Earthquake disaster scene identification method and system based on deep learning

Abstract

The invention belongs to the technical field of earthquake disaster scene recognition, and discloses an earthquake disaster scene recognition method based on deep learning. The method comprises the following specific steps: S1, data acquisition and preprocessing; S1.1, multi-source heterogeneous data acquisition and establishment of a comprehensive database containing seismic waveform data, surface deformation data, building structure data, geographic information data and historical disaster record data; through fusion of a 3D convolutional network, a graph attention mechanism, a space-time LSTM and an adaptive cross-modal attention fusion technology, combined modeling of a seismic waveform space-time evolution law, an earth surface deformation space distribution characteristic, a building group topology vulnerability and disaster chain time sequence association is realized, the characterization capability of a complex nonlinear disaster mode is effectively improved, and the method has the advantages of high adaptability and high reliability. And disaster assessment response time is shortened to a sub-second level through mixed precision quantification and edge computing deployment, and high recognition accuracy is still kept in a scene with strong noise and data missing in combination with a multi-task classifier and a physical constraint verification mechanism.

Description

Technical Field

[0001] The present invention belongs to the technical field of earthquake disaster scenario recognition, and specifically relates to an earthquake disaster scenario recognition method and system based on deep learning. Background Art

[0002] Earthquake disaster scenario identification refers to the comprehensive and scientific identification and assessment of potential earthquake disaster scenarios through systematic analysis of the possible scenarios, processes and impacts of earthquake disasters, providing a basis for the formulation of disaster prevention and mitigation strategies, emergency plans and resource allocation.

[0003] Traditional earthquake disaster assessment methods mainly rely on physical model simulation and statistical empirical analysis, and have significant technical bottlenecks. Among them, methods based on seismic wave propagation equations and structural mechanics models are limited by the accuracy of parameter assumptions and are difficult to adapt to the nonlinear dynamic characteristics under the coupling of complex geological environments and building complexes; statistical learning methods based on threshold discrimination face the disaster of feature dimension when processing multi-source heterogeneous data, and cannot effectively integrate the correlation laws between waveform time series characteristics, spatial deformation fields and building topology, resulting in coarse granularity in disaster factor identification; at the same time, existing data fusion technologies mostly use simple weighted superposition or rule splicing, lacking cross-modal feature alignment and deep semantic association. Modeling results in insufficient analysis of the physical coupling mechanism between InSAR deformation data, BIM structural parameters and real-time waveform data. In addition, traditional assessment systems rely on offline numerical simulation and manual interpretation, and have low computational efficiency when processing high-resolution remote sensing images and massive sensor data. Response delays of more than one minute cannot meet the real-time decision-making needs of emergency command. In addition, the disaster chain prediction model based on fixed rules has weak generalization capabilities for complex scenarios such as abnormal source mechanisms and secondary disaster interactions, and is prone to misjudgment when data is missing or there is noise interference. These limitations seriously restrict the accuracy and timeliness of earthquake disaster scenario identification, and therefore need to be improved. Summary of the Invention

[0004] The purpose of the present invention is to provide a method and system for earthquake disaster scenario identification based on deep learning to solve the problems raised in the above background technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for identifying earthquake disaster scenarios based on deep learning, the specific steps of which are as follows:

[0006] S1. Data collection and preprocessing

[0007] S1.1 Multi-source heterogeneous data collection

[0008] Establish a comprehensive database containing earthquake waveform data, surface deformation data, building structure data, geographic information data, and historical disaster record data. Specifically, it includes: P-wave and S-wave raw waveform data recorded by the seismic network, with a sampling frequency of no less than 100Hz; millimeter-level surface deformation data acquired by InSAR satellite remote sensing; high-resolution disaster site image data acquired by drone aerial photography; structural parameter data of urban building information models (BIM); historical earthquake intensity distribution maps and disaster damage assessment reports;

[0009] S1.2 Data Standardization

[0010] In the data standardization processing stage, it is necessary to build a unified data processing framework with spatiotemporal consistency. First, all geospatial data are accurately aligned using the WGS84 global geodetic system. The geographic reference deviations caused by different sensors and observation platforms are eliminated through coordinate conversion algorithms to ensure the accurate matching of multiple data such as seismic station coordinates, satellite remote sensing images, and building model positions in three-dimensional space. On this basis, a global time synchronization mechanism based on Coordinated Universal Time (UTC) is established to implement millisecond-level time alignment for heterogeneous time series data such as seismic waveform recorders, satellite transit times, and drone collection times, and to build a data association system under a continuous spatiotemporal coordinate system. For high-resolution aerial images, disaster relief images, and other disaster relief images, a global time synchronization mechanism is established based on Coordinated Universal Time (UTC). For unstructured data such as text reports from disaster sites, an image vectorization algorithm is used to convert raster data into vector layers with clear topological relationships. Natural language processing techniques are also used to extract key semantic features from the text and encode them into structured vectors. Finally, multi-dimensional data normalization is used to eliminate physical dimension differences. Numerical data such as earthquake waveform amplitude, surface deformation values, and building structural parameters are scaled using Min-Max normalization and Z-score normalization methods, respectively. Categorical variables are then encoded using one-hot encoding and embedded representation. This ultimately forms a standardized dataset with unified spatiotemporal benchmarks, standardized data structures, and balanced feature distribution, laying a high-quality data foundation for the subsequent construction of deep learning models.

[0011] S2. Deep learning model construction

[0012] S2.1. Hybrid Architecture Neural Network

[0013] In the multimodal feature extraction network design during the deep learning model construction phase, it is necessary to build a collaborative hybrid neural network architecture to fully exploit the multi-source features of earthquake disaster data;

[0014] A three-dimensional convolutional neural network is used to process spatiotemporally continuous seismic waveform data. A convolution kernel with adaptive receptive field adjustment is designed, with a size of 5×5×3. A hierarchical spatiotemporal feature extraction module is used to capture key seismic phase features such as P-wave first arrival and S-wave amplitude changes. A dilated convolution is also integrated with a dilation rate of 2 to expand the temporal context perception range of the waveform sequence.

[0015] For high-resolution remote sensing image data, an improved U-Net architecture is constructed. ResNet-50 is used as the backbone network in the encoder part and a channel attention mechanism is introduced. Multi-scale feature fusion is implemented in the decoder stage. Low-level texture features are aggregated with high-level semantic features across layers through skip connections. A directional feature enhancement module is designed specifically for disaster markers such as collapsed buildings and surface cracks.

[0016] When processing building complex structural data, a graph convolutional network is used to establish a topological relationship analysis model. Individual buildings are abstracted into graph nodes, with node attributes including parameters such as structure type, floor height, and construction year. Spatial adjacency relationships are constructed through Voronoi diagrams as the basis for edge connections. A graph attention layer is used to dynamically learn the mechanical interaction weights between buildings, and a graph pooling operation based on physical constraints is designed to preserve key structural vulnerability characteristics.

[0017] A bidirectional long-short-term memory network is constructed to capture the temporal dynamics of disaster evolution. In the temporal dimension, 128 hidden units are set and a hierarchical propagation mechanism is adopted. The forward layer captures the propagation characteristics of the earthquake source rupture process, and the backward layer reversely analyzes the disaster chain reaction path. The contribution of features from different time phases is weighted and fused through a temporal attention mechanism. Ultimately, a multimodal feature expression system is formed that takes into account spatial topology, temporal evolution, and physical mechanisms. This provides a basic feature representation with strong discriminability for subsequent cross-modal feature fusion.

[0018] S2.2 Feature Fusion

[0019] After the initial extraction of multimodal features, there is a significant modality gap between the original heterogeneous features. To effectively integrate these cross-domain features, it is necessary to design an adaptive attention mechanism to integrate the deep features of multi-source heterogeneous data.

[0020] First, a cross-modal feature alignment layer is used to eliminate heterogeneity in data such as earthquake waveforms, remote sensing images, and building structures. The dynamic time warping (DTW) algorithm is used to align temporal features of different sampling rates, and a feature projection matrix is ​​constructed to achieve semantic matching in high-dimensional space. On this basis, a multi-head self-attention mechanism is introduced to establish an inter-modal correlation matrix. Through parallel computing, multi-dimensional dependency patterns such as the interaction between spatial deformation fields and building complex topology, and the causal relationship between earthquake source rupture processes and surface responses are captured. A differentiable feature selection gating unit is further developed. Based on the gated recurrent network (GRU), the contribution of each modal feature is dynamically evaluated, redundant features are softly masked, and a channel attention weighting strategy is used to focus on key disaster identification features. Finally, a hierarchical feature aggregation module is used to map multi-scale features to a unified semantic space. Depthwise separable convolution is used to compress feature dimensions, and residual connections are combined to retain the original information to generate a 128-dimensional fused feature vector. This vector simultaneously encodes the physical mechanism, spatial distribution, and temporal evolution characteristics of the disaster, providing a highly discriminative input representation for subsequent classification tasks.

[0021] S2.3 Scenario Recognition Classifier

[0022] The 128-dimensional high-density feature vector generated through adaptive feature fusion not only retains the physical properties of the original data but also contains the underlying laws of disaster evolution. To transform these abstract features into actionable disaster scenario judgments, a hybrid classification decision-making system is constructed, and joint reasoning of disaster factors is achieved through a multi-task learning framework.

[0023] First, a deep residual network (ResNet-50) is adopted as the backbone architecture. Skip connections are used to alleviate the gradient vanishing problem. A global average pooling layer is used to map the fused features to the disaster level space, achieving initial classification of intensity from VI to XI. For abnormal disaster modes, such as earthquake swarms and induced landslides, a support vector data description (SVDD) is introduced to construct a hypersphere decision boundary, and kernel function mapping is used to detect dangerous scenarios that deviate from the normal distribution. A conditional random field (CRF) post-processing module is designed to encode spatial continuity constraints such as the degree of building damage and the direction of surface rupture zones into energy functions. The graph cut algorithm is used to optimize the regional consistency of the classification results. Finally, a multi-task learning framework is constructed. By sharing underlying features and collaboratively designing independent task heads, the triple recognition results of disaster type, intensity level, and impact range are simultaneously output. The confidence of each prediction result is evaluated through the uncertainty quantification module, forming a decision output that is both physically interpretable and engineering practical.

[0024] S3, model training and optimization

[0025] S3.1. Training Strategy Design

[0026] During the model training phase, a progressive optimization strategy is adopted to improve learning efficiency and generalization ability: first, the network parameters are unfrozen in stages, with the feature extraction layer prioritized for training to capture fundamental patterns, followed by joint optimization of the classifier parameters. A curriculum learning mechanism is introduced to gradually transition from single focal mechanism samples to multi-type composite disaster scenarios. Adversarial training is simultaneously implemented to enhance the model's robustness to noise interference by generating adversarial perturbation samples. In response to the long-tail distribution characteristics of earthquake disaster data, a dynamic weighted cross-entropy loss function is designed to adaptively adjust the weight coefficient based on the frequency of class samples to balance the class imbalance problem in the mainshock and aftershock identification task.

[0027] S3.2 Hyperparameter Optimization

[0028] After completing the basic architecture design of the training strategy, improving model performance will rely on accurate exploration of the hyperparameter space. To this end, an intelligent parameter adjustment system is built to achieve the global optimal configuration of model parameters through multi-dimensional optimization strategies.

[0029] A Bayesian optimization algorithm was used to search for learning rate combinations (with a base learning rate range of 1e-5 to 1e-3), combined with Neural Architecture Search (NAS) to determine the Pareto optimal solution for network depth and convolution kernel size. A genetic algorithm was used to optimize the batch size (32-256) and the L2 regularization coefficient (1e-4 to 1e-2), and a multi-generation population evolution was used to select a parameter set that resists overfitting. A hyperparameter response surface analysis model was constructed to quantify the impact of parameter interactions on recognition accuracy and inference speed, ultimately determining the global optimal configuration that balances computational efficiency (FLOPs ≤ 15G) and classification performance (F1-score ≥ 0.92).

[0030] S4: Scenario Identification and Decision Support Stage

[0031] S4.1, Real-time Inference Engine

[0032] To meet the real-time demands of earthquake emergency response, an edge computing-based inference system was designed. This system converts floating-point models into 8-bit integers using mixed-precision quantization technology, enabling FPGA hardware acceleration deployment. A streaming data processing pipeline was constructed to support parallel access and online updates of multi-source sensor data. A lightweight inference engine was integrated to ensure sub-second response latency (<500ms). Simultaneously, an uncertainty quantization module was developed, utilizing the Monte Carlo Dropout method to output prediction confidence, providing a reliability assessment basis for decision-making.

[0033] S4.2 Visualization Interaction Platform

[0034] Build a 3D disaster situation awareness system, using WebGL technology to achieve real-time browser rendering, support dynamic overlay display of earthquake focal mechanism spheres, surface deformation fields, and building damage heat maps, and trace back the key evolution process from 10 seconds before to 30 minutes after the disaster.

[0035] S4.3. Decision support

[0036] An intelligent auxiliary decision-making chain is established to automatically generate emergency response plans based on scenario recognition results, and rescue route planning is optimized by combining real-time road conditions and aftershock probability predictions. At the same time, the probability and spatiotemporal distribution of derivative risks such as landslides and barrier lakes are predicted through the LSTM network to provide early warning of secondary disasters. A building reinforcement priority list is output, and a weighted ranking is performed based on the comprehensive structural damage index, population density, and economic value to provide a quantitative decision-making basis for post-disaster reconstruction.

[0037] Preferably, in order to address the common sample imbalance problem in earthquake disaster scenario identification as described in S1.1, it is necessary to construct a multi-level composite data enhancement system. First, based on the conditional generative adversarial network (CGAN) framework, by inputting historical earthquake parameter condition constraints and combining the latent space feature decoupling technology, rare disaster scenario samples with physical rationality are synthesized, including rare cases such as high-intensity earthquakes with low probability of occurrence and secondary disasters in special geological structure areas.

[0038] Preferably, after the disaster scenario samples are generated, spatial enhancement processing is performed on the multi-source observation data on this basis, using geometric deformation operations such as random rotation, translation, and affine transformation, and introducing texture perturbation strategies such as Gaussian noise, random occlusion, and color perturbation to enhance the model's robustness to sensor errors and on-site environmental interference.

[0039] Preferably, the angle range of the random rotation is ±30°, the maximum offset of the translation is 20%, the signal-to-noise ratio of the Gaussian noise is controlled at 30-50dB, the occlusion ratio of the random occlusion is 15%-40%, and the HSV space of the color disturbance is adjusted by ±10%.

[0040] Preferably, in order to address the common sample imbalance problem in earthquake disaster scenario identification as described in S1.1, an improved Mixup algorithm is further applied to implement cross-category data fusion. By regulating the linear interpolation coefficient λ (Beta distribution parameter α=0.4), adjacent intensity level samples are smoothly transitioned in the feature space to construct transitional training samples with continuous disaster intensity changes; at the same time, a three-dimensional spatiotemporal data cube is constructed, continuous seismic phase waveform data (time step 0.1 second) is stacked along the time dimension, and the kilometer-level gridded surface deformation field (resolution 100m×100m) is integrated in the spatial dimension. The spatiotemporal joint features are extracted through a three-dimensional convolution kernel (5×5×3 size), and the data resampling technology is used to expand the time series length to 2-3 times of the original data. Finally, a comprehensive data enhancement scheme covering geometric transformation, noise injection, feature mixing, and spatiotemporal expansion is formed, which increases the scale of training samples by 3-5 times, significantly improving the model's representation ability for long-tail distribution data and the generalization performance of complex disaster scenarios.

[0041] Preferably, after the system optimization of the hyperparameter space as described in S3.2, in order to ensure the generalization ability of the model in complex earthquake scenarios, a multidimensional evaluation system including physical constraint verification is established. The system not only verifies the theoretical performance of the model, but also focuses on examining its applicability in actual earthquake engineering; a multidimensional comprehensive evaluation system is constructed, and the data set is partitioned according to the epicenter location and occurrence time through a spatiotemporal cross-validation strategy to avoid overfitting and test the generalization ability; a quantitative indicator system including disaster identification accuracy, false alarm rate, spatial positioning error, and response delay is formulated, and a physical constraint verification mechanism is introduced at the same time, and the seismic wave propagation equation and the law of conservation of energy are used to check the rationality of the prediction results, and the adversarial sample generation technology is combined to test the robustness of the model under noise interference and data anomalies, and finally a credibility evaluation report that takes into account both data-driven performance and compliance with physical laws is formed.

[0042] The deep learning-based earthquake disaster scenario recognition system includes the following modules:

[0043] Data acquisition and preprocessing module: used to integrate multi-source heterogeneous data such as earthquake waveforms, surface deformation, building structures, geographic information, and historical disaster records, and preprocess the data;

[0044] Deep learning model building module: used to design a hybrid neural network architecture, integrate an adaptive attention mechanism to achieve cross-modal feature fusion, and build a multi-task classifier to simultaneously output disaster type, intensity level, and impact range;

[0045] Model training module: used to design model training strategies and use Bayesian optimization and genetic algorithms to search for hyperparameters. It also introduces a physical constraint verification mechanism to establish a multi-dimensional evaluation system to ensure that model prediction results conform to earthquake propagation laws and actual engineering needs.

[0046] Visualization Interaction Module: This module is used for real-time overlay display of earthquake focal mechanism spheres, surface deformation fields, and building damage heat maps. It also uses a timeline to trace back the key evolution processes before and after the disaster, providing an intuitive spatiotemporal analysis interface.

[0047] Decision support module: used to generate emergency response plans.

[0048] Preferably, the model training module is based on a multimodal heterogeneous data fusion framework, and integrates 3D convolutional networks, U-Net, graph convolutional networks and bidirectional LSTM through a hybrid architecture neural network to construct a multi-level feature extraction system. It adopts an adaptive attention fusion mechanism to realize cross-modal alignment and correlation matrix modeling of earthquake waveform spatiotemporal characteristics, surface deformation spatial characteristics, building complex topological characteristics and disaster evolution time series characteristics. It combines deep residual networks, support vector data description and conditional random fields to construct a hybrid classification decision system, and uses a multi-task learning framework to simultaneously optimize disaster type identification, intensity classification and impact range prediction tasks. It realizes the dynamic generation of 128-dimensional high-discriminative feature vectors through differentiable feature selection gating units, and finally forms an end-to-end earthquake disaster scenario recognition model with physical law embedding and spatiotemporal continuity constraints.

[0049] Preferably, the auxiliary decision-making module can optimize rescue path planning and resource scheduling strategies, output a building reinforcement priority list based on the comprehensive structural damage index, population density and economic value, and provide quantitative decision-making support for post-disaster reconstruction.

[0050] The beneficial effects of the present invention are as follows:

[0051] By integrating 3D convolutional networks, graph attention mechanisms, spatiotemporal LSTMs, and adaptive cross-modal attention fusion technology, we have achieved joint modeling of the spatiotemporal evolution of earthquake waveforms, the spatial distribution characteristics of surface deformation, the topological vulnerability of building complexes, and the temporal correlation of disaster chains, effectively improving the representation capabilities of complex nonlinear disaster patterns. Through mixed-precision quantization and edge computing deployment, we have shortened the disaster assessment response time to sub-seconds. Combined with multi-task classifiers and physical constraint verification mechanisms, we have maintained high recognition accuracy even in scenarios with strong noise and data loss. Relying on the standardized integration of multi-source heterogeneous data and a visual decision-making chain, we have formed a full-process closed loop from millimeter-level deformation monitoring, minute-level disaster assessment to dynamic rescue dispatch, significantly enhancing the efficiency of analyzing high-dimensional, multi-scale disaster data and the scientific nature of emergency decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is a system block diagram of the present invention. DETAILED DESCRIPTION

[0053] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0054] like Figure 1 As shown, the embodiment of the present invention provides an earthquake disaster scenario recognition method based on deep learning, and the specific steps are as follows:

[0055] S1. Data collection and preprocessing

[0056] S1.1 Multi-source heterogeneous data collection

[0057] Establish a comprehensive database containing earthquake waveform data, surface deformation data, building structure data, geographic information data, and historical disaster record data. Specifically, it includes: P-wave and S-wave raw waveform data recorded by the seismic network, with a sampling frequency of no less than 100Hz; millimeter-level surface deformation data acquired by InSAR satellite remote sensing; high-resolution disaster site image data acquired by drone aerial photography; structural parameter data of urban building information models (BIM); historical earthquake intensity distribution maps and disaster damage assessment reports;

[0058] S1.2 Data Standardization

[0059] In the data standardization processing stage, it is necessary to build a unified data processing framework with spatiotemporal consistency. First, all geospatial data are accurately aligned using the WGS84 global geodetic system. The geographic reference deviations caused by different sensors and observation platforms are eliminated through coordinate conversion algorithms to ensure the accurate matching of multiple data such as seismic station coordinates, satellite remote sensing images, and building model positions in three-dimensional space. On this basis, a global time synchronization mechanism based on Coordinated Universal Time (UTC) is established to implement millisecond-level time alignment for heterogeneous time series data such as seismic waveform recorders, satellite transit times, and drone collection times, and to build a data association system under a continuous spatiotemporal coordinate system. For high-resolution aerial images, disaster relief images, and other disaster relief images, a global time synchronization mechanism is established based on Coordinated Universal Time (UTC). For unstructured data such as text reports from disaster sites, an image vectorization algorithm is used to convert raster data into vector layers with clear topological relationships. Natural language processing techniques are also used to extract key semantic features from the text and encode them into structured vectors. Finally, multi-dimensional data normalization is used to eliminate physical dimension differences. Numerical data such as earthquake waveform amplitude, surface deformation values, and building structural parameters are scaled using Min-Max normalization and Z-score normalization methods, respectively. Categorical variables are then encoded using one-hot encoding and embedded representation. This ultimately forms a standardized dataset with unified spatiotemporal benchmarks, standardized data structures, and balanced feature distribution, laying a high-quality data foundation for the subsequent construction of deep learning models.

[0060] S2. Deep learning model construction

[0061] S2.1. Hybrid Architecture Neural Network

[0062] In the multimodal feature extraction network design during the deep learning model construction phase, it is necessary to build a collaborative hybrid neural network architecture to fully exploit the multi-source features of earthquake disaster data;

[0063] A three-dimensional convolutional neural network is used to process spatiotemporally continuous seismic waveform data. A convolution kernel with adaptive receptive field adjustment is designed, with a size of 5×5×3. A hierarchical spatiotemporal feature extraction module is used to capture key seismic phase features such as P-wave first arrival and S-wave amplitude changes. A dilated convolution is also integrated with a dilation rate of 2 to expand the temporal context perception range of the waveform sequence.

[0064] For high-resolution remote sensing image data, an improved U-Net architecture is constructed. ResNet-50 is used as the backbone network in the encoder part and a channel attention mechanism is introduced. Multi-scale feature fusion is implemented in the decoder stage. Low-level texture features are aggregated with high-level semantic features across layers through skip connections. A directional feature enhancement module is designed specifically for disaster markers such as collapsed buildings and surface cracks.

[0065] When processing building complex structural data, a graph convolutional network is used to establish a topological relationship analysis model. Individual buildings are abstracted into graph nodes, with node attributes including parameters such as structure type, floor height, and construction year. Spatial adjacency relationships are constructed through Voronoi diagrams as the basis for edge connections. A graph attention layer is used to dynamically learn the mechanical interaction weights between buildings, and a graph pooling operation based on physical constraints is designed to preserve key structural vulnerability characteristics.

[0066] A bidirectional long-short-term memory network is constructed to capture the temporal dynamics of disaster evolution. In the temporal dimension, 128 hidden units are set and a hierarchical propagation mechanism is adopted. The forward layer captures the propagation characteristics of the earthquake source rupture process, and the backward layer reversely analyzes the disaster chain reaction path. The contribution of features from different time phases is weighted and fused through a temporal attention mechanism. Ultimately, a multimodal feature expression system is formed that takes into account spatial topology, temporal evolution, and physical mechanisms. This provides a basic feature representation with strong discriminability for subsequent cross-modal feature fusion.

[0067] S2.2 Feature Fusion

[0068] After the initial extraction of multimodal features, there is a significant modality gap between the original heterogeneous features. To effectively integrate these cross-domain features, it is necessary to design an adaptive attention mechanism to integrate the deep features of multi-source heterogeneous data.

[0069] First, a cross-modal feature alignment layer is used to eliminate heterogeneity in data such as earthquake waveforms, remote sensing images, and building structures. The dynamic time warping (DTW) algorithm is used to align temporal features of different sampling rates, and a feature projection matrix is ​​constructed to achieve semantic matching in high-dimensional space. On this basis, a multi-head self-attention mechanism is introduced to establish an inter-modal correlation matrix. Through parallel computing, multi-dimensional dependency patterns such as the interaction between spatial deformation fields and building complex topology, and the causal relationship between earthquake source rupture processes and surface responses are captured. A differentiable feature selection gating unit is further developed. Based on the gated recurrent network (GRU), the contribution of each modal feature is dynamically evaluated, redundant features are softly masked, and a channel attention weighting strategy is used to focus on key disaster identification features. Finally, a hierarchical feature aggregation module is used to map multi-scale features to a unified semantic space. Depthwise separable convolution is used to compress feature dimensions, and residual connections are combined to retain the original information to generate a 128-dimensional fused feature vector. This vector simultaneously encodes the physical mechanism, spatial distribution, and temporal evolution characteristics of the disaster, providing a highly discriminative input representation for subsequent classification tasks.

[0070] S2.3 Scenario Recognition Classifier

[0071] The 128-dimensional high-density feature vector generated through adaptive feature fusion not only retains the physical properties of the original data but also contains the underlying laws of disaster evolution. To transform these abstract features into actionable disaster scenario judgments, a hybrid classification decision-making system is constructed, and joint reasoning of disaster factors is achieved through a multi-task learning framework.

[0072] First, a deep residual network (ResNet-50) is adopted as the backbone architecture. Skip connections are used to alleviate the gradient vanishing problem. A global average pooling layer is used to map the fused features to the disaster level space, achieving initial classification of intensity from VI to XI. For abnormal disaster modes, such as earthquake swarms and induced landslides, a support vector data description (SVDD) is introduced to construct a hypersphere decision boundary, and kernel function mapping is used to detect dangerous scenarios that deviate from the normal distribution. A conditional random field (CRF) post-processing module is designed to encode spatial continuity constraints such as the degree of building damage and the direction of surface rupture zones into energy functions. The graph cut algorithm is used to optimize the regional consistency of the classification results. Finally, a multi-task learning framework is constructed. By sharing underlying features and collaboratively designing independent task heads, the triple recognition results of disaster type, intensity level, and impact range are simultaneously output. The confidence of each prediction result is evaluated through the uncertainty quantification module, forming a decision output that is both physically interpretable and engineering practical.

[0073] S3, model training and optimization

[0074] S3.1. Training Strategy Design

[0075] During the model training phase, a progressive optimization strategy is adopted to improve learning efficiency and generalization ability: first, the network parameters are unfrozen in stages, with the feature extraction layer prioritized for training to capture fundamental patterns, followed by joint optimization of the classifier parameters. A curriculum learning mechanism is introduced to gradually transition from single focal mechanism samples to multi-type composite disaster scenarios. Adversarial training is simultaneously implemented to enhance the model's robustness to noise interference by generating adversarial perturbation samples. In response to the long-tail distribution characteristics of earthquake disaster data, a dynamic weighted cross-entropy loss function is designed to adaptively adjust the weight coefficient based on the frequency of class samples to balance the class imbalance problem in the mainshock and aftershock identification task.

[0076] S3.2 Hyperparameter Optimization

[0077] After completing the basic architecture design of the training strategy, improving model performance will rely on accurate exploration of the hyperparameter space. To this end, an intelligent parameter adjustment system is built to achieve the global optimal configuration of model parameters through multi-dimensional optimization strategies.

[0078] A Bayesian optimization algorithm was used to search for learning rate combinations (with a base learning rate range of 1e-5 to 1e-3), combined with Neural Architecture Search (NAS) to determine the Pareto optimal solution for network depth and convolution kernel size. A genetic algorithm was used to optimize the batch size (32-256) and the L2 regularization coefficient (1e-4 to 1e-2), and a multi-generation population evolution was used to select a parameter set that resists overfitting. A hyperparameter response surface analysis model was constructed to quantify the impact of parameter interactions on recognition accuracy and inference speed, ultimately determining the global optimal configuration that balances computational efficiency (FLOPs ≤ 15G) and classification performance (F1-score ≥ 0.92).

[0079] S4: Scenario Identification and Decision Support Stage

[0080] S4.1, Real-time Inference Engine

[0081] To meet the real-time demands of earthquake emergency response, an edge computing-based inference system was designed. This system converts floating-point models into 8-bit integers using mixed-precision quantization technology, enabling FPGA hardware acceleration deployment. A streaming data processing pipeline was constructed to support parallel access and online updates of multi-source sensor data. A lightweight inference engine was integrated to ensure sub-second response latency (<500ms). Simultaneously, an uncertainty quantization module was developed, utilizing the Monte Carlo Dropout method to output prediction confidence, providing a reliability assessment basis for decision-making.

[0082] S4.2 Visualization Interaction Platform

[0083] Build a 3D disaster situation awareness system, using WebGL technology to achieve real-time browser rendering, support dynamic overlay display of earthquake focal mechanism spheres, surface deformation fields, and building damage heat maps, and trace back the key evolution process from 10 seconds before to 30 minutes after the disaster.

[0084] S4.3. Decision support

[0085] An intelligent auxiliary decision-making chain is established to automatically generate emergency response plans based on scenario recognition results, and rescue route planning is optimized by combining real-time road conditions and aftershock probability predictions. At the same time, the probability and spatiotemporal distribution of derivative risks such as landslides and barrier lakes are predicted through the LSTM network to provide early warning of secondary disasters. A building reinforcement priority list is output, and a weighted ranking is performed based on the comprehensive structural damage index, population density, and economic value to provide a quantitative decision-making basis for post-disaster reconstruction.

[0086] Through the standardized integration of multi-source heterogeneous data, a hybrid neural network architecture is constructed to realize spatiotemporal feature extraction, an adaptive attention mechanism is used to fuse cross-modal features, a multi-task classifier is designed to synchronously output disaster type, intensity and range, and a combination of progressive training strategy and Bayesian hyperparameter optimization to improve model performance, an edge computing engine is deployed to achieve sub-second real-time reasoning, and by establishing a visualization platform and an intelligent decision-making chain, a full-process closed-loop management is achieved from disaster identification, dynamic evolution analysis to emergency resource scheduling, effectively improving the accuracy of earthquake disaster assessment, response speed and scientific decision-making.

[0087] Among them, S1.1 addresses the common sample imbalance problem in earthquake disaster scenario identification and requires the construction of a multi-level composite data enhancement system. First, based on the conditional generative adversarial network (CGAN) framework, by inputting historical earthquake parameter constraints and combining latent space feature decoupling technology, rare disaster scenario samples with physical rationality are synthesized, including rare cases such as high-intensity earthquakes with low probability of occurrence and secondary disasters in special geological structure areas.

[0088] Through the conditional generative adversarial network, physically reasonable rare disaster scenario samples are generated, which effectively alleviates the sample imbalance problem and improves the model's ability to identify low-probability high-intensity earthquakes and special geological secondary disasters.

[0089] Among them, after the disaster scenario samples are generated, spatial enhancement processing is performed on the multi-source observation data on this basis, using geometric deformation operations such as random rotation, translation, and affine transformation, and introducing texture perturbation strategies such as Gaussian noise, random occlusion, and color perturbation to enhance the model's robustness to sensor errors and on-site environmental interference.

[0090] A composite data augmentation strategy combining geometric deformation and texture perturbation enhances the model's adaptability to sensor errors, environmental occlusion, and lighting changes, improving generalization performance in complex field environments.

[0091] Among them, the angle range of random rotation is ±30°, the maximum offset of translation is 20%, the signal-to-noise ratio of Gaussian noise is controlled at 30-50dB, the occlusion ratio of random occlusion is 15%-40%, and the HSV space of color perturbation is adjusted by ±10%.

[0092] By precisely controlling the range of data augmentation parameters, the enhancement effect can be maximized while maintaining physical rationality, avoiding feature distortion caused by excessive deformation or noise interference.

[0093] Among them, S1.1 further applies an improved Mixup algorithm to implement cross-category data fusion to address the common sample imbalance problem in earthquake disaster scenario identification. By regulating the linear interpolation coefficient λ (Beta distribution parameter α = 0.4), adjacent intensity level samples are smoothly transitioned in the feature space to construct transitional training samples with continuous disaster intensity changes. At the same time, a three-dimensional spatiotemporal data cube is constructed, continuous seismic phase waveform data are stacked along the time dimension (time step 0.1 second), and the kilometer-level gridded surface deformation field (resolution 100m×100m) is integrated in the spatial dimension. The spatiotemporal joint features are extracted using a three-dimensional convolution kernel (size 5×5×3), and data resampling technology is used to extend the time series length to 2-3 times the original data. Finally, a comprehensive data enhancement scheme covering geometric transformation, noise injection, feature mixing, and spatiotemporal expansion is formed, which increases the training sample size by 3-5 times, significantly improving the model's ability to represent long-tail distribution data and its generalization performance for complex disaster scenarios.

[0094] An improved Mixup algorithm and spatiotemporal data resampling technology are used to construct continuous disaster intensity transition samples and expand the time series length, thereby enhancing the model's ability to characterize the continuous process and weak characteristics of disaster evolution.

[0095] Among them, after the system optimization of the hyperparameter space in S3.2, in order to ensure the generalization ability of the model in complex earthquake scenarios, a multi-dimensional evaluation system including physical constraint verification is established. This system not only verifies the theoretical performance of the model, but also focuses on examining its applicability in actual earthquake engineering; a multi-dimensional comprehensive evaluation system is constructed, and the data set is partitioned according to the epicenter location and occurrence time through a spatiotemporal cross-validation strategy to avoid overfitting and test the generalization ability; a quantitative indicator system including disaster identification accuracy, false alarm rate, spatial positioning error, and response delay is formulated, and a physical constraint verification mechanism is introduced at the same time. The seismic wave propagation equation and the law of conservation of energy are used to verify the rationality of the prediction results, and the adversarial sample generation technology is combined to test the robustness of the model under noise interference and data anomalies, and finally a credibility evaluation report that takes into account both data-driven performance and compliance with physical laws is formed.

[0096] Through physical constraint verification and multi-dimensional evaluation system, we ensure that the model prediction results conform to the laws of earthquake propagation, and combine spatiotemporal cross-validation and adversarial testing to improve reliability and robustness in engineering applications.

[0097] The deep learning-based earthquake disaster scenario recognition system includes the following modules:

[0098] Data acquisition and preprocessing module: used to integrate multi-source heterogeneous data such as earthquake waveforms, surface deformation, building structures, geographic information, and historical disaster records, and preprocess the data;

[0099] Deep learning model building module: used to design a hybrid neural network architecture, integrate an adaptive attention mechanism to achieve cross-modal feature fusion, and build a multi-task classifier to simultaneously output disaster type, intensity level, and impact range;

[0100] Model training module: used to design model training strategies and use Bayesian optimization and genetic algorithms to search for hyperparameters. It also introduces a physical constraint verification mechanism to establish a multi-dimensional evaluation system to ensure that model prediction results conform to earthquake propagation laws and actual engineering needs.

[0101] Visualization Interaction Module: This module is used for real-time overlay display of earthquake focal mechanism spheres, surface deformation fields, and building damage heat maps. It also uses a timeline to trace back the key evolution processes before and after the disaster, providing an intuitive spatiotemporal analysis interface.

[0102] Decision support module: used to generate emergency response plans.

[0103] Through the data acquisition and preprocessing module, deep learning model construction module, model training module, visualization interaction module and auxiliary decision-making module, the full process of earthquake disaster data collection, feature analysis, model training and emergency response is automatically managed. The physical constraint verification and spatiotemporal cross-validation of the model training module are combined to ensure the practicality of the model engineering. The accuracy of disaster situation awareness and the efficiency of emergency response are improved by designing the visualization interaction module and auxiliary decision-making module.

[0104] Among them, the model training module is based on a multimodal heterogeneous data fusion framework. It integrates 3D convolutional networks, U-Net, graph convolutional networks and bidirectional LSTM through a hybrid architecture neural network to construct a multi-level feature extraction system. It uses an adaptive attention fusion mechanism to realize cross-modal alignment and correlation matrix modeling of earthquake waveform spatiotemporal characteristics, surface deformation spatial characteristics, building complex topological characteristics and disaster evolution time series characteristics. It combines deep residual networks, support vector data description and conditional random fields to construct a hybrid classification decision system. It uses a multi-task learning framework to simultaneously optimize disaster type identification, intensity classification and impact range prediction tasks. It realizes the dynamic generation of 128-dimensional high-discriminative feature vectors through differentiable feature selection gating units, and finally forms an end-to-end earthquake disaster scenario recognition model with physical law embedding and spatiotemporal continuity constraints.

[0105] Through the multimodal heterogeneous data fusion framework and hybrid neural network architecture, deep feature interactive modeling of earthquake waveforms, surface deformation, building topology and temporal evolution is achieved. Combined with physical constraints and gating units, highly discriminative features are dynamically generated to form an end-to-end disaster identification system with both spatiotemporal continuity and mechanism explainability.

[0106] Among them, the auxiliary decision-making module can optimize rescue route planning and resource scheduling strategies, and output a building reinforcement priority list based on the comprehensive structural damage index, population density and economic value, providing quantitative decision-making support for post-disaster reconstruction.

[0107] By dynamically optimizing rescue routes and quantitatively evaluating building reinforcement priorities, and integrating multi-dimensional indicators such as structural damage, population density, and economic value, we provide a scientific decision-making basis for post-disaster emergency response and reconstruction planning, significantly improving resource scheduling efficiency and the ability to protect the safety of life and property.

[0108] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0109] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for identifying earthquake disaster scenarios based on deep learning, characterized in that: The specific steps are as follows: S1. Data collection and preprocessing S1.1 Multi-source heterogeneous data collection Establish a comprehensive database containing earthquake waveform data, surface deformation data, building structure data, geographic information data, and historical disaster record data. Specifically, it includes: P-wave and S-wave raw waveform data recorded by the seismic network, with a sampling frequency of no less than 100Hz; millimeter-level surface deformation data acquired by InSAR satellite remote sensing; high-resolution disaster site image data acquired by drone aerial photography; structural parameter data of urban building information models (BIM); historical earthquake intensity distribution maps and disaster damage assessment reports; S1.2 Data Standardization In the data standardization processing stage, it is necessary to build a unified data processing framework with spatiotemporal consistency. First, all geospatial data are accurately aligned using the WGS84 global geodetic system. The geographic reference deviations caused by different sensors and observation platforms are eliminated through coordinate conversion algorithms to ensure the accurate matching of multiple data such as seismic station coordinates, satellite remote sensing images, and building model positions in three-dimensional space. On this basis, a global time synchronization mechanism based on Coordinated Universal Time (UTC) is established to implement millisecond-level time alignment for heterogeneous time series data such as seismic waveform recorders, satellite transit times, and drone collection times, and to build a data association system under a continuous spatiotemporal coordinate system. For high-resolution aerial images, disaster relief images, and other disaster relief images, a global time synchronization mechanism is established based on Coordinated Universal Time (UTC). For unstructured data such as text reports from disaster sites, an image vectorization algorithm is used to convert raster data into vector layers with clear topological relationships. Natural language processing techniques are also used to extract key semantic features from the text and encode them into structured vectors. Finally, multi-dimensional data normalization is used to eliminate physical dimension differences. Numerical data such as earthquake waveform amplitude, surface deformation values, and building structural parameters are scaled using Min-Max normalization and Z-score normalization methods, respectively. Categorical variables are then encoded using one-hot encoding and embedded representation. This ultimately forms a standardized dataset with unified spatiotemporal benchmarks, standardized data structures, and balanced feature distribution, laying a high-quality data foundation for the subsequent construction of deep learning models. S2. Deep learning model construction S2.

1. Hybrid Architecture Neural Network In the multimodal feature extraction network design during the deep learning model construction phase, it is necessary to build a collaborative hybrid neural network architecture to fully exploit the multi-source features of earthquake disaster data; A three-dimensional convolutional neural network is used to process spatiotemporally continuous seismic waveform data. A convolution kernel with adaptive receptive field adjustment is designed, with a size of 5×5×3. A hierarchical spatiotemporal feature extraction module is used to capture key seismic phase features such as P-wave first arrival and S-wave amplitude changes. A dilated convolution is also integrated with a dilation rate of 2 to expand the temporal context perception range of the waveform sequence. For high-resolution remote sensing image data, an improved U-Net architecture is constructed. ResNet-50 is used as the backbone network in the encoder part and a channel attention mechanism is introduced. Multi-scale feature fusion is implemented in the decoder stage. Low-level texture features are aggregated with high-level semantic features across layers through skip connections. A directional feature enhancement module is designed specifically for disaster markers such as collapsed buildings and surface cracks. When processing building complex structural data, a graph convolutional network is used to establish a topological relationship analysis model. Individual buildings are abstracted into graph nodes, with node attributes including parameters such as structure type, floor height, and construction year. Spatial adjacency relationships are constructed through Voronoi diagrams as the basis for edge connections. A graph attention layer is used to dynamically learn the mechanical interaction weights between buildings, and a graph pooling operation based on physical constraints is designed to preserve key structural vulnerability characteristics. A bidirectional long-short-term memory network is constructed to capture the temporal dynamics of disaster evolution. In the temporal dimension, 128 hidden units are set and a hierarchical propagation mechanism is adopted. The forward layer captures the propagation characteristics of the earthquake source rupture process, and the backward layer reversely analyzes the disaster chain reaction path. The contribution of features from different time phases is weighted and fused through a temporal attention mechanism. Ultimately, a multimodal feature expression system is formed that takes into account spatial topology, temporal evolution, and physical mechanisms. This provides a basic feature representation with strong discriminability for subsequent cross-modal feature fusion. S2.2 Feature Fusion After the initial extraction of multimodal features, there is a significant modality gap between the original heterogeneous features. To effectively integrate these cross-domain features, it is necessary to design an adaptive attention mechanism to integrate the deep features of multi-source heterogeneous data. First, a cross-modal feature alignment layer is used to eliminate heterogeneity in data such as earthquake waveforms, remote sensing images, and building structures. The dynamic time warping (DTW) algorithm is used to align temporal features of different sampling rates, and a feature projection matrix is ​​constructed to achieve semantic matching in high-dimensional space. On this basis, a multi-head self-attention mechanism is introduced to establish an inter-modal correlation matrix. Through parallel computing, multi-dimensional dependency patterns such as the interaction between spatial deformation fields and building complex topology, and the causal relationship between earthquake source rupture processes and surface responses are captured. A differentiable feature selection gating unit is further developed. Based on the gated recurrent network (GRU), the contribution of each modal feature is dynamically evaluated, redundant features are softly masked, and a channel attention weighting strategy is used to focus on key disaster identification features. Finally, a hierarchical feature aggregation module is used to map multi-scale features to a unified semantic space. Depthwise separable convolution is used to compress feature dimensions, and residual connections are combined to retain the original information to generate a 128-dimensional fused feature vector. This vector simultaneously encodes the physical mechanism, spatial distribution, and temporal evolution characteristics of the disaster, providing a highly discriminative input representation for subsequent classification tasks. S2.3 Scenario Recognition Classifier The 128-dimensional high-density feature vector generated through adaptive feature fusion not only retains the physical properties of the original data but also contains the underlying laws of disaster evolution. To transform these abstract features into actionable disaster scenario judgments, a hybrid classification decision-making system is constructed, and joint reasoning of disaster factors is achieved through a multi-task learning framework. First, a deep residual network (ResNet-50) is adopted as the backbone architecture. Skip connections are used to alleviate the gradient vanishing problem. A global average pooling layer is used to map the fused features to the disaster level space, achieving initial classification of intensity from VI to XI. For abnormal disaster modes, such as earthquake swarms and induced landslides, a support vector data description (SVDD) is introduced to construct a hypersphere decision boundary, and kernel function mapping is used to detect dangerous scenarios that deviate from the normal distribution. A conditional random field (CRF) post-processing module is designed to encode spatial continuity constraints such as the degree of building damage and the direction of surface rupture zones into energy functions. The graph cut algorithm is used to optimize the regional consistency of the classification results. Finally, a multi-task learning framework is constructed. By sharing underlying features and collaboratively designing independent task heads, the triple recognition results of disaster type, intensity level, and impact range are simultaneously output. The confidence of each prediction result is evaluated through the uncertainty quantification module, forming a decision output that is both physically interpretable and engineering practical. S3, model training and optimization S3.

1. Training Strategy Design During the model training phase, a progressive optimization strategy is adopted to improve learning efficiency and generalization ability: first, the network parameters are unfrozen in stages, with the feature extraction layer prioritized for training to capture fundamental patterns, followed by joint optimization of the classifier parameters. A curriculum learning mechanism is introduced to gradually transition from single focal mechanism samples to multi-type composite disaster scenarios. Adversarial training is simultaneously implemented to enhance the model's robustness to noise interference by generating adversarial perturbation samples. In response to the long-tail distribution characteristics of earthquake disaster data, a dynamic weighted cross-entropy loss function is designed to adaptively adjust the weight coefficient based on the frequency of class samples to balance the class imbalance problem in the mainshock and aftershock identification task. S3.2 Hyperparameter Optimization After completing the basic architecture design of the training strategy, improving model performance will rely on accurate exploration of the hyperparameter space. To this end, an intelligent parameter adjustment system is built to achieve the global optimal configuration of model parameters through multi-dimensional optimization strategies. A Bayesian optimization algorithm was used to search for learning rate combinations (with a base learning rate range of 1e-5 to 1e-3), combined with Neural Architecture Search (NAS) to determine the Pareto optimal solution for network depth and convolution kernel size. A genetic algorithm was used to optimize the batch size (32-256) and the L2 regularization coefficient (1e-4 to 1e-2), and a multi-generation population evolution was used to select a parameter set that resists overfitting. A hyperparameter response surface analysis model was constructed to quantify the impact of parameter interactions on recognition accuracy and inference speed, ultimately determining the global optimal configuration that balances computational efficiency (FLOPs ≤ 15G) and classification performance (F1-score ≥ 0.92). S4: Scenario Identification and Decision Support Stage S4.1, Real-time Inference Engine To meet the real-time demands of earthquake emergency response, an edge computing-based inference system was designed. This system converts floating-point models into 8-bit integers using mixed-precision quantization technology, enabling FPGA hardware acceleration deployment. A streaming data processing pipeline was constructed to support parallel access and online updates of multi-source sensor data. A lightweight inference engine was integrated to ensure sub-second response latency (<500ms). Simultaneously, an uncertainty quantization module was developed, utilizing the Monte Carlo Dropout method to output prediction confidence, providing a reliability assessment basis for decision-making. S4.2 Visual Interaction Platform Build a 3D disaster situation awareness system, using WebGL technology to achieve real-time browser rendering, support dynamic overlay display of earthquake focal mechanism spheres, surface deformation fields, and building damage heat maps, and trace back the key evolution process from 10 seconds before to 30 minutes after the disaster. S4.

3. Decision support An intelligent auxiliary decision-making chain is established to automatically generate emergency response plans based on scenario recognition results, and rescue route planning is optimized by combining real-time road conditions and aftershock probability predictions. At the same time, the probability and spatiotemporal distribution of derivative risks such as landslides and barrier lakes are predicted through the LSTM network to provide early warning of secondary disasters. A building reinforcement priority list is output, and a weighted ranking is performed based on the comprehensive structural damage index, population density, and economic value to provide a quantitative decision-making basis for post-disaster reconstruction.

2. The earthquake disaster scenario recognition method based on deep learning according to claim 1 is characterized in that: As described in S1.1, to address the common sample imbalance problem in earthquake disaster scenario identification, it is necessary to build a multi-level composite data enhancement system. First, based on the conditional generative adversarial network (CGAN) framework, by inputting historical earthquake parameter conditional constraints and combining the latent space feature decoupling technology, rare disaster scenario samples with physical rationality are synthesized, including rare cases such as high-intensity earthquakes with low probability of occurrence and secondary disasters in special geological structure areas.

3. The earthquake disaster scenario recognition method based on deep learning according to claim 2, characterized in that: After the disaster scenario samples are generated, spatial enhancement processing is performed on the multi-source observation data on this basis, using geometric deformation operations such as random rotation, translation, and affine transformation, and introducing texture perturbation strategies such as Gaussian noise, random occlusion, and color perturbation to enhance the model's robustness to sensor errors and on-site environmental interference.

4. The earthquake disaster scenario recognition method based on deep learning according to claim 3 is characterized in that: The angle range of the random rotation is ±30°, the maximum offset of the translation is 20%, the signal-to-noise ratio of the Gaussian noise is controlled at 30-50dB, the occlusion ratio of the random occlusion is 15%-40%, and the HSV space of the color disturbance is adjusted by ±10%.

5. The earthquake disaster scenario recognition method based on deep learning according to claim 1, characterized in that: As described in S1.1, to address the common sample imbalance problem in earthquake disaster scenario identification, an improved Mixup algorithm is further applied to implement cross-category data fusion. By regulating the linear interpolation coefficient λ (Beta distribution parameter α = 0.4), a smooth transition process is performed on adjacent intensity level samples in the feature space to construct transitional training samples with continuous disaster intensity changes; At the same time, a three-dimensional spatiotemporal data cube is constructed, continuous seismic phase waveform data are stacked along the time dimension (time step of 0.1 second), and the kilometer-level gridded surface deformation field is integrated in the spatial dimension (resolution 100m×100m). The spatiotemporal joint features are extracted through a three-dimensional convolution kernel (5×5×3 size), and data resampling technology is used to expand the time series length to 2-3 times of the original data. Finally, a comprehensive data enhancement scheme covering geometric transformation, noise injection, feature mixing, and spatiotemporal expansion is formed, which increases the training sample size by 3-5 times, significantly improving the model's ability to represent long-tail distribution data and the generalization performance of complex disaster scenarios.

6. The earthquake disaster scenario recognition method based on deep learning according to claim 1, characterized in that: After the systematic optimization of the hyperparameter space as described in S3.2, in order to ensure the generalization ability of the model in complex earthquake scenarios, a multi-dimensional evaluation system including physical constraint verification is established. This system not only verifies the theoretical performance of the model, but also focuses on examining its applicability in actual earthquake engineering; a multi-dimensional comprehensive evaluation system is constructed, and the data set is partitioned according to the epicenter location and occurrence time through a spatiotemporal cross-validation strategy to avoid overfitting and test the generalization ability; a quantitative indicator system including disaster identification accuracy, false alarm rate, spatial positioning error, and response delay is formulated, and a physical constraint verification mechanism is introduced at the same time. The seismic wave propagation equation and the law of conservation of energy are used to verify the rationality of the prediction results, and the adversarial sample generation technology is combined to test the robustness of the model under noise interference and data anomalies, and finally a credibility evaluation report that takes into account both data-driven performance and compliance with physical laws is formed.

7. The earthquake disaster scenario recognition system based on deep learning is characterized by: Includes the following modules: Data acquisition and preprocessing module: used to integrate multi-source heterogeneous data such as earthquake waveforms, surface deformation, building structures, geographic information, and historical disaster records, and preprocess the data; Deep learning model building module: used to design a hybrid neural network architecture, integrate an adaptive attention mechanism to achieve cross-modal feature fusion, and build a multi-task classifier to simultaneously output disaster type, intensity level, and impact range; Model training module: used to design model training strategies and use Bayesian optimization and genetic algorithms to search for hyperparameters. It also introduces a physical constraint verification mechanism to establish a multi-dimensional evaluation system to ensure that model prediction results conform to earthquake propagation laws and actual engineering needs. Visualization Interaction Module: This module is used for real-time overlay display of earthquake focal mechanism spheres, surface deformation fields, and building damage heat maps. It also uses a timeline to trace back the key evolution processes before and after the disaster, providing an intuitive spatiotemporal analysis interface. Decision support module: used to generate emergency response plans.

8. The deep learning-based earthquake disaster scenario recognition system according to claim 7, characterized in that: The model training module is based on a multimodal heterogeneous data fusion framework. It integrates 3D convolutional networks, U-Net, graph convolutional networks and bidirectional LSTM through a hybrid architecture neural network to construct a multi-level feature extraction system. It uses an adaptive attention fusion mechanism to achieve cross-modal alignment and correlation matrix modeling of earthquake waveform spatiotemporal characteristics, surface deformation spatial characteristics, building complex topological characteristics and disaster evolution time series characteristics. It combines deep residual networks, support vector data description and conditional random fields to construct a hybrid classification decision system. It uses a multi-task learning framework to simultaneously optimize disaster type identification, intensity classification and impact range prediction tasks. It realizes the dynamic generation of 128-dimensional high-discriminative feature vectors through differentiable feature selection gating units, and finally forms an end-to-end earthquake disaster scenario recognition model with physical law embedding and spatiotemporal continuity constraints.

9. The deep learning-based earthquake disaster scenario recognition system according to claim 7, characterized in that: The auxiliary decision-making module can optimize rescue path planning and resource scheduling strategies, comprehensively integrate structural damage index, population density and economic value to output a building reinforcement priority list, and provide quantitative decision-making support for post-disaster reconstruction.

Images