Landslide disaster multi-element correlation analysis method, device and equipment and storage medium

Abstract

The application provides a landslide disaster multi-element correlation analysis method, device and equipment and a storage medium, relates to the technical field of geological disaster analysis, and comprises the following steps: constructing a landslide knowledge base through multi-source landslide disaster data; data channel information is used to reflect the objective characteristics of a landslide event, and knowledge channel information is used to reflect the semantic logic of the landslide event; obtaining a landslide heterogeneous graph according to the data channel information and the knowledge channel information, wherein the landslide heterogeneous graph comprises nodes and edges; embedding and aligning multi-modal data of the landslide heterogeneous graph by using a data source label to obtain low-dimensional feature vectors; obtaining an updated landslide heterogeneous graph according to all the low-dimensional feature vectors based on a cross-modal attention mechanism; and performing correlation analysis on the updated landslide heterogeneous graph based on a GHT heterogeneous graph neural network correlation analysis algorithm to obtain a target landslide heterogeneous graph. The application realizes improved representation capability of landslide events and potential correlation relationships.

Description

Technical Field

[0001] This invention relates to the field of geological disaster analysis technology, and more specifically, to a method, apparatus, equipment, and storage medium for multi-factor correlation analysis of landslide disasters. Background Technology

[0002] Analyzing landslide disasters can capture early warning signals, effectively guiding the accuracy of landslide prediction. Traditional landslide disaster analysis methods mainly rely on single data sources for modeling. However, due to the high complexity and systematic nature of landslide formation mechanisms, a single data source cannot accurately depict landslide patterns. Compared to single-data-source modeling, studying and analyzing the potential correlations between multi-source heterogeneous information on landslide events can significantly improve landslide prediction accuracy. However, existing methods generally lack effective integration of expert knowledge, causal reasoning chains, and complex evolutionary mechanisms, resulting in significant deficiencies in the accuracy and interpretability of disaster correlation modeling. Summary of the Invention

[0003] The present invention aims to solve at least one of the above-mentioned problems.

[0004] To address the aforementioned problems, this invention provides a method, apparatus, equipment, and storage medium for multi-factor correlation analysis of landslide disasters.

[0005] In a first aspect, the present invention provides a method for multi-factor correlation analysis of landslide disasters, including:

[0006] A landslide knowledge base is constructed using multi-source landslide disaster data. The landslide knowledge base includes data channel information and knowledge channel information. The data channel information is used to reflect the objective characteristics of landslide events, and the knowledge channel information is used to reflect the semantic logic of landslide events.

[0007] A landslide heterogeneous graph is obtained based on the data channel information and the knowledge channel information, wherein the landslide heterogeneous graph includes nodes and edges;

[0008] The multimodal data of the landslide heterogeneous map are embedded and aligned using data source labels to obtain low-dimensional feature vectors;

[0009] Based on the cross-modal attention mechanism, an updated landslide heterogeneous map is obtained from all the aforementioned low-dimensional feature vectors;

[0010] Based on the GHT heterogeneous graph neural network correlation analysis algorithm, the updated landslide heterogeneous graph is subjected to correlation analysis to obtain the target landslide heterogeneous graph, wherein the target landslide heterogeneous graph is used to reflect the correlation of multiple factors of landslide disaster.

[0011] Optionally, the data source label includes a data channel label and a knowledge channel label. The step of embedding and aligning the multimodal data of the landslide heterogeneous map using the data source label to obtain a low-dimensional feature vector includes:

[0012] The multimodal data of the landslide heterogeneous map is encoded and mapped using the data channel labels and the knowledge channel labels to obtain embedded feature vectors;

[0013] Based on the cross-modal contrastive learning method, modal alignment is performed on the embedded feature vector to obtain the low-dimensional feature vector.

[0014] Optionally, the data channel labels include image data labels and time-series data labels, and the embedded feature vectors include image embedding vectors and time-series embedding vectors; the process of encoding and mapping the multimodal data of the landslide heterogeneous map using the data channel labels and the knowledge channel labels to obtain the embedded feature vectors includes:

[0015] When the data channel label is the image data label, the multimodal data of the landslide heterogeneous map is input into a pre-trained ResNet visual model for feature extraction to obtain the image embedding vector. The pre-trained ResNet visual model is used to process image data using a deep convolutional neural network.

[0016] When the data channel label is the time series data label, the multimodal data of the landslide heterogeneous map is input into the LSTM model for learning to obtain the time series embedding vector, wherein the LSTM model is used to capture the long-term dependencies of the time series data.

[0017] Optionally, the knowledge channel label includes text data labels and expert knowledge rule labels, and the embedded feature vector includes text embedding vectors and semantic embedding vectors; the step of encoding and mapping the multimodal data of the landslide heterogeneous map using the data channel label and the knowledge channel label to obtain the embedded feature vector includes:

[0018] When the knowledge channel label is the text data label, the multimodal data of the landslide heterogeneous map is input into the pre-trained RoBERTa language model for encoding to obtain the text embedding vector, wherein the pre-trained RoBERTa language model is used to perform lexicalization on the text data;

[0019] When the knowledge channel label is the expert knowledge rule label, the TransE knowledge graph embedding method is used to map the multimodal data of the landslide heterogeneous graph to a low-dimensional space to obtain the semantic embedding vector.

[0020] Optionally, the low-dimensional feature vectors include text low-dimensional vectors, image low-dimensional vectors, and time-series low-dimensional vectors; obtaining the updated landslide heterogeneous map based on all the low-dimensional feature vectors includes:

[0021] The text low-dimensional vector, the image low-dimensional vector, and the temporal low-dimensional vector corresponding to each node are weighted and fused to obtain a single embedding vector.

[0022] The updated landslide heterogeneity map is obtained based on the single embedding vector;

[0023] The node update process includes:

[0024] ,

[0025] in, This is the weighted fusion result of low-dimensional vector edges. The text is a low-dimensional vector. The low-dimensional vector of the image, The time series low-dimensional vector, , and These are the weight parameters.

[0026] Optionally, the step of performing correlation analysis on the updated landslide heterogeneity map to obtain the target landslide heterogeneity map includes:

[0027] Multi-head heterogeneous attention calculation is performed on each node of the updated landslide heterogeneous graph to obtain the corresponding attention weights;

[0028] Based on the channel weight control mechanism, the nodes of the updated landslide heterogeneous map are aggregated according to all the attention weights to obtain the target landslide heterogeneous map.

[0029] The channel weight control mechanism includes:

[0030] ,

[0031] in, Let the output of node i be represented. and The dynamic contribution rates of the data channel information and the knowledge channel information are respectively. Let i be the embedding vector of the data channel information of node i. Let be the embedding vector of node i in the knowledge channel information, softmax be the normalization function, and U be the linear transformation matrix.

[0032] Optionally, the nodes include landslide event nodes, cause nodes, geological condition nodes, time nodes, spatial nodes, and text nodes.

[0033] Secondly, the present invention provides a multi-factor correlation analysis device for landslide disasters, comprising:

[0034] A landslide knowledge base construction module is used to construct a landslide knowledge base using multi-source landslide disaster data. The landslide knowledge base includes data channel information and knowledge channel information. The data channel information is used to reflect the objective characteristics of landslide events, and the knowledge channel information is used to reflect the semantic logic of landslide events.

[0035] A landslide heterogeneous map acquisition module is used to obtain a landslide heterogeneous map based on the data channel information and the knowledge channel information, wherein the landslide heterogeneous map includes nodes and edges;

[0036] Low-dimensional feature vector acquisition is used to embed and align the multimodal data of the landslide heterogeneous map using data source labels to obtain low-dimensional feature vectors;

[0037] A cross-modal attention mechanism module is used to obtain an updated landslide heterogeneous map based on all the low-dimensional feature vectors, using a cross-modal attention mechanism.

[0038] The correlation analysis module is used to perform correlation analysis on the updated landslide heterogeneous map based on the GHT heterogeneous graph neural network correlation analysis algorithm to obtain the target landslide heterogeneous map, wherein the target landslide heterogeneous map is used to reflect the correlation of multiple factors of landslide disaster.

[0039] Thirdly, the present invention provides an electronic device, including a memory and a processor;

[0040] The memory is used to store computer programs;

[0041] The processor is configured to implement the landslide disaster multi-factor correlation analysis method as described in the first aspect when executing the computer program.

[0042] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the landslide disaster multi-factor correlation analysis method as described in the first aspect.

[0043] The beneficial effects of the landslide disaster multi-element correlation analysis method, apparatus, equipment, and storage medium of the present invention are as follows: A landslide knowledge base is constructed using multi-source landslide disaster data, which is divided into data channel information and knowledge channel information. Data channel information reflects the objective characteristics of landslide events, while knowledge channel information reflects the semantic logic of landslide events, improving the accuracy of landslide disaster correlation modeling. Different types of entities are abstracted as nodes, and their interactions or dependencies are represented as edges, resulting in a landslide heterogeneous graph that characterizes the complex correlation structure among multiple elements in the landslide disaster system. Data source labels are used to embed and align the multimodal data of the landslide heterogeneous graph, obtaining low-dimensional feature vectors, enhancing the consistency and comparability of feature expressions, and laying the foundation for subsequent fusion analysis. Based on a cross-modal attention mechanism, an updated landslide heterogeneous graph is obtained based on all low-dimensional feature vectors, achieving embedding alignment and fusion, ensuring the reasonable expression of various types of information. Based on the GHT heterogeneous graph neural network correlation analysis algorithm, in-depth correlation analysis is performed on the updated heterogeneous graph, thereby improving the representation ability of landslide events and their potential correlations. Attached Figure Description

[0044] Figure 1 This is a flowchart illustrating a multi-factor correlation analysis method for landslide disasters according to an embodiment of the present invention;

[0045] Figure 2 This is a schematic diagram of the structure of a landslide disaster multi-factor correlation analysis device according to an embodiment of the present invention;

[0046] Figure 3 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0047] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Although some embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present invention. It should be understood that the accompanying drawings and embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0048] It should be understood that the various steps described in the method embodiments of the present invention may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present invention is not limited in this respect.

[0049] The term "comprising" and its variations as used herein are open-ended, meaning "including but not limited to"; the term "based on" means "at least partially based on"; the term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments"; and the term "optionally" means "optional embodiments". Definitions of other terms will be given in the description below. It should be noted that the concepts of "first," "second," etc., mentioned in this invention are used only to distinguish different devices, modules, or units, and are not intended to limit the order of functions performed by these devices, modules, or units or their interdependencies.

[0050] It should be noted that the terms "a" and "a plurality of" used in this invention are illustrative rather than restrictive. Those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0051] like Figure 1 As shown in the figure, an embodiment of the present invention provides a multi-factor correlation analysis method for landslide disasters, including:

[0052] Step 110: Construct a landslide knowledge base using multi-source landslide disaster data. The landslide knowledge base includes data channel information and knowledge channel information. The data channel information is used to reflect the objective characteristics of landslide events, and the knowledge channel information is used to reflect the semantic logic of landslide events.

[0053] Specifically, the multi-source landslide disaster data includes remote sensing imagery, digital elevation models (DEMs), geological structure maps, soil types, hydrological and meteorological data, land use types, NDVI vegetation cover index data, and road and mine distribution. It also includes prior knowledge data, such as expert experience rules, landslide classification standards, and historical landslide event data. First, the multi-source landslide disaster data is preprocessed. Based on the preprocessed data, a landslide knowledge base is constructed to achieve structured integration and dual-channel organization of disaster information. The landslide knowledge base uses landslide disaster points as basic units. For each landslide instance, its attribute information is uniformly organized and divided into two categories: data channel information and knowledge channel information, to achieve structured input for heterogeneous graph modeling. The data channel information primarily originates from objective observation and monitoring data, including the spatial location (latitude and longitude coordinates) of the landslide point, the time of occurrence, the landslide type (e.g., shallow landslide, deep landslide), the landslide volume, the landslide elevation difference, the sliding distance, the slope, the aspect, the surface curvature, the geological structure type, the soil type, the vegetation cover, the previous rainfall, the cumulative rainfall, and other environmental and geomorphological factors. This data channel information is obtained through remote sensing interpretation, digital elevation model analysis, hydrological and meteorological data statistics, and geological map extraction, possessing clear spatiotemporal coordinate attributes and quantitative characteristics. The knowledge channel information originates from manual annotation, expert rule extraction, and literature summarization results. It is structured using methods such as rule extraction, entity relationship annotation, and map modeling, including the type of triggering factors (e.g., heavy rainfall, seismic disturbance, artificial engineering excavation), the landslide evolution stage (e.g., incubation period, critical period, instability period), and potential causal relationships (e.g., "lithological combination A + slope greater than..."). (e.g., "corresponding to high risk"), links to similar historical events, and expert experience tags (e.g., "typical rainfall-induced landslide").

[0054] Step 120: Obtain a landslide heterogeneous graph based on the data channel information and the knowledge channel information, wherein the landslide heterogeneous graph includes nodes and edges.

[0055] Specifically, each edge in the landslide heterogeneous graph is labeled as either a data edge or a knowledge edge based on its source, and is accompanied by a weight value and a confidence parameter to reflect the strength and determinism of the relationship. Data import of data channel information and knowledge channel information adopts a pattern-driven automated mapping process. Based on predefined node types and relationship templates, the system automatically extracts corresponding entities and their relationships from databases (including meteorological, geological, remote sensing, and expert knowledge bases, etc.), and generates node and edge records through scripted ETL (Extract-Transform-Load) rules. This achieves automated import and synchronous updates of data and knowledge channels, thus forming a heterogeneous graph construction mechanism driven by structural constraints and data content. Storage can be implemented based on graph databases such as Neo4j, using a label-relationship model to store node and edge types, and ensuring traceable and scalable heterogeneous graph management through indexes (uid, spatial index, timestamp) and constraints (uniqueness, required attributes).

[0056] Step 130: Embed and align the multimodal data of the landslide heterogeneous map using data source labels to obtain a low-dimensional feature vector.

[0057] Specifically, the multimodal data in the landslide heterogeneous map is embedded and aligned using data source labels, image features are extracted using an image pre-trained network, time series such as rainfall are encoded using a time series modeling method, text is encoded using a pre-trained language model, and causal links and expert rules are mapped using a knowledge graph embedding method, thus transforming the multimodal data into corresponding low-dimensional feature vectors.

[0058] Step 140: Based on the cross-modal attention mechanism, obtain the updated landslide heterogeneous map according to all the said low-dimensional feature vectors.

[0059] Specifically, a cross-modal alignment mechanism is proposed, based on contrastive learning InfoNCE loss and cross-modal attention mechanism, to ensure the consistency of the representation of the same landslide event in different modal feature spaces.

[0060] Step 150: Based on the GHT heterogeneous graph neural network correlation analysis algorithm, perform correlation analysis on the updated landslide heterogeneous graph to obtain the target landslide heterogeneous graph, wherein the target landslide heterogeneous graph is used to reflect the correlation of multiple factors of landslide disaster.

[0061] Specifically, an improved heterogeneous graph neural network (DC-HGT model) based on dual-channel collaborative driving is proposed for correlation analysis. A channel-aware node representation update mechanism is introduced to separately perceive data channel embedding and knowledge channel embedding. Furthermore, a channel weight adjustment module is incorporated during the aggregation process to dynamically adjust the contribution of different channels to the node representation. This outputs multi-perspective representations of landslide events and inference results regarding potential relationships, thus achieving correlation analysis of landslide disasters.

[0062] In this embodiment, a landslide knowledge base is constructed using multi-source landslide disaster data, divided into data channel information and knowledge channel information. Data channel information reflects the objective characteristics of landslide events, while knowledge channel information reflects their semantic logic, improving the accuracy of landslide disaster correlation modeling. Different types of entities are abstracted as nodes, and their interactions or dependencies are represented as edges, resulting in a landslide heterogeneous graph that characterizes the complex correlation structure among multiple elements in the landslide disaster system. Data source labels are used to embed and align the multimodal data of the landslide heterogeneous graph, obtaining low-dimensional feature vectors. This enhances the consistency and comparability of feature representation, laying the foundation for subsequent fusion analysis. Based on a cross-modal attention mechanism, an updated landslide heterogeneous graph is obtained from all low-dimensional feature vectors, achieving embedding alignment and fusion to ensure the reasonable expression of various types of information. Based on the GHT heterogeneous graph neural network correlation analysis algorithm, in-depth correlation analysis is performed on the updated heterogeneous graph, thereby improving the representation ability of landslide events and their potential correlations.

[0063] Optionally, the data source label includes a data channel label and a knowledge channel label. The step of embedding and aligning the multimodal data of the landslide heterogeneous map using the data source label to obtain a low-dimensional feature vector includes:

[0064] The multimodal data of the landslide heterogeneous map is encoded and mapped using the data channel labels and the knowledge channel labels to obtain embedded feature vectors;

[0065] Based on the cross-modal contrastive learning method, modal alignment is performed on the embedded feature vector to obtain the low-dimensional feature vector.

[0066] Specifically, by differentially encoding and embedding information from multimodal sources in the data and knowledge channels of heterogeneous graphs, and combining this with an adaptive alignment mechanism, the representation spaces between modalities are fused and unified, thus providing a unified low-dimensional vector representation for subsequent graph correlation analysis of landslide events. In the multimodal feature embedding design, different embedding techniques are employed to encode and map the multimodal feature representations present in the data and knowledge channels, ensuring that different modal features can be compared, aligned, and collaboratively learned in the same space. To achieve collaborative expression of multimodal feature embedding representations in the feature space, a modality alignment mechanism is further introduced to achieve alignment between different modalities. The InfoNCE loss method, a cross-modal contrastive learning method, is used to calculate the similarity between different modalities, ensuring that the embedding vectors of the same event are close in different modal spaces, while maintaining the discriminative power between vectors of different events. The objective is to maximize the cosine similarity of the embedding vectors of text, image, and time-series data of the same event, thereby minimizing the distance between different modal embedding vectors of the same event in the shared space and maximizing the distance between different events, thus achieving the alignment of multimodal embedding vectors in the shared space.

[0067] In some more specific embodiments, each landslide record can be formatted as follows:

[0068] [Landslide Number] LS202308001;

[0069] [Data Channel];

[0070] [Date] June 5, 2021;

[0071] [Location] Xinzhou City, Shanxi Province;

[0072] [Location] E, N;

[0073] [Type] Shallow landslide;

[0074] [volume] ;

[0075] [slope] ;

[0076] [Slope Aspect] SSE;

[0077] [Geological type] Mudstone interbedded with sandstone;

[0078] [Cumulative rainfall over three days] 125 mm;

[0079] [Knowledge Channel];

[0080] [Triggering Factors] Continuous heavy rainfall + engineering disturbance;

[0081] [Empirical Rule Matching] Weak rock formations combined with continuous rainfall correspond to a high risk of instability;

[0082] [Expert Note] The area has well-developed soft rock layers, making it prone to rainfall-induced landslides.

[0083] Each landslide disaster record forms a dual-channel composite expression in the knowledge base, consisting of data channel information and knowledge channel information. This enables comprehensive, multi-level, and structured modeling of landslide events, providing complete information support for subsequent heterogeneous graph node definition, edge construction, and semantic alignment.

[0084] In this optional embodiment, different embedding techniques are used to encode and map data for each modality, ensuring that data from different sources can be compared and collaboratively learned in the same space. Introducing a modality alignment mechanism and using a cross-modal contrastive learning method effectively aligns information from different modalities into a common feature space, which helps improve the accuracy of subsequent steps.

[0085] Optionally, the data channel labels include image data labels and time-series data labels, and the embedded feature vectors include image embedding vectors and time-series embedding vectors; the process of encoding and mapping the multimodal data of the landslide heterogeneous map using the data channel labels and the knowledge channel labels to obtain the embedded feature vectors includes:

[0086] When the data channel label is the image data label, the multimodal data of the landslide heterogeneous map is input into a pre-trained ResNet visual model for feature extraction to obtain the image embedding vector. The pre-trained ResNet visual model is used to process image data using a deep convolutional neural network.

[0087] When the data channel label is the time series data label, the multimodal data of the landslide heterogeneous map is input into the LSTM model for learning to obtain the time series embedding vector, wherein the LSTM model is used to capture the long-term dependencies of the time series data.

[0088] Specifically, for image data such as terrain features and remote sensing images near landslide disasters, a ResNet pre-trained visual model is used for feature extraction. This method utilizes a deep convolutional neural network to process landslide disaster-related images, mapping image features into low-dimensional vectors while preserving spatial structure information in the images. For time-series data such as rainfall, temperature, and wind speed, an LSTM model is used to learn their inherent temporal characteristics, capturing long-term dependencies in the time-series data and generating temporal embedding vectors for each time point, which can reflect the temporal dynamics of landslide events.

[0089] Optionally, the knowledge channel label includes text data labels and expert knowledge rule labels, and the embedded feature vector includes text embedding vectors and semantic embedding vectors; the step of encoding and mapping the multimodal data of the landslide heterogeneous map using the data channel label and the knowledge channel label to obtain the embedded feature vector includes:

[0090] When the knowledge channel label is the text data label, the multimodal data of the landslide heterogeneous map is input into the pre-trained RoBERTa language model for encoding to obtain the text embedding vector, wherein the pre-trained RoBERTa language model is used to perform lexicalization on the text data;

[0091] When the knowledge channel label is the expert knowledge rule label, the TransE knowledge graph embedding method is used to map the multimodal data of the landslide heterogeneous graph to a low-dimensional space to obtain the semantic embedding vector.

[0092] Specifically, for textual data such as descriptions of landslide events, causal analysis reports, expert tags, and textual information of similar cases, a RoBERTa pre-trained language model is used to encode the textual information related to landslide disasters. This model performs tokenization on the textual data, inputting it into the pre-trained RoBERTa model to extract [CLS] vectors or perform average pooling to obtain a fixed-dimensional text embedding vector. This embedding vector can fully capture the semantic information in the textual data, and is used for comparative learning and alignment with data from other modalities. For expert knowledge and rules in the knowledge channel, such as causal rules, expert tags, causal links, and similar case links, due to their strong semantic dependencies and unstructured prior features, the TransE knowledge graph embedding method is used to map the causal chains related to landslides and the causal information labeled by experts into a low-dimensional space, preserving the relational information in the knowledge graph and ensuring the semantic differences and logical consistency between different entities.

[0093] In this optional embodiment, a multimodal embedding and alignment method is proposed to extract the structural features and semantic information contained in the multi-type nodes and edges in the heterogeneous map of landslide disasters, and to realize the unified representation and collaborative learning of data and knowledge through dual channels, thereby achieving deep low-dimensional representation of multi-source and multimodal data.

[0094] Optionally, the low-dimensional feature vectors include text low-dimensional vectors, image low-dimensional vectors, and time-series low-dimensional vectors; obtaining the updated landslide heterogeneous map based on all the low-dimensional feature vectors includes:

[0095] The text low-dimensional vector, the image low-dimensional vector, and the temporal low-dimensional vector corresponding to each node are weighted and fused to obtain a single embedding vector.

[0096] The updated landslide heterogeneity map is obtained based on the single embedding vector;

[0097] The node update process includes:

[0098] ,

[0099] in, This is the weighted fusion result of low-dimensional vector edges. The text is a low-dimensional vector. The low-dimensional vector of the image, The time series low-dimensional vector, , and These are the weight parameters.

[0100] Specifically, after the alignment process is complete, a cross-modal attention mechanism is introduced to perform weighted fusion of the embedding vectors from each modality. The specific steps are as follows: Weighted Fusion. The embedding vectors from each modality are weighted using a gradient descent weighting function. Weight parameters are set. , and These are used to control the importance of embedding vectors for text, image, and time-series data, respectively. Performed at the node level, this unifies the multimodal features (text, image, time-series) of the same landslide event node into a single embedding representation for use by the subsequent heterogeneous graph relationship learning module. To further enhance semantic consistency between different modal embeddings, a cross-modal regularization term is introduced. This regularization term ensures that the features of the same event remain consistent across all modalities by constraining the distribution of different modal embeddings in the low-dimensional space. The design of the cross-modal regularization term includes:

[0101] ,

[0102] in, For cross-modal loss function, and Let be the embedding vectors of the same landslide event in the i-th and j-th modes, respectively. By minimizing this regularization term, we can ensure that the distribution of embeddings in different modes tends to be consistent.

[0103] In this optional embodiment, a cross-modal alignment mechanism is proposed, which is based on contrastive learning InfoNCE loss and cross-modal attention mechanism to ensure the consistency of the representation of the same landslide event in different modal feature spaces.

[0104] Optionally, the step of performing correlation analysis on the updated landslide heterogeneity map to obtain the target landslide heterogeneity map includes:

[0105] Multi-head heterogeneous attention calculation is performed on each node of the updated landslide heterogeneous graph to obtain the corresponding attention weights;

[0106] Based on the channel weight control mechanism, the nodes of the updated landslide heterogeneous map are aggregated according to all the attention weights to obtain the target landslide heterogeneous map.

[0107] The channel weight control mechanism includes:

[0108] ,

[0109] in, Let the output of node i be represented. and The dynamic contribution rates of the data channel information and the knowledge channel information are respectively. Let i be the embedding vector of the data channel information of node i. Let be the embedding vector of node i in the knowledge channel information, softmax be the normalization function, and U be the linear transformation matrix.

[0110] Specifically, traditional GNN information aggregation methods struggle to capture the complex relationships between different types of nodes in heterogeneous graphs, ignoring semantic differences arising from network structural heterogeneity and failing to effectively model multi-hop paths, structural dependencies, and semantic dependencies. Therefore, we employ Heterogeneous Graph Transformer (HGT) to fuse information from different types of nodes in our constructed complex heterogeneous graph network, aggregating network node information flows to obtain potential causal structures and learn slippery association patterns. HGT introduces a type-aware multi-head attention mechanism, suitable for graph structures with multiple nodes and multiple edge types. Each edge type r has an independent attention function. When calculating the aggregation of information from neighbor u by node v, the multi-head attention calculation includes:

[0111] ,

[0112] in, Let u, v, and r represent the attention score, and u, v, and r represent the neighbor node, the target node, and the relationship type between the neighbor node and the target node, respectively. and These are the feature vectors of the neighbor node u and the target node v, respectively. and Node types , The corresponding weight matrix can map the features of different types of nodes to a unified feature space. The square root of the feature dimension is used to scale the attention score. Interactions between the same pair of nodes receive different attention weights depending on the relationship type, reflecting the significant impact of type differences on information transmission. This formula calculates the attention relevance score of neighbor node u to target node v under relationship type r, measuring the importance of information interaction between them. Message passing combines information through the attention of different heads, including:

[0113] ,

[0114] in, For combined information under relation type r, Let r be the set of neighbors connected by relations r and v. The normalized attention weight of node u to node v under relation type r is obtained by normalizing the above formula using softmax. For node type The corresponding value mapping matrix is ​​used to extract feature information from neighboring nodes. The input features of the neighboring nodes are used to obtain the final result. That is, the message vector aggregated by node v from the neighborhood of relation type r reflects the comprehensive information received by node v under this type of relation. After merging the multi-head attention, type-sensitive residual updates and normalization are performed to obtain the final output embedding:

[0115] ,

[0116] in, This represents the embedding of node v at layer l+1. To combine information, Let v be the embedding representation of point v in layer l, LayerNorm be the layer normalization, and FFN be the feedforward network. The model can perceive the interaction between different path structures (metapaths) and different modal information. The network learns the contextual semantic embedding of each node through multiple rounds of information propagation.

[0117] In the model inference stage, to achieve the acquisition of the correlation between multiple elements of landslide disaster, this invention adopts a relation-aware heterogeneous graph Transformer (GHT) structure. First, the multimodal features of nodes such as event, geology, geomorphology, rainfall, soil, and vegetation are mapped to a unified embedding space, and the channel label (data channel / knowledge channel) to which the node belongs is introduced as additional input. For each pair of connected nodes... Through relation type r and edge attributes (weights) Credibility The calculation of its relation-aware attention coefficient includes:

[0118] ,

[0119] in, , Let i and j be the node features of the i-th and j-th nodes. , For a specific transformation matrix of the relation, , These are the weighting coefficients. This formula comprehensively considers node semantic similarity, edge attributes, and knowledge credibility. Attention coefficient The square root of the feature dimension. Softmax normalization is applied. Furthermore, to reflect the dual-channel collaborative driving characteristic, a channel weight adjustment mechanism is introduced during the node aggregation stage:

[0120] ,

[0121] in, Let the output of node i be represented. and The dynamic contribution rates of the data channel information and the knowledge channel information are respectively. Let i be the embedding vector of the data channel information of node i. Let be the embedding vector of node i in the knowledge channel information, and softmax be the normalization function. The node fusion representation is obtained after multi-layer aggregation. The heterogeneous graph edge weights are updated based on the attention weights between node pairs to construct the output graph. ,in The weights represent the strength of the disaster correlation among multiple landslide elements. In this way, the system not only obtains the implicit correlations between multiple source elements, but also dynamically reflects the impact of information from different channels on the disaster propagation path, enabling interpretable modeling of the potential causal relationships between landslide events.

[0122] Optionally, the nodes include landslide event nodes, cause nodes, geological condition nodes, time nodes, spatial nodes, and text nodes.

[0123] Specifically, the nodes include: landslide event nodes (Event), trigger nodes (Trigger), geological condition nodes (Geo), time nodes (Time), spatial nodes (Location), and text nodes (Text). Regarding edge construction, the landslide heterogeneous mapping package introduces multiple edge types to reflect the similarity between disasters and disaster nodes, and the correlation between disasters and attribute data nodes. Edge types include: Event-Trigger, Event-Geo, Event-Time, Event-Location, Event-Text, and similarity edges between events (Event-Event), etc. In terms of node representation, for different types of nodes such as landslide events, triggers, geomorphic conditions, time, location, and text, type identifiers (type-id), globally unique identifiers (uid), and property dictionaries are defined respectively. Numerical attributes (such as slope, cumulative rainfall, and curvature) are directly used as continuous features after Z-score standardization. Categorical / enumerated attributes (such as lithology, soil type, and land use) are encoded using learnable embedding vectors. Text attributes (event descriptions, expert rules, and causal chains) are processed by a domain text encoder to obtain dense semantic vectors. Spatial nodes are mainly based on latitude, longitude, altitude, and topographically derived factors, and are accompanied by coordinate reference systems and spatial indexes (such as GeoHash / QuadKey). Time nodes use timestamps, periodic location encoding (year / season / day / hour), and event window annotations. Landslide event nodes simultaneously store channel identifiers for both data channels and knowledge channels. The data includes its source metadata (source-id, collection date, data quality level), and during initialization, the multi-source attributes are concatenated and mapped to a fixed-dimensional vector using a type-specific encoder. Regarding edge representation, a directed multi-relation edge set R = {Event-Trigger, Event-Geo, Event-Time, Event-Location, Even-Text, Event-Event} is defined. Each edge records its edge type (r-type), direction, source channel label (data / knowledge), weight w, and confidence level. It can also include relation-specific attributes (such as spectral distance / cosine similarity of event-event similarity edges, spatial distance and buffer level of event-location edges, and time delay of event-time edges). (e.g., matching scores of event-text edges); To support relation-aware feature transfer, the system configures an independent relation transformation matrix / affine mapping Wr and optional semantic attention parameters for each type of relation, used to implement relation-based message passing and weighted aggregation in subsequent graph learning; simultaneously, it constructs a typed adjacency list and a sparse tensor. To enable efficient access and batch processing, each relation r belongs to a set R. At the graph schema level, a constraint is set that "event nodes must be connected to at least one data edge and one knowledge edge simultaneously" to ensure the coordinated expression of dual-channel information within the structure.

[0124] In some more specific embodiments, data preprocessing for multi-source landslide disaster data includes: for remote sensing image data, radiometric calibration, atmospheric correction, projection transformation, and image fusion are performed to obtain surface observation data under a unified spatial reference system; for DEM data, resampling, smoothing, and boundary clipping are performed, and secondary topographic factors such as slope, aspect, slope length, and curvature are further derived; for geological structure map data, vector format unification and element classification are performed, and after standardizing structural information such as faults and folds, it is converted into raster data and spatially registered with other data; for soil type data, a unified classification system is used, and attributes such as soil thickness, texture, and water content are normalized and then rasterized, aligned with topographic data; for hydrological and meteorological data, spatial interpolation and time series reconstruction methods are used to fill in missing measurements and construct factor sequences such as rainfall, temperature, and evapotranspiration under a unified time scale. For land use data, classification, simplification, and recoding are performed based on remote sensing interpretation results to ensure consistency in data caliber across different years, and time series overlay is used to capture changes in human activities. For NDVI vegetation index data, the maximum value synthesis method is used to eliminate cloud and fog interference, and the mean and fluctuation rate of vegetation cover are calculated based on seasonal scales to reflect vegetation dynamics. For road and mine distribution data, topological checks and cleaning are first performed on vector data to extract road grades, buffer zone influence ranges, and mining area disturbance ranges, and then unified projection and rasterization are applied. For historical landslide point data, cross-validation is performed through field survey records and remote sensing interpretation, and spatial positioning and attribute standardization are completed based on geographic coordinates to form a spatiotemporal database of landslide events. For prior knowledge data, including expert experience rules, landslide classification standards, and causal link descriptions, they are unified into structured semantic representations and mapped to knowledge graph nodes or tag systems. Subsequently, the cleaned data of the same type were standardized using the Z-score method to eliminate the differences in units and magnitudes of different physical quantities. Then, the Bayesian estimation method was used to collaboratively organize the multi-source data, aligning text data, image data and time series data to a unified feature space to achieve the fusion representation of multimodal data, thereby providing high-quality input for subsequent dual-channel collaborative modeling and landslide disaster correlation analysis.

[0125] like Figure 2As shown in the figure, an embodiment of the present invention provides a multi-factor correlation analysis device for landslide disasters, comprising:

[0126] The landslide knowledge base construction module 10 is used to construct a landslide knowledge base through multi-source landslide disaster data. The landslide knowledge base includes data channel information and knowledge channel information. The data channel information is used to reflect the objective characteristics of landslide events, and the knowledge channel information is used to reflect the semantic logic of landslide events.

[0127] The landslide heterogeneous map acquisition module 20 is used to obtain a landslide heterogeneous map based on the data channel information and the knowledge channel information, wherein the landslide heterogeneous map includes nodes and edges;

[0128] 30 low-dimensional feature vectors are obtained, which are used to embed and align the multimodal data of the landslide heterogeneous map using data source labels to obtain low-dimensional feature vectors;

[0129] The cross-modal attention mechanism module 40 is used to obtain the updated landslide heterogeneous map based on all the said low-dimensional feature vectors according to the cross-modal attention mechanism.

[0130] The correlation analysis module 50 is used to perform correlation analysis on the updated landslide heterogeneous map based on the GHT heterogeneous graph neural network correlation analysis algorithm to obtain the target landslide heterogeneous map, wherein the target landslide heterogeneous map is used to reflect the correlation of multiple factors of landslide disaster.

[0131] like Figure 3 As shown in the figure, an electronic device 300 provided in this embodiment of the invention includes a memory 310 and a processor 320; the memory 310 is used to store a computer program; the processor 320 is used to implement the landslide disaster multi-factor correlation analysis method as described above when the computer program is executed.

[0132] Alternatively, an electronic device 300 includes a memory 310 and a processor 320 coupled to the memory 310; the memory 310 is configured to store a computer program; and the processor 320 is configured to perform the following operations when the computer program is executed:

[0133] A landslide knowledge base is constructed using multi-source landslide disaster data. The landslide knowledge base includes data channel information and knowledge channel information. The data channel information is used to reflect the objective characteristics of landslide events, and the knowledge channel information is used to reflect the semantic logic of landslide events.

[0134] A landslide heterogeneous graph is obtained based on the data channel information and the knowledge channel information, wherein the landslide heterogeneous graph includes nodes and edges;

[0135] The multimodal data of the landslide heterogeneous map are embedded and aligned using data source labels to obtain low-dimensional feature vectors;

[0136] Based on the cross-modal attention mechanism, an updated landslide heterogeneous map is obtained from all the aforementioned low-dimensional feature vectors;

[0137] Based on the GHT heterogeneous graph neural network correlation analysis algorithm, the updated landslide heterogeneous graph is subjected to correlation analysis to obtain the target landslide heterogeneous graph, wherein the target landslide heterogeneous graph is used to reflect the correlation of multiple factors of landslide disaster.

[0138] This invention provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the multi-factor correlation analysis method for landslide disasters as described above.

[0139] Alternatively, a non-volatile computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform the following operations:

[0140] A landslide knowledge base is constructed using multi-source landslide disaster data. The landslide knowledge base includes data channel information and knowledge channel information. The data channel information is used to reflect the objective characteristics of landslide events, and the knowledge channel information is used to reflect the semantic logic of landslide events.

[0141] A landslide heterogeneous graph is obtained based on the data channel information and the knowledge channel information, wherein the landslide heterogeneous graph includes nodes and edges;

[0142] The multimodal data of the landslide heterogeneous map are embedded and aligned using data source labels to obtain low-dimensional feature vectors;

[0143] Based on the cross-modal attention mechanism, an updated landslide heterogeneous map is obtained from all the aforementioned low-dimensional feature vectors;

[0144] Based on the GHT heterogeneous graph neural network correlation analysis algorithm, the updated landslide heterogeneous graph is subjected to correlation analysis to obtain the target landslide heterogeneous graph, wherein the target landslide heterogeneous graph is used to reflect the correlation of multiple factors of landslide disaster.

[0145] The present invention will now be described an electronic device 300 that can serve as a server or client of the present invention, which is an example of a hardware device that can be applied to various aspects of the present invention. Electronic device 300 is intended to represent various forms of digital electronic computer devices, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. Electronic device 300 can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0146] Electronic device 300 includes a computing unit that can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) or a computer program loaded from a storage unit into random access memory (RAM). The RAM may also store various programs and data required for device operation. The computing unit, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0147] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc. In this application, the units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of the present invention according to actual needs. Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units can be implemented in hardware or as software functional units.

[0148] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.

Claims

1. A landslide disaster multi-element correlation analysis method, characterized by, The application relates to a landslide knowledge base construction method based on multi-source landslide disaster data. The landslide knowledge base comprises data channel information and knowledge channel information, wherein the data channel information is used for reflecting objective characteristics of a landslide event, and the knowledge channel information is used for reflecting semantic logic of the landslide event. A landslide heterogeneous graph is obtained according to the data channel information and the knowledge channel information, wherein the landslide heterogeneous graph comprises nodes and edges. Low-dimensional feature vectors are obtained by embedding and aligning multi-modal data of the landslide heterogeneous graph by using data source labels. An updated landslide heterogeneous graph is obtained according to all the low-dimensional feature vectors based on a cross-modal attention mechanism. A target landslide heterogeneous graph is obtained by performing correlation analysis on the updated landslide heterogeneous graph based on a GHT heterogeneous graph neural network correlation analysis algorithm, wherein the target landslide heterogeneous graph is used for reflecting landslide disaster multi-element correlation. Multi-head heterogeneous attention calculation is performed on each node of the updated landslide heterogeneous graph to obtain corresponding attention weights. The target landslide heterogeneous graph is obtained by aggregating each node of the updated landslide heterogeneous graph according to all the attention weights based on a channel weight regulation mechanism. The channel weight regulation mechanism comprises: ， wherein, is an output representation of a node i, and are dynamic contribution degrees of the data channel information and the knowledge channel information, respectively, is an embedding vector of node i in the data channel information, is an embedding vector of node i in the knowledge channel information, and softmax is a normalization function, and U is a linear transformation matrix.

2. The landslide disaster multi-element correlation analysis method according to claim 1, characterized by, The data source labels comprise data channel labels and knowledge channel labels. The data channel labels and the knowledge channel labels are used for encoding and mapping multi-modal data of the landslide heterogeneous graph to obtain embedded feature vectors. The embedded feature vectors are aligned in modes based on a cross-modal contrast learning method to obtain the low-dimensional feature vectors.

3. The landslide disaster multi-element correlation analysis method according to claim 2, characterized by, The data channel labels comprise image data labels and time sequence data labels, and the embedded feature vectors comprise image embedded vectors and time sequence embedded vectors. When the data channel labels are the image data labels, multi-modal data of the landslide heterogeneous graph is input into a pre-trained ResNet visual model for feature extraction to obtain the image embedded vectors, wherein the pre-trained ResNet visual model is used for image data processing by using a deep convolutional neural network. When the data channel labels are the time sequence data labels, multi-modal data of the landslide heterogeneous graph is input into an LSTM model for learning to obtain the time sequence embedded vectors, wherein the LSTM model is used for capturing long-term dependence of time sequence data.

4. The landslide disaster multi-element correlation analysis method according to claim 3, characterized by, The knowledge channel labels comprise text data labels and expert knowledge rule labels, and the embedded feature vectors comprise text embedded vectors and semantic embedded vectors. The data channel labels and the knowledge channel labels are used for encoding and mapping multi-modal data of the landslide heterogeneous graph to obtain embedded feature vectors. When the knowledge channel label is the text data label, the multi-modal data of the landslide heterogeneous graph is input into a pre-trained RoBERTa language model for encoding to obtain the text embedding vector, wherein the pre-trained RoBERTa language model is used for word tokenization processing on text data; When the knowledge channel label is the expert knowledge rule label, a TransE knowledge graph embedding method is used to map the multi-modal data of the landslide heterogeneous graph to a low-dimensional space to obtain the semantic embedding vector.

5. The landslide disaster multi-element correlation analysis method according to claim 1, characterized by, The low-dimensional feature vectors include text low-dimensional vectors, image low-dimensional vectors, and time sequence low-dimensional vectors; and the updated landslide heterogeneous graph obtained according to all the low-dimensional feature vectors includes: The text low-dimensional vector, the image low-dimensional vector, and the time sequence low-dimensional vector corresponding to each node are weighted and fused to obtain a single embedding vector; The updated landslide heterogeneous graph is obtained according to the single embedding vector; The node updating process includes: ， wherein, is a low-dimensional vector edge-weighted fusion result, is the text low-dimensional vector, is the image low-dimensional vector, is the time-series low-dimensional vector, , and is a weight parameter.

6. The landslide disaster multi-element correlation analysis method according to claim 1, characterized by, The nodes include landslide event nodes, cause nodes, geological condition nodes, time nodes, space nodes, and text nodes.

7. A landslide disaster multi-element correlation analysis device characterized by comprising: It includes: A landslide knowledge base construction module is configured to construct a landslide knowledge base through multi-source landslide disaster data, wherein the landslide knowledge base includes data channel information and knowledge channel information, the data channel information is used to reflect the objective characteristics of a landslide event, and the knowledge channel information is used to reflect the semantic logic of the landslide event; A landslide heterogeneous graph acquisition module is configured to obtain a landslide heterogeneous graph according to the data channel information and the knowledge channel information, wherein the landslide heterogeneous graph includes nodes and edges; A low-dimensional feature vector acquisition is configured to embed and align multi-modal data of the landslide heterogeneous graph by using a data source label to obtain low-dimensional feature vectors; A cross-modal attention mechanism module is configured to obtain an updated landslide heterogeneous graph according to all the low-dimensional feature vectors based on a cross-modal attention mechanism; An association analysis module is configured to perform association analysis on the updated landslide heterogeneous graph based on a GHT heterogeneous graph neural network association analysis algorithm to obtain a target landslide heterogeneous graph, wherein the target landslide heterogeneous graph is used to reflect landslide disaster multi-element association, including: Multi-head heterogeneous attention calculation is performed on each node of the updated landslide heterogeneous graph to obtain corresponding attention weights; Based on a channel weight regulation mechanism, all the attention weights are used to aggregate nodes of the updated landslide heterogeneous graph to obtain the target landslide heterogeneous graph. The channel weight regulation mechanism includes: ， wherein, is an output representation of a node i, and are dynamic contribution degrees of the data channel information and the knowledge channel information, respectively, is an embedding vector of node i in the data channel information, is an embedding vector of node i in the knowledge channel information, and softmax is a normalization function, and U is a linear transformation matrix.

8. An electronic device, comprising: It includes a memory and a processor. The memory is configured to store a computer program. The processor is configured to implement the landslide disaster multi-element association analysis method of any one of claims 1 to 6 when executing the computer program.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, and when the computer program is executed by the processor, the landslide disaster multi-element association analysis method of any one of claims 1 to 6 is implemented.

Images