CN120673558A

CN120673558A - Urban ground collapse intelligent early warning method and system based on multi-source factor fusion

Info

Publication number: CN120673558A
Application number: CN202511164021.1A
Authority: CN
Inventors: 卢克; 刘明; 张剑; 昝友明; 崔苗苗
Original assignee: Shandong Fengshi Information Technology Co ltd
Current assignee: Shandong Fengshi Information Technology Co ltd
Priority date: 2025-08-20
Filing date: 2025-08-20
Publication date: 2025-09-19

Abstract

The present invention relates to an intelligent early warning method and system for urban ground collapse based on multi-source factor fusion, which belongs to the field of urban disaster early warning technology. PS-InSAR and SBAS-InSAR are used to process and calculate deformation values respectively, and the deformation time series extracted by the two processing methods are verified and analyzed, and the qualified vector results are used to generate a settlement map; a settlement rate threshold is set to perform preliminary ground collapse early warning identification according to the deformation; the deformation time series and the collected multi-source data are divided into dynamic factors and static factors, and a multi-channel weighted spatiotemporal graph structure with the monitoring area grid units as nodes is constructed, and the spatiotemporal graph neural differential attention network is used to fuse multi-source data and predict the comprehensive risk probability; and dual-channel fusion analysis is performed. The fusion process of the present invention comprehensively depicts the nonlinear coupling and spatiotemporal dynamic relationship between disaster-causing factors, and can achieve high-precision, low-false-alarm, and strong generalization prediction of urban ground collapse risks.

Description

Urban ground collapse intelligent early warning method and system based on multi-source factor fusion

Technical Field

The invention relates to an intelligent early warning method and system for urban ground collapse based on multi-source factor fusion, in particular to an intelligent early warning method and system for urban ground collapse based on time sequence synthetic aperture radar interferometry (InSAR) and multi-source remote sensing data fusion and combining a space-time diagram nerve differential attention network, and belongs to the technical field of urban disaster early warning and artificial intelligence.

Background

Along with the acceleration of the urban process, the development intensity of underground space is continuously increased, and subways, underground pipe galleries, water supply and drainage pipe networks and large-scale infrastructures are densely distributed, so that the urban geological environment is more complex and fragile. The ground subsidence and subsidence events are frequently generated in part of cities and cause serious threat to public safety, traffic operation and stability of infrastructure due to the influence of continuous or extreme rainfall, underground pipe network leakage, foundation softening, traffic dynamic load, geological structure and other multi-factor coupling effects.

Currently, synthetic aperture radar interferometry (InSAR) techniques, particularly the time series method of permanent scatterer InSAR (PS-InSAR) and small baseline set InSAR (SBAS-InSAR), have been widely used for millimeter-level surface deformation monitoring. However, in urban complex terrains and variable climatic environments, a single InSAR technology is susceptible to atmospheric residual errors, phase unwrapping errors, data loss and the like, and uncertainty exists in monitoring results. In addition, most of the existing multi-source data fusion methods only carry out simple weighted combination on prediction results at a model output layer, and are difficult to deeply describe nonlinear coupling relations among multi-mode factors and space-time dynamic evolution modes thereof, and have obvious defects in capturing hidden and progressive collapse risks.

In the field of urban ground collapse early warning, another key difficulty of multisource fusion modeling is irregular sampling time sequence data processing and space dependency modeling. The prior method generally adopts a sequence modeling assumption of time intervals, is difficult to effectively utilize data with different sources such as SAR images, meteorological monitoring, geological investigation and the like and inconsistent sampling periods, and meanwhile, has a complex space topological relation in urban geological environment, and a traditional convolution or full-connection structure cannot directly model the non-European space structure, so that the cross-regional risk conduction and neighborhood interaction are not depicted sufficiently.

In addition, most of the existing early warning systems depend on a single risk judging channel, lack of a mechanism for dual verification of physical deformation evidence and a multi-factor coupling mechanism, and are easy to miss and misinformation in hidden collapse or short-time emergencies.

Disclosure of Invention

Aiming at the defects of the existing urban ground collapse early warning method in the aspects of multi-source data fusion, space-time dynamic relation modeling, irregular sampling processing, risk judging mechanism and the like, the invention provides the urban ground collapse intelligent early warning method based on multi-source factor fusion, realizes high-precision, low-false-alarm and forced prediction of urban ground collapse risk, and provides reliable technical support for urban safety management.

The technical scheme adopted by the invention is as follows:

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

s1, acquiring multi-temporal SAR remote sensing image data of the same place, adopting two time series synthetic aperture radar interferometry technologies PS-InSAR and SBAS-InSAR to respectively process and calculate deformation values, and extracting respective deformation time sequences;

S2, performing space consistency verification and time sequence feature cross verification analysis on the deformation time sequences extracted by the two processing methods, and importing a verified PS-InSAR processing vector result into an ArcGIS to draw a time sequence InSAR earth surface deformation rate diagram, namely generating a sedimentation diagram;

s3, setting a sedimentation rate threshold value, and carrying out preliminary ground collapse early warning identification according to a sedimentation diagram obtained by InSAR deformation monitoring;

S4, carrying out multi-source remote sensing data acquisition and GIS space data acquisition, dividing the InSAR deformation time sequence and the acquired multi-source data into dynamic factors and static factors, constructing a multi-channel weighted space-time diagram structure taking a monitoring area grid unit as a node, carrying out multi-source data fusion by using a space-time diagram nerve differential attention network (Spatio-Temporal Graph Neural ODE with Attention, ST-GNODEA) and predicting the comprehensive risk probability;

In a space-time diagram nerve differential attention network, firstly, a monitoring area is rasterized into nodes, irregular sampling multisource time sequence data of each node is modeled in continuous time by using a nerve regular differential equation, attention weights of each node and neighbor nodes (including self) thereof on dynamic factor characteristics are calculated through a graph attention network with self-loop connection, then, hidden state vectors of the dynamic factors and static factors are used for representing input factor-level multimode attention modules, importance allocation and nonlinear coupling modeling among factors are realized, collapse risk probability of the node is output by the fused node space-time characteristics through a full-connection prediction layer, and all node risks form a global-oriented risk distribution map;

s5, carrying out double-channel fusion research and judgment, and carrying out ground collapse early warning recognition by combining the InSAR deformation threshold value and the comprehensive risk probability predicted by ST-GNODEA.

In the method, the permanent scatterer PS-InSAR processing in the step S1 comprises registration, interference processing and deformation information extraction, and the small baseline set InSAR processing SBAS-InSAR comprises registration, interference processing, phase unwrapping, atmospheric residual filtering and deformation information extraction.

Step S2, space consistency verification is carried out, gridding matching is carried out, a research area is divided into grid units with the same size, sedimentation rate average values of PS-InSAR and SBAS-InSAR are respectively extracted, and then a rate difference value Deltav of two methods in each grid is calculated:

,

If Deltav is less than or equal to 5mm/a and the spatial correlation coefficient R ² is more than or equal to 0.85, judging that the regions are consistent, otherwise, marking the regions as 'regions to be verified', and manually carrying out verification processing.

And the time sequence characteristic cross verification analysis is carried out, wherein the coherence coefficient is larger than 0.3, the pixel clusters need to cover a continuous area, meanwhile, PS points and SBAS pixel clusters with the same geographic position are ensured, and the deformation time sequence data D _PS (t) and D _SBAS (t) of the PS points and the SBAS pixel clusters are extracted. The data represent deformation values of the point at corresponding SAR image acquisition time points, and the discrete data point connection is plotted into a deformation-time relation graph in the visualization. The similarity of the two sequences D _PS (t) and D _SBAS (t) is calculated by a Dynamic Time Warping (DTW) algorithm:

,

The Similarity is a Similarity index of the deformation time sequence, the larger the value is in a value range of [0, 1], the more consistent the evolution trend of the two sequences is, the DTW_distance is the minimum path accumulation Distance between two points of the sequences calculated by a dynamic time warping algorithm, and the time sequence evolution is considered consistent when the length of the sequences is len and the Similarity is more than 0.9.

The sedimentation rate threshold in the step S3 is set to be that the annual sedimentation rate is > 30 mm/a, the cumulative sedimentation is >10 mm in 30 days, the daily sedimentation is >2mm, and once the threshold is exceeded, a risk marking mechanism is started.

In order to solve the problems that the existing multi-source data fusion method only carries out simple weighted combination on a model output layer and is difficult to capture nonlinear coupling of multi-mode factors and space-time dynamic relations thereof, the invention provides a space-time diagram nerve differential attention network (Spatio-Temporal Graph Neural ODE with Attention, ST-GNODEA), a monitoring area grid is used as a node, a diagram structure is constructed by combining space topology and infrastructure connectivity, continuous time modeling is carried out on irregularly sampled multi-source time sequence data by using a nerve regular differential equation, and end-to-end interpretable multi-source data fusion is realized by a diagram attention mechanism and multi-mode factor attention, thereby improving early warning accuracy and generalization capability, and the specific process of multi-source data fusion is as follows:

(1) And (3) constructing a graph structure:

Dividing the monitoring area into N nodes v _i, wherein the node characteristics comprise:

InSAR deformation time sequence Rainfall sequenceEtc.;

static factors such as soil layer thickness, pipe network density, terrain gradient, land utilization type and the like;

Establishing multi-channel weighted edges between nodes M is an edge type, ij represents an edge from node i to node j, and the edge weight is comprehensively calculated by the space distance, the geological similarity and the infrastructure information;

(2) Modeling in continuous time:

for each node, mapping the discrete observations to a continuous time hidden state using a neural ordinary differential equation (Neural ODE):

,

Wherein the method comprises the steps of For the hidden state of the node v _i at the time t, f _θ is a trainable neural network, x _i (t) represents an input feature vector of the node i at the time t, and the method can adapt to the condition that the sampling interval of the sensor such as InSAR is uneven;

(3) Spatial information interaction:

Drawing a graph attention network (GAT) on the hidden state, and calculating the attention weight of the neighbor node:

,

Where a is the attention weight of the person, 、Vector representation of node features after linear transformation LeakyReLU is a leak ReLU activation function, allowing inputs less than 0 to preserve certain gradients, avoiding dead neurons,Normalizing all neighbor nodes (including the neighbor nodes), so that the weight sum is 1, and e _ij is an edge feature vector between the node i and the node j;

And (3) information aggregation:

,

Wherein the method comprises the steps of For the set of neighbors of node i,In order for the attention to be weighted,For a matrix of linear transformation weights,The original feature vector of the node j is obtained by graph annotation force network space aggregationAs a dynamic factor hidden state, entering a factor level attention calculation module together with static factor characteristics;

(4) Multimodal factor attention fusion:

① Node input and dynamic hidden state

Recording deviceFor node i, obtaining dynamic hidden state vector through space information interaction at time t, the vector has coded neighbor node and side weight informationFor the mth static factor vector of node i, the dynamic factor at time i can be represented byIs obtained directly from the partial components of (a);

② Factor level attention calculation

For each node i, the importance weights of its factors are calculated at time t:

,

Where u is the attention scoring vector, tanh is the hyperbolic tangent activation function, W _d is the weight matrix mapping the dynamic factors, For the dynamic hidden state vector obtained by the node i through space information interaction at the time t, W _s is a weight matrix for mapping static factors,M is the number representing the current factor mode, m' represents the index to all mode factor sets and is the traversal variable in summation;

③ Fusion feature construction

Weighting the dynamic and static information by using factor weights to obtain fused features:

,

Wherein the method comprises the steps of For the fused node modal feature vector, W _m is the feature fusion weight matrix of the mth factor,Attention weight being the mth factor of time t node i.

④ Risk prediction

Fused feature vectorsInputting the node characteristics into a prediction layer, inputting the fused node characteristics into the prediction layer, and outputting the collapse risk probability node risk probability of the nodes through a sigmoid function:

,

Wherein the method comprises the steps of For node risk probability, σ is Sigmoid activation function, W _p is risk prediction weight vector, and b _p is prediction bias term. When the system takes the grid as a unit for early warning, the node risk probabilities can be summarized according to weights.

(5) Uncertainty estimation and online self-adaptive update can also be performed after prediction is completed:

Confidence estimation method based on evidence theory is introduced, and prediction uncertainty is calculated And combining the risk probability with an uncertainty threshold value to perform early warning triggering, so as to avoid false alarm under low confidence, wherein a method for estimating the uncertainty based on Monte Carlo Dropout and a Bayesian neural network is introduced into an output layer:

,

Wherein the method comprises the steps of The degree of uncertainty is predicted and the degree of uncertainty,For the M-th sampling prediction result, M is the Monte Carlo sampling times,Is mean value prediction.

Incremental learning strategy online adaptive update implementation by combining Elastic Weight Consolidation (EWC) with experience playback (Experience Replay):

a. triggering small batch online updating when the performance of the internal model in a short-term window (24 days) is reduced to exceed a set threshold;

b. Neural ODE parameters are updated through small step optimization, and simultaneously, the EWC regular term protects important parameters and prevents historical knowledge from being forgotten.

According to the invention, a dual-channel fusion judging mechanism is introduced in early warning judgment, and the InSAR deformation threshold channel and the ST-GNODEA comprehensive risk probability channel jointly act to realize multi-layer protection combining quick response and accurate prediction.

The two-channel judging rule in the step S5 is as follows:

If the InSAR sedimentation rate value exceeds a set threshold, marking that the physical deformation reaches a critical state, and regarding the physical deformation as a potential risk;

setting ST-GNODEA predicted risk greater than 0.8 as high risk threshold, and when the comprehensive risk probability is high And is also provided withWhen the disaster risk caused by multi-source factor coupling enters a high-risk zone, the disaster risk zone is regarded as a high-risk zone;

when the conditions are met at the same time, triggering a high-priority early warning signal, and performing space visual output and warning prompt.

The invention further aims to provide an intelligent urban ground collapse early warning system based on multi-source factor fusion, which comprises the following components:

The data acquisition unit is used for performing multi-time-phase SAR image data acquisition, multi-source remote sensing data acquisition and GIS space data acquisition;

The data preprocessing unit is used for preprocessing the multi-source data acquired by the data acquisition unit, and comprises SAR image registration, interference processing, phase unwrapping, atmospheric delay correction, multi-source data coordinate unification and resampling, missing value filling, outlier rejection and the like, so as to generate a time-space aligned multi-source factor data set;

Based on an InSAR deformation monitoring unit, adopting two time series synthetic aperture radar interferometry technologies PS-InSAR and SBAS-InSAR to respectively process multi-temporal SAR image data, calculating deformation values to form deformation time series, and carrying out consistency verification on the deformation time series extracted by the two processing methods to generate a settlement map;

The self-adaptive multi-source factor fusion modeling unit is used for constructing a node characteristic set containing dynamic factors and static factors based on multi-source remote sensing data, GIS space data and InSAR deformation results, constructing a multi-channel weighted space-time diagram structure taking a monitoring area grid unit as a node, and realizing end-to-end multi-mode factor fusion and continuous time modeling by adopting a space-time diagram nerve differential attention network ST-GNODEA;

The model performance evaluation and optimization unit monitors the prediction performance index of the space-time diagram nerve differential attention network model under a sliding window in real time, triggers an online updating mechanism when the performance is reduced to exceed a threshold value, performs small-batch incremental training by using new acquired data and dynamically adjusts the factor attention weight and Neural ODE parameters;

The dual-channel studying and judging unit is used for carrying out early warning studying and judging according to the sedimentation rate threshold value set by InSAR deformation and the comprehensive risk probability of ST-GNODEA risk probability;

and the early warning unit is used for sending early warning to the area exceeding the sedimentation rate threshold value and having high predicted risk probability.

The beneficial effects of the invention are as follows:

(1) The urban area is continuously deformation monitored by utilizing two time sequence synthetic aperture radar interferometry techniques, ground subsidence time sequence extraction with millimeter-level precision is realized, deformation priori information is provided for subsequent collapse risk analysis, the problem that the traditional D-InSAR is greatly influenced by atmospheric noise and phase unwrapping in urban complex terrain is solved by combining PS-InSAR and SBAS-InSAR technologies, and the limitation of a single method is eliminated and the reliability of deformation monitoring results is improved by virtue of complementary analysis of PS-InSAR and SBAS-InSAR.

(2) And (3) carrying out fusion modeling on multi-source remote sensing and environmental factors, fusing InSAR deformation, rainfall time sequence, geological parameters, underground pipe network density, DEM, gradient, land utilization and other multi-source factors, constructing a time-space aligned rasterized sample library, and realizing collaborative characterization of dynamic factors and static factors. The method breaks through the dead zone of the traditional model on the depiction of the multi-factor coupling disaster-causing mechanism, combines GIS space analysis and normalization processing, realizes unified coding and seamless fusion of the heterologous data, and provides high-quality input for subsequent intelligent modeling.

(3) The invention combines the neural ordinary differential equation with the graph attention network and the factor attention mechanism to realize continuous time modeling of multi-source data and interpretable factor level fusion. The graph attention mechanism introduces spatial proximity and infrastructure connectivity, captures the interaction between the propagation path of surface instability and the neighborhood, and comprehensively characterizes nonlinear coupling and space-time dynamic relationship between disaster factors in the fusion process.

(4) The online self-adaptive updating guarantees long-term stability, an incremental learning strategy combining elastic weight consolidation and experience playback is adopted, and small-batch online updating is triggered when the performance degradation of the model is detected, so that the rapid absorption of new knowledge and the reservation of historical knowledge are guaranteed. The strategy can adapt to different urban geological conditions and climate change environments, and the early warning failure risk caused by data distribution drift is remarkably reduced.

(5) The multi-dimensional confidence assessment improves the reliability of decision making, and the risk probability and uncertainty index are obtained by estimating the prediction distribution through a Bayesian method and a plurality of sampling methods combined with Monte Carlo Dropout. When the uncertainty is higher than a set threshold, the system automatically sends out a prompt of 'focus attention', so that low confidence prediction is prevented from directly triggering early warning, and the controllability and the robustness of early warning quality are realized.

(6) The dual-channel early warning mechanism realizes synchronous verification of a physical deformation signal and a disaster-causing mechanism through dual-channel fusion of a deformation threshold channel and an ST-GNODEA risk probability channel, so that early warning accuracy is remarkably improved, and false alarm rate is reduced. The deformation threshold channel is precisely focused, key evidence of surface instability is directly captured, and the ST-GNODEA risk probability channel identifies progressive hidden danger through multi-source factor fusion. The two are cooperated to exert force, so that false alarm and false alarm caused by construction vibration are effectively restrained, monitoring accuracy is guaranteed, and generalization capability and long-term stability of the model under different cities and environmental conditions are enhanced.

Drawings

FIG. 1 is a flow chart of the method of the present invention;

FIG. 2 is a representation of the result of monitoring the surface deformation of a certain area in Shenzhen city based on the time sequence InSAR, a. PS-InSAR processing result, b. SBAS-InSAR processing result;

FIG. 3 shows the result of performing ground collapse risk calculation on a region in Shenzhen city by using the ST-GNODEA model in the embodiment of the present invention.

Detailed Description

The invention will be further illustrated with reference to specific examples.

Embodiment 1 is an intelligent early warning method for urban ground collapse based on multi-source factor fusion, comprising the following steps:

S1, acquiring multi-time-phase SAR remote sensing image data of the same place, adopting two time sequence synthetic aperture radar interferometry technologies PS-InSAR and SBAS-InSAR to respectively process and calculate deformation values, and extracting respective deformation time sequences:

Firstly, continuous deformation monitoring is carried out on an urban area by using two time series synthetic aperture radar interferometry technologies, namely a permanent scatterer InSAR and a small baseline set InSAR, so that ground subsidence time sequence extraction with millimeter-level precision is realized, and deformation priori information is provided for subsequent collapse risk analysis.

The core principle of InSAR is that the image pairs acquired by SAR satellites at different times are utilized to extract the relative deformation information of the ground object in the radar wave propagation direction through interference processing. In the invention, the combination of PS-InSAR and SBAS-InSAR technology overcomes the problem that the traditional D-InSAR is greatly influenced by atmospheric noise and phase unwrapping in urban complex terrain.

The PS-InSAR processing comprises registration, interference processing and deformation information extraction, and the SBAS-InSAR processing comprises registration, interference processing, phase unwrapping, atmospheric residual filtering and deformation information extraction.

(1) Acquiring multi-phase SAR image data, and carrying out accurate registration:

And taking continuous time series SAR images from the satellite platform, selecting a main image, establishing a master-slave relation according to a set baseline threshold value, registering the images, and ensuring comparability between pixel pairs of the same ground object.

(2) Interference treatment:

Each pair of interference pairs (i.e., the SAR image pair consisting of the primary image and the secondary image) is subjected to interference processing to generate an interferogram, the formula is as follows:

,

Wherein the method comprises the steps of Is the earth phase caused by the earth's curvature,For the phase caused by the relief of the terrain,For the phase of the deformation,For the phase caused by the atmospheric propagation delay,Is the systematic and registration error. By means of the high-precision digital elevation model, the flat ground phase caused by the earth curvature and the phase caused by the topography fluctuation can be eliminated, and a differential interference pattern can be obtained.

(3) The original differential phase is subjected to spatial domain and time domain filtering, the signal to noise ratio of the phase is improved, then phase unwrapping processing is carried out, and continuous physical phase quantity is recovered, wherein the formula is as follows:

,

Wherein the method comprises the steps of Is a phase unwrapping function, i.e. the wrapping phase is reduced to a continuous phase by an algorithm. Common practice is toIndicating the wrapping phase (WRAPPED PHASE), i.e., the phase value that is limited to the (-pi, pi) range after interference processing, since the radar measured phase has periodicity, it can only measure the relative change over a range of wavelengths, and out of range "wrapping".

(4) The deformation calculation and LOS conversion, using the unwrapped interference phase, calculate the cumulative deformation along the radar line of sight (LOS) direction, as follows:

,

Wherein the method comprises the steps of For the deformation in the line of sight, lambda is the radar wavelength,Is the unwrapped interference phase.

Of radar measurementsIs the amount of change in the satellite to ground point range. Since the ground subsidence is mainly vertical movement, it is necessary to move the ground according to the incident angle theta of the radar beamProjecting to the vertical direction to obtain vertical deformation:

,

Where θ is the angle of incidence of the radar.

(5) Extracting the deformation time sequence of PS-InSAR and SBAS-InSAR:

PS-InSAR: point targets (i.e., permanent scatterers) with stable scattering properties are identified by statistical analysis and a long-time series of coherent image pairs are created for them. And the deformation time sequence D (t) of the target point is extracted through differential interference, track error fitting and atmospheric noise modeling, so that the method has high time resolution and noise suppression capability.

The SBAS-InSAR builds a plurality of interference atlases with small space baselines and time baselines to form a underdetermined equation set, and reconstructs a deformation time sequence of each pixel through Singular Value Decomposition (SVD), so that the method is suitable for scenes with wide areas and incondensable scatterers.

S2, performing space consistency verification and time sequence feature cross verification analysis on the deformation time sequences extracted by the two processing methods, and importing a vector result processed by PS-InSAR which is qualified in verification into an ArcGIS to draw a time sequence InSAR earth surface deformation rate diagram, namely generating a settlement diagram:

And through the complementary analysis of the PS-InSAR and the SBAS-InSAR, the limitation of a single method is eliminated, and the reliability of a deformation monitoring result is improved.

Space consistency verification, gridding matching is performed firstly, a research area is divided into 500m multiplied by 500m grid units, sedimentation rate average values of PS-InSAR and SBAS-InSAR are respectively extracted, and then a rate difference value Deltav of two methods in each grid is calculated:

,

Wherein v _PS is the sedimentation rate average value of PS-InSAR extraction, v _SBAS is the sedimentation rate average value of SBAS-InSAR extraction, if Deltav is less than or equal to 5mm/a and the spatial correlation coefficient R ² is more than or equal to 0.85, the uniform region is judged, otherwise, the uniform region is marked as a region to be verified, and verification treatment is carried out manually.

Time sequence characteristic cross-validation analysis, namely selecting PS points and SBAS pixel clusters with coherence coefficients larger than 0.3 and pixel clusters needing to cover continuous areas, simultaneously ensuring the PS points and the SBAS pixel clusters in the same geographic position, extracting deformation time sequence data D _PS (t) and D _SBAS (t) of the PS points and the SBAS pixel clusters, representing deformation values of the PS points at corresponding SAR image acquisition time points, and usually connecting and drawing the discrete data points into a deformation-time relation graph during visualization, wherein the similarity of two sequences D _PS (t) and D _SBAS (t) is calculated through a Dynamic Time Warping (DTW) algorithm:

,

The Similarity is a Similarity index of the deformation sequences, the value range is 0 and 1, and the larger the value is, the more consistent the evolution trend of the two sequences is. Dtw_distance is the minimum path cumulative Distance between two points of the sequence calculated by the dynamic time warping algorithm. D _PS is a deformation time sequence of PS-InSAR extraction. D _SBAS is a deformation time sequence extracted by SBAS-InSAR. len is the sequence length, and the time sequence evolution is considered consistent when the similarity is > 0.9.

S3, setting a sedimentation rate threshold value, and carrying out preliminary ground collapse early warning recognition according to a sedimentation diagram obtained by InSAR deformation monitoring:

an early warning threshold (annual sedimentation rate >30 mm/a, 30-day cumulative sedimentation > 10 mm, daily average sedimentation >2 mm) is set, and a risk marking mechanism is started once the InSAR result exceeds the threshold.

S4, carrying out multi-source remote sensing data acquisition and GIS space data acquisition, dividing the InSAR deformation time sequence and the acquired multi-source data into dynamic factors and static factors, constructing a multi-channel weighted space-time diagram structure taking a monitoring area grid unit as a node, carrying out multi-source data fusion by using a space-time diagram nerve differential attention network (Spatio-Temporal Graph Neural ODE with Attention, ST-GNODEA) and predicting the comprehensive risk probability:

The method divides the monitoring area into N nodes v _i, node characteristics are composed of dynamic factors and static factors, and space and functional connection among the nodes are described through multi-channel weighted edges.

(1) Factor profile system:

According to the method, the space-time sample is constructed, inSAR deformation data are fused with a plurality of external remote sensing and geological factors, and a ground collapse risk prediction data set is constructed. The factor involved includes:

① Surface deformation

The accumulated settlement of the earth surface extracted by the time sequence InSAR technology reflects the compression, displacement or instability trend of the earth body in the monitoring period. The factor is a direct deformation precursor of collapse risk, and is particularly sensitive to progressive soil loss and sudden karst collapse. The data is input in a grid time sequence form, and the spatial resolution is consistent with the SAR image.

② Rainfall time series

Accumulated rainfall data generated based on interpolation of meteorological satellites and ground rainfall stations. Rainfall is a core dynamic factor for inducing collapse, short-term heavy rain reduces shear strength by increasing soil saturation and pore water pressure, and long-term rainfall aggravates foundation softening and karst erosion. The data requires construction of a timing grid contemporaneous with InSAR to capture the hysteresis effect of the rain intensity-deformation response.

③ Density of underground pipe network

And calculating the underground pipeline in a unit area by utilizing the GIS database of the urban infrastructure. The soil loss risk of the high pipe network density area is obviously increased due to the ageing and cracking of the pipeline and leakage and scouring, and the high pipe network density area is particularly used for the water supply and drainage pipe network dense area. The factors are input in a static grid layer, and the spatial resolution is consistent with that of an early warning system.

④ Road density

Road coverage area kernel density data generated based on road network vector data. The road dense area accelerates roadbed fatigue due to repeated action of traffic dynamic load, and meanwhile, the construction of underground pipe gallery leaves holes to easily cause collapse. The factor is needed to correlate heavy vehicle traffic frequency data to improve prediction accuracy.

⑤ Land use type

Land cover class codes derived from remote sensing classification and planning data. Different land utilization types imply different disaster exposure.

⑥ Soil layer thickness

And the vertical thickness of the weak soil layer is obtained through geological drilling and electromagnetic surveying. When the thickness of the weak layer is more than 5 meters, soil mass is easy to liquefy or compress and subside under the seepage and vibration effects, and the critical deformation threshold is obviously reduced. The factor is input in the form of a spatial interpolation grid, the resolution being matched to the system.

⑦ Distance from main river

The euclidean distance of the grid unit to the nearest river is calculated based on the hydrologic network GIS data. The close-range area is strongly influenced by river bank scouring and underground water level fluctuation, and especially the side erosion of the river bed in the flood season is easy to trigger the collapse of the bank slope. The data is weighted and optimized by combining river width and flow rate.

⑧ Digital Elevation Model (DEM)

Ground level Cheng Shange data from LiDAR and stereo navigation. The depression may accelerate foundation softening due to water accumulation.

⑨ Gradient factor

And meanwhile, the rainfall runoff speed of the high-gradient area is increased, and the surface soil erosion is aggravated. This factor needs to be coupled with lithology data to evaluate the slip resistance.

Each factor is rasterized by GIS means to unify resolution. All factor data are normalized:

,

for each grid cell, a time series sample is constructed, the true label is from the collapse event point of the history, marked 1, and the non-collapse area is 0.

(2) Multi-source factor data fusion:

① And (3) constructing a graph structure:

node definition each node represents a permanent scattering point (PS point) or grid element, to which both dynamic and static factors correspond. The dynamic factors comprise InSAR deformation time sequence and rainfall time sequence, and the static factors comprise underground pipe network density, road density, land utilization type, soil layer thickness, river distance, DEM and gradient.

Edge relation construction, namely establishing multi-channel weighted edgesM is an edge type, ij represents an edge from node i to node j, and the edge weight is comprehensively calculated by the space distance, the geological similarity and the infrastructure network information to form a multi-scale space topology.

The multisource factors are modeled jointly with the graph structure in a unified end-to-end framework.

② Modeling in continuous time:

for dynamic factors such as InSAR deformation time sequence, rainfall time sequence and the like, neural ODE is adopted to represent a continuous time hidden state h _i (t):

,

Wherein the method comprises the steps of For the hidden state of the node v _i at the time t, f _θ is a trainable neural network, x _i (t) represents an input characteristic vector of the node i at the time t, the method can adapt to the condition that the sampling interval of the sensor such as InSAR is uneven, neural ODE allows state evolution to be directly calculated on irregular time intervals, so that different sampling frequencies of different data sources are adapted, and high-precision interpolation or extrapolation prediction can be performed when new data arrives.

③ Spatial information interaction:

Drawing a graph attention network (GAT) in a hidden state, calculating the attention weight of a neighbor node, and carrying out message transfer on the space and the functional topology between the nodes by adopting the GAT in a continuous time hidden state:

,

multiple risk propagation paths, such as groundwater seepage diffusion caused by pipeline rupture, runoff paths for slope rainfall infiltration and the like, can be captured simultaneously through the multi-channel edge structure.

And (3) information aggregation:

,

Wherein the method comprises the steps of For the set of neighbors of node i,In order for the attention to be weighted,For a matrix of linear transformation weights,The original feature vector of the node j is obtained by graph annotation force network space aggregationAs a dynamic factor hidden state, a factor level attention calculation module is entered along with static factor features.

④ Multimodal factor attention fusion:

a. Node input and dynamic hidden state

Recording deviceFor node i, obtaining dynamic hidden state vector through space information interaction at time t, the vector has coded neighbor node and side weight informationFor the mth static factor vector of node i, the dynamic factor at time i can be represented byIs obtained directly.

B. Factor level attention calculation

,

Where u is the attention scoring vector, tanh is the hyperbolic tangent activation function, W _d is the weight matrix mapping the dynamic factors, For the dynamic hidden state vector obtained by the node i through space information interaction at the time t, W _s is a weight matrix for mapping static factors,M is the number representing the current factor mode, m' represents the index to all mode factor sets, is the traversal variable in summation, and is recordedThe change of the factor(s) can intuitively explain the main driving factors of the prediction risk at a certain moment, such as the significant increase of the weight of rainfall factors during heavy rain, and the long-term high weight of pipe network factors in dense areas of underground pipe networks.

C. Fusion feature construction

,

D. risk prediction

,

Wherein the method comprises the steps of For node risk probability, σ is Sigmoid activation function, W _p is risk prediction weight vector, and b _p is prediction bias term. When the system takes the grid as a unit for early warning, node risk probabilities can be summarized according to weights;

By recording The change of the factor(s) can intuitively explain the main driving factors of the prediction risk at a certain moment, such as the significant increase of the weight of rainfall factors during heavy rain, and the long-term high weight of pipe network factors in dense areas of underground pipe networks.

E. Prediction uncertainty estimation

In order to quantify the reliability of the prediction result, the uncertainty estimation method based on the Monte Carlo Dropout and the Bayesian neural network is introduced into an output layer:

,

Wherein the method comprises the steps of The degree of uncertainty is predicted and the degree of uncertainty,For the M-th sampling prediction result, M is the Monte Carlo sampling times,Is mean value prediction. Uncertainty is used for assisting a two-channel early warning mechanism, when risk probability is high but uncertainty is large, the region is listed into a key attention area, high-level early warning is not triggered directly, and false alarm risk is reduced.

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

a. loss function design

Adopting weighted two-classification cross entropy loss and combining sample unbalance to adjust positive and negative class weights:

,

Where L is a loss value representing the difference between model predictions and real labels. N is the total number of samples. y _i is a real label, and the value is 0 or 1. The probability value predicted by the model represents the probability that the sample belongs to the positive class, and the value range is [0, 1]. W _pos is a positive class weight used to adjust the importance of the positive class sample in the loss calculation. W _neg is a negative class weight used to adjust the importance of the negative class sample in the loss calculation.

B. Online learning and incremental updating

In order to ensure that the system adapts to dynamically-changed geology and environmental conditions, the invention introduces an incremental learning strategy of elastic weight consolidation and experience playback under a single model architecture:

,

Where L _new is the loss function of the new data, F _i is the diagonal elements of the information matrix, For the optimal parameters of the historical task, θ _i is the differential equation parameter of node i, and λ is the regularization coefficient.

Experience playback, namely mixing and sampling new data with historical key samples, and keeping generalization capability of space-time characteristics;

Triggering increment updating when the early warning accuracy rate in the sliding window is reduced by more than 2% or significant data distribution drift occurs;

periodic calibration, namely performing full parameter calibration once every 30 days.

C. Parameter optimization and training stability

Updating model weights by gradient descent optimization algorithm when new monitoring data is acquired

An improved version of the Adam optimizer, adamW, is adopted, a weight attenuation mechanism is introduced to prevent overfitting, limit the gradient norm to be not more than 1.0, prevent gradient explosion, and the FP16 precision acceleration training is adopted, and meanwhile, the numerical stability of the FP32 precision is maintained.

D. Knowledge preservation and environmental adaptation

In the aspect of knowledge reservation, the characteristic weight which greatly contributes to the historical early warning event is identified, the parameter change is restrained by utilizing L2 regularization, independent output layers are arranged for different geological blocks and share a bottom layer characteristic extractor in multi-scene adaptation, and on distribution alignment, the characteristic distribution difference of a source domain and a new target domain is reduced by means of maximum average difference loss, so that smooth transfer learning is achieved.

E. model performance real-time assessment

And (3) multi-scale performance trend monitoring and self-adaptive maintenance, and constructing 24-day (short-term), 72-day (medium-term) and 144-day (long-term) multi-scale sliding window performance monitoring frames based on SAR image acquisition periods, wherein the multi-scale sliding window performance monitoring frames are used for capturing timeliness changes and seasonal trends simultaneously.

The trend analysis method combines linear regression slope, mann-Kendall monotonicity test and CUSUM variable point detection to realize performance trend identification and mutation point discovery.

The grading early warning threshold value is that the short-term F1 fraction is reduced by more than 1 percent (slight), the middle-term accuracy is reduced by more than 3 percent (obvious), and the long-term recall rate is reduced by more than 5 percent (serious).

Adaptive maintenance strategy:

slight, quick increment fine adjustment;

Experience playback + EWC incremental training;

and (3) seriously rolling back to a stable model snapshot and comprehensively retraining.

The mechanism realizes dynamic monitoring and hierarchical maintenance of the performance of the ST-GNODEA model, and ensures long-term stability of prediction accuracy.

The mechanism can realize long-term monitoring and self-adaptive optimization of single model performance without relying on multi-sub model comparison, and ensure the prediction precision and stability of ST-GNODEA under the conditions of data distribution change, seasonal fluctuation or sudden environmental interference.

S5, carrying out double-channel studying, judging and early warning according to a sedimentation 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 _i (t) of each grid, and the invention provides a two-channel studying and judging rule as follows:

if the InSAR sedimentation value exceeds a set threshold value, marking that the physical deformation reaches a critical state, and regarding the physical deformation as a potential risk;

setting ST-GNODEA predicted risk greater than 0.8 as high risk threshold, and when the comprehensive risk probability is high And is also provided withWhen the method is used, high risk early warning is directly triggered;

When the conditions are met at the same time, triggering a high-priority early warning signal, and performing space visual output and warning prompt, wherein the early warning triggering conditions are shown in the following table.

TABLE 1 early warning trigger conditions

。

And selecting 78-scene Sentinel-1 image data from 2 days of 2022, 6 months to 17 days of 2025, 5 months and 17 days, and monitoring PS-InSAR and SBAS-InSAR of a certain region in Shenzhen city to obtain the surface deformation data of the region as shown in figure 2.

Monitoring shows that the local maximum sedimentation rate of the Futian area reaches-48.94 mm/a, the precision error is +/-3 mm, and the method meets the first-level monitoring standard of engineering measurement Specification (GB 50026-2007). Meanwhile, the subsidence of the Futian area has space aggregation, and the subsidence high-value area is in 'binuclear distribution'.

The sedimentation center is located in the range of 500 meters around the joint inspection building, and is related to the combined action of dynamic load superposition underground pipe network leakage of 2.38 ten thousand times (2023 years of data) of deep harbor cross-border trucks in a day.

The settlement rate under roads such as deep south major road, shore major road and the like exceeds-30 mm/a, and the induced settlement main cause can induce soil layer stress release due to subway tunnel shallow burying construction.

The fused analysis based on the time sequence InSAR monitoring and ST-GNODEA output is shown in figure 3, and the median of the regional global ground collapse risk probability is up to 0.48, which is significantly higher than the safety threshold. The InSAR monitoring result shows that the sedimentation effect of the sentry port is most obvious, and the risk of collapse is high. The ST-GNODEA results show that urban main roads such as deep south major roads, side major roads and the like form a high-risk gathering area, the grid with the comprehensive risk probability of more than 0.8 occupies higher area, but the risk of the front guard ports in the areas with serious settlement is relatively lower, because the front guard ports are mainly used for import and export trade, and the life and commute influence of urban personnel is smaller. If non-living areas, mainly the gate of the guard, are ignored, the result is misjudged. And through the double-channel monitoring of the time sequence InSAR settlement monitoring threshold and the ST-GNODEA comprehensive risk threshold, the urban ground collapse early warning accuracy is improved.

The conclusion is verified by spatial correlation analysis with the historic collapse event kernel density distribution. The result shows that the high risk prediction area is [ ]) The pearson correlation coefficient with the historic collapse event high nuclear density hot zone (nuclear density value > 0.85) reaches r=0.91 (p < 0.001), indicating that the proposed integrated model has better performance in high risk zone identification.

In conclusion, the early warning model accuracy verification shows that the pearson correlation coefficient of the prediction result and the historical collapse space distribution reaches 0.91 (p < 0.001), the recall rate of the high-risk area is 92.3%, and the false alarm rate is controlled below 5%. The demonstration result fully shows that the method has obvious scientificity and engineering effectiveness in urban ground collapse early warning by fusing the time sequence InSAR deformation and multisource factor coupling mechanism based on the ST-GNODEA model, and has obvious technical advantages and application value compared with the traditional single model method.

Embodiment 2 is an intelligent urban ground collapse early warning system based on multi-source factor fusion, which realizes the whole-flow closed-loop management from data acquisition and processing to risk early warning, and the specific functional modules are as follows:

the data acquisition unit is used for acquiring multi-time-phase SAR image data, multi-source remote sensing data and GIS space data;

The data preprocessing unit is used for preprocessing the data acquired by the data acquisition unit;

The invention is further described above in connection with embodiments, to which the scope of protection of the invention is not limited.

Claims

1. An intelligent early warning method for urban ground collapse based on multi-source factor fusion is characterized by comprising 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 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 alignment, interferometry processing and deformation information extraction; the small baseline set InSAR processing SBAS-InSAR includes alignment, interferometry 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 first gridded and matched, the study area is divided into grid cells of equal size, the sedimentation rate means of PS-InSAR and SBAS-InSAR are extracted respectively, and then the rate difference Δ v of the two methods in each grid is calculated:

,

If Δ v ≤5 mm/a and the spatial correlation coefficient R ² ≥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, in the time series feature cross-validation analysis described in step S2, a pixel cluster with a coherence coefficient greater than 0.3 and covering a continuous area is selected, and the PS points and SBAS pixel clusters are ensured to be in the same geographical location. Their respective deformation time series data D _PS (t) and D _SBAS (t) are extracted, and the similarity of the two sequences D _PS (t) and D _SBAS (t) is calculated using a dynamic time warping algorithm:

,

Similarity is the similarity index of the deformation time series, ranging from 0 to 1. A larger value indicates a more consistent evolution trend between 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 settlement rate threshold of step S3 is set to annual settlement rate>30 mm/a, 30-day cumulative settlement>10 mm, and daily average settlement>2 mm. Once the threshold is 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 multi-source data fusion process in step S4 is as follows:

(1) Graph structure construction:

The monitoring area is divided into N nodes v _i , and the 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 is the hidden state of node vi 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 ;

(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 is the Leaky ReLU activation function, which 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 the edge feature vector between node i and node j ;

Aggregate information:

,

in is the neighbor set of node i , is the attention weight, is the linear transformation weight matrix, is the original feature vector of node j , obtained by spatial aggregation of the graph attention network 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 is the dynamic hidden state vector obtained by node i at time t through spatial information interaction, which has encoded the neighbor nodes and edge weight information; is the mth static factor vector of node i , and the dynamic factor at time i can be expressed by Part of the weight is obtained directly;

② Factor-level attention calculation,

For each node i , calculate the importance weight of each factor at time t :

,

Where u is the attention score vector, tanh is the hyperbolic tangent activation function, Wd is the weight _matrix of the mapping dynamic factor, is the dynamic hidden state vector obtained by node i at time t through spatial information interaction, Ws is the weight matrix of the mapping static _factor , is the mth static factor vector of node i , m is the number representing 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, Wm is the feature _fusion weight matrix of the mth factor, is the attention weight of the mth factor of node i at time t ;

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, Wp is the risk prediction weight vector, _and _bp 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 in step S4 includes uncertainty estimation and online adaptive update after the prediction.

8. The intelligent early warning method for urban ground collapse based on multi-source factor fusion according to claim 7 is characterized in that an uncertainty estimation method based on Monte Carlo Dropout and Bayesian neural network is introduced in the output layer:

,

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 and judgment rule 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. An intelligent early warning system for urban ground collapse based on multi-source factor fusion is characterized by including:

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.