A method for coupling extreme random tree-information content of land subsidence risk assessment

Abstract

This invention discloses a land subsidence risk assessment method coupled with extreme random trees and information content, belonging to the field of geological disaster prevention and control technology. The novel land subsidence risk assessment model coupled with extreme random trees and information content described in this invention is based on synthetic aperture radar interferometry image data, employing time-series InSAR technology to acquire long-term regional surface deformation information; it overlays multi-source geoscientific information such as basic geology, hydrogeology, and regional dynamic / static loads representing human activities to construct a land subsidence machine learning model based on extreme random trees; on this basis, the weight factors output by the model are substituted into the information content model to conduct regional land subsidence risk assessment and dynamically identify the spatial distribution of regional land subsidence risk. This invention can effectively improve the accuracy and reliability of regional land subsidence risk assessment, enhance the level of land subsidence prevention and disaster early warning, and provide certain technical support for regional subsidence prevention and control.

Description

Technical Field

[0001] This invention relates to the field of geological disaster prevention technology, specifically to a ground subsidence risk assessment method coupled with extreme random tree-information. Background Technology

[0002] Land subsidence is a geological phenomenon caused by the combined effects of natural and anthropogenic factors, leading to the consolidation and compression of loose strata and a decrease in surface elevation. Currently, this problem occurs in over 200 countries and regions worldwide. In my country, it is mainly concentrated in the Beijing-Tianjin-Hebei region, the Fenwei Basin, the Pearl River Delta, and the Yangtze River Delta, with significant variations. The Beijing Plain area, due to factors such as groundwater over-extraction, urban construction, and geology, has become one of the regions with the most severe subsidence globally. Land subsidence affects the earthquake resistance of buildings, exacerbates ground fissures, threatens rail transit, and restricts urban construction and economic development. Therefore, conducting regional subsidence risk assessments is crucial for precise prevention and control.

[0003] Currently, there are various evaluation models for land subsidence geological hazards, mainly including the analytic hierarchy process (AHP), information content model, entropy method, neural network model, and comprehensive index method. Among them, expert scoring method, AHP and its improved methods are limited by data conditions, labor costs and the degree of refinement. The evaluation results rely too much on the subjective experience of experts, lack objectivity, and are difficult to meet the needs of accurate evaluation. Machine learning models such as logistic regression, random forest, and artificial neural networks have the advantages of high prediction accuracy and strong objectivity, but the reliability of the evaluation is easily affected by the selection of evaluation indicators, data processing methods and data quality, resulting in insufficient stability. In addition, most of the above-mentioned models focus on static analysis of the study area and lack the ability to capture and assess the spatiotemporal dynamic changes of land subsidence. They cannot adapt to the characteristics of the superposition of multiple factors and dynamic evolution of land subsidence, and cannot comprehensively reflect the dynamic distribution characteristics of regional land subsidence risk.

[0004] To address the shortcomings of existing land subsidence risk assessment models, such as insufficient objectivity, high data-dependent reliability, lack of spatiotemporal dynamic analysis capabilities, and low physical interpretability, this invention proposes a land subsidence risk assessment method coupled with extreme random tree-information. This invention innovatively applies the ERT-IV algorithm to land subsidence risk assessment, enabling the model to possess the advantages of long-term spatiotemporal dynamic analysis while effectively compensating for the shortcomings of traditional models in terms of low physical interpretability, thus providing more scientific technical support for precise prevention and control of land subsidence. Summary of the Invention

[0005] The purpose of this invention is to dynamically identify the spatiotemporal evolution information of regional land subsidence disaster risk, qualitatively reveal the evolutionary characteristics and changing trends of regional land subsidence disaster risk in time series and spatial dimensions, and quantify the differences in regional land subsidence disaster risk levels. A dynamic risk assessment model for land subsidence is proposed, providing a scientific basis for the precise prevention and efficient management of regional land subsidence disasters, and contributing to the rational planning and safe development of cities.

[0006] 1. The technical problem to be solved by the present invention

[0007] The purpose of this invention is to propose a ground subsidence risk assessment method coupled with extreme random tree-information to solve the following problems existing in the prior art: (1) Existing ground subsidence risk assessment models are not objective enough and their reliability is easily affected by interference. The assessment results deviate significantly from the actual geological laws. Traditional land subsidence risk assessment methods, such as expert scoring and analytic hierarchy process (AHP), rely excessively on expert subjective experience and lack objective quantitative data-driven evidence, making them unsuitable for precise assessment. While machine learning assessment models, such as logistic regression and ordinary random forest, possess a certain degree of objectivity, their reliability is easily affected by the selection of assessment indicators, data processing methods, and data quality, resulting in insufficient model stability. Furthermore, these machine learning models suffer from low physical interpretability, failing to clearly quantify the actual contribution of each influencing factor to land subsidence, thus limiting the scientific validity and reference value of the assessment results.

[0008] (2) Existing models are based on static analysis and lack the ability to capture the spatiotemporal dynamic changes of land subsidence, thus failing to adapt to its dynamic evolution characteristics. Land subsidence is a geological phenomenon characterized by the superposition of multiple factors and dynamic evolution in time and space under the combined influence of natural and human factors. However, most existing evaluation models are based on static data of the study area. They cannot identify abrupt changes in the land subsidence rate through time series data, reveal the changing trend of subsidence risk over time, or incorporate the spatial differences in surface deformation into the evaluation system. They cannot accurately identify the spatial distribution characteristics of regional land subsidence risk, cannot fully reflect the dynamic evolution law of land subsidence risk, and cannot provide practical technical support for the precise prevention and control of regional subsidence.

[0009] 2. Technical Solution To achieve the above objectives, this scheme first employs spatial analysis and mathematical statistics techniques, combined with geological conditions of the study area, to determine the optimal grid scale for studying subsidence. Secondly, it optimizes and integrates new technologies such as Google Earth Engine (GEE) and machine learning, comprehensively considering multi-source spatial data including groundwater level, compressible layer thickness, and OSM roads in the study area, to construct a multi-factor driven ERT-IV land subsidence risk assessment model. This invention can provide a scientific basis for the prevention and control of land subsidence disasters and has significant practical application value for the rational planning of urban development and the prevention and control of urban land subsidence. The specific technical solution of this invention is as follows: A land subsidence risk assessment method coupled with extreme random tree-information, comprising the following steps; S1. Data Acquisition and Preprocessing: Acquire long-term synthetic aperture radar image data covering the study area. Use time-series synthetic aperture radar interferometry to obtain deformation information along the line of sight on the ground. Use measured GPS or leveling point data to verify the time-series InSAR monitoring results. Convert the line-of-sight values ​​to vertical values ​​to obtain high-precision vertical deformation data of permanent scattering points. Obtain the average annual subsidence rate of the study area through Kriging interpolation / inverse distance weighting interpolation methods. S2. Segmented analysis of settlement time series based on Mann-Kendall mutation test: The time series annual average settlement rate obtained in S1 is segmented using the Mann-Kendall mutation test method; S3. Constructing a coupled extreme random tree and information content model: Establish a ground subsidence machine learning model based on extreme random trees in the study area, embed the trained extreme random tree model into the information content model calculation framework, realize the coupling of extreme random trees and information content model, and construct a regional extreme random tree-information content coupled model. S4. Multi-source data processing and acquisition of weight factors for regional land subsidence: This includes comprehensive basic geological, hydrogeological, and regional dynamic / static load data representing human activities. Using spatial analysis and mathematical statistics techniques, the contribution of each weight factor in the regional extreme random tree land subsidence machine learning model is obtained. S5. Construct a novel ground subsidence risk assessment model coupled with extreme random tree-information: Input the annual average subsidence rate calculated in S1 and the weight factor contribution obtained in S4 into the extreme random tree-information coupled model constructed in S3 to establish a novel ground subsidence risk assessment model coupled with extreme random tree-information in the region. S6. Model Accuracy Validation and Result Output: Using the Receiver Operating Characteristic (ROC) curve and area under the curve (AUC) as accuracy evaluation indicators, we will conduct an accuracy analysis of the novel ground subsidence risk assessment model with coupled extreme random tree-information obtained in S5, verify the feasibility of the model, and output the validation results.

[0010] Preferably, the Mann-Kendall mutation detection method mentioned in S2 specifically includes the following steps: S2.1. Represent a time series of length n as follows: (1) In the formula, n is the sequence length, corresponding to the number of sample years; For the first Annual observations; initialize time series and complete data preprocessing.

[0011] S2.2 Constructing the ordered series and calculating the statistics Represented as: (2) In the formula, It is a symbolic function; For the first Annual observations; use equation (2) to calculate the sequence of events at each time point, and solve for its mean and variance; S2.3, Positive Standardized Statistic Represented as: (3) In the formula, The mean of the ordered sequence. The variance is calculated using equation (3) for the forward statistical sequence. .

[0012] S2.4, Reverse Statistical Sequence Represented as: (4) In the formula, The statistics are calculated for the reverse sequence; the reverse statistical sequence is obtained using equation (4). .

[0013] S2.5, The mutation determination criteria are expressed as follows: (5) In the formula, 1.96 is the significance level. The corresponding critical value; and The two curves are If the critical lines intersect inside each other and satisfy equation (5), then the first... The year was a year of significant change.

[0014] S2.6. Recalculate the statistics using formula (2) in S2.2 and update the mutation judgment results in real time; S2.7. Determine whether the termination condition is met based on the result obtained in S2.6. If it is met, exit and output the year of the mutation; otherwise, repeat the operation of S2.2-S2.6.

[0015] Preferably, when processing radar observation image data, step S1 requires the use of radar image processing software, which is... StamPS, Sarproz and GAMMA .

[0016] Preferably, the calculations of equations (1)-(5) can be obtained by Matlab, Python, C++, etc., and the calculations of the spatial analysis technology and mathematical statistics technology mentioned in S4 are obtained by ArcGIS, Matlab, and Excel software.

[0017] Preferably, the extreme random tree-information land subsidence coupling model mentioned in S3 specifically includes the following steps: S3.1. Divide the dataset based on the results of S2: (6) In the formula, For a complete dataset of ground subsidence; The number of segments to divide the MK mutation; For the first Segment dataset; For the first The dataset is divided into several sub-datasets according to the year of mutation using Equation (6), thus completing the data segmentation preprocessing.

[0018] S3.2 Construct an Extreme Random Tree (ERT) machine learning model for ground subsidence and calculate the weights of influencing factors: (7) In the formula, For the first The weights of each ground subsidence influencing factor; For the first The first decision tree Information gain of each factor; For the first The sample size of each decision tree; Let be the total number of impact factors. Run the extreme random tree model and output the weight vector of all impact factors using equation (7). .

[0019] S3.3 Substitute the weights output by the extreme random tree into the information content model: (8) In the formula, For the first The weighted information content of each influencing factor; The first in the land subsidence risk zone The probability of each factor occurring; The first in the entire study area The probability of occurrence of each factor. The weights are calculated using equation (8). By incorporating information content calculation, the coupled model is constructed.

[0020] S3.4 Calculate the total information content value using the running information coupling model: (9) In the formula, This represents the total information content value for a single sample point. By iterating through all sample points, the total information content sequence is obtained. ( (Total number of sample points).

[0021] S3.5. Segment the total information content sequence using the natural breakpoint method: (10) In the formula, The preset number of segments is (usually 4 to 5 segments, corresponding to low, low-medium, medium-high, and high risks). For the first The variance of a quantum set of information segments; , The first The lower and upper breakpoint values ​​of the segment. Using equation (10) to minimize the variance within each segment, the optimal breakpoint threshold is solved, and the information sequence is divided into segments. Each risk level range.

[0022] S3.6 Determine the model termination condition and output the result: (11) In the formula, For the first The total information content of each iteration; The convergence threshold (take) If the convergence condition of equation (11) is met, the breakpoint thresholds for each risk level and the corresponding ground subsidence risk zoning results will be output; otherwise, the S3.2–S3.5 operations will be repeated.

[0023] Preferably, the weighting factor in S4 is obtained by processing basic geological and remote sensing data, and the specific processing method includes: A1. Groundwater level data processing methods: A1.1 Constructing an irregular triangular network (TIN) from groundwater contour lines, specifically including: A1.1.1 Establishing an irregular triangular network (TIN) using contour lines as feature lines: Using contour lines as structural lines, utilizing the feature points of contour lines, and using contour lines as constraint line segments; A1.1.2. Method for automatically adding feature points and optimizing the Triangular Irregular Network (TIN): This method essentially discretizes the isoline contours to establish a Triangular Irregular Network (TIN), but uses the method of adding feature points to eliminate the flat triangles in the TIN, and uses the method of optimizing the TIN to eliminate unreasonable triangulation; the addition of feature points requires the automatic extraction of certain algorithms, and the principles of these algorithms are based on the topological relationship of the original isoline contours; A1.1.3 Method for directly generating irregular triangular mesh (TIN) from discrete points of isostatic lines: directly discretize the isostatic lines, and then use a common TIN generation algorithm; A1.2. Convert to raster data: Use interpolation to convert the triangular mesh (TIN) into raster data. Convert to a raster; specify a height or NoData value for each output cell, depending on whether the cell center falls within the interpolation region of the Triangulated Irregular Network (TIN); this method can use LINEAR or NATURAL_NEIGHBORS; linear interpolation displays the TIN triangles as planes; specify the height for each output cell by finding triangles that fall in 2D space and calculating the position of the cell center relative to the triangle plane; natural neighborhood interpolation produces smoother results than linear interpolation, using a region-based weighting scheme for the nearest TIN nodes found in all directions around the center of each output cell; hard break lines in the TIN will affect the results because the natural neighborhood surface is not continuous and smooth when crossing the hard break line; the sampling distance variable represents the output cell size and can be defined in one of two ways; the OBSERVATIONS keyword is followed by an integer indicating the number of columns required in the output; cell size is determined based on the output range and the number of columns; OBSERVATIONS is used by default. 250; The CELLSIZE keyword provides the explicit cell size, followed by the desired size, which can be given in floating-point form; The Z-factor variable is used to convert the z-units of the output raster; the output height will be multiplied by this value; The default Z-factor variable is calculated to convert the z-units to the same units of measurement as the x and y units if and only if the x, y, and z units are defined in the spatial reference of the input TIN and are all standard projected units (e.g., meters and feet); If the x, y, and z units are the same, the value 1.0 is used; If no units are defined, the value 1.0 is used. A1.3. Convert 30m resolution raster data into a point layer and create a point in the output feature class; these points will be located at the center of the cell they represent; NoData cells will not be converted into points; the cell size of the input raster can be arbitrary and can belong to any valid raster dataset; calculate the intersection and take the average to obtain the groundwater level data information of each grid. A2. Compressible layer data processing methods: The compressible layer thickness contour data is converted into point data and spatial IDW interpolation is performed. A linear weighted combination of a set of sampled points is used to determine the cell value. The weights are an inverse distance function, and the surface being interpolated has a local dependent variable; it mainly depends on the power value of the inverse distance. The power parameter controls the influence of known points on the interpolation based on their distance from the output point. The power parameter is a positive real number with a default value of 2. By defining higher power values, the nearest points can be further emphasized; therefore, neighboring data will be most affected, and the surface will become more detailed (less smooth). As the power increases, the interpolation value will gradually approach the value of the nearest sampled point. Specifying smaller power values ​​will have a greater impact on surrounding points that are farther away, resulting in a smoother surface. Since the inverse distance weighting formula is not related to any actual physical process, it is impossible to determine whether a particular power value is too large; as a general rule, a power of 30 is considered too large and is therefore not recommended. If the distance or power value is large, it may generate erroneous results; the power value that produces the lowest minimum mean absolute error can be considered the optimal power value. This allows for the acquisition of raster data of the compressible layer thickness in the study area. A3. Method for processing distance data from fault zones: First, based on the distribution data of fault zones in the study area, a raster image of the distance to the fault zones was generated using the Euclidean distance method, with the pixel size set to 240 m. Then, the distance values ​​of each grid cell to the fault zone were extracted using a zonal statistics tool (the zonal statistics are displayed in a table) as the geological structural influence factor of that grid.

[0024] A4. Road Data Processing Methods: Based on existing road data classification and considering the differences in average vehicle speeds across various road types in my country, the weights for different road types are determined. A5. IBI Data Processing Method: First, IBI data for the study area was acquired through the Google Earth Engine (GEE) platform and resampled into 30 m resolution raster images. Then, the average IBI value within each grid cell was obtained using a zonal statistical method as a quantitative indicator of the regional static load.

[0025] Preferably, the novel ground subsidence risk assessment model coupled with extreme random trees and information content mentioned in S5 specifically includes the following steps: S5.1 Organize the model input parameter set: (12) In the formula, This is a set of parameters for factors affecting ground subsidence. These are the groundwater levels for the four layers; The thicknesses of the three compressible layers are respectively; This refers to the distance from the fault zone. Road network density; The IBI building index is used. The average annual settlement rate calculated using S1 is also imported. ( (Total number of sample points), weighting factors obtained from S4 .

[0026] S5.2, Integrating annual average settlement rate and weighting factors: (13) In the formula, For the first The weighted adjusted weights of each factor; Let be the average annual settlement rate of the study area. Equation (13) is used to incorporate the spatial differences in the average annual settlement rate into the weighting factor, and the final weight vector is obtained after correction. .

[0027] S5.3 Core formula for constructing a regional settlement coupling risk assessment model: (14) In the formula, For the first Ground subsidence risk assessment value for each sample point; For the first The nth sample point The probability of each factor occurring in the settlement risk zone; For the first The global probability of each factor occurring in the study area; The contribution coefficient to the average annual settlement rate; ~ The 10 input factors corresponding to Equation (12) are: four-layer groundwater level, three-layer compressible layer, fault zone distance, road network density, and IBI index.

[0028] S5.4 Training an extreme random tree model to obtain factor weights: Calculate the initial weights of 10 factors using Python code. and corrected weights .

[0029] S5.5, Substitute the information content into the model to calculate the risk value: Adjust weights Substitute into equation (14) and iterate through all sample points to calculate. The sequence of ground subsidence risk values ​​for the study area was obtained.

[0030] S5.6 Validate the effectiveness of the novel coupling model (using the receiver operating characteristic curve and area under the curve as indicators of model accuracy): (15) In the formula, The area under the ROC curve; (True positive rate) is the correct identification rate of settlement risk areas; The false positive rate (ROC) is the probability that a non-sedimentation risk area is mistakenly identified as a risk area. The ROC curve uses... For the horizontal axis, Plot on the vertical axis. The closer the value is to 1, the higher the model accuracy; if If the model accuracy threshold is met, the model is considered valid; otherwise, return to step S5.2 for adjustment. The value is then recalculated and the corrected weights are adjusted.

[0031] S5.7 Output a new ground subsidence risk assessment model: Output model structure (Equation 14), final corrected weights of the 10 types of factors Contribution coefficient of average annual settlement rate ROC curve and Values, and risk assessment values ​​for each sample point in the study area. The construction of a new coupling model was completed.

[0032] Preferably, the calculations of equations (1) to (15) are performed using Python or Matlab.

[0033] Compared with existing technologies, this invention provides a ground subsidence risk assessment method coupled with extreme random tree-information, which has the following beneficial effects: (1) Solve the problem of subjectivity in traditional models and improve the objectivity of evaluation; This scheme abandons evaluation methods that rely too heavily on subjective human experience, such as expert scoring and analytic hierarchy process. Instead, it integrates extreme random tree machine learning models and information content models. By using multi-source geoscientific data to drive the automatic calculation of influencing factor weights, it combines spatial analysis and mathematical statistics techniques to quantify the contribution of each factor to land subsidence. This eliminates human judgment bias from a technical perspective, making the regional land subsidence risk assessment results more consistent with actual geological and human activity patterns.

[0034] (2) To compensate for the shortcomings in the stability of machine learning models and enhance the reliability of evaluation; To address the issue that machine learning models such as logistic regression and ordinary random forest are susceptible to the influence of indicator selection and data processing methods, this solution first segments the sedimentation time series data through the Mann-Kendall mutation test, then constructs an extreme random tree model based on the segmentation results, and embeds the model weights into the information content model to form a coupled framework. This approach retains the high prediction accuracy of machine learning models while leveraging the mathematical principles of the information content model to improve the physical interpretability of the model, reducing the interference of data quality on the evaluation results, and achieving a dual improvement in stability and accuracy.

[0035] (3) Break through the limitations of static analysis and realize dynamic spatiotemporal evaluation of settlement risk; Existing models mostly rely on static data of the study area, which cannot capture the spatiotemporal dynamic evolution characteristics of land subsidence. This scheme uses long-term InSAR technology to obtain time series data of regional surface deformation, combines abrupt change detection to identify abrupt change nodes in subsidence rate, incorporates the spatial difference of annual average subsidence rate into the weighting factor correction, and dynamically calculates the risk assessment value of each sample point through coupled model. This can reveal the changing trend of subsidence risk in the time dimension and accurately identify its spatial distribution characteristics, adapting to the characteristics of multi-factor superposition and dynamic evolution of land subsidence.

[0036] (4) Integrate multi-technology and multi-source data to improve evaluation efficiency and practical value; This solution integrates multiple technologies, including temporal InSAR, Google Earth Engine (GEE), extreme random trees, and information content models. It fuses multi-source data such as radar imagery, GPS / leveling points, basic geology, hydrogeology, and regional dynamic / static loads representing human activities. Through standardized computational workflows and software tools (ArcGIS, Python, Matlab, etc.), it achieves integrated data processing, model building, and accuracy verification, significantly saving time and economic costs in urban geological disaster prevention and control. Simultaneously, the model's output results, such as risk zoning, factor weights, and accuracy reports, provide precise scientific basis for urban planning, construction, and safe operation, improving the practical application effectiveness of land subsidence prevention and disaster early warning.

[0037] In summary, this invention utilizes radar imagery data, firstly employing temporal InSAR technology to acquire regional land subsidence information; secondly, it integrates spatial analysis, mathematical statistics, and other techniques, combined with regional geological conditions, to determine the optimal grid scale for subsidence in the study area; finally, it integrates new technologies such as Google Earth Engine (GEE) and machine learning, comprehensively considering multi-source spatial data including groundwater level, compressible layer thickness, and OSM roads in the study area, to construct a multi-factor driven ERT-IV land subsidence risk assessment model, and conducts a spatiotemporal dynamic risk assessment of regional land subsidence. Compared with traditional models and methods, this invention more accurately achieves a spatiotemporal dynamic risk assessment of regional land subsidence, providing a more efficient and precise basis for the prevention and control of regional land subsidence disasters. This method can effectively save time and economic costs in disaster prevention and control, and can also provide important references for urban planning, geological disaster prevention and control, and sustainable development. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating a ground subsidence risk assessment model that couples extreme random trees with information content, as proposed in this invention. Figure 2 This is a schematic diagram of the time-series InSAR processing flow of a ground subsidence risk assessment model that couples extreme random trees and information content, as proposed in this invention. Figure 3 This is a schematic diagram of the ROC curve and AUC analysis of the study area in Embodiment 2 of the present invention; Figure 4 This is a stacked area diagram showing the risk levels of the study area in Embodiment 2 of the present invention; Figure 5 This is a graph showing the average risk level and trend of the study area in Embodiment 2 of the present invention. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0040] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0041] Example 1: Please refer to Figure 1 A ground subsidence risk assessment method coupled with extreme random tree-information includes the following steps; S1. Data Acquisition and Preprocessing; Long-term synthetic aperture radar (SAR) imagery data covering the study area is acquired. Temporal InSAR technology is used to obtain deformation information along the line of sight in the region. Measured GPS or leveling data are used to verify the temporal InSAR monitoring results. The line-of-sight values ​​are converted to vertical values ​​to obtain high-precision vertical deformation data of permanent scattering points. When obtaining the regional annual average subsidence rate using interpolation methods such as Kriging interpolation / inverse range weighting, radar image processing software such as StamPS, Sarproz, and GAMMA is required to complete the data processing.

[0042] S2. Combining the Mann-Kendall test method with spatial patterns of rapid regional changes; Using the Mann-Kendall mutation test method, the years of abrupt changes in the annual average settlement rate calculated in S1 are identified. The specific steps are as follows: S2.1. Represent a time series of length n as follows: (1) In the formula, n=15 (2010-2024); Given the observations for year i, initialize the time series and perform data preprocessing such as outlier removal and missing value imputation.

[0043] S2.2 Constructing the ordered series and calculating the statistics : (2) In the formula, It is a symbolic function; Let j be the observation value in year j; use equation (2) to calculate the sequence of time points and solve for its mean. With variance .

[0044] S2.3 Calculate the positive standardized statistic : (3) In the formula, The mean of the ordered sequence. The variance is calculated using equation (3) for the forward statistical sequence. .

[0045] S2.4 Calculate the inverse statistical sequence : (4) In the formula, The statistics are calculated for the reverse sequence; the reverse statistical sequence is obtained using equation (4). .

[0046] S2.5 Mutation determination criteria: (5) In the formula, 1.96 is the significance level. The corresponding critical value; if and If the two curves intersect within the ±1.96 critical line and satisfy equation (5), then year k is a year of significant abrupt change.

[0047] S2.6. Recalculate the statistics using formula (2) and update the mutation point judgment results in real time; S2.7. Determine whether the termination condition is met based on the result obtained in S2.6. If it is met, exit and output the mutation year; otherwise, repeat the operation of S2.2-S2.6.

[0048] S3. Construct a coupled extreme random tree and information content model; A ground subsidence machine learning model based on extreme random trees was constructed for the study area. The trained extreme random tree model was embedded into the information content model computation framework to achieve coupling between the extreme random tree and the information content model, forming an extreme random tree-information content coupled model. The specific steps are as follows: S3.1. Divide the dataset based on the mutation results of S2: (6) In the formula, D is the complete ground settlement dataset; M=2, which is the number of segments divided by the MK mutation; This is the m-th segment of the dataset; Let m be the year of the mutation, where , The dataset is divided into several subsets according to the year of mutation using equation (6) to complete the data segmentation preprocessing.

[0049] S3.2 Construct an extreme random tree machine learning model for ground subsidence and calculate the weights of influencing factors: (7) In the formula, Let be the weight of the j-th ground subsidence influencing factor; Let be the information gain of the j-th factor in the t-th decision tree; Let J be the sample size of the t-th decision tree; J = 10, which is the total number of influencing factors. Run the extreme random tree model and output the weight vector of all influencing factors using equation (7). .

[0050] S3.3 Substitute the weights output by the extreme random tree into the information content model: (8) In the formula, The weighted information content value of the j-th influence factor; Let be the probability of the j-th factor occurring in the land subsidence risk zone; Let be the probability of the j-th factor appearing in the entire study area. The weights are calculated using equation (8). Integrate information content calculation to construct a coupled model.

[0051] S3.4 Calculate the total information content value using the running information coupling model: (9) In the formula, This represents the total information content value for a single sample point. By iterating through all sample points, the total information content sequence is obtained. (N is the total number of sample points).

[0052] S3.5. Segment the total information content sequence using the natural breakpoint method: (10) In the formula, K is the preset number of segments (usually 4 to 5 segments, corresponding to low, low-medium, medium-high, and high risks). Let be the variance of the k-th information quantum set; , These are the lower and upper breakpoint values ​​of the k-th segment, respectively. Equation (10) is used to minimize the variance within each segment, solve for the optimal breakpoint threshold, and divide the information sequence into K risk level intervals.

[0053] S3.6 Determine the model termination condition and output the result: (11) In the formula, This represents the total information content value in the k-th iteration; The convergence threshold (take) If the convergence condition of equation (11) is met, the breakpoint thresholds for each risk level and the corresponding ground subsidence risk zoning results will be output; otherwise, the S3.2–S3.5 operations will be repeated.

[0054] S4. Multi-source data processing and acquisition of weighting factors for regional ground subsidence; By integrating multi-source spatial data (groundwater level, compressible layer thickness, fault zone distance, etc.) and using technologies such as ArcGIS (spatial analysis) and Matlab / Excel (mathematical statistics), the contributions of each influencing factor in the extreme random tree machine learning model of regional land subsidence were obtained.

[0055] S5. Construct a novel ground subsidence risk assessment model that couples extreme random trees with information content; Substituting the annual average settlement rate calculated in S1 and the contribution of the weighting factors obtained in S4 into the coupled extreme random tree and information content model in S3, a new land subsidence risk assessment model based on this coupled model is constructed. The specific steps are as follows: S5.1 Organize the model input parameter set: (12) In the formula, X is the set of parameters for ground subsidence influencing factors; These are the groundwater levels for the four layers; The thicknesses of the three compressible layers are respectively; Distance from the fault zone (m); Road network density (km / km²); The IBI building index is used. The average annual settlement rate calculated using S1 is also imported. (N is the total number of sample points), the weighting factor obtained from S4 .

[0056] S5.2, Integrating annual average settlement rate and weighting factors: (13) In the formula, The weighted adjusted weight for the j-th factor; Let be the average annual settlement rate of the study area. Equation (13) is used to incorporate the spatial differences in the average annual settlement rate into the weighting factor, and the final weight vector is obtained after correction. .

[0057] S5.3, Core formula for risk assessment of the coupled model: (14) In the formula, Let be the ground subsidence risk assessment value for the i-th sample point; Let be the probability of the j-th factor appearing in the settlement risk zone in the i-th sample point; Let be the global probability of the j-th factor in the study area; The contribution coefficient of the annual average settlement rate; j=1~10 corresponds to the 10 types of input factors in equation (12) (four-layer groundwater level, three-layer compressible layer, fault zone distance, road network density, IBI index).

[0058] S5.4 Training an Extremely Random Tree Model to Obtain Factor Weights: An extreme randomly generated tree model is built using Python's sklearn library. Parameters are set, and the initial weights w and adjusted weights for the 10 classes of factors are calculated. .

[0059] S5.5, Substitute into the information content model to calculate the risk value: Adjust the weights. Substitute into equation (14) and iterate through all sample points to calculate. The sequence of ground subsidence risk values ​​for the study area was obtained.

[0060] S5.6 Validate the effectiveness of the novel coupling model (using ROC curves and AUC values ​​as model accuracy validation metrics): (15) In the formula, AUC is the area under the ROC curve; TPR (True Positive Rate) is the correct identification rate of settlement risk areas; FPR (False Positive Rate) is the probability of a non-settlement risk area being mistakenly identified as a risk area. The ROC curve is plotted with FPR on the horizontal axis and TPR on the vertical axis. The closer the AUC value is to 1, the higher the model accuracy. In this study, AUC ≥ 0.85 is set as the model accuracy threshold. If this threshold is met, the model is considered effective; otherwise, return to S5.2 to adjust the λ value (e.g., λ ± 0.05) and recalculate the corrected weights.

[0061] S5.7 Output a new land subsidence risk assessment model: Output the core formula of the model (Equation 14) and the final corrected weights of 10 types of factors. The contribution coefficient of annual average settlement rate λ, ROC curve and AUC value, and risk assessment values ​​of each sample point in the study area. The construction of a new coupling model was completed.

[0062] The calculations of equations (1)-(15) can be implemented using Python, Matlab and C++. In this study, the statistical calculations were completed using Python's pandas and scipy libraries. Spatial analysis and mathematical statistics calculations were completed using ArcGIS 10.8, Matlab R2022b and Excel 2019.

[0063] S6. Model accuracy verification and result output; ROC curve and AUC value are selected as accuracy parameters to perform accuracy analysis on the novel ground subsidence risk assessment model coupled with extreme random tree-information constructed in S5. If AUC ≥ 0.85, the feasibility of the model is verified, and the verification results (including risk zoning map and accuracy report) are output. If AUC < 0.85, return to S5.2 to adjust the λ value and retrain the model until the accuracy requirements are met.

[0064] This invention utilizes radar interferometric imagery data and employs time-series InSAR technology to acquire regional land subsidence information. Based on this, it integrates basic geology, hydrogeology, regional dynamic / static loads representing human activities, and combines the Mann-Kendall subsidence abrupt change test results to construct an ERT-IV land subsidence risk assessment model with multiple weighted factors (groundwater level, compressible layer thickness, fault zone, and dynamic / static loads). This model is used to conduct dynamic risk assessments of regional land subsidence, providing a scientific basis for the prevention and control of land subsidence disasters.

[0065] Example 2: Based on Example 1, but with a difference, this example selects a 2 km section along Beijing Metro Line 6 as the study area. Using GAMMA software, radar remote sensing images from 2010-2020 (RADARSAT-2) and 2020-2024 (Sentinel-1A) are processed to obtain temporal surface deformation results for the study area. Leveling data is used for verification, and line-of-sight displacement is converted into vertical displacement information. The risk assessment model of this invention is used to obtain the dynamic spatiotemporal evolution of ground subsidence risk in the study area. The main steps are as follows: Step 1: Using 81 RADARSAT-2 and 60 Sentinel-1A images covering the Beijing Subway Line 6 from 2010 to 2024, the line-of-sight surface deformation information for the study area from 2010 to 2016, 2017 to 2020, and 2021 to 2024 was obtained using the GAMMA software and the time-series InSAR method, and then converted into vertical deformation. The InSAR deformation results were verified using leveling point monitoring data. Based on this, the long-term time-series surface deformation information for the study area from 2010 to 2024 was obtained through time-series fusion.

[0066] Table 1. Surface Deformation Monitoring Data Set

[0067] Step Two: First, using spatial analysis, mathematical statistics and other techniques, combined with the geological conditions of the study area, the optimal grid scale for studying settlement was determined to be 240 meters. Taking into account the basic geological and hydrogeological conditions of the study area, the study area was divided into four sub-regions.

[0068] Step 3: The inverse distance weighted interpolation method was used to obtain the annual average subsidence rate at the grid scale in the study area from 2010 to 2024. The Mann-Kendall mutation test and mathematical statistics were used to obtain the ground subsidence mutation information in the study area and identify the segmented characteristics of the ground subsidence evolution. Based on this, an extreme random tree ground subsidence simulation model of the study area was established and embedded with the information content model calculation framework to construct a regional extreme random tree-information content ground subsidence coupling model.

[0069] By integrating platforms such as GEE, ArcGIS, and EXCEL, data on various factors affecting the study area (compressible layer thickness, groundwater level, distance from the fault zone, road OSM, and IBI, etc.) were processed to construct an extreme random tree settlement characteristic dataset for the study area (Table 2).

[0070] Table 2. Extremely Random Tree Settlement Feature Dataset;

[0071] Step Four: Ground subsidence information and each weight factor are embedded into the established ERT-IV ground subsidence risk assessment model. The optimal parameters of the model are determined by random search and grid search methods. Based on this, the weight data of each influencing factor and the results of a new ground subsidence risk assessment model coupled with extreme random tree-information are obtained.

[0072] Table 3. Dataset of optimal parameters for the model;

[0073] Step 5: Conduct model accuracy analysis (ROC, AUC) to determine the accuracy of the ERT-IV settlement risk assessment model. For example... Figure 3 The ROC curves and AUC analysis of the study area are shown below; The verification results show that the AUC value of the risk assessment model of the present invention is between 0.87 and 0.99, which shows excellent model accuracy and confirms the feasibility and effectiveness of the present invention in land subsidence risk assessment.

[0074] like Figure 4 and Figure 5 As shown, this invention overcomes the limitations of traditional static risk assessment models and provides a dynamic spatiotemporal risk assessment scheme for land subsidence. It can dynamically identify the spatiotemporal evolution of subsidence risk, adapting to the physical process of multi-factor dynamic superposition and the characteristics of dynamic evolution in land subsidence. This provides a more efficient and accurate basis for regional land subsidence disaster prevention and control. This invention can effectively save time and economic costs in disaster prevention and control, and also provides important reference for urban planning, geological disaster prevention and control, and sustainable development.

[0075] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A ground subsidence risk assessment method coupled with extreme random tree-information, characterized in that, Includes the following steps; S1. Data Acquisition and Preprocessing: Acquire long-term synthetic aperture radar (SAR) image data covering the study area. Use time-series SAR interferometry and small baseline interferometry to obtain surface deformation information along the line of sight. Use measured GPS or leveling data to verify the time-series InSAR monitoring results. Convert the line-of-sight values ​​to vertical values ​​to obtain high-precision vertical deformation data of permanent scatterer points. Obtain the annual average subsidence rate of the study area through Kriging interpolation / inverse distance weighting interpolation methods. S2. Segmented analysis of settlement time series based on Mann-Kendall mutation test: The time series annual average settlement rate obtained in S1 is segmented using the Mann-Kendall mutation test method; S3. Constructing a coupled extreme random tree and information content model: Establish a ground subsidence machine learning model based on extreme random trees in the study area, embed the trained extreme random tree model into the information content model calculation framework, realize the coupling of extreme random trees and information content model, and construct a regional extreme random tree-information content coupled model. S4. Multi-source data processing and acquisition of weight factors for regional land subsidence: This includes comprehensive basic geological, hydrogeological, and regional dynamic / static load data representing human activities. Using spatial analysis and mathematical statistics techniques, the contribution of each weight factor in the regional extreme random tree land subsidence machine learning model is obtained. S5. Construct a novel ground subsidence risk assessment model coupled with extreme random tree-information: Input the annual average subsidence rate calculated in S1 and the weight factor contribution obtained in S4 into the extreme random tree-information coupled model constructed in S3 to establish a novel ground subsidence risk assessment model coupled with extreme random tree-information in the region. S6. Model Accuracy Validation and Result Output: Using the Receiver Operating Characteristic (ROC) curve and area under the curve (AUC) as accuracy evaluation indicators, we will conduct an accuracy analysis of the novel ground subsidence risk assessment model with coupled extreme random tree-information obtained in S5, verify the feasibility of the model, and output the validation results.

2. The ground subsidence risk assessment method coupled with extreme random tree-information as described in claim 1, characterized in that, The Mann-Kendall mutation detection method in S2 specifically includes the following steps: S2.

1. Define n as the length of the time series, specifically expressed as: （1） In the formula, n is the sequence length, corresponding to the number of sample years; For the first Annual observations; initialize time series and complete data preprocessing; S2.2 Constructing the ordered series and calculating the statistics The The expression is: （2） In the formula, It is a symbolic function; x j Represented as the first Annual observations; use equation (2) to calculate the sequence of events at each time point, and solve for its mean and variance; S2.3, Define the positive standardized statistic as The The expression is: （3） In the formula, The mean of the ordered sequence. The variance is calculated using equation (3) for the forward statistical sequence. ; S2.4, Define the reverse statistical sequence as The The expression is: （4） In the formula, The statistics are calculated for the reverse sequence; the reverse statistical sequence is obtained using equation (4). . S2.5, where the statistic is... The mutation determination criteria are expressed as follows: （5） In the formula, 1.96 is the significance level. The corresponding critical value; and The two curves are If the critical lines intersect inside each other and satisfy equation (5), then the first... The year was a year of significant change; S2.

6. Recalculate the statistic using formula (2) in S2.2 and update the statistic mutation judgment result in real time; S2.

7. Determine whether the termination condition is met based on the result obtained in S2.

6. If it is met, exit and output the year of the mutation; otherwise, repeat the operation of S2.2–S2.

6.

3. The ground subsidence risk assessment method coupled with extreme random tree-information as described in claim 1, characterized in that, The acquisition of radar observation image data in S1 requires the use of radar image processing software, including StamPS, Sarproz, and GAMMA.

4. The ground subsidence risk assessment method coupled with extreme random tree-information as described in claim 2, characterized in that, The calculation methods of formulas (1) to (5) in S2.1 to S2.5 include calculations using Matlab, Python, and C++. The spatial analysis and mathematical statistics calculation software in S4 includes ArcGIS, Matlab, and Excel.

5. The ground subsidence risk assessment method coupled with extreme random tree-information as described in claim 1, characterized in that, The construction of the extreme random tree-information coupling model in S3 specifically includes the following steps: S3.

1. Divide the dataset based on the results of S2: （6） In the formula, For a complete dataset of ground subsidence; for MK The number of segments into which the mutation occurs; For the first Segment dataset; For the first The dataset is divided into several sub-datasets according to the mutation years using Equation (6) to complete the data segmentation preprocessing. S3.2 Construct an Extreme Random Tree (ERT) machine learning model for ground subsidence and calculate the weights of influencing factors: （7） In the formula, For the first The weights of each ground subsidence influencing factor; For the first The first decision tree Information gain of each factor; For the first The sample size of each decision tree; The total number of impact factors; run the extreme random tree model and output the weight vector of all impact factors using equation (7). ; S3.3 Substitute the weights output by the extreme random tree into the information content model: （8） In the formula, For the first The weighted information content of each influencing factor; The first in the land subsidence risk zone The probability of each factor occurring; The first in the entire study area The probability of occurrence of each factor. The weights are calculated using equation (8). By incorporating information content calculation, the coupled model is constructed. S3.4 Calculate the total information content value using the running information coupling model: （9） In the formula, The total information content value for a single sample point is given; by iterating through all sample points, the total information content sequence is obtained. ( (Total number of sample points). S3.

5. Segment the total information content sequence using the natural breakpoint method: （10） In the formula, The preset number of segments; For the first The variance of a quantum set of information segments; , The first The lower and upper breakpoint values ​​of the segment; using equation (10) to minimize the variance within each segment, the optimal breakpoint threshold is solved, and the information sequence is divided into segments. One risk level range; S3.6 Determine the model termination condition and output the result: （11） In the formula, For the first The total information content of each iteration; The threshold is the convergence threshold. If the convergence condition of equation (11) is met, the breakpoint threshold of each risk level and the corresponding ground subsidence risk zoning result are output. Otherwise, the S3.2–S3.5 operation is repeated.

6. The ground subsidence risk assessment method coupled with extreme random tree-information as described in claim 1, characterized in that, The novel ground subsidence risk assessment model in S5, which couples extreme random trees with information content, specifically includes the following steps: S5.1 Organize the model input parameter set: （12） In the formula, This is a set of parameters for factors affecting ground subsidence. These are the groundwater levels for the four layers; The thicknesses of the three compressible layers are respectively; This refers to the distance from the fault zone. Road network density; The IBI building index is used; the average annual settlement rate calculated using S1 is also imported. ,in The total number of sample points and the weighting factor obtained by S4 ; S5.2, Integrating annual average settlement rate and weighting factors: （13） In the formula, For the first The weighted adjusted weights of each factor; The average annual settlement rate of the study area is given; the spatial difference of the average annual settlement rate is incorporated into the weighting factor using equation (13), and the final weight vector is obtained after correction. ; S5.3, Core formula for risk assessment of the coupled model: （14） In the formula, For the first Ground subsidence risk assessment value for each sample point; For the first The nth sample point The probability of each factor occurring in the settlement risk zone; For the first The global probability of each factor occurring in the study area; The contribution coefficient to the average annual settlement rate; ~ The 10 types of input factors correspond to equation (12); S5.4 Training an extreme random tree model to obtain factor weights: Calculate the initial weights of 10 factors using Python code. and corrected weights ; S5.5, Substitute the information content into the model to calculate the risk value: Adjust weights Substitute into equation (14) and iterate through all sample points to calculate. The sequence of ground subsidence risk values ​​for the study area was obtained. S5.

6. Validate the effectiveness of the novel coupling model by using the receiver operating characteristic (ROC) curve and area under the curve (AUC) as validation metrics for model accuracy: （15） In the formula, The area under the ROC curve; (True positive rate) is the correct identification rate of settlement risk areas; (False positive rate) is the probability that a non-sedimentation risk area is mistakenly identified as a risk area; the ROC curve is based on... For the horizontal axis, Plot on the vertical axis. The closer the value is to 1, the higher the model accuracy; if If the model accuracy threshold is met, the model is considered valid; otherwise, return to step S5.2 for adjustment. The value is then recalculated and the corrected weights are adjusted. S5.7 Output a new ground subsidence risk assessment model: The model structure and the final corrected weights of the 10 factors are output through equation (14). Contribution coefficient of average annual settlement rate ROC curve and Values, and risk assessment values ​​for each sample point in the study area. The construction of a new coupling model was completed.

7. The ground subsidence risk assessment method coupled with extreme random tree-information according to any one of claims 6 is wherein the calculation of equations (1) to (15) is performed by Python or Matlab.

Images