Intelligent Risk Assessment Method and System for Urban Ground Collapse Integrating Multi-Source Spatiotemporal Data
By fusing multi-source spatiotemporal data using the MS-GATBN model, the problem of capturing dynamic disaster-causing processes in urban ground collapse risk assessment was solved, achieving high-precision and interpretable risk assessment and improving the foresight and reliability of urban safety management.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHENZHEN NATURAL RESOURCES & REAL ESTATE ASSESSMENT & DEV RES CENT (SHENZHEN GEOLOGICAL ENVIRONMENT MONITORING CENT)
- Filing Date
- 2026-02-04
- Publication Date
- 2026-05-26
AI Technical Summary
Existing technologies are insufficient to effectively capture dynamic disaster-causing processes in urban ground collapse risk assessment. They lack the ability to fuse multi-source spatiotemporal data, and traditional models are not reliable enough in risk propagation path modeling and prediction, resulting in inaccurate assessment results and poor interpretability.
The MS-GATBN model is constructed to achieve dynamic risk assessment through multi-source spatiotemporal data processing, feature enhancement, sample construction, and intelligent modeling processes. This model includes multi-scale temporal feature learning, conditional feature fusion, anti-diffusion graph attention network, and uncertainty quantification module, fusing static and dynamic features to generate high-precision, interpretable risk assessment results.
It significantly improves the spatiotemporal accuracy and reliability of urban ground collapse risk assessment, enables real-time perception and dynamic adjustment of risk factors, breaks through the limitations of timeliness and correlation analysis of traditional methods, and supports the accurate assessment and prevention of urban ground collapse.
Smart Images
Figure CN122089062A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of urban public safety and geological risk assessment technology, specifically to an intelligent risk assessment method and system for urban ground collapse that integrates multi-source spatiotemporal data. Background Technology
[0002] With the rapid advancement of urbanization, the intensity of development and utilization of urban underground space is increasing daily. Ground collapse disasters, triggered by the coupling of multiple factors such as complex hydrogeological conditions, aging underground pipe networks, frequent construction projects, and extreme rainfall events, have become a major hidden danger threatening urban public safety and the stable operation of infrastructure. These disasters are often sudden, insidious, and cascading, and their causal mechanisms are complex, involving the spatiotemporal interaction of static geological environment and dynamic external disturbances, posing a great challenge to accurate prediction and effective prevention and control.
[0003] Currently, ground collapse risk assessment mainly relies on static sensitivity assessment methods based on geographic information systems, such as the analytic hierarchy process (AHP), frequency ratio models, and machine learning models like logistic regression and random forests. These methods typically spatially correlate collapse events with a series of static influencing factors (such as soil type, pipeline density, and distance from rivers) to generate static risk zoning maps. However, these methods have significant limitations: First, they are essentially statistical inductions of historical event distribution patterns, making it difficult to effectively characterize the dynamic evolution of risk over time, especially failing to capture the coupling effect between dynamic triggering factors such as rainfall and the static vulnerable environment; second, traditional models generally lack the ability to explicitly model the spatial propagation path of disasters along infrastructure networks, ignoring the topological characteristics of risk propagation; furthermore, these methods are insufficient in processing multi-source heterogeneous spatiotemporal data, constructing high-quality training sample sets, and quantifying prediction uncertainties, resulting in limited model generalization and decision support reliability in practical applications.
[0004] In recent years, some studies have attempted to introduce deep learning models such as convolutional neural networks and graph neural networks to capture spatiotemporal dependencies. However, when faced with complex mechanisms, sparse data, and strong geographical constraints, such as ground collapse, these studies still suffer from problems such as poor fit between model structure and physical mechanisms, insufficient analysis of multi-timescale driving factors, overly idealized spatial relationship modeling lacking geographical rationality, and difficulty in assessing the confidence level of prediction results. Therefore, there is an urgent need for a new intelligent risk assessment method that can deeply integrate multi-source spatiotemporal data, accurately characterize dynamic disaster-causing processes, and possess high reliability and interpretability to support the accurate assessment and proactive prevention of urban ground collapse. Summary of the Invention
[0005] The purpose of this invention is to address the shortcomings of existing technologies, such as insufficient modeling capabilities for dynamically coupled disaster-causing processes, inadequate utilization of spatiotemporal features, and low reliability of risk assessment. This invention provides an intelligent risk assessment method and system for urban ground collapse that integrates multi-source spatiotemporal data. By constructing an integrated data processing, feature enhancement, sample construction, and intelligent modeling workflow, this invention achieves high-precision and interpretable dynamic risk assessment.
[0006] The objective of this invention is achieved through the following technical solution: The first aspect of this invention provides a method for intelligent risk assessment of urban ground collapse that integrates multi-source spatiotemporal data, comprising the following steps: S1. Collect multi-source spatiotemporal data related to urban ground collapse from multiple heterogeneous data sources, and generate spatially aligned and time-aligned static feature sequences and dynamic feature sequence data through unified spatial normalization and standardization processing. S2. Using dynamic spatiotemporal feature enhancement methods, the original rainfall data in the dynamic feature sequence data is reconstructed by multi-time-scale indexation to extract two types of disaster-causing features: short-term intensity index and medium-term cumulative index. S3. Based on static and dynamic feature sequence data, a negative sample set with stable geological conditions and distinguishable features is generated by adopting a negative sample precise sampling strategy based on sample feature similarity. S4. By using a multi-source feature fusion method, static feature sequences and dynamic feature sequence data, short-term intensity index and medium-term cumulative index, as well as the spatial coordinate information of negative sample sets are integrated into a structured three-dimensional spatiotemporal sample dataset. S5. Construct an MS-GATBN model to achieve end-to-end collapse risk prediction; the MS-GATBN model includes a multi-scale temporal feature learning module, a conditional feature fusion module, an anti-diffusion graph attention network, and an uncertainty quantification module; S6. Train the MS-GATBN model based on the spatiotemporal sample dataset to obtain a trained MS-GATBN model; S7. Input the current spatiotemporal data of the area to be evaluated into the trained MS-GATBN model, output the final predicted probability of collapse risk of each spatial unit in the future period and its uncertainty estimate, and generate the intelligent risk assessment result of urban ground collapse.
[0007] Furthermore, the unified spatial normalization and standardization process specifically includes: For spatial data in multi-source spatiotemporal data, unified spatial projection and rasterization are used to transform them to the same coordinate system and the same spatial resolution. For dynamic time-series data in multi-source spatiotemporal data, time alignment and spatial interpolation are performed to ensure that the spatiotemporal dimensions remain consistent. For all numerical features in multi-source spatiotemporal data, Z-score standardization or max-min normalization is performed. Finally, spatially and temporally aligned static and dynamic feature sequence data are obtained.
[0008] Furthermore, the short-term intensity index is calculated using the following expression: In the formula, coordinate position In time t Short-term intensity index of rainfall, coordinate position In time t The original monthly-scale rainfall at that time This is the threshold for short-term rainfall intensity; The medium-term cumulative index is calculated using the following expression: In the formula, coordinate position In time t The medium-term cumulative index of rainfall, Indicates coordinate position In time t Cumulative rainfall in the first 6 months; This is the threshold for accumulated rainfall over a medium period.
[0009] Furthermore, the negative sample accurate sampling strategy based on sample feature similarity specifically includes the following sub-steps: S31. Random generation of candidate negative samples: Candidate negative sample points are generated within the effective spatial region based on system grid sampling; S32. Spatial Buffer Exclusion: Establish a buffer with radius R centered on each historical collapse point, and use the spatial indexing method to exclude all candidate negative sample points within the buffer; where the historical collapse point is the positive sample point. S33. Feature Space Clustering Screening: K-means clustering is performed on the static and dynamic feature sequences corresponding to the positive sample points. The random forest algorithm is used to quantify the global importance of each static and dynamic feature to calculate the weighted Euclidean distance. For each candidate negative sample point remaining after the spatial buffer is excluded, the weighted Euclidean distance to all cluster centers is calculated. Only candidate negative sample points whose distance to all cluster centers is greater than the corresponding cluster radius are retained. S34. Similarity-weighted probability selection: Calculate the cosine similarity between the candidate negative sample points selected by feature space clustering and all positive sample points, and perform stratification. Combine the time-series risk factor to dynamically adjust the sampling weight of the candidate negative sample points in each stratum. Finally, perform stratified weighted random sampling to obtain a negative sample set with stable geological conditions and distinguishable features.
[0010] Furthermore, step S4 specifically includes the following sub-steps: The static feature sequence and dynamic feature sequence data, short-term intensity index and medium-term cumulative index, and spatial coordinate information of sample points are aligned and extracted according to the sample points; the sample points include positive sample points and negative sample points in the negative sample set; For each sample point, its dynamic temporal features are constructed by extracting its dynamic feature sequence data for T consecutive time steps before the target time t, and timestamp information is added. All static feature sequence data, dynamic time-series features, short-term intensity index, medium-term cumulative index, timestamp information and spatial coordinate information are concatenated in the time-series dimension to form a complete feature vector for each sample. The final structured three-dimensional spatiotemporal sample dataset is constructed, with dimensions of sample quantity, time step and sample features. The sample features include static features, dynamic temporal features, periodic features and coordinate features. The periodic features include short-term intensity index and medium-term cumulative index.
[0011] Furthermore, in the conditional feature fusion module, a conditional vector is extracted from the static features by sequentially passing through a one-dimensional convolutional layer, a global average pooling layer, and a fully connected layer. The conditional vector dynamically outputs weights and biases on a sample-by-sample basis through independent mapping. In the multi-scale temporal feature learning module, short-term conditional causal convolutional layers, mid-term causal convolutional layers, and long-term causal convolutional layers are used in parallel to extract temporal feature representations at the monthly, semi-annual, and yearly scales from high-dimensional dynamic convolutional features. When extracting the temporal feature representation at the monthly scale, weights and bias terms are used as conditional inputs to the short-term conditional causal convolutional layer. After concatenating the temporal feature representations at the monthly, semi-annual, and yearly scales, they are then enhanced by seasonal position encoding, bidirectional long short-term memory network, and multi-head self-attention mechanism. Finally, the multi-scale temporal representation features are projected through a fully connected layer. The multi-scale temporal representation features are then fused and dimensionality-reduced with the conditional vector to obtain the dimensionality-reduced fused features. In the anti-diffusion graph attention network, coordinate features and dimensionality-reduced fused features are processed by projected distance. Then, an anti-diffusion adjacency matrix that integrates geographical constraints and feature similarity is constructed by introducing a spatial constraint gating mechanism. A spatial association weight matrix is generated by dynamic weight optimization of graph attention. The dimensionality-reduced fused features are used as input to the first layer of the graph attention network. The spatial association weight matrix is used as a differentiable neighborhood information aggregation weight through two layers of graph attention network to realize graph information aggregation and iterative optimization of neighborhood risk features. The residual connection preserves the uniqueness of nodes and outputs deep abstract features. In the uncertainty quantification module, the deep abstract features are calibrated by global pooling, output optimization and temperature scaling to obtain the temperature-scaled calibrated deep features; uncertainty weights are obtained from the temperature-scaled calibrated deep features by uncertainty gating vector; and the uncertainty weights are used to obtain the final prediction probability through the hybrid probability output layer to obtain the prediction result of urban road surface collapse risk.
[0012] Further, step S6 specifically includes: The spatiotemporal sample dataset is divided into training set, validation set and test set according to the time sequence and proportions. In the model parameter learning phase, the MS-GATBN model is trained on the training set. Static features, high-dimensional dynamic convolutional features, and coordinate features from the training set are input into the MS-GATBN model to obtain the final predicted probability of collapse risk. Based on the final predicted probability of collapse risk and the corresponding true label, the binary cross-entropy loss function is calculated. With minimizing the binary cross-entropy loss function as the optimization objective, the Adam optimizer is used to iteratively update the internal parameters of the MS-GATBN model. During the hyperparameter tuning phase, the MS-GATBN model is evaluated on the validation set, and the model performance is comprehensively evaluated using precision, recall, F1 score, and AUC. The hyperparameters of the MS-GATBN model are adjusted based on the evaluation results of the validation set, and the adjusted hyperparameters are fed back into the training process to update the model. Repeat the above training and validation process, continuously optimize hyperparameters based on the model performance evaluation results of the validation set and the early stopping mechanism until the best-performing MS-GATBN model is obtained; evaluate the performance of the best-performing MS-GATBN model on the test set, and obtain an unbiased estimate of the model's generalization performance based on precision, recall, F1 score and AUC, thus obtaining the trained MS-GATBN model.
[0013] Furthermore, the high-dimensional dynamic convolutional features are obtained through the following process: After seasonal location encoding of periodic features, monthly scale encoding expansion is performed, and then features are concatenated together with standardized dynamic temporal features. Finally, high-dimensional dynamic convolutional features are obtained through dynamic convolution.
[0014] Further, step S7 specifically includes: The current spatiotemporal data of the area to be evaluated is input into the trained MS-GATBN model. Combined with the uncertainty quantification module, a raster map of the probability index of ground subsidence susceptibility in the area is generated. Each raster value represents the probability of subsidence risk in the spatial unit in a specific future period. Then, based on the overall distribution of data throughout the year within the road mask area, the susceptibility results of the area are divided into multiple levels using quantiles as a unified threshold, forming ground subsidence risk zones along the road. Finally, a dynamic risk assessment level map at the urban road network level with high confidence is generated, resulting in an intelligent risk assessment of urban ground subsidence.
[0015] A second aspect of this invention provides a system for implementing the above-described intelligent risk assessment method for urban ground collapse that integrates multi-source spatiotemporal data, comprising: The multi-source data acquisition and preprocessing module is used to collect multi-source spatiotemporal data related to urban ground collapse from multiple heterogeneous data sources, and generate spatially aligned and time-aligned static feature sequence and dynamic feature sequence data through unified spatial normalization and standardization processing. The dynamic feature enhancement module is used to perform multi-time-scale indexation reconstruction of the original rainfall data in dynamic feature sequence data using dynamic spatiotemporal feature enhancement methods, and to extract two types of disaster-causing features: short-term intensity index and medium-term cumulative index. The negative sample sampling module is used to generate a set of negative samples with stable geological conditions and distinguishable features based on a precise negative sample sampling strategy based on sample feature similarity, using static feature sequence and dynamic feature sequence data. The feature fusion and dataset construction module is used to integrate static feature sequences and dynamic feature sequence data, short-term intensity index and medium-term cumulative index, and spatial coordinate information of negative sample sets into a structured three-dimensional spatiotemporal sample dataset through multi-source feature fusion methods. The MS-GATBN model construction module is used to construct the MS-GATBN model to achieve end-to-end collapse risk prediction; the MS-GATBN model includes a multi-scale temporal feature learning module, a conditional feature fusion module, an anti-diffusion graph attention network, and an uncertainty quantification module. The model training and evaluation module is used to train the MS-GATBN model based on the spatiotemporal sample dataset to obtain the trained MS-GATBN model; The risk assessment and output module is used to generate the final predicted probability and uncertainty estimate of the collapse risk of each spatial unit in the area to be assessed in the future time period based on the trained MS-GATBN model, and generate intelligent risk assessment results for urban ground collapse.
[0016] The beneficial effects of this invention are as follows: Firstly, through a dynamic risk assessment mechanism, this invention significantly improves the accuracy and foresight of urban safety risk management. Secondly, by integrating multi-source spatiotemporal data, it achieves continuous perception and dynamic weight adjustment of risk factors, capturing real-time changes and interactions of risk elements, thus overcoming the limitations of traditional static assessments in terms of timeliness and correlation analysis. Thirdly, the technical path of data fusion-intelligent judgment-confidence assessment established by this invention transforms risk identification from single-point and lagging to global and forward-looking approaches, assisting managers in achieving early warning and intervention for abnormal risks. Fourthly, by integrating multi-source spatiotemporal data with multi-scale temporal analysis and an artificial intelligence model driven by anti-diffusion spatial constraint mechanisms, this invention overcomes the bottleneck of traditional methods in spatiotemporal dynamic coupling modeling, significantly improving the spatiotemporal accuracy, mechanism interpretability, and prediction reliability of risk assessment. Fifthly, by optimizing the risk identification mechanism and efficiency, this invention enhances the ability of multi-risk coupling analysis, providing technical support for the transformation of urban ground collapse management from "human-based" to "intelligent-based" methods. Attached Figure Description
[0017] Figure 1 This is a flowchart of an intelligent risk assessment method for urban ground collapse that integrates multi-source spatiotemporal data; Figure 2 This is the architecture flowchart of the MS-GATBN model. Detailed Implementation
[0018] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention as detailed in the appended claims.
[0019] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The singular forms “a,” “the,” and “the” used in this invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.
[0020] It should be understood that although the terms first, second, third, etc., may be used in this invention to describe various information, this information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first information may also be referred to as second information without departing from the scope of this invention, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."
[0021] The present invention will now be described in detail with reference to the accompanying drawings. Unless otherwise specified, the features of the following embodiments and implementations can be combined with each other.
[0022] See Figure 1 The intelligent risk assessment method for urban ground collapse that integrates multi-source spatiotemporal data of the present invention specifically includes the following steps: S1. Acquisition and unified normalization of multi-source spatiotemporal data: Collect multi-source spatiotemporal data related to urban ground collapse from multiple heterogeneous data sources, and generate spatially aligned and time-aligned static feature sequences and dynamic feature sequences through unified spatial normalization and standardization processing.
[0023] Furthermore, data sources include geological survey databases, underground pipeline survey data, municipal engineering construction records, and meteorological observation network data. Multi-source spatiotemporal data includes urban geological conditions, pipeline defects, road networks, engineering construction records, and rainfall observation data.
[0024] Furthermore, through unified spatial normalization and standardization processing, spatially and temporally aligned static and dynamic feature sequence data are generated. Specifically, for spatial data in multi-source spatiotemporal data, unified spatial projection and rasterization processing are used to transform them to the same coordinate system and the same spatial resolution. For dynamic time-series data in multi-source spatiotemporal data, such as monthly rainfall data, temporal alignment and spatial interpolation are performed to ensure consistency in spatiotemporal dimensions. For all numerical features in multi-source spatiotemporal data, Z-score standardization or min-max normalization processing is performed to eliminate the influence of dimensions, providing high-quality, standardized input data for subsequent model training, i.e., spatially and temporally aligned static and dynamic feature sequence data.
[0025] S2. Enhancement and Engineering of Dynamic Spatiotemporal Features: To achieve a more accurate characterization of the disaster-causing mechanism of rainfall, the dynamic spatiotemporal feature enhancement method is used to reconstruct the original rainfall data in the dynamic feature sequence data obtained in step S1 using multi-time-scale indexation, and to extract two types of disaster-causing driving features: short-term intensity index and medium-term cumulative index.
[0026] Furthermore, the dynamic spatiotemporal feature enhancement method specifically includes: processing the original monthly-scale rainfall data in the dynamic feature sequence data obtained in step S1. Perform multi-timescale indexation reconstruction and generate a short-term intensity index. With medium-term cumulative index The calculation of the short-term intensity index. It is calculated using the following expression: In the formula, coordinate position In time t The original monthly-scale rainfall at that time; The short-term rainfall intensity threshold is determined based on the statistical quantiles (e.g., 95th percentile) of historical extreme rainfall events in the region. It is used to distinguish between ordinary rainfall and heavy rainfall events and to highlight the instantaneous impact of heavy rainfall events. coordinate position In time t The short-term intensity index of rainfall is used to identify whether the monthly rainfall exceeds the short-term rainfall intensity threshold. This indicates the short-term triggering effect of rainfall in the current month, which helps the model learn and identify short-term rainfall intensity states more effectively, especially identifying those areas with short-term heavy rainfall that require high attention.
[0027] Among them, the medium-term cumulative index It is calculated using the following expression: In the formula, Indicates coordinate position In time t Cumulative rainfall in the first 6 months; The threshold for accumulated medium-term rainfall was set based on the analysis of regional soil saturation characteristics and historical landslide data. coordinate position In time t The medium-term cumulative rainfall index is used to identify whether the cumulative rainfall over the previous 6 months exceeds the medium-term cumulative rainfall threshold. The short-term intensity index and the medium-term accumulation index together reflect the sustained impact of accumulated rainfall on soil saturation, facilitating the model to learn and identify the effects of accumulated rainfall more effectively, especially identifying areas affected by accumulated rainfall that require high attention. These two indices together constitute an enhanced feature reflecting the multi-scale disaster-driving capacity of rainfall.
[0028] It should be understood that, since the data obtained in step S1 contains categorical rasters (such as geological types), the 0-9 integer encoding must be retained. Other continuous features are also mapped to 0-9 for a unified standard. Therefore, in hour, ;exist hour, Similarly, when hour, ;when hour, .
[0029] S3. Negative sample precise sampling strategy based on sample feature similarity: In order to construct a high-quality training sample set, based on the static feature sequence and dynamic feature sequence data obtained in step S1, a negative sample precise sampling strategy based on sample feature similarity is adopted to generate a negative sample set with stable geological conditions and distinguishable features.
[0030] Furthermore, the precise negative sample sampling strategy based on sample feature similarity generates negative samples by employing a hierarchical weighted feature space sampling method, specifically including the following sub-steps: S31. Random generation of candidate negative samples: A large number of candidate negative sample points are generated based on system grid sampling within the effective spatial area.
[0031] S32. Spatial Buffer Exclusion: A buffer zone with a radius of R (in this embodiment, R is set to 500 meters) is established with each historical collapse point (i.e., positive sample point) as the center. The spatial indexing method is used to efficiently exclude all candidate negative sample points in the buffer zone to avoid sample confusion caused by spatial autocorrelation and ensure that the negative samples collected can truly represent stable areas that are not prone to collapse.
[0032] S33. Feature Space Clustering and Screening: First, K-means clustering is performed on the static and dynamic feature sequences corresponding to the positive sample points. Collapse points with similar disaster-causing backgrounds are grouped into one category, resulting in several clusters. Then, a random forest algorithm is used to train on the clustering results, quantifying the global importance of each static and dynamic feature, and calculating the weighted Euclidean distance accordingly. Finally, for each candidate negative sample point remaining after spatial buffer exclusion, its weighted Euclidean distance to all cluster centers is calculated. Only candidate negative sample points whose distance to all cluster centers is greater than the corresponding cluster radius are retained, i.e., candidate negative sample points far from the cluster center regions of the positive sample points are retained, ensuring that negative samples also deviate from the typical patterns of positive samples in the feature space.
[0033] In this process, when training the Random Forest algorithm on the clustering results, cluster labels are used as supervisory information to train the Random Forest classification model. During training, all static and dynamic features corresponding to each positive sample are used as input features, and the Random Forest classification model learns the feature boundaries that distinguish different clusters. After the Random Forest classification model is trained, the feature importance score of each static and dynamic feature to the clustering results is quantified through a built-in feature importance evaluation mechanism (Gini impurity reduction), i.e., global importance. This reflects the discriminative ability of each feature in distinguishing different clusters; the higher the score, the more important the feature's discriminative role in the feature space. Subsequently, a weighted Euclidean distance is constructed using the feature importance scores. Specifically, each feature dimension is assigned its corresponding importance weight when calculating the distance. For any sample point, the weighted Euclidean distance to a certain cluster center is calculated as follows: first, the squared difference between the sample and the cluster center on each feature dimension is calculated, then multiplied by the corresponding feature importance weight, and finally, the summation and square root of all weighted difference squares are taken. This calculation process emphasizes the influence of highly important feature dimensions in the distance metric, making the distance more reflective of substantial differences in the feature space.
[0034] S34. Similarity-Weighted Probability Selection: Calculate the cosine similarity between candidate negative sample points selected through feature space clustering and all positive sample points. Based on the cosine similarity, divide the candidate negative sample points into three levels: low, medium, and high. The cosine similarity for the low level is less than 0.3, for the medium level it is 0.3-0.7, and for the high level it is greater than 0.7. Combined with temporal risk factors (such as monthly rainfall intensity), dynamically adjust the sampling weights of candidate negative sample points within each level. Finally, perform stratified weighted random sampling to obtain a geologically stable and feature-distinguishable negative sample set, ensuring that the negative sample set is representative in spatial distribution, feature composition, and temporal context, and is distinguishable from the positive samples.
[0035] S4. Multi-source feature fusion and sample dataset construction: Through the multi-source feature fusion method, the static feature sequence and dynamic feature sequence data obtained in step S1, the dynamic enhancement features (i.e., short-term intensity index and medium-term cumulative index) obtained in step S2, and the spatial coordinate information of the negative sample set in step S3 are integrated into a structured three-dimensional spatiotemporal sample dataset.
[0036] Specifically, firstly, the static feature sequences (such as adverse geological distribution and pipeline density) obtained after step S1 are aligned and extracted with the dynamic feature sequence data (such as the original monthly-scale rainfall time series data), the short-term intensity index and medium-term cumulative index obtained in step S2, and the spatial coordinate information of the sample points. The sample points include positive sample points and negative sample points from the negative sample set generated in step S3. Then, for each sample point, its dynamic temporal features are constructed by extracting its dynamic feature sequence data for T consecutive time steps before the target time t, and timestamp information is added. Subsequently, all static feature sequence data, dynamic temporal features, short-term intensity index, medium-term cumulative index, timestamp information, and spatial coordinate information are concatenated along the temporal dimension to form a complete feature vector for each sample. Finally, a structured three-dimensional spatiotemporal sample dataset is constructed, with dimensions of sample quantity, time step, and sample features, providing a directly input data structure for subsequent deep learning models.
[0037] Furthermore, in the spatiotemporal sample dataset, the sample features include static features, dynamic temporal features, periodic features, and coordinate features. The periodic features include short-term intensity index and medium-term cumulative index.
[0038] S5, MS-GATBN Model Architecture Design and Functional Implementation: The MS-GATBN model is constructed to achieve end-to-end collapse risk prediction. Its structure is as follows: Figure 2 As shown, the MS-GATBN model includes a multi-scale temporal feature learning module, a conditional feature fusion module, an anti-diffusion graph attention network, and an uncertainty quantization module.
[0039] In this embodiment, the conditional feature fusion module extracts conditional vectors from static features by sequentially passing through a one-dimensional convolutional layer, a global average pooling layer, and a fully connected layer, which can be represented as: In the formula, Represents a condition vector. Indicates static characteristics, This represents a one-dimensional convolutional layer. This indicates a global average pooling layer. This represents a fully connected layer. The conditional vector dynamically outputs the weights of a one-dimensional causal convolution (i.e., a short-run conditional causal convolution layer) with a kernel size of 1 on a sample-by-sample basis through independent mapping. and bias terms .
[0040] In this embodiment, in the multi-scale temporal feature learning module, one-dimensional causal convolutions with kernel sizes of 1, 6, and 12 are used as short-term conditional causal convolutional layers, medium-term causal convolutional layers, and long-term causal convolutional layers, respectively. These layers extract temporal feature representations at monthly, semi-annual, and yearly scales in parallel from the high-dimensional dynamic convolutional features. Specifically, when extracting temporal feature representations at the monthly scale, a conditional feature fusion module integrates static features with high-dimensional dynamic convolutional features. Conditional causal convolution is used to learn and identify short-term triggering patterns of accidents. Weights are first obtained through the conditional feature fusion module. and bias terms Then through weights and bias terms High-dimensional dynamic convolutional features are modulated and activated using the Swish activation function to generate short-term temporal representations, i.e., monthly-scale temporal feature representations. This enables the MS-GATBN model to adapt its feature extraction of short-term rainfall responses to the static environmental features of the samples, as can be expressed by the following formula: In the formula, Representation of temporal features at a monthly scale; and These represent the weights and biases output by the conditional feature fusion module, respectively. Represents high-dimensional dynamic convolutional features; This represents a one-dimensional causal convolution with a kernel size of 1, i.e., a short-term conditional causal convolution layer; This represents the Swish activation function, used to perform a nonlinear transformation on the output of the causal convolution. The temporal feature representations at the six-month and yearly scales only combine dynamic temporal features, utilizing standard causal convolution to learn and identify the medium- to long-term temporal evolution patterns of accidents. The calculation process is implemented through the following expression: In the formula, Indicated on the time scale k The following temporal feature representation, k This indicates the kernel size corresponding to different time scales. k=6,12. The temporal feature representations at monthly, semi-annual, and yearly scales are concatenated, then sequentially enhanced through seasonal location encoding, a bidirectional Long Short-Term Memory (LSTM) network, and a multi-head self-attention mechanism. Finally, the multi-scale temporal representation features are projected through a fully connected layer to obtain the input features. Subsequently, the multi-scale temporal representation features and conditional vectors are fused and dimensionality reduced to obtain dimensionality-reduced fused features, which serve as input for subsequent modules. This design allows the conditional convolutional kernels at short-term scales to adapt to the static environment of each sample, thereby enhancing the sensitivity to short-term triggering factors. Convolutional operations at medium and long-term scales employ standard causal convolutions without conditional constraints, focusing on capturing broader temporal trends. This hybrid strategy achieves a balance between adaptability and parameter efficiency.
[0041] In this embodiment, in the anti-diffusion graph attention network, firstly, the coordinate features and the dimensionality-reduced fused features are processed by projected distance. Then, an anti-diffusion adjacency matrix that integrates geographical constraints and feature similarity is constructed by introducing a spatial constraint gating mechanism. Finally, a spatial association weight matrix is generated through graph attention dynamic weight optimization. The spatial constraint gating mechanism is implemented through the following expression: In the formula, This represents a spatially constrained mask, where r is the geographical distance between sampling points. To define the critical radius of the core interaction area. It is a sparsity threshold that controls the interaction of samples over long distances. This indicates that the value specified in parentheses is the larger of the two values. For trainable sparsity factors, For the maximum spatial correlation distance, Temperature parameters used to adjust boundary sharpness. Critical radius of the core interaction region. A core influence zone was defined, within which neighboring samples ensure strong correlation and maintain the inherent consistency of the spatial structure; simultaneously, a trainable sparse threshold was established. It can dynamically expand the interaction boundary, enabling the MS-GATBN model to adaptively identify and capture valid long-distance connections based on the data; spatial constraint mask. and sparsity threshold The synergistic effect of these elements constitutes a smooth, differentiable spatial filter. This filter, on the one hand, restricts risk propagation within a reasonable range of physical reality, and on the other hand, maintains a stable gradient flow during training. Further, the spatial constraint mask... Compared with feature similarity-based matrices Combined with risk adaptive items The synergistic effect prevents the model from over-relying on local information, forming the optimized final adjacency matrix, i.e., the spatial correlation weight matrix, which can be expressed as: In the formula, This represents the optimized final adjacency matrix, i.e., the spatial correlation weight matrix; This is a matrix constructed based on feature similarity. For a time causal matrix constructed based on time series causal relationships, future information leakage is prevented by restricting each time step to only focus on the current and past time steps. This indicates that the calculated association strength is standardized to ensure the smooth and efficient training of subsequent networks. The spatial association weight matrix integrates spatial constraints and feature perception capabilities, serving as the optimized graph structure input to subsequent Graph Attention Network (GAT) layers, ensuring that information aggregation is geographically reasonable. Using the generated spatial association weight matrix as differentiable neighborhood information aggregation weights, two GAT layers are used to achieve graph information aggregation and iterative optimization of neighborhood risk features. Residual connections preserve node uniqueness, outputting deep abstract features. Specifically, using the spatial association weight matrix as the base graph structure, it optimizes the graph structure by eliminating geographically unreasonable connections before the standard attention mechanism operates. Therefore, the attention weights of the GAT layer are only calculated within the selected meaningful neighborhoods. This mechanism ensures that the information aggregation of the GAT layer is based on real-world geographical features, while the inherent attention mechanism of GAT further optimizes the importance weights of these candidate neighbors based on feature compatibility.
[0042] In this embodiment, a two-stage uncertainty calibration mechanism is integrated into the uncertainty quantification module. This mechanism synergistically combines temperature scaling and uncertainty gating probability fusion to provide reliable confidence assessment results for risk-sensitive predictions. Its core is the deep abstract features ultimately output by the adversarial diffusion graph attention network through global pooling, output optimization, and temperature scaling. A calibration process is performed to calibrate the entropy value of the output distribution. This calibration process is achieved through the following expression: In the formula, This represents the deep features after temperature scaling calibration. T The temperature parameter is trainable and adjusts the confidence output of the Softmax function by scaling the magnitude of deep features without changing the predicted class. Higher temperatures ( T >1) will produce a smoother probability distribution (with higher uncertainty), while lower temperatures ( T A distribution with a value less than 1 will produce a sharper (higher confidence) distribution. Subsequently, a 32-dimensional uncertainty-gated vector is used. Deep features after temperature scaling calibration By capturing uncertainty signals, the uncertainty weights of the MS-GATBN model output probability are obtained. The calculation formula is as follows: In the formula, Represents the gating vector for uncertainty The arithmetic mean of all elements is calculated. Finally, the final predicted probability is obtained through a hybrid probability output layer; that is, the final predicted probability of the MS-GATBN model is the same as the original output probability of the MS-GATBN model. With uniform prior distribution probability The weighted sum of is expressed as: In the formula, The final predicted probability is the output of the mixed probability. The original output probability before uncertainty quantification. The probability is a discrete uniform prior distribution. for and The convex combination of the model yields the prediction results for urban road surface collapse risk based on the final predicted probability. This mechanism effectively calibrates the confidence level of the MS-GATBN model output, resulting in more confident output results.
[0043] S6. Training and performance evaluation of the MS-GATBN model: The MS-GATBN model is trained based on the spatiotemporal sample dataset. At the same time, the stability, hyperparameter sensitivity, classification accuracy and generalization ability of the MS-GATBN model are evaluated to obtain a well-trained MS-GATBN model.
[0044] Specifically, the MS-GATBN model is trained and evaluated end-to-end based on a spatiotemporal sample dataset. First, the spatiotemporal sample dataset is divided into training, validation, and test sets in chronological order, with a sample size ratio of 8:1:1. Then, during the model parameter learning phase, the MS-GATBN model is trained on the training set. Static features, high-dimensional dynamic convolutional features, and coordinate features from the training set are input into the MS-GATBN model. The final predicted probability of collapse risk is obtained through its conditional feature fusion module, multi-scale temporal feature learning module, anti-diffusion graph attention network, and uncertainty quantification module. Based on the final predicted probability of collapse risk and the corresponding true labels, a binary cross-entropy loss function is calculated. Minimizing the binary cross-entropy loss function is the optimization objective, and the Adam optimizer is used to iteratively update the internal parameters of the MS-GATBN model. Subsequently, in the hyperparameter tuning phase, the MS-GATBN model is evaluated on the validation set, and its performance is comprehensively assessed using metrics such as precision, recall, F1 score, and AUC (area under the ROC curve). Based on the validation set evaluation results, the hyperparameters of the MS-GATBN model are adjusted, and the adjusted hyperparameters are fed back into the training process to update the model. This training and validation process is repeated, continuously optimizing the hyperparameters based on the model performance evaluation results from the validation set and the early stopping mechanism until the optimal-performing MS-GATBN model is obtained. The optimal-performing MS-GATBN model is then evaluated on the test set, and an unbiased estimate of the model's generalization performance is obtained based on metrics such as precision, recall, F1 score, and AUC, resulting in the trained MS-GATBN model.
[0045] Furthermore, high-dimensional dynamic convolutional features are obtained through the following process: seasonal location encoding of periodic features is performed, followed by monthly scale encoding expansion, and then feature concatenation is performed together with standardized dynamic temporal features. Finally, high-dimensional dynamic convolutional features are obtained through dynamic convolution.
[0046] S7. Output of intelligent risk assessment results based on the trained MS-GATBN model: Input the current spatiotemporal data of the area to be assessed into the trained MS-GATBN model, output the final predicted probability of collapse risk of each spatial unit in the future period and its uncertainty estimate, and generate intelligent risk assessment results for urban ground collapse.
[0047] Specifically, the current spatiotemporal data of the area to be assessed is input into the trained MS-GATBN model. Combined with the uncertainty quantification module, a raster map of the area's ground subsidence susceptibility probability index is generated. Each raster value represents the probability of subsidence risk for that spatial unit within a specific future time period, ranging from 0 to 1. Subsequently, based on the overall distribution of data throughout the year within the road mask area, using quantiles as a unified threshold, the susceptibility result of the area (i.e., the output of the MS-GATBN model) is divided into multiple levels (e.g., 5 levels), forming ground subsidence risk zoning along the road. Finally, a dynamic risk assessment level map at the urban road network level with high confidence is generated. This map can be used for early dynamic risk assessment of urban ground subsidence, and further used to guide municipal inspections, emergency resource allocation, and intelligent risk assessment of urban ground subsidence.
[0048] It is worth mentioning that the present invention also provides a system for implementing the above-mentioned intelligent risk assessment method for urban ground collapse that integrates multi-source spatiotemporal data. The system includes a multi-source data acquisition and preprocessing module, a dynamic feature enhancement module, a negative sample sampling module, a feature fusion and dataset construction module, an MS-GATBN model construction module, a model training and evaluation module, and a risk assessment and output module.
[0049] In this embodiment, the multi-source data acquisition and preprocessing module is used to collect multi-source spatiotemporal data related to urban ground collapse from multiple heterogeneous data sources, and generate spatially aligned and time-aligned static feature sequence data and dynamic feature sequence data through unified spatial normalization and standardization processing.
[0050] In this embodiment, the dynamic feature enhancement module is used to perform multi-time-scale indexation reconstruction of the original rainfall data in the dynamic feature sequence data using the dynamic spatiotemporal feature enhancement method, and extract two types of disaster-driving features: short-term intensity index and medium-term cumulative index.
[0051] In this embodiment, the negative sample sampling module is used to generate a set of negative samples with stable geological conditions and distinguishable features based on static feature sequence and dynamic feature sequence data and a negative sample precision sampling strategy based on sample feature similarity.
[0052] In this embodiment, the feature fusion and dataset construction module is used to integrate static feature sequences and dynamic feature sequence data, short-term intensity index and medium-term cumulative index, and spatial coordinate information of negative sample sets into a structured three-dimensional spatiotemporal sample dataset through a multi-source feature fusion method.
[0053] In this embodiment, the MS-GATBN model building module is used to build the MS-GATBN model to achieve end-to-end collapse risk prediction; the MS-GATBN model includes a multi-scale temporal feature learning module, a conditional feature fusion module, an anti-diffusion graph attention network, and an uncertainty quantification module.
[0054] In this embodiment, the model training and evaluation module is used to train the MS-GATBN model based on the spatiotemporal sample dataset to obtain the trained MS-GATBN model.
[0055] In this embodiment, the risk assessment and output module is used to generate the final predicted probability and uncertainty estimate of the collapse risk of each spatial unit in the area to be assessed in the future time period based on the trained MS-GATBN model, and generate the intelligent risk assessment result of urban ground collapse.
[0056] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for intelligent risk assessment of urban ground collapse that integrates multi-source spatiotemporal data, characterized in that, Includes the following steps: S1. Collect multi-source spatiotemporal data related to urban ground collapse from multiple heterogeneous data sources, and generate spatially aligned and time-aligned static feature sequences and dynamic feature sequence data through unified spatial normalization and standardization processing. S2. Using dynamic spatiotemporal feature enhancement methods, the original rainfall data in the dynamic feature sequence data is reconstructed by multi-time-scale indexation to extract two types of disaster-causing features: short-term intensity index and medium-term cumulative index. S3. Based on static and dynamic feature sequence data, a negative sample set with stable geological conditions and distinguishable features is generated by adopting a negative sample precise sampling strategy based on sample feature similarity. S4. By using a multi-source feature fusion method, static feature sequences and dynamic feature sequence data, short-term intensity index and medium-term cumulative index, as well as the spatial coordinate information of negative sample sets are integrated into a structured three-dimensional spatiotemporal sample dataset. S5. Construct an MS-GATBN model to achieve end-to-end collapse risk prediction; the MS-GATBN model includes a multi-scale temporal feature learning module, a conditional feature fusion module, an anti-diffusion graph attention network, and an uncertainty quantification module; S6. Train the MS-GATBN model based on the spatiotemporal sample dataset to obtain a trained MS-GATBN model; S7. Input the current spatiotemporal data of the area to be evaluated into the trained MS-GATBN model, output the final predicted probability of collapse risk of each spatial unit in the future period and its uncertainty estimate, and generate the intelligent risk assessment result of urban ground collapse.
2. The intelligent risk assessment method for urban ground collapse that integrates multi-source spatiotemporal data according to claim 1, characterized in that, The unified spatial normalization and standardization process specifically includes: For spatial data in multi-source spatiotemporal data, unified spatial projection and rasterization are used to transform them to the same coordinate system and the same spatial resolution. For dynamic time-series data in multi-source spatiotemporal data, time alignment and spatial interpolation are performed to ensure that the spatiotemporal dimensions remain consistent. For all numerical features in multi-source spatiotemporal data, Z-score standardization or max-min normalization is performed. Finally, spatially and temporally aligned static and dynamic feature sequence data are obtained.
3. The intelligent risk assessment method for urban ground collapse that integrates multi-source spatiotemporal data according to claim 1, characterized in that, The short-term intensity index is calculated using the following expression: In the formula, coordinate position In time t Short-term intensity index of rainfall, coordinate position In time t The original monthly-scale rainfall at that time This is the threshold for short-term rainfall intensity; The medium-term cumulative index is calculated using the following expression: In the formula, coordinate position In time t The medium-term cumulative index of rainfall, Indicates coordinate position In time t Cumulative rainfall in the first 6 months; This is the threshold for accumulated rainfall over a medium period.
4. The intelligent risk assessment method for urban ground collapse that integrates multi-source spatiotemporal data according to claim 1, characterized in that, The precise negative sample sampling strategy based on sample feature similarity specifically includes the following sub-steps: S31. Random generation of candidate negative samples: Candidate negative sample points are generated within the effective spatial region based on system grid sampling; S32. Spatial Buffer Exclusion: Establish a buffer with radius R centered on each historical collapse point, and use the spatial indexing method to exclude all candidate negative sample points within the buffer; where the historical collapse point is the positive sample point. S33. Feature Space Clustering Screening: K-means clustering is performed on the static and dynamic feature sequences corresponding to the positive sample points. The random forest algorithm is used to quantify the global importance of each static and dynamic feature to calculate the weighted Euclidean distance. For each candidate negative sample point remaining after the spatial buffer is excluded, the weighted Euclidean distance to all cluster centers is calculated. Only candidate negative sample points whose distance to all cluster centers is greater than the corresponding cluster radius are retained. S34. Similarity-weighted probability selection: Calculate the cosine similarity between the candidate negative sample points selected by feature space clustering and all positive sample points, and perform stratification. Combine the time-series risk factor to dynamically adjust the sampling weight of the candidate negative sample points in each stratum. Finally, perform stratified weighted random sampling to obtain a negative sample set with stable geological conditions and distinguishable features.
5. The intelligent risk assessment method for urban ground collapse that integrates multi-source spatiotemporal data according to claim 1, characterized in that, Step S4 specifically includes the following sub-steps: The static feature sequence and dynamic feature sequence data, short-term intensity index and medium-term cumulative index, and spatial coordinate information of sample points are aligned and extracted according to the sample points; the sample points include positive sample points and negative sample points in the negative sample set; For each sample point, its dynamic temporal features are constructed by extracting its dynamic feature sequence data for T consecutive time steps before the target time t, and timestamp information is added. All static feature sequence data, dynamic time-series features, short-term intensity index, medium-term cumulative index, timestamp information and spatial coordinate information are concatenated in the time-series dimension to form a complete feature vector for each sample. The final structured three-dimensional spatiotemporal sample dataset is constructed, with dimensions of sample quantity, time step and sample features. The sample features include static features, dynamic temporal features, periodic features and coordinate features. The periodic features include short-term intensity index and medium-term cumulative index.
6. The intelligent risk assessment method for urban ground collapse that integrates multi-source spatiotemporal data according to claim 1, characterized in that, In the conditional feature fusion module, a conditional vector is extracted from the static features through a one-dimensional convolutional layer, a global average pooling layer, and a fully connected layer in sequence. The conditional vector dynamically outputs weights and bias terms on a sample-by-sample basis through independent mapping. In the multi-scale temporal feature learning module, short-term conditional causal convolutional layers, mid-term causal convolutional layers, and long-term causal convolutional layers are used to extract temporal feature representations at the monthly, semi-annual, and yearly scales in parallel from high-dimensional dynamic convolutional features. When extracting the temporal feature representation at the monthly scale, weights and bias terms are used as conditional inputs to the short-term conditional causal convolutional layer. After concatenating the temporal feature representations at the monthly, semi-annual, and yearly scales, they are then enhanced by seasonal position encoding, bidirectional long short-term memory network, and multi-head self-attention mechanism. Finally, the multi-scale temporal representation features are obtained by projection through a fully connected layer. Multi-scale temporal representation features and conditional vectors are fused and dimensionality reduced to obtain dimensionality-reduced fused features. In the anti-diffusion graph attention network, coordinate features and dimensionality-reduced fused features are processed by projected distance. Then, an anti-diffusion adjacency matrix that integrates geographical constraints and feature similarity is constructed by introducing a spatial constraint gating mechanism. A spatial association weight matrix is generated by dynamic weight optimization of graph attention. The dimensionality-reduced fused features are used as input to the first layer of the graph attention network. The spatial association weight matrix is used as a differentiable neighborhood information aggregation weight through two layers of graph attention network to realize graph information aggregation and iterative optimization of neighborhood risk features. The residual connection preserves the uniqueness of nodes and outputs deep abstract features. In the uncertainty quantification module, the deep abstract features are calibrated by global pooling, output optimization and temperature scaling to obtain the temperature-scaled calibrated deep features; uncertainty weights are obtained from the temperature-scaled calibrated deep features by uncertainty gating vector; and the uncertainty weights are used to obtain the final prediction probability through the hybrid probability output layer to obtain the prediction result of urban road surface collapse risk.
7. The intelligent risk assessment method for urban ground collapse that integrates multi-source spatiotemporal data according to claim 1, characterized in that, Step S6 specifically includes: The spatiotemporal sample dataset is divided into training set, validation set and test set according to the time sequence and proportions. In the model parameter learning phase, the MS-GATBN model is trained on the training set. Static features, high-dimensional dynamic convolutional features, and coordinate features from the training set are input into the MS-GATBN model to obtain the final predicted probability of collapse risk. Based on the final predicted probability of collapse risk and the corresponding true label, the binary cross-entropy loss function is calculated. With minimizing the binary cross-entropy loss function as the optimization objective, the Adam optimizer is used to iteratively update the internal parameters of the MS-GATBN model. During the hyperparameter tuning phase, the MS-GATBN model is evaluated on the validation set, and the model performance is comprehensively evaluated using precision, recall, F1 score, and AUC. The hyperparameters of the MS-GATBN model are adjusted based on the evaluation results of the validation set, and the adjusted hyperparameters are fed back into the training process to update the model. Repeat the above training and validation process, continuously optimize hyperparameters based on the model performance evaluation results of the validation set and the early stopping mechanism until the best-performing MS-GATBN model is obtained; evaluate the performance of the best-performing MS-GATBN model on the test set, and obtain an unbiased estimate of the model's generalization performance based on precision, recall, F1 score and AUC, thus obtaining the trained MS-GATBN model.
8. The intelligent risk assessment method for urban ground collapse that integrates multi-source spatiotemporal data according to claim 6 or 7, characterized in that, The high-dimensional dynamic convolutional features are obtained through the following process: After seasonal location encoding of periodic features, monthly scale encoding expansion is performed, and then features are concatenated together with standardized dynamic temporal features. Finally, high-dimensional dynamic convolutional features are obtained through dynamic convolution.
9. The intelligent risk assessment method for urban ground collapse that integrates multi-source spatiotemporal data according to claim 1, characterized in that, Step S7 specifically includes: The current spatiotemporal data of the area to be evaluated is input into the trained MS-GATBN model. Combined with the uncertainty quantification module, a raster map of the probability index of ground subsidence susceptibility in the area is generated. Each raster value represents the probability of subsidence risk in the spatial unit in a specific future period. Then, based on the overall distribution of data throughout the year within the road mask area, the susceptibility results of the area are divided into multiple levels using quantiles as a unified threshold, forming ground subsidence risk zones along the road. Finally, a dynamic risk assessment level map at the urban road network level with high confidence is generated, resulting in an intelligent risk assessment of urban ground subsidence.
10. A system for implementing the intelligent risk assessment method for urban ground collapse that integrates multi-source spatiotemporal data as described in any one of claims 1-9, characterized in that, include: The multi-source data acquisition and preprocessing module is used to collect multi-source spatiotemporal data related to urban ground collapse from multiple heterogeneous data sources, and generate spatially aligned and time-aligned static feature sequence and dynamic feature sequence data through unified spatial normalization and standardization processing. The dynamic feature enhancement module is used to perform multi-time-scale indexation reconstruction of the original rainfall data in dynamic feature sequence data using dynamic spatiotemporal feature enhancement methods, and to extract two types of disaster-causing features: short-term intensity index and medium-term cumulative index. The negative sample sampling module is used to generate a set of negative samples with stable geological conditions and distinguishable features based on a precise negative sample sampling strategy based on sample feature similarity, using static feature sequence and dynamic feature sequence data. The feature fusion and dataset construction module is used to integrate static feature sequences and dynamic feature sequence data, short-term intensity index and medium-term cumulative index, and spatial coordinate information of negative sample sets into a structured three-dimensional spatiotemporal sample dataset through multi-source feature fusion methods. The MS-GATBN model construction module is used to construct the MS-GATBN model to achieve end-to-end collapse risk prediction; the MS-GATBN model includes a multi-scale temporal feature learning module, a conditional feature fusion module, an anti-diffusion graph attention network, and an uncertainty quantification module. The model training and evaluation module is used to train the MS-GATBN model based on the spatiotemporal sample dataset to obtain the trained MS-GATBN model; The risk assessment and output module is used to generate the final predicted probability and uncertainty estimate of the collapse risk of each spatial unit in the area to be assessed in the future time period based on the trained MS-GATBN model, and generate intelligent risk assessment results for urban ground collapse.