Real-time monitoring method for visualizing ground rebound deformation based on insar technology

Abstract

The application discloses a ground rebound deformation visualization real-time monitoring method based on InSAR technology, and belongs to the technical field of ground monitoring, which comprises the following steps: obtaining ground rebound deformation real-time monitoring multi-source data based on InSAR technology and processing and analyzing the data to extract ground real rebound deformation information; constructing a deep learning ground rebound deformation intelligent prediction model to real-time predict the ground rebound state at the current time and realize real-time monitoring of ground rebound deformation; integrating a 3D stratum model, multi-source remote sensing data and real-time data large screen of early warning levels to realize dynamic monitoring and visualization display of ground rebound deformation and underground water level change. The application solves the problem that existing monitoring data is lagging behind and is difficult to meet the real-time monitoring and immediate early warning requirements. The application can realize dynamic monitoring and visualization display of ground rebound deformation and underground water level change by eliminating monitoring lag through time series InSAR technology and a deep learning model.

Description

Technical Field

[0001] This invention relates to the field of surface monitoring technology, specifically to a method for real-time visualization monitoring of surface rebound deformation based on InSAR technology. Background Technology

[0002] Rising groundwater levels can lead to increased pore water pressure and reduced effective stress, which in turn can cause soil expansion and rebound deformation on the ground surface. Although rebound is different from settlement disaster, uneven rebound deformation may generate additional stress on the foundations of high-speed railways, underground utility tunnels and large buildings, causing structural safety hazards.

[0003] Traditional D-InSAR technology is greatly affected by spatiotemporal decoherence and suffers from monitoring data lag (long satellite revisit cycle and time-consuming processing), making it difficult to meet the needs of real-time monitoring and immediate early warning. Summary of the Invention

[0004] The purpose of this invention is to provide a real-time visualization monitoring method for surface rebound deformation based on InSAR technology. By using time-series InSAR technology and deep learning models to eliminate monitoring lag, dynamic monitoring and visualization of surface rebound deformation and groundwater level changes can be achieved. This provides high-precision technical support for geological disaster early warning, safety assurance of major projects, and resilient city construction, and solves the problems mentioned in the background technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A real-time monitoring method for surface rebound deformation visualization based on InSAR technology includes the following steps:

[0007] S1: Based on InSAR technology, satellite remote sensing data, POD precision orbit data, DEM data and surface auxiliary data are acquired to form multi-source data for real-time monitoring of surface rebound deformation;

[0008] S2: Based on the ENVI-SARscape module, multi-source data of real-time monitoring of surface rebound deformation are processed and analyzed to extract the real rebound deformation information of the surface.

[0009] S3: Construct a deep learning-based intelligent prediction model for surface rebound deformation, predict the amount of surface rebound deformation and groundwater level changes within the time interval of satellite data, predict the surface rebound state at the current moment in real time, and realize real-time monitoring of surface rebound deformation.

[0010] S4: A real-time data dashboard integrating 3D geological models, multi-source remote sensing data, and early warning levels, enabling dynamic monitoring and visualization of surface rebound deformation and groundwater level changes.

[0011] Among them, the nonlinear attenuation weighted adjustment of the water level state characteristic values ​​corresponding to each independent moment in the continuous groundwater level change is based on the hydrogeological consolidation hysteresis compensation factor, generating a corrected water level hysteresis feature matrix with the a priori physical mechanism properties of soil mechanics. The corrected water level hysteresis feature matrix is ​​used to guide the Attention layer to output the accurate time step weight distribution state, eliminating the long-period physical deviation error caused by pure data driving.

[0012] Preferably, the multi-source data for real-time monitoring of surface rebound deformation includes satellite remote sensing data, POD precision orbit data, DEM data, and surface auxiliary data;

[0013] The satellite remote sensing data is used to provide C-band SAR images with high revisit cycles, which is the basis for time-series monitoring of surface rebound deformation.

[0014] The POD precision orbit data is used to correct satellite orbit errors and ensure the geometric accuracy of the interferometry processing;

[0015] The DEM data is used to remove terrain phase and simulate the effect of terrain undulation on the rebound signal;

[0016] The surface auxiliary data is used for auxiliary monitoring of surface rebound deformation, including groundwater level monitoring well data, geological borehole data, and three-dimensional stratigraphic structure data.

[0017] Preferably, the multi-source data of real-time monitoring of surface rebound deformation is processed and analyzed based on the ENVI-SARscape module, and the following operations are performed:

[0018] Preprocessing of multi-source data from real-time monitoring of surface rebound deformation, including multi-view processing, filtering, registration, and interferogram generation;

[0019] Spatiotemporal matching of multi-source data for real-time monitoring of surface rebound deformation is performed, and SAR images from different time phases are accurately registered with POD orbits and DEM data to ensure pixel-level correspondence.

[0020] Temporal InSAR analysis was performed on multi-source data of real-time monitoring of surface rebound deformation. SBAS-InSAR technology was used to extract information on small surface deformations over long time series and to obtain the cumulative deformation and deformation rate fields.

[0021] Preferably, the multi-source data from real-time monitoring of surface rebound deformation are preprocessed, and the following operations are performed:

[0022] The real-time monitoring data of surface rebound deformation from multiple sources is converted into the internal format of SARscape, and satellite remote sensing data and POD precision orbit data are imported. The orbit information is automatically written into the metadata and satellite position error is corrected.

[0023] The pixels of the two SAR images are geometrically aligned, the vertical baselines of the two SAR images are calculated, an initial geometric model is constructed using orbital data for coarse registration, and image feature point matching is used for fine registration. The slave image is resampled onto the geometric grid of the master image to generate the registered slave image SLC data. At this point, the pixels of the master and slave images are in one-to-one correspondence.

[0024] The registered master and slave images are multiplied by their complex conjugates to obtain the amplitude reflecting the characteristics of the ground features and the phase reflecting the slant range difference. When generating the interferogram, the flat phase is simulated using POD fine track data and DEM data and then subtracted to remove the flat phase fringes caused by topographic relief, so that the residual phase reflects the surface deformation and elevation error.

[0025] In the interferogram generation dialog box, the number of views in the range and azimuth directions is directly set to balance the resolution of the range and azimuth directions, making the pixels close to squares. Speckle noise is reduced and the signal-to-noise ratio is improved through averaging. The interferogram is divided into small windows, with the filtering window size being 5x5. The spectrum is smoothed in the frequency domain. The filtering intensity is adaptively adjusted by estimating local coherence, with weak filtering in areas of high coherence and strong filtering in areas of low coherence.

[0026] Preferably, SBAS-InSAR technology is used to extract long-term surface micro-deformation information and obtain cumulative deformation and deformation rate fields, and the following operations are performed:

[0027] Import the Sentinel-1 SLC time series dataset with orbit correction, establish pairing relationships between images, and automatically generate a complex network topology diagram. Each image has at least one in-and-out connection to ensure the continuity of the time series and generate a connection diagram.

[0028] Registration and interferogram generation are performed in batches based on the connection map. The flat terrain effect is removed using the precision track data, the terrain phase is removed using the external DEM, and noise reduction is performed based on multi-view processing and adaptive filtering to improve the signal-to-noise ratio.

[0029] The average coherence coefficient map of all interference pairs is calculated, and a threshold is set so that pixels with an average coherence coefficient higher than the threshold are included in subsequent calculations. Among them, densely built areas and some stable bare soil are retained, while water areas and construction sites with drastic changes are excluded.

[0030] Based on the Delaunay MCF algorithm and using coherence diagrams to guide the unwrapping path, the propagation of phase jump error is prevented, the wrapped phase is restored to the true phase, and the residual orbital error polynomial is fitted using the selected high coherence points to eliminate the systematic phase trend and remove the long-wave error caused by the small error of the POD orbit, preventing it from being misjudged as regional rebound.

[0031] An observation equation is established, and the atmospheric phase is separated by high-pass filtering in the time domain and low-pass filtering in the spatial domain, taking advantage of the low-frequency correlation and temporal uncorrelation of the atmospheric phase. The minimum norm solution is obtained by singular value decomposition or least squares method, and the linear deformation rate field and DEM residual are obtained.

[0032] Preferably, a deep learning-based intelligent prediction model for surface rebound deformation is constructed to predict the amount of surface rebound deformation and groundwater level changes within the satellite data time interval, and the following operations are performed:

[0033] Based on historical monitoring data, groundwater level change and surface deformation rate were selected as core characteristic factors.

[0034] A related hybrid deep learning algorithm is used to train a time series prediction model to predict the amount of surface rebound deformation and groundwater level change within the time interval of satellite data.

[0035] By using the Long Short-Term Memory (LSTM) network to capture the lag effect in time series data and combining it with an attention mechanism to dynamically allocate the weights of different influencing factors, we can achieve accurate prediction of deformation values ​​at future moments.

[0036] By continuously optimizing model parameters and improving prediction accuracy using newly acquired satellite data, a closed loop of monitoring, prediction, and correction can be achieved, thereby eliminating the monitoring lag problem caused by the time difference of satellite data.

[0037] Preferably, the construction of a deep learning-based intelligent prediction model for surface rebound deformation further includes:

[0038] High-frequency groundwater level data are unified to the InSAR time point through moving average or interpolation methods, and InSAR monitoring points are spatially correlated with the data from the nearest groundwater monitoring well.

[0039] Core features including groundwater level change, historical surface deformation rate and cumulative deformation were selected, as well as auxiliary surface features including stratigraphic lithology parameters, rainfall and POD orbital correction residuals. Since there is a physical time lag between surface rebound and groundwater level rise, a sliding window was introduced when constructing features to use groundwater level changes at continuous time intervals as input to capture the lag response.

[0040] The model employs an Attention-LSTM hybrid architecture, receives multivariate time series data, automatically calculates the weight distribution of different time steps and features based on the Attention layer to highlight key influencing factors, extracts time series features based on the LSTM layer, and maps the high-dimensional features extracted by the LSTM to deformation prediction values, outputting the amount of surface rebound in the future time.

[0041] Preferably, to achieve dynamic monitoring and visualization of surface rebound deformation and groundwater level changes, the following operations are performed:

[0042] Based on pre-processed multi-source data of real-time monitoring of surface rebound deformation and intelligent prediction model of surface rebound deformation, a large screen of real-time data of groundwater level rebound and surface rebound deformation was developed.

[0043] By integrating geological data, a 3D stratigraphic model is constructed to clearly display stratigraphic features; a data fusion and visualization module is developed to enable the linked display of the 3D stratigraphic model with multi-source data and early warning levels, and supports interactive functions such as multi-scale zooming, regional filtering, and single-point query; and the data rendering algorithm is optimized to ensure smooth operation of the large screen under massive amounts of data.

[0044] Preferably, when constructing the features, a sliding window is introduced. After taking the continuous changes in groundwater level as input to capture the hysteresis response, the following processing operations are included:

[0045] Obtain the equivalent thickness of the cohesive soil layer and the vertical permeability coefficient corresponding to the InSAR monitoring points contained in the multi-source data;

[0046] Combining the preset time span value corresponding to the sliding window, and the water level state characteristic values ​​corresponding to each independent moment in the continuous groundwater level change, the hydrogeological consolidation hysteresis compensation factor corresponding to each independent moment is calculated. The calculation formula is as follows:

[0047] ;

[0048] In the formula, This represents the hydrogeological consolidation hysteresis compensation factor; This represents the pre-defined shape fitting parameters; This represents the equivalent thickness of the cohesive soil layer. This represents the numerical value of the vertical permeability coefficient; This represents the preset time span value; Represents the exponential function with the natural constant as the base; The logarithmic function is defined with the natural constant as its base. Represents the natural constant; This represents a pre-defined hydrological scaling constant; This represents the water level status characteristic value corresponding to the independent time point;

[0049] Finally, based on the calculated hydrogeological consolidation lag compensation factor, the water level state feature values ​​corresponding to each independent time moment in the continuous groundwater level changes input into the Attention-LSTM hybrid model architecture are nonlinearly attenuated and weighted to generate a corrected water level lag feature matrix with prior physical mechanism properties of soil mechanics. The corrected water level lag feature matrix is ​​used to guide the Attention layer to output an accurate time step weight distribution state, eliminating the long-period physical deviation error caused by pure data-driven approaches.

[0050] Preferably, the strong filtering operation for regions with low coherence includes the following constraint processing steps:

[0051] The temporal coherence fluctuation variance contained in each interferogram generated during the preprocessing operation of multi-source data for real-time monitoring of surface rebound deformation is obtained, and the capillary rise height value corresponding to each interferogram is obtained.

[0052] Determine whether the temporal coherence fluctuation variance and the capillary rise height value synchronously exceed a preset surge judgment threshold;

[0053] When it is determined that the temporal coherence fluctuation variance and the capillary rise height value simultaneously exceed the surge determination threshold, the target interference window corresponding to the surge determination threshold is marked as an abnormal window caused by signal masking due to sudden changes in soil moisture.

[0054] For the abnormal window, the default logical constraint of performing strong filtering when the coherence is low is removed, and the entangled phase gradient vector corresponding to the pixel point whose coherence meets the requirements at the edge of the abnormal window is extracted.

[0055] The spatial topological phase reconstruction compensation operation is performed on the low coherence pixels inside the abnormal window using the entangled phase gradient vector, in order to replace the default strong filtering process and prevent the real small surface rebound initial deformation signal caused by the sudden increase in surface water content from being excessively smoothed out.

[0056] When it is determined that the variance of temporal coherence fluctuation and the value of capillary rise height do not synchronously exceed the surge determination threshold, the operating principle of maintaining weak filtering in areas with high coherence and strong filtering in areas with low coherence is maintained.

[0057] Compared with the prior art, the beneficial effects of the present invention are:

[0058] This invention focuses on satellite remote sensing data, relying on the ENVI-SARscape module to complete multi-source data preprocessing and spatiotemporal matching, and using time-series InSAR technology and deep learning models to eliminate monitoring lag. It develops a real-time data dashboard that integrates 3D geological models, multi-source remote sensing data, and early warning levels, enabling dynamic monitoring and visualization of surface rebound deformation and groundwater level changes, and providing high-precision technical support for geological disaster early warning, safety assurance of major projects, and resilient city construction. Attached Figure Description

[0059] Figure 1 This is a flowchart of the real-time monitoring method for surface rebound deformation visualization based on InSAR technology according to the present invention. Detailed Implementation

[0060] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] To address the issues of existing D-InSAR technology being significantly affected by spatiotemporal decoherence and suffering from data lag (long satellite revisit cycles and time-consuming processing), thus failing to meet the needs of real-time monitoring and immediate early warning, please refer to [link to relevant documentation]. Figure 1 This embodiment provides the following technical solution:

[0062] A real-time monitoring method for surface rebound deformation visualization based on InSAR technology includes:

[0063] Real-time monitoring data of surface rebound deformation was acquired using InSAR technology. The data was then processed and analyzed using the ENVI-SARscape module to extract the true rebound deformation information of the surface.

[0064] In this embodiment, the multi-source data for real-time monitoring of surface rebound deformation includes satellite remote sensing data, POD precision orbit data, DEM data, and surface auxiliary data;

[0065] The satellite remote sensing data (Sentinel 1 radar imagery) is used to provide C-band SAR imagery with a high revisit period (12 days / 6 days), which is the basis for time-series monitoring of surface rebound deformation.

[0066] The POD precision orbit data is used to correct satellite orbit errors and ensure the geometric accuracy of the interferometry processing;

[0067] The DEM data (such as SRTM data / TanDEM-X data) is used to remove terrain phase and simulate the effect of terrain undulation on rebound signal;

[0068] The surface auxiliary data is used for auxiliary monitoring of surface rebound deformation, including groundwater level monitoring well data, geological borehole data, and three-dimensional stratigraphic structure data.

[0069] In this embodiment, the multi-source data of real-time monitoring of surface rebound deformation is processed and analyzed based on the ENVI-SARscape module, and the following operations are performed:

[0070] Preprocessing of multi-source data from real-time monitoring of surface rebound deformation includes multi-view processing, filtering (removing speckle noise), registration, and interferogram generation;

[0071] Spatiotemporal matching of multi-source data for real-time monitoring of surface rebound deformation is performed, and SAR images from different time phases are accurately registered with POD orbits and DEM data to ensure pixel-level correspondence.

[0072] Temporal InSAR analysis was performed on multi-source data of real-time monitoring of surface rebound deformation. SBAS-InSAR (small baseline set technique) was used to extract information on small surface deformation over long time series and to obtain the cumulative deformation and deformation rate fields.

[0073] In this embodiment, the multi-source data for real-time monitoring of surface rebound deformation are preprocessed by performing the following operations:

[0074] The real-time monitoring data of surface rebound deformation from multiple sources is converted into the internal format of SARscape, and satellite remote sensing data and POD precision orbit data are imported. The orbit information is automatically written into the metadata and satellite position error is corrected.

[0075] The pixels of the two SAR images are geometrically aligned, the vertical baselines of the two SAR images are calculated, an initial geometric model is constructed using orbital data for coarse registration, and image feature point matching is used for fine registration. The slave image is resampled onto the geometric grid of the master image to generate the registered slave image SLC data. At this point, the pixels of the master and slave images are in one-to-one correspondence.

[0076] The registered master and slave images are multiplied by their complex conjugates to obtain the amplitude reflecting the characteristics of the ground features and the phase reflecting the slant range difference. When generating the interferogram, the flat phase is simulated using POD fine track data and DEM data and then subtracted to remove the flat phase fringes caused by topographic relief, so that the residual phase reflects the surface deformation and elevation error.

[0077] In the interferogram generation dialog box, the number of views in the range and azimuth directions is directly set to balance the resolution of the range and azimuth directions, making the pixels close to squares. Speckle noise is reduced and the signal-to-noise ratio is improved through averaging. The interferogram is divided into small windows, with the filtering window size being 5x5. The spectrum is smoothed in the frequency domain. The filtering intensity is adaptively adjusted by estimating local coherence, with weak filtering in areas of high coherence and strong filtering in areas of low coherence.

[0078] In this embodiment, SBAS-InSAR technology is used to extract long-term surface micro-deformation information and obtain cumulative deformation and deformation rate fields, and the following operations are performed:

[0079] Import the Sentinel-1 SLC time series dataset with orbit correction, establish pairing relationships between images, and automatically generate a complex network topology diagram. Each image has at least one in-and-out connection to ensure the continuity of the time series and generate a connection diagram.

[0080] Registration and interferogram generation are performed in batches based on the connection map. The flat terrain effect is removed using the precision track data, the terrain phase is removed using the external DEM, and noise reduction is performed based on multi-view processing and adaptive filtering to improve the signal-to-noise ratio.

[0081] The average coherence coefficient map of all interference pairs is calculated, and a threshold is set so that pixels with an average coherence coefficient higher than the threshold are included in subsequent calculations. Among them, densely built areas and some stable bare soil are retained, while water areas and construction sites with drastic changes are excluded.

[0082] Based on the Delaunay MCF algorithm and using coherence diagrams to guide the unwrapping path, the propagation of phase jump error is prevented, the wrapped phase is restored to the true phase, and the residual orbital error polynomial is fitted using the selected high coherence points to eliminate the systematic phase trend and remove the long-wave error caused by the small error of the POD orbit, preventing it from being misjudged as regional rebound.

[0083] An observation equation is established, and the atmospheric phase is separated by high-pass filtering in the time domain and low-pass filtering in the spatial domain, taking advantage of the low-frequency correlation and temporal uncorrelation of the atmospheric phase. The minimum norm solution is obtained by singular value decomposition or least squares method, and the linear deformation rate field and DEM residual are obtained.

[0084] A deep learning-based intelligent prediction model for surface rebound deformation is constructed to predict the amount of surface rebound deformation and groundwater level changes within the time interval of satellite data. The model predicts the surface rebound state at the current moment in real time, determines the intelligent prediction result of surface rebound deformation, and realizes real-time monitoring of surface rebound deformation.

[0085] In this embodiment, a deep learning-based intelligent prediction model for surface rebound deformation is constructed to predict the amount of surface rebound deformation and groundwater level changes within the satellite data time interval, and the following operations are performed:

[0086] Based on historical monitoring data, groundwater level change and surface deformation rate were selected as core characteristic factors.

[0087] A related hybrid deep learning algorithm is used to train a time series prediction model to predict the amount of surface rebound deformation and groundwater level change within the time interval of satellite data.

[0088] By using the Long Short-Term Memory (LSTM) network to capture the lag effect in time series data and combining it with an attention mechanism to dynamically allocate the weights of different influencing factors, we can achieve accurate prediction of deformation values ​​at future moments.

[0089] By continuously optimizing model parameters and improving prediction accuracy using newly acquired satellite data, a closed loop of monitoring, prediction, and correction can be achieved, thereby eliminating the monitoring lag problem caused by the time difference of satellite data.

[0090] Specifically, InSAR monitoring results often lag behind the actual occurrence time (satellite transit period + data processing time). Therefore, a deep learning-based intelligent prediction model for surface rebound deformation is constructed. This model utilizes a Long Short-Term Memory (LSTM) network to capture the lag effect in the time series and combines an attention mechanism to dynamically allocate the weights of different influencing factors, thereby achieving accurate prediction of deformation values ​​at future moments. By learning the nonlinear lag relationship between groundwater level changes and surface rebound response, the model can predict the surface rebound state at the current moment based on current groundwater level data, filling the data gap during satellite revisit periods and achieving real-time monitoring.

[0091] In this embodiment, constructing a deep learning-based intelligent prediction model for surface rebound deformation further includes:

[0092] High-frequency groundwater level data are unified to the InSAR time point through moving average or interpolation methods, and InSAR monitoring points are spatially correlated with the data from the nearest groundwater monitoring well.

[0093] Core features including groundwater level change, historical surface deformation rate and cumulative deformation were selected, as well as auxiliary surface features including stratigraphic lithology parameters, rainfall and POD orbital correction residuals. Since there is a physical time lag between surface rebound and groundwater level rise, a sliding window was introduced when constructing features to use groundwater level changes at continuous time intervals as input to capture the lag response.

[0094] The model employs an Attention-LSTM hybrid architecture, receives multivariate time series data, automatically calculates the weight distribution of different time steps and features based on the Attention layer to highlight key influencing factors, extracts time series features based on the LSTM layer, and maps the high-dimensional features extracted by the LSTM to deformation prediction values, outputting the amount of surface rebound in the future time.

[0095] Specifically, the development of an intelligent prediction model integrating physical mechanisms and deep learning addresses the 21-day time difference issue in satellite data (especially POD precision orbit data). A deep learning-based intelligent prediction model for surface rebound deformation is constructed. First, based on historical monitoring data, groundwater level changes and surface deformation rates are selected as core feature factors. Second, a related hybrid deep learning algorithm is used to train the time-series prediction model to predict surface rebound deformation and groundwater level changes within the satellite data time intervals. Finally, newly acquired satellite data is used to continuously optimize the model parameters, minimizing the monitoring lag caused by the satellite data time difference.

[0096] A large real-time data dashboard integrating 3D geological models, multi-source remote sensing data, and early warning levels enables dynamic monitoring and visualization of surface rebound deformation and groundwater level changes.

[0097] In this embodiment, dynamic monitoring and visualization of surface rebound deformation and groundwater level changes are achieved by performing the following operations:

[0098] Based on pre-processed multi-source data of real-time monitoring of surface rebound deformation and intelligent prediction model of surface rebound deformation, a large screen of real-time data of groundwater level rebound and surface rebound deformation was developed.

[0099] By integrating geological data, a 3D stratigraphic model is constructed to clearly display stratigraphic characteristics. In particular, a true three-dimensional geological model is constructed using borehole data to display the lithology and physical and mechanical properties of soil layers at different depths in three-dimensional space, and to intuitively present the main strata where rebound occurs.

[0100] Develop a data fusion and visualization module to enable the linked display of 3D geological models, multi-source data, and early warning levels, and support interactive functions such as multi-scale zooming, regional filtering, and single-point query; and optimize the data rendering algorithm to ensure smooth operation of the large screen under massive data.

[0101] Specifically, when displaying multi-source data fusion, an InSAR deformation rate color map is overlaid on the surface layer, visually showing the distribution of rebound areas through color depth; the underground layer displays the changes in groundwater level isosurfaces in three dimensions, simulating the water level rise process through animation; clicking on any location simultaneously pops up the surface rebound curve and the groundwater level change curve, visually showing the correlation and hysteresis effect between the two through a dual Y-axis chart, and setting thresholds based on rebound rate and cumulative rebound amount to divide into four levels of early warning: blue, yellow, orange, and red. Combined with the location of major projects, an engineering safety risk heat map is generated to automatically identify high-risk areas of uneven rebound.

[0102] In summary, focusing on satellite remote sensing data, relying on the ENVI-SARscape module to complete multi-source data preprocessing and spatiotemporal matching, and using time-series InSAR technology and deep learning models to eliminate monitoring lag, a real-time data dashboard integrating 3D geological models, multi-source remote sensing data, and early warning levels was developed. This enables dynamic monitoring and visualization of surface rebound deformation and groundwater level changes, providing high-precision technical support for geological disaster early warning, safety assurance of major projects, and resilient city construction.

[0103] In this embodiment, a sliding window is introduced when constructing features. After taking the continuous changes in groundwater level as input to capture the hysteresis response, the following processing operations are included:

[0104] Obtain the equivalent thickness of the cohesive soil layer and the vertical permeability coefficient corresponding to the InSAR monitoring points contained in the multi-source data;

[0105] Combining the preset time span value corresponding to the sliding window, and the water level state characteristic values ​​corresponding to each independent moment in the continuous groundwater level change, the hydrogeological consolidation hysteresis compensation factor corresponding to each independent moment is calculated. The calculation formula is as follows:

[0106] ;

[0107] In the formula, This represents the hydrogeological consolidation hysteresis compensation factor; This represents the pre-defined shape fitting parameters; This represents the equivalent thickness of the cohesive soil layer. This represents the numerical value of the vertical permeability coefficient; This represents the preset time span value; Represents the exponential function with the natural constant as the base; The logarithmic function is defined with the natural constant as its base. Represents the natural constant; This represents a pre-defined hydrological scaling constant; This represents the water level status characteristic value corresponding to the independent time point;

[0108] Finally, based on the calculated hydrogeological consolidation lag compensation factor, the water level state feature values ​​corresponding to each independent time moment in the continuous groundwater level changes input into the Attention-LSTM hybrid model architecture are nonlinearly attenuated and weighted to generate a corrected water level lag feature matrix with prior physical mechanism properties of soil mechanics. The corrected water level lag feature matrix is ​​used to guide the Attention layer to output an accurate time step weight distribution state, eliminating the long-period physical deviation error caused by pure data-driven approaches.

[0109] The working principle and beneficial effects of the above technical solution are as follows:

[0110] It is understandable that this is for obtaining the equivalent thickness value of cohesive soil layers. With vertical permeability coefficient The processing steps involve a numerical mapping relationship between surface rebound deformation and underground pore water pressure dissipation parameters. The data parsing component retrieves the regional geological borehole database via a network communication interface. The system extracts a three-dimensional spatial coordinate matrix containing known exploration points and corresponding stratigraphic thickness attribute column vectors from the regional geological borehole database. The system calls the spatial interpolation algorithm module to perform regional thickness attribute estimation. The system constructs a spatial distance tensor and iteratively calculates the spatial Euclidean distance scalar between the target synthetic aperture radar monitoring point and the known exploration points. For calculating the spatial Euclidean distance scalar, the system extracts the longitude coordinates of the first and last exploration points, performs subtraction, and calculates the square value; it also extracts the latitude coordinates of the first and last exploration points, performs subtraction, and calculates the square value; and it extracts the elevation coordinates of the first and last exploration points, performs subtraction, and calculates the square value. These three sets of squared values ​​are then added together, and the arithmetic square root function is called to calculate the spatial Euclidean distance scalar. The system sets a spatial lag step size parameter, dividing all calculated spatial Euclidean distance scalars into corresponding distance discrete intervals. The system traverses each distance discrete interval, extracts all combinations of double-ended exploration points within the corresponding interval, and calculates the difference in formation thickness attribute corresponding to the double-ended exploration points. The difference in formation thickness is squared, and all the squared results within the interval are summed. This summation is then divided by the total logarithm of the combination of exploration points at both ends within the interval to calculate the experimental semivariogram. The experimental semivariogram values ​​corresponding to all discrete distance intervals are then integrated to generate a discrete set of experimental semivariogram points.

[0111] Furthermore, the system invokes a spherical fitting model to perform nonlinear parameter fitting on the experimental semivariance discrete point set. The system uses an iterative method to update the fitting parameters, solving for the nugget constant, sill constant, and range constant. Using the mathematical model of the spherical semivariance function output from the fitting, the system calculates the spatial autovariance between each known exploration point, constructing a spatial autovariance matrix. Simultaneously, the system calculates the spatial covariance between the target synthetic aperture radar (SAR) monitoring point and each known exploration point, generating a spatial covariance column vector. The system introduces Lagrange multipliers to construct an unbiased estimation constraint linear equation system. The system performs matrix inversion on the spatial autovariance matrix, multiplies the inverted matrix with the spatial covariance column vector, and solves for the weight coefficient column vector corresponding to the target SAR monitoring point. After transposing the weight coefficient column vector, the system performs a vector inner product operation with the stratigraphic thickness attribute column vector of the known exploration points, deriving and outputting the three-dimensional stratigraphic columnar distribution state matrix directly below the target SAR monitoring point. The system sets a physical threshold for permeability coefficient calibration and performs layer-by-layer analysis on the three-dimensional stratigraphic columnar distribution matrix. The system removes data from coarse sand and gravel layers with permeability coefficients exceeding the calibration threshold, retaining the physical thickness values ​​of geological layers such as silt and silty clay. The system reads the initial void ratio parameter for the corresponding retained geological layers, divides the physical thickness value by the initial void ratio parameter, and performs a hydraulic impedance conversion operation. Finally, the converted thickness values ​​of all retained geological layers are summed using a scalar algorithm to output the equivalent thickness value of the cohesive soil layer. Equivalent thickness of cohesive soil layer It has the standard absolute length dimension.

[0112] For obtaining the vertical permeability coefficient value The processing flow involves the data parsing component receiving sensor datasets from multi-stage variable head pumping tests conducted within in-situ exploration wells. These datasets include time-series arrays of dynamic drawdown amplitudes within aquifers at different depths, as well as discrete vectors of rate changes indicating the return of the confined head to its initial equilibrium state after pumping ceases. The system extracts the dynamic drawdown amplitude time-series array and the discrete vectors of rate changes, importing them into the unsteady flow analytical formula calculation module. Based on unsteady seepage theory, the system constructs a set of nonlinear partial differential equations for seepage control, including well function parameters and reservoir coefficient parameters. The system sets initial inference tensors for hydraulic conductivity and reservoir coefficient. It then uses the nonlinear least squares method to iteratively solve the residuals of the nonlinear partial differential equations for seepage control through multiple iterations. Within each independent iterative loop step, the system calculates the residual column vector between the measured dynamic drawdown amplitude and the theoretically predicted drawdown amplitude. Finally, the system calculates the partial derivative matrices of the residual column vector with respect to the inference tensors for hydraulic conductivity and reservoir coefficient. The system generates an approximate Hessian matrix by multiplying the transpose of the partial derivative matrix by the partial derivative matrix itself. The system then solves for the inverse of the approximate Hessian matrix and multiplies it by the transpose of the partial derivative matrix and the residual column vector to calculate the parameter update step vector. This parameter update step vector is then superimposed onto the existing hydroconductivity and water storage coefficient prediction tensors. Iterative calculations continue until the L2 norm of the residual column vector falls below the set convergence standard parameter. After convergence, the system extracts the obtained hydroconductivity prediction tensor. The system reads the aquifer thickness parameter corresponding to the in-situ exploration well, divides the obtained hydroconductivity prediction tensor by the aquifer thickness parameter, and inversely outputs the water flow conductivity index data within a specific cross-sectional area in the vertical direction. The system assigns the aforementioned water flow conductivity index data to the vertical permeability coefficient value. Vertical permeability coefficient value The constitutive relation that maps pore water pressure dissipation over time has a composite dimensional property of the ratio of length dimension to time dimension.

[0113] Specifically, this is combined with the preset time span value corresponding to the sliding window. And the water level state characteristic values ​​at each independent moment in the continuous changes of groundwater level. The system performs feature discretization and time window alignment operations. The microporous network within the formation causes frictional resistance to water flow, resulting in a lag in the conversion of pore water pressure into effective stress in the soil skeleton. The system incorporates a dynamic sliding window extraction mechanism during data preprocessing. The system sets a preset time span value. Set the preset time span value The revisit period parameters of the synthetic aperture radar (SAR) satellite Earth observations are kept identical, and synchronous numerical corrections are performed using the mean dissipation time constant of soil consolidation extracted from historical hydrological data. The system aligns the time step of the hybrid deep learning model to the sampling frequency of the radar interferogram sequence, ensuring the temporal synchronization of the model's input sequence. Preset time span values... It has a time dimension.

[0114] Furthermore, for the water level state characteristic values ​​corresponding to each independent time point... The extraction step involves continuously acquiring the aquifer head height state parameter of the aquifer by deploying a piezoresistive water level sensor array at the depth of the target underground aquifer. The sensor array outputs a time-series sequence of raw discrete analog level signals. This time-series sequence is transmitted to an analog-to-digital converter module and converted into a high-frequency digital discrete coded sequence. The high-frequency digital discrete coded sequence is then imported into the built-in adaptive filtering and noise reduction matrix operation module. During the initialization phase, the algorithm module sets the state transition matrix and the system process evolution noise covariance matrix. Within the discrete time step progressive loop, the algorithm module executes prediction inference and update correction instructions. In the prediction inference phase, the system multiplies the posterior optimal state estimation vector of the previous historical time step by the state transition matrix to derive the prior state estimation vector of the current independent time step; the system multiplies the state transition matrix by the posterior error covariance matrix of the previous historical time step, then multiplies it by the transpose of the state transition matrix, and superimposes it with the system process evolution noise covariance matrix to generate the prior error covariance matrix of the current independent time step. In the update and correction phase, the system uses the observation matrix, prior error covariance matrix, and measurement noise covariance matrix to solve for the filter gain matrix. The system subtracts the product of the observation matrix and the prior state estimation vector from the actual observed signal vector, outputting the state observation residual vector. The system multiplies the filter gain matrix by the state observation residual vector and adds the resulting product to the prior state estimation vector, completing the update of the posterior optimal state estimation vector for the current independent time step. After filtering and smoothing, a continuous sequence of water level changes reflecting the continuous fluctuations of the underground physical water body is output. The system operates according to a preset time span. By using a specific step size, the continuous sequence of water level changes is discretely sampled and truncated to generate water level state characteristic values ​​for each independent time point. Water level characteristics at independent moments. It represents the absolute change in water head pressure at a corresponding time point, and has the dimension of length.

[0115] Understandably, the calculation module utilizes the obtained equivalent thickness value of the cohesive soil layer. and vertical permeability coefficient values The hydrogeological consolidation hysteresis compensation factor corresponding to each independent time point was derived through calculation formula. .

[0116] The algebraic structure of the above calculation formula contains a product operation of two sets of sub-modules. The first term consists of components expressed as natural constants. Exponential function with base Module; the sub-term is composed of natural constants. Logarithmic function with base as Module. Dimensionless shape fitting parameters Pre-set adjustment constants; dimensionless hydrological scaling constants This is a pre-set proportional adjustment constant.

[0117] Specifically, the arithmetic logic operation module executes the formula algebra solution process. The system reads the vertical permeability coefficient value, which has a composite dimension of length and time. And preset time span values ​​with time units. The system imports the two sets of values ​​into the multiplication calculator and performs a multiplication operation. Based on the rules of physical units, the time dimensions cancel each other out, and the multiplication operation outputs a scalar value of the theoretical permeability path with a length dimension. The system then reads the equivalent thickness of the cohesive soil layer with a length dimension. As the numerator operand, the theoretical penetration path scalar value is used as the denominator operand, and a division algebraic mapping operation is performed. The spatial length dimensions cancel each other out, and the division algebraic mapping operation outputs a dimensionless thickness penetration ratio feature. The system reads the dimensionless shape fitting parameters. After being converted to a negative state, it undergoes a multiplication operation with the dimensionless thickness penetration ratio characteristic, generating a decaying independent variable for the exponential function. The system calls the series expansion polynomial mathematics module, executing the operation using the natural constant. Exponential function with base The solver command calculates and outputs the floating-point value of the physical hysteresis decay constant. Dimensionless shape fitting parameters. The loss function is calibrated and updated based on local historical measured deformation data through backpropagation.

[0118] Furthermore, the system reads the water level state characteristic values ​​corresponding to independent time points with length dimensions. As a constant input as the dividend, the equivalent thickness of the cohesive soil layer, which also has the dimension of length, is read. As a divisor input constant, the system issues a floating-point division instruction, canceling out the length dimensions. The division instruction outputs the dimensionless physical state ratio constant node value. The system reads the dimensionless hydrological scaling constant. The system performs a multiplication and superposition operation on the node values ​​of the dimensionless physical state ratio constant to generate an intermediate temporary register scalar. The system then modifies this intermediate temporary register scalar with the natural constant. The system executes an addition summation constant offset instruction to generate a tensor representing the argument constants of the logarithmic operation function. The system then calls the mathematical operations library to execute the operation using natural constants. Logarithmic function with base as The mapping solution calculates and outputs the floating-point value of the hydraulic load rebound deformation scaling constant. The system performs a multiplication operation on the floating-point value of the physical hysteresis attenuation constant and the floating-point value of the hydraulic load rebound deformation scaling constant, and outputs the hydrogeological consolidation hysteresis compensation factor defined under the absolute dimensionless calibration state. .

[0119] Understandably, the system is based on hydrogeological consolidation lag compensation factors. The control variables within the data flow system of the hybrid model architecture are modified. The system memory management unit stores all water level status feature values ​​corresponding to independent time points. The system encapsulates a multidimensional temporal feature input tensor data stream according to temporal encoding rules. This multidimensional temporal feature input tensor data stream is then transmitted to the Long Short-Term Memory (LSTM) network hierarchical architecture for forward sequence inference. For each independent time step iteration of the LTM network hidden layer state transition, the system extracts the water level state feature value corresponding to the current independent time step. And the calculated hydrogeological consolidation hysteresis compensation factor The system will incorporate hydrogeological consolidation lag compensation factors. The matrix is ​​expanded and replicated into a compensation factor bias tensor matrix of equal dimension. The system utilizes the compensation factor bias tensor matrix to represent the water level state characteristic values ​​at independent time points. The resulting state tensor undergoes nonlinear decay weighted adjustment. The system calls the built-in tensor product function to perform element-wise multiplication, generating a corrected water level lag feature matrix carrying the soil mechanical decay characteristics. The system inputs the corrected water level lag feature matrix into the attention mechanism layer, guiding the attention mechanism layer to output the time-step weight distribution state.

[0120] Specifically, within the computation cycle of the hidden layer cell unit of the Long Short-Term Memory Network, the system extracts the hidden layer state feedback tensor of the previous historical time step and the water level state feature value corresponding to the current independent time step. The system constructs a temporal input feature tensor. It performs a horizontal concatenation operation on the spatial scale between the hidden layer state feedback tensor and the temporal input feature tensor to generate a joint feedforward feature column vector array. The system extracts the static connection weight matrices of the forget gate, input gate, and output gate obtained during model training. It then performs a spatial linear inner product projection operation on the joint feedforward feature column vector array and the static connection weight matrix of the forget gate, and superimposes the forget gate-specific bias constant vector to generate the initial forgotten feature tensor. The system adjusts the initial forgotten feature tensor and the initial input feature tensor using the compensation factor bias tensor matrix. Finally, the system performs a tensor product dot product operation on the initial forgotten feature tensor and the compensation factor bias tensor matrix to generate an intermediate forgotten feature tensor; and performs a tensor product dot product operation on the initial input feature tensor and the mathematical inverse matrix of the compensation factor bias tensor matrix to generate an intermediate input feature tensor. The system calls the nonlinear activation function module to map the intermediate forgetting feature tensor, the intermediate input feature tensor, and the initial output feature tensor, and outputs the forgetting gate control filtering probability matrix, the input gate control filtering probability matrix, and the output gate control probability matrix, respectively.

[0121] Furthermore, the system performs an inner product projection operation on the feedforward feature column vector array and the cell state connection weight matrix, adds a cell state bias constant fine-tuning term, and then inputs it into the hyperbolic tangent nonlinear function module for mapping and scaling to generate a tensor of incremental feature parameters of candidate neuron state content. The system calls the array multiplication algorithm interface, using the forget gate to control the filtering probability matrix and perform an element-wise multiplication algebraic filtering operation on the global backbone memory cell state feature tensor passed from the previous historical time step; it also uses the input gate to control the filtering probability matrix and perform an element-wise multiplication algebraic activation operation on the tensor of incremental feature parameters of candidate neuron state content. The system performs parallel matrix addition to sum the tensor results output from the two multiplication operations, completing the iterative update of the neuron state feature tensor. Based on the updated neuron state feature tensor and the output gate control probability matrix, the system generates the hidden layer feedback state tensor for the current independent time step.

[0122] Specifically, the system concatenates the entire sequence of hidden layer feedback state tensors output cyclically at consecutive time steps to generate a three-dimensional state tensor matrix, which is then transmitted to the adaptive attention mechanism parameter configuration mapping layer module. This module pre-deploys a parameter set including the query feature projection transformation parameter weight matrix, the key index feature projection transformation parameter weight matrix, and the value-based physical feature projection transformation parameter weight matrix. The system performs spatial linear multiplication transformation mapping inner product dot product operations on the three-dimensional state tensor matrix with the query feature projection transformation parameter weight matrix, the key index feature projection transformation parameter weight matrix, and the value-based physical feature projection transformation parameter weight matrix, respectively, generating a multi-dimensional query mapping feature column vector matrix, a multi-dimensional key index feature comparison column vector matrix, and a multi-dimensional value-based physical feature expression column vector matrix.

[0123] The system transposes the multi-dimensional key index feature comparison column vector matrix and inputs it into a dot product inner product mathematical calculation array along with the multi-dimensional query mapping feature column vector matrix to generate a cross-correlation scoring evaluation matrix. The system then uses the arithmetic square root constant of the feature dimensions to perform scaling operations on the cross-correlation scoring evaluation matrix, generating a scaled evaluation scoring matrix. This scaled evaluation scoring matrix is ​​input into the normalized probabilistic activation response function mathematical processing module. This module performs exponential operations on each element within the scaled evaluation scoring matrix and divides the result of each exponential operation by the sum of all exponential operation results, performing probability allocation operations to generate a time-step weight distribution state parameter weight matrix. The system extracts the time-step weight distribution state parameter weight matrix and performs weighted matrix multiplication and tensor dimension fusion concatenation calculations with the multi-dimensional physical feature expression column vector matrix to output a corrected water level lag feature matrix. Finally, the system uses this corrected water level lag feature matrix through a fully connected mapping layer to output a tensor of intelligent prediction results for surface rebound deformation.

[0124] It is understandable that the target survey grid area is defined by a thick layer of clay deposits at the bottom. The system calculates and outputs the equivalent thickness of the cohesive soil layer. Exceeding the set thickness constant, the vertical permeability coefficient value is obtained through inversion. The value is below the set permeability constant. When the arithmetic logic microprocessor executes the calculation process, it generates operands whose values ​​fall within the negative range, causing the hydrogeological consolidation hysteresis compensation factor to... Floating-point values ​​converge to zero. After the compensation factor bias tensor matrix is ​​input into the long short-term memory network gating adjustment calculation stage, the values ​​of elements within the input gate control filtering probability matrix decrease, reducing the proportion of the surge water level feature tensor written into the cell state feature tensor main memory. The weighting coefficient values ​​of the time-step weight distribution state parameter weight matrix within the attention mechanism layer synchronously decrease at the corresponding precipitation occurrence time. Based on this, the prediction model outputs a smooth surface rebound deformation prediction curve tensor. The target survey grid range is set within a shallow, coarse sand permeable geological area. The system calculates the equivalent thickness of the cohesive soil layer. The vertical permeability coefficient value is below the set thickness constant. Higher than the set permeability constant. Hydrogeological consolidation hysteresis compensation factor. Floating-point values ​​drift towards higher parameter ranges. The compensation factor bias tensor matrix, after being input into the network computation stage, ensures that the input gate control filtering probability matrix maintains the numerical extraction access channel. The generated water head surge feature tensor is written into the cell state feature tensor main memory; the weighting coefficients of the time step weight distribution state parameter weight matrix maintain a set proportion distribution at the corresponding time. Based on this, the prediction model outputs the corresponding surface rebound deformation prediction curve tensor.

[0125] In this embodiment, the strong filtering operation for regions with low coherence includes the following constraint processing steps:

[0126] The temporal coherence fluctuation variance contained in each interferogram generated during the preprocessing operation of multi-source data for real-time monitoring of surface rebound deformation is obtained, and the capillary rise height value corresponding to each interferogram is obtained.

[0127] Determine whether the temporal coherence fluctuation variance and the capillary rise height value synchronously exceed a preset surge judgment threshold;

[0128] When it is determined that the temporal coherence fluctuation variance and the capillary rise height value simultaneously exceed the surge determination threshold, the target interference window corresponding to the surge determination threshold is marked as an abnormal window caused by signal masking due to sudden changes in soil moisture.

[0129] For the abnormal window, the default logical constraint of performing strong filtering when the coherence is low is removed, and the entangled phase gradient vector corresponding to the pixel point whose coherence meets the requirements at the edge of the abnormal window is extracted.

[0130] The spatial topological phase reconstruction compensation operation is performed on the low coherence pixels inside the abnormal window using the entangled phase gradient vector, in order to replace the default strong filtering process and prevent the real small surface rebound initial deformation signal caused by the sudden increase in surface water content from being excessively smoothed out.

[0131] When it is determined that the variance of temporal coherence fluctuation and the value of capillary rise height do not synchronously exceed the surge determination threshold, the operating principle of maintaining weak filtering in areas with high coherence and strong filtering in areas with low coherence is maintained.

[0132] The working principle and beneficial effects of the above technical solution are as follows:

[0133] Specifically, the system data processor retrieves the temporal coherence fluctuation variance parameter matrix contained in each interferogram. The system extracts the master-slave radar interferometric complex image tensor matrix of multi-scene registration status from the time-series database. The system configures a spatial motion estimation reading window to read the microwave scattered echo amplitude and phase values ​​pixel by pixel within the complex image tensor matrix. The system uses the interferometric coherence coefficient formula to calculate the mathematical expectation of the complex conjugate product of the master and slave images, the mathematical expectation of the amplitude energy of the master image, and the mathematical expectation of the amplitude energy of the slave image within the motion estimation reading window. The system performs an absolute value operation on the mathematical expectation of the complex conjugate product of the master and slave images and divides it by the arithmetic square root of the product of the mathematical expectations of the amplitude energy of the master image and the slave image to generate independent coherence coefficients for specific spatial pixel locations. For specific spatial pixel locations, the system extracts multiple independent coherence coefficients in chronological order along the time axis and constructs a floating-point vector of the coherence coefficient temporal evolution sequence. The system calls the mean calculation library function for the floating-point vector of the coherence coefficient time evolution sequence to calculate the expected value of the local temporal coherence coefficient mean. The system then performs subtraction operations on the coherence coefficients of each period contained within the floating-point vector of the coherence coefficient time evolution sequence with the expected value of the local temporal coherence coefficient mean, and squares the resulting difference sequence. The system sums all the squared results and divides them by the total number of time series samples to calculate the temporal coherence fluctuation variance value corresponding to the spatial pixels. Finally, the system traverses the coordinates of all mapped pixel nodes within the target monitoring area to generate a two-dimensional temporal coherence fluctuation variance distribution matching matrix.

[0134] Furthermore, during the computation process, the system initiates an inversion operation to obtain the capillary rise height values ​​corresponding to each interferogram. The system data bus retrieves the spatially continuous distribution grid data matrix of planar rainfall generated through spatial interpolation from the network server; simultaneously, it retrieves the tensor of shallow stratum soil volumetric water content parameters fed back by the sensor monitoring network, and the physical property parameter matrix of vadose zone soil particle size distribution in the target area extracted from the stratum drilling database. The system inputs the aforementioned multi-source heterogeneous parameter set into the solution modules of the partial differential equations of water movement and the capillary rise formula. The system establishes a discrete grid partitioning node sequence coordinate system in the vertical elevation coordinate dimension. For each layer of spatial finite difference grid control volume within the vadose zone partitioning sequence, the system derives the unsaturated hydraulic conductivity variable matrix parameters and soil-water matrix potential gradient calculation vector corresponding to each layer of grid control volume based on the soil particle size distribution physical property parameter matrix. The system calls the numerical center finite difference discretization algorithm program. The system calculates the infiltration water exchange flux matrix between a specific target grid control volume and its adjacent grid control volume directly above it. During the calculation, the unsaturated hydraulic conductivity variable and matrix potential parameters of the upper grid control volume are extracted, and the water convection flux at the contact interface is solved using a finite difference formula. Simultaneously, the system calculates the effluent water exchange flux matrix between the specific target grid control volume and its adjacent grid control volume directly below it. Based on the infiltration water exchange flux matrix and the effluent water exchange flux matrix, the system performs matrix subtraction to obtain the net water flux retention matrix within the specific target grid control volume.

[0135] Understandably, based on the physical constraint that the partial derivative of soil volumetric water content with respect to time is equal to the net water flux retention matrix, the system employs a forward Euler numerical approximation scheme to perform time-scale discretization and establish a system of nonlinear algebraic equations. The system uses the spatially continuous distribution grid data matrix of planar rainfall as the upper boundary condition of the top-level grid control volume; and the soil volumetric water content parameter tensor as the bottom-level initial state condition of the deep grid control volume. The system uses the pursuit method to solve the tridiagonal sparse matrix equations and performs nonlinear iterative derivation. The system outputs a spatial distribution tensor model of soil volumetric water content within the vertical stratigraphic profile at independent measurement times. The system iterates through the floating-point scalar values ​​of water content within the spatial distribution tensor model of soil volumetric water content; the system compares these floating-point scalar values ​​with the physical upper limit threshold of water content in the local saturated state. The system extracts the highest longitudinal elevation spatial coordinates at the location that meets the upper limit physical threshold standard position, performs an arithmetic subtraction calculation operation between the aforementioned highest longitudinal elevation spatial coordinates and the longitudinal coordinate parameters of the static reference surface, extracts the scalar numerical results output by the difference operation, and generates a numerical array of capillary rise height corresponding to each interferogram.

[0136] Specifically, the system determines whether the time-series coherence variance and capillary rise height values ​​simultaneously exceed preset surge thresholds. The system configures a first surge threshold and a second surge threshold in the storage area. The system logic control unit calls a dual-channel comparator. The comparator reads the time-series coherence variance and the first surge threshold, performing a numerical verification operation; the comparator simultaneously reads the capillary rise height and the second surge threshold, performing corresponding numerical verification operations. The system uses a Boolean logic operation module to combine and determine the two sets of verification results. When the time-series coherence variance is greater than the first surge threshold, and the capillary rise height value within the corresponding coordinate grid area is greater than the second surge threshold, the Boolean logic operation module performs an AND gate operation, outputting a truth control level signal. When the system outputs a truth control level signal, the system determines that the time-series coherence variance and capillary rise height values ​​simultaneously exceed the surge threshold. Based on the coordinate positions corresponding to the true value control level signal, the system marks the target interference window region, composed of clusters of continuously out-of-bounds pixels, as an anomalous window mask binarized distribution feature matrix. For the pixel coordinate clusters contained in the anomalous window mask binarized distribution feature matrix, the system blocks the data preprocessing thread's call instructions to the wide-window strong filtering component code library. The system removes the constraint that forces a strong filtering process if coherence is low. The system extracts the wrapped phase gradient vector data corresponding to pixel nodes located outside the boundary buffer of the anomalous window mask matrix and whose coherence coefficient values ​​are greater than the set safety tolerance evaluation benchmark.

[0137] Furthermore, the system performs boundary contour extraction operations through morphological operations to obtain the wrapped phase gradient vector matrix data corresponding to pixels with acceptable coherence at the edge of the abnormal window. The system establishes a ring-connected dilatational morphological operator structure matrix with a search step size. The system transforms the pixel set of the outer contour of the binary distribution feature matrix of the abnormal window mask into a seed dilatation coordinate group. The system uses the ring-connected dilatational morphological operator structure matrix to perform dilatation and expansion processing on the seed dilatation coordinate group towards the surrounding mapping grid. The dilatation and expansion processing includes: aligning the central anchor point of the ring-connected dilatational morphological operator structure matrix to the traversed pixel positions of the seed dilatation coordinate group, calculating the logical OR value of pixels within the area covered by the operator structure matrix, and assigning the logical OR value to the output grid image corresponding to the central anchor point. The system extracts and generates a set of network coordinate nodes for the dilatation and expansion transition region. The system reads the measured coherence coefficient evaluation values ​​of each pixel coordinate in the set of network coordinate nodes for the dilatation and expansion transition region from the original interferometric complex number buffer storage area. The system retrieves the coherence quality preservation evaluation threshold for comparison and verification. The system removes pixel sets whose measured coherence coefficient evaluation values ​​do not meet the evaluation threshold, and saves the discrete clusters of high-coherence observation feature nodes that meet the conditions. For the pixel coordinates within the saved discrete clusters of high-coherence observation feature nodes, the system reads the floating-point data of the unwrapped complex phase angle from radar measurements. The system's computational core performs micro-discrete differential calculations along the radar ranging time slant range direction and the satellite azimuth direction, respectively. The system uses the central finite difference approximation formula to perform spatial discretization calculations of the rate of change of difference quotients on adjacent pixel nodes, generating spatial two-dimensional partial derivative differential rate parameters. The system synthesizes the partial derivative differential rate parameters in different geometric directions into vectors, outputting a spatially wrapped phase gradient multidimensional vector array parameter containing physical information about the direction and magnitude of the wrapped phase partial derivatives.

[0138] Understandably, the system utilizes the entangled phase gradient vector to perform spatial topological phase reconstruction compensation inference on low-coherence pixels within the anomaly window. Spatial weighted interpolation algorithms are prone to inducing phase jump unwrapping errors when processing entangled folded phases restricted to negative and positive pi values. The spatial topological phase reconstruction compensation operation follows the closed-loop integral law of physical geometry for two-dimensional continuous manifolds to perform network path planning solutions. Specifically, the system uses the spatial entangled phase gradient multidimensional vector information array parameters extracted from the discrete community of high-coherence observation feature nodes as the derivative physical boundary constraint condition variables for solving the Poisson partial differential equation continuous network model. Within the spatially connected directed transfer graph theory grid topology model virtually constructed by the system, anomalously low-coherence pixels are mapped and defined as discrete network nodes in the graph theory model. The physical phase difference generational value between adjacent four-neighbor network position nodes is assigned as the physical arc edge data of the directed connectivity transmission network with flow load capacity constraints. For isolated residual physical singularity nodes that experience wrap-up unwrapping jumps within the network model, the system performs loop integral verification. The system calculates the algebraic summation of phase differences along a closed integral path composed of adjacent pixel network nodes. The system classifies and labels residual singularity nodes based on the positive or negative polarity of the summation result. The system establishes positive charge source nodes and negative charge termination drain nodes within the network topology model. The system constructs a graph-theoretic cost function with unidirectional connectivity constraints on the transmission network. The system inputs coherence characteristic evaluation parameters and spatial gradient variation values ​​into the graph-theoretic cost function as network transmission compensation cost parameters.

[0139] Specifically, the system scheduling and calculation module executes a unidirectional flow optimization algorithm. The system iteratively executes the connected network node path search operator to find a global network connected path branch tangent distribution structure diagram that minimizes the power consumption of all positive and negative polarity residual singularity nodes and the network flow path transmission cost. Based on the established global network connected path branch tangent distribution structure diagram, the system cuts the corresponding directed arc edges according to the mutual algebraic transmission physical equations. The system uses the control equation matrix containing the surrounding partial derivative gradients as the solution boundary and runs the preconditional conjugate gradient dimensionality reduction solution module. Through a step-by-step iterative elimination process, the system derives the geometric complex network reconstructed phase surface distribution matrix that satisfies the topological changes of the surrounding strata. The system extracts the geometric complex network reconstructed phase surface distribution matrix, replaces the invalid phase pixel grid data blocks corresponding to the abnormal windows in memory, and completes the radar signal repair and filling operation.

[0140] Furthermore, when the system's underlying logic judgment module calculates a verification failure level signal based on the comparison of temporal coherence fluctuation variance and capillary rise height, the system confirms that neither of these parameters synchronously exceeds the pre-set surge judgment benchmark parameter. The system terminates the interception operation task. The system follows the loaded and set main workflow of interferometric image enhancement preprocessing. For coherence regions where the coherence evaluation score is higher than the set retention baseline, the system calls and executes a frequency domain smoothing weak filtering calculation unit with a low calculation window span size parameter; for low coherence distribution regions where the coherence evaluation score is equal to or lower than the set retention baseline, the system calls and executes a Gaussian suppression strong filtering elimination operation program with a set filter window convolution kernel size.

[0141] Understandably, the scenario involves scheduling irrigation for a regional area of ​​farmland. The shallow soil pores become saturated with water during irrigation. A synthetic aperture radar (SAR) remote sensing satellite performs a scanning mission during the operation period. On the interferogram, the farmland-covered area exhibits decoherent noise pixel characteristics. The system retrieves the calculated variance of temporal coherence fluctuations for each pixel node in the farmland area and compares it with the first surge threshold for verification, confirming that the value exceeds the limit. The system retrieves the capillary rise height parameter and compares it, confirming that it exceeds the second surge threshold. The system outputs a dual-path verification signal, delineating the irrigated farmland-covered area as an abnormal window mask array, and stops noise reduction filtering. The system extracts the entangled phase gradient tensor information of physical points around the farmland mask array as boundary variables for solving the topological phase equation. The system performs mesh connectivity topological unwrapping network path reconstruction calculations on the internal noise points. A complex matrix replaces the abnormal pixels, outputting the surface rebound and subsidence monitoring results.

[0142] For example, the target monitoring area might cover urban underground tunneling construction projects. The rotating cutterhead cutting through the soil and rock layers transmits mechanical wave energy to the ground, causing high-frequency vibration and displacement deformation. This vibration displacement causes phase shift and misalignment in the interference, causing the temporal coherence fluctuation variance to exceed the first surge threshold. The capillary rise height calculated by the system remains stable within the constant safe range of the envelope, confirming that it has not exceeded the second surge threshold. The system determines that the dual control state has not been triggered. The system abandons the masking obstruction extraction operation and calls the preset Gaussian smoothing strong filtering noise reduction function code. The system performs smoothing and noise reduction operations to remove the floating-point information of speckle noise pixels caused by mechanical vibration, generating a stable low-frequency smoothed map matrix for the front-end display module to read.

[0143] The conventional data processing algorithms and classical mathematical equations involved in the embodiments of this invention and the above calculation process, such as the spherical semivariance function model and unbiased estimation constraint linear equation system for formation thickness estimation, the nonlinear partial differential equation system for seepage control and its nonlinear least squares solution process for permeability coefficient inversion, the state estimation and error covariance matrix filtering update algorithm for time-series water level denoising, the partial differential equation system for water movement and the forward Euler numerical approximation scheme for calculating capillary rise height, and the Poisson partial differential equation and preconditional conjugate gradient solution algorithm for spatial topological phase reconstruction, are all existing well-known mathematical tools and conventional basic data processing methods familiar to those skilled in the art. The embodiments of this invention focus on illustrating the logical calling relationships, business flow rules, and overall architecture design of various multi-source physical parameters and feature tensors in the intelligent prediction model and dynamic monitoring system. To highlight the core innovative contributions of this invention and make the specification more concise, the underlying formula derivation and standard algebraic analytical process of the above-mentioned basic mathematical theoretical models and classical partial differential equations will not be elaborated here.

[0144] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0145] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for real-time visualization and monitoring of surface rebound deformation based on InSAR technology, characterized in that, Includes the following steps: S1: Based on InSAR technology, satellite remote sensing data, POD precision orbit data, DEM data and surface auxiliary data are acquired to form multi-source data for real-time monitoring of surface rebound deformation; S2: Based on the ENVI-SARscape module, multi-source data of real-time monitoring of surface rebound deformation are processed and analyzed to extract the real rebound deformation information of the surface. S3: Construct a deep learning-based intelligent prediction model for surface rebound deformation, predict the amount of surface rebound deformation and groundwater level changes within the time interval of satellite data, predict the surface rebound state at the current moment in real time, and realize real-time monitoring of surface rebound deformation. S4: A real-time data dashboard integrating 3D geological models, multi-source remote sensing data, and early warning levels, enabling dynamic monitoring and visualization of surface rebound deformation and groundwater level changes. Among them, the water level state characteristic values ​​corresponding to each independent moment in the continuous groundwater level change are nonlinearly attenuated and weighted based on the hydrogeological consolidation lag compensation factor, generating a corrected water level lag feature matrix with the a priori physical mechanism properties of soil mechanics. The corrected water level lag feature matrix is ​​used to guide the Attention layer to output the accurate time step weight distribution state, eliminating the long-period physical deviation error caused by pure data driving. Building a deep learning-based intelligent prediction model for surface rebound deformation also includes: High-frequency groundwater level data are unified to the InSAR time point through moving average or interpolation methods, and InSAR monitoring points are spatially correlated with the data from the nearest groundwater monitoring well. Core features including groundwater level change, historical surface deformation rate and cumulative deformation were selected, as well as auxiliary surface features including stratigraphic lithology parameters, rainfall and POD orbital correction residuals. Since there is a physical time lag between surface rebound and groundwater level rise, a sliding window was introduced when constructing features to use groundwater level changes at continuous time intervals as input to capture the lag response. The model adopts an Attention-LSTM hybrid architecture, receives multivariate time series data, automatically calculates the weight distribution of different time steps and different features based on the Attention layer to highlight key influencing factors, extracts time series features based on the LSTM layer, and maps the high-dimensional features extracted by the LSTM to deformation prediction values, outputting the amount of surface rebound in the future time. To achieve dynamic monitoring and visualization of surface rebound deformation and groundwater level changes, the following operations are performed: Based on pre-processed multi-source data of real-time monitoring of surface rebound deformation and intelligent prediction model of surface rebound deformation, a large screen of real-time data of groundwater level rebound and surface rebound deformation was developed. Integrate geological data to construct a 3D stratigraphic model, clearly displaying stratigraphic features; develop a data fusion and visualization module to achieve linked display of the 3D stratigraphic model with multi-source data and early warning levels, and support interactive functions such as multi-scale zooming, regional filtering, and single-point query; and optimize the data rendering algorithm to ensure smooth operation of the large screen under massive data. When constructing features, a sliding window is introduced. After capturing the hysteresis response by taking continuous changes in groundwater level as input, the following processing operations are covered: Obtain the equivalent thickness of the cohesive soil layer and the vertical permeability coefficient corresponding to the InSAR monitoring points contained in the multi-source data; Combining the preset time span value corresponding to the sliding window, and the water level state characteristic values ​​corresponding to each independent moment in the continuous groundwater level change, the hydrogeological consolidation hysteresis compensation factor corresponding to each independent moment is calculated. The calculation formula is as follows: ； In the formula, This represents the hydrogeological consolidation hysteresis compensation factor; This represents the pre-defined shape fitting parameters; This represents the equivalent thickness of the cohesive soil layer. This represents the numerical value of the vertical permeability coefficient; This represents the preset time span value; Represents the exponential function with the natural constant as the base; The logarithmic function is defined with the natural constant as its base. Represents the natural constant; This represents a pre-defined hydrological scaling constant; This represents the water level status characteristic value corresponding to the independent time point; Finally, based on the calculated hydrogeological consolidation lag compensation factor, the water level state feature values ​​corresponding to each independent time moment in the continuous groundwater level changes input into the Attention-LSTM hybrid model architecture are nonlinearly attenuated and weighted to generate a corrected water level lag feature matrix with prior physical mechanism properties of soil mechanics. The corrected water level lag feature matrix is ​​used to guide the Attention layer to output an accurate time step weight distribution state, eliminating the long-period physical deviation error caused by pure data-driven approaches.

2. The method for real-time visualization and monitoring of surface rebound deformation based on InSAR technology according to claim 1, characterized in that, The real-time monitoring data for surface rebound deformation from multiple sources includes satellite remote sensing data, POD precision orbit data, DEM data, and surface auxiliary data. The satellite remote sensing data is used to provide C-band SAR images with high revisit cycles, which is the basis for time-series monitoring of surface rebound deformation. The POD precision orbit data is used to correct satellite orbit errors and ensure the geometric accuracy of the interferometry processing; The DEM data is used to remove terrain phase and simulate the effect of terrain undulation on the rebound signal; The surface auxiliary data is used for auxiliary monitoring of surface rebound deformation, including groundwater level monitoring well data, geological borehole data, and three-dimensional stratigraphic structure data.

3. The method for real-time visualization and monitoring of surface rebound deformation based on InSAR technology according to claim 2, characterized in that, Based on the ENVI-SARscape module, multi-source data from real-time monitoring of surface rebound deformation are processed and analyzed to perform the following operations: Preprocessing of multi-source data from real-time monitoring of surface rebound deformation, including multi-view processing, filtering, registration, and interferogram generation; Spatiotemporal matching of multi-source data for real-time monitoring of surface rebound deformation is performed, and SAR images from different time phases are accurately registered with POD orbits and DEM data to ensure pixel-level correspondence. Temporal InSAR analysis was performed on multi-source data of real-time monitoring of surface rebound deformation. SBAS-InSAR technology was used to extract information on small surface deformations over long time series and to obtain the cumulative deformation and deformation rate fields.

4. The method for real-time monitoring of surface rebound deformation based on InSAR technology according to claim 3, characterized in that, Preprocess the multi-source data from real-time monitoring of surface rebound deformation by performing the following operations: The real-time monitoring data of surface rebound deformation from multiple sources is converted into the internal format of SARscape, and satellite remote sensing data and POD precision orbit data are imported. The orbit information is automatically written into the metadata and satellite position error is corrected. The pixels of the two SAR images are geometrically aligned, the vertical baselines of the two SAR images are calculated, an initial geometric model is constructed using orbital data for coarse registration, and image feature point matching is used for fine registration. The slave image is resampled onto the geometric grid of the master image to generate the registered slave image SLC data. At this point, the pixels of the master and slave images are in one-to-one correspondence. The registered master and slave images are multiplied by their complex conjugates to obtain the amplitude reflecting the characteristics of the ground features and the phase reflecting the slant range difference. When generating the interferogram, the flat phase is simulated using POD fine track data and DEM data and then subtracted to remove the flat phase fringes caused by topographic relief, so that the residual phase reflects the surface deformation and elevation error. In the interferogram generation dialog box, the number of views in the range and azimuth directions is directly set to balance the resolution of the range and azimuth directions, making the pixels close to squares. Speckle noise is reduced and the signal-to-noise ratio is improved through averaging. The interferogram is divided into small windows, with the filtering window size being 5x5. The spectrum is smoothed in the frequency domain. The filtering intensity is adaptively adjusted by estimating local coherence, with weak filtering in areas of high coherence and strong filtering in areas of low coherence.

5. The method for real-time visualization and monitoring of surface rebound deformation based on InSAR technology according to claim 4, characterized in that, The SBAS-InSAR technique was used to extract information on minute surface deformations over long time series and to obtain the cumulative deformation and deformation rate fields. The following operations were performed: Import the Sentinel-1 SLC time series dataset with orbit correction, establish pairing relationships between images, and automatically generate a complex network topology diagram. Each image has at least one in-and-out connection to ensure the continuity of the time series and generate a connection diagram. Registration and interferogram generation are performed in batches based on the connection map. The flat terrain effect is removed using the precision track data, the terrain phase is removed using the external DEM, and noise reduction is performed based on multi-view processing and adaptive filtering to improve the signal-to-noise ratio. The average coherence coefficient map of all interference pairs is calculated, and a threshold is set so that pixels with an average coherence coefficient higher than the threshold are included in subsequent calculations. Among them, densely built areas and some stable bare soil are retained, while water areas and construction sites with drastic changes are excluded. Based on the Delaunay MCF algorithm and using coherence diagrams to guide the unwrapping path, the propagation of phase jump error is prevented, the wrapped phase is restored to the true phase, and the residual orbital error polynomial is fitted using the selected high coherence points to eliminate the systematic phase trend and remove the long-wave error caused by the small error of the POD orbit, preventing it from being misjudged as regional rebound. An observation equation is established, and the atmospheric phase is separated by high-pass filtering in the time domain and low-pass filtering in the spatial domain, taking advantage of the low-frequency correlation and temporal uncorrelation of the atmospheric phase. The minimum norm solution is obtained by singular value decomposition or least squares method, and the linear deformation rate field and DEM residual are obtained.

6. The method for real-time visualization and monitoring of surface rebound deformation based on InSAR technology according to claim 5, characterized in that, A deep learning-based intelligent prediction model for surface rebound deformation is constructed to predict the amount of surface rebound deformation and groundwater level changes within satellite data time intervals, and the following operations are performed: Based on historical monitoring data, groundwater level change and surface deformation rate were selected as core characteristic factors. A related hybrid deep learning algorithm is used to train a time series prediction model to predict the amount of surface rebound deformation and groundwater level change within the time interval of satellite data. By using the Long Short-Term Memory (LSTM) network to capture the lag effect in time series data and combining it with an attention mechanism to dynamically allocate the weights of different influencing factors, we can achieve accurate prediction of deformation values ​​at future moments. By continuously optimizing model parameters and improving prediction accuracy using newly acquired satellite data, a closed loop of monitoring, prediction, and correction can be achieved, thereby eliminating the monitoring lag problem caused by the time difference of satellite data.

7. The method for real-time visualization and monitoring of surface rebound deformation based on InSAR technology according to claim 4, characterized in that, For strong filtering operations in regions of low coherence, the following constraint handling steps are included: The temporal coherence fluctuation variance contained in each interferogram generated during the preprocessing operation of multi-source data for real-time monitoring of surface rebound deformation is obtained, and the capillary rise height value corresponding to each interferogram is obtained. Determine whether the temporal coherence fluctuation variance and the capillary rise height value synchronously exceed a preset surge judgment threshold; When it is determined that the temporal coherence fluctuation variance and the capillary rise height value simultaneously exceed the surge determination threshold, the target interference window corresponding to the surge determination threshold is marked as an abnormal window caused by signal masking due to sudden changes in soil moisture. For the abnormal window, the default logical constraint of performing strong filtering when the coherence is low is removed, and the entangled phase gradient vector corresponding to the pixel point whose coherence meets the requirements at the edge of the abnormal window is extracted. The spatial topological phase reconstruction compensation operation is performed on the low coherence pixels inside the abnormal window using the entangled phase gradient vector, in order to replace the default strong filtering process and prevent the real small surface rebound initial deformation signal caused by the sudden increase in surface water content from being excessively smoothed out. When it is determined that the variance of temporal coherence fluctuation and the value of capillary rise height do not synchronously exceed the surge determination threshold, the operating principle of maintaining weak filtering in areas with high coherence and strong filtering in areas with low coherence is maintained.

Images