High-speed rail surface deformation LSTM prediction method and device based on PS point metadata feature clustering

Abstract

The invention discloses a high-speed rail surface deformation LSTM prediction method based on PS point metadata feature clustering. The method comprises the following steps: step 1, extracting PS point metadata; step 2, screening key features of PS point metadata; step 3, dividing the PS points into a plurality of clusters; 4, performing trend fitting on the deformation time sequence data of each PS point, and determining a deformation trend type according to a fitting residual error; 5, aggregating PS points which are spatially adjacent and have similar characteristics into a same proton region; step 6, inputting the time sequence deformation quantity of the PS point for prediction; and step 7, evaluating the prediction precision through the error index, and outputting a deformation prediction result. According to the prediction method provided by the invention, PS point information can be effectively utilized, the prediction precision of surface deformation is remarkably improved, and important theoretical reference and data support are provided for surface subsidence monitoring of linear projects such as high-speed rails.

Description

Technical Field

[0001] The present invention relates to the technical field of geological disaster simulation and prediction. More specifically, the present invention relates to a high-speed rail surface deformation LSTM prediction method and device based on PS point metadata feature clustering. Background Art

[0002] Ground subsidence is a common geological disaster that poses a serious threat to urban infrastructure. With the acceleration of the urbanization process, the construction of high-speed railway projects has become increasingly intensive. High-speed rail has become the main mode of transportation for people, and its safe and stable operation is of vital importance. Therefore, high-precision surface settlement monitoring and prediction of high-speed rail and its surrounding areas are key links to ensure its safe operation.

[0003] Traditional surface settlement monitoring methods, such as leveling, Global Navigation Satellite System (GNSS) measurement, and sensor methods, although having advantages in some aspects, also face many limitations. These methods are often restricted by factors such as the environment and weather, consuming a large amount of human, material, and time costs, and having a low sampling rate, insufficient positioning accuracy, and limited monitoring range, making it difficult to meet the requirements of large-scale and high-precision surface settlement monitoring of high-speed rail. Synthetic Aperture Radar Interferometry (InSAR) technology, as an active microwave remote sensing technology, has the advantages of all-weather, large-scale, and high-precision monitoring, providing a new way for large-scale surface deformation monitoring. Among them, the Persistent Scatterer Interferometric Synthetic Aperture Radar (PS-InSAR) technology overcomes the problems of temporal and geometric decorrelation by identifying persistent scatterers, increasing the density of observation points, and meeting the requirements of time-series settlement monitoring.

[0004] Although InSAR technology has shown many advantages in settlement monitoring, the existing InSAR surface deformation prediction methods still have some limitations. Currently, InSAR surface deformation prediction methods are mainly divided into three categories: physical models, mathematical models, and machine learning models. Traditional physical model predictions often rely on parameters that are difficult to obtain, resulting in limited prediction accuracy. For example, MULDER et al. used a numerical weather prediction (NWP) model to predict the atmospheric delay signal in InSAR, but this model is difficult to accurately separate the atmospheric signal from other signals, leading to deviations in the prediction results. Zeng et al. constructed an integrated driving model based on physical fusion (IDNPM) to evaluate the likelihood of post-earthquake landslides, but this model has a strong dependence on the parameters of the Newmark physical model. On the other hand, mathematical models are difficult to fully capture the complexity of actual deformations, and there are deviations in the prediction results. For example, XIONG et al. used an exponential function model and the Asaoka method to predict the settlement of reclamation areas and determine the time when the settlement tends to stabilize, but it is not applicable to all types of reclamation areas. Yin et al. used the cumulative distribution function (CDF) of the Weibull distribution to simulate and predict the time evolution of surface uplift, but this method lacks in-depth exploration of the settlement mechanism. Zhang et al. used the statistical distribution of deformation and the Hurst exponent to predict the deformation trend and effectively predicted the stability trend of landslides. However, the applicability of the Hurst exponent under other geological and landslide type conditions remains to be verified. In recent years, machine learning methods have shown potential in deformation prediction, but most existing studies focus on single-point prediction and large-scale prediction, and there are still deficiencies. For example, HILL et al. randomly selected points with seasonal characteristics for SARIMA and LSTM predictions, which limits the general applicability of the research results. Some research scholars used GM-SVR, Transformer, Grey-Markov models to achieve large-scale spatial settlement predictions. Although some research scholars also used models such as H-LSTM, Conv-LSTM, DACLnet for predictions considering spatial information, these research methods still lack comprehensive information utilization and effective consideration of spatial correlation and the characteristic information of permanent scatterer (PS) point metadata.

[0005] Therefore, how to improve the large-scale and high-precision surface deformation prediction ability while fully utilizing the advantages of InSAR technology and considering the complexity of surface deformation, spatial correlation, and the characteristic information of PS point metadata remains a key problem to be solved urgently. Summary of the Invention

[0006] An object of the present invention is to solve at least the above problems and provide at least the advantages described later.

[0007] Another object of the present invention is to provide a high-speed rail surface deformation LSTM prediction method based on PS point metadata feature clustering, which meets the urgent need for accurate settlement prediction in the safe operation of high-speed rail and overcomes the limitation that the traditional InSAR monitoring method fails to fully utilize PS point information in complex surface deformation and time series prediction.

[0008] To achieve these objects and other advantages of the present invention, a high-speed rail surface deformation LSTM prediction method based on PS point metadata feature clustering is provided, which includes the following steps:

[0009] Step 1: Obtain multi-temporal synthetic aperture radar (SAR) time series data of the target area, process it through PS-InSAR technology to obtain PS points in the target area, and further extract PS point metadata, including the deformation time series data of PS points and the time series deformation amounts of PS points;

[0010] Step 2: Calculate the feature importance scores of the PS point metadata extracted in Step 1 based on the random forest model, and screen the key features of the PS point metadata;

[0011] Step 3: Perform clustering analysis on the key features screened in Step 2 to complete the clustering division of PS points;

[0012] Step 4: Fit the trend of the deformation time series data of each PS point, and determine its deformation trend type according to the fitting residuals;

[0013] Step 5: Combine the clustering division results in Step 3 and the deformation trend types in Step 4, and aggregate PS points that are spatially adjacent and have the same deformation trend into the same homogeneous sub-region;

[0014] Step 6: Build a long short-term memory (LSTM) neural network model in each sub-region, and input the time series deformation amounts of PS points for prediction;

[0015] Step 7: Evaluate the prediction accuracy through error indicators and output the deformation prediction results.

[0016] Preferably, in the high-speed rail surface deformation LSTM prediction method based on PS point metadata feature clustering, the PS-InSAR technology processing in Step 1 includes the following steps:

[0017] S11: SAR image registration and interferogram generation;

[0018] S12: Use the SRTM 30m digital elevation model (DEM) to remove the topographic phase error;

[0019] S13: Separate the atmospheric delay phase through a spatio-temporal filtering method;

[0020] S14. Screen PS points based on the coherence threshold, and retain the PS points with a coherence value greater than 0.85;

[0021] S15. Generate time-series surface deformation.

[0022] Preferably, in the high-speed rail surface deformation LSTM prediction method based on PS point metadata feature clustering, the parameter settings of the random forest model in step two are as follows: the number of decision trees is 100, the minimum number of leaf nodes is 5, and the division ratio of the training set to the test set is 7:3.

[0023] Preferably, in the high-speed rail surface deformation LSTM prediction method based on PS point metadata feature clustering, the method for determining the number of clusters in step three includes:

[0024] S31. Calculate the sum of squared errors within clusters (SSE) corresponding to different numbers of clusters based on the elbow method;

[0025] S32. Select the inflection point position where the SSE decline trend significantly slows down as the optimal number of clusters, (ii) select the inflection point position where the SSE decline trend significantly slows down as the optimal number of clusters;

[0026] S33. Iteratively optimize the distribution of cluster centers through the three-dimensional space Euclidean distance, and divide the PS points into multiple cluster clusters. Preferably, in the high-speed rail surface deformation LSTM prediction method based on PS point metadata feature clustering, the Euclidean distance calculation formula is:

[0027]

[0028] where (x1, x2, x3) are the three eigenvalue of the point target, and (c1, c2, c3) are the three eigenvalue of the cluster center.

[0029] Preferably, in the high-speed rail surface deformation LSTM prediction method based on PS point metadata feature clustering, the sub-region aggregation method in step five specifically includes:

[0030] S51. Construct a spatial proximity network of PS points based on the Delaunay triangulation algorithm;

[0031] S52. Set the spatial proximity distance threshold and the deformation trend consistency threshold, and merge the adjacent PS points that meet the threshold into the same sub-region;

[0032] S53. Use morphological dilation operation to fill the isolated points in the sub-region to ensure spatial continuity.

[0033] Preferably, the high-speed rail surface deformation LSTM prediction method based on PS point metadata feature clustering further includes step eight, dynamic early warning, which includes the following steps:

[0034] S81. Establish a sliding update mechanism for model parameters, and retain the most recent 24 periods of data as the training set each time new SAR images are obtained;

[0035] S82. Set the deformation warning threshold to trigger a graded warning signal when the cumulative deformation exceeds ±15 mm.

[0036] Preferably, in the high-speed rail surface deformation LSTM prediction method based on PS point metadata feature clustering, the dynamic warning step further includes:

[0037] S83. Analyze the long-range dependence of the deformation sequence based on the Hurst index and correct the warning threshold;

[0038] S84. When the cumulative deformation triggers a warning, combine the spatial overlay analysis of the geographic information system (GIS) to generate a settlement risk heat map and an engineering reinforcement suggestion plan;

[0039] S85. Introduce a transfer learning mechanism and use the historical deformation data of adjacent areas to optimize the local model parameters.

[0040] The present invention also provides a high-speed rail surface deformation LSTM prediction device based on PS point metadata feature clustering, which includes: a processor and a memory; wherein, the memory is used to store a computer program; the processor is used to load and execute the computer program to implement the above prediction method.

[0041] The present invention has at least the following beneficial effects:

[0042] The present invention proposes an LSTM prediction method based on PS point metadata feature clustering (PSFC-LSTM) aiming to improve the prediction accuracy of ground settlement of high-speed rail lines. This method first extracts the key features of PS point metadata and uses the K-Means algorithm to cluster them into three clusters. At the same time, the deformation trend of PS points is fitted and divided into four types (sine type, linear rising type, linear stable type, and linear falling type). Then, combining the clustering results of PS points and the deformation trend fitting results, the study area is divided into 12 sub-areas. In each sub-area, the LSTM model is used to predict the cumulative deformation amount respectively. The method provided by the present invention not only considers the PS point metadata features but also takes into account the prediction method of the deformation trend, effectively improving the prediction accuracy. Using the method of the present invention, 28 scenes of Sentinel-1 SAR data from October 17, 2020, to October 24, 2021, were selected to conduct PS-InSAR monitoring and cumulative deformation amount prediction on the Taijiao Railway (Jincheng section) in China. The results show that for the prediction of the cumulative deformation amount in the last four periods of the study period, compared with the traditional LSTM method, the prediction RMSE of the PSFC-LSTM method of the present invention is reduced by an average of 13.8%, and the MAE is reduced by an average of 26.0%, R2 The average improvement is 4.4%. Therefore, the PSFC-LSTM method proposed by the present invention can effectively utilize PS point information, significantly improve the prediction accuracy of ground deformation, and provide important theoretical reference and data support for the ground settlement monitoring of linear projects such as high-speed railways.

[0043] Other advantages, objectives and features of the present invention will be partially reflected by the following description, and partially will be understood by those skilled in the art through the research and practice of the present invention. Brief Description of the Drawings

[0044] Figure 1 It is a schematic flow chart of the high-speed railway ground deformation LSTM prediction method based on PS point metadata feature clustering according to the present invention;

[0045] Figure 2 It is a technical flow chart of the high-speed railway ground deformation LSTM prediction method based on PS point metadata feature clustering according to the present invention;

[0046] Figure 3 It is a structure diagram of the PSFC-LSTM model of the present invention;

[0047] Figure 4 It is a statistical bar chart of the importance results of PS point metadata features in Example 1 of the present invention;

[0048] Figure 5 It is an elbow graph of clusters in Example 1 of the present invention;

[0049] Figure 6 The clustering result based on features in Example 1 of the present invention;

[0050] Figure 7 It is a deformation trend graph of PS points in Example 1 of the present invention;

[0051] Figure 8 It is a fitting result based on the deformation trend in Example 1 of the present invention;

[0052] Figure 9 It is a comparison graph of PS point deformation prediction results in Example 1 of the present invention;

[0053] Figure 10 It is a residual distribution graph between the predicted values and the true values of the PSFC-LSTM prediction method and the traditional method in Example 1 of the present invention;

[0054] Figure 11 It is a scatter plot of the predicted values and the true values of the PSFC-LSTM prediction method and the traditional method in Example 1 of the present invention;

[0055] Figure 12 It is a graph of the deformation and prediction curve changes of four typical points in Example 1;

[0056] Figure 13 This is the result graph of the PS point deformation prediction by the RFC-LSTM and TY-LSTM methods in Embodiment 2 of the present invention;

[0057] Figure 14 This is the prediction residual distribution graph corresponding to the RFC-LSTM and TY-LSTM methods in Embodiment 2 of the present invention;

[0058] Figure 15 This is the scatter plot between the predicted values and the true values of various prediction methods in Embodiment 2 of the present invention;

[0059] Figure 16 This is the deformation and prediction curve change graph of four typical points in Embodiment 2 of the present invention. Detailed implementation manners

[0060] The following further describes the present invention in detail with reference to the accompanying drawings and embodiments, so that those skilled in the art can implement it according to the description in the specification.

[0061] It should be understood that the terms such as "having", "comprising" and "including" used in the present invention do not exclude the existence or addition of one or more other elements or their combinations.

[0062] It should be noted that the experimental methods described in the following implementation schemes are all conventional methods unless otherwise specified, and the reagents and materials can be obtained from commercial channels unless otherwise specified.

[0063] In the description of the present invention, the orientation or positional relationship indicated by terms such as "transverse", "longitudinal", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc. is based on the orientation or positional relationship shown in the accompanying drawings. It is only for the convenience of describing the present invention and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation to the present invention.

[0064] As Figure 1 shown, the present invention provides a high-speed rail surface deformation LSTM prediction method based on PS point metadata feature clustering, which includes the following steps:

[0065] Step 1: Obtain the synthetic aperture radar (SAR) time series data of the target area, process it through the PS-InSAR technology to obtain the PS points in the target area, and further extract the PS point metadata information, including the time series deformation amount of the PS points (i.e., the cumulative deformation amount at each time point - the observed data), and the metadata such as the spatial coordinates, elevation and coherence coefficient of the PS points;

[0066] The processing of PS-InSAR technology includes the following steps:

[0067] S11. SAR image registration and interferogram generation;

[0068] S12. Using the SRTM 30m digital elevation model (DEM) to remove topographic phase errors;

[0069] S13. Separating the atmospheric delay phase through spatio-temporal filtering methods;

[0070] S14. Screening PS points based on the coherence threshold, and retaining PS points with a coherence value greater than 0.85;

[0071] S15. Generating time-series surface deformation. Step 2: Calculating the feature importance scores based on the random forest model and screening the key features of the PS point metadata; calculating the feature importance scores of the PS point metadata extracted in Step 1 based on the random forest model, and selecting the three key features with the highest importance scores, including longitude, reference ellipsoid height, and latitude; the parameter settings of the random forest model are: the number of decision trees is 100, the minimum number of leaf nodes is 5, and the division ratio of the training set to the test set is 7:3;

[0072] Step 3: Conducting cluster analysis on the key features screened in Step 2 to complete the cluster division of PS points; for the key features selected in Step 2, using the K-Means clustering algorithm combined with the elbow method to determine the optimal number of clusters: calculating the within-cluster sum of squared errors corresponding to different numbers of clusters, and taking the inflection point of the change of the sum of squared errors with the number of clusters as the optimal number of clusters to complete the cluster division of PS points;

[0073] The method for determining the number of clusters in the cluster division process includes:

[0074] S31. Calculating the within-cluster sum of squared errors corresponding to different numbers of clusters based on the elbow method;

[0075] S32. Selecting the inflection point position where the SSE (sum of squared errors) decreases significantly and slowly as the optimal number of clusters;

[0076] S33. Iteratively optimizing the distribution of cluster centers through the three-dimensional spatial Euclidean distance, and dividing the PS points into multiple cluster clusters, that is, completing the cluster division of PS points; the Euclidean distance calculation formula is:

[0077]

[0078] where (x1, x2, x3) are the three eigenvalue of the point target, and (c1, c2, c3) are the three eigenvalue of the cluster center;

[0079] Step 4: Perform trend fitting on the deformation time series data of each PS point, and determine its deformation trend type according to the fitting residuals; further divide the PS points into 12 sub-regions according to the deformation trend classification results, where each clustering cluster corresponds to 4 deformation trend types, namely sine type, linear ascending type, linear stable type, and linear descending type;

[0080] Step 5: Combine the clustering division results in Step 3 and the deformation trend types in Step 4, and aggregate the PS points with spatial proximity and consistent deformation trends into homogeneous sub-regions; the specific sub-region aggregation method includes:

[0081] S51: Construct a spatial proximity network of PS points based on the Delaunay triangulation algorithm;

[0082] S52: Set the spatial proximity distance threshold and the deformation trend consistency threshold, and merge the adjacent PS points that meet the thresholds into the same sub-region;

[0083] S53: Use morphological dilation operation to fill the isolated points in the sub-region to ensure spatial continuity;

[0084] Step 6: Build a long short-term memory (LSTM) neural network model in each sub-region, and input the time series deformation amounts of PS points for prediction;

[0085] Step 7: Evaluate the prediction accuracy through error metrics, and output the deformation prediction results; the error metrics include root mean square error, mean absolute error, and coefficient of determination

[0086] The present invention proposes a new method for predicting surface deformation LSTM based on PS point metadata feature clustering (PSFC-LSTM), and the specific technical process is as Figure 2 shown. This method mainly includes the following key steps: First, based on the acquired SAR data covering the study area (target area), perform PS-InSAR ground settlement information processing to obtain the PS point data of the target area, and further extract the metadata information of the PS points. Subsequently, perform feature selection, PS point clustering, deformation type discrimination, and model prediction in sequence. Finally, use three indicators, namely root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R 2 ) to quantitatively evaluate the reliability of the prediction method proposed by the present invention.

[0087] As a preferred technical solution, the surface deformation LSTM prediction method based on PS point metadata feature clustering of the present invention further includes Step 8: Dynamic early warning includes the following steps:

[0088] S81: Establish a sliding update mechanism for model parameters, and retain the latest 24 periods of data as the training set each time a new SAR image is acquired;

[0089] S82. Set the deformation warning threshold to trigger a graded warning signal when the cumulative deformation exceeds ±15 mm.

[0090] S83. Analyze the long-term dependence of the deformation sequence based on the Hurst exponent and correct the warning threshold.

[0091] S84. When the cumulative deformation triggers a warning, combine the spatial overlay analysis of the Geographic Information System (GIS) to generate a settlement risk heat map and an engineering reinforcement suggestion plan.

[0092] S85. Introduce a transfer learning mechanism and use the historical deformation data of adjacent areas to optimize the local model parameters.

[0093] Setting dynamic warnings can trigger warnings according to the prediction results and take corresponding measures in a timely manner.

[0094] The present invention also provides a surface deformation LSTM prediction device based on the clustering of PS point metadata features, which is a device including a processor (CPU / MCU / SOC) and a memory (ROM / RAM), such as: a desktop computer, a portable computer, a smart phone, etc. In particular, a computer program is stored in the memory, and when the processor loads and executes the computer program, it can implement all or part of the steps of the above prediction method.

[0095] I. The PS-InSAR (Persistent Scatterer InSAR) technology selects stable scatterers that can maintain high coherence over a long time and a long spatial baseline as targets, effectively overcoming the influence of spatio-temporal decorrelation on the interference signal, thereby accurately obtaining surface deformation information. These stable scatterers are usually buildings, bare rocks, etc., which are insensitive to spatio-temporal baseline decorrelation and less affected by speckle noise, and maintain high coherence for a long time.

[0096] Before performing the prediction processing of the time-series InSAR data, it is first necessary to preprocess the original data, including data import and cropping to the study area to reduce the data volume for easy storage and subsequent processing. The key steps of PS-InSAR processing include: precise image registration to ensure the accurate correspondence between pixels; using a high-precision digital elevation model (DEM) to remove the influence of the topographic phase; removing the flat-earth phase to eliminate the phase error caused by the earth's curvature; selecting high-quality PS points based on coherence analysis; using methods such as spatio-temporal filtering to remove the atmospheric delay error; and finally, obtaining the time series of surface deformation through time-series analysis.

[0097] The modeling expression in the deformation stage is shown in Equation (3).

[0098]

[0099] Among them, represents the interference phase observed in the interferogram, represents the phase error caused by the reference ellipsoid, represents the phase error caused by terrain undulation, represents the surface deformation in the satellite line-of-sight direction, represents the atmospheric delay error, represents the random noise.

[0100] The selection of PS point metadata features is to explore the contribution degree of the original attributes of PS points to surface deformation. As shown in Table 1, the original attribute information of PS points can describe the surface deformation monitoring results from multiple perspectives, including data quality, geometric position, line-of-sight geometry, and elevation correction information, so as to ensure the reliability, particularity, and accuracy of the data. Among them, the attributes evaluating data quality include coherence and quality index. The higher the coherence, the higher the accuracy of the deformation signal; the larger the quality index, the better the stability of the intensity of PS points in the time series, with smaller changes, the more reliable the interference signal, and the more accurate the deformation measurement. The geometric position attributes include longitude, latitude, height above the reference ellipsoid, azimuth and range coordinates in the pixel coordinate system, and geoid height. The geometric position attributes are an important part of the PS-InSAR processing results. PS points at different positions have different deformation characteristics and are the basis for deformation analysis. The line-of-sight geometry attributes include the line-of-sight incidence angles in the azimuth and vertical directions. The line-of-sight geometry will affect the direction of deformation measurement. By correcting the deformation results measured at different satellite incidence angles, more real deformation rates can be obtained. The elevation correction information includes the average accuracy of elevation measurement. Since the surface deformation measurement is essentially a distance change and the elevation value is closely related to the distance change, the elevation accuracy directly affects the accuracy of deformation.

[0101] Table 1 PS point metadata

[0102]

[0103]

[0104] II. Random forest feature selection

[0105] The diverse metadata attribute information of PS points is all related to the deformation of PS points, which increases the complexity of the model. Therefore, extracting the key features with greater influence among them can achieve data dimensionality reduction, improve the model processing efficiency without losing important information. In the present invention, a random forest model is adopted to achieve the feature extraction of PS points.

[0106] The random forest model is a learner based on decision trees. The principle of this model is relatively simple. Its core idea is to select the most favored category from the results of multiple decision trees through a voting mechanism to construct a "forest". Suppose the random forest consists of m trees, and the mapping relationship between the input and output of the model is Y = f(X), where the input vector of the model is X = {x1, x2, …, x n}. For each tree, there is a corresponding output Y1 = f1(X), Y2 = f2(X), …, Y m = f m (X). The output of the final model is obtained by calculating the average of the output results of m trees. During the generation of decision trees, the model selects samples according to the principle of random sampling. In addition, the model also ranks the importance (γ) of the n input variables X and normalizes the ranking results so that the sum of γ1, γ2, …, γ n is 1, that is

[0107]

[0108] According to relevant research, the top three of the normalized importance values are the selected features.

[0109] III. K-Means Clustering

[0110] Figure 3 Shows the structure of the PSFC-LSTM model. This model first performs K-Means clustering on important features. The K-Means clustering algorithm is an unsupervised learning method aimed at dividing point targets into K clusters. The core idea of this algorithm is iterative optimization, minimizing the distance between point targets within a cluster and the cluster center to achieve clustering. The K-Means algorithm has certain advantages in processing multi-dimensional data. Its main steps are as follows:

[0111] ① Initialization

[0112] Randomly select K point targets as the initial cluster centers.

[0113] ② Assignment

[0114] Calculate the distance between each point target and all cluster centers. In three-dimensional space, this is usually done through the Euclidean distance, and the distance formula is:

[0115]

[0116] where (x1, x2, x3) are the three feature values of the point target, and (c1, c2, c3) are the three feature values of the cluster center.

[0117] Assign each point target to the nearest cluster center to form K clusters.

[0118] ③ Update

[0119] For each cluster, calculate the average value of all point targets in the cluster, and this average value becomes the new cluster center. The average value of each feature is

[0120]

[0121] where C i are all point targets in the i-th cluster, N i is the number of points in the i-th cluster, and x is the point target.

[0122] ④ Iterate

[0123] Repeat the assignment and update steps until the preset number of iterations is reached.

[0124] In this process, the elbow method is used to select the number of clusters. The elbow method is a common technique for determining the value of K in cluster analysis. Its core idea is to observe the change trend of the sum of squared errors within clusters (SSE) under different numbers of clusters, and find an optimal K value, which is usually the point where the slope of the curve begins to slow down significantly.

[0125] IV. Discrimination of Deformation Signal Trends

[0126] In the deformation signal discrimination part, it is assumed that the changes of the surface deformation signal in the time dimension mainly show three trends: sinusoidal, linear, and exponential trends. Among them, the sinusoidal trend mainly reflects the periodic changes in surface deformation caused by seasonal factors and can be simulated by complex harmonics; the linear trend is usually related to the long-term and uniform settlement or uplift changes of the surface; while the exponential trend reflects the process of rapid growth or decay of surface deformation when some geological activities are relatively intense. The signal discrimination function is

[0127]

[0128] where f1(t) represents the fitted sine function, f2(t) represents the fitted linear function, f3(t) represents the exponential function fitting, t represents time, A represents the amplitude, ω is the angular frequency, is the phase, a is the slope, b is the intercept, K is the initial value, and α is the growth rate.

[0129] To find the best fitting function, the sum of squared residuals (RSS) is used as an index to evaluate the signal error, that is, each function corresponds to an RSS value,

[0130]

[0131] where RSSn Represents the RSS index of each sub-function, y i Represents the actual value of the i-th observation, Represents the corresponding fitted value of the model.

[0132] Select minRSS n The sub-function represented by is the best fitting function. Using the selected best fitting function as the selection criterion, the present invention finally confirms four deformation trends: sine type, linear rising type, linear stable type, and linear falling type.

[0133] V. LSTM Model

[0134] The present invention uses the LSTM (Long Short-Term Memory) model to predict the deformation amount of points with the same deformation trend within each cluster. In the LSTM model part, assume that x represents the input, h represents the output, and the state units at different time steps are represented by t-1, t, and t+1 respectively. Each unit contains key structures such as a forget gate, an input gate, and an output gate. The forget gate forgets relatively unimportant information in the model from the memory unit, which is determined by the following formula:

[0135] f t = sigmoid(W f ·[h t-1 , x t +b f ) (7)

[0136] Among them, h t-1 represents the output of the previous state unit, x t represents the input at state t, W f represents the weight, and b f represents the bias value.

[0137] The input gate determines the value to be updated, and the calculation formula is as follows

[0138] i t = sigmoid(W i ·[h t-1 , x t +b i ) (8)

[0139]

[0140] Among them, W i and b i represent the weight and bias value respectively, and W C and b C represent the weight and bias value of the new candidate value respectively.

[0141] The cell unit is updated as follows

[0142]

[0143] The output gate processes the output h through the sigmoid function and the tanh function t , and the calculation formula is as follows

[0144] O t = sigmoid(W o ·[h t-1 , x t +b o ) (11)

[0145] h t = O t ·tanh(C t ) (12)

[0146] VI. Precision Evaluation Indexes

[0147] The prediction precision of the present invention is evaluated by indexes such as root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R 2 ).

[0148]

[0149]

[0150]

[0151] Among them, y i is the true value of the deformation amount of each PS point, is the predicted value of the corresponding deformation amount, is the average value of the true values of the deformation amounts of all PS points, and m is the number of all PS points.

[0152] Example 1

[0153] In this example, the Jincheng section of the Taiyuan-Jiaozuo High-Speed Railway is selected as the target area, and the Sentinel-1 SAR ascending orbit image dataset covering the Taiyuan-Jiaozuo High-Speed Railway (Jincheng section) published on the website of the European Space Agency (ESA) with a time span from October 17, 2020 to October 24, 2021 is used. This dataset contains 28 images in the Interferometric Wide Swath (IWS) mode, and the absolute orbit number is 40260. All these data use the C band (wavelength 5.6 cm), with a resolution of 5 m × 20 m and a single polarization (VV) observation mode. The detailed information on the image data parameters is shown in Table 2.

[0154] The purpose of using the precise orbit ephemerides (POD) of the Sentinel-1A satellite is to correct orbit errors. In addition, in order to remove the terrain effect, the SRTM 30m DEM provided by the National Aeronautics and Space Administration (NASA) was used as external reference data.

[0155] Table 2 Basic parameter information of Sentinel-1A image data

[0156]

[0157] Deformation prediction results

[0158] Using the prediction method of the present invention, the random forest regression method is adopted to analyze the relationship between the attribute values of ground PS points and the deformation rate, so as to identify the key features that have a significant impact on deformation. A random forest model is constructed from the scikit-learn library in Python, and the feature importance ranking is obtained through rf.feature_importances. Among them, the parameter settings of the random forest model are shown in Table 3 (n_estimators represents the number of decision trees, min_samples_leaf represents the minimum number of leaf nodes, and the ratio of the training set to the test set is 7:3). The number of decision trees is 100, the minimum number of child leaves is 5, and the ratio of the training set to the test set is 7:3.

[0159] Table 3 Parameter settings of the random forest model

[0160]

[0161] After the random forest calculates the ranking of the relationship between the attribute features of PS points and the deformation rate, the importance results of PS point metadata features are obtained (as shown in Figure 4 shown, Figure 4 the importance values and rankings of PS point attributes. Among them, the red columns represent the selected important features, including longitude, height above the reference ellipsoid, and latitude, and the light blue columns represent other attribute data.). It can be seen from Figure 4 that the contribution of the feature importance of the deformation rate of PS points along the Taiyuan-Jiaozuo Railway (Jincheng section) is in turn: longitude, height above the reference ellipsoid, and latitude. This result strongly proves that geospatial information has a significant impact on surface subsidence. Therefore, the present invention will use these three key features for clustering analysis and combine the elbow method to determine the appropriate number of clusters, so as to better fit the change trend of deformation signals in different regions.

[0162] To determine the number of PS point clusters, the elbow method is used to select the best number, and the result is as shown in Figure 5 ( Figure 5 Cluster elbow graph. Among them, when K = 3, the number of clusters is the best) shown. InFigure 5 Among them, the horizontal axis represents the number of different clusters, and the vertical axis represents the sum of squared errors within the clusters (SSE) after clustering. When the number of clusters is 3, it corresponds to the position where the downward trend of the SSE value begins to slow down significantly, representing the balance between the model complexity and the fitting effect. After the value of K is greater than 3, as the value of K increases, although the SSE value will gradually decrease because more clusters can better fit the data, it also increases the model complexity and is prone to overfitting. Therefore, the "elbow" position (K = 3) of the curve in the elbow method principle is considered to be the optimal number of clusters. All PS points in the study area are clustered into 3 clusters, and all possible combinations of clustered PS points are continuously divided. The three important features of the PS points (longitude, height above the reference ellipsoid, and latitude) are used as indicators to evaluate the error within the clusters until the error within each clustering cluster is minimized, and then this combination of PS point clustering is the optimal clustering result ( Figure 6 The clustering result based on features. Cluster 1 is the PS points in the east, Cluster 2 is the PS points in the south, and Cluster 3 is the PS points in the north.). Figure 6 It shows that the PS points are clustered into 3 clusters, which are distributed in the south, east, and north of the figure respectively, and their proportions are 17.95%, 56.04%, and 26.01% respectively. This indicates that the feature similarity within each cluster is relatively high, while the difference between clusters is relatively large.

[0163] In order to further divide the similar PS points to achieve a better prediction result. The deformation trends of all PS points in the time domain are respectively fitted as sine type, exponential type, and linear type. For a single PS point, the best fitting effect has the smallest sum of squared residuals (RSS), and thus the corresponding best fitting trend type of each PS point is determined. The fitting types of the deformation signal change trends of PS points are as Figure 7 shown. Among them, the purple curve represents the sine type (SIN), the blue curve represents the linear rising type (LINE_up), the yellow curve represents the linear stable type (LINE_no), and the pink curve represents the linear falling type (LINE_down). As Figure 7 can be seen, in this study area, no trend points consistent with the exponential type of deformation are found. Therefore, the deformation trends of PS points are divided into four types: sine type, linear rising type, linear stable type, and linear falling type.

[0164] Figure 8 shows the spatial distribution map of the classification results of the deformation trends. As Figure 8From the fitting results based on the deformation trend, it can be seen that the sine-type trend points (SIN) do not have an obvious spatial distribution pattern, showing a relatively scattered characteristic, with a certain degree of randomness, accounting for 9.37%. The linearly decreasing trend points (LINE_down) are mainly clustered, concentrated in the areas with obvious subsidence, accounting for 8.59%. The linearly stable trend points (LINE_no) are distributed more, mainly around the linearly decreasing trend points, accounting for 35.01%. And the linearly increasing trend points (LINE_up) show a general upward trend in deformation during this period, accounting for 47.03%.

[0165] Based on the comprehensive PS point clustering and fitting results, the study area is divided into 12 sub-regions. Inside each sub-region, the cumulative deformation amounts of PS points with different deformation trend types are predicted respectively. Among the 28-phase image results, 70% of the data is selected as the training set, and the remaining data is used as the test set. Taking the first 24-phase data from October 17, 2020 to September 6, 2021 as the true values, the surface deformation of the last 4 phases from September 18, 2021 to October 24, 2021 is predicted. Figure 9 A comparison chart of the deformation prediction results of the last four-phase PS points is shown. The first row is the true deformation results of the corresponding images, the second row is the results predicted by the method proposed in the present invention, and the third row is the results predicted by the traditional LSTM model. The results show that both methods can predict the deformation trend in a large range and are consistent with the real data. However, the traditional LSTM model has deficiencies in the prediction accuracy of local areas, such as Figure 9 the local uplift area shown by the red circle and the partial subsidence area shown by the black circle, and their prediction results are significantly deviated from the actual observed values. In sharp contrast, the method proposed in the present invention effectively solves this problem and significantly improves the accuracy of local area deformation prediction.

[0166] Figure 10 It is the prediction residual of PS points from September 18, 2021 to October 24, 2021 for the Taijiao Railway (Jincheng section). The first row represents the residual predicted by using the PSFC-LSTM model proposed in the present invention, and the second row represents the residual predicted by using the traditional LSTM model. Figure 10 It shows the residual distribution between the predicted values and the true values of the two methods. The first row is the residual between the prediction result of the method proposed in the present invention and the true value, and the second row is the residual between the prediction result of the traditional LSTM model and the true value. The results show that the prediction method proposed in the present invention is characterized by fewer noise points, more uniform color distribution and relatively smaller residuals in the figure, indicating that the LSTM prediction method for surface deformation based on PS point metadata feature clustering has better prediction performance than the traditional method.

[0167] To evaluate the prediction accuracy of the proposed method, in this embodiment, the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R 2 ) are used as evaluation indicators. Figure 11 Figure 4 shows the scatter plot comparison of the predicted values and true values of the prediction method provided by the present invention and the traditional method (scatter plots of predictions at different time periods). In the figure, the red line represents the y = x reference line, the green dot set represents the prediction results of the traditional method, and the yellow dot set represents the prediction results of the method of the present invention. The closer the scatter points are to the y = x line, the more accurate the prediction results are and the smaller the deviation from the true values. The results show that among the four-period prediction results, the prediction point set (green) of the traditional method is relatively scattered, and some points deviate far from the y = x line; while the prediction point set (yellow) of the method of the present invention is more concentrated, and most points are located near the y = x line. Linear fitting is performed on the two point sets respectively. The fitting line of the traditional method is the green line (Fit1), and the fitting line of the method of the present invention is the yellow line (Fit2). The closer the slope of the fitting line is to 1, the higher the reliability of the prediction results. Table 4 (where 1-4 are sub-regions in cluster result cluster1, 5-8 are sub-regions in cluster2, and 9-12 are sub-regions in cluster3. 1, 5, 9 are sine-type trend deformations, 2, 6, 10 are linear decline-type deformations, 3, 7, 11 are linear rise-type deformations, and 4, 8, 12 are linear stability deformations.) gives the statistical results of the error indicators predicted by the method of the present invention for each sub-region. The maximum RMSE of the four-period prediction results is 2.45, and the maximum MAE is 1.63. Table 5 compares the errors of the predicted cumulative settlement amounts of PS points in the entire study area after fusing 12 sub-regions between the method of the present invention and the traditional method. The results show that in the prediction of the four-period cumulative settlement amount, the RMSE of the method of the present invention is lower than 1.75, the MAE is lower than 1.03, and the R 2 The minimum value reaches 0.86; while the RMSE of the traditional method is lower than 2.02, the MAE is lower than 1.37, and the R 2 The minimum value reaches 0.81. Compared with the traditional method, the RMSE of the method of the present invention is reduced by an average of 13.8%, the MAE is reduced by an average of 26.0%, and the R 2 is increased by an average of 4.4%. These results indicate that the PSFC-LSTM method proposed by the present invention is significantly superior to the traditional LSTM model and has higher prediction accuracy.

[0168] Table 4 Statistical table of sub-region prediction errors.

[0169]

[0170] Table 5 Prediction error table for the entire study area

[0171]

[0172] In addition, in four typical deformation regions of the Taijiao Railway (Jincheng section) in this embodiment, four typical deformation points were randomly selected for cumulative deformation analysis. These four points are located in Sanjia Town, Gaoping City, Liushukou Town, Zezhou County, the southeast of the urban area, and Hexi Town, Gaoping City in sequence. Figure 12 The deformation and prediction curves of these four typical points are shown. Among them, the blue line represents the true cumulative deformation of the PS points, the red dotted line with dots represents the prediction results of the PSFC-LSTM model proposed by the present invention, and the green dotted line with dots represents the prediction results of the LSTM model. Figure 12 Typical point deformation prediction trend curve graph. Among them, Figure 12 (a) represents the cumulative settlement amount and prediction results of a random point in Sanjia Town, Figure 12 (b) represents the cumulative settlement amount and prediction results of a random point in Liushukou Town, Figure 12 (c) represents the cumulative settlement amount and prediction results of a random point in the southeast of the urban area, Figure 12 (d) represents the cumulative settlement amount and prediction results of a random point in Hexi Town.

[0173] Figure 12 As shown in (a), during the research period, the overall deformation of random point 1 in Sanjia Town showed a settlement trend. The settlement suddenly increased in July 2021 and then recovered. In the stage after September 2021, the prediction results were similar to the true results. The maximum cumulative deformation of this point was -13.6 mm. The average RMSE of the prediction results of the present invention was 1.25, and the MAE was 1.06. The RMSE of the prediction results of the traditional method was 1.33, and the MAE was 1.80. Figure 12 As shown in (b), the cumulative deformation of random point 2 in Liushukou Town was relatively large, showing an overall settlement trend. The prediction results after September 2021 were roughly consistent with the true deformation trend. The maximum cumulative settlement of this point was -17.3 mm. The average RMSE of the prediction results of the present invention was 0.90, and the MAE was 0.86. The RMSE of the prediction results of the traditional method was 1.79, and the MAE was 2.39. Figure 12 As shown in (c), random point 3 in the southeast of the urban area generally showed a settlement trend. The initial deformation of this point was relatively slow, suddenly decreased in July 2021, then recovered, and the prediction results after September were approximately the same as the true values. The maximum cumulative deformation of this point was -12.3 mm. The average RMSE of the prediction results of the present invention was 0.83, and the MAE was 0.65. The RMSE of the prediction results of the traditional method was 1.36, and the MAE was 2.05. Figure 12As shown in (d), the random point 4 in Hexi Town mainly shows a steady downward trend. There is a sudden drop and then a rebound in May 2021. After September 2021, the predicted values and the true values maintain a good trend consistency. The maximum cumulative deformation at this point reaches -20.4 mm. The average RMSE of the prediction results of the present invention is 1.51, and the MAE is 1.29. The RMSE of the prediction results of the traditional method is 1.30, and the MAE is 1.72. The RMSE and MAE errors of the four typical deformation points are shown in Table 6. Through the analysis of the prediction results of the typical points, it shows that the model proposed by the present invention also shows good prediction effects in local single points. Among them, the deformation prediction of the typical point in the southeastern part of the urban area located in the typical settlement area performs the best, while the prediction effect of the typical point inside Hexi Town is relatively poor.

[0174] Table 6 Prediction error table of typical deformation points

[0175]

[0176]

[0177] The present invention uses Sentinel-1 SAR images to conduct permanent scatterer interferometric measurement (PS-InSAR) on the surface settlement along the Taijiao Railway (Jincheng section) in China from October 17, 2020 to October 24, 2021. On this basis, a long short-term memory network model (PSFC-LSTM) based on the clustering of permanent scatterer (PS) point metadata features is proposed for predicting the cumulative settlement amount, and its prediction results are compared and analyzed with the prediction results of the traditional single long short-term memory network model (LSTM). The results show that in the four-phase prediction results, the maximum root mean square error (RMSE) of the sub-region prediction is 2.45, and the maximum mean absolute error (MAE) is 1.63. In the entire study area, compared with the traditional LSTM model, the RMSE of the PSFC-LSTM model proposed by the present invention is reduced by 13.8% on average, and the MAE is reduced by 26.0% on average, and the coefficient of determination (R 2 ) is increased by 4.4% on average.

[0178] The prediction results of this method in the study area show certain superiority over traditional methods and are consistent with the expected experimental results. The reason for this result may be that after feature selection, clustering, and classification of the PS point metadata, the data points within the sub-region have high feature similarity, thus reducing the interference of other types of data points. The machine learning model precisely uses the internal correlation of the data to learn the rules and then improves the prediction accuracy. Therefore, the PSFC-LSTM model proposed in this invention shows good prediction effects. However, the maximum RMSE and maximum MAE errors of the sub-region predictions mentioned above both appear in sub-region 5 and sub-region 9. The PS points within these two sub-regions are both fitted to a sine-type trend, and their prediction accuracy is relatively lower than that of other sub-regions. The reason for this phenomenon may be that the LSTM model has better adaptability to linear deformation trends than sine-type trends.

[0179] Example 2

[0180] To better verify the advantages and reliability of the prediction method of this invention, two comparative prediction methods are set in this example. One is to select important features from PS points using the random forest method for clustering and partitioning prediction, that is, the RFC-LSTM prediction method. The other is to predict according to different deformation trend types by fitting the deformation trend of PS points, that is, the TY-LSTM prediction method. The target area for prediction is the Taijiao Railway (Jincheng section), and the time period is from September 18, 2021, to October 24, 2021. The prediction results of the two comparative prediction methods are as Figure 13 ; Figure 13 shows the results of PS point deformation prediction based on the RFC-LSTM and TY-LSTM methods. The first row is the RFC-LSTM prediction method, and the second row is the TY-LSTM prediction method. Both methods effectively predict the true deformation of the ground surface and improve the traditional prediction results.

[0181] The distribution of prediction residuals corresponding to the RFC-LSTM and TY-LSTM methods is as Figure 14 shown. The first row is the distribution map of RFC-LSTM prediction residuals, and the second row is the distribution map of TY-LSTM prediction residuals. The numerical ranges of the prediction residuals of both methods are smaller than those of the traditional prediction method, and there are fewer noise points in the residual distribution, showing better effects compared with the traditional prediction method.

[0182] The scatter plots between the predicted values and the true values of the PSFC-LSTM prediction method provided by this invention, the two comparative prediction methods (RFC-LSTM, TY-LSTM), and the traditional prediction method are as Figure 15As shown, the scatter plots of PS point predictions during September 18, 2021, September 30, 2021, October 12, 2021, and October 24, 2021, for the Taijiao Railway (Jincheng section). Among them, the red dashed line represents the line y = x, indicating the optimal fitting effect. Green represents the predicted scatter points of the traditional method, yellow represents the predicted scatter points of the PSFC-LSTM method proposed in this invention, blue represents the predicted scatter points of the RFC-LSTM method, and black represents the predicted scatter points of the TY-LSTM method. From Figure 15 it can be seen that the scatter points of the three improved methods, PSFC-LSTM, RFC-LSTM, and TY-LSTM, have a smaller distribution range than the traditional method, and the slope of the traditional fitting line is smaller, which means that the gap from the true value is larger. The fitting line of the method proposed in this invention is closest to the y = x reference line, with a better effect. The prediction errors of the entire study area are evaluated using RMSE, MAE, and R 2 These three indicators. Table 7 shows the prediction error table. From the prediction results in the four time periods, it can be seen that the PSFC-LSTM prediction method proposed in this invention shows the best prediction effect, while the effects of the RFC-LSTM and TY-LSTM prediction methods are between the traditional prediction method and the prediction method proposed in this invention. Specifically, the RMSE of the RFC-LSTM method is reduced by 3.1%, the MAE is reduced by 2.4%, and R 2 is increased by 0.9%. The RMSE of the TY-LSTM method is reduced by 3.2%, the MAE is reduced by 2.8%, and R 2 is increased by 1.2%. The prediction effects of these two methods are slightly improved. However, the prediction effect of the method proposed in this invention after combining the two is better. This shows that the PSFC-LSTM method has a better prediction effect.

[0183] Table 7 Prediction error table of four prediction methods

[0184]

[0185]

[0186] The prediction curve change diagrams of four typical deformation points are as Figure 16As shown (where red represents the PSFC-LSTM prediction method proposed in the present invention, green represents the traditional prediction method, gray represents the RFC-LSTM prediction method, and orange represents the TY-LSTM prediction method), the corresponding prediction errors are shown in Table 8. For deformation point 1, the RMSE predicted by the RFC-LSTM method is 1.23 and the MAE is 1.11; the RMSE predicted by the TY-LSTM method is 1.31 and the MAE is 1.07. For deformation point 2, the RMSE predicted by the RFC-LSTM method is 1.03 and the MAE is 0.93; the RMSE predicted by the TY-LSTM method is 0.91 and the MAE is 0.80. For deformation point 3, the RMSE predicted by the RFC-LSTM method is 0.73 and the MAE is 0.58; the RMSE predicted by the TY-LSTM method is 0.78 and the MAE is 0.69. For deformation point 4, the RMSE predicted by the RFC-LSTM method is 1.73 and the MAE is 1.49; the RMSE predicted by the TY-LSTM method is 1.61 and the MAE is 1.47.

[0187] Table 8 Prediction Error Table for Typical Deformation Points

[0188]

[0189] In order to accurately predict the surface settlement along and around high-speed railways, fully consider the complexity and spatial correlation of deformation, and improve the high-precision surface deformation prediction ability, the present invention proposes a new surface deformation LSTM prediction method based on PS point metadata feature clustering. The main research results are as follows:

[0190] (1) The LOS-direction deformation rate in the study area from October 17, 2020 to October 24, 2021 was obtained using the PS-InSAR technology, and its variation range was from -20.913 mm / year to 28.087 mm / year.

[0191] (2) The prediction method proposed in the present invention showed significant superiority in the surface deformation prediction during the period from September 18 to October 24, 2021. The RMSE values of the prediction results of the prediction method of the present invention were all less than 1.75, and the MAE values were all less than 1.03, and the minimum value of R 2 reached 0.86. Compared with the traditional LSTM model, the RMSE of the PSFC-LSTM model was reduced by an average of 13.8%, the MAE was reduced by an average of 26.0%, and R 2 was increased by an average of 4.4%, fully confirming that the prediction method of the present invention has higher accuracy in surface deformation prediction compared with the prior art.

[0192] The achievements of the present invention provide a new method and practical case for the field of InSAR surface subsidence prediction, provide data support for the safe operation of linear projects such as high-speed railways and highways, and provide theoretical guidance for urban development planning.

[0193] The number of devices and the processing scale described here are used to simplify the description of the present invention. Applications, modifications, and variations of the present invention will be apparent to those skilled in the art.

[0194] Although the embodiments of the present invention have been disclosed above, it is not limited to the applications listed in the specification and embodiments. It can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily achieved. Therefore, without departing from the general concept defined by the claims and the equivalent scope, the present invention is not limited to specific details and the illustrated and described examples here.

Claims

1. The LSTM prediction method for high-speed rail surface deformation based on PS point metadata feature clustering is characterized by: The following steps are involved: Step 1: Obtain the synthetic aperture radar time series data of the target area in the SAR image, obtain the PS points of the target area through PS-InSAR technology processing, and further extract the metadata of the PS points, including the deformation time series data of the PS points and the time series deformation variables of the PS points; Step 2: Calculate the feature importance score of the PS point metadata extracted in step 1 based on the random forest model, and select the key features of the PS point metadata; Step 3: Perform cluster analysis on the key features selected in step 2 to complete the clustering of PS points; Step 4: Perform trend fitting on the deformation time series data of each PS point, and determine its deformation trend type based on the fitting residual; Step 5: Combining the clustering results in step 3 and the deformation trend type in step 4, the PS points that are spatially adjacent and have the same deformation trend are aggregated into homogeneous sub-regions; Step 6: Build a long short-term memory neural network model in each sub-region and input the time series shape variable of the PS point for prediction; Step 7: Evaluate the prediction accuracy through error indicators and output the deformation prediction results.

2. The high-speed railway surface deformation LSTM prediction method based on PS point metadata feature clustering as claimed in claim 1 is characterized in that: The PS-InSAR technology processing in step 1 includes the following steps: S11, SAR image registration and interferogram generation; S12, using SRTM 30m digital elevation model to remove terrain phase error; S13, separating the atmospheric delay phase by spatiotemporal filtering method; S14, screening PS points based on the coherence threshold, and retaining PS points with coherence values ​​greater than 0.85; S15. Generate time series surface deformation.

3. The high-speed railway surface deformation LSTM prediction method based on PS point metadata feature clustering as claimed in claim 1 is characterized in that: The parameters of the random forest model in step 2 are set as follows: the number of decision trees is 100, the minimum number of leaf nodes is 5, and the ratio of training set to test set is 7:

3.

4. The high-speed railway surface deformation LSTM prediction method based on PS point metadata feature clustering as claimed in claim 1, characterized in that: The method for determining the number of clusters in step 3 includes: S31, calculating the intra-cluster error sum of squares corresponding to different cluster numbers based on the elbow method; S32, select the inflection point where the SSE downward trend slows down significantly as the optimal number of clusters; S33. The cluster center distribution is optimized iteratively through the Euclidean distance in three-dimensional space, and the PS points are divided into multiple clusters.

5. The high-speed railway surface deformation LSTM prediction method based on PS point metadata feature clustering as claimed in claim 4 is characterized in that: The Euclidean distance calculation formula is: Among them, (x1, x2, x3) are the three eigenvalues ​​of the point target, and (c1, c2, c3) are the three eigenvalues ​​of the cluster center.

6. The high-speed railway surface deformation LSTM prediction method based on PS point metadata feature clustering as claimed in claim 1, characterized in that: The sub-region aggregation method in step 5 specifically includes: S51. Construct a spatial proximity network of PS points based on the Delaunay triangulation algorithm; S52, setting a spatial proximity distance threshold and a deformation trend consistency threshold, and merging adjacent PS points that meet the threshold into the same sub-region; S53. Use morphological dilation operation to fill isolated points in the sub-region to ensure spatial continuity.

7. The high-speed railway surface deformation LSTM prediction method based on PS point metadata feature clustering as claimed in claim 1, characterized in that: The method further includes step eight, dynamic early warning, which includes the following steps: S81. Establish a sliding update mechanism for model parameters, and retain the most recent 24 periods of data as a training set each time a new SAR image is acquired; S82. Set the deformation warning threshold to trigger a graded warning signal when the cumulative deformation exceeds ±15 mm.

8. The high-speed railway surface deformation LSTM prediction method based on PS point metadata feature clustering as claimed in claim 7, characterized in that: The dynamic warning of step eight also includes: S83, analyze the long-range dependence of deformation sequence based on Hurst index and modify the warning threshold; S84. When the cumulative deformation triggers an early warning, a settlement risk heat map and engineering reinforcement recommendations are generated in combination with the geographic information system (GIS) spatial overlay analysis; S85. Introduce a transfer learning mechanism to optimize local model parameters using historical deformation data of neighboring areas.

9. The LSTM prediction device for high-speed railway surface deformation based on PS point metadata feature clustering is characterized by: include: A processor and a memory; wherein the memory is used to store a computer program; and the processor is used to load and execute the computer program to implement the prediction method as described in any one of claims 1-8.

Images