Geological disaster automatic identification method and system based on multi-source remote sensing data

Abstract

The invention relates to a geological disaster automatic identification method and system based on multi-source remote sensing data, and belongs to the technical field of geological disaster monitoring, and the method comprises the steps: collecting the multi-source remote sensing data of a to-be-detected region, carrying out the cross-modal registration, and generating a registered multi-source data set; carrying out multi-modal feature extraction based on the multi-source data set and carrying out space-time correlation analysis to obtain a multi-modal feature map; performing dynamic weight distribution on the multi-modal feature map, generating a fusion feature vector, inputting the fusion feature vector into a pre-trained geological disaster prediction model, and outputting a geological disaster probability map; performing binarization segmentation on the geological disaster probability graph to obtain a potential disaster area mask; and according to the geological disaster probability map and the potential disaster area mask, based on a preset joint determination rule of the surface deformation rate and the gradient characteristics, carrying out risk grade division on the potential disaster area to obtain a risk grade distribution map of the to-be-detected area. The disaster prediction precision can be improved, and the emergency response capability is enhanced.

Description

Technical Field

[0001] The present application relates to the technical field of geological disaster monitoring, and in particular to a method and system for automatically identifying geological disasters based on multi-source remote sensing data. Background Art

[0002] Geological hazards are natural disasters caused by factors such as geological conditions, meteorological changes, or human activities. They typically include landslides, debris flows, ground subsidence, and mountain collapses. These disasters not only have a significant impact on human economic and social activities, but also cause severe damage to the natural environment. In recent years, with the acceleration of urbanization, many cities and villages have been built in mountainous or coastal areas with complex geological conditions, significantly increasing the risk of geological hazards in these areas. Especially in areas with high rainfall and poor geological conditions, disasters such as landslides and debris flows are frequent, posing a serious threat to the safety of people's lives and property.

[0003] Currently, remote sensing technology, especially satellite remote sensing technology, has become an important means of monitoring and early warning of geological disasters. Remote sensing data is acquired through platforms such as satellites or drones, can cover large geographical areas, and provide real-time or near-real-time data. However, because geological disasters are often sudden and unpredictable, different types of disasters have different triggering mechanisms and development processes. For example, the occurrence of landslides is closely related to factors such as slope, geological structure, and rainfall, while the occurrence of debris flows may be affected by the combined influence of surface vegetation, soil moisture, and precipitation. Common risk assessment methods are often unable to accurately assess the probability and severity of disasters based on the interaction of multiple factors, resulting in low accuracy of emergency response and disaster prevention measures. Summary of the Invention

[0004] In order to improve the accuracy of disaster prediction, this application provides a method and system for automatically identifying geological disasters based on multi-source remote sensing data.

[0005] In the first aspect, the present application provides a method for automatically identifying geological hazards based on multi-source remote sensing data, which adopts the following technical solutions: A method for automatically identifying geological hazards based on multi-source remote sensing data, the method comprising: Collecting multi-source remote sensing data of the area to be measured; wherein the multi-source remote sensing data includes optical image data, synthetic aperture radar data and digital elevation model data; Performing cross-modal registration on the multi-source remote sensing data to generate a registered multi-source dataset; Extracting multimodal features and performing spatiotemporal correlation analysis based on the multi-source data set to obtain a multimodal feature map; wherein the multimodal features include vegetation cover index features of the optical image data, surface deformation rate of the synthetic aperture radar data, and slope features of the digital elevation model data; Dynamically assigning weights to the multimodal feature maps based on a channel attention mechanism to generate a fused feature vector; Inputting the fused feature vector into a pre-trained geological disaster prediction model and outputting a geological disaster probability map; Binarize and segment the geological disaster probability map according to a preset threshold value to obtain a potential disaster area mask; According to the geological hazard probability map and the potential hazard area mask, based on the preset joint judgment rule of the surface deformation rate and slope characteristics, the potential hazard area is divided into risk levels to obtain a risk level distribution map of the area to be tested.

[0006] By adopting the above technical solution, based on the precise collection and registration of multi-source remote sensing data, combined with the application of dynamic weight allocation and deep learning models, we achieve efficient and accurate automatic identification of geological hazards. Through feature extraction and spatiotemporal correlation analysis, we can comprehensively and accurately assess the risk of geological hazards, generate hazard probability maps, and classify risk levels based on surface deformation rate and slope, providing a scientific basis for emergency management. This application has important application value in improving disaster prediction accuracy and enhancing emergency response capabilities.

[0007] Optionally, the step of extracting multimodal features and performing spatiotemporal correlation analysis based on the multi-source data set to obtain a multimodal feature map includes: Based on the multi-source data set, extracting vegetation cover index characteristics of the optical image data, surface deformation rate of the synthetic aperture radar data, and slope characteristics of the digital elevation model data respectively; A spatiotemporal correlation matrix is ​​constructed, and a cross-modal correlation analysis is performed on the vegetation cover index characteristics, surface deformation rate and slope characteristics through a graph neural network to generate a multimodal feature map.

[0008] By adopting the above technical solutions, the spatiotemporal correlation matrix and graph neural network can capture the complex relationship between disaster characteristics and generate multimodal feature maps, providing a reliable basis for disaster early warning. The generated multimodal feature maps provide accurate spatial distribution information for the identification and prediction of geological disasters, which helps to take emergency measures in a timely manner and reduce disaster losses.

[0009] Optionally, the steps of constructing a spatiotemporal correlation matrix, performing cross-modal correlation analysis on the vegetation cover index characteristics, surface deformation rate, and slope characteristics through a graph neural network, and generating a multimodal feature map include: Performing spatiotemporal alignment on the vegetation cover index feature, the surface deformation rate, and the slope feature to generate a spatiotemporal aligned rasterized feature matrix; Based on the grid feature matrix, each grid unit is defined as a graph node; the graph node attributes include vegetation cover index characteristics, surface deformation rate and slope characteristics; Calculate the feature similarity weight and spatial adjacency weight between each graph node and perform weighted fusion to construct a spatiotemporal correlation matrix; Input the spatiotemporal correlation matrix into the graph neural network, update the node implicit features through neighborhood node sampling and feature aggregation, and iterate based on the multi-layer graph neural network to generate an updated node feature matrix; Reconstructing the node feature matrix at each time point into a two-dimensional feature map according to the spatial coordinates of the grid cells; The two-dimensional feature maps of multiple time slices are temporally fused to obtain a multimodal feature map.

[0010] By adopting the above technical solutions, based on spatiotemporal alignment, feature aggregation and time series fusion, the prediction accuracy of geological disaster risks is effectively improved, so that the potential risks of disasters can be accurately captured in different time and space dimensions, providing scientific and accurate decision-making support for disaster warning and emergency response.

[0011] Optionally, the step of dynamically allocating weights on the multimodal feature map based on a channel attention mechanism to generate a fused feature vector includes: Determining, based on a preset channel mapping table, a feature type corresponding to each channel in the multimodal feature map; Normalizing each channel in the multimodal feature map to generate a standardized multimodal feature map; Performing global feature extraction on the standardized multimodal feature map to generate a channel-level statistical vector; Inputting the channel-level statistical vector into a classification model and outputting a probability distribution of disaster types; Determine the dominant disaster type according to the probability distribution and generate a corresponding disaster type code; Constructing a joint feature vector based on the disaster type coding and the standardized multimodal feature map; Mapping the joint feature vector through a weight generation network to generate a channel weight vector; Performing channel-level weighting on the channel weight vector and the standardized multimodal feature map to generate a weighted feature map; The weighted feature map is spatially compressed to generate a fused feature vector.

[0012] By employing this technical solution, multimodal feature maps are dynamically weighted, effectively combining hazard type information with various geological hazard-related features (such as deformation rate, slope, and NDVI) to generate precise hazard identification feature vectors. Through dynamic weight assignment and feature weighting, the system can accurately identify different types of geological hazards (such as landslides, debris flows, and collapses).

[0013] Optionally, the method further includes a step of training the geological disaster prediction model, the training step including: Acquire a sample data set and perform data preprocessing; wherein the sample data set includes a fused feature vector and a binary label map of a historical disaster area; Dividing the sample data set into a training set, a validation set, and a test set; The training set is input into a pre-built convolutional neural network for training, the error between the predicted probability map and the true label map is calculated by the back propagation algorithm, the gradient is calculated layer by layer, and the weights and biases of the convolutional neural network are updated; Validating the convolutional neural network based on the validation set, evaluating the performance of the convolutional neural network and adjusting model hyperparameters according to the validation results; The prediction ability of the adjusted convolutional neural network is tested based on the test set to obtain the trained geological disaster prediction model.

[0014] By adopting the above technical solution, building a high-quality sample data set and using convolutional neural networks (CNN) for training, and combining verification and testing steps to optimize and evaluate the model, we ultimately obtained an accurate geological disaster prediction model. This ensures efficient data-driven design and powerful computing capabilities, enabling the prediction model to provide highly accurate and stable disaster prediction results in practical applications.

[0015] Optionally, the step of classifying the potential disaster areas into risk levels based on the geological hazard probability map and the potential disaster area mask and a preset joint determination rule of the surface deformation rate and slope characteristics to obtain a risk level distribution map of the area to be measured includes: The surface deformation rate map and slope characteristic map are obtained based on the multimodal characteristic map; Clipping the surface deformation rate map and slope characteristic map based on the potential disaster area mask to extract the deformation rate matrix and slope matrix within the potential disaster area mask; spatially aligning the geological hazard probability map, deformation rate matrix, and slope matrix; Load the preset risk level classification rule configuration file, analyze the threshold parameters of surface deformation rate and slope, and obtain the judgment condition function; Traversing each grid cell of the geological hazard probability map, determining areas where the probability value is greater than a preset probability threshold as potential hazard areas; Obtaining surface deformation rate data and slope characteristic data of the potential disaster area and performing a determination based on the determination condition function, dividing the risk level of the potential disaster area into risk levels, and obtaining a risk level coding matrix; A risk level distribution map of the area to be tested is generated according to the risk level coding matrix.

[0016] By adopting this technical solution, a risk classification system for geological hazard areas is implemented based on a combined determination method of deformation rate and slope. Through steps such as spatial alignment, rule loading, and pixel-by-pixel risk assessment, this system accurately and efficiently classifies potential hazard areas according to pre-defined rules, ultimately generating a visual risk level distribution map. This provides a reliable risk assessment tool for geological hazard monitoring systems, supporting disaster warning and decision-making, thereby improving the timeliness and accuracy of disaster response.

[0017] In a second aspect, the present application provides a geological disaster automatic identification system based on multi-source remote sensing data, which adopts the following technical solutions: A geological disaster automatic identification system based on multi-source remote sensing data, the identification system comprising: An acquisition module is used to acquire multi-source remote sensing data of the area to be measured; wherein the multi-source remote sensing data includes optical image data, synthetic aperture radar data and digital elevation model data; A registration module, configured to perform cross-modal registration on the multi-source remote sensing data to generate a registered multi-source dataset; A multimodal feature map generation module is configured to extract multimodal features and perform spatiotemporal correlation analysis based on the multi-source data set to obtain a multimodal feature map; wherein the multimodal features include vegetation cover index features of the optical image data, surface deformation rate of the synthetic aperture radar data, and slope features of the digital elevation model data; A feature fusion module, configured to dynamically assign weights to the multimodal feature maps based on a channel attention mechanism to generate a fused feature vector; A geological disaster prediction module is used to input the fused feature vector into a pre-trained geological disaster prediction model and output a geological disaster probability map; A binary segmentation module is used to perform binary segmentation on the geological disaster probability map according to a preset threshold value to obtain a potential disaster area mask; The risk level classification module is used to classify the potential disaster area into risk levels based on the geological disaster probability map and the potential disaster area mask, based on the preset joint judgment rules of the surface deformation rate and slope characteristics, and obtain a risk level distribution map of the area to be tested.

[0018] Optionally, the multimodal feature map generation module includes: a multimodal feature extraction unit, configured to extract, based on the multi-source data set, vegetation cover index features of the optical image data, surface deformation rate of the synthetic aperture radar data, and slope features of the digital elevation model data; The cross-modal correlation analysis unit is used to construct a spatiotemporal correlation matrix, perform cross-modal correlation analysis on the vegetation cover index characteristics, surface deformation rate and slope characteristics through a graph neural network, and generate a multimodal feature map.

[0019] In a third aspect, the present application provides a computer device that adopts the following technical solution: A computer device comprises a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method according to the first aspect.

[0020] In a fourth aspect, the present application provides a computer-readable storage medium, which adopts the following technical solution: A computer-readable storage medium stores a computer program capable of being loaded by a processor and executing any one of the methods in the first aspect. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a first flow chart of a method for automatic identification of geological hazards based on multi-source remote sensing data in one of the embodiments of the present application.

[0022] Figure 2 This is a second flow chart of a method for automatic identification of geological hazards based on multi-source remote sensing data in one of the embodiments of the present application.

[0023] Figure 3 This is a third flow chart of a method for automatic identification of geological hazards based on multi-source remote sensing data in one of the embodiments of the present application.

[0024] Figure 4 This is a fourth flow chart of a method for automatic identification of geological hazards based on multi-source remote sensing data in one of the embodiments of the present application.

[0025] Figure 5 This is the fifth flow chart of a method for automatic identification of geological hazards based on multi-source remote sensing data in one of the embodiments of the present application.

[0026] Figure 6 This is the sixth flow chart of a method for automatic identification of geological disasters based on multi-source remote sensing data in one of the embodiments of the present application. DETAILED DESCRIPTION

[0027] In order to make the purpose, technical solutions and advantages of this application more clear, the following Figure 1-6 It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application.

[0028] The embodiment of the present application discloses a method for automatically identifying geological hazards based on multi-source remote sensing data.

[0029] Reference Figure 1 , a method for automatic identification of geological hazards based on multi-source remote sensing data, the identification method includes: Step S101, collecting multi-source remote sensing data of the area to be measured; Among them, multi-source remote sensing data includes optical image data, synthetic aperture radar data and digital elevation model data; Specifically, when identifying geological hazards, remote sensing technology provides an effective means of acquiring a variety of data about the area being measured. Multi-source remote sensing data refers to data from different sensors, each with its own unique advantages, and can provide important information about geological hazards at multiple levels.

[0030] In the embodiments of this application, optical imaging data can provide detailed, clear images of the ground, particularly vegetation cover. Synthetic aperture radar (SAR) data, by transmitting and receiving electromagnetic waves, can penetrate clouds and rain to monitor surface deformation. Digital elevation models (DEMs) provide three-dimensional topographic data of the area being measured, including information on geographic features such as height and slope. Combining this data allows for a more comprehensive understanding and assessment of potential geological hazard risks.

[0031] Step S102, performing cross-modal registration on the multi-source remote sensing data to generate a registered multi-source dataset; Because these data come from different sensors and are often acquired at different times, angles, and scales, accurately aligning and unifying them into a common coordinate system is crucial for disaster identification. The goal of registration is to ensure that the spatial positions of different datasets are perfectly aligned to facilitate subsequent feature extraction and analysis.

[0032] In some embodiments, commonly used registration methods include image registration algorithms, such as affine transformation, rigid transformation, etc., among which affine transformation is particularly suitable for processing geometric distortion of large-scale images, can maintain the linearity and proportional relationship within the image, and ensure that the error is less than 1.5 pixels, thereby achieving high-precision registration effect.

[0033] It is understandable that accurately registering remote sensing data from different sources ensures consistency within the same spatial reference system, allowing all types of data to be analyzed within the same map space. This not only improves the accuracy of subsequent analysis but also lays the foundation for multi-source data fusion, ensuring that different features (such as surface deformation, slope, and vegetation cover) can be effectively combined to provide comprehensive information on geological hazards.

[0034] Step S103: extracting multimodal features and performing spatiotemporal correlation analysis based on the multi-source data sets to obtain a multimodal feature map; Among them, multimodal features include vegetation cover index features from optical image data, surface deformation rate from synthetic aperture radar data, and slope features from digital elevation model data; It's important to note that the occurrence of geological disasters depends not only on static terrain information but also on dynamic environmental changes. Multimodal feature extraction extracts important features from different types of remote sensing data that aid in disaster identification. These features can be categorized as vegetation cover index from optical images, surface deformation rate from SAR data, and slope characteristics from DEM data.

[0035] Specifically, the vegetation cover index characteristic of optical image data can reflect the health of vegetation in a region, as vegetation growth is closely related to the occurrence of disasters such as landslides and mudslides. SAR data, by capturing the rate of surface deformation, can reveal whether the ground is shifting or deforming, which is crucial for predicting disasters such as landslides. DEM data provides elevation information within a region, enabling the calculation of characteristics such as slope and aspect. These factors also play a significant role in the occurrence of geological disasters, especially in mountainous and hilly areas, where excessive slope is a key factor in landslides and other disasters.

[0036] Spatiotemporal correlation analysis correlates these extracted features with temporal changes. It considers the temporal variations in remote sensing data and can capture dynamic trends in how certain features change over time. For example, the rate of surface deformation in SAR images varies with events like rainfall or earthquakes. Vegetation changes in optical imagery are often correlated with seasonal precipitation or soil moisture variations. Slope changes in DEM data may be due to long-term erosion or other natural factors. Spatiotemporal correlation analysis allows these trends to be modeled and predicted, more accurately identifying the probability and risk of disasters.

[0037] It is understandable that by analyzing multiple characteristic dimensions, potential disaster areas can be identified and sufficient evidence can be provided for subsequent disaster risk assessment. Spatiotemporal correlation analysis, by capturing the relationship between geological hazards and time and space, helps reveal the precursors and trends of disasters, thereby achieving early warning of disasters.

[0038] Step S104: dynamically assign weights to the multimodal feature maps based on the channel attention mechanism to generate a fused feature vector; The channel attention mechanism, among other things, prioritizes the most important features in geological hazard identification by assigning different weights to features from different channels. For example, in landslide identification, surface deformation rate and slope are key features, so these can be given higher weights. The channel attention mechanism automatically adjusts the weightings based on the hazard type to optimize the feature fusion process. However, since different hazard types (such as landslides and debris flows) may exhibit distinct characteristics, the fusion of multi-source data requires dynamic adjustment of the weights of each feature based on the specific scenario. For example, landslide hazards prioritize surface deformation rate, while debris flows may be more sensitive to slope and vegetation cover.

[0039] It is understandable that dynamically adjusting the contribution of features based on the type of disaster improves the accuracy of feature fusion. This approach avoids the "blind weighting" problem found in traditional feature weighting methods, ensuring that each feature contributes the most to the disaster identification model, thereby improving the accuracy of disaster identification.

[0040] Step S105, inputting the fused feature vector into a pre-trained geological disaster prediction model and outputting a geological disaster probability map; Among them, geological disaster prediction models can adopt deep neural networks, especially convolutional neural networks (CNN), which can efficiently process complex geological disaster data. These models can automatically learn the potential patterns in the input data without manual feature design, thereby improving recognition accuracy.

[0041] In the embodiments of the present application, a deep learning model is trained to generate a probability map that shows the likelihood of geological disasters occurring at different locations in the area being tested. The value at each location represents the probability of a disaster occurring in that area, with larger values ​​indicating a higher probability of disaster occurring in that area. During training, the deep learning model learns from a large amount of historical disaster data and can automatically adjust its model parameters, enabling the model to accurately assess the risk of geological disasters when encountering new areas and data.

[0042] Step S106, performing binary segmentation on the geological disaster probability map according to a preset threshold to obtain a potential disaster area mask; By performing binary segmentation on the geological hazard probability map, the area to be tested can be divided into potential disaster areas and non-disaster areas. The key to this step is setting an appropriate threshold to extract high-probability areas from the probability map, forming a disaster area mask. In some embodiments, the threshold can be selected based on pre-defined risk criteria, such as surface deformation rate and slope.

[0043] Step S107 , based on the geological hazard probability map and the potential hazard area mask, and based on the preset joint determination rule of surface deformation rate and slope characteristics, the potential hazard area is classified into risk levels to obtain a risk level distribution map of the area to be tested.

[0044] Specifically, based on predefined criteria (such as a combined determination rule for surface deformation rate and slope), potential disaster areas are classified into risk levels. Surface deformation rate and slope characteristics are key factors in the occurrence of geological hazards, particularly in the identification of landslides and debris flows, as these two characteristics can reflect the vulnerability of an area. Using this combined determination rule, regions can be classified into high, medium, and low risk zones based on the relationship between surface deformation rate and slope.

[0045] Among them, the generated risk level distribution map not only helps in real-time monitoring and assessment of disasters, but also provides a basis for disaster emergency management. The timely identification of high-risk areas can enable relevant departments to take preventive measures to reduce losses caused by disasters.

[0046] It should be noted that in the examples of this application, repeated introduction of the vegetation cover index feature at the rule level to reduce data redundancy may lead to coupling with the model level features, reducing the credibility of the classification results. Furthermore, surface deformation rate (millimeter-level displacement) and slope characteristics (terrain stability) have clear engineering significance and are easily associated with disaster mechanisms (such as landslide mechanics models), making them more suitable as classification indicators.

[0047] The above implementation achieves efficient and accurate automatic identification of geological hazards through the precise collection and registration of multi-source remote sensing data, combined with dynamic weight allocation and the application of deep learning models. Through feature extraction and spatiotemporal correlation analysis, it is possible to comprehensively and accurately assess geological hazard risks, generate hazard probability maps, and classify risk levels based on surface deformation rate and slope, providing a scientific basis for emergency management. This application has important application value in improving disaster prediction accuracy and enhancing emergency response capabilities.

[0048] Reference Figure 2 As an implementation of step S103, the steps of extracting multimodal features and performing spatiotemporal correlation analysis based on multi-source data sets to obtain a multimodal feature map include: Step S201 , based on the multi-source data set, extracting vegetation cover index features of optical image data, surface deformation rate of synthetic aperture radar data, and slope features of digital elevation model data respectively; The Normalized Difference Vegetation Index (NDVI) is a commonly used index that uses the difference in reflectance between the red and near-infrared bands to measure vegetation health and coverage. NDVI values ​​range from -1 to +1, with positive values ​​indicating green vegetation and negative values ​​indicating non-vegetated areas (such as water bodies and bare soil). In geological hazard identification, changes in vegetation are closely correlated with hazard occurrence. For example, a decrease in vegetation can signal soil erosion or a precursor to landslides.

[0049] Synthetic Aperture Radar (SAR) is a remote sensing technology that acquires surface information by emitting electromagnetic waves and receiving reflected signals. SAR data is robust in a wide range of meteorological conditions, including cloud cover and rain, making it uniquely advantageous for geological hazard monitoring. Interferometry techniques (such as Small Baseline Alignment Interferometry (SBAS-InSAR)) using SAR data can calculate surface deformation rates. This rate reflects ground displacement over time and is commonly used to monitor subsidence, landslides, and other geological hazards.

[0050] Digital Elevation Model (DEM) data provides three-dimensional elevation information of the earth's surface. Slope is a crucial geographic feature in geological hazard prediction. Slope indicates the degree of inclination of the ground, typically expressed as an angle in degrees. Areas with steeper slopes, particularly in mountainous areas, are more susceptible to hazards such as landslides because the steep terrain increases the risk of soil sliding. By calculating slope characteristics from DEM data, high-risk areas can be identified.

[0051] Step S202: construct a spatiotemporal correlation matrix, perform cross-modal correlation analysis on vegetation cover index characteristics, surface deformation rate, and slope characteristics through a graph neural network, and generate a multimodal feature map.

[0052] Among them, the spatiotemporal correlation matrix is ​​constructed based on the fusion of multi-source data, and captures the spatial and temporal characteristics by establishing the relationship between various data sources.

[0053] In one of the embodiments of the present application, the spatiotemporal correlation matrix is ​​input into a graph neural network, which updates node features through neighborhood node sampling and feature aggregation, and reconstructs the updated node feature matrix into a two-dimensional feature graph. The node feature matrix at each time point corresponds to a spatial distribution graph. The feature graphs of multiple time slices are temporally fused to obtain a static multimodal feature graph.

[0054] In the above implementation, the spatiotemporal correlation matrix and graph neural network can capture the complex relationship between disaster characteristics and generate multimodal feature maps, providing a reliable basis for disaster warning. The generated multimodal feature maps provide accurate spatial distribution information for the identification and prediction of geological disasters, which helps to take emergency measures in a timely manner and reduce disaster losses.

[0055] Reference Figure 3 As an implementation method of step S202, the steps of constructing a spatiotemporal correlation matrix, performing cross-modal correlation analysis on vegetation cover index characteristics, surface deformation rate, and slope characteristics through a graph neural network, and generating a multimodal feature map include: Step S301: aligning the vegetation cover index feature, the surface deformation rate, and the slope feature in time and space to generate a rasterized feature matrix aligned in time and space; Specifically, in geological hazard identification, multi-source remote sensing data (such as optical imagery, synthetic aperture radar data, and digital elevation model data) provides a wealth of information for analysis. However, the acquisition of these data often varies in time and space. Therefore, it is first necessary to perform spatiotemporal alignment on these data to ensure that the spatial locations of different data sources are consistent and their temporal resolution is uniform.

[0056] In an embodiment of the present application, by extracting the vegetation cover index (NDVI) of optical image data, the surface deformation rate of SAR data, and the slope characteristics of DEM data, and unifying these features to the same temporal resolution (e.g., 30 days) and the same spatial resolution (e.g., 10m×10m grid unit), a spatiotemporally aligned rasterized feature matrix can be generated.

[0057] Step S302: Based on the rasterized feature matrix, each grid cell is defined as a graph node; the graph node attributes include vegetation cover index characteristics, surface deformation rate and slope characteristics; When constructing the spatiotemporal correlation matrix, each grid cell is defined as a node in the graph, and the attributes of each node include the vegetation cover index characteristics, surface deformation rate and slope characteristics of the grid cell.

[0058] Step S303: Calculate the feature similarity weight and spatial adjacency weight between each graph node and perform weighted fusion to construct a spatiotemporal correlation matrix; Specifically, feature similarity weights are measured by calculating the cosine similarity between node attribute vectors, indicating the degree of feature similarity between nodes. Spatial adjacency weights are defined based on the spatial positional relationship between grid cells. If two grid cells are spatially adjacent, the spatial adjacency weight is 1, otherwise it is 0. The spatiotemporal association matrix describes the feature and spatial relationships between nodes and is a key input for graph neural network processing, effectively reflecting node similarity and spatial adjacency.

[0059] For example, in monitoring a mountainous region, two adjacent grid cells (for example, one with a slope of 30 degrees and the other with a slope of 32 degrees) will have a strong spatial adjacency weight due to their close spatial location. Furthermore, if their deformation rates are similar, their feature similarity weight will also be high. By combining these weights, the spatiotemporal correlation matrix can reflect the multi-dimensional relationships between grid cells.

[0060] Step S304: Input the spatiotemporal correlation matrix into the graph neural network, update the node implicit features through neighborhood node sampling and feature aggregation, and iterate based on the multi-layer graph neural network to generate an updated node feature matrix; Graph neural networks (GNNs) update node implicit features through neighborhood node sampling and feature aggregation. The input spatiotemporal correlation matrix provides the GNN with information about the relationships between nodes and their edges. The GNN first selects each node's neighboring nodes through neighborhood node sampling. It then aggregates the features of these neighboring nodes, typically using a weighted average to calculate the weighted mean of all neighboring node features. After feature aggregation, node features are updated using a fully connected layer and a ReLU activation function to generate new node implicit features. Through the iterative processing of multi-layer GNNs, node features are continuously updated, learning the correlations between cross-modal features and effectively improving the accuracy of geological disaster prediction.

[0061] It's understandable that the feature aggregation mechanism of graph neural networks can enhance the features of each node through neighborhood information, capturing more complex patterns. This is particularly true when influenced by spatial adjacency and feature similarity, helping to learn potential disaster risk patterns. Multi-layer GNN iterations further enhance this learning capability, making disaster predictions more accurate for each node in the graph.

[0062] Step S305, reconstructing the node feature matrix at each time point into a two-dimensional feature map according to the spatial coordinates of the grid cells; The updated node feature matrix is ​​reconstructed into a two-dimensional feature map based on the spatial coordinates of each grid cell. During the reconstruction process, the spatial structure information is retained, so that the node feature matrix at each time point corresponds to a spatial distribution map.

[0063] Step S306: Perform temporal fusion on the two-dimensional feature maps of multiple time slices to obtain a multimodal feature map.

[0064] Specifically, through time series fusion operations (such as maximum pooling), the two-dimensional feature maps of multiple time slices are fused, and the maximum response value of the feature in each time slice is retained. The resulting static multimodal feature map can effectively integrate the features of different time points and reflect more stable disaster risk characteristics.

[0065] It can be understood that time series fusion allows for a unified representation of feature information across multiple time slices by retaining the most significant features at each time point. Static multimodal feature maps can better demonstrate the potential risk of geological hazards in a region, providing rich information for final disaster prediction.

[0066] In the above implementation, based on spatiotemporal alignment, feature aggregation and time series fusion, the prediction accuracy of geological disaster risks is effectively improved, so that the potential risks of disasters can be accurately captured in different time and space dimensions, providing scientific and accurate decision support for disaster warning and emergency response.

[0067] Reference Figure 4 As an implementation of step S104, the step of dynamically allocating weights to the multimodal feature maps based on the channel attention mechanism to generate a fused feature vector includes: Step S401: determining the feature type corresponding to each channel in the multimodal feature map based on a preset channel mapping table; Specifically, when processing geological hazard data, a multimodal feature map consists of multiple channels, each representing a different geological hazard-related feature, such as deformation rate, slope, and vegetation cover index. The role of each channel in the multimodal feature map is directly related to the feature it represents. Therefore, based on a preset channel mapping table, each channel can be assigned a specific feature type identifier. These feature types help subsequent processing steps understand the data meaning of each channel and provide a basis for the dynamic weight allocation of channels.

[0068] For example, suppose a multimodal feature map has four channels, representing deformation rate, slope, vegetation cover index, and terrain roughness. Using a pre-set channel mapping table, the first channel is determined to correspond to deformation rate, the second channel to slope, the third channel to vegetation cover index, and the fourth channel to terrain roughness.

[0069] Step S402: normalize each channel in the multimodal feature map to generate a standardized multimodal feature map; Normalization is the process of unifying the numerical ranges of different features to avoid uneven results caused by excessively large or small eigenvalues ​​in subsequent calculations. In this step, normalization linearly scales the eigenvalues ​​of each channel to a value between 0 and 1. A commonly used normalization method is min-max normalization, which scales the eigenvalues ​​so that the minimum value corresponds to 0 and the maximum value corresponds to 1.

[0070] Step S403: performing global feature extraction on the standardized multimodal feature map to generate a channel-level statistical vector; Specifically, global feature extraction is used to extract global statistical information from the feature map of each channel. Common global feature extraction methods include global average pooling or global max pooling. These methods pool the entire feature map and extract a single statistic (such as the average or maximum value), forming a channel-level statistical vector that describes the overall information of each channel.

[0071] It can be understood that global feature extraction reduces the dimensionality of spatial information through pooling operations, but retains the global characteristics of the channel. This is crucial for subsequent disaster classification and feature fusion because it transforms the information of each channel into a compact and easy-to-process representation.

[0072] Step S404: input the channel-level statistical vector into the classification model and output the probability distribution of disaster types; Disaster classification is performed by inputting channel-level statistical vectors into a classification model (such as a convolutional neural network (CNN)). Based on the input feature vectors, the classification model learns the relationship between each feature and the disaster type and outputs a probability distribution for each disaster type. Disaster types can include landslides, debris flows, and collapses.

[0073] Step S405: determining the dominant disaster type according to the probability distribution and generating a corresponding disaster type code; The probability distribution output by the classification model can be used to determine the dominant disaster type. Specifically, the disaster type with the highest probability is selected as the dominant disaster type. Based on the dominant disaster type, the system generates a corresponding disaster type code for subsequent feature fusion and weighting operations.

[0074] Step S406: constructing a joint feature vector based on the disaster type coding and the standardized multimodal feature map; Among them, a joint feature vector is constructed based on the coding of the dominant disaster type and the standardized multimodal feature map. The joint feature vector combines the disaster type coding with the standardized features of each channel to provide comprehensive information for subsequent feature weighting. This step enables different types of disasters to obtain targeted feature weighting, thereby improving the adaptability and accuracy of the model.

[0075] Step S407, mapping the joint feature vector through a weight generation network to generate a channel weight vector; The weight generation network (usually composed of a fully connected layer and a sigmoid activation function) maps the joint feature vector to generate a channel weight vector, which describes the importance of each channel under a specific disaster type.

[0076] It is understandable that the weight generation network can assign appropriate weights to each channel by learning the relationship between disaster types and various feature channels, ensuring that features related to disaster types receive more attention. The dynamic allocation of weights helps the model to perform feature fusion more accurately and enhance disaster identification capabilities.

[0077] Step S408, performing channel-level weighting on the channel weight vector and the standardized multimodal feature map to generate a weighted feature map; The standardized multimodal feature maps are weighted at the channel level using a channel weight vector. Each channel's eigenvalue is multiplied by its corresponding weight to generate a weighted feature map. This weighting operation ensures that features related to the dominant disaster type receive sufficient attention in the model, highlighting the most relevant features and reducing the influence of irrelevant features, thereby improving disaster identification accuracy.

[0078] Step S409: perform spatial compression on the weighted feature map to generate a fused feature vector.

[0079] The weighted feature map is spatially compressed using global maximum pooling or other pooling operations. Pooling extracts the maximum response value from the feature map, compressing the two-dimensional feature map into a one-dimensional feature vector. This feature vector serves as input in subsequent classification and represents the disaster potential of the area.

[0080] In this implementation, dynamic weighting of multimodal feature maps effectively combines hazard type information with various geological hazard-related features (such as deformation rate, slope, and NDVI) to generate precise hazard identification feature vectors. Through dynamic weighting and feature weighting, the system can accurately identify different types of geological hazards (such as landslides, debris flows, and avalanches).

[0081] Reference Figure 5 As a further implementation of the method for automatic identification of geological hazards, a training step of a geological hazard prediction model is also included. The training step includes: Step S501: Acquire a sample data set and perform data preprocessing; wherein the sample data set includes a fused feature vector and a binary label map of a historical disaster area; Specifically, the sample dataset consists of fused feature vectors of historical disaster areas and a binary label map. The fused feature vectors are derived from multiple data sources (such as SAR data, optical imagery, and DEM data). Feature fusion is performed using a channel-wise attention mechanism to generate a high-dimensional feature representation. This feature representation effectively captures the spatial and temporal information related to disasters. The binary label map generates labels corresponding to the feature vectors by marking historical disaster areas as "1" (disaster area) or "0" (non-disaster area).

[0082] In addition, data cleaning, standardization, and normalization are key operations during data preprocessing. For example, normalization linearly scales feature values ​​to between 0 and 1 to reduce the impact of scale differences between different data sources and ensure that the model can effectively process different features.

[0083] Step S502: Divide the sample data set into a training set, a validation set, and a test set; Among them, the training set is used to train the model, the validation set is used to evaluate the performance of the model and perform tuning, and the test set is used for the final model performance evaluation.

[0084] Step S503: Input the training set into the pre-built convolutional neural network for training. The error between the predicted probability map and the true label map is calculated by the back propagation algorithm, the gradient is calculated layer by layer, and the weights and biases of the convolutional neural network are updated. The training data is fed into a convolutional neural network (CNN) for training. CNN is a deep learning model widely used in image processing tasks. Through multiple convolutional layers, CNN extracts feature information from the input image and gradually reduces the spatial dimension, focusing on higher-level semantic information. Using the backpropagation algorithm, the model calculates the error between the output predicted probability map and the true label map. The model then calculates the gradient layer by layer and updates the model's weights and biases. Convolutional neural network training can effectively learn the spatial structure in multimodal feature maps and optimize model parameters through backpropagation. Backpropagation and gradient descent algorithms enable the model to gradually adjust weights, thereby minimizing prediction error and improving prediction accuracy.

[0085] For example, assuming that the training set contains different types of geological disasters (such as landslides and mudslides), CNN will automatically identify the characteristics of different disasters during the learning process and update the network parameters through backpropagation so that the final output can accurately identify the area where the disaster occurred.

[0086] Step S504, verifying the convolutional neural network based on the validation set, evaluating the performance of the convolutional neural network and adjusting the model hyperparameters according to the validation results; After model training is complete, the model needs to be validated using a validation set. This validation set is used to evaluate the model's performance on data outside the training set. Common evaluation metrics include Intersection over Union (IoU), Precision, and Recall. During this process, if the model's performance is unsatisfactory (for example, a low IoU), hyperparameters can be adjusted to optimize the model. Common hyperparameter tuning methods include changing the learning rate, increasing the training dataset, and adding regularization.

[0087] It is understandable that by evaluating the validation set, model weaknesses can be discovered and adjusted in a timely manner, thereby improving the accuracy and robustness of the model. Hyperparameter adjustment helps avoid overfitting or underfitting of the model and improve its performance in practical applications.

[0088] Step S505: Testing the prediction capability of the adjusted convolutional neural network based on the test set to obtain a trained geological disaster prediction model.

[0089] After model training and validation, the model is finally tested using the test set. The test set consists of data not used in training and validation, and is designed to assess the model's generalization ability on unknown data. In this step, the model's output is compared with the labeled graph of the test set, and performance metrics such as the F1 score and Kappa coefficient are calculated to determine the model's predictive accuracy. Highly accurate predictions indicate that the model can successfully identify geological hazard areas, enabling effective intervention by the disaster warning system.

[0090] In the above implementation, a high-quality sample data set is constructed and trained using a convolutional neural network (CNN). The model is optimized and evaluated in conjunction with verification and testing steps, ultimately resulting in an accurate geological disaster prediction model. This ensures efficient data-driven design and powerful computing capabilities, enabling the prediction model to provide highly accurate and stable disaster prediction results in practical applications.

[0091] Reference Figure 6 As an implementation method of step S107, based on the geological hazard probability map and the potential hazard area mask, the risk level of the potential hazard area is divided based on a preset joint determination rule of surface deformation rate and slope characteristics, and the step of obtaining a risk level distribution map of the area to be tested includes: Step S601, obtaining a surface deformation rate map and a slope characteristic map based on the multimodal characteristic map; Surface deformation rate and slope characteristics are two crucial factors in the identification and assessment of geological hazards. Multimodal feature maps are derived by fusing multi-source remote sensing data. They incorporate data features from different sensors, such as surface deformation, slope, and vegetation cover. The surface deformation rate map is calculated using synthetic aperture radar (SAR) data and reflects ground displacement over time. The slope feature map, on the other hand, calculates terrain slope using digital elevation model (DEM) data and represents the degree of ground inclination. Through processing of the multimodal feature maps, two key feature maps are ultimately extracted: the surface deformation rate map and the slope feature map.

[0092] Step S602: clipping the surface deformation rate map and slope characteristic map based on the potential disaster area mask to extract the deformation rate matrix and slope matrix within the potential disaster area mask; Specifically, during the disaster risk assessment process, a potential disaster area mask is created by segmenting the potential disaster area from the geological hazard probability map using a set probability threshold (e.g., 0.5). This mask allows us to extract the area of ​​interest and avoid processing irrelevant areas. In this step, the surface deformation rate map and slope characteristic map are cropped based on the mask, extracting only the deformation rate and slope data within the mask. This reduces computational effort, improves subsequent processing efficiency, and ensures accurate analysis of the disaster area.

[0093] Step S603: spatially aligning the geological hazard probability map, deformation rate matrix, and slope matrix; Because different data sources (geological hazard probability maps, deformation rate maps, and slope maps) may originate from different sensors or devices, they often differ in temporal and spatial resolution. Therefore, spatial alignment is essential to ensure that these data can be effectively processed within the same geographic coordinate system. The goal of spatial alignment is to ensure that the coordinate systems, resolutions, and grid boundaries of different layers are consistent, enabling seamless integration and computation of these different data sources.

[0094] Step S604: Load the preset risk level classification rule configuration file, analyze the threshold parameters of the surface deformation rate and slope, and obtain the judgment condition function; During the risk classification process, pre-set risk classification rules are used to determine the criteria for different risk levels. These rules include threshold parameters for surface deformation rate and slope, which are used to classify high-risk, medium-risk, and low-risk areas. A rule configuration file is a format that stores these threshold conditions, typically a text file. By parsing this configuration file, the program can automatically load the threshold parameters and convert them into executable numerical judgment conditions.

[0095] Step S605, traversing each grid cell of the geological hazard probability map, and determining that the area with a probability value greater than a preset probability threshold is a potential disaster area; In geological hazard risk assessment, a geological hazard probability map provides the probability of a disaster occurring in each region. By setting a probability threshold (e.g., 0.5), regions within the probability map can be classified as potential disaster areas or non-disaster areas. Regions with probability values ​​greater than or equal to the threshold are considered potential disaster areas, while those with probability values ​​less than or equal to the threshold are considered non-disaster areas. By traversing each grid cell, disaster areas can be gradually screened and marked.

[0096] Step S606: Obtaining surface deformation rate data and slope characteristic data of the potential disaster area and performing a determination based on a determination condition function to classify the risk level of the potential disaster area and obtain a risk level coding matrix; Once potential disaster areas are identified, surface deformation rate and slope characteristic data for these areas are obtained, and the risk level of each grid cell can be classified according to the previously loaded judgment condition function (such as threshold rules for deformation rate and slope). Based on different combinations of deformation rate and slope, the area will be classified as high risk, medium risk, or low risk. For example, for a potential disaster area, assuming its deformation rate is 35mm / year and its slope is 28°, which meets the criteria for a high risk area, the area will be marked as high risk.

[0097] Step S607: generating a risk level distribution map of the area to be tested according to the risk level coding matrix.

[0098] Once the risk classification is complete, the risk level coding matrix can be converted into a final risk level distribution map. This distribution map visualizes the risk level (high, medium, and low) for each grid cell as a different color based on the original geographic coordinate system, allowing users to intuitively understand the disaster risk level of each area. The generated distribution map not only includes the risk level of each area but also contains necessary geographic metadata (such as coordinate system and spatial resolution) to ensure data availability and accuracy.

[0099] In the above implementation, a joint determination method based on deformation rate and slope is used to categorize geological hazard areas into risk levels. Through spatial alignment, rule loading, and pixel-by-pixel risk assessment, this method accurately and efficiently categorizes potential hazard areas according to predefined rules, ultimately generating a visual risk level distribution map. This provides a reliable risk assessment tool for geological hazard monitoring systems, supporting disaster warning and decision-making, thereby improving the timeliness and accuracy of disaster response.

[0100] The embodiment of the present application also discloses a geological disaster automatic identification system based on multi-source remote sensing data.

[0101] An automatic geological disaster identification system based on multi-source remote sensing data, the identification system includes: An acquisition module is used to acquire multi-source remote sensing data of the area to be measured; wherein the multi-source remote sensing data includes optical image data, synthetic aperture radar data and digital elevation model data; The registration module is used to perform cross-modal registration on multi-source remote sensing data and generate a registered multi-source dataset; A multimodal feature map generation module is used to extract multimodal features from multi-source datasets and perform spatiotemporal correlation analysis to generate a multimodal feature map. The multimodal features include vegetation cover index features from optical image data, surface deformation rate from synthetic aperture radar data, and slope features from digital elevation model data. Feature fusion module, which is used to dynamically assign weights to multimodal feature maps based on the channel attention mechanism to generate a fused feature vector; The geological disaster prediction module is used to input the fused feature vector into the pre-trained geological disaster prediction model and output a geological disaster probability map; A binary segmentation module is used to perform binary segmentation on the geological disaster probability map according to a preset threshold value to obtain a potential disaster area mask; The risk level classification module is used to classify the potential disaster area into risk levels based on the geological disaster probability map and the potential disaster area mask, based on the preset joint judgment rules of surface deformation rate and slope characteristics, and obtain the risk level distribution map of the area to be tested.

[0102] As an implementation of the multimodal feature map generation module, it specifically includes: A multimodal feature extraction unit is used to extract vegetation cover index features from optical image data, surface deformation rate from synthetic aperture radar data, and slope features from digital elevation model data based on multi-source data sets; The cross-modal correlation analysis unit is used to construct a spatiotemporal correlation matrix, perform cross-modal correlation analysis on vegetation cover index characteristics, surface deformation rate and slope characteristics through graph neural network, and generate a multimodal feature map.

[0103] The embodiment of the present application is a geological disaster automatic identification system based on multi-source remote sensing data that can implement any of the above-mentioned automatic identification methods, and the specific working processes of each module in the automatic identification system can refer to the corresponding processes in the above-mentioned method embodiments.

[0104] In the several embodiments provided in this application, it should be understood that the provided methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for example, the division of a module is merely a logical functional division, and in actual implementation, other division methods may be used, such as combining or integrating multiple modules into another system, or ignoring or not implementing certain features.

[0105] The embodiment of the present application also discloses a computer device.

[0106] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, a method for automatically identifying geological hazards based on multi-source remote sensing data as described above is implemented.

[0107] The embodiment of the present application also discloses a computer-readable storage medium.

[0108] A computer-readable storage medium stores a computer program that can be loaded by a processor and executed by any one of the above-mentioned methods for automatic identification of geological hazards based on multi-source remote sensing data.

[0109] Among them, computer-readable storage media can be any tangible medium that contains or stores a program that can be used by or in combination with an instruction execution system, apparatus or device; the program code contained on the computer-readable medium can be transmitted using any appropriate medium, including but not limited to wireless, wire, optical cable, RF, etc., or any suitable combination of the above.

[0110] It should be noted that, in the above embodiments, the description of each embodiment has different emphases. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0111] The above are all preferred embodiments of the present application and are not intended to limit the scope of protection of this application. Unless otherwise specified, any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features. In other words, unless otherwise specified, each feature is merely an example of a series of equivalent or similar features.

Claims

1. A method for automatic identification of geological hazards based on multi-source remote sensing data, characterized in that: The identification method comprises: Collecting multi-source remote sensing data of the area to be measured; wherein the multi-source remote sensing data includes optical image data, synthetic aperture radar data and digital elevation model data; Performing cross-modal registration on the multi-source remote sensing data to generate a registered multi-source dataset; Extracting multimodal features and performing spatiotemporal correlation analysis based on the multi-source data set to obtain a multimodal feature map; wherein the multimodal features include vegetation cover index features of the optical image data, surface deformation rate of the synthetic aperture radar data, and slope features of the digital elevation model data; Dynamically assigning weights to the multimodal feature maps based on a channel attention mechanism to generate a fused feature vector; Inputting the fused feature vector into a pre-trained geological disaster prediction model and outputting a geological disaster probability map; Binarize and segment the geological disaster probability map according to a preset threshold value to obtain a potential disaster area mask; According to the geological hazard probability map and the potential hazard area mask, based on the preset joint judgment rule of the surface deformation rate and slope characteristics, the potential hazard area is divided into risk levels to obtain a risk level distribution map of the area to be tested.

2. The method for automatic identification of geological hazards based on multi-source remote sensing data according to claim 1, characterized in that: The steps of extracting multimodal features and performing spatiotemporal correlation analysis based on the multi-source data set to obtain a multimodal feature map include: Based on the multi-source data set, extracting vegetation cover index characteristics of the optical image data, surface deformation rate of the synthetic aperture radar data, and slope characteristics of the digital elevation model data respectively; A spatiotemporal correlation matrix is ​​constructed, and a cross-modal correlation analysis is performed on the vegetation cover index characteristics, surface deformation rate and slope characteristics through a graph neural network to generate a multimodal feature map.

3. The method for automatic identification of geological hazards based on multi-source remote sensing data according to claim 2, characterized in that: The steps of constructing a spatiotemporal correlation matrix, performing cross-modal correlation analysis on the vegetation cover index characteristics, surface deformation rate, and slope characteristics through a graph neural network, and generating a multimodal feature map include: Performing spatiotemporal alignment on the vegetation cover index feature, the surface deformation rate, and the slope feature to generate a spatiotemporal aligned rasterized feature matrix; Based on the grid feature matrix, each grid unit is defined as a graph node; the graph node attributes include vegetation cover index characteristics, surface deformation rate and slope characteristics; Calculate the feature similarity weight and spatial adjacency weight between each graph node and perform weighted fusion to construct a spatiotemporal correlation matrix; Input the spatiotemporal correlation matrix into the graph neural network, update the node implicit features through neighborhood node sampling and feature aggregation, and iterate based on the multi-layer graph neural network to generate an updated node feature matrix; Reconstructing the node feature matrix at each time point into a two-dimensional feature map according to the spatial coordinates of the grid cells; The two-dimensional feature maps of multiple time slices are temporally fused to obtain a multimodal feature map.

4. The method for automatic identification of geological hazards based on multi-source remote sensing data according to claim 1, characterized in that: The steps of dynamically allocating weights to the multimodal feature map based on the channel attention mechanism to generate a fused feature vector include: Determining, based on a preset channel mapping table, a feature type corresponding to each channel in the multimodal feature map; Normalizing each channel in the multimodal feature map to generate a standardized multimodal feature map; Performing global feature extraction on the standardized multimodal feature map to generate a channel-level statistical vector; Inputting the channel-level statistical vector into a classification model and outputting a probability distribution of disaster types; Determine the dominant disaster type according to the probability distribution and generate a corresponding disaster type code; Constructing a joint feature vector based on the disaster type coding and the standardized multimodal feature map; Mapping the joint feature vector through a weight generation network to generate a channel weight vector; Performing channel-level weighting on the channel weight vector and the standardized multimodal feature map to generate a weighted feature map; The weighted feature map is spatially compressed to generate a fused feature vector.

5. The method for automatic identification of geological hazards based on multi-source remote sensing data according to claim 1, characterized in that: The method further includes a training step of the geological disaster prediction model, wherein the training step includes: Acquire a sample data set and perform data preprocessing; wherein the sample data set includes a fused feature vector and a binary label map of a historical disaster area; Dividing the sample data set into a training set, a validation set, and a test set; The training set is input into a pre-built convolutional neural network for training, the error between the predicted probability map and the true label map is calculated by the back propagation algorithm, the gradient is calculated layer by layer, and the weights and biases of the convolutional neural network are updated; Validating the convolutional neural network based on the validation set, evaluating the performance of the convolutional neural network and adjusting model hyperparameters according to the validation results; The prediction ability of the adjusted convolutional neural network is tested based on the test set to obtain the trained geological disaster prediction model.

6. A method for automatic identification of geological hazards based on multi-source remote sensing data according to any one of claims 1 to 5, characterized in that: The steps of classifying the potential disaster areas into risk levels based on the geological hazard probability map and the potential disaster area mask and the preset joint determination rule of the surface deformation rate and slope characteristics to obtain the risk level distribution map of the area to be tested include: The surface deformation rate map and slope characteristic map are obtained based on the multimodal characteristic map; Clipping the surface deformation rate map and slope characteristic map based on the potential disaster area mask to extract the deformation rate matrix and slope matrix within the potential disaster area mask; spatially aligning the geological hazard probability map, deformation rate matrix, and slope matrix; Load the preset risk level classification rule configuration file, analyze the threshold parameters of surface deformation rate and slope, and obtain the judgment condition function; Traversing each grid cell of the geological hazard probability map, determining areas where the probability value is greater than a preset probability threshold as potential hazard areas; Obtaining surface deformation rate data and slope characteristic data of the potential disaster area and performing a determination based on the determination condition function, dividing the risk level of the potential disaster area into risk levels, and obtaining a risk level coding matrix; A risk level distribution map of the area to be tested is generated according to the risk level coding matrix.

7. A geological disaster automatic identification system based on multi-source remote sensing data, characterized in that: The identification system comprises: An acquisition module is used to acquire multi-source remote sensing data of the area to be measured; wherein the multi-source remote sensing data includes optical image data, synthetic aperture radar data and digital elevation model data; A registration module, configured to perform cross-modal registration on the multi-source remote sensing data to generate a registered multi-source dataset; A multimodal feature map generation module is configured to extract multimodal features and perform spatiotemporal correlation analysis based on the multi-source data set to obtain a multimodal feature map; wherein the multimodal features include vegetation cover index features of the optical image data, surface deformation rate of the synthetic aperture radar data, and slope features of the digital elevation model data; A feature fusion module, configured to dynamically assign weights to the multimodal feature maps based on a channel attention mechanism to generate a fused feature vector; A geological disaster prediction module is used to input the fused feature vector into a pre-trained geological disaster prediction model and output a geological disaster probability map; A binary segmentation module is used to perform binary segmentation on the geological disaster probability map according to a preset threshold value to obtain a potential disaster area mask; The risk level classification module is used to classify the potential disaster area into risk levels based on the geological disaster probability map and the potential disaster area mask, based on the preset joint judgment rules of the surface deformation rate and slope characteristics, and obtain a risk level distribution map of the area to be tested.

8. The automatic geological disaster identification system based on multi-source remote sensing data according to claim 7 is characterized in that: The multimodal feature map generation module includes: a multimodal feature extraction unit, configured to extract, based on the multi-source data set, vegetation cover index features of the optical image data, surface deformation rate of the synthetic aperture radar data, and slope features of the digital elevation model data; The cross-modal correlation analysis unit is used to construct a spatiotemporal correlation matrix, perform cross-modal correlation analysis on the vegetation cover index characteristics, surface deformation rate and slope characteristics through a graph neural network, and generate a multimodal feature map.

9. A computer device, characterized in that: The method comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method according to any one of claims 1 to 6 when executing the program.

10. A computer-readable storage medium, characterized in that: A computer program is stored which can be loaded by a processor and execute the method according to any one of claims 1 to 6.

Images