A Method for Assessing Urban Groundwater Inrush Risk Based on 3D Mesh Model and XGBoost Algorithm
By combining a 3D mesh model with the XGBoost algorithm, the limitations of data and the linearization of models in urban groundwater inrush risk assessment are solved, local features are finely characterized, and high-precision risk assessment and real-time decision support are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- 北京超维创想信息技术有限公司
- Filing Date
- 2025-03-25
- Publication Date
- 2026-05-26
AI Technical Summary
Existing technologies for assessing the risk of sudden groundwater inrush in cities suffer from limitations in data, linearization of models, neglect of local features, and insufficient visualization and decision support, leading to biased assessments, misjudgments, or omissions.
By combining a 3D mesh model with the XGBoost algorithm, geological and environmental information is obtained from multi-source heterogeneous data. The mesh units are dynamically divided, interactive features are extracted, a risk assessment model is constructed, and real-time visualization decision support is provided through a 3D GIS platform.
It enables accurate risk assessment of complex geological conditions, improves the accuracy of hazardous area identification and the real-time nature of assessment, and provides intuitive decision support.
Smart Images
Figure CN120278516B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of urban groundwater inrush risk assessment technology, and in particular to a method for urban groundwater inrush risk assessment based on a three-dimensional mesh model and the XGBoost algorithm. Background Technology
[0002] Urban groundwater inrush is a common geological hazard that poses a serious threat to urban infrastructure, residents' lives and property. Current assessment methods for urban groundwater inrush have the following shortcomings:
[0003] Data limitations: Traditional methods rely on empirical formulas and limited static data (such as borehole data), which are insufficient to reflect the spatiotemporal dynamics of complex geological conditions, leading to one-sided risk assessments.
[0004] Model linearization defects: Existing technologies mostly use linear regression or simple statistical models, which cannot effectively handle the nonlinear relationship between geological factors and sudden surge risk, resulting in large prediction errors.
[0005] Ignoring local features: Most methods assess the region as a whole, ignoring local geological heterogeneity such as faults and fissures, leading to missed or misjudged high-risk areas.
[0006] Insufficient visualization and decision support: The assessment results are mostly presented in two-dimensional charts, lacking interactive three-dimensional spatial analysis functions, making it difficult to provide intuitive guidance for disaster prevention in engineering.
[0007] To address the aforementioned issues, this invention proposes a method that integrates 3D mesh modeling and the XGBoost algorithm. This method uses high-resolution meshes to characterize geological heterogeneity, combines machine learning to handle complex nonlinear relationships, and integrates spatial statistics and dynamic visualization techniques to achieve accurate risk assessment and real-time decision support. Summary of the Invention
[0008] Therefore, the purpose of this invention is to provide a method for assessing the risk of urban groundwater inrush based on a three-dimensional mesh model and the XGBoost algorithm, so as to at least solve the above problems.
[0009] The technical solution adopted in this invention is as follows:
[0010] A method for assessing the risk of urban groundwater inrush based on a three-dimensional mesh model and the XGBoost algorithm, the method comprising the following steps:
[0011] S1. Through multi-source heterogeneous data interfaces, real-time geological data, environmental data and historical disaster data of the study area are obtained. The geological data includes the three-dimensional spatial attributes of stratigraphy, fault zone distribution and karst caves, while the environmental data includes dynamic changes in groundwater level, temporal characteristics of rainfall intensity and factors induced by human activities.
[0012] S2. Based on the geographic coordinate system, the study area is divided into three-dimensional grid units of equal size. Each unit is dynamically bound to geological attribute data, and a heterogeneous geological structure model containing rock layer dip angle, shear strength and fracture density is generated by three-dimensional geological modeling software.
[0013] S3. Perform feature extraction and fusion on the three-dimensional mesh unit, extract the interaction features of the main control factors and the inducing factors, wherein the main control factors include groundwater level gradient, fault activity and lithological permeability coefficient, and the inducing factors include rainfall accumulation effect, construction disturbance intensity and seismic wave propagation attenuation coefficient, and optimize the input feature set using a dynamic weighting method based on feature importance.
[0014] S4. Constructing a groundwater inrush risk assessment model based on the XGBoost algorithm:
[0015] S5. Input real-time data into the trained model, output the surge probability value of each grid cell, use an adaptive threshold segmentation algorithm to remove isolated high-risk cells based on spatial continuity constraints, and perform spatial autocorrelation analysis based on Moran's index to divide contiguous danger zones.
[0016] S6. Interactive Decision Support: Through the 3D Geographic Information System (3D GIS) platform, the risk heat map and geological profile overlay view are dynamically rendered, and a real-time early warning module is integrated. When the volume of the dangerous area exceeds the preset safety threshold, a multi-level response mechanism is triggered.
[0017] Furthermore, in step S2, the size of the three-dimensional grid unit is adaptively adjusted according to the complexity of the geological structure. The fault zone and karst area adopt a 1-meter high-resolution grid, while the homogeneous rock layer area adopts a 5-10 meter grid. The grid division is dynamically coupled with the groundwater flow field simulation results.
[0018] Furthermore, in step S3, the dynamic weighting method specifically includes:
[0019] The contribution of each feature to the prediction of surge risk is calculated using the feature split gain value Gain generated during the training of the XGBoost model. The gain value is defined as the total gain of the feature when it is a split node in all trees.
[0020] The contribution rate is the percentage of each feature gain value to the total gain value, calculated using the following formula:
[0021]
[0022] Where F is the total number of features, features with a contribution rate higher than the preset threshold are assigned a weight of 1.2-1.5 times, and redundant features are randomly discarded using Dropout.
[0023] Furthermore, the specific steps for constructing a groundwater inrush risk assessment model based on the XGBoost algorithm are as follows:
[0024] S41. Define the loss function as a binary logical loss function:
[0025] Furthermore, a second-order Taylor expansion is introduced to approximate the gradient of the loss function to accelerate gradient calculation;
[0026] S42. Use tree structure regularization term:
[0027]
[0028] Where γ is the minimum gain threshold for leaf node splitting, ranging from 0.1 to 1.0; λ is the L2 regularization coefficient, ranging from 0.01 to 0.5; T is the total number of leaf nodes in a single tree; w j The weight value of the j-th leaf node is used to control the model complexity by combining early stopping and Bayesian hyperparameter tuning.
[0029] S43. Quantify the contribution of each feature to the surge risk based on the SHAP value, and dynamically adjust the feature weights.
[0030] Furthermore, in step S4, a spatiotemporal cross-validation strategy is adopted during model training: historical data is divided into training and validation sets according to time series, and data in the same spatial grid cell are strictly isolated during the training and validation phases to avoid overfitting caused by data leakage.
[0031] Furthermore, in step S5, the adaptive threshold segmentation algorithm is as follows:
[0032] High-risk clusters are identified based on local spatial statistics, and those that meet the following criteria are considered.
[0033]
[0034] Clustering is performed on the grid cells, where x i Let i be the risk value of the i-th grid cell. s represents the mean risk value of all grid cells within the study area. 2 Z represents the sample variance of the risk value. a w is the significance level threshold. ij This is the spatial weight matrix.
[0035] Furthermore, the 3D GIS platform supports multi-dimensional comparative analysis functions, including:
[0036] The real-time risk prediction results are overlaid and compared with the spatial distribution of historical surge events;
[0037] Based on the virtual borehole function, the correlation between geological attributes and risk values of any profile can be displayed;
[0038] Generate a time-series animation of risk evolution to simulate the sudden surge's diffusion path and impact range.
[0039] Furthermore, it also includes model iterative optimization steps:
[0040] Using an online learning framework, sensor monitoring data is received in real time. The Focal Loss function is employed to weight high-risk events in newly added samples, where:
[0041] The Focal Loss function is defined as FL(p t )=-a t (1-p t ) γ log(p t ), where p t Let a be the predicted probability of the model for the sample. t Here, γ is the category weight coefficient, and γ is the focusing parameter;
[0042] The criteria for determining high-risk events are: grid units within 50 meters of historical sudden surge events, or units where the groundwater level change rate in real-time monitoring data exceeds 10 cm / h.
[0043] Compared with the prior art, the beneficial effects of the present invention are:
[0044] 1. The present invention proposes a method for assessing the risk of urban groundwater inrush based on a three-dimensional mesh model and the XGBoost algorithm. By integrating remote sensing, geological survey, meteorological monitoring and human activity data and updating them in real time through an interface, it solves the problem of single data in traditional methods.
[0045] 2. The present invention proposes a method for assessing the risk of urban groundwater inrush based on a three-dimensional mesh model and the XGBoost algorithm. By using dynamic mesh division at the 1-10 meter level, it can finely characterize the local features of high-risk areas such as faults and karst areas, thereby improving the accuracy of dangerous area identification.
[0046] 3. The present invention proposes a method for assessing the risk of urban groundwater inrush based on a three-dimensional mesh model and the XGBoost algorithm. The method uses the XGBoost algorithm combined with second-order Taylor expansion to accelerate gradient calculation and improve model training efficiency. The method also introduces SHAP value feature interpretation analysis and scientifically allocates the weights of key factors (such as groundwater level gradient) to further improve prediction accuracy. Attached Figure Description
[0047] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only preferred embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0048] Figure 1 This is a schematic diagram of the overall process of an urban groundwater inrush risk assessment method based on a three-dimensional mesh model and the XGBoost algorithm proposed in an embodiment of the present invention. Detailed Implementation
[0049] The principles and features of the present invention are described below with reference to the accompanying drawings. The listed embodiments are only used to explain the present invention and are not intended to limit the scope of the present invention.
[0050] Reference Figure 1 This invention provides a method for assessing the risk of urban groundwater inrush based on a three-dimensional mesh model and the XGBoost algorithm. The method includes the following steps:
[0051] S1. Through multi-source heterogeneous data interfaces, real-time geological data, environmental data and historical disaster data of the study area are obtained. The geological data includes the three-dimensional spatial attributes of stratigraphy, fault zone distribution and karst caves, while the environmental data includes dynamic changes in groundwater level, temporal characteristics of rainfall intensity and factors induced by human activities.
[0052] S2. Based on the geographic coordinate system, the study area is divided into three-dimensional grid units of equal size. Each unit is dynamically bound to geological attribute data, and a heterogeneous geological structure model containing rock layer dip angle, shear strength and fracture density is generated by three-dimensional geological modeling software.
[0053] S3. Perform feature extraction and fusion on the three-dimensional mesh unit, extract the interaction features of the main control factors and the inducing factors, wherein the main control factors include groundwater level gradient, fault activity and lithological permeability coefficient, and the inducing factors include rainfall accumulation effect, construction disturbance intensity and seismic wave propagation attenuation coefficient, and optimize the input feature set using a dynamic weighting method based on feature importance.
[0054] S4. Constructing a groundwater inrush risk assessment model based on the XGBoost algorithm:
[0055] S5. Input real-time data into the trained model, output the surge probability value of each grid cell, use an adaptive threshold segmentation algorithm to remove isolated high-risk cells based on spatial continuity constraints, and perform spatial autocorrelation analysis based on Moran's index to divide contiguous danger zones.
[0056] S6. Interactive Decision Support: Through the 3D Geographic Information System (3D GIS) platform, the risk heat map and geological profile overlay view are dynamically rendered, and a real-time early warning module is integrated. When the volume of the dangerous area exceeds the preset safety threshold, a multi-level response mechanism is triggered.
[0057] For example, this method mainly includes six steps: data acquisition, 3D mesh generation and geological modeling, feature extraction and fusion, risk assessment model construction, model application and risk classification, and interactive decision support. For S1 data collection, geological data (such as stratigraphic lithology and fault zone distribution), environmental data (such as groundwater level changes and rainfall intensity), and historical disaster data (such as past groundwater inrush events) of a certain urban study area can be obtained in real time through a multi-source heterogeneous data interface. For S2 3D mesh generation and geological modeling, the study area can be divided into multiple 3D mesh units based on a geographic coordinate system. In fault zones and karst areas, due to the complex geological structure, a high-resolution mesh of 1 meter level is used; while in homogeneous rock strata areas, a mesh of 5-10 meters level is used. These mesh units are dynamically bound to geological attribute data, and a heterogeneous geological structure model containing information such as rock strata dip angle and shear strength is generated through 3D geological modeling software. For S3 feature extraction and fusion, the main controlling factors (such as groundwater level gradient and fault activity) and inducing factors (such as rainfall accumulation effect and construction disturbance intensity) can be extracted from the 3D mesh units. The contribution of each feature to the prediction of sudden surge risk was calculated using the feature splitting gain values generated during XGBoost model training. Features with high contributions were assigned higher weights, while redundant features were discarded. For S4: Building a risk assessment model, a groundwater sudden surge risk assessment model was constructed based on the XGBoost algorithm. During model training, the gradient of the second-order Taylor expansion approximate loss function was introduced to accelerate computation, and a tree structure regularization term was used to control model complexity. Furthermore, the contribution of each feature to the sudden surge risk was quantified based on SHAP values, and the feature weights were dynamically adjusted. For S5: Model application and risk classification, real-time data was input into the trained model, outputting the sudden surge probability value for each grid cell. An adaptive threshold segmentation algorithm was used to eliminate isolated high-risk cells based on spatial continuity constraints, and spatial autocorrelation analysis was performed based on the Moran index to classify contiguous danger zones. For S6: Interactive decision support, a risk heatmap and geological profile overlay view were dynamically rendered using a 3D Geographic Information System (3D GIS) platform. It supports multi-dimensional comparative analysis functions, such as overlaying and comparing real-time risk prediction results with the spatial distribution of historical surge events, displaying the correlation between geological attributes and risk values of any profile, and generating time-series animations of risk evolution.
[0058] In step S2, the size of the three-dimensional mesh unit is adaptively adjusted according to the complexity of the geological structure. A high-resolution mesh of 1 meter level is used in the fault zone and karst area, while a mesh of 5-10 meters level is used in the homogeneous rock layer area. The mesh division is dynamically coupled with the simulation results of the groundwater flow field.
[0059] For example, in conducting a groundwater inrush risk assessment in a complex geological area of a city, if the area includes multiple fault zones and karst regions, as well as large areas of homogeneous rock strata, a high-resolution grid of 1 meter can be used for the mesh generation of fault zones and karst regions. In these geologically complex areas, high-resolution grids can more accurately capture the geological features of fault zones and karst regions, including minute variations in rock strata and the distribution of fractures. This is crucial for subsequent risk assessment, as these areas are often high-risk zones for groundwater inrush. For the mesh generation of homogeneous rock strata regions: a 5-10 meter grid can be used. Compared to fault zones and karst regions, the geological structure of homogeneous rock strata is relatively simple, so a larger grid size can be used. This helps reduce computational load and improve assessment efficiency. Regarding the dynamic coupling of mesh generation and groundwater flow field simulation: the results of groundwater flow field simulation are considered when generating the mesh. By dynamically coupling grid generation and groundwater flow field simulation, the impact of groundwater flow on the risk of inrush can be assessed more accurately. For example, in areas where groundwater flow is faster, a denser grid may be used to more accurately capture changes in water flow.
[0060] In step S3, the dynamic weighting method specifically includes:
[0061] The contribution of each feature to the prediction of surge risk is calculated using the feature split gain value Gain generated during the training of the XGBoost model. The gain value is defined as the total gain of the feature when it is a split node in all trees.
[0062] The contribution rate is the percentage of each feature gain value to the total gain value, calculated using the following formula:
[0063]
[0064] Where F is the total number of features, features with a contribution rate higher than the preset threshold are assigned a weight of 1.2-1.5 times, and redundant features are randomly discarded using Dropout.
[0065] For example, a dynamic weighting method is used to calculate the contribution of each feature to the prediction of sudden surge risk, and the feature weights are adjusted according to the contribution. It is assumed that a series of controlling and inducing factors, including groundwater level gradient, fault activity, and cumulative rainfall effects, are extracted from the three-dimensional grid cells. During the XGBoost model training process, the total gain (i.e., feature split gain value) of each feature when it acts as a split node in all trees is calculated. Then, based on these gain values, the contribution of each feature to the prediction of sudden surge risk is calculated, and features with contributions higher than a preset threshold are assigned higher weights (e.g., 1.2-1.5 times), while redundant features are discarded. This allows for more accurate capture of key features and improves the accuracy of risk assessment.
[0066] The specific steps involved in constructing a groundwater inrush risk assessment model based on the XGBoost algorithm are as follows:
[0067] S41. Define the loss function as a binary logical loss function:
[0068] Furthermore, a second-order Taylor expansion is introduced to approximate the gradient of the loss function to accelerate gradient calculation;
[0069] S42. Use tree structure regularization term:
[0070]
[0071] Where γ is the minimum gain threshold for leaf node splitting, ranging from 0.1 to 1.0; λ is the L2 regularization coefficient, ranging from 0.01 to 0.5; T is the total number of leaf nodes in a single tree; w j The weight value of the j-th leaf node is used to control the model complexity by combining early stopping and Bayesian hyperparameter tuning.
[0072] S43. Quantify the contribution of each feature to the surge risk based on the SHAP value, and dynamically adjust the feature weights.
[0073] For example, a binary logistic loss function is defined, and a second-order Taylor expansion is introduced to approximate the gradient of the loss function to accelerate gradient calculation. A tree-structured regularization term is used to control model complexity, including setting parameters such as the minimum gain threshold for leaf node splitting and the L2 regularization coefficient. Simultaneously, early stopping and Bayesian hyperparameter tuning are combined to further optimize the model. Finally, the contribution of each feature to the surge risk is quantified based on SHAP values, and the feature weights are dynamically adjusted according to these contributions. This allows the model to continuously learn and optimize during training, improving the accuracy and stability of risk assessment.
[0074] In step S4, a spatiotemporal cross-validation strategy is adopted during model training: historical data is divided into training and validation sets according to time series, and data in the same spatial grid cell are strictly isolated during the training and validation phases to avoid overfitting caused by data leakage.
[0075] For example, to ensure the model's generalization ability and avoid overfitting, a spatiotemporal cross-validation strategy can be used for model training. For instance, historical data can be divided into training and validation sets according to time series, ensuring that data within the same spatial grid cell is strictly isolated during training and validation. This prevents the model from accessing data in the validation set during training, thus avoiding overfitting caused by data leakage. Through spatiotemporal cross-validation, the model's performance can be evaluated more accurately, and the optimal model parameters can be selected for subsequent risk assessment.
[0076] In step S5, the adaptive threshold segmentation algorithm is as follows:
[0077] High-risk clusters are identified based on local spatial statistics, and those that meet the following criteria are considered.
[0078]
[0079] Clustering is performed on the grid cells, where x i Let i be the risk value of the i-th grid cell. s represents the mean risk value of all grid cells within the study area. 2 Z represents the sample variance of the risk value. a w is the significance level threshold. ij This is the spatial weight matrix.
[0080] For example, after outputting the surge probability value for each grid cell, an adaptive threshold segmentation algorithm is used to remove isolated high-risk cells and delineate contiguous danger zones. This algorithm identifies high-risk clusters based on local spatial statistics and clusters grid cells that meet specific conditions, including the risk value of the grid cell, the mean and variance of the risk values of all grid cells within the study area. Through the adaptive threshold segmentation algorithm, isolated high-risk cells can be effectively removed, and contiguous danger zones can be delineated, providing decision-makers with a more intuitive risk distribution map.
[0081] The 3D GIS platform supports multi-dimensional comparative analysis functions, including:
[0082] The real-time risk prediction results are overlaid and compared with the spatial distribution of historical surge events;
[0083] Based on the virtual borehole function, the correlation between geological attributes and risk values of any profile can be displayed;
[0084] Generate a time-series animation of risk evolution to simulate the sudden surge's diffusion path and impact range.
[0085] For example, by using a 3D Geographic Information System (3D GIS) platform that supports multi-dimensional comparative analysis, decision-makers are provided with a wealth of visualization tools and data analysis methods. Real-time risk prediction results can be overlaid and compared with the spatial distribution of historical surge events, showing the correlation between geological attributes and risk values of any profile, and generating time-series animations of risk evolution, so that decision-makers can more intuitively understand the risk distribution and changing trends, thereby making more accurate decisions.
[0086] This embodiment also includes a model iterative optimization step:
[0087] Using an online learning framework, sensor monitoring data is received in real time. The Focal Loss function is employed to weight high-risk events in newly added samples, where:
[0088] The Focal Loss function is defined as FL(p t )=-a t (1-p t ) γ log(p t ), where p t Let a be the predicted probability of the model for the sample. t Here, γ is the category weight coefficient, and γ is the focusing parameter;
[0089] The criteria for determining high-risk events are: grid units within 50 meters of historical sudden surge events, or units where the groundwater level change rate in real-time monitoring data exceeds 10 cm / h.
[0090] For example, to improve the accuracy and reliability of the model, an online learning framework can be used for iterative optimization. This framework can receive sensor monitoring data in real time and use the Focal Loss function to weight high-risk events in newly added samples. High-risk events include grid cells within a certain range surrounding historical groundwater inrush events or cells in real-time monitoring data where the rate of change in groundwater level exceeds a set threshold. By continuously updating and optimizing the model, the accuracy and real-time performance of risk assessment can be continuously improved, providing stronger support for early warning and prevention of urban groundwater inrush risks.
[0091] 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, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for assessing the risk of urban groundwater inrush based on a three-dimensional mesh model and the XGBoost algorithm, characterized in that, The method includes the following steps: S1. Through multi-source heterogeneous data interfaces, real-time geological data, environmental data and historical disaster data of the study area are obtained. The geological data includes the three-dimensional spatial attributes of stratigraphy, fault zone distribution and karst caves, while the environmental data includes dynamic changes in groundwater level, temporal characteristics of rainfall intensity and factors induced by human activities. S2. Based on the geographic coordinate system, the study area is divided into three-dimensional grid units of equal size. Each unit is dynamically bound to geological attribute data, and a heterogeneous geological structure model containing rock layer dip angle, shear strength and fracture density is generated by three-dimensional geological modeling software. The size of the three-dimensional grid unit is adaptively adjusted according to the complexity of the geological structure. A 1-meter high-resolution grid is used in the fault zone and karst area, and a 5-10 meter grid is used in the homogeneous rock layer area. The grid division is dynamically coupled with the groundwater flow field simulation results. S3. Perform feature extraction and fusion on the three-dimensional mesh unit, extract the interaction features of the main control factors and the inducing factors, wherein the main control factors include groundwater level gradient, fault activity and lithological permeability coefficient, and the inducing factors include rainfall accumulation effect, construction disturbance intensity and seismic wave propagation attenuation coefficient, and optimize the input feature set using a dynamic weighting method based on feature importance. S4. Construct a groundwater inrush risk assessment model based on the XGBoost algorithm, specifically including the following steps: S41. Define the loss function as a binary logical loss function: Furthermore, a second-order Taylor expansion is introduced to approximate the gradient of the loss function in order to accelerate gradient calculation. S42. Use tree structure regularization term: ， in, This is the minimum gain threshold for leaf node splitting, with a value ranging from 0.1 to 1.
0. This is the L2 regularization coefficient, with a value ranging from 0.01 to 0.5; This represents the total number of leaf nodes in a single tree. For the first The weight values of each leaf node are combined with early stopping and Bayesian hyperparameter tuning to control model complexity. S43. Quantify the contribution of each feature to the surge risk based on the SHAP value, and dynamically adjust the feature weights; S5. Input real-time data into the trained model, output the surge probability value of each grid cell, use an adaptive threshold segmentation algorithm to remove isolated high-risk cells based on spatial continuity constraints, and perform spatial autocorrelation analysis based on Moran's index to divide contiguous danger zones. S6. Interactive Decision Support: Through the 3D Geographic Information System (3D GIS) platform, the risk heat map and geological profile overlay view are dynamically rendered, and a real-time early warning module is integrated. When the volume of the dangerous area exceeds the preset safety threshold, a multi-level response mechanism is triggered.
2. The method according to claim 1, characterized in that, In step S3, the dynamic weighting method specifically includes: The contribution of each feature to the prediction of surge risk is calculated using the feature split gain value Gain generated during the training of the XGBoost model. The gain value is defined as the total gain of the feature when it is a split node in all trees. The contribution rate is the percentage of each feature gain value to the total gain value, calculated using the following formula: Where F is the total number of features, features with a contribution rate higher than the preset threshold are assigned a weight of 1.2-1.5 times, and redundant features are randomly discarded using Dropout.
3. The method according to claim 1, characterized in that, In step S4, a spatiotemporal cross-validation strategy is adopted during model training: historical data is divided into training and validation sets according to time series, and data in the same spatial grid cell are strictly isolated during the training and validation phases to avoid overfitting caused by data leakage.
4. The method according to claim 1, characterized in that, In step S5, the adaptive threshold segmentation algorithm is as follows: High-risk clusters are identified based on local spatial statistics, and those that meet the following criteria are considered. Clustering is performed on the grid cells, where, For the first Risk value of each grid cell, The mean risk value of all grid cells within the study area. The sample variance of the risk value. The significance level threshold, This is the spatial weight matrix.
5. The method according to claim 1, characterized in that, In step S6, the 3D GIS platform supports multi-dimensional comparative analysis functions, including: The real-time risk prediction results are overlaid and compared with the spatial distribution of historical surge events; Based on the virtual borehole function, the correlation between geological attributes and risk values of any profile can be displayed; Generate a time-series animation of risk evolution to simulate the sudden surge's diffusion path and impact range.
6. The method according to claim 1, characterized in that, It also includes model iterative optimization steps: Using an online learning framework, sensor monitoring data is received in real time. The Focal Loss function is employed to weight high-risk events in newly added samples, where: The Focal Loss function is defined as follows: ,in, The model predicts the probability of a sample. For category weight coefficients, For focusing parameters; The criteria for determining high-risk events are: grid units within 50 meters of historical sudden surge events, or units where the groundwater level change rate in real-time monitoring data exceeds 10 cm / h.