Intelligent early warning method and system for urban ground collapse based on multi-source factor fusion

Abstract

The invention relates to an intelligent early warning method and system for urban ground collapse based on multi-source factor fusion, and belongs to the technical field of urban disaster early warning. A PS-InSAR and an SBAS-InSAR are adopted to process and calculate deformation values respectively, deformation time sequences extracted through the two processing methods are verified and analyzed, and a settlement graph is generated through vector results which are verified to be qualified; setting a settlement rate threshold value, and carrying out preliminary ground collapse early warning identification according to deformation; dividing a dynamic factor and a static factor for the deformation time sequence and the collected multi-source data, constructing a multi-channel weighted space-time diagram structure taking a monitoring area grid unit as a node, and performing multi-source data fusion and predicting a comprehensive risk probability by using a space-time diagram neural differential attention network; and carrying out dual-channel fusion study and judgment. According to the method, the nonlinear coupling and space-time dynamic relation between disaster-inducing factors is comprehensively described in the fusion process, and high-precision, low-false-alarm and strong-generalization prediction of the urban ground collapse risk can be achieved.

Description

Technical Field

[0001] The present invention relates to an intelligent early warning method and system for urban ground collapse based on multi-source factor fusion, and in particular to a ground collapse early warning method and system based on time-series synthetic aperture radar interferometry (InSAR) and multi-source remote sensing data fusion, combined with a spatiotemporal graph neural differential attention network, belonging to the field of urban disaster early warning and artificial intelligence technology. Background Art

[0002] With the acceleration of urbanization, underground space development is intensifying. The dense distribution of subways, underground pipelines, water supply and drainage networks, and other large-scale infrastructure has made the urban geological environment more complex and fragile. Due to the combined effects of factors such as persistent or extreme rainfall, underground pipeline leakage, softening of the ground, dynamic traffic loads, and geological structures, ground subsidence and collapse events are becoming frequent in some cities, posing a serious threat to public safety, transportation operations, and infrastructure stability.

[0003] Currently, synthetic aperture radar interferometry (InSAR) technology, particularly time series methods such as persistent scatterer InSAR (PS-InSAR) and small baseline set InSAR (SBAS-InSAR), has been widely used for millimeter-scale surface deformation monitoring. However, in complex urban terrain and variable climate environments, single InSAR technology is susceptible to atmospheric residuals, phase unwrapping errors, and data loss, resulting in uncertainty in monitoring results. Furthermore, most existing multi-source data fusion methods simply perform a weighted combination of prediction results at the model output layer, making it difficult to deeply characterize the nonlinear coupling relationships between multimodal factors and their spatiotemporal dynamic evolution patterns. Consequently, they still have significant shortcomings in capturing hidden and progressive collapse risks.

[0004] Another key challenge in multi-source fusion modeling for urban ground collapse early warning lies in processing irregularly sampled time series data and modeling spatial dependencies. Existing methods generally assume uniform time intervals in their modeling, making it difficult to effectively utilize data from disparate sources like SAR imagery, meteorological monitoring, and geological surveys, often with inconsistent sampling periods. Furthermore, the complex spatial topology inherent in urban geological environments cannot be directly modeled by traditional convolutional or fully connected architectures, resulting in inadequate characterization of cross-regional risk transmission and neighborhood interactions.

[0005] In addition, most existing early warning systems rely on a single risk assessment channel and lack a dual verification mechanism based on physical deformation evidence and multi-factor coupling mechanisms, making it easy for missed reports and false alarms to occur in hidden collapses or short-term emergencies. Summary of the Invention

[0006] The purpose of the present invention is to address the shortcomings of existing urban ground collapse warning methods in multi-source data fusion, spatiotemporal dynamic relationship modeling, irregular sampling processing, and risk judgment mechanism, and to provide an intelligent urban ground collapse warning method based on multi-source factor fusion to achieve high-precision, low false alarm, and strong generalization prediction of urban ground collapse risks, providing reliable technical support for urban safety management.

[0007] The technical solution adopted by the present invention is: The intelligent early warning method for urban ground collapse based on multi-source factor fusion includes the following steps: S1. Collect multi-temporal SAR remote sensing image data at the same location, use two time series synthetic aperture radar interferometry techniques (PS-InSAR and SBAS-InSAR) to calculate deformation values, and extract their respective deformation time series. S2. Perform spatial consistency verification and temporal feature cross-validation analysis on the deformation time series extracted by the two processing methods. Import the verified qualified PS-InSAR processing vector results into ArcGIS to draw a time-series InSAR surface deformation rate map, that is, generate a settlement map. S3. Set a settlement rate threshold to perform preliminary ground collapse warning identification based on the settlement map obtained from InSAR deformation monitoring; S4. Collect multi-source remote sensing data and GIS spatial data. Classify the InSAR deformation time series and the collected multi-source data into dynamic and static factors. Construct a multi-channel weighted spatiotemporal graph structure with grid cells in the monitoring area as nodes. Use a spatio-temporal graph neural ODE with attention (ST-GNODEA) to fuse the multi-source data and predict the overall risk probability. In the spatiotemporal graph neural differential attention network, the monitoring area is first gridded into nodes, and the irregularly sampled multi-source time series data of each node is modeled in continuous time using neural ordinary differential equations. Through a graph attention network with self-loop connections, the attention weights of each node and its neighboring nodes (including itself) on the dynamic factor features are calculated. The hidden state vector of the dynamic factor and the static factor representation are then input into the factor-level multimodal attention module to achieve importance distribution and nonlinear coupling modeling between factors. The fused node spatiotemporal features are output through a fully connected prediction layer to determine the collapse risk probability of the node. The risk of each node forms a risk distribution map for the entire domain. S5. Conduct dual-channel fusion analysis and identify ground collapse warnings by combining the InSAR deformation threshold with the comprehensive risk probability predicted by ST-GNODEA.

[0008] In the above method, the permanent scatterer PS-InSAR processing described in step S1 includes registration, interferometric processing and deformation information extraction; the small baseline set InSAR processing SBAS-InSAR includes registration, interferometric processing, phase unwrapping, atmospheric residual filtering and deformation information extraction.

[0009] The spatial consistency verification described in step S2 is to first perform grid matching, divide the study area into grid cells of the same size, extract the mean sedimentation rate of PS-InSAR and SBAS-InSAR respectively, and then calculate the rate difference Δ between the two methods in each grid v : , If Δ v ≤5mm / a and spatial correlation coefficient R 2 If the value is ≥0.85, it is determined to be a consistent area; otherwise, it is marked as a "pending verification area" and manually verified.

[0010] The cross-validation analysis of temporal features selected the PS points and SBAS pixel clusters with a coherence coefficient greater than 0.3 and covering a continuous area, and ensured that they were in the same geographical location, and extracted their respective deformation time series data D PS (t) and D SBAS (t). These data represent the deformation value of the point at the time when the corresponding SAR image was acquired. When visualizing, these discrete data points are connected to draw a deformation-time relationship curve. The two series D are calculated by the dynamic time warping (DTW) algorithm. PS (t) and D SBAS Similarity of (t): , Similarity is the similarity index of the deformation time series, with a value range of [0, 1]. A larger value indicates a more consistent evolution trend of the two series. DTW_Distance is the minimum path cumulative distance between two points in the sequence calculated by the dynamic time warping algorithm. len is the sequence length. A similarity greater than 0.9 is considered to indicate consistent time series evolution.

[0011] The sedimentation rate thresholds in step S3 above are set as annual sedimentation rate > 30 mm / a, cumulative sedimentation in 30 days > 10 mm, and average daily sedimentation > 2 mm. Once the thresholds are exceeded, the risk marking mechanism is activated.

[0012] To address the problem that existing multi-source data fusion methods only perform simple weighted combination at the model output layer and fail to capture the nonlinear coupling of multimodal factors and their spatiotemporal dynamic relationships, this paper proposes a spatio-temporal graph neural ODE with attention (ST-GNODEA). This network grids the monitoring area into nodes, constructs a graph structure combining spatial topology and infrastructure connectivity, and uses neural ordinary differential equations to perform continuous-time modeling of irregularly sampled multi-source time series data. Furthermore, the graph attention mechanism and multimodal factor attention are combined to achieve end-to-end interpretable multi-source data fusion, thereby improving the accuracy and generalization ability of early warnings. The specific process of multi-source data fusion is as follows: (1) Graph structure construction: Divide the monitoring area into N nodes v i , node characteristics include: Dynamic Factor: InSAR Deformation Time Series , rainfall sequence wait; Static factors: soil thickness, pipe network density, terrain slope, land use type, etc.; Establish multi-channel weighted edges between nodes , m is the edge type, ij Represents a slave node i To Node j The edge weight is calculated by the spatial distance, geological similarity and infrastructure information; (2) Continuous-time modeling: For each node, a Neural ODE is used to map discrete observations into a continuous-time hidden state: , in For nodes v i The hidden state at time t is f θ is a trainable neural network, x i ( t ) represents a node i At the moment t The input feature vector of this method can adapt to the situation where the sampling interval of sensors such as InSAR is uneven; (3) Spatial information interaction: The graph attention network (GAT) is introduced on the hidden state to calculate the attention weights of neighbor nodes: , Where a is the attention weight, 、 is the vector representation of the node feature after linear transformation, Leaky Re LU Leaky ReLU activation function allows input less than 0 to retain a certain gradient to avoid dead neurons. Normalize all neighbor nodes (including itself) so that the sum of weights is 1. e ij is a node i With node j The edge eigenvector between ; Aggregate information: , in For nodes i The neighbor set of is the attention weight, is the linear transformation weight matrix, node j The original feature vector of As the dynamic factor hidden state, it enters the factor-level attention calculation module together with the static factor features; (4) Multimodal factor attention fusion: ①Node input and dynamic hidden state remember For nodes i At the moment t The dynamic hidden state vector obtained by spatial information interaction has encoded the neighbor nodes and edge weight information; For nodes i No. m static factor vectors, dynamic factors in time i The expression can be expressed by Part of the weight is obtained directly; ② Factor-level attention calculation For each node i , at the moment t The importance weights of each factor are calculated as follows: , Where u is the attention score vector, tanh is the hyperbolic tangent activation function, W d is the weight matrix mapping the dynamic factors, For nodes i At the moment t The dynamic hidden state vector obtained through spatial information interaction, W s is the weight matrix mapping the static factors, For nodesi No. m static factor vectors; m is the number of the current factor mode, m′ represents the index of all modal factor sets, and is the traversal variable in the summation; ③ Fusion feature construction Use factor weights to weight dynamic and static information to obtain fused features: , in is the fused node modal feature vector, W m For the m The feature fusion weight matrix of factors, For the moment t node i No. m The attention weight of the factor.

[0013] ④Risk prediction The fused feature vector Input to the prediction layer, input the fused node features into the prediction layer, and output the node collapse risk probability through the sigmoid function. Node risk probability: , in is the node risk probability, σ is the Sigmoid activation function, W p is the risk prediction weight vector, b p is the prediction bias term. When the system issues warnings in grid units, the node risk probabilities can be aggregated by weight.

[0014] (5) After the prediction is completed, uncertainty estimation and online adaptive update can be performed: Introducing a confidence estimation method based on evidence theory to calculate prediction uncertainty , and combine the risk probability and uncertainty threshold to trigger early warning to avoid false alarms under low confidence: introduce the uncertainty estimation method based on Monte Carlo Dropout and Bayesian neural network in the output layer: , in Prediction uncertainty, is the prediction result of the mth sampling, M is the number of Monte Carlo sampling, is the mean prediction.

[0015] The incremental learning strategy combining Elastic Weight Consolidation (EWC) and Experience Replay is used to achieve online adaptive updates: a. When the model performance drops below a set threshold within a short-term window (24 days), a small batch online update is triggered; b. Neural ODE parameters are updated through small step optimization, while the EWC regularization term protects important parameters and prevents the forgetting of historical knowledge.

[0016] The present invention introduces a dual-channel fusion analysis mechanism in early warning judgment, in which the InSAR deformation threshold channel and the ST-GNODEA comprehensive risk probability channel work together to achieve multi-layer protection that combines rapid response with accurate prediction.

[0017] The dual-channel analysis rule described in step S5 above is: If the InSAR sedimentation rate value exceeds the set threshold, it indicates that the physical deformation has reached a critical state and is considered a potential risk; The ST-GNODEA predicted risk is set to be greater than 0.8 as the high risk threshold. and When , it indicates that the disaster risk caused by the coupling of multiple source factors has entered the high-risk range and is considered a high-risk area; When the above conditions are met at the same time, a high-priority warning signal is triggered, and spatial visualization output and alarm prompts are performed.

[0018] Another object of the present invention is to provide an intelligent early warning system for urban ground collapse based on multi-source factor fusion, comprising: Data acquisition unit, used for multi-temporal SAR image data acquisition, multi-source remote sensing data acquisition and GIS spatial data acquisition; The data preprocessing unit preprocesses the multi-source data acquired by the data acquisition unit, including SAR image registration, interferometric processing, phase unwrapping, atmospheric delay correction, multi-source data coordinate unification and resampling, missing value filling and outlier removal, to generate a spatiotemporally aligned multi-source factor data set; Based on the InSAR deformation monitoring unit, two time series synthetic aperture radar interferometry techniques, PS-InSAR and SBAS-InSAR, are used to process multi-temporal SAR image data, calculate deformation values, and form deformation time series. The consistency of the deformation time series extracted by the two processing methods is verified to generate a settlement map. The adaptive multi-source factor fusion modeling unit constructs a node feature set containing dynamic and static factors based on multi-source remote sensing data, GIS spatial data, and InSAR deformation results. It also constructs a multi-channel weighted spatiotemporal graph structure with the monitoring area grid cells as nodes, and uses the spatiotemporal graph neural differential attention network ST-GNODEA to achieve end-to-end multimodal factor fusion and continuous time modeling. The model performance evaluation and optimization unit monitors the prediction performance indicators of the spatiotemporal graph neural differential attention network model under the sliding window in real time. When the performance drops below a threshold, the online update mechanism is triggered to use newly collected data for small-batch incremental training and dynamically adjust the factor attention weights and Neural ODE parameters. The dual-channel analysis unit performs early warning analysis based on the subsidence rate threshold set by InSAR deformation and the comprehensive risk probability of ST-GNODEA risk probability; The early warning unit issues warnings for areas that exceed the sedimentation rate threshold and have a high predicted risk probability.

[0019] The beneficial effects of the present invention are: (1) Two time series synthetic aperture radar interferometry techniques are used to continuously monitor deformation in urban areas, achieving millimeter-level precision in the extraction of ground subsidence time series, and providing deformation prior information for subsequent collapse risk analysis. The combination of PS-InSAR and SBAS-InSAR technology overcomes the problem that traditional D-InSAR is greatly affected by atmospheric noise and phase unwrapping in complex urban terrain. Through the complementary analysis of PS-InSAR and SBAS-InSAR, the limitations of a single method are eliminated and the reliability of deformation monitoring results is improved.

[0020] (2) Multi-source remote sensing and environmental factor fusion modeling integrates InSAR deformation, rainfall time series, geological parameters, underground pipe network density, DEM, slope, land use, and other multi-source factors to construct a spatiotemporally aligned rasterized sample library and achieve collaborative characterization of dynamic and static factors. This method overcomes the blind spots of traditional models in depicting multi-factor coupled disaster mechanisms. Combined with GIS spatial analysis and normalization processing, it achieves unified encoding and seamless integration of heterogeneous data, providing high-quality input for subsequent intelligent modeling.

[0021] (3) This paper combines neural ordinary differential equations with graph attention networks and factor attention mechanisms to achieve continuous-time modeling and interpretable factor-level fusion of multi-source data. The graph attention mechanism introduces spatial proximity and infrastructure connectivity to capture the propagation path of surface instability and the interaction between neighborhoods. The fusion process comprehensively depicts the nonlinear coupling and spatiotemporal dynamic relationships between disaster-causing factors.

[0022] (4) Online adaptive updates ensure long-term stability. An incremental learning strategy combining elastic weight consolidation and experience replay is employed. When model performance degradation is detected, small batch online updates are triggered to ensure rapid absorption of new knowledge and retention of historical knowledge. This strategy can adapt to different urban geological conditions and climate change environments, significantly reducing the risk of early warning failures caused by data distribution drift.

[0023] (5) Multi-dimensional confidence assessment improves decision reliability. By combining the Bayesian method with Monte Carlo Dropout, multiple sampling estimates of the predicted distribution are used to obtain risk probability and uncertainty indicators. When the uncertainty exceeds the set threshold, the system automatically issues a "focus attention" prompt to prevent low-confidence predictions from directly triggering warnings, thereby achieving controllable and robust warning quality.

[0024] (6) The dual-channel early warning mechanism achieves simultaneous verification of physical deformation signals and disaster-causing mechanisms through the dual-channel fusion of the deformation threshold channel and the ST-GNODEA risk probability channel, thereby significantly improving early warning accuracy and reducing false alarm rates. The deformation threshold channel precisely focuses on capturing key evidence of surface instability, while the ST-GNODEA risk probability channel identifies progressive hazards through the fusion of multiple sources. The two work together to effectively suppress false alarms and false positives caused by construction vibration, ensure monitoring accuracy, and enhance the model's generalization and long-term stability under different urban and environmental conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 is a flow chart of the method of the present invention; Figure 2 The surface deformation monitoring results of a certain district in Shenzhen based on time-series InSAR in an embodiment of the present invention are as follows: a. PS-InSAR processing results, b. SBAS-InSAR processing results; Figure 3 This is the result of calculating the ground collapse risk probability in a certain district of Shenzhen using the ST-GNODEA model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0026] The present invention is further described below with reference to specific embodiments.

[0027] Example 1 An intelligent early warning method for urban ground collapse based on multi-source factor fusion includes the following steps: S1. Collect multi-temporal SAR remote sensing image data at the same location, use two time series synthetic aperture radar interferometry techniques (PS-InSAR and SBAS-InSAR) to calculate deformation values, and extract their respective deformation time series: Firstly, two time series synthetic aperture radar interferometry techniques - permanent scatterer InSAR and small baseline set InSAR - were used to continuously monitor deformation in urban areas, achieving millimeter-level precision in the extraction of ground subsidence time series, providing deformation prior information for subsequent collapse risk analysis.

[0028] The core principle of InSAR is to use pairs of images acquired by SAR satellites at different times to extract information about the relative deformation of ground objects along the direction of radar wave propagation through interferometry. In this paper, the combination of PS-InSAR and SBAS-InSAR technologies overcomes the significant impact of atmospheric noise and phase unwrapping on traditional D-InSAR in complex urban terrain.

[0029] PS-InSAR processing includes registration, interferometry processing and deformation information extraction; SBAS-InSAR processing includes registration, interferometry processing, phase unwrapping, atmospheric residual filtering and deformation information extraction.

[0030] (1) Collect multi-temporal SAR image data and perform precise registration: A continuous time series of SAR images is obtained from the satellite platform, a master image is selected, a master-slave relationship is established according to the set baseline threshold, and the images are registered to ensure comparability between pixel pairs of the same ground object.

[0031] (2) Interference processing: Interference processing is performed on each interferometric image pair (i.e., a SAR image pair consisting of a master image and a slave image) to generate an interferogram. The formula is as follows: , in is the flat Earth phase caused by the Earth's curvature, is the phase caused by terrain undulation, is the deformation phase, is the phase caused by atmospheric propagation delay, The high-precision digital elevation model can eliminate the flat ground phase caused by the curvature of the earth and the phase caused by the terrain undulation, thereby obtaining a differential interferogram.

[0032] (3) The original differential phase is filtered in the spatial and temporal domains to improve the phase signal-to-noise ratio, and then the phase is unwrapped to restore the continuous physical phase quantity. The formula is as follows: , in Is the phase unwrapping function, that is, the algorithm is used to restore the wrapped phase to a continuous phase. Common represents the wrapped phase, that is, the phase value limited to the range of (−π,π] after interference processing. Since the phase measured by radar is periodic, it can only measure relative changes within a wavelength range. If it exceeds the range, it will "wrap around".

[0033] (4) Deformation calculation and LOS conversion: The accumulated deformation along the radar line of sight (LOS) is calculated using the unwrapped interference phase. The formula is as follows: , in is the deformation along the sight line direction, λ is the radar wavelength, is the interference phase after unwrapping.

[0034] Radar measurement It is the change in slant distance from the satellite to the ground point. Since ground subsidence mainly moves in the vertical direction, it is necessary to calculate the slope of the ground according to the incident angle of the radar beam. θ Will Projected to the vertical direction, the vertical deformation is obtained : , in θ is the incident angle of the radar.

[0035] (5) PS-InSAR and SBAS-InSAR deformation time series extraction: PS-InSAR: This system uses statistical analysis to identify point targets with stable scattering properties (i.e., permanent scatterers) and creates long-term coherent image pairs. Through differential interferometry, orbit error fitting, and atmospheric noise modeling, it extracts the target point deformation time series D(t), achieving high temporal resolution and noise suppression.

[0036] SBAS-InSAR constructs multiple interferograms with small spatial and temporal baselines, forming an underdetermined set of equations. The deformation time series of each pixel is reconstructed through singular value decomposition (SVD). It is suitable for scenes with wide areas and sparse scatterers.

[0037] S2. Verify the spatial consistency and cross-validate the temporal characteristics of the deformation time series extracted by the two processing methods. Import the vector results of the verified PS-InSAR processing into ArcGIS to draw a time-series InSAR surface deformation rate map, that is, to generate a settlement map: Through the complementary analysis of PS-InSAR and SBAS-InSAR, the limitations of a single method are eliminated and the reliability of deformation monitoring results is improved.

[0038] To verify spatial consistency, we first performed grid matching and divided the study area into 500m×500m grid cells. We extracted the mean sedimentation rates of PS-InSAR and SBAS-InSAR respectively, and then calculated the rate difference Δ between the two methods within each grid. v : , in v PS is the mean sedimentation rate extracted by PS-InSAR, v SBAS is the mean sedimentation rate extracted by SBAS-InSAR, if Δ v ≤5mm / a and spatial correlation coefficient R 2 If the value is ≥0.85, it is determined to be a consistent area; otherwise, it is marked as a "pending verification area" and manually verified.

[0039] Cross-validation analysis of temporal features was performed. The coherence coefficient was greater than 0.3 and the pixel clusters covered a continuous area. At the same time, the PS points and SBAS pixel clusters were ensured to be at the same geographical location. Their respective deformation time series data D were extracted. PS (t) and D SBAS (t), these data represent the deformation value of the point at the time when the corresponding SAR image is obtained. When visualizing, these discrete data points are usually connected to draw a deformation-time relationship curve. The dynamic time warping (DTW) algorithm is used to calculate the two series D PS (t) and D SBAS Similarity of (t): , Similarity is the similarity index of the deformation sequences, ranging from 0 to 1. A larger value indicates a more consistent evolution trend between the two sequences. DTW_Distance is the minimum path cumulative distance between two points in the sequence calculated by the dynamic time warping algorithm. D PS Deformation time series extracted by PS-InSAR. D SBAS is the deformation time series extracted by SBAS-InSAR. len is the length of the sequence. A similarity > 0.9 is considered to indicate consistent temporal evolution.

[0040] S3. Set the settlement rate threshold and perform preliminary ground collapse warning identification based on the settlement map obtained from InSAR deformation monitoring: An early warning threshold is set (annual subsidence rate > 30 mm / a, cumulative subsidence in 30 days > 10 mm, and average daily subsidence > 2 mm). Once the InSAR results exceed the threshold, the risk marking mechanism is activated.

[0041] S4. Collect multi-source remote sensing data and GIS spatial data. Classify the InSAR deformation time series and the collected multi-source data into dynamic and static factors. Construct a multi-channel weighted spatiotemporal graph structure with grid cells in the monitoring area as nodes. Use a spatio-temporal graph neural ODE with attention (ST-GNODEA) to fuse the multi-source data and predict the overall risk probability. This method divides the monitoring area into N nodes v i ,The node characteristics are composed of dynamic factors and static factors, and the spatial and functional connections between nodes are characterized through multi-channel weighted edges.

[0042] (1) Factor characteristic system: This paper constructs a ground collapse risk prediction dataset by building a spatiotemporal sample and fusing InSAR deformation data with multiple external remote sensing and geological factors. The factors involved include: ①Surface deformation Cumulative surface settlement, extracted using time-series InSAR technology, reflects soil compression, displacement, or instability trends during the monitoring period. This factor is a direct deformation precursor to collapse risk and is particularly sensitive to progressive soil loss and sudden karst collapse. Data is input as a raster time series with a spatial resolution consistent with SAR imagery.

[0043] ② Rainfall time series Cumulative rainfall data generated by interpolation from meteorological satellites and ground-based rain gauges. Rainfall is a core dynamic factor inducing collapse: short-term heavy rainfall reduces shear strength by increasing soil saturation and pore water pressure; long-term rainfall exacerbates foundation softening and karst erosion. The data must be constructed as a time series grid synchronized with InSAR to capture the lag effect of rainfall intensity-deformation response.

[0044] ③Underground pipe network density The number of underground pipelines per unit area is calculated using the urban infrastructure GIS database. Areas with high pipe density significantly increase the risk of soil loss due to aging, cracking, leakage, and erosion, especially in densely populated water supply and drainage networks. This factor is input as a static raster layer with a spatial resolution consistent with the early warning system.

[0045] ④Road density Kernel density data for road coverage generated from road network vector data. In densely populated areas, repeated dynamic loads from traffic accelerate roadbed fatigue, while voids left behind from underground pipeline corridor construction can easily lead to collapse. This factor should be correlated with heavy vehicle traffic frequency data to improve prediction accuracy.

[0046] ⑤Land use type Land cover class codes derived from remote sensing classification and planning data. Different land use types imply different hazard exposures.

[0047] ⑥Soil layer thickness The vertical thickness of the weak soil layer, obtained through geological drilling and electromagnetic surveys. When the weak layer is thicker than 5 meters, the soil is susceptible to liquefaction or compression settlement due to seepage and vibration, significantly reducing the critical deformation threshold. This factor is input as a spatially interpolated raster with a resolution that matches the system.

[0048] ⑦ Distance to major rivers The Euclidean distance from the grid cell to the nearest river, calculated based on hydrological network GIS data. Nearby areas are strongly affected by riverbank erosion and groundwater level fluctuations, especially during flood season, when lateral erosion of the riverbed can easily trigger bank collapse. This data needs to be weighted and optimized based on river width and flow velocity.

[0049] ⑧Digital Elevation Model (DEM) Surface elevation raster data derived from LiDAR and stereo aerial surveys. Water accumulation in low-lying areas can accelerate ground softening.

[0050] ⑨Slope factor The surface slope angle is calculated using the Sobel operator from DEM data. Areas with slopes greater than 10° are subject to concentrated shear stress, making them prone to slope instability. Furthermore, rainfall runoff in high-slope areas increases velocity, exacerbating topsoil erosion. This factor needs to be coupled with lithologic data to assess anti-slide capacity.

[0051] All factors are rasterized by GIS to achieve uniform resolution. All factor data are normalized: , A time series sample is constructed for each grid cell. The true label comes from the collapse event point in the historical record, which is marked as 1, and the non-collapsed area is marked as 0.

[0052] (2) Multi-source factor data fusion: ①Graph structure construction: Node Definition: Each node represents a permanent scattering point (PS point) or grid cell, and is associated with both dynamic and static factors. Dynamic factors include: InSAR deformation time series and rainfall time series; static factors include: underground pipe density, road density, land use type, soil thickness, distance to rivers, DEM, and slope.

[0053] Edge relationship construction: establishing multi-channel weighted edges , m is the edge type, ij represents an edge from node i to node j, and the edge weight is calculated comprehensively by spatial distance, geological similarity and infrastructure network information to form a multi-scale spatial topology.

[0054] Multi-source factors and graph structures are jointly modeled in a unified end-to-end framework.

[0055] ②Continuous time modeling: For dynamic factors such as InSAR deformation time series and rainfall time series, Neural ODE is used to represent them as continuous time hidden state h i ( t ): , in For nodes v i The hidden state at time t is f θ is a trainable neural network, x i ( t ) represents a node i At the moment t The NeuralODE method can adapt to the situation where the sampling interval of sensors such as InSAR is uneven; NeuralODE allows the state evolution to be directly calculated at irregular time intervals, thereby adapting to the different sampling frequencies of different data sources and performing high-precision interpolation or extrapolation prediction when new data arrives.

[0056] ③Spatial information interaction: The Graph Attention Network (GAT) is introduced on the hidden state to calculate the attention weights of neighbor nodes. In the continuous time hidden state, GAT is used to transmit messages between the spatial and functional topologies of nodes: , Where a is the attention weight, 、 is the vector representation of the node feature after linear transformation, Leaky Re LU Leaky ReLU activation function allows input less than 0 to retain a certain gradient to avoid dead neurons. Normalize all neighbor nodes (including itself) so that the sum of weights is 1. e ij is a node i With node j The edge eigenvector between ; The multi-channel edge structure can simultaneously capture multiple risk transmission paths, such as groundwater seepage diffusion caused by pipe rupture and runoff path of slope rainfall infiltration.

[0057] Aggregate information: , in For nodes i The neighbor set of is the attention weight, is the linear transformation weight matrix, node j The original feature vector of As the dynamic factor hidden state, it enters the factor-level attention calculation module together with the static factor features.

[0058] ④Multimodal factor attention fusion: a. Node input and dynamic hidden state remember For nodes i At the moment t The dynamic hidden state vector obtained by spatial information interaction has encoded the neighbor nodes and edge weight information; For nodes i No. m static factor vectors, dynamic factors in time i The expression can be expressed by Part of the component is obtained directly.

[0059] b. Factor-level attention calculation For each node i , at the moment t The importance weights of each factor are calculated as follows: , Where u is the attention score vector, tanh is the hyperbolic tangent activation function, W d is the weight matrix mapping the dynamic factors, For nodes i At the moment t The dynamic hidden state vector obtained through spatial information interaction, W s is the weight matrix mapping the static factors, For nodes i No. m static factor vectors; m is the number of the current factor mode, m′ represents the index of all modal factor sets, and is the traversal variable in the summation; by recording The changes in the risk factor can intuitively explain the main driving factors of the predicted risk at a certain moment. For example, the weight of the rainfall factor increases significantly during heavy rain, and the weight of the pipeline factor is high for a long time in areas with dense underground pipeline networks.

[0060] c. Fusion feature construction Use factor weights to weight dynamic and static information to obtain fused features: , in is the fused node modal feature vector, W m For the m The feature fusion weight matrix of factors, For the moment t node i No. m The attention weight of the factor.

[0061] d. Risk prediction The fused feature vector Input to the prediction layer, input the fused node features into the prediction layer, and output the node collapse risk probability through the sigmoid function. Node risk probability: , in is the node risk probability, σ is the Sigmoid activation function, W p is the risk prediction weight vector, b p is the prediction bias term. When the system issues warnings in grid units, the node risk probability can be summarized by weight; By Record The changes in the risk factor can intuitively explain the main driving factors of the predicted risk at a certain moment. For example, the weight of the rainfall factor increases significantly during heavy rain, and the weight of the pipeline factor is high for a long time in areas with dense underground pipeline networks.

[0062] e. Estimation of prediction uncertainty In order to quantify the reliability of the prediction results, the present invention introduces an uncertainty estimation method based on Monte Carlo Dropout and Bayesian neural network in the output layer: , in Prediction uncertainty, is the prediction result of the mth sampling, M is the number of Monte Carlo sampling, The uncertainty is used to assist the dual-channel early warning mechanism. When the risk probability is high but the uncertainty is large, the area is designated as a "key focus area" rather than directly triggering a high-level warning, reducing the risk of false alarms.

[0063] (3) ST-GNODEA model training and online optimization strategy: a. Loss function design Use weighted binary cross entropy loss and adjust the weights of positive and negative classes in combination with sample imbalance: , inL is the loss value, which represents the difference between the model prediction and the true label. N is the total number of samples. y i is the true label, which takes the value of 0 or 1. W is the probability value predicted by the model, indicating the probability that the sample belongs to the positive class, and its value range is [0, 1]. pos is the positive class weight, which is used to adjust the importance of positive samples in loss calculation. neg is the negative class weight, which is used to adjust the importance of negative class samples in loss calculation.

[0064] b. Online learning and incremental updates To ensure that the system adapts to dynamically changing geological and environmental conditions, the present invention introduces an incremental learning strategy of elastic weight consolidation plus experience replay under a single model architecture: , in L new is the loss function for new data, F i are the diagonal elements of the information matrix, is the optimal parameter for the historical task, θ i is a node i The differential equation parameters, λ is the regularization coefficient.

[0065] Experience replay: Mixing new data with historical key samples to maintain the generalization ability of spatiotemporal features; Update trigger conditions: Incremental updates are triggered when the warning accuracy within the sliding window drops by more than 2% or when significant data distribution drift occurs; Periodic calibration: Perform full parameter calibration every 30 days.

[0066] c. Parameter optimization and training stability When new monitoring data is obtained, the model weights are updated through the gradient descent optimization algorithm AdamW, an improved version of the Adam optimizer, is used. It introduces a weight decay mechanism to prevent overfitting, limits the gradient norm to no more than 1.0 to prevent gradient explosion, and uses FP16 precision to accelerate training while maintaining the numerical stability of FP32 precision.

[0067] d. Knowledge retention and environmental adaptation In terms of knowledge retention, by identifying the feature weights that contribute greatly to historical warning events, L2 regularization is used to constrain their parameter changes; in multi-scenario adaptation, independent output layers are set for different geological blocks and the underlying feature extractor is shared; in distribution alignment, the maximum mean difference loss is used to narrow the feature distribution differences between the source domain and the new target domain to achieve smooth transfer learning.

[0068] e. Real-time evaluation of model performance Multi-scale performance trend monitoring and adaptive maintenance: Based on the SAR image acquisition cycle, a 24-day (short-term), 72-day (medium-term), and 144-day (long-term) multi-scale sliding window performance monitoring framework is constructed to simultaneously capture timeliness changes and seasonal trends.

[0069] Trend analysis method: Combine linear regression slope, Mann-Kendall monotonicity test, and CUSUM change point detection to identify performance trends and discover mutation points.

[0070] Graded warning thresholds: short-term F1 score drop > 1% (minor), medium-term precision drop > 3% (significant), long-term recall drop > 5% (serious).

[0071] Adaptive maintenance strategy: Slight: Quick incremental fine-tuning; Significant: Experience replay + EWC incremental training; Critical: Roll back to a stable model snapshot and fully retrain.

[0072] This mechanism realizes dynamic monitoring and hierarchical maintenance of the ST-GNODEA model performance, ensuring the long-term stability of prediction accuracy.

[0073] This mechanism can achieve long-term monitoring and adaptive optimization of single model performance without relying on multi-submodel comparison, ensuring the prediction accuracy and stability of ST-GNODEA under changes in data distribution, seasonal fluctuations or sudden environmental interference.

[0074] S5. Dual-channel assessment and early warning based on the subsidence rate threshold set by InSAR deformation and the comprehensive risk probability of the ST-GNODEA model: The final output of the system is the collapse risk probability P of each grid i (t), the present invention proposes a dual-channel analysis rule as follows: If the InSAR sedimentation value exceeds the set threshold, it indicates that the physical deformation has reached a critical state and is considered a potential risk; The ST-GNODEA predicted risk is set to be greater than 0.8 as the high risk threshold. and When the high-risk warning is triggered directly; When the above conditions are met at the same time, a high-priority warning signal is triggered, and spatial visualization output and alarm prompts are performed. The warning trigger conditions are shown in the following table.

[0075] Table 1 Warning trigger conditions .

[0076] A total of 78 Sentinel-1 image data from June 2, 2022 to May 17, 2025 were selected to conduct PS-InSAR and SBAS-InSAR monitoring of a district in Shenzhen, and the surface deformation data of the district were obtained. Figure 2 shown.

[0077] Monitoring shows that the maximum local subsidence rate in Futian District reached -48.94 mm / year, with an accuracy of ±3 mm, meeting the Level 1 monitoring standard of the "Engineering Surveying Code" (GB50026-2007). It also shows that subsidence in Futian District exhibits spatial clustering, with high-value areas exhibiting a "dual-core distribution."

[0078] Huanggang Port Area: The subsidence center is located within 500 meters of the port joint inspection building, and is related to the combined effects of the dynamic load of 23,800 cross-border trucks traveling between Shenzhen and Hong Kong per day (data from 2023) and underground pipeline leakage.

[0079] The settlement rate under the city's core main roads, such as Shennan Avenue and Binhe Avenue, exceeds -30mm / a. The main cause of the settlement may be the stress release of the soil induced by the shallow buried construction of the subway tunnel.

[0080] Fusion analysis based on time series InSAR monitoring and ST-GNODEA output Figure 3 As shown, the median probability of ground collapse risk in the entire area reached 0.48, significantly exceeding the safety threshold. InSAR monitoring results indicate that the Huanggang Port area has the most significant subsidence effect, posing a significant collapse risk. ST-GNODEA results indicate that urban arterial roads such as Shennan Avenue and Binhe Avenue constitute a high-risk cluster, with a high proportion of grids with a comprehensive risk probability greater than 0.8. However, the risk at Huanggang Port, an area of ​​severe subsidence, is relatively low. This is because Huanggang Port is primarily engaged in import and export trade and is less affected by urban life and commuting. Ignoring non-living areas, such as Huanggang Port, would lead to misjudgment. Using dual-channel monitoring using time-series InSAR subsidence monitoring thresholds and ST-GNODEA comprehensive risk thresholds is beneficial for improving the accuracy of urban ground collapse warnings.

[0081] This conclusion was verified by spatial correlation analysis with the kernel density distribution of historical collapse events. The results showed that the high risk prediction area ( ) and the high kernel density hotspots of historical collapse events (kernel density value > 0.85) reached a Pearson correlation coefficient of r = 0.91 (p < 0.001), indicating that the proposed integrated model has better performance in identifying high-risk areas.

[0082] In summary, the accuracy verification of the early warning model shows a Pearson correlation coefficient of 0.91 (p < 0.001) between the prediction results and the historical spatial distribution of collapses, a 92.3% recall rate for high-risk areas, and a false alarm rate below 5%. These empirical results fully demonstrate that the ST-GNODEA model, by integrating time-series InSAR deformation with a multi-source factor coupling mechanism, possesses significant scientific and engineering effectiveness in urban ground collapse early warning, and possesses significant technical advantages and application value compared to traditional single-model approaches.

[0083] Example 2 An intelligent early warning system for urban ground collapse based on multi-source factor fusion realizes closed-loop management of the entire process from data acquisition and processing to risk early warning. The specific functional modules are as follows: Data acquisition unit, which collects multi-temporal SAR image data, multi-source remote sensing data, and GIS spatial data; A data preprocessing unit, for preprocessing the data collected by the data collection unit; Based on the InSAR deformation monitoring unit, two time series synthetic aperture radar interferometry techniques, PS-InSAR and SBAS-InSAR, were used to process multi-temporal SAR image data, calculate deformation values, and form deformation time series. The consistency of the deformation time series extracted by the two processing methods was verified to generate a settlement map. The adaptive multi-source factor fusion modeling unit constructs a node feature set containing dynamic and static factors based on multi-source remote sensing data, GIS spatial data, and InSAR deformation results. It also constructs a multi-channel weighted spatiotemporal graph structure with the monitoring area grid cells as nodes, and uses the spatiotemporal graph neural differential attention network ST-GNODEA to achieve end-to-end multimodal factor fusion and continuous time modeling. The model performance evaluation and optimization unit monitors the prediction performance indicators of the spatiotemporal graph neural differential attention network model under the sliding window in real time. When the performance drops below a threshold, the online update mechanism is triggered to use newly collected data for small-batch incremental training and dynamically adjust the factor attention weights and Neural ODE parameters. The dual-channel analysis unit performs early warning analysis based on the subsidence rate threshold set by InSAR deformation and the comprehensive risk probability of ST-GNODEA risk probability; The early warning unit issues warnings for areas that exceed the sedimentation rate threshold and have a high predicted risk probability.

[0084] The above is a further description of the present invention in conjunction with the embodiments, and the protection scope of the present invention is not limited thereto.

Claims

1. An intelligent early warning method for urban ground collapse based on multi-source factor fusion is characterized by: The steps are as follows: S1. Collect multi-temporal SAR remote sensing image data at the same location, use two time series synthetic aperture radar interferometry techniques (PS-InSAR and SBAS-InSAR) to calculate deformation values, and extract their respective deformation time series. S2. Perform spatial consistency verification and temporal feature cross-validation analysis on the deformation time series extracted by the two processing methods. Import the verified qualified PS-InSAR processing vector results into ArcGIS to draw a time-series InSAR surface deformation rate map, that is, generate a settlement map. S3. Set a settlement rate threshold to perform preliminary ground collapse warning identification based on the settlement map obtained from InSAR deformation monitoring; S4. Collect multi-source remote sensing data and GIS spatial data. Classify the InSAR deformation time series and the collected multi-source data into dynamic and static factors. Construct a multi-channel weighted spatiotemporal graph structure with the monitoring area grid cells as nodes. Use the spatiotemporal graph neural differential attention network (ST-GNODEA) to fuse multi-source data and predict the comprehensive risk probability. In the spatiotemporal graph neural differential attention network, the monitoring area is first gridded into nodes, and the irregularly sampled multi-source time series data of each node is modeled in continuous time using neural ordinary differential equations. Through a graph attention network with self-loop connections, the attention weights of each node and its neighboring nodes on the dynamic factor features are calculated. The hidden state vector of the dynamic factor and the static factor representation are then input into the factor-level multimodal attention module to achieve importance distribution and nonlinear coupling between factors. The fused node spatiotemporal features are output through a fully connected prediction layer to determine the collapse risk probability of the node. The risk of each node forms a global risk distribution map. S5. Conduct dual-channel fusion analysis and identify ground collapse warnings by combining the InSAR deformation threshold with the comprehensive risk probability predicted by ST-GNODEA.

2. The intelligent early warning method for urban ground collapse based on multi-source factor fusion according to claim 1 is characterized in that: The permanent scatterer PS-InSAR processing described in step S1 includes registration, interferometric processing and deformation information extraction; the small baseline set InSAR processing SBAS-InSAR includes registration, interferometric processing, phase unwrapping, atmospheric residual filtering and deformation information extraction.

3. The intelligent early warning method for urban ground collapse based on multi-source factor fusion according to claim 1 is characterized in that: The spatial consistency verification described in step S2 is to first perform grid matching, divide the study area into grid cells of the same size, extract the mean sedimentation rate of PS-InSAR and SBAS-InSAR respectively, and then calculate the rate difference Δ between the two methods in each grid v : ， If Δ v ≤5mm / a and spatial correlation coefficient R 2 If the value is ≥0.85, it is determined to be a consistent area; otherwise, it is marked as a "pending verification area" and manually verified.

4. The intelligent early warning method for urban ground collapse based on multi-source factor fusion according to claim 1 is characterized in that: The time series feature cross validation analysis described in step S2 selects the PS points and SBAS pixel clusters with a coherence coefficient greater than 0.3 and covering a continuous area, and ensures that they are in the same geographical location, and extracts their respective deformation time series data D PS (t) and D SBAS (t), calculate the two sequences D by dynamic time warping algorithm PS (t) and D SBAS Similarity of (t): ， Similarity is the similarity index of the deformation time series, with a value range of [0, 1]. A larger value indicates a more consistent evolution trend of the two series. DTW_Distance is the minimum path cumulative distance between two points in the sequence calculated by the dynamic time warping algorithm. len is the sequence length. A similarity greater than 0.9 is considered to indicate consistent time series evolution.

5. The intelligent early warning method for urban ground collapse based on multi-source factor fusion according to claim 1 is characterized in that: The sedimentation rate thresholds in step S3 are set as annual sedimentation rate > 30 mm / a, cumulative sedimentation in 30 days > 10 mm, and daily average sedimentation > 2 mm. Once the thresholds are exceeded, the risk marking mechanism is activated.

6. The intelligent early warning method for urban ground collapse based on multi-source factor fusion according to claim 1 is characterized in that: The specific process of multi-source data fusion in step S4 is as follows: (1) Graph structure construction: Divide the monitoring area into N nodes v i , node characteristics include: Dynamic factors: InSAR deformation time series, rainfall series; Static factors: soil thickness, pipe network density, terrain slope, and land use type; Establish multi-channel weighted edges between nodes , m is the edge type, ij represents an edge from node i to node j, and the edge weight is calculated comprehensively by spatial distance, geological similarity and infrastructure information; (2) Continuous-time modeling: For each node, a neural ordinary differential equation is used to map discrete observations into a continuous-time hidden state: ， in For nodes v i The hidden state at time t is f θ is a trainable neural network, x i ( t ) represents a node i At the moment t The input feature vector of (3) Spatial information interaction: Introduce the graph attention network on the hidden state and calculate the attention weights of neighbor nodes: ， Where a is the attention weight, 、 is the vector representation of the node feature after linear transformation, Leaky Re LU Leaky ReLU activation function allows input less than 0 to retain a certain gradient to avoid dead neurons. Normalize all neighboring nodes including itself so that the sum of weights is 1. e ij is a node i With node j The edge eigenvector between ; Aggregate information: ， in For nodes i The neighbor set of is the attention weight, is the linear transformation weight matrix, For nodes j The original feature vector of As the dynamic factor hidden state, it enters the factor-level attention calculation module together with the static factor features; (4) Multimodal factor attention fusion: ①Node input and dynamic hidden state, remember For nodes i At the moment t The dynamic hidden state vector obtained by spatial information interaction has encoded the neighbor nodes and edge weight information; For nodes i No. m static factor vectors, dynamic factors in time i The expression can be expressed by Part of the weight is obtained directly; ② Factor-level attention calculation, For each node i , at the moment t The importance weights of each factor are calculated as follows: ， Where u is the attention score vector, tanh is the hyperbolic tangent activation function, W d is the weight matrix mapping the dynamic factors, For nodes i At the moment t The dynamic hidden state vector obtained through spatial information interaction, W s is the weight matrix mapping the static factors, For nodes i No. m A static factor vector, m is the number of the current factor mode, m′ represents the index of all modal factor sets, and is the traversal variable in the summation; ③ Fusion feature construction, Use factor weights to weight dynamic and static information to obtain fused features: ， in is the fused node modal feature vector, W m For the m The feature fusion weight matrix of factors, For the moment t node i No. m The attention weight of each factor; ④ Risk prediction, The fused feature vector Input to the prediction layer, input the fused node features into the prediction layer, and output the node collapse risk probability through the sigmoid function: ， in is the node risk probability, σ is the Sigmoid activation function, W p is the risk prediction weight vector, b p is the prediction bias term.

7. The intelligent early warning method for urban ground collapse based on multi-source factor fusion according to claim 6 is characterized in that: The multi-source data fusion described in step S4 includes uncertainty estimation and online adaptive update after the prediction is completed.

8. The intelligent early warning method for urban ground collapse based on multi-source factor fusion according to claim 7 is characterized in that: In the output layer, an uncertainty estimation method based on Monte Carlo Dropout and Bayesian neural network is introduced: ， in is the prediction uncertainty, is the prediction result of the mth sampling, M is the number of Monte Carlo sampling, is the mean prediction; An incremental learning strategy combining elastic weight consolidation and experience replay is adopted for online adaptive update.

9. The intelligent early warning method for urban ground collapse based on multi-source factor fusion according to claim 1 is characterized in that: The dual-channel analysis rule described in step S5 is: If the InSAR sedimentation rate value exceeds the set threshold, it indicates that the physical deformation has reached a critical state and is considered a potential risk; Set the ST-GNODEA predicted risk greater than 0.8 as the high risk threshold, when the predicted risk probability And uncertainty When , it indicates that the disaster risk caused by the coupling of multiple source factors has entered the high-risk range and is considered a high-risk area; When the above conditions are met at the same time, a high-priority warning signal is triggered, and spatial visualization output and alarm prompts are performed.

10. The intelligent early warning system for urban ground collapse based on multi-source factor fusion is characterized by: include: Data acquisition unit, used for multi-temporal SAR image data acquisition, multi-source remote sensing data acquisition and GIS spatial data acquisition; A data preprocessing unit, which preprocesses the multi-source data acquired by the data acquisition unit; Based on the InSAR deformation monitoring unit, two time series synthetic aperture radar interferometry techniques, PS-InSAR and SBAS-InSAR, are used to process multi-temporal SAR image data, calculate deformation values, and form deformation time series. The consistency of the deformation time series extracted by the two processing methods is verified to generate a settlement map. The adaptive multi-source factor fusion modeling unit constructs a node feature set containing dynamic and static factors based on multi-source remote sensing data, GIS spatial data, and InSAR deformation results. It also constructs a multi-channel weighted spatiotemporal graph structure with the monitoring area grid cells as nodes, and uses the spatiotemporal graph neural differential attention network ST-GNODEA to achieve end-to-end multimodal factor fusion and continuous time modeling. The model performance evaluation and optimization unit monitors the prediction performance indicators of the spatiotemporal graph neural differential attention network model under the sliding window in real time. When the performance drops below a threshold, the online update mechanism is triggered to use newly collected data for small batch incremental training and dynamically adjust the factor attention weights and neural ordinary differential equation parameters. The dual-channel analysis unit performs early warning analysis based on the subsidence rate threshold set by InSAR deformation and the comprehensive risk probability of ST-GNODEA risk probability; The early warning unit issues warnings for areas that exceed the sedimentation rate threshold and have a high predicted risk probability.

Images