High-precision topographic change monitoring and geological disaster early warning image analysis system

Abstract

The invention, which belongs to the technical field of image analysis and geological disaster early warning, discloses a high-precision topographic change monitoring and geological disaster early warning image analysis system comprising a deformation spatio-temporal feature sensing module, a geomechanics constraint optimization module, a multi-scale disaster evolution prediction module and a self-adaptive early warning decision module. Through deep coupling and closed-loop feedback between modules, dynamic matching of deformation confidence and mechanical constraint weight, physical enhancement of stress distribution and disaster prediction, and dual-path parameter optimization driven by early warning performance are realized, and the system adopts an InSAR technology and deep learning fusion to extract millimeter-level deformation. The physical embedded neural network is utilized to realize anomaly recognition under mechanical constraints, multi-scale disaster evolution is predicted through space-time convolution Transform, and the early warning accuracy rate can reach 95% or above after closed-loop iterative optimization.

Description

Technical Field

[0001] This invention belongs to the field of terrain change monitoring and geological disaster early warning technology, involving the interdisciplinary fields of image analysis, remote sensing monitoring, deep learning, and geomechanics. Specifically, it relates to a high-precision terrain change monitoring and geological disaster early warning image analysis system based on InSAR technology, deep learning, and physical constraints. This system achieves millimeter-level deformation monitoring and geological disaster early warning 7-15 days in advance through the deep coupling and closed-loop collaboration of four core modules: deformation spatiotemporal feature perception, geomechanical constraint optimization, multi-scale disaster evolution prediction, and adaptive early warning decision-making. It can be widely applied to scenarios such as surface subsidence monitoring in mining areas, landslide early warning in mountainous areas, dam safety monitoring, and urban land subsidence prevention and control. Background Technology

[0002] Geological disasters such as landslides, ground subsidence, and debris flows seriously threaten people's lives and property, as well as the stable operation of infrastructure. Topographic changes are an important precursor signal of geological disasters. High-precision topographic change monitoring can achieve early warning of geological disasters, buying valuable time for disaster prevention and control.

[0003] Existing technology CN120298398B discloses a terrain change monitoring system based on image analysis. This system employs techniques such as multi-source image fusion, preprocessing, dual-channel deep convolutional network analysis, point cloud processing, and early warning to achieve terrain change monitoring. However, the existing technology has the following shortcomings:

[0004] First, existing technologies employ a linear, sequential processing flow, where the image acquisition, preprocessing, analysis, generation, and early warning modules operate independently. These modules only exchange data without deep coupling, resulting in overall system performance limited by the simple summation of their components, failing to achieve synergistic effects. For example, the parameters of the preprocessing module are fixed and cannot be dynamically adjusted based on subsequent analysis results, leading to poor preprocessing performance in complex scenarios.

[0005] Secondly, existing technologies lack geomechanical constraint mechanisms, relying solely on image analysis and statistical learning methods to identify topographic changes, failing to integrate geological structural features and mechanical laws into the analysis process. This leads to the system easily misclassifying abnormal data that conforms to statistical patterns but violates mechanical laws as real deformation, or missing minute deformation signals that conform to mechanical laws but deviate from statistical patterns, affecting the accuracy and reliability of monitoring. While existing dual-channel deep convolutional networks can extract texture, elevation, and temporal features, they lack consideration for geological structural constraints and mechanical equilibrium conditions, resulting in insufficient monitoring accuracy in stress concentration areas and geologically weak zones.

[0006] Third, existing technologies employ fixed warning thresholds and linear warning logic, failing to adaptively adjust based on warning effectiveness. In practical applications, geological conditions, climate environments, and human activities vary significantly across different regions, making fixed thresholds ill-suited to diverse monitoring scenarios, resulting in high false alarm and false negative rates. Furthermore, existing technologies lack a feedback optimization mechanism from warning results to monitoring parameters, preventing the system from learning and improving through historical warning data, thus hindering continuous improvement in warning performance.

[0007] Fourth, the processing logic of each module in the existing technology is independent, lacking closed-loop feedback and collaborative optimization mechanisms. For example, the terrain change parameters and risk levels calculated by the early warning module cannot be fed back to the preceding feature extraction and analysis modules. This prevents the system from optimizing feature extraction strategies and analysis algorithms based on the early warning effect, resulting in the entire system operating in an open-loop state, unable to achieve adaptive optimization and continuous performance improvement. Summary of the Invention

[0008] The purpose of this invention is to address the problems of existing topographic change monitoring systems, such as independent modules, lack of mechanical constraints, inability to adaptively optimize, and limited early warning performance. This invention provides a high-precision topographic change monitoring and geological disaster early warning image analysis system. Through deep coupling and closed-loop feedback of a deformation spatiotemporal feature perception module, a geomechanical constraint optimization module, a multi-scale disaster evolution prediction module, and an adaptive early warning decision module, this system achieves millimeter-level deformation monitoring, anomaly identification under mechanical constraints, multi-scale disaster prediction, and adaptive early warning decision-making, significantly improving the accuracy, lead time, and robustness of geological disaster early warnings.

[0009] To achieve the aforementioned objectives, this invention provides a high-precision topographic change monitoring and geological disaster early warning image analysis system. This system employs a modularly coupled closed-loop collaborative architecture, comprising four core modules: a deformation spatiotemporal feature perception module, a geomechanical constraint optimization module, a multi-scale disaster evolution prediction module, and an adaptive early warning decision module. These modules form an organic whole through deep coupling at the parameter, state, and logic levels, achieving adaptive optimization through a dual-path feedback closed loop, and achieving synergistic performance enhancement through a multi-dimensional collaborative mechanism.

[0010] The deformation spatiotemporal feature perception module employs a fusion technique of InSAR phase unwrapping and deep learning to extract millimeter-level deformation information, and generates a deformation spatiotemporal feature map and deformation consistency confidence score based on a spatiotemporal attention mechanism. The core innovation of this module lies in combining traditional InSAR technology with deep neural networks. It leverages the powerful fitting capabilities of neural networks to solve the problems of phase unwrapping and atmospheric delay correction, while simultaneously introducing a spatiotemporal attention mechanism to identify the consistency region of the deformation field, providing a reliability metric for subsequent mechanical constraints.

[0011] The geomechanical constraint optimization module is deeply coupled with the deformation spatiotemporal feature perception module. It receives deformation spatiotemporal feature maps and deformation consistency confidence scores, constructs mechanical constraints based on the geological structural characteristics of the monitored area, and uses a physical embedded neural network to verify the mechanical rationality of the deformation field. The core innovation of this module lies in integrating geomechanical principles into a deep learning framework. By embedding mechanical equilibrium equations into the loss function, the predicted deformation field conforms to both observational data and mechanical laws, effectively identifying and correcting abnormal deformations that violate mechanical laws.

[0012] The multi-scale disaster evolution prediction module is coupled with the geomechanical constraint optimization module. It uses a spatiotemporal convolutional Transformer network to extract multi-scale spatiotemporal patterns of deformation evolution and embeds the stress distribution of geological structures as a constraint into the attention calculation mechanism of the Transformer. The core innovation of this module lies in combining statistical learning with physical mechanisms, enabling the prediction model to capture both data-driven statistical regularities and physical constraints, significantly improving prediction accuracy and generalization ability.

[0013] The adaptive early warning decision-making module is coupled with the multi-scale disaster evolution prediction module and the deformation spatiotemporal feature perception module. It calculates an early warning accuracy index based on the early warning response effect and the actual occurrence of the disaster, and feeds this index back to the deformation spatiotemporal feature perception module and the geomechanical constraint optimization module, forming a dual-path feedback closed loop. The core innovation of this module lies in using early warning performance as a feedback signal to drive the adaptive optimization of system parameters, enabling the entire system to have self-learning capabilities. The early warning accuracy can continuously improve with the increase of the monitoring period.

[0014] Compared with the prior art, the present invention has the following beneficial effects:

[0015] First, this invention breaks through the limitations of traditional linear serial architectures by deeply coupling modules. Deformation consistency confidence directly controls the weighting of mechanical constraints, dynamically matching the strength of mechanical constraints with the reliability of deformation measurements. The stress distribution of geological structures is embedded in the attention calculation for disaster prediction, allowing the prediction process to consider both statistical laws and physical mechanisms. Dual-path feedback signals simultaneously optimize the two core modules—feature perception and mechanical constraints—forming a mutually reinforcing synergistic optimization mechanism. Experiments show that the early warning accuracy of the module-coupled synergistic system is 28 percentage points higher than the simple superposition of modules used independently, demonstrating significant synergistic efficiency.

[0016] Secondly, this invention innovatively integrates geomechanical constraints into a deep learning framework, achieving an organic combination of data-driven and physics-driven approaches through physical embedding of neural networks. Mechanical equilibrium equation constraints are embedded in the loss function of the neural network, introducing prior mechanical knowledge during forward propagation, ensuring that the predicted deformation field conforms to both observational data and mechanical laws. Comparative experiments show that the anomaly identification accuracy of the deformation field after mechanical constraint optimization is improved by 35%, effectively solving the problem of physical inconsistencies that are easily generated by traditional purely data-driven methods.

[0017] Third, this invention achieves adaptive optimization of the system through a dual-path feedback closed loop. The early warning accuracy index is simultaneously fed back to the deformation spatiotemporal feature perception module and the geomechanical constraint optimization module. The former adjusts the spatiotemporal perception parameters and attention mechanism weights, while the latter corrects the geological structure identification threshold and the mechanical parameter inference model. The closed-loop iteration enables the system to have self-learning capabilities. After 3-5 monitoring cycles, the early warning accuracy can be improved from the initial 75% to over 95%, an improvement of 20 percentage points, while the early warning accuracy of traditional open-loop systems remains basically unchanged.

[0018] Fourth, the multi-scale disaster evolution prediction module of this invention extracts multi-scale spatiotemporal patterns such as long-term trends, seasonal cycles, and short-term fluctuations through a spatiotemporal convolutional Transformer network, and embeds the stress distribution of geological structures as a physical constraint into attention calculation. This prediction framework combining statistics and physics enables the system to simultaneously capture complex nonlinear evolution laws and basic physical and mechanical laws, improving prediction accuracy by 25 percentage points compared to purely statistical methods, and extending the early warning period from 3-5 days in traditional methods to 7-15 days.

[0019] Fifth, the system of this invention has achieved remarkable results in multiple practical application scenarios. In monitoring surface subsidence in mining areas, the system achieves millimeter-level deformation monitoring accuracy, with an early warning accuracy rate of 92% and a false alarm rate reduced to below 5%. In landslide early warning in mountainous areas, the system successfully issued early warnings for more than 85% of landslide events 7-15 days in advance. In urban ground subsidence monitoring, the system identified 18 subsidence risk points that were missed by traditional methods. These application results demonstrate the high precision, high reliability, and strong adaptability of the system of this invention.

[0020] In summary, this invention solves the problems of existing technologies, such as independent modules, lack of mechanical constraints, and inability to adaptively optimize, by deeply coupling and closed-loop synergy of four core modules: deformation spatiotemporal feature perception, geomechanical constraint optimization, multi-scale disaster evolution prediction, and adaptive early warning decision-making. It achieves millimeter-level deformation monitoring, anomaly identification under mechanical constraints, multi-scale disaster prediction, and adaptive early warning decision-making, significantly improving the accuracy, lead time, and robustness of geological disaster early warning, and providing strong technical support for natural resource protection and disaster risk management. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall architecture of the system of the present invention.

[0022] Figure 2 This is a detailed structural diagram of the deformation spatiotemporal feature perception module.

[0023] Figure 3 This is a detailed structural diagram of the geomechanical constraint optimization module.

[0024] Figure 4 This is a detailed structural diagram of the multi-scale disaster evolution prediction module.

[0025] Figure 5 This is a detailed structural diagram of the adaptive early warning decision module. Detailed Implementation

[0026] Please refer to the attached document. Figures 1-5 The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0027] like Figure 1 As shown, the present invention provides a high-precision topographic change monitoring and geological disaster early warning image analysis system, which includes four core modules: deformation spatiotemporal feature perception module 1, geomechanical constraint optimization module 2, multi-scale disaster evolution prediction module 3, and adaptive early warning decision module 4.

[0028] like Figure 2 As shown, the deformation spatiotemporal feature perception module 1 includes an InSAR phase processing unit, a spatiotemporal feature extraction unit, a deformation consistency evaluation unit, and a spatiotemporal fusion unit.

[0029] The InSAR phase processing unit operates as follows: First, it acquires multi-source SAR image data of the monitored area, including C-band and L-band SAR data from different satellite platforms such as Sentinel-1 and ALOS-2. Then, it registers different temporal SAR images of the same monitored area to generate interferograms. The phase information of the interferograms is then processed. Includes terrain phase Deformation phase Atmospheric phase and noise phase :

[0030] ,

[0031] in: To observe the phase, For terrain phase components, For deformation phase components, For atmospheric delayed phase components, This represents the noise phase component.

[0032] A phase unwrapping algorithm based on a deep neural network is employed to eliminate 2π ambiguity. Specifically, a U-Net-based phase unwrapping network is constructed. This network takes a wrapped phase map and a coherence coefficient map as input and outputs the unwrapped true phase. The network is trained on a large amount of simulated and real data to learn the recognition and correction patterns of phase transitions. The loss function of the phase unwrapping network is:

[0033] ,

[0034] in: For phase unwrapping loss, Total number of pixels For the first The predicted phase of each pixel. For the first The true phase of each pixel For gradient operators, The gradient loss weights are used to optimize both the absolute value of the phase and the phase gradient, ensuring that the unwrapping result is both accurate and smooth.

[0035] An atmospheric delay correction network is used to remove the influence of atmospheric disturbances. Atmospheric delay is mainly caused by changes in tropospheric water vapor content, exhibiting spatial correlation but temporal randomness. A spatiotemporal convolutional neural network is constructed, which takes a multi-temporal interferogram sequence as input, learns the spatiotemporal distribution pattern of atmospheric delay, and outputs an atmospheric delay phase estimate. The deformation phase after atmospheric delay correction is:

[0036] ,

[0037] in: The corrected deformation phase. This represents the estimated atmospheric delay phase.

[0038] Through the aforementioned phase unwrapping and atmospheric delay correction processes, the InSAR phase processing unit generates a high-precision phase deformation field, with a deformation measurement accuracy of 1-2 mm.

[0039] The spatiotemporal feature extraction unit works as follows: It converts the phase deformation field into a deformation time series along the LOS (line of sight). For each pixel within the monitoring area, it obtains its... Deformation observation sequence of each time phase A 3D convolutional neural network is used to extract the spatial distribution and temporal evolution features of the deformation.

[0040] The input to a 3D convolutional neural network is a deformable spatiotemporal data volume. ,in and For spatial dimensions, The temporal dimension is considered. The network contains multiple 3D convolutional layers, and the convolutional kernels of each layer perform convolution operations simultaneously in both the spatial and temporal dimensions. A convolution operation is defined as:

[0041] ,

[0042] in: For the first Feature map of the layer For the first Feature map of the layer For the first The convolutional kernel weights of the layer, For the first Layer bias, The activation function is ReLU. Spatial and temporal indices for the convolution kernel. These are the spatial and temporal coordinates of the feature map.

[0043] Through multi-layer 3D convolution, the network extracts multi-scale spatiotemporal features of deformation step by step, forming a multi-scale spatiotemporal feature pyramid. ,in It exhibits shallow, fine-grained characteristics. It represents deep semantic features.

[0044] The deformation consistency assessment unit operates as follows: Based on a multi-scale spatiotemporal feature pyramid, a spatiotemporal attention mechanism is used to identify consistent and anomalous regions in the deformation field. This spatiotemporal attention mechanism comprises two branches: temporal attention and spatial attention.

[0045] The time attention calculation process involves analyzing the deformation time series and calculating the importance weight for each time step. First, global average pooling is performed on the feature map in the spatial dimension to obtain the time series feature vector. Then, the temporal attention weights are calculated using a two-layer fully connected network:

[0046] ,

[0047] in: For time attention weight vectors, This is the weight matrix of the fully connected layer. For bias vectors, and This is the activation function.

[0048] Spatial attention calculation process: Analyze the spatial distribution of deformation and calculate the importance weight of each spatial location. First, perform global average pooling and max pooling on the feature map in the time dimension to obtain two spatial feature maps. and Then, the two feature maps are concatenated along the channel dimension, and spatial attention weights are calculated using a convolutional layer:

[0049] ,

[0050] in: This is a spatial attention weight map. For convolution operations, This indicates a splicing operation.

[0051] The spatiotemporal joint attention matrix is ​​calculated by the outer product of temporal attention and spatial attention:

[0052] ,

[0053] in: For spatiotemporal joint attention matrix, This is for outer product operations.

[0054] Based on the spatiotemporal joint attention matrix, the deformation consistency confidence score for each pixel is calculated. The confidence score calculation considers the consistency of that pixel within its spatiotemporal neighborhood.

[0055] ,

[0056] in: For pixels The confidence level of deformation consistency. For pixels The spatiotemporal neighborhood, For pixels Deformation time series, For neighborhood points Deformation time series, This is the similarity calculation function (using cosine similarity). For pixels The spatiotemporal attention weights.

[0057] Based on the above calculations, the deformation consistency assessment unit generates a confidence distribution map. Areas with high confidence indicate that the deformation signal is reliable, while areas with low confidence may have noise interference or abnormal deformation.

[0058] The spatiotemporal fusion unit works as follows: Multi-scale spatiotemporal features are weighted and fused based on the confidence distribution map. For each layer of features in the multi-scale feature pyramid... Weighted according to confidence level:

[0059] ,

[0060] in: This is the weighted feature map. For element-wise multiplication, This is an upsampling operation (upsampling the confidence map to the same spatial resolution as the feature map).

[0061] Then, the weighted features from each layer are fused to generate a deformation spatiotemporal feature map:

[0062] ,

[0063] in: This is a spatiotemporal feature map of deformation. For the first The fusion weights of layer features are determined by learnable parameters.

[0064] The deformation spatiotemporal feature perception module 1 ultimately outputs a deformation spatiotemporal feature map. confidence distribution plot These two outputs are passed as key parameters to the geomechanical constraint optimization module 2, where the confidence distribution map... It directly determines the weight configuration of subsequent mechanical constraints, achieving deep coupling at the parameter level.

[0065] like Figure 3 As shown, the geomechanical constraint optimization module 2 includes a geological structure identification unit, a mechanical parameter inference unit, a physical constraint verification unit, and a stress distribution calculation unit. This module is deeply coupled with the deformation spatiotemporal feature perception module 1, receiving deformation spatiotemporal feature maps. confidence distribution plot As a key input.

[0066] The working process of the geological structure identification unit is as follows: based on deformation spatiotemporal feature maps Graph Convolutional Networks (GCNs) were used to identify geological structural elements such as faults, fissures, and rock strata interfaces in the monitoring area. First, the monitoring area was divided into... Each grid cell is used to construct a spatial adjacency graph. ,in A set of nodes (representing grid cells). It is a set of edges (representing the connection relationship between adjacent units).

[0067] For each node Extract its deformation spatiotemporal features as node features Graph convolutional networks aggregate neighborhood information and update node features through a message-passing mechanism:

[0068] ,

[0069] in: For the first Layer The feature vector of each node For nodes The set of neighboring nodes, For the first The weight matrix of the layer, For activation function, For aggregation functions (using summation or mean aggregation).

[0070] Through multi-layer graph convolution, the network learns high-order neighborhood information and global structural information for each node. Finally, a classification layer predicts whether each node belongs to geological structural elements such as faults, fractures, or rock-layer interfaces.

[0071] ,

[0072] in: For nodes The probability distribution of belonging to different geological structure categories. This is the classification layer weight matrix. denoted as the total number of layers in the graph convolutional network.

[0073] Based on the classification results, a geological structure topology map is generated. The map identifies the main geological structural elements and their spatial distribution within the monitoring area.

[0074] The working process of the mechanical parameter inference unit is as follows: Based on the geological structure topology map Based on historical geological survey data, the mechanical parameters of various geological structural units are inferred. These mechanical parameters include the elastic modulus. Poisson's ratio Cohesion and internal friction angle .

[0075] First, the monitoring area was divided into several geological structural units based on the geological structure topology map. For each geological structural unit, the statistical distribution of mechanical parameters of similar geological structures was retrieved from historical geological exploration databases. Then, using a Bayesian inference method, the posterior mechanical parameter distribution was inferred by combining prior distributions and observational data (deformation characteristics).

[0076] ,

[0077] in: A set of mechanical parameters For observational data (deformation spatiotemporal characteristics). For the posterior distribution, Let be the likelihood function. As a prior distribution, As evidence.

[0078] By solving the posterior distribution using the Markov chain Monte Carlo (MCMC) sampling method, the estimated values ​​of mechanical parameters and their uncertainty intervals for each geological structural unit are obtained.

[0079] The working process of the physical constraint verification unit is as follows: The deformation spatiotemporal feature map is... As observational data, mechanical equilibrium equations are constructed by combining mechanical parameters, and the stress-strain field is solved by physical embedded neural networks. Abnormal deformations that do not meet the mechanical equilibrium conditions are identified and corrected.

[0080] The mechanical equilibrium equations are based on the theory of continuum mechanics, and the equilibrium equations are as follows:

[0081] ,

[0082] in: For stress tensor, For body force vector, It is a divergence operator.

[0083] The constitutive relation (stress-strain relationship) follows Hooke's law for linear elastic materials:

[0084] ,

[0085] in: For strain tensor, The elasticity matrix is ​​derived from the elastic modulus. Compared to Poisson Sure, This is a double dot product operation.

[0086] The strain-displacement relationship is as follows:

[0087] ,

[0088] in: It is a displacement vector. This is the gradient operator.

[0089] Construct a physically embedded neural network that uses spatial coordinates and time Input, output displacement field The network calculates the strain field through an automatic differentiation mechanism. and stress field The network's loss function includes an observation data fitting term and a physical constraint term:

[0090] ,

[0091] in: For the total loss function, The fitting loss for the observed data, For the residual loss of the equilibrium equation, For boundary condition residual loss, and This is the loss weighting coefficient.

[0092] The fitting loss for the observed data is:

[0093] ,

[0094] in: The number of observation points. For the first Predicted displacement at each observation point For the first The observed displacement of each observation point For the first The confidence level of each observation point (confidence distribution map output by deformation spatiotemporal feature perception module 1) ).

[0095] Note the confidence level here. The weight coefficients are directly used as weights in the loss function. Observation points with high confidence receive greater weights during the optimization process, while observation points with low confidence receive smaller weights, thereby achieving parameter-level deep coupling between the deformation spatiotemporal feature perception module 1 and the geomechanical constraint optimization module 2.

[0096] The residual loss of the equilibrium equation is:

[0097] ,

[0098] in: The number of sampling points is determined by sampling points evenly within the monitoring area.

[0099] Boundary condition residual loss is defined according to the specific boundary conditions, for example, free surface boundary conditions are... ,in This is the boundary normal vector.

[0100] By training a physical embedding neural network, the displacement field output by the network is made to both fit the observed data and satisfy the mechanical equilibrium equations and boundary conditions. After training, the network outputs an optimized deformation field under mechanical constraints. .

[0101] For observing deformation fields Anomalies that do not meet the conditions of mechanical equilibrium are identified through comparison. and Identification:

[0102] ,

[0103] in: The threshold for anomaly detection is set, and anomalies may be measurement errors or real local abnormal deformations.

[0104] The working process of the stress distribution calculation unit is as follows: based on the corrected deformation field The stress distribution of the geological structure is calculated using strain-displacement relationships and constitutive relations. First, the strain field is calculated:

[0105] ,

[0106] Then, based on the mechanical parameters of each geological structural unit, the stress field is calculated using constitutive relations:

[0107] ,

[0108] Calculate the principal stresses and their directions, and identify stress concentration regions. The stress concentration index is defined as:

[0109] ,

[0110] in: For point Stress concentration index This is the maximum principal stress at that point. This represents the average maximum principal stress in the monitored area.

[0111] To identify potential failure surfaces, the Mohr-Coulomb failure criterion is used to assess failure risk for geotechnical materials.

[0112] ,

[0113] in: For shear stress, It is normal stress. For cohesion, The internal friction angle is defined as the angle of stress at a point. When the stress state at a point approaches or exceeds the failure criterion, that point is considered to have a potential risk of failure.

[0114] Geomechanical constraint optimization module 2 ultimately outputs the optimized deformation field under mechanical constraints. and stress distribution in geological structures These two outputs are passed as key parameters to the multi-scale disaster evolution prediction module 3, including the stress distribution of geological structures. The attention calculation mechanism embedded in the prediction network achieves deep coupling at the state level.

[0115] like Figure 4As shown, the multi-scale disaster evolution prediction module 3 includes a multi-scale feature decoupling unit, a spatiotemporal convolutional coding unit, a mechanical enhancement Transformer unit, and a disaster risk assessment unit. This module is coupled with the geomechanical constraint optimization module 2 and receives the optimized deformation field. and stress distribution in geological structures As a key input.

[0116] The working process of the multi-scale feature decoupling unit is as follows: The optimized deformation field... The time series data is decomposed into long-term trend components, seasonal cycle components, and short-term fluctuation components. The Seasonal Trend Decomposition (STL) method is used for this decomposition.

[0117] ,

[0118] in: For a moment Deformation observations, As a component of long-term trends, As a seasonal cyclical component, This represents the short-term fluctuation component (residual).

[0119] Long-term trend components It reflects the slow, cumulative deformation process of geological structures, mainly controlled by long-term factors such as tectonic stress and soil creep. Seasonal periodic component. This reflects deformation fluctuations related to periodic factors such as climate, rainfall, and temperature. Short-term fluctuation component. It includes random noise and sudden deformation events.

[0120] For each grid point within the monitoring area, the time series of three components are extracted to form a multi-scale deformation feature representation: .

[0121] The spatiotemporal convolutional coding unit works as follows: it uses a spatiotemporal convolutional network to extract spatial and temporal dependencies for each temporal scale component. The spatiotemporal convolutional network consists of multiple stacked spatiotemporal convolutional blocks, each of which contains spatial and temporal convolutional layers.

[0122] Spatial convolutional layers extract the spatial distribution pattern of the deformation field, using 2D convolution operations:

[0123] ,

[0124] in: For the input feature map, For spatial convolution kernels, This is the output of a spatial convolution.

[0125] Temporal convolutional layers extract the temporal evolution patterns of deformation sequences, and causal dilated convolution is used to ensure that predictions rely solely on historical information.

[0126] ,

[0127] in: For temporal convolution kernels, This is the output of the temporal convolution.

[0128] Spatiotemporal convolutional encoding is performed on the long-term trend component, seasonal cycle component, and short-term fluctuation component respectively to generate multi-scale spatiotemporal encoded features:

[0129] ,

[0130] ,

[0131] ,

[0132] in: These are the spatiotemporal coding features of the three components. This represents a spatiotemporal convolutional network.

[0133] The three coding features are concatenated along the channel dimension to form a comprehensive multi-scale spatiotemporal coding feature:

[0134] ,

[0135] The working process of the mechanical enhancement Transformer unit is as follows: receiving multi-scale spatiotemporal encoded features. and the stress distribution of geological structures Attention calculation mechanism embedded in Transformer as a constraint.

[0136] The self-attention mechanism of the standard Transformer is calculated as follows:

[0137] ,

[0138] in: For querying the matrix, The key matrix, For value matrices, The dimension of the key vector.

[0139] This invention innovatively embeds geological structural stress distribution into attention calculation, incorporating stress information as a physical constraint into the attention weight calculation:

[0140] ,

[0141] in: The stress distribution matrix of the geological structure (including the stress field) Organized in matrix form according to spatial location) This is the mechanical constraint weighting coefficient, which controls the intensity of the influence of stress constraints on attention calculation.

[0142] By stress distribution By embedding attention calculations, the Transformer considers geomechanical principles while learning statistical patterns of deformation evolution. Regions with high stress concentrations (larger stress values) receive higher attention weights, and the model focuses more on the deformation evolution trends of these high-risk areas.

[0143] The mechanically enhanced Transformer consists of multiple encoder and decoder layers. The encoder layer receives multi-scale spatiotemporal encoded features. and stress distribution Output encoding representation The decoder layer predicts the deformation field for future time periods based on the encoded representation:

[0144] ,

[0145] ,

[0146] in: For decoder output, For the predicted future deformation field.

[0147] The working process of the disaster risk assessment unit is as follows: based on the predicted future deformation field and stress evolution trend (Calculated by predicting the deformation field), combined with disaster threshold criteria to assess the probability of occurrence and the scope of impact of geological disasters such as landslides, ground collapses, and debris flows.

[0148] For landslide disasters, the slope stability coefficient is used as the criterion:

[0149] ,

[0150] in: The stability coefficient, For cohesion, It is normal stress. It is the internal friction angle. For shear stress. When The slope was in an unstable state and there was a risk of landslide.

[0151] For ground subsidence disasters, cumulative settlement and settlement rate are used as criteria:

[0152] ,

[0153] in: This represents the cumulative vertical settlement. For the settling rate, and The threshold value is used.

[0154] For debris flow disasters, multiple factors such as rainfall, slope, and deformation rate should be considered comprehensively:

[0155] ,

[0156] in: This represents the probability of a debris flow occurring. Rainfall intensity, For slope, For deformation rate, This is a probabilistic prediction model (which can employ logistic regression or a neural network).

[0157] Risk assessment is performed on each grid point within the monitoring area to generate a spatiotemporal prediction map of disaster evolution. This map identifies the spatial distribution of high-risk, medium-risk, and low-risk areas and their evolutionary trends over time.

[0158] Simultaneously, based on the prediction uncertainty of the Transformer model (obtained through multiple random sampling or ensemble predictions), the prediction confidence interval is calculated:

[0159] ,

[0160] in: The standard deviation is the prediction value, and the confidence interval reflects the reliability of the prediction result.

[0161] The multi-scale disaster evolution prediction module 3 ultimately outputs a spatiotemporal prediction map of disaster evolution and a prediction confidence interval. These outputs are then passed to the adaptive early warning decision module 4 for risk assessment and early warning triggering.

[0162] like Figure 5 As shown, the adaptive early warning decision-making module 4 includes a risk assessment unit, a dynamic threshold adjustment unit, an early warning triggering unit, a feedback evaluation unit, and a dual-path feedback unit. This module is coupled with the multi-scale disaster evolution prediction module 3 and the deformation spatiotemporal feature perception module 1 to form a complete closed-loop feedback system.

[0163] The risk assessment unit operates as follows: Based on the spatiotemporal prediction map of disaster evolution, the comprehensive disaster risk index for each monitoring point is calculated. The risk index comprehensively considers the probability of occurrence of multiple disaster types.

[0164]

[0165] in: For point The comprehensive risk index, These represent the probabilities of occurrence for landslides, ground subsidence, and debris flows, respectively. The weighting coefficient is determined based on the geological environmental characteristics of the monitoring area.

[0166] Based on the prediction confidence interval, determine the range of uncertainty for the risk level:

[0167] ,

[0168] in: As an indicator of risk uncertainty, This represents the width of the prediction interval for the probability of each disaster type.

[0169] The dynamic threshold adjustment unit operates as follows: Based on historical early warning data, it statistically analyzes the actual disaster occurrence rate corresponding to different risk levels and dynamically adjusts the risk level determination threshold. An early warning database is established to record the risk index, warning level, warning time, and subsequent actual disaster occurrence for each early warning.

[0170] Statistical analysis of historical data was used to calculate the early warning accuracy rate (true positive rate) and false positive rate (false alarm rate) corresponding to different risk index thresholds:

[0171] ,

[0172] ,

[0173] in: The risk index threshold, The true positive rate, The false positive rate, This represents the number of true positive results (correct warning). This represents the number of false negatives (underreporting). This represents the number of false positives (misreports). This represents the number of true negatives (correct results will not trigger a warning).

[0174] Using ROC curve analysis, select... The threshold that is maximized is taken as the optimal threshold:

[0175] ,

[0176] Different risk index thresholds are set for different warning levels (such as red warning, orange warning, and yellow warning):

[0177] ,

[0178] The dynamic threshold adjustment unit recalculates the optimal threshold based on the latest historical data at regular intervals (e.g., every month), thereby enabling dynamic adjustment and continuous optimization of the warning threshold.

[0179] The early warning triggering unit operates as follows: It iterates through all grid points within the monitoring area and determines whether the risk index of each point exceeds the threshold for the corresponding level. When the risk index of a certain area exceeds the threshold, an early warning signal is generated. The early warning signal contains the following information:

[0180] Warning levels (red / orange / yellow);

[0181] Warning location (geographic coordinates and extent of high-risk areas);

[0182] Disaster types (landslides / ground subsidence / debris flows);

[0183] Expected time of occurrence (based on deformation evolution trend prediction);

[0184] Scope of impact (area of ​​high-risk areas and threatened population and infrastructure);

[0185] Early warning confidence level (based on prediction confidence interval);

[0186] Warning signals are sent to relevant management departments and residents in affected areas through various means such as SMS, telephone, and online platforms, providing decision support for disaster prevention and emergency response.

[0187] The feedback evaluation unit works as follows: It records the actual occurrence of disasters after each warning and calculates the warning performance indicators. For each warning, it monitors whether a disaster occurs in the area during the observation period after the warning (e.g., 15 days), noting the actual type, scale, and timing of the disaster.

[0188] Calculate the following early warning performance indicators:

[0189] Early warning accuracy:

[0190] ,

[0191] Early warning time:

[0192] ,

[0193] in: The actual time of the disaster This refers to the time the warning was issued.

[0194] False alarm rate:

[0195] ,

[0196] Missed report rate:

[0197] ,

[0198] The above performance indicators are compared with the preset target values ​​to analyze the current early warning performance level of the system and its changing trend.

[0199] The working process of the dual-path feedback unit is as follows: This is the key unit for achieving closed-loop collaborative optimization in this invention. The dual-path feedback unit generates feedback adjustment signals based on the early warning performance indicators, and transmits them to the deformation spatiotemporal feature perception module 1 and the geomechanical constraint optimization module 2, respectively, forming a dual-path parameter optimization closed loop.

[0200] The specific implementation process of the dual-path feedback mechanism is as follows:

[0201] Feedback path 1: Optimize the parameters of deformation spatiotemporal feature perception module 1.

[0202] When false alarm rate A value that is too high (exceeding the target value) indicates that the system's ability to suppress noise and non-catastrophic deformation is insufficient, requiring an increase in the rigor of deformation consistency verification. Specific adjustment strategies:

[0203] Increase the threshold parameter in the spatiotemporal attention mechanism to enhance the suppression of low-confidence regions;

[0204] Adjust the regularization strength of the feature extraction network to reduce its sensitivity to noise signals;

[0205] Similarity calculation threshold for modified deformation consistency evaluation units :

[0206] ,

[0207] in: For learning rate, The target is the false alarm rate.

[0208] When the false negative rate When the value is too high (exceeding the target value), it indicates that the system's sensitivity to small deformation signals is insufficient, and the sensitivity of feature extraction needs to be enhanced. Specific adjustment strategies:

[0209] Lowering the threshold parameter in the spatiotemporal attention mechanism enhances the ability to capture signals of minute deformations;

[0210] Adjust the filtering intensity of the InSAR phase processing unit to find a better balance between noise reduction and detail preservation;

[0211] Adjust the activation threshold parameters of the feature extraction network.

[0212] When the early warning time If the reading is insufficient, it indicates that the system's ability to identify the accelerated deformation phase is inadequate, and its sensitivity to the rate of deformation change needs to be enhanced. Specific adjustment strategies:

[0213] Add extraction of the second derivative of deformation (acceleration) to the spatiotemporal feature extraction unit.

[0214] Adjust the weighting of the time attention mechanism, giving higher weight to recent deformations.

[0215] Feedback path 2: Optimize the parameters of geomechanical constraint optimization module 2.

[0216] When the early warning time If the sensitivity is insufficient, it indicates that the mechanical model is not sensitive enough to identifying stress accumulation and critical failure states, and the sensitivity parameters of the mechanical model need to be adjusted. Specific adjustment strategies:

[0217] Lowering the stress concentration threshold makes the system more sensitive to stress anomalies;

[0218] Adjusting the weights of physical constraint terms in a physical embedding neural network :

[0219] ,

[0220] in: For learning rate, Provide early warnings for targets.

[0221] When the accuracy of the early warning If the results are insufficient, it indicates a discrepancy in the identification of geological structures and the inference of mechanical parameters, requiring correction of the relevant parameters. Specific adjustment strategies:

[0222] Based on the actual location and type of the disaster, the accuracy of geological structure identification is analyzed in reverse, and the classification threshold of the graph convolutional network is adjusted accordingly.

[0223] Based on the actual stress state at the time of the disaster, the prior distribution of the inference model is modified by correcting the mechanical parameters;

[0224] Update the mechanical parameter database and incorporate new disaster cases into the training samples;

[0225] A collaborative optimization strategy is adopted for parameter adjustments of the two modules to ensure that parameter adjustments improve early warning performance while maintaining system stability. Specifically, a multi-objective optimization framework is used:

[0226] ,

[0227] ,

[0228] in: This is the parameter set for the deformation spatiotemporal feature sensing module 1. This is the parameter set for geomechanical constraint optimization module 2. These are the weighting coefficients for each performance indicator. This involves setting step size constraints for parameter adjustments to prevent drastic parameter changes from causing system instability.

[0229] Through the aforementioned dual-path feedback closed loop, the system achieves adaptive optimization based on early warning performance. After 3-5 monitoring cycles (each cycle lasting 1-3 months) of closed-loop iteration, the system's early warning accuracy can be improved from the initial 75% to over 95%, the false alarm rate can be reduced from 20% to below 5%, and the early warning time can be extended from 5 days to 10-15 days.

[0230] The system of this invention achieves a significant synergistic effect through the deep coupling and closed-loop feedback of four core modules, which is specifically reflected in the following aspects:

[0231] Synergistic effect 1: Dynamic matching of deformation confidence and mechanical constraint weights.

[0232] Deformation consistency confidence level generated by deformation spatiotemporal feature perception module 1 The weight configuration of physical constraint terms in the geomechanical constraint optimization module 2 is directly controlled. In the loss function of the physical embedded neural network, the weight of the observation data fitting loss is the confidence level. :

[0233] ,

[0234] This dynamic weighting mechanism ensures that: for observation points with high confidence (reliable deformation signals), mechanical constraints fit the observation data more closely; for observation points with low confidence (potentially containing noise), mechanical constraints rely more on physical laws for correction.

[0235] Comparative experiments show that after adopting dynamic weight configuration of confidence levels, the accuracy of deformation field anomaly identification increased from 68% to 91%, an improvement of 23 percentage points. In contrast, the accuracy of fixed weight configuration was only 70%, demonstrating the significant advantage of the dynamic matching mechanism.

[0236] Synergistic effect 2: Physical enhancement of stress distribution and disaster prediction.

[0237] Geomechanical constraint optimization module 2 outputs geological structure stress distribution Transformer attention computation embedded in multi-scale disaster evolution prediction module 3:

[0238] ,

[0239] This physical enhancement mechanism allows the prediction model to consider both statistical regularities and physical mechanisms. Comparative experiments show that the prediction accuracy of the pure statistical method (without embedding stress distribution) is 72%; the prediction accuracy after embedding stress distribution is 90%, an improvement of 18 percentage points.

[0240] More importantly, the physically enhanced prediction model has a stronger generalization ability. In new monitoring areas outside the training data, the prediction accuracy can still be maintained above 85%, while the accuracy of the pure statistical method drops to 55%.

[0241] Synergistic effect 3: Synergistic optimization of dual-path feedback.

[0242] The dual-path feedback signal of the adaptive early warning decision module 4 simultaneously optimizes the parameters of the deformation spatiotemporal feature perception module 1 and the geomechanical constraint optimization module 2, enabling the entire system to have adaptive learning capabilities.

[0243] In a long-term monitoring experiment in a mining area, the initial warning accuracy rate of the system was 75%, and the false alarm rate was 18%. After three monitoring cycles (each two months) of dual-path feedback optimization: at the end of the first cycle, the warning accuracy rate increased to 82%, and the false alarm rate decreased to 14%; at the end of the second cycle, the warning accuracy rate increased to 88%, and the false alarm rate decreased to 9%; at the end of the third cycle, the warning accuracy rate increased to 93%, and the false alarm rate decreased to 5%.

[0244] In contrast, systems employing single-path feedback (optimizing only the feature perception module or only the mechanical constraint module) only improve warning accuracy by 8-10 percentage points, far less than the 18 percentage point improvement achieved with dual-path feedback. This demonstrates the 1+1>2 effect of dual-path collaborative optimization.

[0245] Synergistic effect 4: The performance of closed-loop iteration is continuously improved.

[0246] Traditional open-loop systems maintain relatively constant early warning performance and cannot learn from or improve upon historical data. The system of this invention, through a closed-loop feedback mechanism, continuously improves its performance as the monitoring cycle increases.

[0247] Comparative experiments across multiple monitoring areas show that: open-loop system (no feedback optimization): warning accuracy remains stable at 73%–75%; single-cycle closed-loop system (optimized once): warning accuracy improves to 80%–82%; multi-cycle closed-loop system (optimized 3–5 times): warning accuracy improves to 92%–95%.

[0248] The mechanism by which closed-loop iteration continuously optimizes system performance is as follows: after each warning, the system analyzes the shortcomings of the current parameter configuration based on the actual disaster occurrence, and adjusts key parameters such as feature extraction strategy, mechanical constraint strength, and prediction model weights in a targeted manner, so that the system gradually adapts to the specific geological environment and disaster evolution law of the monitoring area.

[0249] In summary, the system of this invention achieves the organic integration of deformation perception, mechanical constraints, disaster prediction, and early warning decision-making through deep module coupling, closed-loop feedback, and collaborative optimization. The modules promote each other and have synergistic effects, forming an overall performance that far exceeds that of individual modules used independently. This provides an innovative technical solution for high-precision terrain change monitoring and geological disaster early warning.

[0250] Example 1: Monitoring of Surface Subsidence in Mining Areas

[0251] In a coal mining area, the system of this invention was used to monitor surface subsidence. The monitoring area covered approximately 30 square kilometers, and Sentinel-1 satellite InSAR monitoring and ground GNSS reference stations were deployed. During six months of system operation, millimeter-level surface subsidence was successfully detected, with a maximum cumulative subsidence of 180 mm. The system provided 12-day advance warnings for three areas of accelerated subsidence, achieving a 100% accuracy rate in warnings. Compared to traditional ground measurement methods, the monitoring efficiency of this invention is increased tenfold, and the cost is reduced by 70%.

[0252] Example 2: Landslide Early Warning in Mountainous Areas

[0253] Along a mountainous highway, the system of this invention was used for landslide monitoring and early warning. The monitoring area covered 60 kilometers of landslide-prone sections of the highway. During one year of operation, the system identified 15 potential landslide hazard points, of which 12 experienced landslides of varying degrees within the following three months, achieving an early warning accuracy rate of 80%. The system issued warnings an average of nine days in advance, providing road management departments with sufficient emergency response time and preventing major traffic accidents and casualties.

[0254] Example 3: Urban Ground Subsidence Monitoring

[0255] In a coastal city, the system of this invention was used for ground subsidence monitoring. The monitoring area covered approximately 200 square kilometers, including the main urban area and newly developed zones. The system identified 18 significant subsidence areas, 8 of which were located in blind spots of the traditional monitoring network. Through multi-scale disaster evolution prediction, the system predicted that the subsidence rate in 3 areas would accelerate significantly within the next year, and recommended strengthening engineering control measures. Actual monitoring verified that the system's prediction accuracy reached 90%, providing a scientific basis for urban planning and infrastructure safety management.

[0256] The above application examples demonstrate that the system of the present invention can achieve high-precision monitoring and reliable early warning in different geological environments and application scenarios, and has broad application value and promotion prospects.

[0257] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A high-precision image analysis system for monitoring topographic changes and providing early warning of geological disasters, characterized in that: The system includes: The deformation spatiotemporal feature perception module is used to acquire multi-source time-series image data of the monitoring area, extract millimeter-level deformation information using InSAR phase unwrapping and deep learning fusion technology, and generate deformation spatiotemporal feature map and deformation consistency confidence based on spatiotemporal attention mechanism, wherein the deformation consistency confidence is used to characterize the reliability of deformation measurement. The geomechanical constraint optimization module, coupled with the deformation spatiotemporal feature perception module, is used to receive the deformation spatiotemporal feature map and deformation consistency confidence, construct mechanical constraint conditions based on the geological structure characteristics of the monitoring area, and use a physical embedded neural network to perform mechanical rationality verification and abnormal deformation identification on the deformation spatiotemporal feature map, generating an optimized deformation field and geological structure stress distribution under mechanical constraints, wherein the deformation consistency confidence directly determines the weight configuration of the mechanical constraint conditions; The multi-scale disaster evolution prediction module, coupled with the geomechanical constraint optimization module, is used to receive the optimized deformation field and the stress distribution of the geological structure. It uses a spatiotemporal convolutional Transformer network to extract the multi-scale spatiotemporal pattern of deformation evolution, and combines geomechanical parameters to predict the deformation trend and disaster risk level in the future period, generating a spatiotemporal prediction map of disaster evolution and a prediction confidence interval. The stress distribution of the geological structure is embedded as a constraint condition in the attention calculation of the prediction network. The adaptive early warning decision module, coupled with the multi-scale disaster evolution prediction module and the deformation spatiotemporal feature perception module, is used to perform risk assessment based on the disaster evolution spatiotemporal prediction map and prediction confidence interval. When the risk level exceeds the dynamic threshold, an early warning is triggered. The module calculates the early warning accuracy index based on the early warning response effect and the actual occurrence of the disaster. The early warning accuracy index is fed back to the deformation spatiotemporal feature perception module to adjust the spatiotemporal perception parameters of feature extraction. At the same time, it is fed back to the geomechanical constraint optimization module to correct the mechanical model parameters, forming a dual-path feedback closed loop of deformation perception optimization and mechanical model correction.

2. The system according to claim 1, characterized in that, The deformation spatiotemporal feature sensing module includes: The InSAR phase processing unit is used to perform interferometric processing on multi-source SAR images. It uses a phase unwrapping algorithm based on deep neural networks to eliminate 2π ambiguity and uses an atmospheric delay correction network to remove the influence of atmospheric disturbances, generating a high-precision phase deformation field. The spatiotemporal feature extraction unit, connected to the InSAR phase processing unit, is used to convert the phase deformation field into a deformation time series, and to extract the spatial distribution features and temporal evolution features of the deformation using a 3D convolutional neural network to generate a multi-scale spatiotemporal feature pyramid. The deformation consistency evaluation unit, connected to the spatiotemporal feature extraction unit, is used to identify consistent and abnormal regions in the deformation field based on the multi-scale spatiotemporal feature pyramid and a spatiotemporal attention mechanism, calculate the deformation consistency confidence of each pixel, and generate a confidence distribution map. The spatiotemporal fusion unit, connected to the spatiotemporal feature extraction unit and the deformation consistency evaluation unit, is used to perform weighted fusion of multi-scale spatiotemporal features based on the confidence distribution map to generate a deformation spatiotemporal feature map.

3. The system according to claim 2, characterized in that, The spatiotemporal attention mechanism is implemented in the following ways: The deformation time series is divided into time windows, and the attention weight of the time dimension is calculated for the deformation data in each time window. The weight reflects the importance of the deformation signal in that time period. The attention weights of the spatial dimensions of the deformation spatial distribution are calculated, and the weights reflect the reliability of the deformation signal at that spatial location. The temporal attention weights and spatial attention weights are jointly calculated to generate a spatiotemporal joint attention matrix, which is used to enhance the consistency of deformation signals and suppress noise interference.

4. The system according to claim 1, characterized in that, The geomechanical constraint optimization module includes: The geological structure identification unit is used to identify geological structural elements such as faults, fissures, and rock strata interfaces in the monitored area based on the deformation spatiotemporal feature map and to generate a geological structure topology map. The mechanical parameter inference unit, connected to the geological structure identification unit, is used to infer the mechanical parameters of each geological structure unit, including elastic modulus, Poisson's ratio, cohesion and internal friction angle, based on the geological structure topology map and historical geological survey data. The physical constraint verification unit, connected to the mechanical parameter inference unit, is used to use the deformation spatiotemporal feature map as observation data, combine it with the mechanical parameters to construct a mechanical equilibrium equation, use a physical embedded neural network to solve the stress-strain field, and identify and correct abnormal deformations that do not meet the mechanical equilibrium conditions. The stress distribution calculation unit, connected to the physical constraint verification unit, is used to calculate the stress distribution state of the geological structure based on the corrected deformation field, and to identify stress concentration areas and potential failure surfaces.

5. The system according to claim 4, characterized in that, The physical embedded neural network implements mechanical constraints in the following way: A mechanical equilibrium equation constraint term is embedded in the loss function of the neural network. This constraint term is based on the finite element mechanical model to calculate the stress-strain deviation. In the forward propagation process of the neural network, prior mechanical knowledge is introduced to limit the range of values ​​for deformation gradient and stress components; By simultaneously optimizing network parameters and mechanical parameters through the backpropagation algorithm, the predicted deformation field can be made to conform to both the observed data and the mechanical equilibrium conditions.

6. The system according to claim 1, characterized in that, The multi-scale disaster evolution prediction module includes: The multi-scale feature decoupling unit is used to decompose the optimized deformation field into long-term trend components, seasonal cycle components and short-term fluctuation components, and extract deformation evolution features at different time scales respectively. The spatiotemporal convolutional coding unit, connected to the multi-scale feature decoupling unit, is used to extract spatial and temporal dependencies of each time-scale component using a spatiotemporal convolutional network to generate multi-scale spatiotemporal coding features. A mechanical enhancement Transformer unit, connected to the spatiotemporal convolutional coding unit, is used to receive the multi-scale spatiotemporal coding features and embed the stress distribution of the geological structure as a constraint condition into the attention calculation mechanism of the Transformer, so that the prediction process considers both statistical laws and physical mechanisms. The disaster risk assessment unit, connected to the mechanical enhancement Transformer unit, is used to assess the probability and impact range of geological disasters such as landslides, ground subsidence, and debris flows based on the predicted future deformation field and stress evolution trend, combined with disaster threshold criteria, and to generate a disaster evolution spatiotemporal prediction map and prediction confidence interval.

7. The system according to claim 6, characterized in that, The attention calculation in the mechanically enhanced Transformer unit uses the following formula: , in: For querying the matrix, The key matrix, For value matrices, This is the stress distribution matrix of the geological structure. The dimension of the key vector. The mechanical constraint weighting coefficient is used in this formula by applying stress distribution. Embedded attention computing enables prediction results to simultaneously reflect statistical patterns and mechanical mechanisms.

8. The system according to claim 1, characterized in that, The adaptive early warning decision module includes: The risk assessment unit is used to calculate the disaster risk index of each monitoring point based on the disaster evolution spatiotemporal prediction map, and to determine the uncertainty range of the risk level according to the prediction confidence interval; A dynamic threshold adjustment unit, connected to the risk assessment unit, is used to statistically analyze the actual disaster occurrence rate corresponding to different risk levels based on historical early warning data, and dynamically adjust the risk level judgment threshold to reduce false alarm rate and false alarm rate. The early warning triggering unit is connected to the dynamic threshold adjustment unit and is used to generate an early warning signal when the risk index exceeds the dynamic threshold, and to determine the early warning level and early warning range according to the risk level and the scope of impact. The feedback evaluation unit, connected to the early warning triggering unit, is used to record the actual occurrence of the disaster after each early warning and to calculate performance indicators such as early warning accuracy, early warning time, and false alarm rate. A dual-path feedback unit, connected to the feedback evaluation unit, the deformation spatiotemporal feature perception module, and the geomechanical constraint optimization module, is used to generate feedback adjustment signals based on the performance indicators. On the one hand, the feedback signals are transmitted to the deformation spatiotemporal feature perception module to adjust the spatiotemporal perception parameters and attention mechanism weights. On the other hand, the feedback signals are transmitted to the geomechanical constraint optimization module to correct the geological structure identification threshold and mechanical parameter inference model, forming a dual-path parameter optimization closed loop driven by early warning performance.

9. The system according to claim 8, characterized in that, The parameter optimization strategy of the dual-path feedback unit is as follows: For the deformation spatiotemporal feature sensing module, when the false alarm rate is too high, the strictness of the spatiotemporal consistency check is increased, and when the false alarm rate is too high, the sensitivity to small deformation signals is enhanced. For the geomechanical constraint optimization module, when the early warning time is insufficient, the sensitivity parameters of the mechanical model are adjusted, and when the early warning accuracy is insufficient, the feature threshold for geological structure identification is corrected. A collaborative optimization strategy is adopted for parameter adjustments of the two modules to ensure that parameter adjustments improve early warning performance while maintaining system stability.

10. The system according to claim 1, characterized in that, The system achieves inter-module synergy in the following ways: The deformation consistency confidence level generated by the deformation spatiotemporal feature perception module directly controls the weight configuration of mechanical constraint conditions in the geomechanical constraint optimization module, so that the strength of mechanical constraints matches the reliability of deformation measurement. The geological structure stress distribution output by the geomechanical constraint optimization module is embedded into the attention calculation of the multi-scale disaster evolution prediction module, so that the prediction process considers both statistical laws and physical mechanisms. The dual-path feedback signal of the adaptive early warning decision module simultaneously optimizes the parameters of the deformation spatiotemporal feature perception module and the geomechanical constraint optimization module.

Images