Intelligent Identification System and Method for Earthquake Surface Fault Zones

Abstract

This invention relates to the field of earthquake disaster emergency monitoring technology, specifically to an intelligent system and method for identifying earthquake surface rupture zones. The system includes modules for heterogeneous data registration, feature extraction and fusion, rupture zone segmentation, active learning, and morphological optimization. It achieves accurate processing of remote sensing data and intelligent identification of rupture zones. The system employs multi-scale feature point matching and adaptive weight registration, improving data registration accuracy by over 30% and overcoming the reliance on human experience in traditional methods. Through multi-modal feature decoupling and attention-guided feature fusion, combined with multi-scale spatial pyramid pooling and deformation-sensitive segmentation, the system effectively identifies rupture zones and assesses prediction uncertainty. The active learning module optimizes training data through Bayesian estimation and a hybrid sampling strategy, while the morphological optimization module utilizes geological constraints and fault topology inference to ultimately output accurate rupture zone distribution and attribute information, improving the intelligence level and efficiency of earthquake disaster emergency monitoring.

Description

Technical Field

[0001] This invention relates to the field of earthquake disaster emergency monitoring technology, specifically to an intelligent identification system and method for earthquake surface rupture zones, used to quickly and accurately identify the spatial distribution and deformation characteristics of surface rupture zones caused by earthquakes. Background Technology

[0002] Earthquake surface rupture zones are crucial for earthquake disaster assessment and monitoring, and their accurate identification is essential for earthquake emergency response, post-disaster reconstruction, and tectonic activity research. Currently, the identification of earthquake surface rupture zones mainly relies on a combination of remote sensing technology and field surveys.

[0003] Existing technologies, such as patent CN114964028B, disclose a method for rapidly interpreting earthquake surface rupture zones using integrated remote sensing. This method comprehensively utilizes multi-source remote sensing data, including high-resolution optical satellite remote sensing images, SAR images, UAV orthophotos, and LiDAR point cloud data, and rapidly and accurately determines the spatial distribution of earthquake surface ruptures and surface deformation characteristics through various image processing techniques.

[0004] However, the existing technical solution has the following shortcomings: ① The data fusion process is highly dependent on human experience, the registration accuracy of remote sensing data from different sources is limited, and the fusion process lacks an adaptive mechanism; ② The accuracy of rupture zone boundary identification is insufficient, especially in complex terrain and vegetated areas, where the boundaries are blurred; ③ The cost of sample acquisition is high, requiring a large number of manual sampling points and labeled data; ④ The ability to identify severely decoherent and multi-segment discontinuous rupture zones is limited; ⑤ It does not make full use of the professional knowledge of rupture zone morphology in seismic geology for guidance.

[0005] Therefore, there is an urgent need for a system and method that can automatically and intelligently identify earthquake surface rupture zones in order to improve identification accuracy, reduce labor costs, and speed up response. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent identification system and method for earthquake surface rupture zones, aiming to solve the technical problems existing in the prior art, such as data fusion relying on human experience, insufficient accuracy of rupture zone boundary identification, high cost of sample acquisition, inability to handle abnormal situations, and lack of morphological knowledge guidance.

[0007] This invention proposes an intelligent identification system for earthquake surface rupture zones, comprising:

[0008] The heterogeneous data registration module is used to receive multi-source remote sensing data, perform heterogeneous data registration using a multi-scale feature point matching algorithm and an adaptive weight registration loss function, and generate registered multi-source data and a registration accuracy evaluation report.

[0009] The feature extraction and fusion module is connected to the heterogeneous data registration module. It is used to receive the registered multi-source data, decompose the features of each data source into edge features, deformation features and context features through a multimodal feature decoupling network, and dynamically fuse complementary features using an attention-guided feature fusion machine to output a fused feature map.

[0010] The rupture zone segmentation module, connected to the feature extraction and fusion module, is used to receive the fused feature map, identify the rupture zone using multi-scale spatial pyramid pooling and deformation-sensitive loss function, and generate preliminary rupture zone segmentation results and prediction uncertainties.

[0011] An active learning module, connected to the rupture zone segmentation module, is used to receive the preliminary rupture zone segmentation results and prediction uncertainties, determine the optimal sampling points based on Bayesian uncertainty estimation and a hybrid sampling strategy, and update the training dataset.

[0012] The morphology optimization module, connected to the active learning module and the rupture zone segmentation module, is used to receive the updated training dataset and the preliminary rupture zone segmentation results, optimize the rupture zone morphology based on the geological constraint energy function and fault system topology reasoning, and output the final earthquake surface rupture zone distribution and attribute information.

[0013] Preferably, the heterogeneous data registration module includes:

[0014] The data preprocessing unit is used to perform radiometric correction, geometric correction, and noise filtering on the received multi-source remote sensing data.

[0015] The feature extraction unit is used to extract robust feature points from different types of remote sensing images;

[0016] The registration loss calculation unit is used to calculate the registration error based on the extracted feature points using an adaptive weighted registration loss function.

[0017] An iterative refinement unit is used to achieve accurate registration of multi-source data through a cascaded architecture from coarse registration to fine registration based on the registration error.

[0018] The accuracy assessment unit is used to quantitatively evaluate the registration results and generate accuracy reports and uncertainty area markers.

[0019] Preferably, the feature extraction and fusion module includes:

[0020] The multimodal encoder unit is used to perform specialized feature encoding on various types of registered data;

[0021] The feature decoupling unit is used to decompose the features of each data source into edge features, deformation features, and context features;

[0022] The attention calculation unit is used to calculate the feature weight coefficients and generate dynamic feature fusion weights.

[0023] The feature fusion unit is used to perform weighted fusion of multi-source features based on the dynamic feature fusion weights;

[0024] The feature consistency enhancement unit is used to force different source features to align in the semantic space through adversarial training.

[0025] Preferably, the fracture zone segmentation module includes:

[0026] The basic feature extraction unit is used to extract the basic representation of the fused feature map;

[0027] Multi-scale feature enhancement unit is used to capture multi-scale features of the fracture zone through spatial pyramid pooling;

[0028] Deformation-sensitive decoding unit is used to enhance the learning weights in areas with large deformation gradients in the rupture zone;

[0029] Regional affinity enhancement units are used to establish long-range spatial dependencies and enhance the spatial continuity of the fracture zone;

[0030] The segmentation result generation unit is used to output the segmentation results of the fracture zone and its probability distribution.

[0031] Preferably, the active learning module includes:

[0032] Uncertainty estimation unit, used to evaluate the uncertainty of model predictions using the Monte Carlo Dropout method;

[0033] The sampling strategy unit is used to calculate the optimal sampling point score based on uncertainty, representativeness, and accessibility.

[0034] Incremental annotation units are used to acquire new sample annotations and update the training dataset;

[0035] The model update unit is used to retrain the recognition model based on the updated training dataset.

[0036] The annotation efficiency evaluation unit is used to calculate the ratio of the increase in the number of annotations to the improvement in accuracy, and to determine whether to continue sampling and annotation.

[0037] Preferably, the morphology optimization module includes:

[0038] Geologically constrained units are used to construct energy functions that integrate physical rules and data-driven predictions.

[0039] Fault topological units are used to establish the network structure of rupture zones and to infer possible fault connections.

[0040] A morphology optimization unit is used for multi-scale morphology optimization based on hierarchical Markov random fields.

[0041] The knowledge base interaction unit is used to extract typical morphological features and regional tectonic background information from the fault knowledge base.

[0042] The results output unit is used to generate the final earthquake surface rupture zone distribution map and attribute report.

[0043] Preferably, the adaptive weight registration loss function is:

[0044] L reg =ɑ·L intensity +β·L geometric +γ·L context ,

[0045] Where α, β, and γ are adaptive weighting coefficients, dynamically adjusted based on data quality; L intensity For intensity similarity loss; L geometric For geometric consistency loss; L context This is for context similarity loss.

[0046] Preferably, the deformation-sensitive loss function is:

[0047]

[0048] Where y is the true label and p is the predicted probability. Let λ be the deformation gradient and λ be the weighting factor.

[0049] Preferably, the multi-source remote sensing data includes: high-resolution optical satellite remote sensing images, SAR images, UAV orthophotos, and LiDAR point cloud data; the earthquake surface rupture zone attribute information includes: rupture zone location, strike, deformation, segmentation characteristics, and uncertainty assessment.

[0050] Intelligent identification methods for earthquake surface rupture zones include:

[0051] Receive multi-source remote sensing data, use a multi-scale feature point matching algorithm and an adaptive weight registration loss function to register heterogeneous data, and generate registered multi-source data and a registration accuracy evaluation report.

[0052] The registered multi-source data is received, and the features of each data source are decomposed into edge features, deformation features and context features through a multimodal feature decoupling network. Complementary features are dynamically fused using an attention-guided feature fusion machine, and a fused feature map is output.

[0053] The fused feature map is received, and multi-scale spatial pyramid pooling and deformation-sensitive loss function are used to identify the fracture zone, generating preliminary fracture zone segmentation results and prediction uncertainty.

[0054] Receive the preliminary rupture zone segmentation results and prediction uncertainties, determine the optimal sampling points based on Bayesian uncertainty estimation and a hybrid sampling strategy, and update the training dataset;

[0055] The system receives the updated training dataset and the preliminary rupture zone segmentation results, optimizes the rupture zone morphology based on the geological constraint energy function and fault system topology reasoning, and outputs the final earthquake surface rupture zone distribution and attribute information.

[0056] The present invention has the following beneficial effects:

[0057] 1. Through heterogeneous data adaptive registration technology, high-precision automatic registration of remote sensing data from different sources was achieved, solving the problem of data fusion relying on human experience in traditional methods and improving registration accuracy by more than 30%;

[0058] 2. By adopting a hierarchical feature extraction and fusion mechanism, complementary information from different data sources is effectively integrated, enhancing the system's ability to identify complex terrain and vegetated areas, and improving the identification accuracy by 35%.

[0059] 3. The uncertainty-guided active learning framework significantly reduces sample acquisition costs, decreasing sample requirements by 65% ​​and substantially reducing manual annotation costs;

[0060] 4. By incorporating knowledge-guided rupture zone morphology optimization algorithms and seismic geology expertise, the system's ability to process multiple discontinuous rupture zones has been enhanced, and the boundary extraction accuracy has been improved to the sub-meter level.

[0061] 5. The system's processing efficiency has been greatly improved, reducing the interpretation time from several hours in traditional methods to minutes, thus meeting the needs of rapid response to earthquake disasters. Attached Figure Description

[0062] Figure 1 This is an overall architecture diagram of the intelligent identification system for earthquake surface rupture zones provided in an embodiment of the present invention;

[0063] Figure 2 This is a structural diagram of the heterogeneous data registration module provided in an embodiment of the present invention;

[0064] Figure 3 This is a structural diagram of the feature extraction and fusion module provided in an embodiment of the present invention;

[0065] Figure 4 This is a structural diagram of the fracture zone segmentation module provided in an embodiment of the present invention;

[0066] Figure 5 This is a structural diagram of the active learning module provided in an embodiment of the present invention;

[0067] Figure 6 This is a structural diagram of the morphology optimization module provided in an embodiment of the present invention;

[0068] Figure 7 A schematic diagram of a multimodal feature decoupling network provided in an embodiment of the present invention;

[0069] Figure 8 A comparison chart of the effects before and after optimization of the geological constraint energy function provided in this embodiment of the invention;

[0070] Figure 9 A flowchart of the intelligent identification method for earthquake surface rupture zones provided in an embodiment of the present invention. Detailed Implementation

[0071] Please refer to the attached document. Figure 1-9 The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0072] Example 1

[0073] like Figure 1 As shown, this embodiment provides an intelligent identification system for earthquake surface rupture zones, including a heterogeneous data registration module 1, a feature extraction and fusion module 2, a rupture zone segmentation module 3, an active learning module 4, and a morphological optimization module 5.

[0074] The heterogeneous data registration module 1 receives multi-source remote sensing data and performs heterogeneous data registration using a multi-scale feature point matching algorithm and an adaptive weighted registration loss function, generating registered multi-source data and a registration accuracy assessment report. The multi-source remote sensing data includes high-resolution optical satellite remote sensing imagery, SAR imagery, UAV orthophotos, and LiDAR point cloud data. The heterogeneous data registration module 1 employs an innovative multi-scale feature point matching algorithm, capable of adapting to differences in scale, resolution, and imaging mechanisms among different source remote sensing data, achieving high-precision registration.

[0075] Feature extraction and fusion module 2 is connected to heterogeneous data registration module 1. It receives registered multi-source data and decomposes the features of each data source into edge features, deformation features, and context features through a multimodal feature decoupling network. It then uses an attention-guided feature fusion engine to dynamically fuse complementary features and output a fused feature map. This module addresses the problem that complementary features from different source data are difficult to fuse effectively, by optimizing the fusion process through a dynamic weighting mechanism.

[0076] The rupture zone segmentation module 3 is connected to the feature extraction and fusion module 2. It receives the fused feature map and uses multi-scale spatial pyramid pooling and a deformation-sensitive loss function to identify the rupture zone, generating preliminary rupture zone segmentation results and prediction uncertainties. This module improves the accuracy of rupture zone boundary identification by capturing multi-scale features of the rupture zone and enhancing the learning weights for large deformation gradient regions.

[0077] The active learning module 4 is connected to the rupture zone segmentation module 3. It receives the preliminary rupture zone segmentation results and prediction uncertainties, determines the optimal sampling points based on Bayesian uncertainty estimation and a hybrid sampling strategy, and updates the training dataset. This module significantly reduces the required sample size and lowers the cost of manual annotation through its uncertainty-guided sampling strategy.

[0078] The morphology optimization module 5 is connected to the active learning module 4 and the fault zone segmentation module 3. It receives updated training datasets and preliminary fault zone segmentation results, optimizes the fault zone morphology based on geologically constrained energy functions and fault system topological reasoning, and outputs the final distribution and attribute information of seismic surface fault zones. This module incorporates seismic geology expertise, optimizing fault zone morphology through physical constraints and graph structure reasoning to improve the geological plausibility of the identification results.

[0079] During system operation, the data flow path between modules is as follows: Multi-source remote sensing data is first processed by heterogeneous data registration module 1 to generate registered multi-source data and registration accuracy assessment report; then, feature extraction and fusion module 2 extracts and fuses features to output fused feature map; next, rupture zone segmentation module 3 performs preliminary identification to generate rupture zone segmentation results and prediction uncertainty; active learning module 4 updates the training dataset based on the uncertainty-guided sampling strategy; finally, morphological optimization module 5 applies geological constraints and topological reasoning to output the final distribution and attribute information of seismic surface rupture zones.

[0080] The above modules work together to form a complete intelligent identification system for earthquake surface rupture zones, realizing fully automated processing from multi-source remote sensing data input to the final rupture zone identification result output.

[0081] Example 2

[0082] like Figure 2 As shown, this embodiment provides a detailed structure of the heterogeneous data registration module. The heterogeneous data registration module 1 includes a data preprocessing unit 11, a feature extraction unit 12, a registration loss calculation unit 13, an iterative refinement unit 14, and an accuracy evaluation unit 15.

[0083] The data preprocessing unit 11 is used to perform radiometric correction, geometric correction, and noise filtering on the received multi-source remote sensing data. For optical images, it mainly performs atmospheric correction, radiometric calibration, and geometric correction; for SAR images, it mainly performs scattering correction, topographic correction, and speckle noise filtering; for LiDAR point cloud data, it mainly performs point cloud filtering and coordinate transformation.

[0084] Feature extraction unit 12 is used to extract robust feature points from different types of remote sensing images. This unit employs a dual-branch feature extraction network to extract robust feature points from different types of images. For optical images, it mainly extracts SIFT feature points and texture features; for SAR images, it mainly extracts KAZE feature points and edge features; and for LiDAR point cloud data, it mainly extracts geometric structure features.

[0085] The registration loss calculation unit 13 is used to calculate the registration error based on the extracted feature points using an adaptive weighted registration loss function. This loss function is expressed as:

[0086] L reg =α·L intensity +β·L geometric +γ·L context ,

[0087] Where α, β, and γ are adaptive weighting coefficients, dynamically adjusted based on data quality; L intensity Intensity similarity loss is used to measure the degree of matching of image grayscale values; L geometric Geometric consistency loss, used to measure the degree of matching of spatial geometry; L context Context similarity loss is used to measure the degree of matching of regional structural features.

[0088] Iterative refinement unit 14 is used to achieve accurate registration of multi-source data based on registration errors through a cascaded architecture from coarse to fine registration. This unit employs an iterative update mechanism.

[0089] T i+1 =T i +δT i ,

[0090] Among them, T i Let δT be the transformation matrix for the i-th iteration. i To refine the increments, a global registration strategy is initially used to obtain a coarse transformation matrix, and then the registration accuracy is gradually improved through multiple iterations.

[0091] The accuracy assessment unit 15 is used to quantitatively evaluate the registration results, generating an accuracy report and uncertainty region markers. This unit uses multiple indicators to evaluate registration accuracy, including root mean square error, registration point matching rate, and geometric consistency index. For areas with low registration accuracy, uncertainty region markers are generated to guide subsequent processing and give special attention to these areas.

[0092] Preferably, the heterogeneous data registration module also establishes a multi-source heterogeneous data registration parameter library, which stores the optimal registration parameter set for different data type pairs (such as optical-SAR, SAR-LiDAR, optical-LiDAR, etc.) and different terrain types (such as plains, mountains, hills, basins, etc.), including feature extraction thresholds, matching tolerances, iteration termination conditions, and registration accuracy indicators, in order to accelerate the registration process and improve accuracy.

[0093] Example 3

[0094] like Figure 3 As shown, this embodiment provides a detailed structure of the feature extraction and fusion module. The feature extraction and fusion module 2 includes a multimodal encoder unit 21, a feature decoupling unit 22, an attention calculation unit 23, a feature fusion unit 24, and a feature consistency enhancement unit 25.

[0095] The multimodal encoder unit 21 is used for specialized feature encoding of various registered data. This unit employs specially designed encoders for different types of data, including optical image encoders, SAR image encoders, and point cloud data encoders. The optical image encoder uses an improved ResNet-50 structure to extract multi-level visual features; the SAR image encoder uses an improved U-Net structure to focus on deformation and texture feature extraction; and the point cloud data encoder uses a PointNet++ structure to extract three-dimensional spatial features.

[0096] The feature decoupling unit 22 is used to decompose the features of each data source into edge features, deformation features, and context features. For example... Figure 7 As shown, this unit employs a multimodal feature decoupling network, achieving decoupling of features at different semantic levels through feature separation and recombination. Feature decoupling is represented as:

[0097] F source =[F edge ,F deform ,F context ],

[0098] Among them, F source F represents the original features of the data source. edge The edge features are represented, mainly including the boundary information F of the rupture zone. deform It represents deformation characteristics, mainly including surface displacement and deformation information; F context This indicates contextual features, which mainly include information about the surrounding terrain and features.

[0099] Attention calculation unit 23 is used to calculate feature weight coefficients and generate dynamic feature fusion weights. This unit employs a mechanism combining channel attention and spatial attention to dynamically adjust the importance of different features. The weight calculation is expressed as follows:

[0100] w i =softmax(MLP(concat[F i Global_context]))

[0101] Among them, w i For feature F i The weight coefficients are: MLP stands for Multilayer Perceptron, Global_context is the global context information, concat represents the feature concatenation operation, and softmax is the normalization function to ensure that the sum of all weights is 1.

[0102] Feature fusion unit 24 is used to perform weighted fusion of multi-source features based on dynamic feature fusion weights. The fusion process is represented as follows:

[0103] F fused =∑(w i ·F i ),

[0104] Among them, F fused For the characteristics after fusion, w i For dynamically generated feature weights, F i These are the features to be fused. This unit dynamically adjusts the fusion weights based on the importance and reliability of the features to ensure that key features are fully represented in the fusion result.

[0105] The feature consistency enhancement unit 25 is used to force different source features to align in the semantic space through adversarial training. This unit achieves feature consistency enhancement by minimizing the distance between different source features. The consistency loss is expressed as:

[0106] L consist =||F optical -F SAR || 2 +||F optical -F LiDAR || 2 +||F SAR -F LiDAR || 2 ,

[0107] Among them, F optical F SAR and F LiDAR Representing the features extracted from optical imagery, SAR imagery, and LiDAR point cloud data, respectively, ||·‖ 2 This represents the squared Euclidean distance. By minimizing this loss function, we can maintain consistency between different source features in the semantic space, thereby improving the fusion effect.

[0108] Preferably, the feature extraction and fusion module also implements a multi-resolution feature pyramid structure, which can perform feature fusion at different scales, adapt to the multi-scale characteristics of the rupture zone, and enhance the system's ability to identify rupture zones of different scales.

[0109] Example 4

[0110] like Figure 4 As shown, this embodiment provides a detailed structure of the rupture zone segmentation module. The rupture zone segmentation module 3 includes a basic feature extraction unit 31, a multi-scale feature enhancement unit 32, a deformation-sensitive decoding unit 33, a region affinity enhancement unit 34, and a segmentation result generation unit 35.

[0111] The basic feature extraction unit 31 is used to extract the basic representation of the fused feature map. This unit uses an improved ResNet-50 as the backbone network to extract multi-level feature representations. Deep features mainly contain semantic information, while shallow features mainly contain detailed information. The combination of the two can comprehensively express the features of the rupture zone.

[0112] Multi-scale feature enhancement unit 32 is used to capture multi-scale features of the fracture zone through spatial pyramid pooling. This unit implements the multi-scale spatial pyramid pooling operation:

[0113] SPP(F)=concat[Pool1(F),Pool2(F),...,Pool n (F)],

[0114] Where SPP represents the spatial pyramid pooling operation, F is the input feature, and Pool... i The `concat` operator represents a feature concatenation operation, indicating pooling operations at different scales. Through multi-scale pooling, the system can simultaneously capture fracture zone features at different scales, enhancing its ability to identify fracture zones with complex morphologies.

[0115] The deformation-sensitive decoding unit 33 is used to enhance the learning weights for regions with large deformation gradients in the rupture zone. This unit employs a deformation-sensitive loss function:

[0116]

[0117] Where y is the true label and p is the predicted probability. Let λ be the deformation gradient, and λ be a weighting factor used to adjust the influence of the deformation gradient on the loss function. This loss function assigns higher learning weights to regions with large deformation gradients, thereby improving the recognition accuracy of severely deformed regions.

[0118] The regional affinity enhancement unit 34 is used to establish long-range spatial dependencies and enhance the spatial continuity of the fracture zone. This unit uses nonlocal operations to establish long-range spatial dependencies:

[0119] F nonlocal (x i )=∑φ(x i ,x j )·g(x j ),

[0120] Among them, F nonlocal (x i ) represents the nonlocal feature at position i, φ(x) i ,x j ) Calculate the similarity between positions i and j, g(x j Let be the feature transformation at position j. Through this nonlocal operation, the system can capture the dependencies between distant locations, which helps to recover discontinuous fracture zones.

[0121] The segmentation result generation unit 35 is used to output the segmentation results of the rupture zone and its probability distribution. This unit converts the feature map into the segmentation result through the decoder network, and at the same time generates a prediction uncertainty map to guide the subsequent active learning process.

[0122] Preferably, the rupture zone segmentation module also integrates a dual-channel segmentation strategy, with one channel responsible for rupture zone location identification and the other channel responsible for deformation prediction. The combination of the two generates a more comprehensive rupture zone characterization.

[0123] Example 5

[0124] like Figure 5 As shown, the active learning module 4 includes an uncertainty estimation unit 41, a sampling strategy unit 42, an incremental annotation unit 43, a model update unit 44, and an annotation efficiency evaluation unit 45.

[0125] Uncertainty estimation unit 41 is used to evaluate the uncertainty of model predictions using the Monte Carlo Dropout method. This unit performs forward propagation multiple times (with Dropout enabled) to calculate the mean and variance of the predictions:

[0126]

[0127] Where μ(x) is the average prediction result, σ 2 (x) represents the prediction variance, f t (x) represents the prediction result of the t-th forward propagation, where T is the number of Monte Carlo samplings. The prediction variance σ 2 (x) serves as a measure of model uncertainty; a larger value indicates higher prediction uncertainty.

[0128] Sampling strategy unit 42 is used to calculate the optimal sampling point score based on uncertainty, representativeness, and accessibility. This unit employs a hybrid sampling strategy, comprehensively considering multiple factors:

[0129] Score(x)=α·Uncertainty(x)+β·Pepresentativeness(x)+γ·Accessibility(x),

[0130] Wherein, Score(x) is the overall score of the sampling point, Uncertainty(x) is the prediction uncertainty, Representativenss(x) is the representativeness of the sample, Accessibility(x) is the actual sampling difficulty, and α, β, and γ are balancing factors used to adjust the importance of each factor. The system selects the point with the highest score as the next sampling point.

[0131] The incremental annotation unit 43 is used to acquire new sample annotations and update the training dataset. Once the optimal sampling points are determined, this unit is responsible for obtaining the true labels for these points and adding the newly annotated samples to the training dataset. Depending on the scenario, the annotation method can be expert annotation, crowdsourced annotation, or automatic annotation.

[0132] The model update unit 44 is used to retrain the recognition model based on the updated training dataset. When the training dataset is updated, this unit uses an incremental learning strategy to update the model parameters, avoiding the computational overhead of a complete retraining. The update strategy includes fine-tuning the model parameters, updating the batch normalization statistics, and adjusting the learning rate.

[0133] The annotation efficiency evaluation unit 45 is used to calculate the ratio of the increase in annotation quantity to the improvement in accuracy, and to determine whether to continue sampling annotation. This unit defines the annotation efficiency index:

[0134]

[0135] Here, ΔAccuracy represents the accuracy improvement, and ΔAnnotation represents the increase in annotation quantity. When this ratio is lower than a preset threshold, it indicates that the marginal benefit of continuing annotation is low, and the system will stop the incremental annotation process.

[0136] Preferably, the active learning module also implements a sampling path planning function, which can plan the optimal sampling path after determining multiple sampling points, thereby minimizing the time and cost of field investigation.

[0137] Example 6

[0138] like Figure 6 As shown, the morphology optimization module 5 includes a geological constraint unit 51, a fault topology unit 52, a morphology optimization unit 53, a knowledge base interaction unit 54, and a result output unit 55.

[0139] Geologically constrained unit 51 is used to construct an energy function that integrates physical rules and data-driven prediction. This unit defines the geologically constrained energy function:

[0140] E(S)=E data (S)+λ1·E continuity (S)+λ2·E smoothness (S)+λ3·E direction (S), where S is the morphology of the fracture zone, and E data For data items, E represents the consistency between the predicted results and the observed data. continuity As a continuity constraint, it encourages the rupture zone to remain continuous; E smoothness E is a smoothness constraint term that suppresses excessively convoluted shapes. direction λ1, λ2, and λ3 are directional constraint terms to ensure that the direction of the rupture zone is consistent with the regional tectonic stress field; λ1, λ2, and λ3 are weighting coefficients used to balance the importance of each constraint term.

[0141] Fault topology unit 52 is used to establish the network graph structure of the rupture zone and infer possible fault connections. This unit represents the rupture zone as a graph structure G = (V, E), where nodes V represent rupture zone segments and edges E represent potential connections. The probability of edge existence is calculated through graph reasoning:

[0142] P(e ij |G)=σ(MLP(concat[h i ,h j ,θ ij ])),

[0143] Wherein, P(e ij |G) represents the edge e given the graph structure G. ij The probability of existence, h i and h j For node features, θ ij The geometric relationships between nodes are represented by MLP (Multilayer Perceptron) and σ is the sigmoid activation function. In this way, the system can infer the possible connections between discontinuous fracture zones.

[0144] The morphology optimization unit 53 is used for multi-scale morphology optimization based on hierarchical Markov random fields. This unit performs morphology optimization at multiple scales:

[0145]

[0146] Among them, S * For the optimized fracture zone morphology, C represents the clue set, and ψ c Let S be the potential function. c For configuration. This unit searches for the rupture zone morphology that minimizes the energy function through an iterative optimization process. For example... Figure 8 As shown, the difference between the before and after morphological optimization is obvious, and the optimized rupture zone morphology is more in line with geological laws.

[0147] Knowledge base interaction unit 54 is used to extract typical morphological features and regional tectonic background information from the fault knowledge base. This unit maintains a knowledge base containing information such as different earthquake types, typical morphological features, and regional tectonic background, providing prior knowledge for morphological optimization. Typical morphological features include linear indices, fault displacement ranges, segmentation characteristics, and secondary structures; regional tectonic background includes stress field direction and fault strike.

[0148] Output unit 55 is used to generate the final earthquake surface rupture zone distribution map and attribute report. This unit outputs standard format rupture zone vector data (GeoJSON format), including attribute information such as location, strike, deformation, segmentation characteristics, and uncertainty assessment, and generates visualization maps and analysis reports.

[0149] Preferably, the morphology optimization module also has a historical data comparison function, which can compare the currently identified rupture zone with similar cases in the historical earthquake database to verify the rationality of the identification results and provide reference information.

[0150] Example 7

[0151] The adaptive weighted registration loss function is applied in the heterogeneous data registration module to guide the accurate registration of multi-source remote sensing data.

[0152] The adaptive weight registration loss function is defined as:

[0153] L reg =α·L intensity +β·L geometric +γ·L context ,

[0154] Where α, β, and γ are adaptive weighting coefficients, dynamically adjusted based on data quality; L intensity For intensity similarity loss L geometric For geometric consistency loss; L context This is for context similarity loss.

[0155] Intensity similarity loss L intensity This is used to measure the similarity between intensity values ​​from different data sources at corresponding locations. Considering the differences in intensity characteristics among different remote sensing data, this loss term uses Normalized Mutual Information (NMI) as a metric.

[0156] L intensity =1-NMI(I1,I2),

[0157] Where I1 and I2 are intensity maps from two data sources, and NMI is normalized mutual information, with a value range of [0,1]. The larger the value, the more similar the two images are.

[0158] Geometric consistency loss L geometric It is used to measure the consistency of the spatial structure after registration, and is calculated through the geometric relationship between feature points:

[0159]

[0160] Where, p i and q i For each corresponding feature point pair, T is the transformation matrix, and N is the number of feature point pairs, ||·|| 2 This represents the square of the Euclidean distance.

[0161] Contextual similarity loss L context The Structural Similarity Index (SSIM) is used to measure the similarity of structural features in the registered regions.

[0162] L context =1-SSIM(T(I1),I2),

[0163] Where T(I1) is the first data source after transformation, I2 is the second data source, and SSIM is the structural similarity index with a value range of [0,1]. The larger the value, the more similar the structures of the two images are.

[0164] The adaptive weighting coefficients α, β, and γ are dynamically adjusted based on data quality and characteristics. For regions with high signal-to-noise ratios, α is increased to emphasize intensity matching; for regions rich in features, β is increased to emphasize geometric consistency; and for regions with complex textures, γ is increased to emphasize structural matching. The weighting coefficients satisfy the constraints α + β + γ = 1 and α, β, γ ≥ 0.

[0165] By using this adaptive weighted registration loss function, the system can adaptively adjust the registration strategy according to the data characteristics, thereby improving the registration accuracy of different types of remote sensing data.

[0166] Example 8

[0167] The deformation-sensitive loss function is applied in the rupture zone segmentation module to improve the recognition accuracy of large deformation gradient regions.

[0168] The deformation-sensitive loss function is defined as:

[0169]

[0170] Where y is the true label and p is the predicted probability. Let λ be the deformation gradient and λ be the weighting factor.

[0171] This is an improved version based on the cross-entropy loss function, by introducing deformation gradients. As a weight adjustment factor, it assigns higher learning weights to regions with large deformation gradients. Deformation gradient The calculation is as follows:

[0172]

[0173] in, and These represent the partial derivatives of the deformation field in the x and y directions, respectively, and can be calculated using the finite difference method. The larger the value, the greater the deformation gradient in that region, and the greater the corresponding weight adjustment.

[0174] The weighting factor λ controls the influence of the deformation gradient on the loss function, and its value range is generally [0, 5]. It can be adjusted appropriately according to the specific application scenario. When λ = 0, the deformation-sensitive loss function degenerates into the standard cross-entropy loss function.

[0175] Preferably, the weighting factor λ can be dynamically adjusted for different types of earthquakes. For earthquakes dominated by strike-slip, λ is set to a larger value to enhance the sensitivity to horizontal displacement gradients; for earthquakes dominated by reverse or normal faults, λ is set to a moderate value to balance the influence of horizontal and vertical displacements.

[0176] By using this deformation-sensitive loss function, the system can focus on areas with large deformation gradients, improve the identification accuracy of these areas, and thus more accurately capture the boundaries and internal structure of earthquake surface rupture zones.

[0177] Example 9

[0178] This embodiment provides a detailed description of multi-source remote sensing data and earthquake surface rupture zone attribute information.

[0179] Multi-source remote sensing data includes: high-resolution optical satellite remote sensing imagery, SAR imagery, UAV orthophotos, and LiDAR point cloud data. These data sources each have their own characteristics, complement each other, and together provide comprehensive information about the Earth's surface.

[0180] High-resolution optical satellite remote sensing imagery mainly includes optical images acquired by satellites such as the Gaofen series (GF-1 / 2 / 6 / 7, etc.), WorldView series, Landsat series, and Sentinel-2. This type of data is characterized by high resolution and rich spectral bands, providing detailed surface texture and spectral information, which helps in the intuitive identification of the location and morphology of surface fracture zones. Preferably, panchromatic images with a spatial resolution better than 1 meter and multispectral images with a spatial resolution better than 4 meters are selected to ensure identification accuracy.

[0181] SAR imagery primarily includes radar images acquired by satellites such as Gaofen-3, Sentinel-1A / B, ALOS-1 / 2, and TerraSAR-X. This type of data offers all-weather, all-day observation capabilities, unaffected by cloud cover or lighting conditions, and can sensitively capture surface deformation information, making it a crucial data source for identifying earthquake surface rupture zones. Preferably, SAR imagery with a ground resolution better than 10 meters is selected to ensure accurate extraction of deformation information.

[0182] UAV orthophotos are near-vertical images acquired by high-resolution cameras mounted on UAVs. This type of data has extremely high resolution (typically better than 0.1 meters) and can provide detailed information on the surface structure, making it an important supplement for the accurate identification of local fault zones.

[0183] LiDAR point cloud data is 3D point cloud data acquired through airborne or ground-based laser scanners. This type of data can provide high-precision 3D information about the Earth's surface structure, especially in vegetated areas. By filtering out vegetation points, true surface information can be obtained, compensating for the deficiencies of optical and SAR data.

[0184] Earthquake surface rupture zone attribute information includes: rupture zone location, strike, deformation, segmentation characteristics, and uncertainty assessment. These attributes collectively describe the spatial distribution and deformation characteristics of the rupture zone and are an important basis for earthquake hazard assessment.

[0185] The location of the rupture zone is represented as vector data, including latitude and longitude coordinates and elevation information, which can be directly displayed and analyzed in the GIS system. The strike refers to the direction of the rupture zone's extension, expressed as an angle, usually with due north as a reference. Deformation includes horizontal and vertical displacements, representing the relative movement of the Earth's surface in the horizontal and vertical directions, respectively. Segmentation features describe the segmentation of the rupture zone, including the number, length, and intervals of segments. Uncertainty assessment provides reliability indicators for the identification results, including location uncertainty, strike uncertainty, and deformation uncertainty.

[0186] Preferably, the system can also output a three-dimensional model of the rupture zone and temporal variation data, providing more information for in-depth research on earthquake mechanisms and monitoring of geological activities.

[0187] Example 10

[0188] like Figure 9 As shown in the illustration, this embodiment provides a detailed description of a method for intelligent identification of earthquake surface rupture zones. The method includes the following steps:

[0189] Step S101: Receive multi-source remote sensing data, use a multi-scale feature point matching algorithm and an adaptive weight registration loss function to register heterogeneous data, and generate registered multi-source data and a registration accuracy evaluation report.

[0190] This step first receives multi-source remote sensing data, including high-resolution optical satellite imagery, SAR imagery, UAV orthophotos, and LiDAR point cloud data. These data are then preprocessed, including radiometric correction, geometric correction, and noise filtering. Next, robust feature points from each data source are extracted using a multi-scale feature point matching algorithm, and heterogeneous data registration is performed based on an adaptive weighted registration loss function. After iterative refinement, registered multi-source data is generated, along with a registration accuracy assessment report and uncertainty region markers.

[0191] Step S102: Receive the registered multi-source data, decompose the features of each data source into edge features, deformation features and context features through a multimodal feature decoupling network, dynamically fuse complementary features using an attention-guided feature fusion machine, and output a fused feature map.

[0192] In this step, the registered multi-source data are first feature-encoded separately to obtain the original feature representations. Then, a multimodal feature decoupling network is used to decompose these features into edge features, deformation features, and context features, achieving feature decoupling at different semantic levels. Next, an attention computation unit generates dynamic feature fusion weights, and the decoupled features are weighted and fused based on these weights. Finally, a feature consistency enhancement unit forces the different source features to align in the semantic space, outputting a high-quality fused feature map.

[0193] Step S103: Receive the fused feature map, use multi-scale spatial pyramid pooling and deformation-sensitive loss function to identify the fracture zone, and generate preliminary fracture zone segmentation results and prediction uncertainty.

[0194] In this step, the basic representations of the fused feature map are first extracted, and then multi-scale features of the rupture zone are captured through multi-scale spatial pyramid pooling. Next, a deformation-sensitive loss function is used to guide model training, assigning higher learning weights to regions with large deformation gradients. Simultaneously, a region affinity enhancement module is used to establish long-range spatial dependencies, enhancing the spatial continuity of the rupture zone. Finally, preliminary rupture zone segmentation results and prediction uncertainties are output.

[0195] Step S104: Receive the preliminary rupture zone segmentation results and prediction uncertainty, determine the optimal sampling points based on Bayesian uncertainty estimation and a hybrid sampling strategy, and update the training dataset.

[0196] In this step, the uncertainty of the model's predictions is first assessed using the Monte Carlo Dropout method. Then, the optimal sampling point score is calculated based on uncertainty, representativeness, and accessibility. Next, the ground truth labels for these sampling points are obtained, the training dataset is updated, and the model is retrained based on the updated dataset. Finally, the ratio of the increase in labeled data to the improvement in accuracy is calculated to determine whether to continue sampling and labeling.

[0197] Step S105: Receive the updated training dataset and preliminary rupture zone segmentation results, optimize the rupture zone morphology based on the geological constraint energy function and fault system topology reasoning, and output the final earthquake surface rupture zone distribution and attribute information.

[0198] In this step, a geologically constrained energy function integrating physical rules and data-driven prediction is first constructed. Then, a rupture zone network structure is established, and possible fault connections are inferred. Next, multi-scale morphological optimization is performed based on a layered Markov random field, while typical morphological features and regional tectonic background information are extracted from the fault knowledge base as optimization guidance. Finally, the final seismic surface rupture zone distribution map and attribute report are output, including information such as rupture zone location, strike, deformation, segmentation characteristics, and uncertainty assessment.

[0199] Preferably, the method further includes a result verification step, which evaluates the accuracy and reliability of the identification results by comparing them with field survey data or known rupture zone data, and further optimizes system parameters and strategies based on the verification results.

[0200] Example 11

[0201] This embodiment provides a practical application case, demonstrating the application effect of the system of the present invention in the response to the 6.2 magnitude earthquake in Jishishan, Gansu in 2023.

[0202] This earthquake occurred in a complex mountainous environment with significant topographic relief and extensive vegetation cover, posing a considerable challenge to traditional methods for identifying surface rupture zones. The intelligent earthquake surface rupture zone identification system of this invention follows the following processing flow:

[0203] First, the system receives multi-source remote sensing data, including Gaofen-2 optical images before and after the earthquake, Sentinel-1A / B SAR images, UAV orthophotos, and local LiDAR point cloud data. Among them, optical and SAR images cover the entire disaster area, while UAV and LiDAR data mainly cover key areas of concern.

[0204] The system received multi-source remote sensing data including: Gaofen-2 optical images before and after the earthquake (spatial resolution better than 1 meter); Sentinel-1A / B SAR images (ground resolution of about 10 meters); UAV orthophotos (resolution better than 0.1 meters); and LiDAR point cloud data of key local areas (point density of about 10 points / square meter). Among them, optical and SAR images cover the entire disaster area (about 200 square kilometers), while UAV and LiDAR data mainly cover key areas of concern (about 20 square kilometers), forming a multi-scale, multi-source data system.

[0205] Next, the system performs heterogeneous data registration, accurately aligning different data sources to a unified coordinate system. During this process, the adaptive weighted registration loss function significantly improves the registration accuracy for forest-covered and steep mountainous areas, increasing the accuracy by approximately 32% compared to traditional methods.

[0206] Then, the system extracts and fuses features from multi-source data. The multimodal feature decoupling network successfully decomposes the features from each data source into complementary semantic features, and the attention-guided feature fusion processor effectively integrates these features, enhancing the system's ability to identify fracture zones in complex environments.

[0207] During the rupture zone segmentation stage, multi-scale spatial pyramid pooling and deformation-sensitive loss functions enable the system to accurately capture rupture zone features of different scales, with a particularly significant improvement in the identification of regions with large deformation gradients. The system successfully identified multiple strike-slip faults, including a main fault approximately 12 kilometers long with a maximum horizontal displacement of about 1.8 meters.

[0208] Through the active learning module, the system selected only 25 optimal sampling points for field validation, saving 75% of the fieldwork compared to the more than 100 sampling points required by traditional methods. The labeled data from these sampling points was used to update the model, further improving recognition accuracy.

[0209] Finally, the morphology optimization module, based on geological constraints and fault topology reasoning, optimized the morphology of the fracture zone, successfully reconstructed the connection relationships of multiple discontinuous faults, and generated a complete fracture zone distribution map and attribute report. The system's final output fracture zone location accuracy reached sub-meter level, and the deformation measurement accuracy was better than 10 centimeters.

[0210] Compared with traditional manual interpretation methods, this system demonstrated significant advantages in the case of the Jishishan earthquake in Gansu:

[0211] Processing time has been reduced from 3 hours using traditional methods to less than 15 minutes; identification accuracy has increased from 60% to 95%; the workload for field verification has been reduced from 100 sample points to 25; location accuracy has improved from 5-10 meters to sub-meter level (approximately 0.8 meters); deformation measurement accuracy is better than 10 centimeters; significantly improving the efficiency of earthquake emergency response. The high-precision rupture zone distribution map provided by the system offers crucial support for subsequent disaster assessment and disaster prevention and mitigation efforts.

[0212] The intelligent identification system and method for earthquake surface rupture zones of this invention achieve rapid and accurate identification of earthquake surface rupture zones by integrating multi-source remote sensing data processing, deep learning, and geological knowledge guidance, and has good industrial applicability. The system can be deployed in earthquake monitoring centers or emergency management departments as an important technical support for earthquake emergency response.

[0213] The system adopts a modular design with standardized interfaces between modules, facilitating system integration and maintenance. It has moderate hardware requirements, running on ordinary high-performance computing platforms without the need for special hardware support. The system boasts a high degree of automation; users only need to provide multi-source remote sensing data, and the system can automatically complete the entire process from data registration to final result output.

[0214] Furthermore, the system is easily expandable and adaptable to different regions and types of earthquake events. It supports incremental updates and online learning, and its performance continuously improves as data accumulates and usage frequency increases. The system outputs results in a standard format, is compatible with mainstream GIS software, and facilitates integration with other systems and data sharing.

[0215] In summary, the system of the present invention is technologically advanced, feasible to implement, and highly applicable, effectively meeting the actual needs of earthquake disaster emergency monitoring and possessing significant industrial practical value.

[0216] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An intelligent identification system for earthquake surface rupture zones, characterized in that, include: The heterogeneous data registration module is used to receive multi-source remote sensing data, perform heterogeneous data registration using a multi-scale feature point matching algorithm and an adaptive weight registration loss function, and generate registered multi-source data and a registration accuracy evaluation report. The feature extraction and fusion module is connected to the heterogeneous data registration module. It is used to receive the registered multi-source data, decompose the features of each data source into edge features, deformation features and context features through a multimodal feature decoupling network, and dynamically fuse complementary features using an attention-guided feature fusion machine to output a fused feature map. The rupture zone segmentation module, connected to the feature extraction and fusion module, is used to receive the fused feature map, identify the rupture zone using multi-scale spatial pyramid pooling and deformation-sensitive loss function, and generate preliminary rupture zone segmentation results and prediction uncertainties. An active learning module, connected to the rupture zone segmentation module, is used to receive the preliminary rupture zone segmentation results and prediction uncertainties, determine the optimal sampling points based on Bayesian uncertainty estimation and a hybrid sampling strategy, and update the training dataset. The morphology optimization module, connected to the active learning module and the rupture zone segmentation module, is used to receive the updated training dataset and the preliminary rupture zone segmentation results, optimize the rupture zone morphology based on the geological constraint energy function and fault system topology reasoning, and output the final earthquake surface rupture zone distribution and attribute information. The geological constraint energy function is: ,in, It is a rupture zone morphology. For data items, this represents the consistency between the prediction results and the observed data; As a continuity constraint, it encourages the rupture zone to remain continuous; As a smoothness constraint, it suppresses excessively convoluted shapes; This is a directional constraint term to ensure that the direction of the rupture zone is consistent with the regional tectonic stress field; These are weighting coefficients used to balance the importance of each constraint term.

2. The intelligent identification system for earthquake surface rupture zones according to claim 1, characterized in that, The heterogeneous data registration module includes: The data preprocessing unit is used to perform radiometric correction, geometric correction, and noise filtering on the received multi-source remote sensing data. The feature extraction unit is used to extract robust feature points from different types of remote sensing images; The registration loss calculation unit is used to calculate the registration error based on the extracted feature points using an adaptive weighted registration loss function. An iterative refinement unit is used to achieve accurate registration of multi-source data through a cascaded architecture from coarse registration to fine registration based on the registration error. The accuracy assessment unit is used to quantitatively evaluate the registration results and generate accuracy reports and uncertainty area markers.

3. The intelligent identification system for earthquake surface rupture zones according to claim 1, characterized in that, The feature extraction and fusion module includes: The multimodal encoder unit is used to perform specialized feature encoding on various types of registered data; The feature decoupling unit is used to decompose the features of each data source into edge features, deformation features, and context features; The attention calculation unit is used to calculate the feature weight coefficients and generate dynamic feature fusion weights. The feature fusion unit is used to perform weighted fusion of multi-source features based on the dynamic feature fusion weights; The feature consistency enhancement unit is used to force different source features to align in the semantic space through adversarial training.

4. The intelligent identification system for earthquake surface rupture zones according to claim 1, characterized in that, The fracture zone segmentation module includes: The basic feature extraction unit is used to extract the basic representation of the fused feature map; Multi-scale feature enhancement unit is used to capture multi-scale features of the fracture zone through spatial pyramid pooling; Deformation-sensitive decoding unit is used to enhance the learning weights in areas with large deformation gradients in the rupture zone; Regional affinity enhancement units are used to establish long-range spatial dependencies and enhance the spatial continuity of the fracture zone; The segmentation result generation unit is used to output the segmentation results of the fracture zone and its probability distribution.

5. The intelligent identification system for earthquake surface rupture zones according to claim 1, characterized in that, The active learning module includes: Uncertainty estimation unit, used to evaluate the uncertainty of model predictions using the Monte Carlo Dropout method; The sampling strategy unit is used to calculate the optimal sampling point score based on uncertainty, representativeness, and accessibility. Incremental annotation units are used to acquire new sample annotations and update the training dataset; The model update unit is used to retrain the recognition model based on the updated training dataset. The annotation efficiency evaluation unit is used to calculate the ratio of the increase in the number of annotations to the improvement in accuracy, and to determine whether to continue sampling and annotation.

6. The intelligent identification system for earthquake surface rupture zones according to claim 1, characterized in that, The morphology optimization module includes: Geologically constrained units are used to construct energy functions that integrate physical rules and data-driven predictions. Fault topological units are used to establish the network structure of rupture zones and to infer possible fault connections. A morphology optimization unit is used for multi-scale morphology optimization based on hierarchical Markov random fields. The knowledge base interaction unit is used to extract typical morphological features and regional tectonic background information from the fault knowledge base. The results output unit is used to generate the final earthquake surface rupture zone distribution map and attribute report.

7. The intelligent identification system for earthquake surface rupture zones according to claim 1, characterized in that, The adaptive weight registration loss function is: ， in, The weighting coefficients are adaptive and dynamically adjusted based on data quality. For intensity similarity loss; This is the geometric consistency loss; This is for context similarity loss.

8. The intelligent identification system for earthquake surface rupture zones according to claim 1, characterized in that, The deformation-sensitive loss function is: , in, For real labels, To predict probabilities, For deformation gradient, This is the weighting factor.

9. The intelligent identification system for earthquake surface rupture zones according to claim 1, characterized in that, The multi-source remote sensing data includes: high-resolution optical satellite remote sensing images, SAR images, UAV orthophotos, and LiDAR point cloud data; the earthquake surface rupture zone attribute information includes: rupture zone location, strike, deformation, segmentation characteristics, and uncertainty assessment.

10. A method for intelligent identification of earthquake surface rupture zones, employing the system described in any one of claims 1-9, characterized in that, include: Receive multi-source remote sensing data, use a multi-scale feature point matching algorithm and an adaptive weight registration loss function to register heterogeneous data, and generate registered multi-source data and a registration accuracy evaluation report. The registered multi-source data is received, and the features of each data source are decomposed into edge features, deformation features and context features through a multimodal feature decoupling network. Complementary features are dynamically fused using an attention-guided feature fusion machine, and a fused feature map is output. The fused feature map is received, and multi-scale spatial pyramid pooling and deformation-sensitive loss function are used to identify the fracture zone, generating preliminary fracture zone segmentation results and prediction uncertainty. Receive the preliminary rupture zone segmentation results and prediction uncertainties, determine the optimal sampling points based on Bayesian uncertainty estimation and a hybrid sampling strategy, and update the training dataset; The system receives the updated training dataset and the preliminary rupture zone segmentation results, optimizes the rupture zone morphology based on the geological constraint energy function and fault system topology reasoning, and outputs the final earthquake surface rupture zone distribution and attribute information.

Images