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

Abstract

The present application relates to a kind of urban ground collapse intelligent early warning method and system based on multi-source factor fusion, belong to urban disaster early warning technical field.Adopt PS-InSAR and SBAS-InSAR respectively to process and calculate deformation value, the deformation time series extracted by two kinds of processing methods are verified and analyzed, and the vector results qualified after verification are generated to generate subsidence chart;Set subsidence rate threshold to preliminarily ground collapse early warning identification according to deformation;Deformation time series and the multi-source data collected are divided into dynamic factor and static factor, construct the multi-channel weighted space-time graph structure with the grid unit of monitoring area as node, utilize space-time graph neural differential attention network to carry out multi-source data fusion and predict comprehensive risk probability;Carry out double-channel fusion research and judge.The present application fully depicts the nonlinear coupling and space-time dynamic relationship between disaster-causing factors in fusion process, can realize high-precision, low false alarm, strong generalization prediction to urban ground collapse risk.

Description

TECHNICAL FIELD

[0001] The present application relates to a kind of urban ground collapse intelligent early warning method and system based on multi-source factor fusion, in particular to a kind of ground collapse early warning method and system based on time sequence synthetic aperture radar interferometry (InSAR) and multi-source remote sensing data fusion, combined with spatiotemporal graph neural differential attention network, belong to urban disaster early warning and artificial intelligence technical field. BACKGROUND

[0002] With the acceleration of urbanization, the development intensity of underground space is increasing, and subway, underground pipe gallery, water supply and drainage pipe network and large-scale infrastructure are densely distributed, making the urban geological environment more complex and fragile. Influenced by the coupling effect of persistent or extreme rainfall, underground pipe network leakage, foundation softening, traffic dynamic load and geological structure, ground subsidence and collapse events occur frequently in some cities, which poses a serious threat to public safety, traffic operation and infrastructure stability.

[0003] At present, synthetic aperture radar interferometry (InSAR) technology, especially permanent scatterer InSAR (PS-InSAR) and small baseline set InSAR (SBAS-InSAR) time series methods, has been widely used in millimeter-level ground deformation monitoring. However, in the complex urban topography and variable climate environment, single InSAR technology is easily affected by atmospheric residual error, phase unwrapping error and data missing, and the monitoring results have uncertainty. In addition, most of the existing multi-source data fusion methods only perform simple weighted combination on the prediction results at the model output layer, which is difficult to deeply depict the nonlinear coupling relationship between multi-modal factors and their spatiotemporal dynamic evolution pattern, and there are still significant deficiencies in capturing hidden and progressive collapse risks.

[0004] In the field of urban ground collapse early warning, another key problem of multi-source fusion modeling is the processing of irregularly sampled time series data and the modeling of spatial dependence. Existing methods generally use equal time interval sequence modeling assumption, which cannot effectively utilize SAR images, meteorological monitoring, geological survey and other data with different sources and inconsistent sampling periods. At the same time, there are complex spatial topological relationships in urban geological environment, and traditional convolution or fully connected structure cannot directly model this non-Euclidean spatial structure, resulting in insufficient description of cross-regional risk transmission and neighborhood interaction.

[0005] In addition, most of the existing early warning systems rely on a single risk determination channel, lack of dual verification mechanism from physical deformation evidence and multi-factor coupling mechanism, and are prone to miss and false alarm in hidden collapse or short-term sudden events. SUMMARY

[0006] The present application aims at the deficiencies of the existing urban ground collapse early warning method in multi-source data fusion, spatio-temporal dynamic relationship modeling, irregular sampling processing and risk judgment mechanism, and provides an urban ground collapse intelligent early warning method based on multi-source factor fusion, realizes high-precision, low-false alarm and strong generalization prediction of urban ground collapse risk, and provides reliable technical support for urban safety management.

[0007] The technical scheme adopted by the present application is:

[0008] The urban ground collapse intelligent early warning method based on multi-source factor fusion comprises the following steps:

[0009] S1. Collecting multi-temporal SAR remote sensing image data of the same place, using two time series synthetic aperture radar interferometry techniques PS-InSAR and SBAS-InSAR to process and calculate deformation values respectively, and extracting respective deformation time series;

[0010] S2. Performing spatial consistency verification and time sequence feature cross verification analysis on the deformation time series extracted by the two processing methods, importing the qualified PS-InSAR processing vector results into ArcGIS to draw a time sequence InSAR ground deformation rate map, i.e. generating a subsidence map;

[0011] S3. Setting a subsidence rate threshold to perform preliminary ground collapse early warning identification according to the subsidence map obtained by InSAR deformation monitoring;

[0012] S4. Collecting multi-source remote sensing data and GIS spatial data, dividing dynamic factors and static factors for the InSAR deformation time series and the collected multi-source data, constructing a multi-channel weighted spatio-temporal graph structure with the monitoring area grid unit as the node, and using a spatio-temporal graph neural differential attention network (Spatio-Temporal Graph Neural ODE with Attention, ST-GNODEA) to fuse multi-source data and predict the comprehensive risk probability;

[0013] In the spatio-temporal graph neural differential attention network, first, the monitoring area is rasterized into nodes, and the irregular sampling multi-source time series data of each node is continuously time-modeled by using a neural differential equation, the attention weight of each node and its neighbor nodes (including itself) on the dynamic factor feature is calculated through a graph attention network with a self-loop connection, then the hidden state vector of the dynamic factor and the static factor representation input factor-level multi-modal attention module, realize the importance allocation and nonlinear coupling modeling between factors, the fused node spatio-temporal features are output through a fully connected prediction layer the collapse risk probability of the node, and the risk of each node forms a risk distribution map facing the whole world;

[0014] S5. Perform dual-channel fusion analysis, combining the InSAR deformation threshold with the comprehensive risk probability predicted by ST-GNODEA to identify ground collapse early warning.

[0015] In the above method, the permanent scatterer PS-InSAR processing in step S1 includes registration, interferometry, and deformation information extraction; the small baseline set InSAR processing SBAS-InSAR includes registration, interferometry, phase unwrapping, atmospheric residual filtering, and deformation information extraction.

[0016] The spatial consistency verification described in step S2 first involves grid matching, dividing the study area into grid cells of equal size, extracting the average settlement rate from PS-InSAR and SBAS-InSAR respectively, and then calculating the rate difference Δ between the two methods within each grid cell. v :

[0017] ,

[0018] If Δ v ≤5mm / a and spatial correlation coefficient R 2 If the value is ≥0.85, it is considered a consistent region; otherwise, it is marked as a "region to be verified" and requires manual verification.

[0019] The aforementioned temporal feature cross-validation analysis selects PS points and SBAS pixel clusters with a coherence coefficient greater than 0.3 and pixel clusters that cover a continuous region, while ensuring they are from the same geographical location, and extracts their respective deformation time series data D. PS (t) and D SBAS (t). These data represent the deformation values ​​of the point at the corresponding SAR image acquisition time. For visualization, these discrete data points are connected to create a deformation-time relationship curve. The dynamic time warping (DTW) algorithm is used to calculate the deformation values ​​of the two sequences. PS (t) and D SBAS Similarity of (t):

[0020] ,

[0021] Similarity is the similarity index of the deformation time series, with a value range of [0, 1]. The larger the value, the more consistent the evolution trend of the two sequences. DTW_Distance is the minimum cumulative path distance between two points in the sequence calculated by the dynamic time warping algorithm. len is the sequence length. When the similarity is > 0.9, the time series evolution is considered to be consistent.

[0022] The settlement rate threshold of the above step S3 is set to be annual settlement rate > 30 mm / a, 30-day cumulative settlement > 10 mm, and daily average settlement > 2 mm, and once the threshold is exceeded, the risk marking mechanism is started.

[0023] In order to solve the problem that the existing multi-source data fusion method only performs simple weighted combination at the model output layer and is difficult to capture the nonlinear coupling of multi-modal factors and their spatio-temporal dynamic relationship, the present application proposes a spatio-temporal graph neural differential attention network (Spatio-Temporal Graph Neural ODE with Attention, ST-GNODEA), which grids the monitoring area into nodes, constructs a graph structure combining spatial topology and infrastructure connectivity, uses a neural differential equation to model the irregularly sampled multi-source time series data in continuous time, and realizes end-to-end interpretable multi-source data fusion through a graph attention mechanism and multi-modal factor attention, thereby improving the accuracy and generalization ability of early warning.

[0024] (1) Graph structure construction:

[0025] The monitoring area is divided into N nodes v i The node features include:

[0026] Dynamic factors: InSAR deformation time series , rainfall series , etc.

[0027] Static factors: soil thickness, pipe network density, terrain slope, land use type, etc.

[0028] Multi-channel weighted edges are established between nodes , m is the edge type, ij represents an edge from node i to node j , and the edge weight is calculated by combining spatial distance, geological similarity and infrastructure information.

[0029] (2) Continuous time modeling:

[0030] For each node, a neural differential equation (Neural ODE) is used to map discrete observations to continuous time hidden states:

[0031] ,

[0032] where is the hidden state of node v i at time t, f θ is a trainable neural network, and x i( t ) represents a node i At any moment t The input feature vector is used to adapt to situations where the sampling interval of sensors such as InSAR is uneven.

[0033] (3) Spatial information interaction:

[0034] Introduce a graph attention network (GAT) in the hidden state to calculate the attention weights of neighboring nodes:

[0035] ,

[0036] Where 'a' is the attention weight, , The vector representation of node features after linear transformation. Leaky Re LU The Leaky ReLU activation function allows inputs less than 0 to retain a certain gradient, thus preventing dead neurons. Normalize all neighboring nodes (including itself) so that the sum of their weights is 1. e ij It is a node i With nodes j The feature vectors of the edges between them;

[0037] Aggregate information:

[0038] ,

[0039] in For nodes i The neighborhood group, For attention weights, The weight matrix is ​​a linear transformation matrix. node j The original feature vectors are obtained by spatial aggregation through a graph attention network. As a dynamic factor latent state, it enters the factor-level attention calculation module together with the static factor features;

[0040] (4) Multimodal factor attention fusion:

[0041] ① Node input and dynamic hidden state

[0042] remember For nodes i At any moment t The dynamic hidden state vector obtained through spatial information interaction encodes neighbor nodes and edge weight information; let be... For nodes i The m A static factor vector, and dynamic factors in time. iThe representation of the node can be obtained directly from the partial components of the node

[0043] ②Factor-level attention calculation

[0044] For each node i , at time t , the importance weight of each factor thereof is calculated:

[0045] ,

[0046] where u is the attention score vector, tanh is the hyperbolic tangent activation function, W d is the weight matrix for mapping dynamic factors, is the node i , at time t , the dynamic hidden state vector obtained through spatial information interaction, W s is the weight matrix for mapping static factors, is the node i , the static factor vector of the m th static factor; m is the number representing the current factor mode, m' represents the index to the set of all modal factors, and i is the traversal variable in summation;

[0047] ③Fusion feature construction

[0048] The dynamic and static information are weighted using the factor weight to obtain the fused feature:

[0049] ,

[0050] where is the fused node modal feature vector, W m is the feature fusion weight matrix of the m th factor, is the attention weight of the t th factor of the node i at time m .

[0051] ④Risk prediction

[0052] The fused feature vector is input to the prediction layer, the fused node feature is input to the prediction layer, and the collapse risk probability of the node is output through the sigmoid function:

[0053] ,

[0054] where ​is the node risk probability, and σ is a sigmoid activation function, W p is the risk prediction weight vector, b p is the prediction bias term. When the system gives an early warning in grid units, the node risk probability can be aggregated by weight.

[0055] (5) After prediction, uncertainty estimation and online adaptive update can also be performed:

[0056] Introduce the confidence estimation method based on evidence theory to calculate the prediction uncertainty , and combine the risk probability and uncertainty threshold to trigger the early warning to avoid false positives under low confidence: Introduce the uncertainty estimation method based on Monte Carlo Dropout and Bayesian neural network in the output layer:

[0057] ,

[0058] wherein the prediction uncertainty, is the mth sampling prediction result, and M is the number of Monte Carlo sampling, is the mean prediction.

[0059] An incremental learning strategy combining elastic weight consolidation (EWC) and experience replay (Experience Replay) is used for online adaptive update:

[0060] a. When the model performance decreases by more than a set threshold in the short-term window (24 days), trigger small batch online update;

[0061] b. Neural ODE parameters are updated by small step optimization, while EWC regularization term protects important parameters to prevent forgetting historical knowledge.

[0062] The invention introduces a dual-channel fusion research and judgment mechanism in the early warning decision, which is jointly acted by the InSAR deformation threshold channel and the ST-GNODEA comprehensive risk probability channel, to realize the combination of rapid response and accurate prediction of multi-layer protection.

[0063] The dual-channel research and judgment rule in the above step S5 is:

[0064] If the InSAR subsidence rate value exceeds the set threshold, it indicates that the physical deformation has reached a critical state, and is considered as a potential risk;

[0065] Set the ST-GNODEA predicted risk greater than 0.8 as the high risk threshold, when the comprehensive risk probability and , it indicates that the multi-source factor coupled disaster risk enters the high-risk interval, and is considered as a high-risk area.

[0066] When the above conditions are met at the same time, a high-priority warning signal is triggered, and spatial visualization output and warning prompts are performed.

[0067] Another object of the present application is to provide a city ground collapse intelligent early warning system based on multi-source factor fusion, comprising:

[0068] A data acquisition unit is configured to acquire multi-temporal SAR image data, multi-source remote sensing data and GIS spatial data.

[0069] A data preprocessing unit is configured to preprocess the multi-source data acquired by the data acquisition unit, including SAR image registration, interference processing, phase unwrapping, atmospheric delay correction, multi-source data coordinate unification and resampling, missing value filling and outlier removal, etc., to generate a multi-source factor dataset aligned in time and space.

[0070] An InSAR deformation monitoring unit is configured to use two time series synthetic aperture radar interferometry techniques PS-InSAR and SBAS-InSAR to process multi-temporal SAR image data, calculate deformation values, form a deformation time series, and perform consistency verification on the deformation time series extracted by the two processing methods to generate a subsidence map.

[0071] An adaptive multi-source factor fusion modeling unit is configured to construct a node feature set containing dynamic factors and static factors based on multi-source remote sensing data, GIS spatial data and InSAR deformation results, construct a multi-channel weighted spatiotemporal graph structure with grid cells in the monitoring area as nodes, and use a spatiotemporal graph neural differential attention network ST-GNODEA to realize end-to-end multi-modal factor fusion and continuous time modeling.

[0072] A model performance evaluation and optimization unit is configured to monitor the prediction performance indicators of the spatiotemporal graph neural differential attention network model under a sliding window in real time, trigger an online update mechanism when the performance decreases by more than a threshold, and use newly acquired data for small-batch incremental training and dynamically adjust factor attention weights and Neural ODE parameters.

[0073] A dual-channel analysis and judgment unit is configured to perform early warning analysis and judgment based on the comprehensive risk probability of the subsidence rate threshold set by InSAR deformation and the risk probability of ST-GNODEA.

[0074] An early warning unit is configured to issue a warning for areas that exceed the subsidence rate threshold and have a large predicted risk probability.

[0075] The present application has the following advantages:

[0076] (1) Using two kinds of time series synthetic aperture radar interferometric technique to monitor the continuous deformation of urban area, realize the millimeter level precision ground subsidence time series extraction, provide 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 topography, through the complementary analysis of PS-InSAR and SBAS-InSAR, the limitations of single method are eliminated, and the reliability of deformation monitoring results is improved.

[0077] (2) Multi-source remote sensing and environmental factor fusion modeling, fusion InSAR deformation, rainfall time series, geological parameters, underground pipe network density, DEM, slope, land use and other multi-source factors, construct the spatially aligned raster sample library, realize the collaborative representation of dynamic factors and static factors. This method breaks through the blind area of traditional model in the description of multi-factor coupling disaster mechanism, combines GIS spatial analysis and normalization processing, realizes the unified coding and seamless fusion of heterogeneous data, and provides high-quality input for subsequent intelligent modeling.

[0078] (3) The neural ordinary differential equation is combined with the graph attention network and the factor attention mechanism to realize continuous time modeling of multi-source data and explainable factor level fusion. The graph attention mechanism introduces spatial proximity and infrastructure connectivity, captures the propagation path and neighborhood interaction of surface instability, and fully describes the nonlinear coupling and spatiotemporal dynamic relationship between disaster-causing factors.

[0079] (4) Online adaptive update ensures long-term stability, adopts the incremental learning strategy combining elastic weight consolidation and experience replay, triggers small batch online update when detecting model performance decline, ensures rapid absorption of new knowledge and retention of historical knowledge. This strategy can adapt to different urban geological conditions and climate change environment, significantly reducing the risk of early warning failure caused by data distribution drift.

[0080] (5) Multi-dimensional confidence assessment improves decision reliability, the prediction distribution is obtained by multiple sampling estimation of prediction distribution through the combination of Bayesian method and Monte Carlo Dropout, and the risk probability and uncertainty index are obtained. When the uncertainty is higher than the set threshold, the system automatically issues a "focus on" prompt to avoid low confidence prediction triggering early warning directly, and realizes the controllability and robustness of early warning quality.

[0081] (6) The double-channel early warning mechanism realizes synchronous verification of physical deformation signals and disaster-causing mechanisms through the fusion of the deformation threshold channel and the ST-GNODEA risk probability channel, thereby significantly improving the early warning accuracy and reducing the false alarm rate. The deformation threshold channel focuses precisely and directly captures the key evidence of ground instability; the ST-GNODEA risk probability channel identifies gradual hazards through multi-source factor fusion. The two channels work together to effectively suppress false alarms and false alarms caused by construction vibration, ensuring monitoring accuracy and enhancing the generalization ability and long-term stability of the model under different urban and environmental conditions. BRIEF DESCRIPTION OF DRAWINGS

[0082] Figure 1 is a flowchart of the method of the present application;

[0083] Figure 2 is a surface deformation monitoring result of a certain district in Shenzhen City based on time series InSAR of an embodiment of the present application, a. PS-InSAR processing result, b. SBAS-InSAR processing result;

[0084] Figure 3 is a ground collapse risk probability calculation result of a certain district in Shenzhen City by the ST-GNODEA model of an embodiment of the present application. DETAILED DESCRIPTION

[0085] The present application will be further described below in combination with specific embodiments.

[0086] Embodiment 1: The urban ground collapse intelligent early warning method based on multi-source factor fusion comprises the following steps:

[0087] S1. Collect multi-temporal SAR remote sensing image data at the same location, and use two time series synthetic aperture radar interferometry techniques PS-InSAR and SBAS-InSAR to process and calculate deformation values, and extract respective deformation time series:

[0088] First, use two time series synthetic aperture radar interferometry techniques - permanent scatterer InSAR and small baseline set InSAR to continuously monitor the deformation of urban areas, achieve millimeter-level precision ground subsidence time series extraction, and provide deformation prior information for subsequent collapse risk analysis.

[0089] The core principle of InSAR is to use image pairs obtained by SAR satellites at different times to extract the relative deformation information of ground objects in the radar wave propagation direction through interference processing. In the present application, the combination of PS-InSAR and SBAS-InSAR techniques overcomes the problem that traditional D-InSAR is greatly affected by atmospheric noise and phase unwrapping in complex urban terrain.

[0090] The PS-InSAR processing includes registration, interferometric processing and deformation information extraction; the SBAS-InSAR processing includes registration, interferometric processing, phase unwrapping, atmospheric residual filtering and deformation information extraction.

[0091] (1) Collecting multi-temporal SAR image data, performing accurate registration:

[0092] Taking a continuous time series of SAR images from a satellite platform, selecting a master image, establishing a master-slave relationship according to a set baseline threshold, and registering the images to ensure comparability between the same ground object pixel pairs.

[0093] (2) Interferometric processing:

[0094] Performing interferometric processing on each pair of interferometric images (i.e. SAR image pairs composed of master and slave images) to generate an interferogram, as follows:

[0095] ,

[0096] Wherein is the flat earth phase caused by the earth's curvature, is the phase caused by the terrain undulation, is the deformation phase, is the phase caused by atmospheric propagation delay, is the system and registration error. By using a high-precision digital elevation model, the flat earth phase caused by the earth's curvature and the phase caused by the terrain undulation can be eliminated, thereby obtaining a differential interferogram.

[0097] (3) Filtering the original differential phase in the spatial and temporal domains to improve the phase signal-to-noise ratio, and then performing phase unwrapping processing to restore continuous physical phase quantities, as follows:

[0098] ,

[0099] Wherein is the phase unwrapping function, i.e. an algorithm that restores the wrapped phase to a continuous phase. Common represents the wrapped phase, i.e. the phase value limited in the range of (−π, π] after interferometric processing. Since the phase measured by radar has periodicity, it can only measure the relative change within one wavelength range, and beyond the range it will be "wrapped".

[0100] (4) Deformation quantity calculation and LOS conversion: using the unwrapped interferometric phase, calculating the cumulative deformation quantity along the radar line of sight (LOS), as follows:

[0101] ,

[0102] Wherein is the deformation along the line-of-sight direction, λ is the radar wavelength, is the unwrapped interferometric phase.

[0103] is the radar measurement is the change of slant range from satellite to ground point. Since ground subsidence is mainly vertical movement, it is necessary to project θ to the vertical direction to obtain the vertical deformation :

[0104] ,

[0105] where θ is the incidence angle of radar.

[0106] (5) PS-InSAR and SBAS-InSAR deformation time series extraction:

[0107] PS-InSAR: Identify point targets with stable scattering characteristics (i.e. permanent scatterers) through statistical analysis, and establish long time series of coherent image pairs for them. Through differential interference, orbit error fitting and atmospheric noise modeling, the deformation time series D(t) of the target point is extracted, which has high time resolution and noise suppression capability.

[0108] SBAS-InSAR constructs multiple interferogram sets with small spatial and temporal baselines, forming underdetermined equations, and reconstructing the deformation time series of each pixel through singular value decomposition (SVD), which is suitable for wide-area and non-dense scatterer scenes.

[0109] S2. Spatial consistency verification and time series feature cross-validation analysis of deformation time series extracted by the two processing methods, and import the qualified PS-InSAR processed vector results into ArcGIS to draw the time series InSAR ground deformation rate map, i.e. generate the subsidence map:

[0110] Through the complementary analysis of PS-InSAR and SBAS-InSAR, the limitations of single method are eliminated, and the reliability of deformation monitoring results is improved.

[0111] Spatial consistency verification, first grid matching, divide the study area into 500m×500m grid cells, extract the average subsidence rate of PS-InSAR and SBAS-InSAR respectively, and then calculate the rate difference Δ v of the two methods in each grid:

[0112] ,

[0113] where v ​PS the mean of the subsidence rate extracted by PS-InSAR, v SBAS the mean of the subsidence rate extracted by SBAS-InSAR, if Δ v ≤ 5mm / a and the spatial correlation coefficient R 2 ≥ 0.85, it is determined as a consistent area; otherwise, it is marked as a "to-be-verified area" and verified by manual processing.

[0114] cross-validation analysis of time series characteristics, select the PS points and SBAS pixel clusters with a coherence coefficient greater than 0.3 and a pixel cluster covering a continuous area, and ensure that they are in the same geographical location, extract their respective deformation time series data D PS (t) and D SBAS (t). These data represent the deformation values of the point at the time points when the corresponding SAR images are acquired, and when visualized, these discrete data points are usually connected to draw a deformation-time relationship curve. The similarity between the two sequences D PS (t) and D SBAS (t) is calculated by the dynamic time warping (DTW) algorithm:

[0115] ,

[0116] where Similarity is the similarity index of the deformation sequence, with a value range of [0, 1], and the larger the value, the more consistent the evolution trend of the two sequences. DTW_Distance is the minimum path cumulative distance between two points calculated by the dynamic time warping algorithm. D PS deformation time series extracted by PS-InSAR. D SBAS deformation time series extracted by SBAS-InSAR. len is the length of the sequence, and a similarity > 0.9 is considered to be consistent in time series evolution.

[0117] S3. Set the subsidence rate threshold to preliminarily identify ground collapse warning based on the subsidence map obtained by InSAR deformation monitoring:

[0118] Set the warning threshold (annual subsidence rate > 30 mm / a, 30-day cumulative subsidence > 10 mm, daily average subsidence > 2 mm), and once the InSAR result exceeds the threshold, the risk marking mechanism is started.

[0119] S4. Perform multi-source remote sensing data collection and GIS spatial data collection, divide the InSAR deformation time series and the collected multi-source data into dynamic factors and static factors, construct a multi-channel weighted space-time graph structure with the monitoring area grid unit as the node, and use the Spatio-Temporal Graph Neural ODE with Attention (ST-GNODEA) to perform multi-source data fusion and predict the comprehensive risk probability:

[0120] The method divides the monitoring area into N nodes v i The node features are composed of dynamic factors and static factors, and the spatial and functional relationship between nodes is described by multi-channel weighted edges.

[0121] (1) Factor feature system:

[0122] The application fuses InSAR deformation data and multiple external remote sensing and geological factors by constructing a space-time sample to construct a ground collapse risk prediction data set. The factors involved include:

[0123] ① Surface deformation

[0124] The surface cumulative subsidence extracted by the time series InSAR technology reflects the compression, displacement or instability trend of the soil body during the monitoring period. This factor is a direct deformation precursor of collapse risk, especially sensitive to gradual soil loss and sudden karst collapse. The data is input in the form of a grid time series, and the spatial resolution is consistent with the SAR image.

[0125] ② Rainfall time series

[0126] The cumulative rainfall data generated based on meteorological satellites and ground rain stations. Rainfall is the core dynamic factor that induces collapse: short-term heavy rain increases soil saturation and pore water pressure, reducing shear strength; long-term rainfall exacerbates foundation softening and karst erosion. The data needs to be constructed in the form of a time series grid synchronized with InSAR to capture the lag effect of rain intensity-deformation response.

[0127] ③ Underground pipe network density

[0128] The unit area underground pipeline calculated using the city infrastructure GIS database. The high pipe network density area has a significantly increased risk of soil loss due to pipe aging and cracking, leakage and erosion, especially in areas with dense water supply and drainage pipe networks. This factor is input in the form of a static grid layer, and the spatial resolution is consistent with the early warning system.

[0129] ④ Road density

[0130] Road coverage area kernel density data generated based on road network vector data. Road dense areas accelerate roadbed fatigue due to repeated traffic load, and underground pipe gallery construction may leave cavities that can easily cause collapse. This factor needs to be associated with heavy vehicle traffic frequency data to improve prediction accuracy.

[0131] ⑤Land use type

[0132] Land cover class coding derived from remote sensing classification and planning data. Different land use types imply different levels of disaster exposure.

[0133] ⑥Soil layer thickness

[0134] Soft soil layer vertical thickness obtained by geological drilling and electromagnetic prospecting. When the thickness of soft layer is greater than 5 meters, the soil is prone to liquefaction or compression settlement under the action of seepage and vibration, significantly reducing the critical deformation threshold. This factor is input in the form of spatial interpolation grid, with resolution matching the system.

[0135] ⑦Distance to major river

[0136] Euclidean distance from raster cell to nearest river calculated based on hydrological network GIS data. Areas close to rivers are strongly affected by riverbank erosion and groundwater level fluctuations, especially lateral erosion of riverbed during flood season, which can trigger slope collapse. Data need to be combined with river width and flow rate for weighted optimization.

[0137] ⑧Digital Elevation Model (DEM)

[0138] Surface elevation grid data derived from LiDAR and stereoscopic aerial photography. Low-lying areas will accelerate the softening of the foundation due to the accumulation of catchment.

[0139] ⑨Slope factor

[0140] Surface slope angle calculated by Sobel operator from DEM data. Areas with slope > 10° are prone to slope instability due to shear stress concentration; at the same time, high slope areas have faster rainfall runoff speed, which exacerbates surface soil erosion. This factor needs to be coupled with lithology data to assess the anti-sliding force.

[0141] Each factor is processed by rasterization through GIS means, with uniform resolution. All factor data is normalized:

[0142] ,

[0143] For each grid cell, a time series sample is constructed, with the true label coming from the historical record of collapse event points, marked as 1, and the non-collapse area as 0.

[0144] (2) Multi-source factor data fusion:

[0145] ① Graph structure construction:

[0146] Node definition: Each node represents a permanent scattering point (PS point) or a grid cell, and both dynamic and static factors correspond to it. The dynamic factors include InSAR deformation time series and rainfall time series, and the static factors include underground pipe network density, road density, land use type, soil thickness, distance from river, DEM, and slope.

[0147] Edge relationship construction: Establish 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 by spatial distance, geological similarity, and infrastructure network information to form a multi-scale spatial topology.

[0148] Multi-source factors and graph structure are jointly modeled in a unified end-to-end framework.

[0149] ② Continuous time modeling:

[0150] For dynamic factors such as InSAR deformation time series and rainfall time series, Neural ODE is used to represent the continuous time hidden state h i ( t ):

[0151] ,

[0152] where is the node v i The hidden state at time t, f θ is a trainable neural network, and x i ( t ) represents the input feature vector of node i at time t This method can adapt to the case where the sampling interval of sensors such as InSAR is not uniform; Neural ODE allows direct calculation of state evolution on irregular time intervals, thus adapting to different sampling frequencies of different data sources, and can perform high-precision interpolation or extrapolation prediction when new data arrives.

[0153] ③ Spatial information interaction:

[0154] Introduce graph attention network (GAT) on the hidden state to calculate the attention weight of the neighbor nodes. On the continuous time hidden state, use GAT to perform message passing between nodes on the spatial and functional topology:

[0155] ,

[0156] where a is the attention weight, , The vector representation of node features after linear transformation. Leaky Re LU The Leaky ReLU activation function allows inputs less than 0 to retain a certain gradient, thus preventing dead neurons. Normalize all neighboring nodes (including itself) so that the sum of their weights is 1. e ij It is a node i With nodes j The feature vector of the edge between them;

[0157] Multi-channel edge structures can simultaneously capture multiple risk propagation paths, such as groundwater seepage and diffusion caused by pipe ruptures, and runoff paths of slope rainfall infiltration.

[0158] Aggregate information:

[0159] ,

[0160] in For nodes i The neighborhood group, For attention weights, The weight matrix is ​​a linear transformation matrix. node j The original feature vectors are obtained by spatial aggregation through a graph attention network. As a dynamic factor hidden state, it enters the factor-level attention calculation module together with the static factor features.

[0161] ④ Multimodal factor attention fusion:

[0162] a. Node Input and Dynamic Hidden State

[0163] remember For nodes i At any moment t The dynamic hidden state vector obtained through spatial information interaction encodes neighbor nodes and edge weight information; let be... For nodes i The m A static factor vector, and dynamic factors in time. i The representation can be derived from A portion of the components is obtained directly.

[0164] b. Factor-level attention calculation

[0165] For each node i At any moment t The importance weights of each factor are calculated below:

[0166] ,

[0167] where u is the attention score vector, tanh is the hyperbolic tangent activation function, W d is the weight matrix for mapping dynamic factors, is the node i At time t is the dynamic hidden state vector obtained through spatial information interaction, W s is the weight matrix for mapping static factors, is the node i is the static factor vector of the m th factor; m is the number representing the current factor modality, m' represents the index for all modality factor sets, is the traversal variable in summation; by recording the changes of , the main driving factors of the predicted risk at a certain time can be intuitively explained, such as the significant increase in the weight of the rainfall factor during heavy rain and the long-term high weight of the pipe network factor in areas with dense underground pipe networks.

[0168] c. Fusion feature construction

[0169] The factor weight is used to weight the dynamic and static information to obtain the fused feature:

[0170] ,

[0171] where is the fused node modality feature vector, W m is the feature fusion weight matrix of the m th factor, is the attention weight of the t th factor of the node i at time m .

[0172] d. Risk prediction

[0173] The fused feature vector is input into the prediction layer, and the fused node feature is input into the prediction layer, and the node collapse risk probability node risk probability is output through the sigmoid function:

[0174] ,

[0175] where is the node risk probability, and σ is the Sigmoid activation function, W p is the risk prediction weight vector, b p is the prediction bias term. When the system warns in grid units, the node risk probability can be summarized according to the weight;

[0176] By recording the changes of , the main driving factors of the predicted risk at a certain moment can be intuitively explained, such as the significant increase in the weight of the rainfall factor during heavy rain and the long-term high weight of the pipe network factor in areas with dense underground pipe networks.

[0177] e. Prediction uncertainty estimation

[0178] To quantify the reliability of the prediction results, the invention introduces an uncertainty estimation method based on Monte Carlo Dropout and Bayesian neural networks in the output layer:

[0179] ,

[0180] where is the prediction uncertainty, is the prediction result of the m-th sampling, and M is the number of Monte Carlo sampling, is the mean prediction. 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 included in the "focus area", rather than directly triggering high-level warning, reducing false positives.

[0181] (3) ST-GNODEA model training and online optimization strategy:

[0182] a. Loss function design

[0183] A weighted binary cross-entropy loss is used, combined with sample imbalance to adjust the positive and negative class weights:

[0184] ,

[0185] where L is the loss value, representing the difference between the model prediction and the true label. N is the total number of samples. y i is the true label, taking the value of 0 or 1. is the probability value predicted by the model, representing the probability that the sample belongs to the positive class, taking the value range of [0, 1]. W pos is the positive class weight, used to adjust the importance of positive class samples in loss calculation. W neg is the negative class weight, used to adjust the importance of negative class samples in loss calculation.

[0186] b. Online learning and incremental update

[0187] To ensure that the system adapts to dynamically changing geological and environmental conditions, the invention introduces an elastic weight consolidation incremental learning strategy with experience replay under a single model architecture:

[0188] ,

[0189] where L new is the loss function for new data, F i is the information matrix diagonal element, is the optimal parameter for historical tasks, θ i is the differential equation parameter of the node i λ is the regularization coefficient.

[0190] Experience replay: Mix sampling new data with historical key samples, maintain the generalization ability of spatio-temporal features;

[0191] Update trigger condition: Trigger incremental update when the early warning accuracy in the sliding window drops by more than 2% or significant data distribution drift occurs;

[0192] Periodic calibration: Full parameter calibration is performed every 30 days.

[0193] c. Parameter optimization and training stability

[0194] When new monitoring data is obtained, update the model weights through gradient descent optimization algorithm

[0195] Use an improved version of the Adam optimizer, AdamW, introduce weight decay mechanism to prevent overfitting, limit the gradient norm to not exceed 1.0 to prevent gradient explosion, use FP16 precision to speed up training, while maintaining the numerical stability of FP32 precision.

[0196] d. Knowledge preservation and environmental adaptation

[0197] In terms of knowledge preservation, identify feature weights that contribute significantly to historical early warning events, and use L2 regularization to constrain their parameter changes; For multi-scenario adaptation, set independent output layers for different geological blocks and share the bottom feature extractor; For distribution alignment, use the maximum mean difference loss to reduce the feature distribution difference between the source domain and the new target domain, achieving smooth transfer learning.

[0198] e. Real-time evaluation of model performance

[0199] Multi-scale performance trend monitoring and adaptive maintenance, based on the SAR image acquisition cycle, construct a 24-day (short-term), 72-day (medium-term), and 144-day (long-term) multi-scale sliding window performance monitoring framework to capture both timeliness changes and seasonal trends.

[0200] Trend analysis method: Combine linear regression slope, Mann-Kendall monotonicity test and CUSUM change point detection to realize performance trend identification and mutation point discovery.

[0201] ​Hierarchical warning thresholds: short-term F1 score drop > 1% (mild), medium-term accuracy drop > 3% (significant), long-term recall drop > 5% (severe).

[0202] Adaptive maintenance strategy:

[0203] Mild: fast incremental fine-tuning;

[0204] Significant: experience replay + EWC incremental training;

[0205] Severe: rollback to stable model snapshot and full retraining.

[0206] This mechanism realizes dynamic monitoring and hierarchical maintenance of the performance of the ST-GNODEA model, ensuring the long-term stability of prediction accuracy.

[0207] This mechanism can realize long-term monitoring and adaptive optimization of the performance of a single model without relying on multiple sub-model comparisons, ensuring the prediction accuracy and stability of ST-GNODEA under data distribution changes, seasonal fluctuations, or sudden environmental disturbances.

[0208] S5. According to the settlement rate threshold set by InSAR deformation and the comprehensive risk probability of the ST-GNODEA model, the double-channel research and judgment warning is carried out:

[0209] The final output of the system is the collapse risk probability P i (t) of each grid, and the invention proposes the following double-channel research and judgment rules:

[0210] If the InSAR settlement value exceeds the set threshold, it indicates that the physical deformation has reached a critical state, which is considered as a potential risk;

[0211] Set the ST-GNODEA prediction risk greater than 0.8 as the high risk threshold, when the comprehensive risk probability and , directly trigger the high-risk warning;

[0212] When the above conditions are met, 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.

[0213] Table 1 Warning trigger conditions

[0214] .

[0215] Select 78 scenes of Sentinel-1 image data from June 2, 2022 to May 17, 2025, and perform PS-InSAR and SBAS-InSAR monitoring on a certain district in Shenzhen City to obtain the surface deformation data of the district as shown in Figure 2 .

[0216] The monitoring shows that the maximum subsidence rate in Futian District is-48.94 mm / a, with a precision error of ±3 mm, meeting the first-level monitoring standard of the Engineering Survey Specification (GB50026-2007). Meanwhile, it can be seen that the subsidence in Futian District has spatial aggregation, and the high-value area of subsidence presents a "double-core distribution".

[0217] Huanggang Port Area: The subsidence center is located within a 500-meter range around the Port Joint Inspection Building, which is related to the combined effect of the dynamic load of 23800 daily truck trips between Shenzhen and Hong Kong (2023 data) and the leakage of underground pipe network.

[0218] Urban core trunk roads: The subsidence rate under roads such as Shennan Boulevard and Binhe Boulevard exceeds-30 mm / a, and the main cause of induced subsidence may be the stress release of soil layer induced by shallow buried construction of subway tunnels.

[0219] Based on the fusion analysis of time-series InSAR monitoring and ST-GNODEA output as shown in Figure 3 , the median of the global ground collapse risk probability in this area is 0.48, which is significantly higher than the safety threshold. According to the InSAR monitoring results, it can be concluded that the subsidence effect of Huanggang Port is the most obvious, and there is a greater risk of collapse. The ST-GNODEA results show that the urban trunk roads such as Shennan Boulevard and Binhe Boulevard form a high-risk aggregation area, and the proportion of grids with a comprehensive risk probability of >0.8 is relatively high, but the risk of Huanggang Port, which is a serious subsidence area, is relatively low. This is because Huanggang Port is mainly engaged in import and export trade, and the impact of urban personnel life and commuting is small. If the non-residential area dominated by Huanggang Port is ignored, it will cause a misjudgment of the results. Through the double-channel monitoring of the threshold of time-series InSAR subsidence monitoring and the threshold of ST-GNODEA comprehensive risk, it is beneficial to improve the accuracy of urban ground collapse early warning.

[0220] This conclusion is verified through spatial correlation analysis with the kernel density distribution of historical collapse events. The results show that the Pearson correlation coefficient between the high-risk prediction area ( ) and the high kernel density hot area of historical collapse events (kernel density value >0.85) reaches r=0.91 (p<0.001), indicating that the proposed integrated model has better performance in identifying high-risk areas.

[0221] In summary, the accuracy verification of the early warning model shows that the Pearson correlation coefficient between the prediction results and the spatial distribution of historical collapses is 0.91 (p<0.001), the recall rate of high-risk areas is 92.3%, and the false positive rate is controlled below 5%. This empirical result fully shows that the ST-GNODEA model based on the fusion of time-series InSAR deformation and multi-source factor coupling mechanism has significant scientificity and engineering effectiveness in urban ground collapse early warning, and has obvious technical advantages and application value compared with traditional single model methods.

[0222] Embodiment 2: The urban ground collapse intelligent early warning system based on multi-source factor fusion realizes the whole-process closed-loop management from data acquisition, processing to risk early warning, and the specific function modules are as follows:

[0223] A data acquisition unit is configured to acquire multi-temporal SAR image data, multi-source remote sensing data, and GIS spatial data.

[0224] A data preprocessing unit is configured to preprocess the data acquired by the data acquisition unit.

[0225] An InSAR deformation monitoring unit is configured to process the multi-temporal SAR image data by using two time series synthetic aperture radar interferometry techniques PS-InSAR and SBAS-InSAR, calculate deformation values, form a deformation time series, and perform consistency verification on the deformation time series extracted by the two processing methods to generate a subsidence map.

[0226] An adaptive multi-source factor fusion modeling unit is configured to construct a node feature set containing dynamic factors and static factors based on multi-source remote sensing data, GIS spatial data, and InSAR deformation results, construct a multi-channel weighted spatiotemporal graph structure with grid cells in a monitoring area as nodes, and realize end-to-end multi-modal factor fusion and continuous time modeling by using a spatiotemporal graph neural differential attention network ST-GNODEA.

[0227] A model performance evaluation and optimization unit is configured to monitor the prediction performance indicators of the spatiotemporal graph neural differential attention network model under a sliding window in real time, trigger an online updating mechanism when the performance decreases by more than a threshold value, perform small-batch incremental training by using newly acquired data, and dynamically adjust factor attention weights and Neural ODE parameters.

[0228] A dual-channel research and judgment unit is configured to perform early warning research and judgment according to a comprehensive risk probability of a subsidence rate threshold set by InSAR deformation and a risk probability of the ST-GNODEA risk probability.

[0229] An early warning unit is configured to issue a warning for an area with a subsidence rate exceeding the threshold and a large predicted risk probability.

[0230] The above is a further description of the present application in combination with the embodiments, and the protection scope of the present application is not limited thereto.

Claims

1. A city ground collapse intelligent early warning method based on multi-source factor fusion, characterized in that, The steps include the following: S1. Collecting multi-temporal SAR remote sensing image data of the same place, using two time series of synthetic aperture radar interferometry techniques PS-InSAR and SBAS-InSAR to calculate the deformation values respectively, and extracting the respective deformation time series; S2. The deformation time series extracted by the two processing methods are verified for spatial consistency and cross-verified for time sequence characteristics, and the qualified PS-InSAR processing vector results are imported into ArcGIS to draw a time sequence InSAR ground deformation rate map, that is, a subsidence map is generated; S3. Setting a subsidence rate threshold to preliminarily identify ground collapse warning according to the subsidence map obtained by InSAR deformation monitoring; S4. Collecting multi-source remote sensing data and GIS spatial data, dividing dynamic factors and static factors for the InSAR deformation time series and the collected multi-source data, constructing a multi-channel weighted space-time graph structure with grid cells in the monitoring area as nodes, and using a space-time graph neural differential attention network ST-GNODEA to fuse multi-source data and predict a comprehensive risk probability; In the space-time graph neural differential attention network, first, the monitoring area is rasterized into nodes, and the irregularly sampled multi-source time series data of each node are continuously time modeled by using a neural differential equation. Through a graph attention network with a self-loop connection, the attention weight of each node and its neighbor nodes on the dynamic factor feature is calculated. Then, the hidden state vector of the dynamic factor and the static factor representation input factor-level multi-modal attention module to realize the importance allocation and nonlinear coupling between factors. The fused node space-time features are output through a fully connected prediction layer to output the collapse risk probability of the node. The risk of each node forms a risk distribution map facing the whole world; S5. Dual-channel fusion analysis, combined with the InSAR deformation threshold and the comprehensive risk probability predicted by ST-GNODEA to identify ground collapse warning.

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

3. The urban ground collapse intelligent early warning method based on multi-source factor fusion according to claim 1, characterized in that, The spatial consistency verification in step S2 is performed by grid matching, dividing the study area into grid cells of the same size, extracting the average subsidence rate of PS-InSAR and SBAS-InSAR respectively, and then calculating the rate difference Δ of the two methods in each grid v : ， If Δ v ≤ 5 mm / a and the spatial correlation coefficient R 2 ≥ 0.85, it is determined as a consistent region; otherwise, it is marked as a "to-be-verified region" and verified by manual processing.

4. The urban ground collapse intelligent early warning method based on multi-source factor fusion according to claim 1, characterized in that, The time series cross-validation analysis in step S2 selects the coherence coefficient greater than 0.3 and the pixel cluster needs to cover the continuous area, and at the same time ensures that it is the PS point and the SBAS pixel cluster in the same geographical position, and extracts their respective deformation time series data D PS (t) and D SBAS (t), the similarity between the two sequences D PS (t) and D SBAS (t) is calculated by the dynamic time warping algorithm: ， Wherein Similarity is the similarity index of the deformation time series, the value range is [0, 1], the larger the value, the more consistent the evolution trend of the two sequences, DTW_Distance is the minimum path cumulative distance between two points of the sequence calculated by the dynamic time warping algorithm, and len is the length of the sequence. When the similarity is greater than 0.9, the time sequence evolution is considered to be consistent.

5. The urban ground collapse intelligent early warning method based on multi-source factor fusion according to claim 1, characterized in that, The subsidence rate threshold of step S3 is set to annual subsidence rate > 30 mm / a, 30-day cumulative subsidence > 10 mm, and daily average subsidence > 2 mm. Once the threshold is exceeded, the risk marking mechanism is started.

6. The urban ground collapse intelligent early warning method based on multi-source factor fusion according to claim 1, 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 The node features include: Dynamic factors: InSAR deformation time series, rainfall sequence; Static factors: soil thickness, pipe network density, terrain slope, land use type; Inter-node multi-channel weighted edge establishment , m is the edge type, ij represents an edge from node i to node j, and the edge weight is calculated based on spatial distance, geological similarity, and infrastructure information (2) Continuous time modeling: For each node, the discrete observations are mapped to continuous-time hidden states using neural ordinary differential equations: ， in For nodes v i In the hidden state at time t f θ For a trainable neural network, x i ( t ) represents a node i At any moment t The input feature vector; (3) Spatial information interaction: Introduce graph attention network on hidden state to calculate attention weight of neighbor nodes: ， where a is the attention weight, , is the vector representation of the node feature after linear transformation, Leaky Re LU is the Leaky ReLU activation function, which allows inputs less than 0 to retain a certain gradient, avoiding dead neurons, normalizes all neighbor nodes including itself, making the weight sum equal to 1, e ij is the edge feature vector between node i and node j ; Information aggregation: ， in For nodes i The neighborhood group, For attention weights, The weight matrix is ​​a linear transformation matrix. For nodes j The original feature vectors are obtained by spatial aggregation through a graph attention network. As a dynamic factor latent state, it enters the factor-level attention calculation module together with the static factor features; (4) Multi-modal factor attention fusion: ① Node input and dynamic hidden state, remember For nodes i At any moment t The dynamic hidden state vector obtained through spatial information interaction encodes neighbor nodes and edge weight information; let be... For nodes i The m A static factor vector, and dynamic factors in time. i The representation can be derived from A portion of the components is obtained directly; ② Factor-level attention calculation, For each node i At time t The importance weight of each factor is calculated at time ， where u is the attention score vector, tanh is the hyperbolic tangent activation function, W d is the weight matrix mapping dynamic factors, is the node i at time t is the dynamic hidden state vector obtained through spatial information interaction, W s is the weight matrix mapping static factors, is the node i is the mth m static factor vector, m is the number representing the current factor modality, m' represents the index to the set of all modality factors, is the traversal variable in summation; ③ Fusion feature construction, Weight the dynamic and static information using factor weights to get the fused features: ， in The fused node modal feature vectors W m For the first m Feature fusion weight matrix of each factor For a moment t node i The m Attention weights for each factor; ④ Risk prediction, The fused feature vector The fused node features are input into the prediction layer, and the collapse risk probability of the node is output by a sigmoid function: ， wherein 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 urban ground collapse intelligent early warning method based on multi-source factor fusion according to claim 6, characterized in that, The multi-source data fusion of step S4 includes uncertainty estimation and online adaptive updating after prediction.

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

9. The urban ground collapse intelligent early warning method based on multi-source factor fusion according to claim 1, characterized in that, The double-channel research and judgment rule in step S5 is: If the InSAR subsidence rate value exceeds the set threshold, it indicates that the physical deformation has reached the critical state, which is considered as potential risk; The high-risk threshold is set as ST-GNODEA prediction risk greater than 0.8, and when the prediction risk probability and the uncertainty is greater than 0.8, it is considered that the multi-source factor coupled disaster risk enters a high-risk interval and is regarded as 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 prompt are performed.

10. An urban ground collapse intelligent early warning system based on multi-source factor fusion, characterized in that, It includes: A data acquisition unit for acquiring multi-temporal SAR image data, multi-source remote sensing data, and GIS spatial data; A data preprocessing unit for preprocessing the multi-source data acquired by the data acquisition unit; An InSAR deformation monitoring unit, which uses two time-series synthetic aperture radar interferometry techniques, PS-InSAR and SBAS-InSAR, to process multi-temporal SAR image data, calculate deformation values, form deformation time series, and perform consistency verification on the deformation time series extracted by the two processing methods to generate a subsidence map; An adaptive multi-source factor fusion modeling unit, which constructs a node feature set containing dynamic factors and static factors based on multi-source remote sensing data, GIS spatial data, and InSAR deformation results, constructs a multi-channel weighted spatio-temporal graph structure with grid cells in the monitoring area as nodes, and uses a spatio-temporal graph neural differential attention network ST-GNODEA to realize end-to-end multi-modal factor fusion and continuous-time modeling; A model performance evaluation and optimization unit that monitors the prediction performance indicators of the spatio-temporal graph neural differential attention network model in a sliding window in real time, triggers an online updating mechanism when the performance drops by more than a threshold, and uses newly collected data for small-batch incremental training and dynamically adjusts factor attention weights and neural ordinary differential equation parameters; A double-channel research and judgment unit that performs early warning research and judgment based on the comprehensive risk probability of the InSAR subsidence rate threshold and the ST-GNODEA risk probability; A warning unit that issues a warning for areas with subsidence rate thresholds and high predicted risk probabilities.

Images