A prediction method for urban land subsidence based on spatiotemporal fusion depth map model

Abstract

The present invention is applicable to the technical field of ground subsidence prediction, and provides a method for predicting urban ground subsidence by spatiotemporal fusion depth map model, including: step S1, data processing based on spatiotemporal feature analysis; step S2, ground subsidence prediction based on STF-GCLSTM model of grouping strategy. The present invention adopts multi-source data input, and extracts spatiotemporal features according to GCN spatial feature extraction model and LSTM temporal feature extraction model, and finally obtains more accurate settlement prediction value through spatiotemporal feature fusion prediction model, and further considers the difference of settlement characteristics between large and small areas, divides the study area into multiple sub-areas through grouping strategy, so that the settlement trend of each sub-area is similar, and then models different sub-areas respectively, so as to reduce the fitting error of the model and improve the prediction accuracy of the model.

Description

Technical Field

[0001] The present invention belongs to the technical field of land subsidence prediction, and in particular relates to a method for predicting urban land subsidence using a spatiotemporal fusion depth map model. Background Art

[0002] Land subsidence is a common engineering geological problem, especially in areas of rapid urbanization and industrialization. It poses a serious threat to infrastructure, buildings and environmental safety. In order to effectively monitor and predict land subsidence, researchers and engineers have developed a variety of technologies, including leveling, GPS, InSAR, etc. Leveling is a traditional method for monitoring land subsidence, which determines the amount of subsidence by measuring the height difference between fixed points. Although this method is simple and accurate, it requires a lot of manpower and time, so it is costly. GPS technology provides a continuous, all-weather monitoring method that can track land subsidence in real time. However, GPS technology has limitations in vertical accuracy and high equipment cost. InSAR (Synthetic Aperture Radar Interferometry) technology can provide large-scale, high-precision land subsidence monitoring. It measures small displacements of the ground surface by analyzing the phase changes of radar signals. InSAR technology is not limited by weather conditions and can cover large areas, but the interpretation and processing of its data are relatively complex.

[0003] In the field of ground subsidence prediction, researchers have explored a variety of methods, including methods based on physical models, methods based on mathematical statistics, and methods based on machine learning.

[0004] The method based on physical model mainly describes the internal mechanism of ground subsidence by considering the relationship between water and soil characteristics and ground subsidence, constructs a physical model to simulate the physical process of subsidence, calculates the deformation and predicts future ground subsidence. This method is often used in the prediction of subsidence caused by resource exploitation such as groundwater and underground construction such as subways. Common models include three-dimensional groundwater flow models. The advantage of the physical model method is that it can explain the causes of subsidence theoretically and has strong physical explanatory power. However, the physical model is highly dependent on actual geological parameters, and the construction of the model often requires a large amount of accurate geological and hydrological data, and the computational complexity of the model is high. Therefore, in some scenarios where data is incomplete or the study area is complex and changeable, the application of physical models may be limited.

[0005] The method based on mathematical statistics mainly describes the time series of settlement based on mathematical statistics, uses statistical methods to establish the regression relationship between ground settlement and time or settlement influencing factors, and predicts future ground settlement by curve fitting. This type of method is often used for settlement prediction in areas where settlement is relatively simple, where ground settlement has been observed for a long time, and where relevant detailed information is available. Common models include Poisson cycle model and gray model. Mathematical statistics methods are usually applicable to areas where the ground settlement process is relatively simple and where long-term observation data is available. In these areas, the future settlement can be predicted more accurately through the regular changes in historical data. The advantage of this method is that it is simple to operate and has relatively low computational cost. However, mathematical statistics methods rely on the integrity and regularity of historical data, and for areas affected by multiple complex factors, their prediction accuracy may be greatly limited, especially when sufficient observation data cannot be obtained, and the prediction results may deviate from reality.

[0006] Based on machine learning methods, traditional machine learning relies on the self-learning ability of the model to optimize model parameters for modeling and prediction, but it requires artificially designed feature engineering. Deep learning retains the self-learning ability of traditional machine learning and characterizes the complex nonlinear relationships in settlement data through the powerful nonlinear fitting ability of neural networks. In the field of settlement prediction, InSAR+deep learning model has become a mainstream method. For example, Liu Qinghao et al. used InSAR and LSTM (Long Short-Term Memory Network) to predict ground settlement in Cangzhou area, and achieved better performance than traditional methods. Ali Radman et al. combined multi-source data including InSAR and multi-deep learning fusion models to model and predict settlement in the vicinity of Lake Urmia in Iran. InSAR+deep learning model has complex diversity in the large-scale settlement trend, which will affect the prediction accuracy of the model. It only regards settlement as time series data, but does not consider its spatial characteristics as spatiotemporal data.

[0007] In addition, the existing deep learning model does not consider the spatial heterogeneity of ground subsidence, and only analyzes subsidence from the perspective of geographical regions. At the same time, ground subsidence has many related factors and complex mechanisms. The prediction model is only built from the subsidence time series data, ignoring the role of human activities. Therefore, in the existing large-area ground subsidence prediction, due to the diversity and complexity of subsidence trends, a single model is difficult to accurately model and predict; at the same time, the spatiotemporal characteristics of subsidence data are often ignored, resulting in insufficient prediction accuracy. Summary of the invention

[0008] In view of the above problems, the purpose of the present invention is to provide a method for predicting urban ground subsidence by spatiotemporal fusion depth map model, aiming to solve the technical problems that the existing model cannot accurately model and the prediction accuracy is not high.

[0009] The present invention adopts the following technical solution:

[0010] The urban land subsidence time-space fusion depth map model prediction method comprises the following steps:

[0011] Step S1: Data processing based on spatiotemporal feature analysis

[0012] Obtain historical monitoring time series data and data on factors affecting settlement, perform spatiotemporal interpolation on the historical monitoring time series data, group the time series data according to the settlement trend, and obtain settlement grouping time series data; extract and screen the data on factors affecting settlement, and construct a local spatial correlation diagram between monitoring points;

[0013] Step S2: Ground subsidence prediction based on the STF-GCLSTM model of grouping strategy

[0014] The screened settlement influencing factor data and the local spatial correlation diagram of the monitoring points are input into the GCN spatial feature extraction model to obtain spatial features; the settlement grouping time series data is input into the LSTM time feature extraction model to obtain the time series dependency features; the spatial features and time series dependency features are fused through the spatiotemporal feature fusion prediction model to output the final prediction value.

[0015] Furthermore, the specific process of step S1 is as follows:

[0016] S11. Obtain historical monitoring time series data and perform spatiotemporal interpolation processing

[0017] The monitoring points obtained by interpreting the administrative boundaries of the study area are cropped, the historical monitoring time series data of the monitoring points in the study area are retained, the distorted and invalid data are eliminated, and the data are interpolated by the spline interpolation method to obtain complete time series data, and then outliers are detected and eliminated. For the eliminated time series data, the spatial neighboring weight matrix is ​​calculated, and spatial weighted interpolation is performed to fill the eliminated time series data. Finally, the time series data are grouped according to the settlement trend characteristics of the time series data to obtain settlement grouped time series data, that is, the study area is divided into multiple sub-areas according to similar settlement trend characteristics;

[0018] S12. Obtain data on factors affecting settlement and perform extraction and factor screening

[0019] The correlation between the influencing factor data was determined by the multicollinearity test method, and the influencing factor data corresponding to the factors with multicollinearity with other factors were eliminated; the correlation between land subsidence and influencing factors was calculated by the Pearson correlation coefficient, and the influencing factor data with little correlation were eliminated;

[0020] S13. Constructing a local spatial correlation diagram between monitoring points

[0021] The K-dimensional tree is used to realize the neighbor search between monitoring points, find the adjacent points of each monitoring point, and the lines between monitoring points and adjacent points constitute edges to form a local spatial association relationship graph.

[0022] Furthermore, in step S13, the distances of the edges in the local spatial association relationship graph are spatially screened, the edges within the target distance range are retained, and then the Pearson correlation coefficient between the detection point and the adjacent points is further calculated, and the edges with correlation coefficients greater than a preset value are retained to obtain the final local spatial association relationship graph.

[0023] Furthermore, the specific process of step S2 is as follows:

[0024] S21. Spatial feature extraction of GCN spatial feature extraction model

[0025] Taking the influencing factor data as the attribute data of the monitoring point, the local spatial association relationship graph data of the monitoring point is input, and the GCN spatial feature extraction model is used to extract the spatial relationship between the monitoring points, that is, the spatial feature, which contains the spatial association information and the influencing factor data information;

[0026] S22, Time-dependent feature extraction of LSTM time feature extraction model

[0027] The time series data obtained by sliding window division is input into the LSTM time feature extraction model to output the time series dependency features;

[0028] S23. Subsidence prediction based on spatiotemporal feature fusion prediction model

[0029] The spatiotemporal features composed of spatial features and time-dependent features are fused by direct feature splicing, and the relationship between spatiotemporal features and predicted values ​​is modeled using a fully connected layer to obtain a spatiotemporal feature fusion prediction model. The model output is the settlement prediction value.

[0030] The beneficial effects of the present invention are:

[0031] Aiming at the specific problems in regional ground subsidence prediction, the present invention obtains time series data, subsidence influencing factor data and local spatial correlation diagram through data processing of spatiotemporal feature analysis. By adopting multi-source data input, spatiotemporal feature extraction according to the GCN spatial feature extraction model and the LSTM temporal feature extraction model, and finally obtaining a more accurate subsidence prediction value through the spatiotemporal feature fusion prediction model, the method of the present invention can automatically mine the effective spatiotemporal features in the subsidence time series data, and introduce the spatiotemporal features into the spatiotemporal feature fusion prediction model for high-precision subsidence prediction.

[0032] Furthermore, taking into account the differences in settlement characteristics between large and small areas, the study area was divided into multiple sub-areas through a grouping strategy, so that the settlement trend of each sub-area was similar. Different sub-areas were then modeled separately to reduce the fitting error of the model and improve the prediction accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a flow chart of the prediction method of urban land subsidence spatiotemporal fusion depth map model;

[0034] Figure 2 It is a flow chart of data processing;

[0035] Figure 3 This is the structural diagram of the STF-GCLSTM model. DETAILED DESCRIPTION

[0036] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0037] In order to illustrate the technical solution of the present invention, a specific embodiment is provided below for illustration.

[0038] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0039] For the traditional model of urban ground subsidence, its generalization ability is limited, and it is difficult to adapt to the subsidence trends in different regions, resulting in a decrease in prediction accuracy when applied in large areas; and ground subsidence is affected by a variety of human activities, such as road traffic, underground engineering construction, etc., and the existing methods fail to fully consider these factors; at the same time, the existing methods often ignore the spatial heterogeneity and time series characteristics of subsidence data, and fail to make full use of this information to improve the prediction accuracy. For this, the present invention provides a prediction method for a spatiotemporal fusion deep map model of urban ground subsidence. The method is based on an adaptive grouping strategy, and a large area is divided into multiple sub-areas, each of which has a similar subsidence trend, thereby improving the generalization ability and prediction accuracy of the model; and integrating multi-source data, such as geology, hydrology, human activities, etc., to more comprehensively understand the causes and trends of ground subsidence. At the same time, deep learning technology is used to fully explore the spatiotemporal characteristics of subsidence data and improve the performance of the prediction model. In order to illustrate the technical solution described in the present invention, it is described below through specific embodiments.

[0040] like Figure 1As shown, the urban land subsidence spatiotemporal fusion depth map model prediction method provided in this embodiment includes the following steps:

[0041] Step S1, data processing based on spatiotemporal feature analysis: obtain historical monitoring time series data and settlement influencing factor data, perform spatiotemporal interpolation processing on the historical monitoring time series data, group the time series data according to the settlement trend, and obtain settlement grouping time series data; extract and screen the settlement influencing factor data, and construct a local spatial correlation relationship diagram between monitoring points.

[0042] The present invention uses historical monitoring time series data of settlement and multi-source settlement influencing factor data. The time series data of ground subsidence is a typical geographic spatiotemporal data with timestamp attributes and spatial location attributes. However, in the field of ground subsidence prediction, the existing technologies regard it as a prediction problem in the time dimension, while ignoring its characteristics in the spatial dimension, so spatial feature analysis is required. In addition to processing the historical monitoring time series data, this step also extracts and screens the data of settlement influencing factors, combines the local spatial correlation relationship diagram of the monitoring points to obtain spatial characteristics, and prepares data for the subsequent time series data input and graph data input of the urban ground subsidence prediction model. Specifically, the implementation process of this step is as follows:

[0043] S11. Obtain historical monitoring time series data and perform spatiotemporal interpolation processing: crop the monitoring points obtained by interpreting the administrative boundaries of the study area, retain the historical monitoring time series data of the monitoring points in the study area, eliminate distorted and invalid data, perform data interpolation through the spline interpolation method, obtain complete time series data, and then perform outlier detection and elimination. For the eliminated time series data, perform spatial weighted interpolation to fill the eliminated time series data by calculating the spatial proximity weight matrix. Finally, group the time series data according to the settlement trend characteristics of the time series data to obtain settlement grouped time series data, that is, divide the study area into multiple sub-areas according to similar settlement trend characteristics.

[0044] This step first obtains historical monitoring time series data. This embodiment uses satellite InSAR data to obtain ground subsidence information through time series analysis. In addition to InSAR data, multiple environmental factor data are also integrated: land cover type data comes from Copernicus Global Land Services, Normalized Difference Vegetation Index (NDVI) data comes from the Digital Earth Open Platform, and Digital Elevation Model (DEM) data is obtained through Alaska satellite facilities. In addition, based on DEM data, ARCGIS (a geographic information mapping software) is used to extract terrain features such as slope, aspect, runoff accumulation, terrain position index (TPI), and terrain wetness index (TWI), providing multi-dimensional environmental factor support for the settlement prediction model.

[0045] Combination Figure 2As shown in the figure, in order to improve the reliability of the historical monitoring time series data of settlement and ensure the high accuracy of the data, the monitoring points obtained by interpreting the administrative boundaries of the study area are cropped, and only the historical monitoring time series data of the monitoring points located in the study area are retained. Then the distorted and invalid data of the monitoring points are eliminated, and the missing data are interpolated using the spline interpolation method, such as cubic spline interpolation, to obtain complete time series data; some outliers that may be caused by interpretation errors are tested for normal distribution, and data exceeding the specified error interval are identified as outliers and eliminated. For the problem of missing data after eliminating outliers, based on the data characteristics of the obvious spatial correlation of settlement time series data, the spatial neighbor weight matrix is ​​calculated according to the geographical location of the monitoring points to represent the spatial relationship between the elements in the data set, and then the monitoring points are spatially weighted interpolated to fill the missing data.

[0046] The spatial neighbor weighting calculation method is as follows:

[0047]

[0048] in represents the spatial weight relationship between the ith monitoring point and its jth neighboring point, i.e., the neighboring weight. n represents the number of all neighboring elements of the ith element, d ij Represents the spatial distance between the ith element and the jth adjacent element. Then, based on the spatial neighbor weight matrix, the removed outliers are spatially weighted. The specific calculation formula is:

[0049]

[0050] Among them, LS i It represents the settlement value of the ith monitoring point in the current period, and n represents the number of neighboring points of the ith monitoring point.

[0051] S12. Obtain data on factors affecting settlement and perform extraction and factor screening: determine the correlation between the factors using the multicollinearity test method, and eliminate the factors that have multicollinearity with other factors; calculate the correlation between ground settlement and the factors using the Pearson correlation coefficient, and eliminate the factors that have little correlation.

[0052] In order to verify whether there is a dependency relationship between the selected influencing factor data and whether there is a certain correlation between the selected influencing factor data and ground subsidence, this step uses the multicollinearity test and Pearson correlation coefficient correlation test methods to quantify their relationship characteristics, so as to conduct a more scientific and reasonable screening of influencing factors.

[0053] The subsidence trend depends on the distribution of influencing factors such as the local environment. In order to extract the spatial characteristics of subsidence distribution, this embodiment selects six types of candidate influencing factors, including ground elevation, slope aspect, land cover type, annual average rainfall and building height, and then obtains the final influencing factor data through value extraction and factor screening to introduce into the model.

[0054] The multicollinearity test method determines the correlation between the influencing factor data. The multicollinearity test includes two indicators: tolerance (TOL) and variance inflation factor (VIF). The specific calculation methods of VIF and TOL are as follows:

[0055]

[0056] in, It represents the coefficient of determination of the regression of other factors with the i-th factor among all factors as the dependent variable. TOL is the reciprocal of VIF, which represents the degree of correlation between factors. Usually, if the VIF value is greater than 10 or the TOL value is less than 0.1, it is considered that there is multicollinearity between the factor and other factors, and it should be removed from the influencing factor data.

[0057] For correlation test, Pearson Correlation Coefficient (PCCs) was selected to calculate the correlation between land subsidence and influencing factors.

[0058] In addition, it is particularly important to note that, considering the differences in settlement characteristics between large and small areas, this step uses a grouping strategy to divide the study area into multiple sub-areas, so that the settlement trend of each sub-area is similar, that is, the settlement trend characteristics group the time series data to obtain settlement grouping time series data. Then, different sub-areas are modeled separately to reduce the fitting error of the model and improve the prediction accuracy of the model.

[0059] S13. Construct a local spatial association relationship graph between monitoring points: Use a K-dimensional tree to implement neighbor search between monitoring points, find the adjacent points of each monitoring point, and connect the lines between the monitoring points and the adjacent points to form an edge, forming a local spatial association relationship graph.

[0060] The relationship graph consists of points and edges. Points are monitoring points, and edges are the edges between monitoring points. In order to effectively mine the spatial correlation between monitoring points, it is necessary to first construct the spatial relationship between all monitoring points. The relationship between monitoring points is displayed in the graph using edges, that is, the edges between monitoring points. The structure composed of monitoring points and edges is called a graph. If there is an edge between monitoring points, the two monitoring points are called each other's adjacent points.

[0061] This step uses the global Moran's index to perform spatial autocorrelation analysis on the sedimentation time series data, which can quantify the similarity of units in spatially adjacent or nearby locations, and then measure whether global spatial autocorrelation exists. Then, the connection between the edges of the monitoring points is established based on the spatial distance as a constraint to ensure the reliability and availability of the data.

[0062] Specifically, the construction of the spatial association relationship graph can be regarded as a neighbor search problem for the N nearest neighbor points of each point. The K-dimensional tree (KD-Tree) can realize the neighbor search of the monitoring point. Compared with the conventional K-Nearest Neighbor (KNN) algorithm, its time complexity is much lower. After finding the neighboring points of each monitoring point, due to the limited number of searches, some edges may be incorrectly connected. Therefore, this step screens the distance of the middle edge of the local spatial association relationship graph by spatial distance (such as 200 meters) to retain the edges within the target distance range. The Pearson correlation coefficient between the detection point and the neighboring points is further calculated, and the correlation coefficient between each monitoring point and its nearest neighbor is calculated, and the edges with a correlation coefficient greater than the preset value are retained. Generally, when the correlation coefficient is greater than 0.3, it is considered to be correlated, when the correlation coefficient is 0.3-0.8, it is weakly correlated, and when the correlation coefficient is 0.8-1, it is strongly correlated.

[0063] Step S2, ground subsidence prediction based on the STF-GCLSTM model of the grouping strategy: input the screened subsidence influencing factor data and the local spatial correlation diagram of the monitoring points into the GCN spatial feature extraction model to obtain spatial features; input the subsidence grouping time series data into the LSTM time feature extraction model to obtain the time series dependency features; the spatial features and the time series dependency features are fused through the spatiotemporal feature fusion prediction model to output the final prediction value.

[0064] In the previous steps, the time series data is grouped. The grouping strategy mainly takes into account the diversity of settlement trends in large-scale ground subsidence, and uses grouping to distinguish settlement trends. In this step, the STF-GCLSTM model takes into account the complex spatiotemporal characteristics of settlement, and uses the advantages of GCN (graph convolutional neural network) and LSTM (long short-term memory network) models to fully explore spatiotemporal characteristics for prediction.

[0065] Specifically, the structure of the STF-GCLSTM model is as follows Figure 3As shown in the figure, it includes the GCN spatial feature extraction model, the LSTM temporal feature extraction model, and the spatiotemporal feature fusion prediction model. The input of the entire model is the constructed local spatial correlation diagram, various settlement influencing factor data, and settlement monitoring historical grouping time series data. The role of spatiotemporal features in settlement prediction is fully considered through the end-to-end automatic feature mining capability of the model, so as to improve the prediction accuracy of the model. The implementation process of this step is as follows:

[0066] S21. Spatial feature extraction of GCN spatial feature extraction model: take influencing factor data as the attribute data of monitoring points, input local spatial correlation relationship graph data of monitoring points, and use GCN spatial feature extraction model to extract the spatial relationship between monitoring points, that is, spatial features, which contain spatial correlation information and influencing factor data information.

[0067] In the GCN spatial feature extraction model, the influencing factor data is used as the attribute data of the monitoring point, and a layer of graph convolutional neural network GCN is used to extract the spatial relationship between the monitoring points.

[0068] For example, assume that the local spatial association graph data of the input monitoring points is the adjacency matrix A and the feature matrix X. The dimension of the adjacency matrix A is [the number of training data input at a time, the number of nodes in the graph, the number of nodes in the graph], and the dimension of the feature matrix X is [the number of training data input at a time, the number of nodes in the graph, the attribute dimension of each node], then the output feature dimension after GCN extraction is [the number of training data input at a time, the number of nodes in the graph, the extracted spatial feature dimension]. Since the input is the local association graph of each monitoring point, only the feature information of the monitoring point needs to be retained, so the actual output is [the number of training data input at a time, 1, the extracted spatial feature dimension]. At this time, the extracted features contain the fusion features of spatial association information and influencing factor data information.

[0069] S22, time series dependency feature extraction of LSTM time feature extraction model: Use sliding window division to obtain time series data and input it into the LSTM time feature extraction model to output time series dependency features.

[0070] The LSTM model is used to extract the time-series dependent features of the settlement time-series data of the monitoring point. The time-series data of complete settlement monitoring is obtained through the preprocessing method of spatiotemporal interpolation, and then the time-series data required for model input is obtained by sliding window division. Assuming that the data dimension of the time-series data is [the number of training data input at a time, the length of the sequence input at a time, the number of input time series], the output feature after extraction by the LSTM time feature extraction model is [the number of training data input at a time, the length of the sequence input at a time, the dimension of the time-series dependent features extracted by LSTM].

[0071] S23. Settlement prediction based on spatiotemporal feature fusion prediction model: The spatiotemporal features composed of spatial features and time-dependent features are fused by direct feature splicing, and the fully connected layer is used to model the relationship between spatiotemporal features and predicted values ​​to obtain the spatiotemporal feature fusion prediction model. The model output is the settlement prediction value.

[0072] In order to construct a complete end-to-end learning process in the spatiotemporal feature fusion prediction model, it is also necessary to establish the feature and prediction value after the spatiotemporal feature extraction in steps S21 and S22. The spatial feature fusion prediction model establishes a mapping relationship between the extracted spatiotemporal features and the final predicted values. Specifically, the spatiotemporal features are first fused. There are many ways to fuse features. In order to reduce feature loss during feature fusion, the fusion strategy adopted in this embodiment is direct feature splicing, and then a simple fully connected layer is used to achieve relationship modeling between spatiotemporal features and predicted values. In addition, the complex interaction relationship between settlement and various settlement influencing factors can be further potentially explored in a data-driven manner. Assume that the dimension of the extracted spatial feature is [number of training data input at a time, spatial feature, 1], the dimension of the extracted temporal dependency feature is [number of training data input at a time, time feature, 1], the dimension of the fused feature after feature fusion is [number of training data input at a time, spatial feature + time feature, 1], and the dimension of the output prediction value after the fully connected layer is [number of training data input at a time, 1].

[0073] In summary, by dividing a large area into multiple sub-areas, the present invention can perform personalized modeling for the subsidence trends in different areas, thereby improving the accuracy of predictions; and by integrating multi-source data such as geology, hydrology, and human activities, the present invention can more comprehensively understand the causes and trends of ground subsidence, thereby improving the accuracy and reliability of predictions. The present invention uses deep learning technology to fully explore the spatiotemporal characteristics of subsidence data, which helps capture complex patterns in the subsidence process and improve the performance of the prediction model.

[0074] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A prediction method for urban land subsidence using a spatiotemporal fusion depth map model, characterized in that: The prediction method comprises the following steps: Step S1: Data processing based on spatiotemporal feature analysis Obtain historical monitoring time series data and data on factors affecting settlement, perform spatiotemporal interpolation on the historical monitoring time series data, group the time series data according to the settlement trend, and obtain settlement grouping time series data; extract and screen the data on factors affecting settlement, and construct a local spatial correlation diagram between monitoring points; Step S2: Ground subsidence prediction based on the STF-GCLSTM model of grouping strategy The screened settlement influencing factor data and the local spatial correlation diagram of the monitoring points are input into the GCN spatial feature extraction model to obtain spatial features; Input the sedimentation group time series data into the LSTM time feature extraction model to obtain the time series dependency features; fuse the spatial features and time series dependency features through the spatiotemporal feature fusion prediction model to output the final prediction value; The specific process of step S1 is as follows: S11. Obtain historical monitoring time series data and perform spatiotemporal interpolation processing The monitoring points obtained by interpreting the administrative boundaries of the study area are cropped, the historical monitoring time series data of the monitoring points in the study area are retained, the distorted and invalid data are eliminated, and the data are interpolated by the spline interpolation method to obtain complete time series data, and then outliers are detected and eliminated. For the eliminated time series data, the spatial neighboring weight matrix is ​​calculated, and spatial weighted interpolation is performed to fill the eliminated time series data. Finally, the time series data are grouped according to the settlement trend characteristics of the time series data to obtain settlement grouped time series data, that is, the study area is divided into multiple sub-areas according to similar settlement trend characteristics; S12. Obtain data on factors affecting settlement and perform extraction and factor screening The correlation between the influencing factor data was determined by the multicollinearity test method, and the influencing factor data corresponding to the factors with multicollinearity with other factors were eliminated; the correlation between land subsidence and influencing factors was calculated by the Pearson correlation coefficient, and the influencing factor data with little correlation were eliminated; S13. Constructing a local spatial correlation diagram between monitoring points The K-dimensional tree is used to realize the neighbor search between monitoring points, find the adjacent points of each monitoring point, and the lines between monitoring points and adjacent points constitute edges to form a local spatial association relationship graph.

2. The urban land subsidence prediction method using a spatiotemporal fusion depth map model as claimed in claim 1, characterized in that: In step S13, the distances of the edges in the local spatial association relationship graph are spatially screened, the edges within the target distance range are retained, and then the Pearson correlation coefficient between the detection point and the adjacent points is further calculated, and the edges with correlation coefficients greater than the preset value are retained to obtain the final local spatial association relationship graph.

3. The urban land subsidence prediction method using a spatiotemporal fusion depth map model as claimed in claim 2, characterized in that: The specific process of step S2 is as follows: S21. Spatial feature extraction of GCN spatial feature extraction model Taking the influencing factor data as the attribute data of the monitoring point, the local spatial association relationship graph data of the monitoring point is input, and the GCN spatial feature extraction model is used to extract the spatial relationship between the monitoring points, that is, the spatial feature, which contains the spatial association information and the influencing factor data information; S22, Time-dependent feature extraction of LSTM time feature extraction model The time series data obtained by sliding window division is input into the LSTM time feature extraction model to output the time series dependency features; S23. Subsidence prediction based on spatiotemporal feature fusion prediction model The spatiotemporal features composed of spatial features and time-dependent features are fused according to the direct feature splicing method, and the fully connected layer is used to model the relationship between spatiotemporal features and predicted values ​​to obtain the spatiotemporal feature fusion prediction model. The model output is the settlement prediction value.

Images