Landslide susceptibility prediction method based on knowledge graph and spatio-temporal feature fusion

Abstract

The present application relates to the technical field of geological disaster monitoring and early warning, and discloses a landslide susceptibility prediction method based on knowledge graph and spatiotemporal feature fusion, which comprises the following steps: extracting a reasoning feature vector from a geological knowledge graph by using a graph neural network; extracting a spatial feature vector from multi-source spatial data by using a convolutional neural network; extracting a time series feature vector from rainfall time series data by using a Transformer model; generating an attention weight based on the reasoning feature vector, and adaptively weighting and fusing the spatial and time series feature vectors by using the weight to obtain a fusion feature vector; and inputting the fusion feature vector into a prediction model to output a landslide occurrence probability. The present application introduces geological prior knowledge in the knowledge graph to guide the fusion process of spatiotemporal features, so that the model can focus on key disaster-causing factor combinations, significantly improving the accuracy and reliability of the prediction, and enhancing the interpretability of the model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of geological disaster monitoring and early warning, in particular to a landslide susceptibility prediction method based on knowledge graph and spatio-temporal feature fusion. BACKGROUND

[0002] Landslide, as a common and destructive geological disaster, poses a serious threat to people's life and property and infrastructure safety. Therefore, conducting landslide susceptibility prediction research, accurately identifying and delineating high-risk areas, has crucial practical significance for disaster prevention and mitigation.

[0003] Landslide susceptibility prediction technology has evolved from traditional statistical models to machine learning models. In recent years, with the development of artificial intelligence technology, deep learning models represented by convolutional neural networks (CNN) and recurrent neural networks (RNN) have been widely applied in this field. These models can automatically extract high-dimensional nonlinear features from multi-source remote sensing data and geographic information data, and show great potential in mining the complex relationship between landslides and predisposing factors, with significantly improved prediction accuracy compared to traditional methods.

[0004] However, existing deep learning prediction methods still face challenges in practical applications. These models usually treat spatial features representing static predisposing environment and temporal features reflecting dynamic triggering process as two independent branches, and only perform simple concatenation at the end of feature extraction. This processing approach largely ignores the complex coupling relationships between different spatio-temporal features, making it difficult to adaptively adjust the attention to different static geological factors (such as lithology) according to dynamically changing triggering conditions (such as rainfall intensity), thereby limiting the further improvement of model prediction performance. In addition, these purely data-driven models do not explicitly incorporate prior knowledge in the field of geology during training, resulting in a possible deviation of their prediction logic from the true disaster-causing mechanism, forming an unexplainable "black box" model, and the prediction reliability in data sparse areas is also limited.

[0005] Therefore, how to deeply combine domain expert knowledge with data-driven models to achieve adaptive and interpretable fusion of heterogeneous spatio-temporal features is a technical problem that needs to be solved in the current landslide susceptibility prediction field. SUMMARY

[0006] Existing landslide susceptibility prediction methods have limitations when processing multi-source heterogeneous data. Specifically, existing technical solutions struggle to effectively combine prior geological knowledge with data-driven deep learning models. Furthermore, the feature fusion methods for static spatial environmental factors and dynamic rainfall-induced factors are relatively simple, typically employing direct stitching without differentiating the contribution of different factors to landslides. This limits the accuracy and reliability of the prediction results.

[0007] Therefore, the present invention aims to provide a landslide susceptibility prediction method that can effectively integrate geological prior knowledge, spatial characteristics and temporal characteristics, so as to solve the technical problem that the prediction accuracy is limited due to insufficient feature fusion mechanism in the prior art.

[0008] To achieve the above objectives, this invention provides a landslide susceptibility prediction method based on the fusion of knowledge graphs and spatiotemporal features, comprising the following steps:

[0009] S1. Extract reasoning feature vectors from the preset geological knowledge map;

[0010] S2. Use convolutional neural networks to extract spatial feature vectors from multi-source spatial data;

[0011] S3. Use the Transformer model to extract temporal feature vectors from rainfall time series data;

[0012] S4. Generate attention weights based on the inference feature vectors, and use the attention weights to perform weighted fusion of the spatial feature vectors and the temporal feature vectors to obtain a fused feature vector;

[0013] S5. Input the fused feature vector into the prediction model and output the probability of landslide occurrence.

[0014] In a specific embodiment, step S1, extracting the inference feature vector, specifically includes: using a graph neural network to perform embedding learning on entities in the geological knowledge graph to obtain entity embedding vectors. Wherein, for any node in the graph... In its first Layer embedding vector The update method is as follows:

[0015] ;

[0016] In the formula, For nodes The set of neighboring nodes, It is an aggregate function. and For the first The learnable weight matrix of the layer, is a nonlinear activation function. Based on the learned final entity embedding vector , the correlation strength between entities is calculated , and combined with the preset state transition path probability , the inference feature vector is generated.

[0017] In a specific embodiment, in step S2, the extraction of the spatial feature vector specifically includes:

[0018] The multi-source spatial data is constructed into a multi-channel input matrix;

[0019] The multi-channel input matrix is processed through at least one convolutional layer and a pooling layer to extract high-dimensional spatial features;

[0020] The high-dimensional spatial features are mapped to the spatial feature vector by a fully connected layer.

[0021] In a specific embodiment, in step S3, the extraction of the time sequence feature vector specifically includes:

[0022] The rainfall time series data is subjected to position encoding processing;

[0023] The position encoding processed data is input into the multi-head self-attention mechanism of the Transformer model to capture the long-term dependence of the sequence, and the time sequence feature vector is generated.

[0024] Preferably, in step S4, the process of generating attention weights specifically includes:

[0025] The inference feature vector is input into a fully connected network, and the output of the network is subjected to softmax function processing to generate the attention weights . The calculation method is:

[0026] ;

[0027] wherein and are learnable parameters of the fully connected network.

[0028] Preferably, in step S4, the process of weighted fusion specifically includes: splicing the spatial feature vector and the time sequence feature vector to obtain a spliced feature vector ; and applying the attention weights to the spliced feature vector by Hadamard product., to obtain a weighted feature; inputting the weighted feature into a full connection layer for nonlinear fusion to obtain the fusion feature vector . The calculation method is as follows:

[0029] ;

[0030] In the formula, represents Hadamard product, represents a full connection layer, is a vector splicing operation.

[0031] In one specific embodiment, in step S5, the prediction model is a random forest model, and the S5 step specifically comprises:

[0032] inputting the fusion feature vector into each decision tree in the random forest model, each decision tree outputs a prediction result ;

[0033] averaging the prediction results of all decision trees to obtain the final landslide occurrence probability . The calculation method is as follows:

[0034] ;

[0035] Preferably, the entities in the geological knowledge graph at least include: topographic factors, lithological factors, rainfall factors and landslide events.

[0036] Preferably, the multi-source spatial data at least include one or more of the following: digital elevation model data, slope data, slope direction data, lithology data, InSAR deformation data.

[0037] In one specific embodiment, the method further comprises the step of: according to a preset probability threshold, dividing the landslide occurrence probability into at least two risk levels of low risk, medium risk and high risk.

[0038] The present application provides a landslide susceptibility prediction method based on knowledge graph and spatio-temporal feature fusion. It has the following advantages:

[0039] 1、The present application extracts a reasoning feature vector from a geological knowledge graph, and uses the vector to generate attention weights for guiding feature fusion, so that the model can quantify and integrate prior knowledge in the field of geology into the prediction process. This way, the fusion of spatial features and temporal features is no longer simply data-driven, but has a guiding nature of geological mechanism, thereby improving the effectiveness of the final fusion feature vector.

[0040] 2、The application adopts convolutional neural network and Transformer model respectively to perform parallel deep feature extraction on static multi-source spatial data and dynamic rainfall time series data. The design can fully and specifically capture the spatial heterogeneity characteristics of landslide disaster environment and the long-term dynamic dependence relationship of key inducing factors (such as rainfall), and provides a comprehensive and deep feature basis for subsequent accurate prediction.

[0041] 3、The application replaces the direct feature splicing method commonly used in the prior art by a knowledge-guided weighted fusion mechanism. The mechanism can dynamically adjust the importance of different source feature dimensions according to geological knowledge, and can focus on integrating information, thereby generating a fusion feature vector with more concentrated information and stronger discrimination, and providing a more reliable input basis for the final landslide susceptibility prediction. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 FIG. 1 is a schematic diagram of geological knowledge graph construction and reasoning feature extraction according to an embodiment of the application;

[0043] Figure 2 FIG. 2 is a structural schematic diagram of a knowledge-guided adaptive feature fusion module according to an embodiment of the application. DETAILED DESCRIPTION

[0044] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the specification of the application. Obviously, the described embodiments are only some of the embodiments of the application, but not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0045] The application provides a landslide susceptibility prediction method based on knowledge graph and spatio-temporal feature fusion. In a specific embodiment, the method is executed on a computing device configured with a specific hardware and software environment.

[0046] The computing device includes but is not limited to at least one central processing unit (CPU), one graphics processing unit (GPU), memory and storage. The method runs in a pre-set software environment, which includes a Linux operating system, a Python3.x programming language environment, and libraries for scientific computing, data processing and model building, such as TensorFlow, PyTorch, GDAL, GeoPandas and Neo4j.

[0047] In the preparation stage of the embodiment, three types of data used for model input need to be obtained and processed first: multi-source spatial data, rainfall time series data and data for constructing a geological knowledge graph.

[0048] Multi-source spatial data are used to characterize the static hazard environment in which landslides occur. These data include, but are not limited to, Digital Elevation Model (DEM) data obtained from the Shuttle Radar Topography Mission (SRTM), as well as slope and aspect data calculated based on DEM data; lithology data extracted from geological maps through vectorization; and surface deformation rate data resolved by Interferometric Synthetic Aperture Radar (InSAR) technology. The above spatial data are stored in raster data formats, such as GeoTIFF format, and have a uniform spatial resolution, such as 30 meters.

[0049] Rainfall time series data as a dynamic triggering factor. This data is derived from ground meteorological stations within the study area or satellite remote sensing products such as the Global Precipitation Measurement (GPM) program, recording rainfall values at specific time intervals (e.g., every hour). The data is stored in the form of records with timestamps.

[0050] The data sources for building the geological knowledge graph include structured data and unstructured data. Structured data includes historical landslide disaster point catalog, table data in regional geological survey reports; unstructured data includes academic papers, exploration reports and other text documents in the field of geological disasters.

[0051] Before feature extraction, a series of preprocessing operations need to be performed on the above raw data to generate model inputs with uniform format and standardization.

[0052] For multi-source spatial data, the preprocessing steps include: projecting all spatial data layers to the same geographic coordinate system (e.g., WGS-84) and projection coordinate system (e.g., UTM); using bilinear interpolation for continuous data such as DEM, slope, and aspect, and using nearest neighbor method for discrete data such as lithology, to ensure that all raster layers have the same spatial resolution and raster alignment; normalizing the values of each layer. For example, using the min-max normalization method to scale the values of each continuous layer to the [0, 1] interval, the calculation formula is:

[0053] ;

[0054] where, is the normalized value, is the original value, and are the minimum and maximum values in the layer, respectively. For categorical data such as lithology, One-Hot Encoding is used to convert it into a multi-dimensional binary vector.

[0055] For rainfall time series data, the preprocessing steps include:

[0056] Linear interpolation is used to fill in the missing data points;

[0057] Based on the original hourly rainfall data, derived time series features such as 24-hour cumulative rainfall, 72-hour cumulative rainfall, etc. are calculated and generated to enrich the expression dimension of time series information.

[0058] Finally, based on the preprocessed data, the samples required for model training are constructed. Each sample corresponds to a grid cell. The grid cell where the landslide has occurred is taken as the positive sample, and from the area where no landslide has occurred in the study area, the same number of grid cells are randomly sampled as negative samples. To ensure the representativeness of the negative samples and improve the discriminant ability of the model, the random sampling can be further limited within a specific area. For example, only in the area where the slope is greater than a certain threshold (such as 5 degrees) but no landslide disaster is recorded, random sampling is carried out to exclude flat areas and other areas that are obviously not prone to landslides, thereby forming a more challenging negative sample set. Each sample contains multi-channel spatial feature data corresponding to the grid cell, as well as rainfall time series data for a period of time before the landslide (e.g., 168 hours).

[0059] Please refer to the attached Figure 1 , attached Figure 1 is a schematic diagram of the construction and reasoning feature extraction of the geological knowledge graph according to an embodiment of the present application. This part corresponds to step S1 in the overall process, and the purpose is to formalize the prior knowledge in the geological field and extract quantized reasoning features that can be used to guide subsequent model training.

[0060] First, the construction of the geological knowledge graph is carried out. In this embodiment, the geological knowledge graph is defined as a directed heterogeneous graph , wherein, is a set of entities, is a set of relationships, is a set of relationship types. The definition of entities is based on the constituent elements of the landslide disaster system, at least including: static predisposing factor entities such as slope, aspect, elevation, lithology, land use type; dynamic inducing factor entities such as hourly rainfall, cumulative rainfall; and disaster event entities such as landslide events, collapse events. The definition of relationships describes the interaction between entities, such as causes (causal relationship), is_located_in (spatial location relationship), has_property (attribute relationship), and positively_correlates_with (positive correlation relationship).

[0061] The knowledge graph construction process is as follows: For the structured data obtained in the aforementioned preparation stage (such as a directory of historical disaster sites), a rule-based template matching method is used to directly extract triples in the form of (entity 1, relation, entity 2). For example, from the disaster site directory, ("a landslide event", is_located_in, "XX region") and ("XX region", has_property, "silty mudstone") are extracted. For unstructured text documents, natural language processing techniques are used. First, a named entity recognition (NER) model is used to identify predefined geological entities in the text. Then, a relation extraction (RE) model is used to determine the relationship type between the identified entity pairs, thereby transforming the unstructured text into structured triple knowledge. All extracted triples are loaded into the graph database to form a complete geological knowledge graph.

[0062] After the knowledge graph is constructed, inference feature vectors are extracted. This process utilizes graph neural networks (GNNs), specifically the GraphSAGE model, to map the symbolic knowledge in the graph into low-dimensional, dense entity embedding vectors. This process updates the vector representation of the central node by iteratively aggregating neighbor node information. For any node in the graph... In its first Layer embedding vector Update using the following formula:

[0063] ;

[0064] In this formula, It is a node In the Layer embedding vector; It is a node The set of neighboring nodes; It is an aggregation function, such as mean aggregation, used to aggregate the embedding vectors of all neighboring nodes. Combined into a single vector; and These are the weight matrices learned during training, used to transform neighbor information and self information; It is a non-linear activation function, such as ReLU. After L layers of iterations, each entity obtains a final embedding vector. , which encodes the semantic information of the entity and its association information with other entities in the graph. To learn meaningful entity embedding vectors, the graph neural network model is trained on an unsupervised or self-supervised task. In the present embodiment, the Link Prediction task is used to train the model. Specifically, a batch of positive triplets (i.e. real existing links) and a batch of negative triplets (non-existing links constructed by randomly replacing the head entity or tail entity) are sampled from the graph, and a loss function (e.g. cross-entropy loss) is defined, the goal of which is to minimize the loss function so that the model can distinguish between real links and non-existing links. All learnable parameters in the model, such as weight matrices , are updated by optimization algorithms such as gradient descent.

[0065] Finally, based on the generated entity embedding vectors, the reasoning feature vector is constructed. This vector is composed of two parts. The first part is the association strength between entities, which quantifies the semantic similarity between key entity pairs in the graph (e.g. a specific lithology and a landslide event). In the present embodiment, the association strength is obtained by calculating the cosine similarity of the two entity embedding vectors:

[0066] ;

[0067] where and are the final embedding vectors of entities and , respectively. The second part is the state transition path probability . This probability is for a predefined disaster-causing path in the graph with clear geological significance, for example, the path from "continuous heavy rainfall" to "soil water saturation" to "high-risk state of landslide". This probability value can be assigned based on historical data statistics or domain expert knowledge. The calculated several key association strength values and state transition path probability values are concatenated to form the final reasoning feature vector . This vector will be used as knowledge-guided information input into the subsequent feature fusion module.

[0068] The function of the spatial feature extraction module is to use a convolutional neural network (CNN) to automatically extract high-dimensional spatial features that can represent the static disaster-prone environment of landslides from multi-source spatial data, corresponding to step S2 in the overall process.

[0069] First, a spatial data matrix for model input is constructed. For each grid cell sample in the study area, a size of a neighborhood window. All the preprocessed multi-source spatial data layers within this window are stacked to form a three-dimensional multi-channel input matrix In this matrix, and represent the height and width of the window, for example, 32; is the number of spatial data sources, i.e., the number of channels, for example, when using DEM, slope, aspect, and lithology (10 dimensions after one-hot encoding) data, = 1 + 1 + 1 + 10 = 13.

[0070] Subsequently, the multi-channel input matrix is input into a pre-defined convolutional neural network. The network is composed of multiple convolutional blocks (Convolutional Block) and a fully connected part in series.

[0071] A typical convolutional block contains a convolutional layer, an activation layer, and a pooling layer in turn. In the convolutional layer, a set of learnable convolutional kernels is used to convolve the output feature maps of the previous layer to extract local spatial patterns. For the output feature maps of the th convolutional block, the calculation process can be represented as:

[0072] ;

[0073] wherein, is the output feature map of the th layer, is the set of convolutional kernels of the th layer, whose size is, for example, 3 x 3; represents the convolution operation; is the corresponding bias term; is a nonlinear activation function. In the present embodiment, the activation function used is the rectified linear unit (ReLU), which is defined as ReLU .

[0074] In the pooling layer, the activated feature maps are down-sampled to reduce the data size, preserve the most significant features, and increase the receptive field of the subsequent convolutional layers. In the present embodiment, max-pooling (Max-Pooling) is used, for example, using a 2 x 2 size pooling window to slide with a step of 2, and taking the maximum value in each window as the output.

[0075] The input matrix The sequence passes through multiple convolutional blocks. As the network deepens, the spatial dimensions (H, W) of the feature map gradually decrease, while the number of channels (i.e., the dimension of the features) usually increases. For example, the network can be configured to include three convolutional blocks with output channel numbers of 32, 64, and 128, respectively.

[0076] After processing through the last convolutional block, the resulting high-dimensional feature map is input to the fully connected portion. First, a flattening operation reshapes the three-dimensional feature map (e.g., with dimensions 4 x 4 x 128) into a one-dimensional long vector. Then, the one-dimensional vector is input to at least one fully connected layer. The fully connected layer performs nonlinear combination and mapping on the flattened features.

[0077] Finally, the output of the last fully connected layer of the network is the spatial feature vector extracted by the spatial feature extraction module . Here, is the preset spatial feature vector dimension, e.g., 128. This vector contains deep abstract features extracted from the original multi-source spatial data to represent the static disaster-prone environment of landslides.

[0078] The function of the time series feature extraction module is to use the Transformer model to extract deep time series features from rainfall time series data that can reflect the dynamic triggering process of landslides, corresponding to step S3 in the overall process.

[0079] First, the time series data sequence for model input is constructed. For each sample, its corresponding preprocessed rainfall time series data is organized into an input matrix . Here, is the length of the time series, e.g., 168 time steps (representing 7 days of hourly data before the landslide); is the feature dimension of each time step, e.g., when each time step contains hourly rainfall and 24-hour cumulative rainfall, = 2.

[0080] Before inputting the Transformer model, first map the dimensional feature vector of each time step to the working dimension of the model (e.g., = 128) through an input embedding layer (a linear transformation layer). Since the Transformer model itself does not have the ability to handle sequence order, a positional encoding (PositionalEncoding) vector needs to be added to the embedded sequence to introduce absolute or relative position information of the time steps. The positional encoding vector is calculated as follows:

[0081] ;

[0082] ;

[0083] in, It is the position index of the time step (from 0 to T-1). It is the dimension index of the encoded vector (from 0 to ...). The position encoding vector is added element-wise to the input embedding vector, and this sum is used as the final input to the Transformer encoder.

[0084] The core of this Transformer model is composed of One (for example, =6) It consists of stacked identical encoder layers. Each encoder layer contains two main sub-layers: a multi-head self-attention mechanism and a position-wise feed-forward network.

[0085] In the multi-head self-attention sublayer, the model calculates the association weights between each time step in the sequence and all other time steps. Specifically, the output sequence of the previous layer is simultaneously linearly transformed three times to obtain the query matrix. Key matrix Sum matrix Then, the attention output is calculated using a scaled dot product attention mechanism. This mechanism is divided into... One (for example, =8) Parallel "heads", each head independently calculates attention, and the calculation formula is:

[0086] ;

[0087] in, It is the first The query, key, and value matrix of the size. It is the dimension of the key vector ( ).Will The outputs of each layer are concatenated and then fused through a linear layer. Residual connections and layer normalization are applied between the output and input of the self-attention sublayer.

[0088] In the positional feed-forward neural network sublayer, the same non-linear transformation is applied to each position vector in the sequence of self-attention sublayer outputs. The network is typically composed of two linear layers and a ReLU activation function. Its output is also connected with its input in a residual connection and layer normalization.

[0089] After the input sequence sequentially passes through the encoder layers, the final output sequence is obtained. In order to obtain a single, fixed-dimensional feature vector representing the entire time sequence, in this embodiment, the vectors of all time steps output by the last encoder layer are subjected to mean pooling, i.e., the element-wise average of all vectors is calculated.

[0090] The vector after mean pooling is the time series feature vector extracted by this module , which deeply encodes the long-term dependencies and dynamic change patterns in the rainfall time series.

[0091] Please refer to the attached Figure 2 , the attached Figure 2 is a structural diagram of a knowledge-guided adaptive feature fusion module according to an embodiment of the present application. The knowledge-guided adaptive feature fusion module is responsible for performing step S4 in the overall process, and its function is to receive the spatial feature vector and the time series feature vector generated by the preceding modules, and use the reasoning feature vector for guidance to effectively fuse the three, and finally generate a high-dimensional fusion feature vector containing multi-source information.

[0092] In this embodiment, the knowledge-guided adaptive feature fusion module receives three input vectors: the spatial feature vector generated by the spatial feature extraction module, the time series feature vector generated by the time series feature extraction module, and the reasoning feature vector generated by the geographic knowledge graph construction and reasoning feature extraction module.

[0093] The fusion process of the module first splices the spatial feature vector and the time series feature vector. Specifically, the and are concatenated along their feature dimensions to form a concatenated feature vector with a dimension of . This vector integrates the original deep features from the static spatial environment and dynamic inducing factors.

[0094] Next, the generation of attention weights is performed in parallel. The reasoning feature vectorto a fully connected network dedicated for generating attention weights, which can be referred to as attention generation network in this embodiment. The network contains a linear layer and a softmax activation function. The linear layer performs a dimension transformation on the input reasoning feature vector, with its output dimension set to be the same as that of the concatenated feature vector, i.e. . Subsequently, the softmax activation function processes the output of the linear layer to generate the final attention weight vector . Its calculation formula is as follows:

[0095] ;

[0096] In this formula, and are the weight matrix and bias term of the linear layer in the attention generation network, both of which are learnable parameters in the model training process. The application of the softmax function ensures that all elements in the output attention weight vector are non-negative and the sum is 1. Each element of the vector corresponds to a weight value of the corresponding dimension in the concatenated feature vector .

[0097] After obtaining the attention weight vector , a weighting operation is performed. The attention weight vector is applied to the concatenated feature vector using element-wise product, i.e., element-by-element multiplication. The output of this operation is a weighted feature vector, in which each dimension of the concatenated feature vector is scaled by its corresponding attention weight.

[0098] Finally, a nonlinear fusion is performed on the weighted feature vector. The vector is input to another fully connected layer (referred to as fusion layer here) for the final feature integration and mapping. The fusion layer performs a nonlinear transformation on the weighted feature to capture the complex interaction between different feature dimensions. The output of this layer is the final fusion feature vector generated by this module, where is the preset final fusion feature dimension. Its calculation formula is as follows:

[0099] ;

[0100] where, denotes the element-wise product operation, denotes the nonlinear transformation operation of the fusion layer. The output vector integrates the spatial and temporal information guided by knowledge and weighted differently, which will serve as the final input for the next stage of the prediction model.

[0101] The function of this module is to receive the fused feature vector generated by the previous module and calculate the final landslide occurrence probability based on this vector, corresponding to step S5 in the overall process. In this embodiment, a Random Forest model is used as the prediction model for performing the probability prediction.

[0102] The Random Forest model is an ensemble learning model composed of multiple independent decision trees. In the model construction and training phase, multiple sub-training sets are generated from the original training sample set through the Bootstrap method, and each decision tree is trained based on a sub-training set. At the same time, when each decision tree node is split, only a subset of all features is randomly selected to find the optimal split point. The main hyperparameters of this model include: the number of decision trees , for example, set to 100; the maximum depth of each decision tree ; and the minimum number of samples required for node splitting .

[0103] In the prediction phase, the fused feature vector generated by the knowledge-guided adaptive feature fusion module is input into the trained Random Forest model. This vector is simultaneously sent to all decision trees in the model. Inside each decision tree, the feature vector starts from the root node and is judged layer by layer according to the feature splitting rule determined in the training phase (i.e., which feature dimension to select and what threshold to use), along a specific path until it reaches a leaf node. Each leaf node corresponds to a preliminary prediction probability based on the statistical analysis of the training samples it contains.

[0104] The final output of the model is the aggregation of the prediction results of all decision trees. Specifically, the preliminary prediction probabilities output by all decision trees are arithmetically averaged to obtain the final landslide occurrence probability , which is a continuous value between 0 and 1. Its calculation formula is:

[0105] ;

[0106] where is the prediction probability output by the th decision tree.

[0107] Finally, this embodiment also includes a risk level division step. According to one or more pre-set probability thresholds, the calculated landslide occurrence probability is mapped to a discrete risk level. For example, two thresholds and (e.g., = 0.3, = 0.7). When = 0.3, the grid cell is determined to be low risk; when = 0.7, it is determined to be medium risk; and when = 0.7, it is determined to be high risk. Finally, the risk level of the grid cell is output as the final result of this prediction.

[0108] To further illustrate the technical solutions of the present application, the complete execution process of the present method will be described below by taking a specific application scenario as an example.

[0109] This embodiment takes the landslide susceptibility of a specific grid cell A in a research area at a specific time point as an example. The grid cell A has its own geographical attributes and historical environmental data.

[0110] First, data preparation and input construction are performed. Multi-source spatial data related to the grid cell A is obtained, a neighborhood window of size 32x32 is constructed centered on the grid cell A, and data of 13 channels including DEM, slope, aspect, and lithology (after one-hot encoding) are stacked to form a multi-channel spatial input matrix of 32x32x13. At the same time, 168 hours of rainfall data of the grid cell A before the time point are obtained, and a time series input matrix of length 168 and feature dimension 2 (e.g., hourly rainfall, 24-hour cumulative rainfall) is constructed.

[0111] Next, the three feature extraction modules work in parallel. In step S1, the extraction of the inference feature vector is performed. According to the attributes of the grid cell A (e.g., its lithology is "silty mudstone" and the slope value is 35 degrees), relevant entities are retrieved from the constructed geological knowledge graph. The pre-trained embedding vectors of the "silty mudstone" entity and the "landslide event" entity are obtained, and the cosine similarity between them is calculated. At the same time, in combination with the regional rainfall characteristics, the preset disaster-causing path probability (e.g., the path probability from "continuous heavy rainfall" to "high risk of landslide") is obtained. These calculated and obtained values are spliced to form the inference feature vector corresponding to the grid cell A.

[0112] In step S2, spatial feature extraction is performed. The 32x32x13 multi-channel spatial input matrix constructed for the grid cell A is input into the pre-trained convolutional neural network model. The data passes through multiple convolutional layers and pooling layers in turn, and finally outputs a spatial feature vector of dimension 128 through a fully connected layer.

[0113] In step S3, the time series feature extraction is performed. The time series input matrix associated with the grid cell A is input into the pre-trained Transformer model after position encoding. The data is processed in the multi-head self-attention mechanism and the feed-forward network of the model, and finally an 128-dimensional time series feature vector is output through the average pooling operation .

[0114] After the feature extraction is completed, in step S4, the three feature vectors are sent into the knowledge-guided adaptive feature fusion module. First, the 128-dimensional and the 128-dimensional are spliced into a 256-dimensional spliced feature vector. At the same time, the is input into the attention generation network to generate a 256-dimensional attention weight vector . Then, the is applied to the spliced feature vector through the Hadamard product, and the result is input into the fusion layer to finally output a fusion feature vector, for example, a 64-dimensional fusion feature vector .

[0115] Finally, in step S5, the probability prediction is performed. The generated 64-dimensional fusion feature vector is input into the trained random forest prediction model. The vector is evaluated in 100 decision trees in the model respectively, and the respective prediction probabilities are output. The output probabilities of all decision trees are arithmetically averaged to obtain the final landslide occurrence probability, for example, the calculation result is = 0.82. According to the preset risk level division threshold (for example, the high risk threshold is 0.7), since 0.82 ≥ 0.7, the landslide risk of the grid cell A at the time point is determined as “high risk”. The result is the final output of the method for this specific grid cell.

Claims

1. A landslide susceptibility prediction method based on the fusion of knowledge graph and spatiotemporal features, characterized in that, Includes the following steps: S1. Extract reasoning feature vectors from the preset geological knowledge map; In step S1, the extraction of the inference feature vector specifically includes: Graph neural networks are used to learn the embedding of entities in the geological knowledge graph to obtain entity embedding vectors. The association strength between entities is calculated based on the entity embedding vector, and the inference feature vector is generated by combining the preset state transition path probability. S2. Use convolutional neural networks to extract spatial feature vectors from multi-source spatial data; S3. Use the Transformer model to extract temporal feature vectors from rainfall time series data; S4. Generate attention weights based on the inference feature vectors, and use the attention weights to perform weighted fusion of the spatial feature vectors and the temporal feature vectors to obtain a fused feature vector; S5. Input the fused feature vector into the prediction model and output the probability of landslide occurrence.

2. The landslide susceptibility prediction method based on knowledge graph and spatiotemporal feature fusion according to claim 1, characterized in that, In step S2, the extraction of spatial feature vectors specifically includes: The multi-source spatial data is constructed into a multi-channel input matrix; The multi-channel input matrix is ​​processed by at least one convolutional layer and a pooling layer to extract high-dimensional spatial features; The high-dimensional spatial features are mapped to spatial feature vectors through a fully connected layer.

3. The landslide susceptibility prediction method based on knowledge graph and spatiotemporal feature fusion according to claim 1, characterized in that, In step S3, the extraction of the temporal feature vector specifically includes: The rainfall time series data is subjected to location encoding processing; The position-encoded data is input into the multi-head self-attention mechanism of the Transformer model to capture long-term dependencies in the sequence and generate the temporal feature vector.

4. The landslide susceptibility prediction method based on knowledge graph and spatiotemporal feature fusion according to claim 1, characterized in that, In step S4, the process of generating attention weights specifically involves: The inference feature vector is input into a fully connected network, and the output of the fully connected network is processed by a softmax function to generate the attention weights.

5. The landslide susceptibility prediction method based on knowledge graph and spatiotemporal feature fusion according to claim 1, characterized in that, In step S4, the weighted fusion process specifically involves: The spatial feature vector and the temporal feature vector are concatenated to obtain the concatenated feature vector; The attention weights are applied to the concatenated feature vector using the Hadamard product to obtain weighted features; The weighted features are input into a fully connected layer for nonlinear fusion to obtain the fused feature vector.

6. The landslide susceptibility prediction method based on knowledge graph and spatiotemporal feature fusion according to claim 1, characterized in that, In step S5, the prediction model is a random forest model, and step S5 specifically includes: The fused feature vector is input into multiple decision trees of the random forest model; The final probability of the landslide is obtained by averaging the predictions of all decision trees.

7. The landslide susceptibility prediction method based on knowledge graph and spatiotemporal feature fusion according to claim 1, characterized in that, In step S1, the entities in the geological knowledge graph include at least: topographic factors, lithological factors, rainfall factors, and landslide events.

8. The landslide susceptibility prediction method based on knowledge graph and spatiotemporal feature fusion according to claim 1, characterized in that, In step S2, the multi-source spatial data includes at least one or more of the following: digital elevation model data, slope data, aspect data, lithology data, and InSAR deformation data.

9. The landslide susceptibility prediction method based on knowledge graph and spatiotemporal feature fusion according to claim 1, characterized in that, It also includes the step of: classifying the probability of landslide occurrence into at least two risk levels among low risk, medium risk and high risk, based on a preset probability threshold.

Images