Roadbed Settlement Time Series Prediction and Three-Dimensional Deformation Monitoring System and Method

Abstract

This invention discloses a roadbed settlement time-series prediction and three-dimensional deformation monitoring system and method, belonging to the field of surface deformation monitoring and prediction technology. The system includes a data acquisition module, a spatiotemporal feature construction module, a prediction model fusion module, a deformation field reconstruction module, a synthetic data generation module, and a monitoring and early warning module. By fusing InSAR satellite remote sensing images and ground micro-deformation sensor data, a spatiotemporally coupled Transformer-GRU prediction model is constructed. A pixel-level settlement rate field algorithm is pioneered, improving the surface deformation monitoring accuracy to 0.5 mm per year. Furthermore, historical settlement images are synthesized through generative adversarial networks to enhance training data, achieving accurate reconstruction and long-term prediction of the roadbed three-dimensional deformation field. This effectively solves the shortcomings of existing technologies in three-dimensional deformation reconstruction, long-term prediction accuracy, and generalization ability under insufficient data conditions, providing strong support for the safe operation of transportation infrastructure.

Description

Technical Field

[0001] This invention relates to the field of surface deformation monitoring and prediction technology, and in particular to a roadbed settlement time-series prediction and three-dimensional deformation monitoring system and method, which can be applied to health monitoring and safety early warning of roadbed engineering such as highways, railways, and airport runways. Background Technology

[0002] With the rapid development of transportation infrastructure construction, roadbed settlement has become a key factor affecting road safety and service life. Traditional roadbed settlement monitoring mainly relies on ground measurement methods, such as leveling and Global Navigation Satellite System (GNSS) measurements. Although these methods offer high accuracy, they suffer from limitations such as limited monitoring points, high costs, and difficulty in achieving large-scale continuous monitoring.

[0003] CN119437159A discloses a method and system for monitoring land subsidence by fusing BeiDou and InSAR data. This technical solution receives positioning data from the BeiDou satellite system and land deformation data acquired by the synthetic aperture radar interferometry system. It constructs a spatiotemporal fusion model to correct the land deformation data, and uses multi-period analysis techniques to determine the development trend of land subsidence and generate a land subsidence distribution map. The solution primarily employs a Kalman filter algorithm to suppress noise in the land deformation data, applies an autoregressive integral moving average model for trend analysis, and combines wavelet transform techniques to analyze subsidence patterns at different frequencies.

[0004] However, the aforementioned existing technologies have the following shortcomings:

[0005] First, existing technologies mainly focus on the monitoring and analysis of two-dimensional surface deformation, lacking the ability to accurately reconstruct the three-dimensional deformation field of the roadbed. Roadbed settlement includes not only vertical subsidence but also horizontal displacement. Existing technologies cannot fully characterize the three-dimensional deformation features of the roadbed, resulting in an insufficient understanding of the settlement mechanism.

[0006] Second, the time series prediction methods used in existing technologies are mainly based on traditional statistical models, such as the autoregressive integral moving average model. These models are difficult to capture the complex nonlinear spatiotemporal dependencies in roadbed settlement data, especially in prediction tasks with long time spans, where the prediction accuracy decreases significantly.

[0007] Third, existing technologies lack the ability to perform fine-grained calculations of pixel-level subsidence velocity fields. Although InSAR technology can provide large-scale deformation monitoring, existing methods do not make full use of InSAR data and fail to reconstruct pixel-level subsidence velocity fields, resulting in insufficient ability to identify local subsidence anomaly areas.

[0008] Fourth, existing technologies suffer from poor generalization ability and predictive stability when data samples are insufficient. Roadbed settlement monitoring typically requires the accumulation of historical data over a long period of time. In the early stages of engineering or when data is lacking, existing methods struggle to provide reliable prediction results.

[0009] Therefore, there is an urgent need for a roadbed settlement time series prediction and three-dimensional deformation monitoring system and method that can achieve pixel-level settlement rate field reconstruction, integrate deep learning time series prediction technology, and have data augmentation capabilities. Summary of the Invention

[0010] To address the problems existing in the prior art, this invention provides a system and method for predicting roadbed settlement over time and monitoring three-dimensional deformation. This invention integrates InSAR satellite remote sensing images with ground micro-deformation sensor data to construct a spatiotemporally coupled Transformer-GRU prediction model. It pioneers a pixel-level settlement rate field algorithm, improving the accuracy of surface deformation monitoring to 0.5 mm per year. Furthermore, it enhances training data by synthesizing historical settlement images using a generative adversarial network, achieving accurate reconstruction and long-term prediction of the roadbed's three-dimensional deformation field.

[0011] In a first aspect, the present invention provides a roadbed settlement time-series prediction and three-dimensional deformation monitoring system, including a data acquisition module, a spatiotemporal feature construction module, a prediction model fusion module, a deformation field reconstruction module, a synthetic data generation module, and a monitoring and early warning module.

[0012] The data acquisition module is used to receive InSAR satellite remote sensing image data and ground microdeformation sensor data, perform pixel-level registration processing on the InSAR satellite remote sensing image data to generate a pixel-level surface deformation dataset, and perform spatiotemporal alignment processing on the ground microdeformation sensor data to generate a ground deformation dataset.

[0013] The spatiotemporal feature construction module divides the pixel-level surface deformation dataset into grids according to the spatial resolution threshold. If the number of pixels in a single grid cell meets the first density range, the spatiotemporal feature vector corresponding to the grid cell is extracted. The spatiotemporal feature vector includes subsidence rate field features, spatial gradient field features, and time series features.

[0014] The prediction model fusion module constructs a spatiotemporally coupled Transformer-GRU prediction model, which includes a Transformer encoder layer and a GRU decoder layer. The Transformer encoder layer is used to capture long-term dependencies in spatiotemporal feature vectors, and the GRU decoder layer is used to predict the settlement trend at future moments.

[0015] The deformation field reconstruction module calculates the pixel-level settlement velocity field based on the prediction results of the Transformer-GRU prediction model. If the settlement velocity of any pixel in the pixel-level settlement velocity field meets the first velocity range, the pixel is marked as a high-risk area, and the three-dimensional deformation field corresponding to the high-risk area is reconstructed.

[0016] The synthetic data generation module constructs a generative adversarial network, which includes a generator network and a discriminator network. The generator network is used to synthesize historical subsidence images, and the discriminator network is used to distinguish between real subsidence images and synthetic historical subsidence images.

[0017] The monitoring and early warning module generates a roadbed settlement monitoring report based on the three-dimensional deformation field. If the maximum cumulative settlement in the three-dimensional deformation field meets the first settlement threshold range or the second settlement threshold range, an early warning signal of the corresponding level is triggered.

[0018] In one possible implementation, the data acquisition module receives multi-temporal InSAR satellite remote sensing images, which include at least 15 SAR images. Persistent scatterer identification is performed on the multi-temporal InSAR satellite remote sensing images to extract a set of persistent scattering points, preferably with a spatial density of 300 to 1000 persistent scattering points per square kilometer. Temporal InSAR processing is performed on the persistent scattering point set to calculate the deformation time series for each persistent scattering point, generating a pixel-level surface deformation dataset. Measurement data from ground-based microdeformation sensors at multiple monitoring points is received; the measurement data includes displacement and timestamps. The measurement data from the ground-based microdeformation sensors are matched with the spatial coordinates of the persistent scattering points. If the matching distance is less than a spatial matching threshold, temporal alignment and fusion are performed; the spatial matching threshold is preferably 20 to 50 meters. This multi-source data fusion method can comprehensively utilize the wide coverage advantage of satellite remote sensing and the high-precision monitoring capability of ground-based sensors, improving the spatial density and temporal resolution of settlement monitoring.

[0019] In one possible implementation, the spatiotemporal feature construction module includes a grid partitioning submodule, a density filtering submodule, and a feature extraction submodule. The grid partitioning submodule divides the pixel-level surface deformation dataset into multiple grid cells based on a spatial resolution threshold, and counts the number of pixels within each grid cell. The density filtering submodule determines whether the number of pixels meets a first density range. If it does, the grid cell is marked as a valid cell; otherwise, it is marked as an invalid cell. The first density range is preferably 10 to 50 pixels per grid cell. The feature extraction submodule extracts spatiotemporal feature vectors from the grid cells marked as valid cells. These vectors include the average settlement rate of pixels within the grid cell, the spatial standard deviation of the settlement rate, the temporal rate of change of the settlement rate, and the settlement gradient with neighboring grid cells. This grid-based spatiotemporal feature extraction method effectively reduces data dimensionality while preserving the spatial distribution pattern and temporal evolution characteristics of settlement.

[0020] In one possible implementation, the settlement rate field features include pixel-level settlement rate and spatial distribution pattern of settlement rate. The pixel-level settlement rate is calculated as follows: a linear regression is performed on the deformation time series of each pixel, and the slope of the regression line is calculated as the pixel-level settlement rate; the pixel-level settlement rate is smoothed using Gaussian filtering, with the filter window size preferably between 3 and 5 pixels; the spatial gradient field of the pixel-level settlement rate is calculated, reflecting the difference in settlement rate between adjacent pixels; if the gradient value at any location in the spatial gradient field conforms to a first gradient range, that location is marked as a settlement anomaly region, with the first gradient range preferably greater than 5 mm / year per 100 meters. The calculation accuracy of the pixel-level settlement rate field is preferably between 0.5 mm / year and 2 mm / year. This high-precision pixel-level analysis can identify local settlement anomalies, providing a more reliable basis for roadbed safety early warning.

[0021] In one possible implementation, the Transformer-GRU prediction model is constructed by: constructing a Transformer encoder layer, which includes a multi-head self-attention mechanism and a feedforward neural network, preferably with 4 to 8 heads in the multi-head self-attention mechanism; inputting spatiotemporal feature vectors into the Transformer encoder layer and calculating the spatiotemporal correlation weights between feature vectors through the multi-head self-attention mechanism; constructing a GRU decoder layer, which includes multiple GRU units, preferably with a hidden layer dimension of 128 to 512 dimensions; using the output of the Transformer encoder layer as the input of the GRU decoder layer, and predicting the settlement values ​​at multiple future time points through the GRU decoder layer, preferably with a time span of 1 to 12 months. If the time span of the spatiotemporal feature vector conforms to a first time range, the Transformer encoder layer uses the first layer configuration. If the time span of the spatiotemporal feature vector conforms to a second time range, the Transformer encoder layer uses the second layer configuration. The first time range is 1 to 3 years, and the second time range is 3 to 10 years. The first layer configuration is 4 to 6 layers, and the second layer configuration is 8 to 12 layers. This spatiotemporally coupled deep learning architecture can simultaneously capture long-term dependencies and short-term fluctuation features in settlement data, significantly improving prediction accuracy.

[0022] It should be noted that the multi-head self-attention mechanism of the Transformer encoder layer can compute the correlation between different spatial locations and different time steps in parallel, which has a stronger ability to model long-term dependencies compared to traditional recurrent neural networks. The GRU decoder layer selectively retains and updates historical information through a gating mechanism, reducing computational complexity while ensuring prediction accuracy. The fusion architecture of the Transformer encoder layer and the GRU decoder layer fully leverages the advantages of both models, forming a spatiotemporally coupled prediction framework.

[0023] In one possible implementation, the processing flow of the deformation field reconstruction module includes: if the settlement rate of a pixel in the pixel-level settlement rate field conforms to a first rate range, extracting the three-dimensional coordinates and settlement rate of the pixel; based on the three-dimensional coordinates and settlement rate, reconstructing the three-dimensional deformation field of the high-risk area using the Kriging interpolation method, which includes calculating the semi-variogram and constructing the interpolation weight matrix; performing spatial smoothing on the three-dimensional deformation field to remove high-frequency noise introduced during the interpolation process; calculating the deformation gradient field and curvature field of the three-dimensional deformation field, whereby the deformation gradient field reflects the spatial variation trend of the deformation and the curvature field reflects the degree of spatial curvature of the deformation. The first rate range is preferably greater than 10 mm / year, indicating that there is significant settlement in the area, requiring close attention and monitoring.

[0024] In one possible implementation, the training process of the synthetic data generation module includes: constructing a generator network using a convolutional neural network architecture, comprising an encoder and a decoder; constructing a discriminator network using a convolutional neural network architecture to distinguish between real subsidence images and synthetic historical subsidence images; randomly sampling real subsidence images from a pixel-level surface deformation dataset and generating synthetic historical subsidence images from random noise vectors; calculating the discriminator network's discrimination output for real subsidence images and its discrimination output for synthetic historical subsidence images, and updating the parameters of the generator and discriminator networks based on the discrimination outputs; if the discriminator network's discrimination accuracy for synthetic historical subsidence images meets a first accuracy range, it indicates that the synthetic historical subsidence images are highly similar to real subsidence images, and the synthetic historical subsidence images are added to the training dataset. The first accuracy range is preferably below 60%, meaning that the discriminator network has difficulty distinguishing between real and synthetic images, at which point the synthetic images have reached sufficiently high quality. This data augmentation method based on generative adversarial networks can effectively alleviate the problem of insufficient historical data and improve the generalization ability and robustness of the prediction model.

[0025] In one possible implementation, the monitoring and early warning module is further used to: calculate the spatial distribution characteristics of roadbed settlement based on the three-dimensional deformation field, including the location of the settlement center, the range of settlement influence, and the settlement morphology; calculate the settlement development trend at future times based on the prediction results of the Transformer-GRU prediction model, including the settlement acceleration and the estimated time when the settlement reaches the danger threshold; generate a roadbed settlement monitoring report, including the current settlement status, historical settlement evolution curves, future settlement prediction curves, early warning level, and recommended treatment measures; if the early warning level is Level 1, send monitoring reminder information to the monitoring terminal; if the early warning level is Level 2, send emergency treatment information to the management terminal. The first settlement threshold range is preferably 30 mm to 50 mm, indicating that the settlement has reached a level requiring attention; the second settlement threshold range is preferably greater than 50 mm, indicating that the settlement has reached a dangerous level and immediate treatment measures are required.

[0026] In one possible implementation, the system further includes an adaptive optimization module for monitoring the prediction error of the Transformer-GRU prediction model. If the prediction error falls within a first error range, the parameters of the Transformer-GRU prediction model remain unchanged. If the prediction error falls within a second error range, a model parameter update process is triggered. The first error range is preferably a root mean square error (RMSE) of less than 3 mm, and the second error range is preferably a RMSE greater than 5 mm. The model parameter update process includes acquiring the latest pixel-level surface deformation dataset and foundation deformation dataset, incrementally training the Transformer-GRU prediction model, and updating the weight parameters of the Transformer encoder layer and the GRU decoder layer. This adaptive optimization mechanism enables the system to continuously adapt to changes in settlement patterns during long-term operation, maintaining the stability of prediction accuracy.

[0027] A second aspect of the present invention provides a method for predicting the time series of roadbed settlement and monitoring three-dimensional deformation based on the above-mentioned system, comprising receiving InSAR satellite remote sensing image data and ground micro-deformation sensor data, performing pixel-level registration processing and spatiotemporal alignment processing; dividing the pixel-level surface deformation dataset into grids according to a spatial resolution threshold, and extracting spatiotemporal feature vectors; constructing a spatiotemporally coupled Transformer-GRU prediction model; calculating the pixel-level settlement rate field based on the prediction results, and reconstructing the three-dimensional deformation field; constructing a generative adversarial network to synthesize historical settlement images; generating a roadbed settlement monitoring report based on the three-dimensional deformation field and triggering an early warning signal.

[0028] The beneficial effects of this invention are as follows:

[0029] First, this invention pioneers a pixel-level settlement rate field algorithm, improving the accuracy of surface deformation monitoring to 0.5 mm / year to 2 mm / year, significantly higher than the accuracy level of existing technologies. By performing refined analysis of the deformation time series of each pixel, local areas of abnormal settlement can be identified, providing a more reliable basis for roadbed safety early warning.

[0030] Second, this invention employs a spatiotemporally coupled Transformer-GRU prediction model, which can simultaneously capture long-term dependencies and short-term fluctuation characteristics in settlement data. The Transformer encoder layer computes the correlation between different spatiotemporal locations in parallel through a multi-head self-attention mechanism, while the GRU decoder layer selectively retains and updates historical information through a gating mechanism. The fusion architecture of the two performs excellently in prediction tasks with a long time span (1 month to 12 months), and the prediction accuracy is significantly better than that of traditional statistical models.

[0031] Third, this invention effectively alleviates the problem of insufficient historical data by synthesizing historical subsidence images using generative adversarial networks. The generator network can learn the distribution characteristics of real subsidence images and generate highly similar synthetic images. These synthetic images can be used as training data to enhance the generalization ability and robustness of the prediction model, especially in the early stages of engineering or in cases of data scarcity, providing more reliable prediction results.

[0032] Fourth, this invention achieves precise reconstruction of the three-dimensional deformation field of the roadbed, enabling monitoring not only vertical settlement but also horizontal displacement. By calculating the deformation gradient field and curvature field, the three-dimensional deformation characteristics of the roadbed can be comprehensively characterized, providing a scientific basis for in-depth analysis of settlement mechanisms and engineering treatment.

[0033] Fifth, this invention establishes a tiered early warning mechanism, triggering different levels of early warning signals based on different threshold ranges of maximum cumulative settlement. The first-level warning (30 mm to 50 mm) reminds relevant personnel to strengthen monitoring, while the second-level warning (greater than 50 mm) requires immediate action. This tiered early warning strategy can effectively prevent the risk of roadbed settlement and ensure the safe operation of transportation infrastructure. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the overall architecture of the roadbed settlement time-series prediction and three-dimensional deformation monitoring system of the present invention.

[0035] Figure 2 This is a schematic diagram of the processing flow of the data acquisition module of the present invention.

[0036] Figure 3 This is a schematic diagram of the spatiotemporal feature construction module of the present invention.

[0037] Figure 4 This is a schematic diagram of the architecture of the Transformer-GRU prediction model of this invention.

[0038] Figure 5 This is a schematic diagram of the processing flow of the deformation field reconstruction module of the present invention.

[0039] Figure 6 This is a flowchart of the method for predicting roadbed settlement time sequence and monitoring three-dimensional deformation according to the present invention. Detailed Implementation

[0040] Please refer to the attached document. Figures 1-6The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings to enable those skilled in the art to better understand the present invention. It should be noted that in the description of the present invention, terms such as "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Furthermore, in the description of the present invention, unless otherwise stated, "a plurality of" means two or more.

[0041] like Figure 1 As shown, the present invention provides a roadbed settlement time-series prediction and three-dimensional deformation monitoring system, including a data acquisition module 1, a spatiotemporal feature construction module 2, a prediction model fusion module 3, a deformation field reconstruction module 4, a synthetic data generation module 5, and a monitoring and early warning module 6.

[0042] Data acquisition module 1 receives InSAR satellite remote sensing image data and ground-based microdeformation sensor data. It performs pixel-level registration processing on the InSAR satellite remote sensing image data to generate a pixel-level surface deformation dataset, and performs spatiotemporal alignment processing on the ground-based microdeformation sensor data to generate a ground deformation dataset. Data acquisition module 1 is the data input interface for the entire system, and its processing quality directly affects the performance of subsequent modules.

[0043] The spatiotemporal feature construction module 2 is connected to the data acquisition module 1. It is used to divide the pixel-level surface deformation dataset into grids according to the spatial resolution threshold. If the number of pixels in a single grid cell meets the first density range, the spatiotemporal feature vector corresponding to the grid cell is extracted. The spatiotemporal feature vector includes settlement rate field features, spatial gradient field features, and time series features. These features can comprehensively characterize the spatiotemporal evolution of roadbed settlement.

[0044] The prediction model fusion module 3 is connected to the spatiotemporal feature construction module 2 to construct a spatiotemporally coupled Transformer-GRU prediction model. The Transformer-GRU prediction model includes a Transformer encoder layer and a GRU decoder layer. The Transformer encoder layer captures long-term dependencies in the spatiotemporal feature vectors, while the GRU decoder layer predicts settlement trends at future times. The prediction model fusion module 3 is one of the core innovations of this invention, achieving accurate prediction of roadbed settlement through deep learning technology.

[0045] The deformation field reconstruction module 4 is connected to the prediction model fusion module 3. Based on the prediction results of the Transformer-GRU prediction model, it calculates the pixel-level settlement velocity field. If the settlement velocity of any pixel in the pixel-level settlement velocity field conforms to a first velocity range, the pixel is marked as a high-risk area, and the corresponding three-dimensional deformation field is reconstructed. The deformation field reconstruction module 4 realizes the conversion from a one-dimensional time series to a three-dimensional spatial field, providing an intuitive tool for the visualization and analysis of settlement.

[0046] The synthetic data generation module 5 is connected to the deformation field reconstruction module 4 to construct a generative adversarial network (GAN). The GAN includes a generator network and a discriminator network. The generator network synthesizes historical subsidence images, and the discriminator network distinguishes between real subsidence images and synthesized historical subsidence images. The synthetic data generation module 5 is another core innovation of this invention, improving the generalization ability of the prediction model through data augmentation technology.

[0047] The monitoring and early warning module 6 is connected to the deformation field reconstruction module 4 and is used to generate a roadbed settlement monitoring report based on the three-dimensional deformation field. If the maximum cumulative settlement in the three-dimensional deformation field meets the first settlement threshold range, a first-level early warning signal is triggered; if the maximum cumulative settlement meets the second settlement threshold range, a second-level early warning signal is triggered. The monitoring and early warning module 6 transforms the system's analysis results into actionable early warning information, providing decision support for engineering management personnel.

[0048] like Figure 2 As shown, the processing flow of data acquisition module 1 includes the following steps:

[0049] Step S11: Receive multi-temporal InSAR satellite remote sensing images, which include at least 15 SAR images. In one specific embodiment, C-band SAR images from the Sentinel-1 satellite can be used, with a time span of 2 years and an interval of 12 days, for a total of 60 SAR images.

[0050] Step S12: Identify persistent scatterers in multi-temporal InSAR satellite remote sensing images and extract persistent scattering point sets. Persistent scatterers refer to targets that maintain coherence in long-term SAR images, typically corresponding to stable surface features such as buildings and rocks. The persistent scatterer identification method involves calculating the amplitude dispersion index and coherence coefficient of the time-series SAR images, selecting pixels with an amplitude dispersion index less than 0.25 and a coherence coefficient greater than 0.7 as persistent scattering points. The spatial density of the persistent scattering point set is preferably 300 to 1000 persistent scattering points per square kilometer. In urban areas, due to the large number of buildings, the persistent scattering point density can reach over 1000 per square kilometer; in suburban or rural areas, the persistent scattering point density may only be around 300 per square kilometer.

[0051] Step S13: Perform temporal InSAR processing on the persistent scattering point set to calculate the deformation time series for each persistent scattering point. Temporal InSAR processing includes differential interferometry, phase unwrapping, and atmospheric delay correction. Differential interferometry extracts surface deformation information by calculating the interferometric phase difference between adjacent SAR images. Phase unwrapping converts the wrapped phase difference into continuous phase values; common methods include the minimum cost flow algorithm and the statistical cost network flow algorithm. Atmospheric delay correction removes the influence of atmospheric water vapor and ionospheric delay on the interferometric phase; this can be achieved using spatiotemporal filtering methods or by combining meteorological data. Through temporal InSAR processing, a deformation time series for each persistent scattering point is generated. The sampling interval for the time series is 12 days, the time span is 2 years, and a total of 60 deformation observations are included.

[0052] Step S14: Receive measurement data from the foundation micro-deformation sensor at multiple monitoring points. The measurement data includes displacement and timestamps. The foundation micro-deformation sensor can be an inclinometer, crack gauge, displacement gauge, etc., and is typically deployed at key locations on the roadbed, such as embankment slopes and bridge abutments. The measurement accuracy of the foundation micro-deformation sensor is typically 0.1 mm to 1 mm, and the sampling frequency is from once per hour to once per day.

[0053] Step S15: Match the measurement data from the ground-based microdeformation sensor with the spatial coordinates of the persistent scattering point. If the matching distance is less than the spatial matching threshold, perform time series alignment and fusion. The spatial matching threshold is preferably 20 to 50 meters, and its selection needs to consider the spatial resolution of InSAR pixels and the representative range of the ground-based sensor. Time series alignment uses a time interpolation method to unify data from different sampling frequencies onto the same time grid. Time series fusion uses a weighted average method, determining the weighting coefficients based on the measurement accuracy of InSAR and the ground-based sensor. The measurement accuracy of InSAR is approximately 2 to 5 millimeters, and the measurement accuracy of the ground-based sensor is approximately 0.1 to 1 millimeter. Therefore, the weighting coefficient for the ground-based sensor is typically set to 0.6 to 0.8, and the weighting coefficient for InSAR is 0.2 to 0.4.

[0054] Through the above processing flow, data acquisition module 1 generates pixel-level surface deformation datasets and foundation deformation datasets, providing a high-quality data foundation for subsequent spatiotemporal feature extraction and predictive modeling.

[0055] like Figure 3 As shown, the spatiotemporal feature construction module 2 includes a grid partitioning submodule 21, a density filtering submodule 22, and a feature extraction submodule 23.

[0056] The grid division submodule 21 divides the pixel-level surface deformation dataset into multiple grid cells based on a spatial resolution threshold, and counts the number of pixels within each grid cell. The spatial resolution threshold is preferably between 5 meters and 20 meters; the selection of this threshold requires a balance between spatial resolution and computational complexity. A smaller spatial resolution threshold (e.g., 5 meters) can provide finer spatial details but increases the number of grid cells and computational load; a larger spatial resolution threshold (e.g., 20 meters) can reduce computational load but may lose some spatial details. In a specific embodiment, for highway subgrade monitoring, the spatial resolution threshold can be set to 10 meters, dividing the entire monitoring area into 10-meter × 10-meter grid cells.

[0057] The density filtering submodule 22 determines whether the number of pixels in each grid cell conforms to the first density range. If it does, the grid cell is marked as a valid cell; otherwise, it is marked as an invalid cell. The first density range is preferably 10 to 50 pixels per grid cell. If the number of pixels in a grid cell is too small (e.g., less than 10), the statistical characteristics of that grid cell may not be stable enough and are easily affected by individual outliers. If the number of pixels is too large (e.g., more than 50), it may be due to an excessively high spatial resolution threshold, causing pixels with different sedimentation characteristics to be mixed in the same grid cell. Density filtering ensures that each valid grid cell contains a sufficient number of pixels with similar characteristics, providing reliable samples for subsequent feature extraction.

[0058] The feature extraction submodule 23 extracts spatiotemporal feature vectors from the grid cells marked as valid cells. The spatiotemporal feature vectors include the following types of features:

[0059] The first type of feature is the average settlement rate of pixels within a grid cell. The average settlement rate of that grid cell is obtained by arithmetically averaging the settlement rates of all pixels within the grid cell. This average settlement rate reflects the overall settlement trend of the region.

[0060] The second feature is the spatial standard deviation of settlement rate. The standard deviation of settlement rate for all pixels within a grid cell is calculated; this standard deviation reflects the spatial dispersion of settlement rate. A larger spatial standard deviation indicates uneven settlement within the grid cell, possibly due to uneven soil composition or uneven load distribution.

[0061] The third type of feature is the time-varying rate of settlement. The time-varying rate of settlement is calculated by differentiating the time series of settlement rates for each pixel, and then averaged across all pixels within the grid cell. The time-varying rate of settlement reflects the trend of acceleration or deceleration of settlement; a positive value indicates that the settlement rate is increasing, and a negative value indicates that the settlement rate is decreasing.

[0062] The fourth type of feature is the settlement gradient with adjacent grid cells. The difference between the average settlement rate of the current grid cell and the average settlement rates of the four adjacent grid cells (east, south, west, and north) is calculated and divided by the distance between the grid cells to obtain the settlement gradient in the four directions. The settlement gradient reflects the spatial trend of settlement variation; a larger settlement gradient indicates a significant spatial discontinuity in the settlement of the area.

[0063] The four types of features mentioned above are combined into a feature vector, which serves as the spatiotemporal feature vector of the grid cell. In a specific embodiment, if four features are extracted for each grid cell (mean settlement rate, spatial standard deviation, rate of change over time, and mean settlement gradient), then the spatiotemporal feature vector has a dimension of 4. If a more detailed feature description is required, other statistical features can be added, such as the maximum, minimum, median, and quartiles of the settlement rate.

[0064] like Figure 4 As shown, the architecture of the Transformer-GRU prediction model includes a Transformer encoder layer 31 and a GRU decoder layer 32.

[0065] The Transformer encoder layer 31 includes a multi-head self-attention mechanism 311 and a feedforward neural network 312. The core idea of ​​the multi-head self-attention mechanism 311 is to calculate the association weights between different positions in the input sequence, thereby capturing long-distance dependencies. The calculation process of the multi-head self-attention mechanism is as follows:

[0066] First, the input spatiotemporal feature vector sequence is represented as: ,in Indicates the first The spatiotemporal feature vector at each moment, For feature dimension, This represents the length of the time series.

[0067] Then, through three linear transformation matrices Map the input sequences to query vectors respectively. Key vector Sum value vector :

[0068] ,

[0069] in: For querying the matrix, The key matrix, For value matrices, To query the transformation matrix, The key transformation matrix is... The value transformation matrix, For the dimensions of keys and queries, The dimension of the value.

[0070] Next, calculate the self-attention weights:

[0071] ,

[0072] in: This is the attention weight matrix. Indicates the first At the nth moment, for the first Attention weight at each moment, The function normalizes each row so that the sum of the weights for each row is 1. This is a scaling factor used to prevent the gradient from vanishing due to an excessively large inner product.

[0073] Finally, the value vector is weighted and summed according to the attention weights to obtain the output of the self-attention:

[0074] ,

[0075] in: This is the self-attention output matrix. Indicates the first The output vector at each time step.

[0076] The multi-head self-attention mechanism executes the single-head attention mechanism multiple times in parallel, with each head using a different transformation matrix, and finally concatenates the outputs of all heads:

[0077] ,

[0078] in: For multi-head self-attention output, For the number of heads, For the first Output of size This indicates a splicing operation. This is the output transformation matrix.

[0079] The optimal number of heads in a multi-head self-attention mechanism is 4 to 8. A higher number of heads allows the model to learn richer spatiotemporal correlation patterns from different representation subspaces, but it also increases computational complexity. In a specific embodiment, for data with a time series length of 24 months, the number of heads can be set to 6, and the feature dimension... , key and query dimensions .

[0080] The feedforward neural network 312 is a two-layer fully connected neural network used to perform non-linear transformations on the output of self-attention:

[0081] ,

[0082] in: The output of the feedforward neural network, This is the first layer weight matrix. This is the first layer bias vector. This is the weight matrix for the second layer. This is the second layer bias vector. The hidden layer dimension of a feedforward neural network is typically set to... , To modify the activation function of the linear unit, it is defined as follows: .

[0083] The Transformer encoder layer also includes residual connections and layer normalization. The complete calculation process is as follows:

[0084] ,

[0085] ,

[0086] in: This is a layer normalization operation used to stabilize the training process. Residual connections can alleviate the gradient vanishing problem in deep networks.

[0087] If the time span of the spatiotemporal feature vectors falls within a first time range (1 to 3 years), the Transformer encoder layer uses a first-level layer configuration (4 to 6 layers); if the time span of the spatiotemporal feature vectors falls within a second time range (3 to 10 years), the Transformer encoder layer uses a second-level layer configuration (8 to 12 layers). This adaptive layer configuration can adjust the model's capacity according to the time span of the data. For data with short time spans, a shallower network can capture the main temporal patterns, while for data with long time spans, a deeper network is needed to model complex long-term dependencies.

[0088] The GRU decoder layer 32 includes multiple GRU units, with the hidden layer dimension of each GRU unit preferably ranging from 128 to 512 dimensions. GRU (Gated Recurrent Unit) is an improved recurrent neural network structure that selectively retains and updates historical information through a gating mechanism. The computation process of a GRU unit is as follows:

[0089] First, calculate the update gate. and reset door :

[0090] ,

[0091] ,

[0092] in: For the first The door is updated at any given moment. For the first The reset door at any given moment, For the first The hidden state at any given moment. For the first The encoder output at each moment. The weight matrix is ​​the hidden state to the gate. The weight matrix is ​​the input to the gate. Let be the bias vector of the gate. For the GRU hidden layer dimension, The Sigmoid activation function is defined as follows: .

[0093] Then, calculate the candidate hidden state. :

[0094] ,

[0095] in: For the first The candidate hidden state at each time step. This is the weight matrix from the hidden state to the candidate hidden state. The weight matrix is ​​input to the candidate hidden state. Let be the bias vector of the candidate hidden state. For element-wise multiplication, Let hyperbolic tangent activation function be defined as follows: .

[0096] Finally, the hidden state at the current moment is obtained by weighted fusion of the historical hidden state and the candidate hidden state according to the update gate:

[0097] ,

[0098] in: For the first The hidden state at each moment, update the door. The ratio of historical information to new information was controlled, when When the value is close to 0, the model tends to retain historical information; when... When the value is close to 1, the model tends to adopt new information.

[0099] The output of the GRU decoder is the predicted settlement values ​​at multiple future time points:

[0100] ,

[0101] in: For the first The predicted settlement value at a given time. To output the weight vector, This is the output bias.

[0102] The Transformer-GRU prediction model is trained using the mean squared error loss function:

[0103] ,

[0104] in: For loss function, The number of training samples. To predict the number of time steps, For the first The sample at the th The predicted value at a future time. For the first The sample at the th The true value of a future moment.

[0105] The model was optimized using the Adam optimizer with an initial learning rate of 0.001, which gradually decreased during training. Batch size ranged from 32 to 64, and the number of training epochs ranged from 100 to 200. In a specific implementation, for a dataset containing 500 grid cells and a 24-month time series, the Transformer encoder layer had 6 layers, each with 6 heads for the multi-head self-attention mechanism, a feature dimension of 256, and the GRU decoder layer had a hidden layer dimension of 256, predicting sedimentation values ​​for the next 6 months. The model achieved a root mean square error (RMSE) of 2.1 mm on the training set and 2.8 mm on the test set, significantly outperforming the traditional ARIMA model (RMSE of 5.2 mm on the test set) and the LSTM model (RMSE of 3.6 mm on the test set).

[0106] like Figure 5 As shown, the processing flow of deformation field reconstruction module 4 includes the following steps:

[0107] Step S41: If the settlement rate of a pixel in the pixel-level settlement rate field conforms to the first rate range, extract the three-dimensional coordinates and settlement rate of the pixel. The first rate range is preferably greater than 10 mm per year, indicating that the pixel has significant settlement. In a specific embodiment, for a 10-kilometer-long highway subgrade, there are approximately 1 million pixels (assuming a pixel resolution of 3 meters × 3 meters), of which approximately 5,000 pixels have a settlement rate greater than 10 mm per year. These pixels are mainly distributed in soft soil foundation sections and sections with high embankment filling.

[0108] Step S42: Based on the three-dimensional coordinates and settlement rate, the three-dimensional deformation field of the high-risk area is reconstructed using the Kriging interpolation method. Kriging interpolation is an optimal interpolation method based on spatial autocorrelation theory. Its core idea is that there is a correlation between nearby points in space, and the closer the distance, the stronger the correlation. Kriging interpolation includes the following steps:

[0109] First, calculate the semivariogram. :

[0110] ,

[0111] in: Lag distance The corresponding semi-mutation function value, The distance is Number of point pairs For position Settlement rate at that location, To and Distance is The settlement rate at the location.

[0112] Then, a suitable semivariogram model is selected for fitting. Commonly used models include the spherical model, the exponential model, and the Gaussian model. The expression for the spherical model is:

[0113] ,

[0114] in: The value of the gold nugget is represented as follows: The semivariogram value over time reflects measurement error and microscale variation. The sill value represents the semivariogram value when spatial correlation reaches a stable state. For variable range, representing the effective influence distance of spatial correlation, when At this point, there is no longer any spatial correlation between the points.

[0115] Next, for the locations where interpolation is needed... Construct the Kriging equations:

[0116] ,

[0117] in: For the first Interpolation weights for known points It is a Lagrange multiplier. The number of known points. Given points and The semivariogram values ​​between For known points and interpolation points The semi-variogram values ​​between.

[0118] Solving the above system of linear equations yields the interpolation weights. Then, calculate the estimated settlement rate at the location to be interpolated:

[0119] ,

[0120] in: For position The estimated settlement rate at that location, For known points Settlement rate at that location.

[0121] In a specific embodiment, for a high-risk area containing 5000 known settlement rates, a spherical model is used to fit the semi-variogram function to obtain the nugget value. sill value Variable range Using the Kriging interpolation method, a three-dimensional deformation field was reconstructed in this region with a spatial resolution of 5 meters, generating settlement rate estimates for approximately 200,000 interpolation points.

[0122] Step S43: Perform spatial smoothing on the three-dimensional deformation field to remove high-frequency noise introduced during interpolation. The spatial smoothing uses a Gaussian filter with a standard deviation set to 10 to 20 meters, which can effectively remove small-scale noise while preserving large-scale settlement patterns.

[0123] Step S44: Calculate the deformation gradient field and curvature field of the three-dimensional deformation field. The deformation gradient field is calculated using the finite difference method, for the position... Settlement rate at ,calculate direction and Deformation gradient in direction:

[0124] ,

[0125] ,

[0126] in: for Deformation gradient in direction, for Deformation gradient in direction, and This is the spatial step size, typically set to the spatial resolution of the grid.

[0127] The magnitude of the deformation gradient is:

[0128] ,

[0129] in: For position The deformation gradient magnitude at a given location reflects the spatial rate of change of the settlement rate at that location.

[0130] The curvature field is calculated using the second derivative, for position... Settlement rate at Calculate curvature :

[0131] ,

[0132] in: For position curvature at that point for The second derivative of the direction, for The second derivative of the direction. Curvature reflects the degree of bending in the spatial distribution of settlement rate; a larger curvature indicates an abrupt change or inflection point in settlement at that location.

[0133] Through the above processing steps, the deformation field reconstruction module 4 generates the three-dimensional deformation field, deformation gradient field, and curvature field of the high-risk area, providing rich spatial information for the visualization analysis and mechanism study of settlement.

[0134] The training process for synthetic data generation module 5 includes the following steps:

[0135] Step S51: Construct a generator network. The generator network adopts a convolutional neural network architecture, including an encoder and a decoder. The encoder maps the random noise vector to a low-dimensional latent representation, and the decoder maps the latent representation to a two-dimensional sedimentation image. The specific structure of the generator network is as follows: the input is a 100-dimensional random noise vector. After being mapped by a fully connected layer, it becomes The feature map is then upsampled progressively through four transposed convolutional layers, with each layer having a kernel size of [size missing]. The stride is 2, the number of channels is 256, 128, 64, and 1 respectively, the activation function is LeakyReLU (negative slope of 0.2), and the last layer uses the Tanh activation function. The output is... Settlement images.

[0136] Step S52: Construct a discriminator network. The discriminator network adopts a convolutional neural network architecture to distinguish between real subsidence images and synthetic historical subsidence images. The specific structure of the discriminator network is as follows: The input is... The sedimentation image was downsampled progressively through 5 convolutional layers, with each layer having a kernel size of [size missing]. The step size is 2, and the number of channels is 64, 128, 256, 512, and 1 respectively. The activation function is LeakyReLU (with a negative slope of 0.2). The last layer uses the Sigmoid activation function, and the output is a single scalar value representing the probability that the input image is a real image.

[0137] Step S53: Randomly sample real subsidence images from the pixel-level surface deformation dataset and generate synthetic historical subsidence images from random noise vectors. In each training iteration, a batch of real subsidence images is randomly sampled from the training dataset, with the batch size set to 32 to 64. Simultaneously, the same number of random noise vectors are sampled from a standard normal distribution and input into the generator network to generate synthetic historical subsidence images.

[0138] Step S54: Calculate the discriminator network's discrimination output for real subsidence images and its discrimination output for synthetic historical subsidence images. The discriminator's loss function uses binary cross-entropy:

[0139] ,

[0140] in: Let the loss function be that of the discriminator. For batch size, For the first A real settlement image, For the first A random noise vector, For the discriminator network, For generator networks, The discriminator's output is the discrimination of the real image. This is the discriminator's output for judging the synthesized image.

[0141] The loss function of the generator is:

[0142] ,

[0143] in: Let be the loss function of the generator. The generator's goal is to maximize the probability that the discriminator will classify the synthesized image as a real image.

[0144] An alternating training strategy is employed. In each training iteration, the generator parameters are first fixed, and the discriminator parameters are updated to minimize the error. Then, with the discriminator parameters fixed, the generator parameters are updated to minimize... Training was performed using the Adam optimizer with a learning rate of 0.0002 and a momentum parameter of... , .

[0145] Step S55: If the discriminator network's accuracy in distinguishing the synthesized historical subsidence images meets the first accuracy range, it indicates that the synthesized historical subsidence images are highly similar to the real subsidence images. The synthesized historical subsidence images are then added to the training dataset. The first accuracy range is preferably below 60%, meaning the discriminator network has difficulty distinguishing between real and synthesized images; at this point, the synthesized images have reached sufficiently high quality. In a specific embodiment, after 200 rounds of training, the discriminator's accuracy in distinguishing the synthesized images drops to 54%, and the generated synthesized subsidence images are visually almost indistinguishable from real subsidence images. Adding the generated 5000 synthesized subsidence images to the original 5000 real subsidence images doubles the size of the training dataset. The Transformer-GRU prediction model is trained on the expanded dataset, and the root mean square error of the test set decreases from 2.8 mm to 2.3 mm, improving prediction accuracy by approximately 18%.

[0146] like Figure 6 As shown, the present invention also provides a method for predicting the time series of roadbed settlement and monitoring three-dimensional deformation based on the above system, the method comprising the following steps:

[0147] Step S1: Receive InSAR satellite remote sensing image data and ground microdeformation sensor data, perform pixel-level registration processing on the InSAR satellite remote sensing image data to generate a pixel-level surface deformation dataset, and perform spatiotemporal alignment processing on the ground microdeformation sensor data to generate a ground deformation dataset.

[0148] Step S2: Divide the pixel-level surface deformation dataset into grids according to the spatial resolution threshold. If the number of pixels in a single grid cell meets the first density range, extract the spatiotemporal feature vector corresponding to the grid cell. The spatiotemporal feature vector includes subsidence rate field features, spatial gradient field features, and time series features. The spatial resolution threshold is 5 meters to 20 meters.

[0149] Step S3: Construct a spatiotemporally coupled Transformer-GRU prediction model. The Transformer-GRU prediction model includes a Transformer encoder layer and a GRU decoder layer. The Transformer encoder layer is used to capture long-term dependencies in the spatiotemporal feature vector, and the GRU decoder layer is used to predict the subsidence trend at future moments. If the time span of the spatiotemporal feature vector conforms to a first time range, the Transformer encoder layer adopts a first layer configuration. If the time span of the spatiotemporal feature vector conforms to a second time range, the Transformer encoder layer adopts a second layer configuration. The first time range is 1 to 3 years, the second time range is 3 to 10 years, the first layer configuration is 4 to 6 layers, and the second layer configuration is 8 to 12 layers.

[0150] Step S4: Based on the prediction results of the Transformer-GRU prediction model, calculate the pixel-level settlement rate field. If the settlement rate of any pixel in the pixel-level settlement rate field meets the first rate range, mark the pixel as a high-risk area and reconstruct the three-dimensional deformation field corresponding to the high-risk area. The calculation accuracy of the pixel-level settlement rate field is 0.5 mm per year to 2 mm per year, and the first rate range is greater than 10 mm per year.

[0151] Step S5: Construct a generative adversarial network, which includes a generator network and a discriminator network. The generator network is used to synthesize historical subsidence images, and the discriminator network is used to distinguish between real subsidence images and synthesized historical subsidence images. If the discriminator network's accuracy in distinguishing synthesized historical subsidence images meets the first accuracy range, the synthesized historical subsidence images are added to the training dataset to enhance the Transformer-GRU prediction model. The first accuracy range is less than 60%.

[0152] Step S6: Generate a roadbed settlement monitoring report based on the three-dimensional deformation field. If the maximum cumulative settlement in the three-dimensional deformation field meets the first settlement threshold range, trigger the first-level early warning signal. If the maximum cumulative settlement meets the second settlement threshold range, trigger the second-level early warning signal. The first settlement threshold range is 30 mm to 50 mm, and the second settlement threshold range is greater than 50 mm.

Claims

1. A roadbed settlement time-series prediction and three-dimensional deformation monitoring system, characterized in that, include: The data acquisition module is used to receive InSAR satellite remote sensing image data and ground micro-deformation sensor data, perform pixel-level registration processing on the InSAR satellite remote sensing image data to generate a pixel-level surface deformation dataset, and perform spatiotemporal alignment processing on the ground micro-deformation sensor data to generate a ground deformation dataset. The spatiotemporal feature construction module is connected to the data acquisition module and is used to divide the pixel-level surface deformation dataset into grids according to a spatial resolution threshold. If the number of pixels in a single grid cell meets the first density range, the spatiotemporal feature vector corresponding to the grid cell is extracted. The spatiotemporal feature vector includes subsidence rate field features, spatial gradient field features, and time series features. The spatial resolution threshold is 5 meters to 20 meters. The prediction model fusion module is connected to the spatiotemporal feature construction module and is used to construct a spatiotemporally coupled Transformer-GRU prediction model, which includes a Transformer encoder layer and a GRU decoder layer. The deformation field reconstruction module is connected to the prediction model fusion module and is used to calculate the pixel-level sedimentation rate field based on the prediction results of the Transformer-GRU prediction model. A synthetic data generation module, connected to the deformation field reconstruction module, is used to construct a generative adversarial network (GAN). The GAN includes a generator network and a discriminator network. The generator network synthesizes historical subsidence images, and the discriminator network distinguishes between real subsidence images and the synthesized historical subsidence images. If the discriminator network's accuracy in distinguishing the synthesized historical subsidence images meets a first accuracy range, the synthesized historical subsidence images are added to the training dataset to enhance the Transformer-GRU prediction model. The first accuracy range is below 60%. The monitoring and early warning module is connected to the deformation field reconstruction module and is used to generate a roadbed settlement monitoring report based on the three-dimensional deformation field. If the maximum cumulative settlement in the three-dimensional deformation field meets the first settlement threshold range, a first-level early warning signal is triggered. If the maximum cumulative settlement meets the second settlement threshold range, a second-level early warning signal is triggered. The first settlement threshold range is 30 mm to 50 mm, and the second settlement threshold range is greater than 50 mm.

2. The system according to claim 1, characterized in that, If the settlement rate of any pixel in the pixel-level settlement rate field conforms to the first rate range, the pixel is marked as a high-risk area, and the three-dimensional deformation field corresponding to the high-risk area is reconstructed. The calculation accuracy of the pixel-level settlement rate field is 0.5 mm per year to 2 mm per year, and the first rate range is greater than 10 mm per year. The Transformer encoder layer is used to capture long-term dependencies in the spatiotemporal feature vector, and the GRU decoder layer is used to predict the subsidence trend at future times. If the time span of the spatiotemporal feature vector conforms to a first time range, the Transformer encoder layer adopts a first layer configuration; if the time span of the spatiotemporal feature vector conforms to a second time range, the Transformer encoder layer adopts a second layer configuration. The first time range is 1 to 3 years, the second time range is 3 to 10 years, the first layer configuration is 4 to 6 layers, and the second layer configuration is 8 to 12 layers. The data acquisition module is specifically used for: Receive multi-temporal InSAR satellite remote sensing images, which include at least 15 SAR images; Persistent scatterers are identified in the multi-temporal InSAR satellite remote sensing images, and a set of persistent scattering points is extracted. The spatial density of the persistent scattering point set is 300 to 1000 persistent scattering points per square kilometer. Temporal InSAR processing is performed on the persistent scattering point set to calculate the deformation time series of each persistent scattering point and generate the pixel-level surface deformation dataset. Receive measurement data from foundation micro-deformation sensors at multiple monitoring points, the measurement data including displacement and timestamp; The measurement data of the ground microdeformation sensor is matched with the spatial coordinates of the persistent scattering point. If the matching distance is less than the spatial matching threshold, the time series is aligned and fused. The spatial matching threshold is 20 meters to 50 meters.

3. The system according to claim 1, characterized in that, The spatiotemporal feature construction module includes: The grid division submodule is used to divide the pixel-level surface deformation dataset into multiple grid units according to the spatial resolution threshold, and to count the number of pixels in each grid unit. The density filtering submodule, connected to the grid division submodule, is used to determine whether the number of pixels conforms to the first density range. If it does, the grid unit is marked as a valid unit; if it does not, the grid unit is marked as an invalid unit. The first density range includes 10 to 50 pixels in each grid unit. The feature extraction submodule, connected to the density filtering submodule, is used to extract the spatiotemporal feature vector from the grid cells marked as valid cells. The spatiotemporal feature vector includes the average settling rate of the pixels in the grid cell, the spatial standard deviation of the settling rate, the time rate of change of the settling rate, and the settling gradient with adjacent grid cells.

4. The system according to claim 1 or 2, characterized in that, The settlement velocity field features include pixel-level settlement velocity and spatial distribution pattern of settlement velocity. The calculation method for the pixel-level settlement velocity is as follows: Linear regression fitting is performed on the deformation time series of each pixel, and the slope of the regression line is calculated as the pixel-level settling rate. The pixel-level sedimentation rate is smoothed by Gaussian filtering, with a filtering window size of 3 to 5 pixels. Calculate the spatial gradient field of the pixel-level settling rate, where the spatial gradient field reflects the difference in settling rate between adjacent pixels; If the gradient value at any location in the spatial gradient field conforms to the first gradient range, the location is marked as a subsidence anomaly region. The first gradient range is greater than 5 mm per year per 100 meters.

5. The system according to claim 1 or 2, characterized in that, The method for constructing the Transformer-GRU prediction model includes: Construct a Transformer encoder layer, which includes a multi-head self-attention mechanism and a feedforward neural network, wherein the number of heads in the multi-head self-attention mechanism is 4 to 8; The spatiotemporal feature vectors are input into the Transformer encoder layer, and the spatiotemporal correlation weights between the feature vectors are calculated through the multi-head self-attention mechanism. Construct a GRU decoder layer, which includes multiple GRU units, and the hidden layer dimension of the GRU units is 128 to 512 dimensions; The output of the Transformer encoder layer is used as the input of the GRU decoder layer, and the GRU decoder layer is used to predict the settlement value at multiple future time points, with the time span of the multiple future time points being 1 to 12 months.

6. The system according to claim 2, characterized in that, The processing flow of the deformation field reconstruction module includes: If the settling rate of a pixel in the pixel-level settling rate field conforms to the first rate range, extract the three-dimensional coordinates and settling rate of the pixel. Based on the three-dimensional coordinates and settlement rate, the three-dimensional deformation field of the high-risk area is reconstructed using the Kriging interpolation method, which includes calculating the semi-variogram function and constructing the interpolation weight matrix. The three-dimensional deformation field is spatially smoothed to remove high-frequency noise introduced during interpolation. The deformation gradient field and curvature field of the three-dimensional deformation field are calculated. The deformation gradient field reflects the spatial variation trend of the deformation, and the curvature field reflects the spatial curvature of the deformation.

7. The system according to claim 1, characterized in that, The training process of the synthetic data generation module includes: The generator network is constructed using a convolutional neural network architecture, including an encoder part and a decoder part; The discriminator network is constructed, and the discriminator network adopts a convolutional neural network architecture to distinguish between real subsidence images and the synthetic historical subsidence images; The real subsidence images are randomly sampled from the pixel-level surface deformation dataset, and the synthetic historical subsidence images are generated from random noise vectors. The discriminator network is calculated to distinguish the real settlement image and the synthetic historical settlement image, and the parameters of the generator network and the discriminator network are updated based on the discrimination output. If the discrimination accuracy of the discriminator network for the synthetic historical subsidence image meets the first accuracy range, it indicates that the synthetic historical subsidence image is highly similar to the real subsidence image, and the synthetic historical subsidence image is added to the training dataset.

8. The system according to claim 1, characterized in that, The monitoring and early warning module is also used for: The spatial distribution characteristics of roadbed settlement are calculated based on the three-dimensional deformation field. The spatial distribution characteristics include the location of the settlement center, the range of settlement influence, and the settlement morphology. The subsidence development trend at future times is calculated based on the prediction results of the Transformer-GRU prediction model. The subsidence development trend includes the subsidence acceleration and the estimated time when the subsidence reaches the danger threshold. Generate the roadbed settlement monitoring report, which includes the current settlement status, historical settlement evolution curves, future settlement prediction curves, early warning levels, and recommended remedial measures; If the warning level is Level 1, a monitoring reminder message will be sent to the monitoring terminal. If the warning level is Level 2, an emergency response message is sent to the management terminal.

9. The system according to claim 1, characterized in that, The system also includes: An adaptive optimization module is used to monitor the prediction error of the Transformer-GRU prediction model. If the prediction error is within a first error range, the parameters of the Transformer-GRU prediction model are kept unchanged. If the prediction error is within a second error range, the model parameter update process is triggered. The first error range is a root mean square error of less than 3 mm, and the second error range is a root mean square error of greater than 5 mm. The model parameter update process includes obtaining the latest pixel-level surface deformation dataset and the foundation deformation dataset, incrementally training the Transformer-GRU prediction model, and updating the weight parameters of the Transformer encoder layer and the GRU decoder layer.

10. A method for predicting the time series of roadbed settlement and monitoring three-dimensional deformation based on the system described in claim 1, using the system described in any one of claims 1-9, characterized in that, include: The system receives InSAR satellite remote sensing image data and ground-based micro-deformation sensor data, performs pixel-level registration processing on the InSAR satellite remote sensing image data to generate a pixel-level surface deformation dataset, and performs spatiotemporal alignment processing on the ground-based micro-deformation sensor data to generate a ground-based deformation dataset. The pixel-level surface deformation dataset is divided into grids according to a spatial resolution threshold. If the number of pixels in a single grid cell meets the first density range, the spatiotemporal feature vector corresponding to the grid cell is extracted. The spatiotemporal feature vector includes subsidence rate field features, spatial gradient field features, and time series features. The spatial resolution threshold is 5 meters to 20 meters. A spatiotemporally coupled Transformer-GRU prediction model is constructed, comprising a Transformer encoder layer and a GRU decoder layer. The Transformer encoder layer is used to capture long-term dependencies in the spatiotemporal feature vector, and the GRU decoder layer is used to predict the subsidence trend at future times. If the time span of the spatiotemporal feature vector conforms to a first time range, the Transformer encoder layer adopts a first layer configuration; if the time span of the spatiotemporal feature vector conforms to a second time range, the Transformer encoder layer adopts a second layer configuration. The first time range is 1 to 3 years, the second time range is 3 to 10 years, the first layer configuration is 4 to 6 layers, and the second layer configuration is 8 to 12 layers. Based on the prediction results of the Transformer-GRU prediction model, a pixel-level settlement rate field is calculated. If the settlement rate of any pixel in the pixel-level settlement rate field conforms to the first rate range, the pixel is marked as a high-risk area, and the three-dimensional deformation field corresponding to the high-risk area is reconstructed. The calculation accuracy of the pixel-level settlement rate field is 0.5 mm per year to 2 mm per year, and the first rate range is greater than 10 mm per year. A generative adversarial network (GAN) is constructed, comprising a generator network and a discriminator network. The generator network synthesizes historical subsidence images, and the discriminator network distinguishes between real subsidence images and the synthesized historical subsidence images. If the discriminator network's accuracy in distinguishing the synthesized historical subsidence images meets a first accuracy range, the synthesized historical subsidence images are added to the training dataset to enhance the Transformer-GRU prediction model. The first accuracy range is below 60%. A roadbed settlement monitoring report is generated based on the three-dimensional deformation field. If the maximum cumulative settlement in the three-dimensional deformation field meets the first settlement threshold range, a first-level early warning signal is triggered. If the maximum cumulative settlement meets the second settlement threshold range, a second-level early warning signal is triggered. The first settlement threshold range is 30 mm to 50 mm, and the second settlement threshold range is greater than 50 mm.

Images