote sensing data includes high-resolution optical images, thermal infrared images, and radar remote sensing images. The preprocessing of the three types of images is as follows: first, the three types of images are spatiotemporally registered, and then radiometric correction, geometric correction, and noise removal are performed in sequence. The spatiotemporal registration is based on the optical image, and the thermal infrared image and radar remote sensing image are aligned with it.

9. The InSAR deformation inversion method according to claim 8, characterized in that, Near-infrared band using optical imaging and the red band of thermal infrared images Calculate the normalized vegetation index for each pixel. ; The thermal response rate of each pixel is calculated using thermal infrared images from a continuous monitoring period. , and These are the temperatures at pixel l corresponding to time m+1 and time m, respectively; The radar remote sensing image is Sentinel-1 data VV / VH dual-polarization data, and the microwave scattering coefficient is read from Sentinel-1 data VV / VH dual-polarization data; The texture features include contrast, entropy, and correlation, all of which are calculated from the optical image using a gray-level co-occurrence matrix.

Images