Building deformation monitoring method based on attention mechanism and collaborative correction

Abstract

This invention discloses a building deformation monitoring method based on attention mechanism and collaborative correction, belonging to the field of building deformation monitoring technology. Using multi-source temporal SAR data as the core, it constructs a full-process monitoring framework encompassing data preprocessing, feature enhancement and matching, deformation inversion, and accuracy verification. First, an innovative preprocessing algorithm eliminates speckle noise and atmospheric delay effects in SAR images. Second, feature extraction and matching strategies are optimized based on building target characteristics to improve target identification accuracy. Third, a temporal deformation inversion fusion model accurately calculates the three-dimensional deformation of the building. Finally, accuracy verification is completed by combining multi-source ground-based measured data and quantitative verification indicators. This invention, employing the aforementioned building deformation monitoring method based on attention mechanism and collaborative correction, effectively solves the problems of low accuracy and poor robustness of traditional SAR deformation monitoring methods in complex scenarios, achieving high-precision, full-process monitoring of building deformation.

Description

Technical Field

[0001] This invention relates to the field of building deformation monitoring technology, and in particular to a building deformation monitoring method based on attention mechanisms and collaborative correction. Background Technology

[0002] Building deformation is a core concern in engineering construction, urban operation and maintenance, and geological disaster prevention and control. Accurate monitoring of deformation trends and magnitudes is crucial to ensuring the safety of building structures. Synthetic Aperture Radar (SAR) technology, with its unique advantages of all-weather, all-time coverage, wide coverage, and high spatial resolution, has become the mainstream technology for building deformation monitoring.

[0003] However, traditional SAR deformation monitoring methods suffer from drawbacks such as susceptibility to interference in complex scenarios, low utilization of time-series data, and insufficient analysis of deformation mechanisms. Summary of the Invention

[0004] The purpose of this invention is to provide a building deformation monitoring method based on attention mechanism and collaborative correction. By adaptive multi-scale noise suppression and atmospheric delay collaborative correction, and target enhancement and time series inversion fusion based on attention mechanism, it effectively solves the problems of low accuracy and poor robustness of traditional SAR deformation monitoring methods in complex scenarios. It realizes high-precision, full-process monitoring of building deformation and can be widely used in urban operation and maintenance, engineering construction, geological disaster prevention and control and other fields, providing reliable technical support for building safety.

[0005] To achieve the above objectives, this invention provides a building deformation monitoring method based on attention mechanisms and collaborative correction, comprising the following steps:

[0006] S1. Collect multi-source data and synthetic aperture radar (SAR) images of the monitoring area and organize them into an unprocessed time-series SAR image set and auxiliary dataset;

[0007] S2. The data of the time-series SAR image set is processed by combining adaptive multi-scale speckle noise suppression and spatiotemporal atmospheric delay model to obtain the processed time-series SAR image set.

[0008] S3. Input the image data from the time-series SAR image set processed in S2 into the building target enhancement network based on the attention mechanism, fuse the auxiliary dataset and optical image features, and extract the core observation points of the building targets;

[0009] S4. Match the core observation points obtained in S3, remove incorrect matching points, and construct a time series observation point set for the building target;

[0010] S5. Input the time series observation point set of S4 into the time series deformation inversion fusion model, combine the structural mechanics parameters and geological environment data, optimize the model parameters through Bayesian inference, invert the three-dimensional deformation amount and time series change curve of the building, analyze the spatiotemporal distribution characteristics and development trend of deformation, identify the deformation anomaly area and output the deformation inversion results.

[0011] S6. Based on the deformation inversion results of S5, combined with multi-source ground measured data and quantitative verification indicators, the accuracy verification is completed.

[0012] Preferably, the specific process of S1 is as follows:

[0013] S11, the SAR image set includes SAR images and accompanying orbital data; the auxiliary dataset includes acquired digital elevation model (DEM) data, acquired medium resolution imaging spectrometer (MODIS) water vapor data, acquired surface meteorological data, and collected geographic information system (GIS) vector data of the monitoring area with a unified coordinate system.

[0014] S12. Perform an integrity check on the SAR image data acquired in S11, and remove images with excessive cloud cover or corrupted data to finally form a time-series SAR image set.

[0015] Preferably, in step S21, SARscape software is used to complete the basic preprocessing orbit correction and radiometric calibration;

[0016] S22. Then, the image with the best imaging quality in the time-series SAR image set of S12 is used as the reference image, and the remaining images are images to be registered. The SIFT algorithm of scale invariant feature transformation is used to extract feature points. The FLANN fast nearest neighbor search packet is used to complete coarse matching. Then, the RANSAC algorithm is used to remove erroneous matching points. Finally, the image is resampled by bilinear interpolation.

[0017] S23. First, perform adaptive multi-scale wavelet threshold denoising, and then combine the local signal-to-noise ratio to dynamically adjust the threshold to achieve differentiated noise suppression.

[0018] S24. Construct a spatiotemporal atmospheric delay model with spatial interpolation and temporal fitting, and integrate MODIS water vapor data and surface meteorological data to achieve collaborative correction;

[0019] S25. The adaptive multi-scale wavelet threshold denoising process of S23 is iterated in conjunction with the spatiotemporal atmospheric delay model of S24 to achieve coordinated optimization of denoising and correction, and the time-series SAR images are converted into a unified format, and the output is the processed time-series SAR image set.

[0020] Preferably, the specific process of adaptive multi-scale wavelet threshold denoising in S23 is as follows:

[0021] S231. Perform N-level wavelet decomposition on the time-series SAR image to obtain the low-frequency component and the high-frequency component of each level.

[0022] S232. For each high-frequency component in each layer, calculate the local signal-to-noise ratio (SNR) of each pixel and define an adaptive threshold. The formula is as follows:

[0023] ;

[0024] in, These are the x and y coordinates of the pixels, respectively. The wavelet decomposition level is denoted as . For the first The noise standard deviation of the high-frequency components of the layer. The total number of image pixels. for Local signal-to-noise ratio at that location This is the adaptive adjustment coefficient;

[0025] S233. Subsequently, an improved soft thresholding function is used to process high-frequency components. The specific formula is as follows:

[0026] ;

[0027] In the formula, These are the original high-frequency coefficients. For the processed high-frequency coefficients, For symbolic functions, It is an exponential function. To avoid extremely small coefficient values ​​where the denominator is 0.

[0028] Preferably, the specific process of constructing the spatiotemporal atmospheric delay model for spatial interpolation and temporal fitting in S24 is as follows:

[0029] S241. Based on MODIS water vapor data from S11, the formula for the space atmospheric delay field is constructed using the Kriging interpolation algorithm as follows:

[0030] ;

[0031] in, for Atmospheric delay in space, This is a valid MODIS observation point. For interpolation weights, For error terms, This represents the number of observation points;

[0032] S242. Combining the surface meteorological data from S11, the spatial atmospheric delay field is time-series optimized to construct a spatiotemporal atmospheric delay model:

[0033] ;

[0034] in, It is a time series. for time The time-space atmospheric delay at that location , They are respectively Temperature and air pressure at any given time , , These are the time series fitting coefficients.

[0035] Preferably, the specific process of S3 is as follows:

[0036] S31. Construct an attention enhancement network based on Transformer;

[0037] S32. Input the data from the processed temporal SAR image set in S25 into the attention enhancement network trained in S31. The feature encoding layer extracts grayscale features, texture features, and polarization features, and concatenates them to form a feature matrix. ;

[0038] S33. The self-attention module focuses on the spatially related areas of buildings, and the cross-attention module integrates GIS contour masks and optical image features to enhance the matching degree. After 1×1 convolution dimensionality reduction, the Otsu adaptive threshold segmentation method is used to obtain the binary mask of buildings. Then, morphological opening operation is used to denoise and closing operation is used to fill holes, and candidate observation points within the building contour are initially extracted.

[0039] S34. Calculate the coherence coefficient and intensity variance of the candidate observation points initially extracted in S33 in the processed time-series SAR image, and retain the points that meet the double threshold as core observation points.

[0040] Preferably, the building target enhancement network based on the Transformer attention mechanism in S31 adopts a three-segment structure of feature encoding layer, attention enhancement layer, and feature fusion layer;

[0041] The feature coding layer in S32 extracts multi-dimensional features of SAR images, including grayscale features. Texture features and polarization characteristics Construct the feature matrix , , For SAR image size, For feature dimension and , They represent , and Each dimension contains the number of independent features;

[0042] In S33, the attention enhancement layer adopts a fusion structure of self-attention module and cross-attention module. The self-attention module calculates the spatial correlation weights between features. The formula is as follows:

[0043] ;

[0044] in, These are query, key, and value matrices, respectively. for or The matrix dimension suppresses background interference by focusing on the spatially related regions of the building target through self-attention;

[0045] Cross-attention module integrates GIS vector data, i.e., building outline masks. With optical image features Calculate cross-modal feature association weights as follows:

[0046] ;

[0047] in, For contour mask diagonal matrix, Dimensions of optical image features;

[0048] In S34, the feature fusion layer uses element-wise addition and convolutional dimensionality reduction to output an enhanced building target feature map, which is then subjected to threshold segmentation and morphological processing to obtain the core observation point set of the building.

[0049] Preferably, the specific process of S5 is as follows:

[0050] S51, the time series observation point set of S4 is adopted using 3 The criteria are used to remove outliers and supplement the remaining time-series observation points with structural-geological prior information to generate a fused dataset.

[0051] S52. Based on the constructed temporal deformation inversion fusion model, the Markov chain Monte Carlo (MCMC) algorithm is used to solve for the optimal solution of the posterior probability to obtain the optimal model parameters, and then the building is inverted to obtain the model parameters.

[0052] S53. Based on the optimal model parameters of S52, substitute them into the deformation model expression to invert the three-dimensional deformation and time-series variation curve of the time-series observation points respectively;

[0053] S54. Analyze the deformation characteristics in both time and space dimensions using the three-dimensional deformation and time-series variation curves from S53, and output the deformation inversion results.

[0054] Preferably, the construction process of the time-series deformation inversion fusion model in S52 is as follows:

[0055] Assume the deformation of the building From dynamic linear trend term Nonlinear mutation term and noise items The composition and deformation model expression are as follows:

[0056] ;

[0057] in, For nonlinear basis functions, Follows a normal distribution ;

[0058] To adapt to different deformation stages, a piecewise function definition is adopted. and :

[0059] ;

[0060] in, For time-series segmentation nodes, , For the first The model parameters of the segment;

[0061] Introducing prior constraints from building structural mechanics and geological environment: This includes the foundation bearing capacity. Groundwater level changes As prior information, construct a Bayesian posterior probability model:

[0062] ;

[0063] in, For the model parameter set, It is a proportional symbol. For time-series observation points, It is the likelihood function and follows a normal distribution. For the prior probability distribution, The fusion dataset of structural-geological prior parameters, i.e. .

[0064] Preferably, in S6, a dual-dimensional verification strategy combining point and surface verification is adopted. Point verification selects several benchmark stations and leveling control points evenly distributed within the monitoring area. The inverted three-dimensional deformation is compared with the measured data point by point, and the root mean square error (RMSE), mean absolute error (MAE), and correlation coefficient are calculated. The accuracy of the inversion was quantified using three core indicators.

[0065] Therefore, the building deformation monitoring method based on attention mechanism and collaborative correction described above has the following advantages compared with the prior art:

[0066] This application features high preprocessing accuracy, effectively improving data quality through collaborative correction algorithms, laying the foundation for precise monitoring; accurate target identification, with an attention mechanism that can accurately extract building targets, adapting to complex urban scenarios; and high deformation inversion accuracy, with a fusion model that can adapt to both linear and nonlinear deformations, improving the monitoring accuracy of multiple types of deformations.

[0067] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0068] Figure 1 This is an overall structural diagram of the building deformation monitoring method based on attention mechanism and collaborative correction of the present invention. Detailed Implementation

[0069] In the description of this invention, it should be noted that the terms "upper," "lower," "inner," "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship in which the product of this invention is usually placed when in use. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention.

[0070] Example

[0071] like Figure 1 As shown, the building deformation monitoring method based on attention mechanism and collaborative correction of the present invention includes the following steps:

[0072] S1. Collect multi-source data and synthetic aperture radar (SAR) images of the monitoring area and organize them into an unprocessed time-series SAR image set and auxiliary dataset;

[0073] S11, the SAR image set includes SAR images and accompanying orbital data; the auxiliary dataset includes acquired digital elevation model (DEM) data, acquired medium resolution imaging spectrometer (MODIS) water vapor data, acquired surface meteorological data, and collected geographic information system (GIS) vector data of the monitoring area with a unified coordinate system.

[0074] Data parameters were determined based on the monitoring scenario (dense urban area / mountainous buildings), monitoring accuracy (routine monitoring / high-precision key monitoring), and monitoring cycle (short-term / long-term). For routine large-scale monitoring, Sentinel-1 IW model data (spatial resolution 5m×20m, time baseline 12 days, single polarization / dual polarization) was selected. For high-precision key building monitoring, TerraSAR-XStripMap model data (spatial resolution 1m×1m, time baseline 3-7 days, full polarization) was selected. Simultaneously, the required auxiliary data were defined: 30m resolution DEM, MODIS atmospheric water vapor data (10km spatial resolution, 3-hour time resolution), surface meteorological time-series data (temperature, air pressure, humidity, 1-hour sampling frequency), and GIS vector data.

[0075] Download SAR images and associated orbital data (precise orbital accuracy must reach the centimeter level) through platforms such as ESA SciHub and DLR; obtain DEM data from geospatial data cloud, MODIS water vapor data from NASA EarthData, and surface meteorological data from local meteorological departments; collect 1:2000 scale GIS vector data (including building outlines and administrative divisions) of the monitoring area and unify the coordinate system (WGS84).

[0076] S12. Perform an integrity check on the SAR image data acquired in S11, remove images with excessive cloud cover or corrupted data, and finally form a time-series SAR image set.

[0077] S2. The data of the time-series SAR image set is processed by combining adaptive multi-scale speckle noise suppression and spatiotemporal atmospheric delay model to obtain the processed time-series SAR image set.

[0078] S21. Use SARscape software to complete basic preprocessing, orbit correction, and radiometric calibration.

[0079] S22. Then, the image with the best imaging quality in the time-series SAR image set of S12 is used as the reference image, and the remaining images are images to be registered. The SIFT algorithm of scale invariant feature transformation is used to extract feature points. The FLANN fast nearest neighbor search packet is used to complete coarse matching. Then, the RANSAC algorithm is used to remove erroneous matching points (1000 iterations, inlier threshold of 2 pixels). Finally, bilinear interpolation is used to complete image resampling to ensure that the registration accuracy is no greater than 0.1 pixels.

[0080] S23. First, perform adaptive multi-scale wavelet threshold denoising, and then combine the local signal-to-noise ratio to dynamically adjust the threshold to achieve differentiated noise suppression.

[0081] S231. Perform N-level wavelet decomposition on the time-series SAR image to obtain the low-frequency component and the high-frequency component of each level.

[0082] S232. For each high-frequency component in each layer, calculate the local signal-to-noise ratio (SNR) of each pixel and define an adaptive threshold. The formula is as follows:

[0083] ;

[0084] in, These are the x and y coordinates of the pixels, respectively. The wavelet decomposition level is denoted as . For the first The noise standard deviation of the high-frequency components of the layer. The total number of image pixels. for Local signal-to-noise ratio at that location The adaptive adjustment coefficient and =2.0;

[0085] S233. Subsequently, an improved soft thresholding function is used to process high-frequency components. The specific formula is as follows:

[0086] ;

[0087] In the formula, These are the original high-frequency coefficients. For the processed high-frequency coefficients, For symbolic functions, It is an exponential function. To avoid taking the smallest coefficient value with a denominator of 0 ;

[0088] S24. Construct a spatiotemporal atmospheric delay model with spatial interpolation and temporal fitting, and integrate MODIS water vapor data and surface meteorological data to achieve collaborative correction;

[0089] S241. Based on MODIS water vapor data from S11, the formula for the space atmospheric delay field is constructed using the Kriging interpolation algorithm as follows:

[0090] ;

[0091] in, for Atmospheric delay in space, This is a valid MODIS observation point. For interpolation weights, For error terms, This represents the number of observation points;

[0092] S242. Combining the surface meteorological data from S11, the spatial atmospheric delay field is time-series optimized to construct a spatiotemporal atmospheric delay model:

[0093] ;

[0094] in, It is a time series. for time The time-space atmospheric delay at that location , They are respectively Temperature and air pressure at any given time , , These are the time series fitting coefficients;

[0095] S25. The adaptive multi-scale wavelet threshold denoising process of S23 is iterated in conjunction with the spatiotemporal atmospheric delay model of S24 to achieve coordinated optimization of denoising and correction, and the time-series SAR images are converted into a unified format to output the processed time-series SAR image set.

[0096] S3. Input the image data from the time-series SAR image set processed in S2 into the building target enhancement network based on the attention mechanism, fuse the auxiliary dataset and optical image features, and extract the core observation points of the building targets;

[0097] S31. Construct an attention enhancement network based on Transformer;

[0098] The building target augmentation network based on the Transformer attention mechanism adopts a three-stage structure consisting of a feature encoding layer, an attention enhancement layer, and a feature fusion layer.

[0099] S32. Input the data from the processed temporal SAR image set in S25 into the attention enhancement network trained in S31. The feature encoding layer extracts grayscale features, texture features, and polarization features, and concatenates them to form a feature matrix. ;

[0100] The feature coding layer extracts multi-dimensional features from SAR images, including grayscale features. Texture features and polarization characteristics Construct the feature matrix , , For SAR image size, For feature dimension and , They represent , and Each dimension contains the number of independent features;

[0101] S33. The self-attention module focuses on the spatially related areas of buildings, and the cross-attention module integrates GIS contour masks and optical image features to enhance the matching degree. After 1×1 convolution dimensionality reduction, the Otsu adaptive threshold segmentation method is used to obtain the binary mask of buildings. Then, morphological opening operation is used to denoise and closing operation is used to fill holes, and candidate observation points within the building contour are initially extracted.

[0102] The attention enhancement layer employs a fusion structure of self-attention and cross-attention modules. The self-attention module calculates the spatial correlation weights between features. The formula is as follows:

[0103] ;

[0104] in, These are query, key, and value matrices, respectively. for or The matrix dimension suppresses background interference by focusing on the spatially related regions of the building target through self-attention;

[0105] Cross-attention module integrates GIS vector data, i.e., building outline masks. With optical image features Calculate cross-modal feature association weights as follows:

[0106] ;

[0107] in, For contour mask diagonal matrix, Dimensions of optical image features;

[0108] S34. Calculate the coherence coefficient (coherence coefficient threshold ≥ 0.75) and intensity variance (intensity variance threshold ≤ 0.1) of the candidate observation points initially extracted in S33 in the processed time-series SAR image. Retain the points that meet the double threshold as core observation points and control the density of core observation points to 0.5-1 per square meter.

[0109] The feature fusion layer uses element-wise addition and convolutional dimensionality reduction to output an enhanced building target feature map, which is then subjected to threshold segmentation and morphological processing to obtain the core observation point set of the building.

[0110] S4. Match the core observation points obtained in S3, remove incorrect matching points, and construct a time series observation point set for the building target;

[0111] S5. Input the time series observation point set of S4 into the time series deformation inversion fusion model, combine the structural mechanics parameters and geological environment data, optimize the model parameters through Bayesian inference, invert the three-dimensional deformation amount and time series change curve of the building, analyze the spatiotemporal distribution characteristics and development trend of deformation, identify the deformation anomaly area and output the deformation inversion results.

[0112] S51, the time series observation point set of S4 is adopted using 3 The criteria are used to remove outliers and supplement the remaining time-series observation points with structural-geological prior information to generate a fused dataset.

[0113] S52. Based on the constructed temporal deformation inversion fusion model, the Markov chain Monte Carlo (MCMC) algorithm is used to solve for the optimal solution of the posterior probability to obtain the optimal model parameters, and then the building is inverted to obtain the model parameters.

[0114] The construction process of the temporal deformation inversion fusion model is as follows:

[0115] Assume the deformation of the building From dynamic linear trend term Nonlinear mutation term and noise items The composition and deformation model expression are as follows:

[0116] ;

[0117] in, For nonlinear basis functions, Follows a normal distribution ;

[0118] To adapt to different deformation stages, a piecewise function definition is adopted. and :

[0119] ;

[0120] in, For time-series segmentation nodes, , For the first The model parameters of the segment;

[0121] Introducing prior constraints from building structural mechanics and geological environment: This includes the foundation bearing capacity. Groundwater level changes As prior information, construct a Bayesian posterior probability model:

[0122] ;

[0123] in, For the model parameter set, It is a proportional symbol. For time-series observation points, It is the likelihood function and follows a normal distribution. For the prior probability distribution, The fusion dataset of structural-geological prior parameters, i.e. ;

[0124] S53. Based on the optimal model parameters of S52, substitute them into the deformation model expression to invert the three-dimensional deformation and time-series variation curve of the time-series observation points respectively;

[0125] S54. Analyze the deformation characteristics in both time and space dimensions using the three-dimensional deformation amount and time-series variation curves from S53, and output the deformation inversion results.

[0126] S6. Based on the deformation inversion results of S5, combined with multi-source ground measured data and quantitative verification indicators, the accuracy verification was completed.

[0127] A dual-dimensional verification strategy combining point and area methods was adopted. Point verification selected several benchmark stations and leveling control points evenly distributed within the monitoring area. The inverted three-dimensional deformation was compared with the measured data point by point, and the root mean square error (RMSE), mean absolute error (MAE), and correlation coefficient were calculated. The accuracy of the inversion was quantified using three core indicators.

[0128] Therefore, this invention adopts the building deformation monitoring method based on attention mechanism and collaborative correction described above. Through adaptive multi-scale noise suppression and atmospheric delay collaborative correction, and target enhancement and time-series inversion fusion based on attention mechanism, it effectively solves the problems of low accuracy and poor robustness of traditional SAR deformation monitoring methods in complex scenarios. It achieves high-precision, full-process monitoring of building deformation and can be widely applied in urban operation and maintenance, engineering construction, geological disaster prevention and control and other fields, providing reliable technical support for building safety.

[0129] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A building deformation monitoring method based on attention mechanism and collaborative correction, characterized in that, Includes the following steps: S1. Collect multi-source data and synthetic aperture radar (SAR) images of the monitoring area and organize them into an unprocessed time-series SAR image set and auxiliary dataset; The auxiliary dataset includes the acquired digital elevation model (DEM) data, the acquired medium resolution imaging spectrometer (MODIS) water vapor data, the acquired surface meteorological data, and the collected geographic information system (GIS) vector data of the monitoring area, with a unified coordinate system. S2. The data of the time-series SAR image set is processed by combining the adaptive multi-scale speckle noise suppression and atmospheric delay co-correction algorithm to obtain the processed time-series SAR image set. The specific process of S2 is as follows: S21. Use SARscape software to complete basic preprocessing, orbit correction, and radiometric calibration. S22. Then, the image with the best imaging quality in the time-series SAR image set of S1 is used as the reference image, and the remaining images are images to be registered. The SIFT algorithm of scale invariant feature transformation is used to extract feature points. The FLANN fast nearest neighbor search packet is used to complete coarse matching. Then, the RANSAC algorithm is used to remove erroneous matching points. Finally, the image is resampled by bilinear interpolation. S23. First, perform adaptive multi-scale wavelet threshold denoising, and then combine the local signal-to-noise ratio to dynamically adjust the threshold to achieve differentiated noise suppression. S24. Construct a spatiotemporal atmospheric delay model with spatial interpolation and temporal fitting, and integrate MODIS water vapor data and surface meteorological data to achieve collaborative correction; The specific process of constructing the spatiotemporal atmospheric delay model for spatial interpolation and temporal fitting in S24 is as follows: S241. Based on the MODIS water vapor data from S1, the formula for the spatial atmospheric delay field is constructed using the Kriging interpolation algorithm as follows: ； in, for Atmospheric delay in space, This is a valid MODIS observation point. For interpolation weights, For error terms, This represents the number of observation points; S242. Combining the surface meteorological data from S1, the spatial atmospheric delay field is temporally optimized to construct a spatiotemporal atmospheric delay model: ； in, It is a time series. for time The time-space atmospheric delay at that location , They are respectively Temperature and air pressure at any given time , , These are the time series fitting coefficients; S25. The adaptive multi-scale wavelet threshold denoising process of S23 is iterated in conjunction with the spatiotemporal atmospheric delay model of S24 to achieve coordinated optimization of denoising and correction, and the time-series SAR images are converted into a unified format to output the processed time-series SAR image set. S3. Input the image data from the time-series SAR image set processed in S2 into the building target enhancement network based on the attention mechanism, fuse the data in the auxiliary dataset with the auxiliary features of optical images, and extract the core observation points of the building targets. The specific process of S3 is as follows: S31. Construct an attention enhancement network based on Transformer; S32. Input the data from the processed temporal SAR image set in S25 into the attention enhancement network trained in S31. The feature encoding layer extracts grayscale features, texture features, and polarization features, and concatenates them to form a feature matrix. ; S33. The self-attention module focuses on the spatially related areas of buildings, and the cross-attention module integrates GIS contour masks and optical image features to enhance the matching degree. After 1×1 convolution dimensionality reduction, the Otsu adaptive threshold segmentation method is used to obtain the binary mask of buildings. Then, morphological opening operation is used to denoise and closing operation is used to fill holes, and candidate observation points within the building contour are initially extracted. Cross-attention module integrates GIS vector data, i.e., building outline masks. With optical image features Calculate cross-modal feature association weights as follows: ； in, For contour mask a diagonal matrix; S34. Calculate the coherence coefficient and intensity variance of the candidate observation points initially extracted in S33 in the processed time-series SAR image, and retain the points that meet the double threshold as core observation points. S4. Match the core observation points obtained in S3, remove incorrect matching points, and construct a time series observation point set for the building target; S5. Input the time series observation point set of S4 into the time series deformation inversion fusion model, combine the structural mechanical parameters and geological environment data as prior constraints of the Bayesian inference optimization model parameters, optimize the model parameters through Bayesian inference, invert the three-dimensional deformation amount and time series change curve of the building, analyze the spatiotemporal distribution characteristics and development trend of deformation, identify the deformation anomaly area and output the deformation inversion results. S6. Based on the deformation inversion results of S5, combined with multi-source ground measured data and quantitative verification indicators, the accuracy verification is completed.

2. The building deformation monitoring method based on attention mechanism and collaborative correction according to claim 1, characterized in that: The specific process of S1 is as follows: S11, the SAR image set includes SAR images and accompanying orbital data; S12. Perform an integrity check on the SAR image data acquired in S11, and remove images with excessive cloud cover or corrupted data to finally form a time-series SAR image set.

3. The building deformation monitoring method based on attention mechanism and collaborative correction according to claim 1, characterized in that: The specific process of adaptive multi-scale wavelet threshold denoising in S23 is as follows: S231. Perform N-level wavelet decomposition on the time-series SAR image to obtain the low-frequency component and the high-frequency component of each level. S232. For each high-frequency component in each layer, calculate the local signal-to-noise ratio (SNR) of each pixel and define an adaptive threshold. The formula is as follows: ； in, These are the x and y coordinates of the pixels, respectively. The wavelet decomposition level is denoted as . For the first The noise standard deviation of the high-frequency components of the layer. The total number of image pixels. for Local signal-to-noise ratio at that location This is the adaptive adjustment coefficient; S233. Subsequently, an improved soft thresholding function is used to process high-frequency components. The specific formula is as follows: ； In the formula, These are the original high-frequency coefficients. For the processed high-frequency coefficients, For symbolic functions, It is an exponential function. To avoid extremely small coefficient values ​​where the denominator is 0.

4. The building deformation monitoring method based on attention mechanism and collaborative correction according to claim 1, characterized in that: The building target augmentation network based on the Transformer attention mechanism in S31 adopts a three-stage structure consisting of a feature encoding layer, an attention enhancement layer, and a feature fusion layer. The feature coding layer extracts multi-dimensional features from SAR images, including grayscale features. Texture features and polarization characteristics Construct the feature matrix , , For SAR image size, For feature dimension and , They represent , and Each dimension contains the number of independent features; The attention enhancement layer adopts a fusion structure of self-attention module and cross-attention module. The self-attention module calculates the spatial correlation weights between features. The formula is as follows: ； in, These are query, key, and value matrices, respectively. for or Dimensions of a matrix and The matrix has consistent dimensions, and background interference is suppressed by focusing on the spatially related areas of the building target through self-attention. The feature fusion layer uses element-wise addition and convolutional dimensionality reduction to output an enhanced building target feature map, which is then subjected to threshold segmentation and morphological processing to obtain the core observation point set of the building.

5. The building deformation monitoring method based on attention mechanism and collaborative correction according to claim 1, characterized in that: The specific process of S5 is as follows: S51, the time series observation point set of S4 is adopted using 3 The criteria are used to remove outliers and supplement the remaining time-series observation points with structural-geological prior information to generate a fused dataset. S52. Based on the constructed temporal deformation inversion fusion model, the optimal model parameters are obtained by solving the optimal solution of the posterior probability using the Markov chain Monte Carlo (MCMC) algorithm. S53. Based on the optimal model parameters of S52, substitute them into the deformation model expression to invert the three-dimensional deformation and time-series variation curve of the time-series observation points respectively; S54. Analyze the deformation characteristics in both time and space dimensions using the three-dimensional deformation and time-series variation curves from S53, and output the deformation inversion results.

6. The building deformation monitoring method based on attention mechanism and collaborative correction according to claim 5, characterized in that: The construction process of the time-series deformation inversion fusion model in S52 is as follows: Assume the deformation of the building From dynamic linear trend term Nonlinear mutation term and noise items The composition and deformation model expression are as follows: ； in, For nonlinear basis functions, Follows a normal distribution ; To adapt to different deformation stages, a piecewise function definition is adopted. and : ； in, For time-series segmentation nodes, 、 For the first The model parameters of the segment; Introducing prior constraints from building structural mechanics and geological environment: This includes the foundation bearing capacity. Groundwater level changes As prior information, construct a Bayesian posterior probability model: ； in, For the model parameter set, It is a proportional symbol. For time-series observation points, It is the likelihood function and follows a normal distribution. For the prior probability distribution, The fusion dataset of structural-geological prior parameters, i.e. .

7. The building deformation monitoring method based on attention mechanism and collaborative correction according to claim 1, characterized in that: S6 employs a dual-dimensional verification strategy combining point and area verification. Point verification selects several benchmark stations and leveling control points evenly distributed within the monitoring area. The inverted three-dimensional deformation is compared point by point with the measured data, and the root mean square error (RMSE), mean absolute error (MAE), and correlation coefficient are calculated. The accuracy of the inversion was quantified using three core indicators.

Images