Landslide space prediction method based on knowledge graph and representation learning

Abstract

The invention provides a landslide space prediction method based on a knowledge graph and representation learning. The landslide space prediction method specifically comprises the following steps: S1, constructing a landslide susceptibility knowledge graph considering mechanism knowledge and spatial-temporal characteristics; a knowledge graph ontology mode is designed, landslide disaster field knowledge modeling considering spatial-temporal characteristics and mechanism knowledge is completed, and on this basis, entities are extracted from multi-source data, relationships are constructed, and a knowledge graph data layer is constructed; and S2, through map semantic feature learning based on a ComplEx model, directed weighted graph construction based on a landslide susceptibility knowledge map, equivalent positive and negative landslide sampling based on a Louvain algorithm and map structural feature learning based on a GraphSAGE model, realizing landslide space prediction based on the knowledge map and representation learning. According to the method, diversified feature inductive learning under the constraint of fewer landslide samples in a complex heterogeneous environment can be realized, and the landslide space distribution probability is predicted through a node classification task.

Description

Technical Field

[0001] The present invention relates to the field of spatial information technology, and in particular to a landslide spatial prediction method based on knowledge graph and representation learning. Background Art

[0002] Existing studies have shown that landslides are affected by the coupling of multiple factors, with hidden locations and unclear disaster-pregnant conditions. Complex and heterogeneous geographical environments promote the formation of diverse landslide characteristics. In order to improve the level of landslide risk management, it is urgent to explore the complex disaster mechanism of landslides, master the diverse landslide characteristics, improve the spatial prediction ability of landslides at the county level, determine the location of landslide-prone areas, and on this basis carry out scientific monitoring, early warning and active prevention to avoid the occurrence of catastrophic disasters.

[0003] Spatial prediction of landslides aims to assess the possibility of landslides in a specific area. Current research at home and abroad is mainly divided into two methods: knowledge-driven and data-driven. Although the knowledge-driven mechanical mechanism model can achieve relatively accurate simulation, it relies on high-precision geological data, and the parameters are difficult to obtain. It is suitable for deterministic prediction within a single body. The empirical rule method selects comprehensive and typical influencing factors and determines the weights based on the knowledge of domain experts, and performs superposition analysis. It is simple to operate and does not rely on data samples, but it is highly subjective, and the expert experience is scattered. It does not fully consider the nonlinear relationship between disaster mechanisms and landslide occurrence. The data-driven method uses influencing factors and historical landslide data to establish a prediction model, which is more suitable for non-deterministic spatial prediction of landslides in a regional scope and is the mainstream of current research. For example, random forests, support vector machines, artificial neural networks, etc., use landslide influencing factors as prediction variables, known landslides as labels, and use a large amount of historical data for training, showing higher prediction accuracy in simple landslide scenarios. However, the prediction method based solely on data still does not fully consider the complex mechanism of landslide disasters, and ignores the interrelationships between influencing factors, such as the clear relationship between topographic conditions affecting vegetation coverage and soil moisture changes, or the potential impact of human activities on landslide development, resulting in insufficient knowledge modeling of complex landslide scenarios. At the same time, the effective association between prediction units that are far away in space is ignored, and only the landslides and their geographical environment characteristics in the local range of the prediction unit are considered, which limits the model's ability to extract and represent features in complex heterogeneous environments. In addition, the model relies on a large amount of high-quality landslide sample data, but the number of landslide samples at the county level is limited. The non-landslide samples generated by traditional random sampling or buffer control sampling methods are often located on steep slopes prone to landslides, which have high uncertainty, further limiting the model's ability to learn diverse landslide and non-landslide features.

[0004] The introduction of knowledge graph technology can effectively model complex mechanism knowledge, integrate multi-source heterogeneous data information, discover the correlation and dependency between data and knowledge, and improve the accuracy of spatial prediction results in complex landslide scenarios. For example, Ge et al. integrated remote sensing image data, geographic element information and domain expert knowledge to construct a knowledge graph of landslide disasters, converted the fitting model of historical landslide disasters and DEM data into semantic reasoning rules, and realized the spatial prediction of landslides based on rule reasoning. Chen et al. used the prediction unit list, historical landslide disaster data and environmental data to enhance the representation of complex landslide scenes, captured the semantic information in the knowledge graph by training the knowledge graph representation learning model (KRL), and realized the landslide susceptibility mapping under sparse landslide samples based on the link prediction algorithm. However, the existing research on knowledge graphs in landslide spatial prediction tasks is still mostly limited to knowledge reasoning based on manually defined rules, which has poor flexibility; the link prediction based on the KRL model only focuses on the learning of local structural features of the knowledge graph, and lacks deep learning of graph structural features. Due to the limited number of landslide samples at the county level and the diverse landslide characteristics, the model's feature extraction and representation capabilities are limited. The graph structure of the knowledge graph can establish effective associations between similar and adjacent regional features at different spatial distances by utilizing nodes and edges. On this basis, the graph neural network is introduced. By defining the depth of the traversal graph, deeper graph structure features can be extracted, thereby improving the ability to summarize and learn diverse landslide and non-landslide features in complex and heterogeneous environments.

[0005] In summary, the main shortcomings of the prior art include:

[0006] (1) Existing knowledge-driven methods have problems such as difficulty in obtaining parameters, and the selection of influencing factors and the determination of their weights rely on expert experience. At the same time, they do not fully consider that the landslide deformation and destruction process is a nonlinear dynamic system that develops and evolves under the coupling of multiple factors such as topography, geology, and meteorology.

[0007] (2) The existing data-driven methods do not fully consider the effective correlation between various influencing factors, and are insufficient in modeling the complex landslide scene knowledge. They are limited to the landslides and their geographical environment characteristics within the local scope of the prediction unit. In addition, the number of landslide samples within the county is limited, and the non-landslide samples obtained by traditional sampling methods have high uncertainty, so the model is prone to overfitting or underfitting.

[0008] (3) Existing methods for spatial landslide prediction based on knowledge graphs are limited to knowledge reasoning based on manually defined rules and lack flexibility. Link prediction methods based on KRL models still lack deep learning of graph structure features.

[0009] To this end, the present invention aims to propose a landslide spatial prediction method based on knowledge graph and representation learning, modeling complex mechanism knowledge, integrating multi-source data, so as to further learn the effective association between data and knowledge. Summary of the invention

[0010] The purpose of the present invention is to propose a landslide spatial prediction method based on knowledge graph and representation learning; this method standardizes the mechanism knowledge and spatiotemporal relationship of landslide disasters, and constructs a landslide field knowledge graph that takes into account the mechanism knowledge and spatiotemporal characteristics; on this basis, the knowledge graph representation learning model (ComplEx) and the graph neural network model (GraphSAGE) are introduced to learn the semantic information related to the complex disaster mechanism, comprehensively consider the influence of different conditional factors on landslides, the cluster characteristics of regional landslides and the geographical similarity characteristics, and deeply learn the structural characteristics of the graph to achieve diversified feature induction learning under complex heterogeneous environments and fewer landslide sample constraints, and predict the spatial distribution probability of landslides through node classification tasks. In addition, the geographical similarity correlation strength and structure are comprehensively considered to construct an equal number of positive and negative landslide sample sets to improve the typicality of non-landslide samples and reduce their uncertainty. Based on the spatial prediction results, landslide susceptibility mapping is carried out to explore the spatial distribution law of regional potential landslides, providing support for early identification and prevention and control management of landslides.

[0011] To achieve the above object, the technical solution of the present invention is: a landslide spatial prediction method based on knowledge graph and representation learning, which specifically includes the following steps:

[0012] S1. Construct a knowledge graph of landslide susceptibility that takes into account mechanism knowledge and spatiotemporal characteristics; design a knowledge graph ontology model, complete the knowledge modeling of landslide disasters that takes into account spatiotemporal characteristics and mechanism knowledge, and on this basis extract entities from multi-source data, build relationships, and realize the construction of the knowledge graph data layer;

[0013] S2. Through graph semantic feature learning based on the ComplEx model, directed weighted graph construction based on the landslide susceptibility knowledge graph, equal positive and negative landslide sampling based on the Louvain algorithm, and graph structural feature learning based on the GraphSAGE model, spatial prediction of landslides based on knowledge graph and representation learning is achieved.

[0014] Preferably, the design knowledge graph ontology model completes the landslide disaster field knowledge modeling taking into account the temporal and spatial characteristics and mechanism knowledge, and is achieved through the design of a model layer, which includes a prediction unit model, a landslide disaster model, a model method model, and a disaster mechanism model.

[0015] Preferably, the design of the mode layer is as follows:

[0016] The tile in the prediction unit mode is the object of landslide spatial prediction. The experimental area is divided into M*N grid cells, which are modeled as entities as prediction units;

[0017] The landslide_hazard in the landslide hazard model is a record of historical landslide hazard events in the experimental area. The spatiotemporal characteristics of the landslide hazard event records are modeled to form a description of the basic attributes of the landslide, including scale, disaster time, and geographical location; where time, as an entity, represents the moment when the landslide hazard event occurs, and the relationship between time entities includes inclusion relationship and order relationship; space, as an entity, represents the geographical location where the landslide hazard event occurs, and the relationship between space entities includes inclusion relationship;

[0018] The prediction_model in the model method mode is a variety of model methods used for landslide spatial prediction tasks, including five types of models: random forest, support vector machine, extreme gradient boosting, multi-layer perceptron and graph neural network; the landslide disaster events associated with tiles and their environmental condition characteristics are used as inputs of the model method, and the spatial probability of landslide occurrence is output;

[0019] The disaster mechanism model describes the disaster environment, analyzes the interaction of influencing factors and their impact on landslides, and designs semantic logical associations, including the design of semantic association relationships in six major categories: topography, geological conditions, vegetation conditions, soil conditions, meteorology and hydrology, and human activities.

[0020] Preferably, the semantic association relationship design of the six categories of topography, geological conditions, vegetation conditions, soil conditions, meteorology and hydrology, and human activities is specifically as follows:

[0021] The definition domain of the topography includes landform type, slope, aspect and altitude; the value range of the influencing relationship of the aspect includes soil type, vegetation coverage and vegetation type, and the value range of the influencing relationship of the altitude includes vegetation type and rainfall;

[0022] The definition domain of the geological conditions includes stratum lithology and geological structure; the value domain of the influence relationship of stratum lithology includes soil type and vegetation type;

[0023] The definition domain of the vegetation condition includes vegetation coverage and vegetation type;

[0024] The definition domain of meteorology and hydrology includes rainfall, water system distribution and terrain humidity index; the value range of the rainfall influence relationship includes vegetation coverage and soil type;

[0025] The definition domain of the soil conditions includes soil type;

[0026] The definition domain of the human activities includes land use and road construction; the value domain of the impact relationship between land use and road construction includes soil type, vegetation coverage and vegetation type.

[0027] Preferably, the construction of the knowledge graph data layer includes the following steps:

[0028] S11. Create a tile coordinate set according to the Mercator projection rules to obtain the prediction unit in the study area. The conversion rules are as follows

[0029]

[0030] In the formula, lon and lat represent the input longitude and latitude coordinates, c and y represent the horizontal and vertical coordinates after the slice transformation, and z represents the zoom level.

[0031] S12. Use the NDVI numerical algorithm to obtain the spatial distribution of surface vegetation from high-resolution satellite images, convert raster and vector data of various scales into GeoJSON files, and calculate the slice index list corresponding to each feature object; directly create a slice index list for discrete attributes, and for continuous attributes, select the appropriate scale discrete attributes to generate a slice index list of the discretized attribute interval;

[0032] S13, converting the above processing results into structured triples, including prediction unit triples, mechanism triples, landslide triples, and time-space triples;

[0033] S14. Map the data instances to the ontology schema to build the landslide susceptibility knowledge graph.

[0034] Preferably, the graph semantic feature learning based on the ComplEx model is specifically as follows:

[0035] The ComplEx knowledge graph representation learning model is used to process binary relations through complex space, realize effective modeling of symmetric and asymmetric relations, learn semantic association information in the knowledge graph, and obtain the embedded representation vectors of all entity nodes; the model scoring function is:

[0036]

[0037] In the formula, The embedding vectors representing entities and relations are both d-dimensional complex vectors. h is the d-dimensional complex vector representation of the head entity, r is the d-dimensional complex vector representation of the relation, and t is the d-dimensional complex vector representation of the tail entity. It represents the conjugate complex number of the tail entity vector, diag(r) represents a diagonal matrix, the diagonal elements are the elements of the relationship vector r, and Re() represents taking the real part of the complex entity.

[0038] Preferably, the directed weighted graph construction based on the landslide susceptibility knowledge graph is specifically as follows:

[0039] Firstly, the frequency ratio model FR is used to quantitatively calculate the spatial variability relationship between landslide occurrence and various types of influencing factors:

[0040]

[0041] In the formula, FR(Value) indicates the importance of each category of influencing factors on landslides. When FR(Value)>1, it indicates that it is favorable for landslides to occur; when FR(Value)<1, it indicates that it is unfavorable for landslides to occur. j is the number of landslide tiles that appear in a certain classification interval for a certain category element, F is the total number of tiles occupied by all landslides in a certain classification interval for a certain category element, and C j Then is the number of tiles of a certain category of elements in a certain classification interval, and C is the total number of tiles in the study area;

[0042] Considering the uncertainty of other semantic associations on landslides, the corresponding edge attribute value is designed to be 1:

[0043]

[0044] In the formula, u and v represent nodes. If FR(value) is included, the attribute value of the edge is the influence degree of each level of influence index calculated by the FR model on the occurrence of landslide, otherwise it is 1;

[0045] Secondly, the intersection matrix between the interior, boundary, and exterior of two geometric figures is constructed through the dimensionally extended nine-intersection model DE-9IM to represent the topological spatial relationship between them:

[0046]

[0047] In the formula, R STR(x,y) Indicates the spatial topological relationship between the prediction unit entities x and y involved; I, B, and E represent the interior, boundary, and exterior of the prediction unit entity, respectively; the dim function represents the return value of the geometric object dimension;

[0048] For spatial adjacency, the edge attribute value is designed to be 1, that is, Adjustments are then made through model training;

[0049] Finally, based on the results of the ComplEx model embedding, the cosine similarity is used to calculate the geographic similarity strength between different prediction units, and the top 8 node units with the highest similarity association strength with each prediction unit are selected to establish a geographic similarity relationship, and the geographic similarity strength is used as the edge attribute value of the similarity relationship:

[0050]

[0051] In the formula, x u ·x v represents the dot product of nodes u and v, ||x|| 2 represents the L2 norm of vector X, Indicates the attribute value of the similarity relationship edge.

[0052] Preferably, the equal positive and negative landslide sampling based on Louvain algorithm is specifically:

[0053] From the perspective of graph structure, the topological structure and strength of geographical similarity associations between different prediction units are comprehensively considered based on the Louvain algorithm, and the region is divided into multiple geographical intervals. By selecting non-landslide samples with the same number of positive samples and the lowest similarity association strength in each partition, it is ensured that negative samples can reflect the typical characteristics under non-landslide conditions, so as to balance the number of positive and negative samples and the typical landslide and non-landslide sample characteristics of various prediction units.

[0054]

[0055] In the formula, Q represents modularity; A ij is an element of the adjacency matrix. If there is an edge between i and j, then A ij =1, otherwise A ij =0; k i and k j are the degrees of nodes i and j respectively, W is the sum of the total weights of all edges in the network; δ(c i ,c j ) is the Kroneckerdelta function, when nodes i and j belong to the same community, δ(c i ,c j )=1, otherwise δ(c i ,c j )=0.

[0056] Preferably, the graph structure feature learning based on the GraphSAGE model is specifically as follows: taking the positive and negative landslide sample data sets and the directed weighted graph as model inputs, updating the node embedding vector representation of the prediction unit based on the trained GraphSAGE model, and converting it into the probability of whether it is a landslide through the softmax activation function; in the construction of the model, the weights of different neighbor nodes are considered in the process of neighbor information aggregation to enhance the inductive learning of node features.

[0057] Preferably, the weights of different neighbor nodes are considered in the neighbor information aggregation process, specifically:

[0058] In the selection of node neighbors, for each node u, the complete neighborhood set Sampling is performed to obtain neighbor node features; when the neighbor node feature information is aggregated, the edge attribute values ​​of various association relationships are As the weight of message passing, a weighted average aggregator is designed to aggregate the neighborhood feature information of the current node u:

[0059]

[0060] In the formula, represents the neighborhood of the current node u, e uv is the edge attribute value, aggregate is the aggregation function, k is the search depth, is the embedding representation of node u in the k+1 layer neighborhood;

[0061] Since the self-correlation of each node u has been added to the graph, the aggregation of neighborhood features New node features Combining its own feature information with the neighborhood feature information of the current node, the update function is used to obtain the aggregated representation of the next layer of node information for information transmission to obtain the final embedded representation of the node;

[0062]

[0063] Where σ is the nonlinear activation function ReLU, W' is the weight matrix of neighbor features at different levels, is the final embedding representation of node u, and b is the offset matrix.

[0064] Compared with the prior art, the present invention has the following beneficial effects:

[0065] (1) The present invention systematically sorts out the knowledge in the field of landslide disasters, expresses the influencing factors of different conditions and their complex interactions with landslides, and realizes the explicit modeling of landslide mechanism knowledge. On this basis, a knowledge graph representation learning model (ComplEx) is introduced to capture the semantic features of the graph and provide data and knowledge interfaces for the landslide spatial prediction task.

[0066] (2) The present invention constructs a graph structure that includes various influencing factors, spatial adjacency, and geographical similarity. The prediction units, influencing factors, etc. are represented by nodes. Geographically similar and spatially adjacent prediction unit nodes are effectively associated by edges, and the strength of the corresponding associations is comprehensively considered, so that the model can more fully consider the interaction between prediction units, the interaction between influencing factors and prediction units and other influencing factors. The graph neural network model (GraphSAGE) is used to aggregate and update the information of surrounding nodes to generate more representative feature vectors, thereby improving the model's ability to summarize and learn landslide (non-landslide) features in complex and heterogeneous environments and better identify landslides.

[0067] (3) Based on the Louvain algorithm, by maximizing the number and weight of similarity edges within the partition and minimizing the correlation between partitions, a geographical partition result with high internal consistency and external difference is formed, and the prediction unit with the lowest similarity in each partition is selected as the negative sample. Compared with the traditional sampling strategy based on buffer distance, this method avoids excessive reliance on spatial distance and more comprehensively considers the actual impact of factors such as terrain and geology on landslides, thereby improving the typicality of non-landslide samples and reducing their uncertainty. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 This is a flow chart of the prediction of the present invention based on knowledge graph and representation learning space;

[0069] Figure 2 This is a schematic diagram of the knowledge graph model layer of the present invention;

[0070] Figure 3 It is a schematic diagram of a landslide susceptibility knowledge graph data layer (part) in a simulation experiment of the present invention;

[0071] Figure 4 The landslide spatial prediction values ​​and their error distribution diagrams of the training data (left) and test data (right) of the five methods in a simulation experiment of the present invention;

[0072] Figure 5 This is a graph showing the landslide susceptibility mapping results of five methods in a simulation experiment of the present invention. DETAILED DESCRIPTION

[0073] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings.

[0074] The present invention proposes a method for spatial prediction of landslides based on knowledge graph and representation learning. Based on remote sensing images, relevant geographic element information, spatiotemporal data of landslide disasters and expert knowledge, the method first systematically sorts out the mechanism of complex landslide disasters, faces the task requirements of spatial prediction of landslides, clarifies the relevant conceptual knowledge such as influencing factors, landslide disasters, prediction units and model methods, designs the knowledge graph pattern layer, models prior knowledge, and on this basis, coordinates multi-source data, forms a mapping between data and knowledge, and constructs a knowledge graph of landslide susceptibility. Then, the ComplEx knowledge graph representation model is used to capture the semantic information related to the disaster mechanism in the graph, and obtain the initial representation vector of the entity node; on this basis, a directed weighted graph containing geographical similarity, spatial adjacency and semantic association relationships is constructed; and a data set of balanced typical positive and negative landslide samples is constructed based on the geographical similarity association structure and strength; then the GraphSAGE graph neural network model is introduced to deeply learn the graph structure information, aggregate and update the representation vector of the entity node, and predict the spatial distribution probability of regional landslides through the node classification task. It specifically includes the following steps:

[0075] S1. Construct a knowledge graph of landslide susceptibility that takes into account mechanism knowledge and spatiotemporal characteristics; design a knowledge graph ontology model, complete the knowledge modeling of landslide disasters that takes into account spatiotemporal characteristics and mechanism knowledge, and on this basis extract entities from multi-source data, build relationships, and realize the construction of the knowledge graph data layer;

[0076] (1) Modeling the landslide domain knowledge that takes into account the temporal and spatial characteristics and mechanism knowledge. Using a top-down approach, starting from the top-level concept, the lower-level concepts are gradually refined, and the concept attributes and relationships are defined to standardize the expression of the data layer and provide a unified model for knowledge computational reasoning.

[0077] (2) Extraction and storage of landslide knowledge. Under the guidance and constraints of the model layer, based on spatiotemporal associations, multi-source data are coordinated, entities are extracted, relationships are constructed, and structured triples are formed, mainly including prediction unit triples, mechanism triples, landslide triples, and spatiotemporal triples. The Neo4j graph database is used, based on the resource description framework, to represent various relevant elements of landslide hazard knowledge in the form of (entity, relationship, entity) and (entity, attribute, attribute value), and stored in a "node-relationship" storage mode.

[0078] S2. Through graph semantic feature learning based on the ComplEx model, directed weighted graph construction based on the landslide susceptibility knowledge graph, equal positive and negative landslide sampling based on the Louvain algorithm, and graph structural feature learning based on the GraphSAGE model, landslide spatial prediction based on knowledge graph and representation learning is achieved. The process method is as follows Figure 1 shown.

[0079] (1) Graph semantic feature learning based on the ComplEx model. Based on the ComplEx knowledge graph representation learning model, the semantic association information such as prediction units and disaster mechanisms in the knowledge graph is learned to obtain the embedded representation vectors of all entity nodes.

[0080] (2) Construction of a directed weighted graph based on the graph spectrum. Considering the impact of different conditional factors on the prediction unit, the impact mechanism between different conditional factors, the geographical similarity between different prediction units, and the strength of spatial adjacent association, a more meaningful graph structure for summarizing landslide susceptibility is constructed. The embedding representation vector results of the entity nodes are used as the node features of the graph, and the attribute values ​​of the edges are used as the edge weights of the graph.

[0081] (3) Equal positive and negative landslide sampling based on Louvain algorithm. Based on the Louvain algorithm, the experimental area is divided into multiple geographical intervals. In each partition, non-landslide samples with the same number as positive samples but the lowest similarity correlation strength are selected to balance the number of positive and negative samples and the typical landslide and non-landslide sample characteristics of each prediction unit.

[0082] (4) Graph structure feature learning based on the GraphSAGE model. Taking into account the multi-category association relationship and the strength of the association, the GraphSAGE model is introduced to design the depth of the traversal graph to achieve inductive learning of node information. The node classification task is used to predict the possibility of landslides in all prediction units.

[0083] The specific steps, experimental analysis and evaluation of the present invention are as follows:

[0084] 1. Construction of landslide susceptibility knowledge graph taking into account spatiotemporal characteristics and mechanism knowledge

[0085] (a) Model layer design

[0086] This paper refers to existing ontologies, consults relevant standards and specifications, such as "Basic Terminology of Natural Disaster Management (GB / T26376-2010)", "Natural Disaster Classification and Code (GB / T 28921-2012)", etc., analyzes a large number of scientific and technological literature related to landslides, conducts research and exchanges with experts in the field, and designs the model layer of the knowledge graph (such as the model layer of the knowledge graph) based on the requirements of the landslide spatial prediction task and taking into account the spatiotemporal characteristics of landslide disasters. Figure 2 ) to clarify the conceptual connotations and relationships of various landslide hazards, influencing factors, prediction units, and model methods, and to achieve explicit modeling of domain prior knowledge. The model layer will be explained in detail below.

[0087] The tile in the prediction unit model is the object of landslide spatial prediction. Since the grid unit has a regular form, it is convenient for computer storage and processing. Therefore, it is considered to divide the experimental area into M*N grid units and use them as the prediction unit modeling entity. The design of the prediction unit model conforms to the first and third laws of geography, that is, the closer the distance, the closer the connection, and similar environments may have closer target characteristics.

[0088] The landslide_hazard in the landslide hazard model is a record of historical landslide disaster events in the experimental area. Consider modeling the spatiotemporal characteristics of landslide disaster event records to form a description of the basic attributes of landslides, including scale, disaster time, geographical location, etc. Among them, time as an entity represents the moment when a landslide disaster event occurs. The relationship between time entities mainly includes inclusion and order relationships. By defining the time model, the potential association between disaster events is explored. Space as an entity represents the geographical location where a landslide disaster event occurs. Through the inclusion relationship between spatial entities, the association between historical landslide disaster events and prediction units can be established.

[0089] The prediction_model in the model method mode refers to various model methods available for landslide spatial prediction tasks, including five types of models: random forest, support vector machine, extreme gradient boosting, multi-layer perceptron and graph neural network. The tile-associated landslide disaster events and their environmental condition characteristics are used as the input of the model method, and the spatial probability of landslide occurrence is the output.

[0090] The disaster mechanism model mainly revolves around the three core concepts of "disaster-pregnant environment, inducing factors, and disaster-bearing bodies". The disaster-pregnant environment is the controlling factor of landslide formation. Inducing factors and disaster-bearing bodies do not change the susceptibility of landslides. Therefore, the present invention focuses on the disaster-pregnant environment, analyzes the interaction of influencing condition elements and their impact on landslides, designs semantic logical associations, models disaster mechanism knowledge, and explores the potential associations between elements on this basis, mainly involving six concepts: topography, geological conditions, vegetation conditions, soil conditions, meteorology and hydrology, and human activities (Table 1).

[0091] Table 1 Design of semantic association relationships involved in landslide mechanism model

[0092]

[0093]

[0094] (b) Data layer construction

[0095] According to the Mercator projection rules, a tile coordinate set is created. To balance computational efficiency and spatial accuracy, the zoom level is selected as 18, and a total of 158,426 prediction units in the study area are obtained. The conversion rules are as follows:

[0096]

[0097] In the formula, lon and lat represent the input longitude and latitude coordinates, x and t represent the horizontal and vertical coordinates after slice transformation, and z represents the zoom level.

[0098] Considering that the knowledge in the knowledge graph is based on object expression, the GeoJSON file adopts a storage form based on feature objects. Therefore, the present invention adopts the NDVI numerical algorithm to obtain the spatial distribution of surface vegetation from high-resolution satellite images with a resolution of 0.8m, converts various scale raster and vector data into GeoJSON files, and calculates the slice index list corresponding to each feature object. Specifically, a slice index list can be directly created for discrete attributes, such as soil type. For continuous attributes, the data should be evaluated first, and the attributes should be discretized based on the selection of a suitable ratio to generate a slice index list of discretized attribute intervals, such as altitude. Then, the above processing results are converted into structured triples, including prediction unit triples, mechanism triples, landslide triples, and space-time triples. By combining prediction unit triples, the description of geographical similarity or spatial adjacent associations between prediction units can be achieved, which involves subsequent model methods and will be explained and constructed in subsequent chapters. The landslide triple is a description of the basic attributes of the landslide hazard, such as (landslide_record_10585, hasScale, small-size). The landslide and prediction unit triples are combined to establish the association between the landslide entity and the prediction unit entity, such as (x_217031y_112184z_18, hasLandslideRecord, landslide_record_10682). The mechanism triplet can realize the description of “disaster-prone environment, inducing factors, and disaster-bearing body”, such as (landslide_record_10682, hasTriggerFactor, rainfall), (landslide_record_10682, hasThreatObject, residential_area); among them, the environment triplet is the basis of the spatial prediction model. By combining the environment triplet and the prediction unit triplet, the environment description of different prediction units can be obtained, such as (x_217031y_112184z_18, hasLanduse, cropland). In addition, by combining landslide triples and time triples, we can understand the multiple time nodes and their sequence of frequent landslide disasters at the same landslide point, and grasp the temporal evolution of landslide disaster events, such as (landslide_record_10682, hasLandslideTime, 2004-7-18) and (landslide_record_10683, hasLandslideTime, 2021-8-4) and (2021-8-4, hasAfter, 2004-7-18). Finally, the data instances are mapped to the ontology schema to realize the construction of the landslide susceptibility knowledge graph.

[0099] Table 2 Symbols and related descriptions used in the present invention, where? is a variable

[0100]

[0101]

[0102] 2. Spatial landslide prediction model based on knowledge graph and representation learning

[0103] (a) Graph semantic feature learning based on ComplEx

[0104] There are many asymmetric relationships in the landslide susceptibility knowledge graph. The present invention adopts the ComplEx knowledge graph representation learning model, processes a large number of binary relationships through complex space, realizes effective modeling of symmetric and asymmetric relationships, efficiently learns semantic association information such as prediction units and disaster mechanisms in the knowledge graph, and obtains the embedded representation vectors of all entity nodes. The model scoring function is as follows:

[0105] As shown in formula 3.

[0106]

[0107] In the formula, The embedding vectors representing entities and relations are both d-dimensional complex vectors. h represents the d-dimensional complex vector representation of the head entity, r represents the d-dimensional complex vector representation of the relation, and t represents the d-dimensional complex representation of the tail entity. It represents the conjugate complex number of the tail entity vector, diag(r) represents a diagonal matrix, the diagonal elements are the elements of the relationship vector r, and Re() represents taking the real part of the complex entity.

[0108] (b) Directed weighted graph construction based on knowledge graph

[0109] First, landslides are the result of the combined effects of multiple conditional factors. There are complex interactions between these factors, and the importance of different influencing factors on landslides is inconsistent. The frequency ratio model (FR) can be used to quantitatively calculate the spatial variability relationship between landslide occurrence and various types of influencing factors, reflecting the difference in the possibility of landslides under different conditional factors, as shown in Formula 4. In addition, considering the uncertainty of other semantic associations (such as hierarchical relationships) on landslides, the corresponding edge attribute value is designed to be 1 (Formula 5).

[0110]

[0111] In the formula, FR(Value) indicates the importance of each category of influencing factors on landslides. When FR(Value)>1, it indicates that it is favorable for landslides to occur; when FR(Value)<1, it indicates that it is unfavorable for landslides to occur. j is the number of landslide tiles that appear in a certain classification interval for a certain category element, F is the total number of tiles occupied by all landslides in a certain classification interval for a certain category element, and C j It is the number of tiles of a certain category of elements in a certain classification interval, and C is the total number of tiles in the study area.

[0112]

[0113] Where u and v represent nodes. If FR(value) is included, the attribute value of the edge is the influence degree of each level of influencing indicators calculated by the FR model on the occurrence of landslide, otherwise it is 1.

[0114] Secondly, landslides generally have regional and group characteristics. The present invention uses the dimensionally extended 9-intersection model (DE-9IM) to construct the intersection matrix between the interior, boundary, and exterior of two geometric figures to represent the topological spatial relationship between them (Formula 6). Considering that the influence between adjacent prediction units has a certain uncertainty, for the spatial adjacent relationship, the edge attribute value is designed to be 1, that is, The model is then adjusted through training.

[0115]

[0116] Where RSTR(x,y) represents the spatial topological relationship between the prediction unit entities x and y involved; I, B, and E represent the interior, boundary, and exterior of the prediction unit entity, respectively; and the dim function represents the return value of the dimension of the geometric object.

[0117] Finally, if the geographical configurations of two regions are more similar, they are more likely to show similar landslide characteristics. Based on the results of the ComplEx model embedding, the characteristics of multiple influencing factors are comprehensively considered, and the cosine similarity is used to calculate the geographical similarity strength between different prediction units (Formula 7). In order to balance the number of spatial adjacent relationships, the first 8 node units with the highest similarity association strength with each prediction unit are selected to establish geographical similarity relationships, and the geographical similarity strength is used as the edge attribute value of the similarity relationship (Formula 8).

[0118]

[0119]

[0120] In the formula, xu ·x v represents the dot product of nodes u and v, ||x|| 2 represents the L2 norm (i.e. Euclidean norm) of vector x, Indicates the attribute value of the hasSimilar relationship edge.

[0121] (c) Equal positive and negative landslide sampling based on Louvain algorithm

[0122] From the perspective of graph structure, the topological structure and strength of geographical similarity associations between different prediction units are comprehensively considered based on the Louvain algorithm (Formula 9), and the region is divided into multiple geographical intervals. By selecting non-landslide samples with the same number of positive samples and the lowest similarity association strength in each partition, it is ensured that negative samples can reflect the typical characteristics under non-landslide conditions, rather than randomly selected uncertain samples, and the number of positive and negative samples as well as the typical landslide and non-landslide sample characteristics of various prediction units are balanced.

[0123]

[0124] In the formula, Q represents modularity; A ij is an element of the adjacency matrix. If there is an edge between i and j, then A ij =1, otherwise A ij =0; k i and k j are the degrees of nodes i and j (i.e., the number of edges connected to the node), W is the sum of the total weights of all edges in the network; δ(c i ,c j ) is the Kroneckerdelta function, when nodes i and j belong to the same community, δ(c i ,c j )=1, otherwise δ(c i ,c j )=0.

[0125] (d) Graph structure feature learning based on GraphSAGE

[0126] The positive and negative landslide sample datasets and directed weighted graphs are used as model inputs. The node embedding vector representation of the prediction unit is updated based on the trained GraphSAGE model, and converted into the probability of whether it is a landslide through the softmax activation function. In the construction of the model, in order to reflect the different effects of different neighbor nodes on the embedding of the central node, the weights (edge ​​attribute values) of different neighbor nodes are considered in the process of neighbor information aggregation to enhance the inductive learning of node features. Specifically, in the selection of node neighbors, the complete neighborhood set of each node u is Sampling is performed to obtain neighbor node features; when the neighbor node feature information is aggregated, the edge attribute values ​​of various association relationships are As the weight of message passing, a weighted average aggregator is designed to aggregate the neighborhood feature information of the current node u (Formula 10). Since the self-correlation of each node u has been added to the graph, the neighborhood feature after aggregation is New node features Combining its own feature information with the neighborhood feature information of the current node, the representation of the aggregated information of the next layer of nodes is obtained through the update function (Formula 11), which is used for information transmission to obtain the final embedded representation of the node.

[0127]

[0128] In the formula, represents the neighborhood of the current node u, e uv is the edge attribute value, aggregate is the aggregation function, k is the search depth, is the embedding representation of node u in the j+1 layer neighborhood.

[0129]

[0130] Where σ is the nonlinear activation function ReLU, W′ is the weight matrix of neighbor features at different levels, is the final embedding representation of node u, and b is the offset matrix.

[0131] 3. Experimental analysis and evaluation. Using the spatiotemporal data of crowd intelligence perception of landslide disasters in xx County as experimental data, the proposed method is applied based on the constructed knowledge graph to predict the spatial distribution probability of regional landslides, draw a landslide susceptibility map, and conduct comparative analysis and evaluation with other model methods.

[0132] Taking xx County as an example, the present invention collects data from different sources for preprocessing, and constructs a regional data set mainly composed of environmental data and landslide inventory data and a domain-related prior knowledge base. Based on the aforementioned knowledge graph construction method, the construction of the landslide susceptibility knowledge graph is realized (e.g. Figure 3 ), according to statistics, there are 159,323 entity nodes and 4,113,128 relationship edges in the knowledge graph.

[0133] Table 3 Statistics of knowledge graph construction results

[0134]

[0135] Based on the above-mentioned landslide spatial prediction method based on knowledge graph and representation learning, landslide spatial prediction is carried out in xx county, and compared with the model method integrating knowledge graph representation learning and RF, SVM, XGBoost, and MLP. The method proposed in this invention has good performance in accuracy, precision, recall, F1-Score, AUC and RMSE (such as Figure 4) shows relative advantages in multiple evaluation indicators, can better predict landslide and non-landslide samples, has strong discrimination ability and higher reliability, and has higher advantages and practical value in landslide spatial prediction tasks.

[0136] Table 4 Comparison of evaluation indicators of five model methods

[0137]

[0138] Based on the prediction results, a landslide susceptibility map of xx County was drawn to visualize the potential risk of landslides in different locations in the region. The natural break method was used to divide the landslide susceptibility level into five levels: very low susceptibility, low susceptibility, moderate susceptibility, high susceptibility, and very high susceptibility. The proportion of each level of susceptibility area in the experimental area and the frequency ratio of historical landslides falling into each level of susceptibility area were also counted (see Appendix). Figure 5 ). The results show that the result of Propoesd_model susceptibility mapping not only has the largest area proportion of extremely high susceptibility area, accounting for about 11.56% of the experimental area, but also has the highest historical landslide distribution frequency ratio, which indicates that Propoesd_model can better balance the recognition ability and estimated area size of landslides and is more suitable for landslide prevention and control management.

[0139] The present invention uses geographical environment, landslide inventory data and expert knowledge as the main data sources, constructs a knowledge graph of landslide susceptibility, takes into account the complex mechanism of action between landslides and their influencing factors, the cluster characteristics of spatial distribution, and the geographical similarity under complex heterogeneous environments, and constructs a landslide spatial prediction model method based on knowledge graph and representation learning, realizes regional landslide susceptibility mapping at the county level, and explores the spatial distribution law of potential landslides. Based on the knowledge graph pattern layer, the prior knowledge in the modeling field is displayed, multi-source data is integrated, and the knowledge graph of landslide susceptibility is constructed, which provides data and knowledge interfaces for subsequent spatial prediction models. Combined with the knowledge graph representation learning model (ComplEx) and the graph neural network model (GraphSAGE), the semantic information and deep graph structure information of the graph are further learned, and the inductive learning of diversified landslide and non-landslide features under complex heterogeneous environments is realized, and higher-precision landslide spatial prediction is achieved through node classification tasks. In addition, from the perspective of graph structure, the Louvain algorithm is used to divide the geographical interval, and a more representative set of equal positive and negative landslide samples can be constructed to promote the feature extraction and representation learning of the model. The results of spatial prediction were visualized using a landslide susceptibility map, and it was found that there were five townships in XX County that were more likely to have landslide hazards, thus defining clear regional targets for risk warning of landslide disasters.

[0140] In summary, the present invention proposes a knowledge graph construction method that takes into account the temporal and spatial characteristics and mechanism knowledge. The present invention fully considers the complex disaster mechanism of landslides, and models the knowledge of landslide disaster mechanism by designing a pattern layer to express the influencing factors of different conditions and their interaction with landslides.

[0141] The present invention also proposes a landslide spatial prediction method based on knowledge graph and representation learning. The present invention fully considers the semantic features and graph structure features of the knowledge graph. The knowledge representation learning model (ComplEx) is integrated to capture the semantic information contained in the graph mechanism and the graph neural network model (GraphSAGE) is used to deeply extract the graph structure information between different prediction units and influencing condition elements, and the strength and structure of similar associations in the geographical environment are taken into account. An equal amount of typical positive and negative landslide sample data sets are constructed, and accurate landslide spatial prediction is achieved in a complex heterogeneous environment and within the county.

[0142] The above are preferred embodiments of the present invention. Any changes made according to the technical solution of the present invention, as long as the resulting functions do not exceed the scope of the technical solution of the present invention, belong to the protection scope of the present invention.

Claims

1. A landslide spatial prediction method based on knowledge graph and representation learning, characterized in that: The specific steps include: S1. Construct a knowledge graph of landslide susceptibility that takes into account the mechanism knowledge and spatiotemporal characteristics; design the knowledge graph ontology model, complete the knowledge modeling of landslide disasters that takes into account the spatiotemporal characteristics and mechanism knowledge, and on this basis extract entities from multi-source data, build relationships, and realize the construction of the knowledge graph data layer; S2. Through graph semantic feature learning based on the ComplEx model, directed weighted graph construction based on the landslide susceptibility knowledge graph, equal positive and negative landslide sampling based on the Louvain algorithm, and graph structural feature learning based on the GraphSAGE model, spatial prediction of landslides based on knowledge graph and representation learning is achieved.

2. The landslide spatial prediction method based on knowledge graph and representation learning according to claim 1 is characterized in that: The design knowledge graph ontology model completes the landslide disaster field knowledge modeling taking into account the temporal and spatial characteristics and mechanism knowledge, and is achieved through the design of a model layer, which includes a prediction unit model, a landslide disaster model, a model method model, and a disaster mechanism model.

3. The landslide spatial prediction method based on knowledge graph and representation learning according to claim 2 is characterized in that: The design of the model layer is as follows: The tile in the prediction unit mode is the object of landslide spatial prediction. The experimental area is divided into M*N grid cells, which are modeled as entities as prediction units; The landslide_hazard in the landslide hazard model is a record of historical landslide hazard events in the experimental area. The spatiotemporal characteristics of the landslide hazard event records are modeled to form a description of the basic attributes of the landslide, including scale, disaster time, and geographical location; where time, as an entity, represents the moment when the landslide hazard event occurs, and the relationship between time entities includes inclusion relationship and order relationship; space, as an entity, represents the geographical location where the landslide hazard event occurs, and the relationship between space entities includes inclusion relationship; The prediction_model in the model method mode is a variety of model methods used for landslide spatial prediction tasks, including five types of models: random forest, support vector machine, extreme gradient boosting, multi-layer perceptron and graph neural network; the landslide disaster events associated with tiles and their environmental condition characteristics are used as inputs of the model method, and the spatial probability of landslide occurrence is output; The disaster mechanism model describes the disaster environment, analyzes the interaction of influencing factors and their impact on landslides, and designs semantic logical associations, including the design of semantic association relationships in six major categories: topography, geological conditions, vegetation conditions, soil conditions, meteorology and hydrology, and human activities.

4. The landslide spatial prediction method based on knowledge graph and representation learning according to claim 3 is characterized in that: The semantic association relationship design of the six categories of topography, geological conditions, vegetation conditions, soil conditions, meteorology and hydrology, and human activities is as follows: The definition domain of the topography includes landform type, slope, aspect and altitude; the value range of the influencing relationship of the aspect includes soil type, vegetation coverage and vegetation type, and the value range of the influencing relationship of the altitude includes vegetation type and rainfall; The definition domain of the geological conditions includes stratum lithology and geological structure; the value domain of the influence relationship of stratum lithology includes soil type and vegetation type; The definition domain of the vegetation condition includes vegetation coverage and vegetation type; The definition domain of meteorology and hydrology includes rainfall, water system distribution and terrain humidity index; the value range of the rainfall influence relationship includes vegetation coverage and soil type; The definition domain of the soil conditions includes soil type; The definition domain of the human activities includes land use and road construction; the value domain of the impact relationship between land use and road construction includes soil type, vegetation coverage and vegetation type.

5. The landslide spatial prediction method based on knowledge graph and representation learning according to claim 1 is characterized in that: The construction of the knowledge graph data layer includes the following steps: S11. Create a tile coordinate set according to the Mercator projection rules to obtain the prediction unit in the study area. The conversion rules are as follows In the formula, lon and lat represent the input longitude and latitude coordinates, x and y represent the horizontal and vertical coordinates after the slice transformation, and z represents the zoom level; S12. Use the NDVI numerical algorithm to obtain the spatial distribution of surface vegetation from high-resolution satellite images, convert raster and vector data of various scales into GeoJSON files, and calculate the slice index list corresponding to each feature object; directly create a slice index list for discrete attributes, and for continuous attributes, select the appropriate scale discrete attributes to generate a slice index list of the discretized attribute interval; S13, converting the above processing results into structured triples, including prediction unit triples, mechanism triples, landslide triples, and time-space triples; S14. Map the data instances to the ontology schema to build the landslide susceptibility knowledge graph.

6. The landslide spatial prediction method based on knowledge graph and representation learning according to claim 1 is characterized in that: The graph semantic feature learning based on the ComplEx model is specifically as follows: The ComplEx knowledge graph representation learning model is used to process binary relations through complex space, realize effective modeling of symmetric and asymmetric relations, learn semantic association information in the knowledge graph, and obtain the embedded representation vectors of all entity nodes; the model scoring function is: In the formula, The embedding vectors representing entities and relations are both d-dimensional complex vectors. h is the d-dimensional complex vector representation of the head entity, r is the d-dimensional complex vector representation of the relation, and t is the d-dimensional complex vector representation of the tail entity. It represents the conjugate complex number of the tail entity vector, diag(r) represents a diagonal matrix, the diagonal elements are the elements of the relationship vector r, and Re() represents taking the real part of the complex entity.

7. The landslide spatial prediction method based on knowledge graph and representation learning according to claim 1 is characterized in that: The directed weighted graph construction based on the landslide susceptibility knowledge graph is specifically as follows: Firstly, the frequency ratio model FR is used to quantitatively calculate the spatial variability relationship between landslide occurrence and various types of influencing condition factors: In the formula, FR(Value) indicates the importance of each category of influencing factors on landslides. When FR(Value)>1, it indicates that it is favorable for landslides to occur; when FR(Value)<1, it indicates that it is unfavorable for landslides to occur. j is the number of landslide tiles that appear in a certain classification interval for a certain category element, F is the total number of tiles occupied by all landslides in a certain classification interval for a certain category element, and C j Then is the number of tiles of a certain category of elements in a certain classification interval, and C is the total number of tiles in the study area; Considering the uncertainty of other semantic associations on landslides, the corresponding edge attribute value is designed to be 1: In the formula, u and v represent nodes. If FR(value) is included, the attribute value of the edge is the influence degree of each level of influence index calculated by the FR model on the occurrence of landslide, otherwise it is 1; Secondly, the intersection matrix between the interior, boundary, and exterior of two geometric figures is constructed through the dimensionally extended nine-intersection model DE-9IM to represent the topological spatial relationship between them: In the formula, R STR(x,y) Indicates the spatial topological relationship between the prediction unit entities x and y involved; I, B, and E represent the interior, boundary, and exterior of the prediction unit entity, respectively; the dim function represents the return value of the geometric object dimension; For spatial adjacency, the edge attribute value is designed to be 1, that is, Adjustments are then made through model training; Finally, based on the results of the ComplEx model embedding, the cosine similarity is used to calculate the geographic similarity strength between different prediction units, and the top 8 node units with the highest similarity association strength with each prediction unit are selected to establish a geographic similarity relationship, and the geographic similarity strength is used as the edge attribute value of the similarity relationship: In the formula, x u ·x v represents the dot product of nodes u and v, ||x|| 2 represents the L2 norm of vector x, Indicates the attribute value of the similarity relationship edge.

8. The landslide spatial prediction method based on knowledge graph and representation learning according to claim 1 is characterized in that: The equal positive and negative landslide sampling based on Louvain algorithm is specifically as follows: From the perspective of graph structure, the topological structure and strength of geographical similarity associations between different prediction units are comprehensively considered based on the Louvain algorithm, and the region is divided into multiple geographical intervals. By selecting non-landslide samples with the same number of positive samples and the lowest similarity association strength in each partition, it is ensured that negative samples can reflect the typical characteristics under non-landslide conditions, so as to balance the number of positive and negative samples and the typical landslide and non-landslide sample characteristics of various prediction units. In the formula, Q represents modularity; A ij is an element of the adjacency matrix. If there is an edge between i and j, then A ij =1, otherwise A ij =0; k i and k j are the degrees of nodes i and j respectively, and W is the sum of the total weights of all edges in the network; is the Kroneckerdelta function, when nodes i and j belong to the same community, δ(c i ,c j )=1, otherwise δ(c i ,c j )=0.

9. The landslide spatial prediction method based on knowledge graph and representation learning according to claim 1 is characterized in that: The graph structure feature learning based on the GraphSAGE model is specifically as follows: taking the positive and negative landslide sample data sets and the directed weighted graph as model input, updating the node embedding vector representation of the prediction unit based on the trained GraphSAGE model, and converting it into the probability of whether it is a landslide through the softmax activation function; in the construction of the model, the weights of different neighbor nodes are considered in the process of neighbor information aggregation to enhance the inductive learning of node features.

10. The landslide spatial prediction method based on knowledge graph and representation learning according to claim 9 is characterized in that: The weights of different neighbor nodes are considered in the process of neighbor information aggregation, specifically: In the selection of node neighbors, for each node u, the complete neighborhood set Perform sampling to obtain neighbor node features; When aggregating neighbor node feature information, the edge attribute values ​​of various association relationships are As the weight of message passing, a weighted average aggregator is designed to aggregate the neighborhood feature information of the current node u: In the formula, represents the neighborhood of the current node u, e uv is the edge attribute value, aggregate is the aggregation function, k is the search depth, is the embedding representation of node u in the k+1 layer neighborhood; Since the self-correlation of each node u has been added to the graph, the aggregation of neighborhood features New node features Combining its own feature information with the neighborhood feature information of the current node, the update function is used to obtain the aggregated representation of the next layer of node information for information transmission to obtain the final embedded representation of the node; Where σ is the nonlinear activation function ReLU, W' is the weight matrix of neighbor features at different levels, is the final embedding representation of node u, and b is the offset matrix.

Images