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

Abstract

The invention relates to the technical field of geological disaster monitoring and early warning, and discloses a landslide susceptibility prediction method based on knowledge map and spatial-temporal feature fusion, and the method comprises the steps: extracting an inference feature vector from a geological knowledge map through a map neural network; extracting a spatial feature vector from the multi-source spatial data by using a convolutional neural network; extracting a time sequence feature vector from the rainfall time sequence data by using a Transform model; generating an attention weight based on the reasoning feature vector, and performing adaptive weighted fusion on the space and time sequence feature vectors by using the weight to obtain a fused feature vector; and inputting the fusion feature vector into the prediction model, and outputting the landslide occurrence probability. The geological priori knowledge in the knowledge graph is introduced to guide the fusion process of the spatial-temporal characteristics, so that the model can focus on the key disaster-inducing factor combination, the accuracy and reliability of prediction are remarkably improved, and the interpretability of the model is enhanced at the same time.

Description

Technical Field

[0001] This invention relates to the field of geological disaster monitoring and early warning technology, specifically a landslide susceptibility prediction method based on the fusion of knowledge graphs and spatiotemporal features. Background Technology

[0002] Landslides, as a common and highly destructive geological hazard, pose a serious threat to people's lives, property, and infrastructure safety. Therefore, conducting landslide susceptibility prediction research and accurately identifying and delineating high-risk areas is of vital 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, such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs), have been widely used in this field. These models can automatically extract high-dimensional nonlinear features from multi-source remote sensing data and geographic information data, showing great potential in uncovering the complex relationship between landslides and disaster-causing factors, and significantly improving prediction accuracy compared to traditional methods.

[0004] However, existing deep learning prediction methods still face challenges in practical applications. These models typically treat spatial features representing the static disaster-causing environment and temporal features reflecting dynamic triggering processes as two independent branches, fusing them at the end of feature extraction through simple concatenation. This approach largely ignores the complex coupling relationships between features in different spatiotemporal dimensions, making it difficult to adaptively adjust the attention given to different static geological factors (such as lithology) according to dynamically changing triggering conditions (such as rainfall intensity), thus limiting further improvements in model prediction performance. Furthermore, these purely data-driven models do not explicitly incorporate prior knowledge from the geological field during training, leading to potential deviations between their prediction logic and the actual disaster-causing mechanisms, forming unexplainable "black box" models, which also limit prediction reliability in data-sparse regions.

[0005] Therefore, how to deeply integrate domain expert knowledge with data-driven models to achieve adaptive and interpretable fusion of heterogeneous spatiotemporal characteristics is a technical challenge that urgently needs to be solved in the field of landslide susceptibility prediction. Summary of the Invention

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

[0009] 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: 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, This is a non-linear activation function. Based on the learned final entity embedding vector... Calculate the correlation strength between entities And combined with the preset state transition path probability Generate the inference feature vector .

[0010] In a specific embodiment, step S2, extracting the spatial feature vector, 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. .

[0011] In a specific embodiment, step S3, extracting 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. .

[0012] Preferably, in step S4, the process of generating attention weights specifically involves: The inference feature vector The input is fed into a fully connected network, and the output of that network is processed by a softmax function to generate the attention weights. The calculation method is as follows: ; In the formula, and These are the learnable parameters for a fully connected network.

[0013] Preferably, in step S4, the weighted fusion process specifically involves: converting the spatial feature vector... and the time-series feature vector The concatenation is performed to obtain the concatenated feature vector. The attention weights are then calculated using the Hadamard product. Applied to the concatenated feature vector The weighted features are obtained; these weighted features are then input into a fully connected layer for nonlinear fusion to obtain the fused feature vector. The calculation method is as follows: ; In the formula, It represents the Hadamah accumulation. Indicates a fully connected layer. This is a vector concatenation operation.

[0014] In a specific embodiment, in step S5, the prediction model is a random forest model, and step S5 specifically involves: The fused feature vector Input into the random forest model Each decision tree Output a prediction result ; The final probability of landslide occurrence is obtained by averaging the predictions from all decision trees. The calculation method is as follows: ; Preferably, the entities in the geological knowledge graph include at least: topographic factors, lithological factors, rainfall factors, and landslide events.

[0015] Preferably, the multi-source spatial data includes at least one or more of the following: digital elevation model data, slope data, aspect data, lithological data, and InSAR deformation data.

[0016] In one specific embodiment, the method further includes the step of: classifying the probability of landslide occurrence into at least two risk levels among low risk, medium risk, and high risk according to a preset probability threshold.

[0017] This invention provides a landslide susceptibility prediction method based on the fusion of knowledge graphs and spatiotemporal features. It has the following beneficial effects: 1. This invention extracts inference feature vectors from geological knowledge graphs and uses these vectors to generate attention weights that guide feature fusion. This enables the model to quantify prior knowledge in the field of geology and integrate it into the prediction process. This approach makes the fusion of spatial and temporal features no longer simply data-driven, but rather guided by geological mechanisms, thereby improving the effectiveness of the final fused feature vectors.

[0018] 2. This invention employs convolutional neural networks and Transformer models to perform parallel deep feature extraction on static multi-source spatial data and dynamic rainfall time-series data, respectively. This design can fully and specifically capture the spatial heterogeneity of the landslide-prone environment, as well as the long-term dynamic dependencies of key triggering factors (such as rainfall), providing a comprehensive and in-depth feature foundation for subsequent accurate prediction.

[0019] 3. This invention replaces the commonly used direct feature splicing method in the prior art with a knowledge-guided weighted fusion mechanism. This mechanism can dynamically adjust the importance of feature dimensions from different sources according to geological knowledge, and integrate information with emphasis, thereby generating a fusion feature vector with more concentrated information and stronger discriminative power, providing a more reliable input basis for the final landslide susceptibility prediction. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating a landslide susceptibility prediction method according to an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the geological knowledge graph construction and reasoning feature extraction according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a knowledge-guided adaptive feature fusion module according to an embodiment of the present invention. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please refer to the appendix. Figure 1 , attached Figure 1 This is a flowchart illustrating a landslide susceptibility prediction method according to an embodiment of the present invention. The present invention provides a landslide susceptibility prediction method based on the fusion of knowledge graphs and spatiotemporal features. In a specific embodiment, this method is executed on a computing device configured with a specific hardware and software environment.

[0023] 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-defined software environment, including a Linux operating system, a Python 3.x programming language environment, and libraries for scientific computing, data processing, and model building, such as TensorFlow, PyTorch, GDAL, GeoPandas, and Neo4j.

[0024] In the preparation phase of this embodiment, it is first necessary to acquire and process three types of data for model input: multi-source spatial data, rainfall time series data, and data for constructing a geological knowledge graph.

[0025] Multi-source spatial data are used to characterize the static environment in which landslides occur. This data includes, but is not limited to: digital elevation model (DEM) data acquired from the Space Shuttle Radar Topography Mission (SRTM), and slope and aspect data calculated based on the DEM data; lithological data extracted from geological maps using vectorization; and surface deformation rate data calculated using Interferometric Synthetic Aperture Radar (InSAR) technology. All of the above spatial data are stored in raster data formats, such as GeoTIFF, and have a uniform spatial resolution, such as 30 meters.

[0026] Rainfall time-series data served as a dynamic inducing factor. This data, derived from surface meteorological stations within the study area or satellite remote sensing products such as the Global Precipitation Measurement Project (GPM), recorded rainfall values ​​at specific time intervals (e.g., hourly). The data was stored as timestamped records.

[0027] The data sources used to construct geological knowledge graphs include both structured and unstructured data. Structured data includes directories of historical landslide hazard sites and tabular data from regional geological survey reports; unstructured data includes text documents such as academic papers and exploration reports in the field of geological hazards.

[0028] Before feature extraction, a series of preprocessing operations need to be performed on the above raw data to generate model input with a uniform and standardized format.

[0029] 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 projected coordinate system (e.g., UTM); resampling continuous data such as DEM, slope, and aspect using bilinear interpolation, and resampling discrete data such as lithology using nearest neighbor method to ensure that all raster layers have the same spatial resolution and raster alignment; and normalizing the values ​​of each layer. For example, using the min-max normalization method, the values ​​of each continuous layer are scaled to the [0,1] interval, and the calculation formula is as follows: ; in, The value is the normalized value. These are the original values. and These represent the minimum and maximum values ​​in the layer, respectively. For categorical data such as lithology, one-hot encoding is used to convert them into multidimensional binary vectors.

[0030] For rainfall time series data, the preprocessing steps include: Missing data points are filled using linear interpolation. Based on the original hourly rainfall data, derived time-series features are calculated and generated, such as the cumulative rainfall in the previous 24 hours and the cumulative rainfall in the previous 72 hours, to enrich the expression dimensions of time-series information.

[0031] Finally, samples are constructed based on the preprocessed data to train the model. Each sample corresponds to a raster cell. Raster cells where landslides have occurred are used as positive samples, and the same number of raster cells are randomly sampled from areas within the study area that have not experienced landslides, selecting the same number as the positive samples as negative samples. To ensure the representativeness of the negative samples and improve the model's discriminative ability, the random sampling can be further limited to a specific area. For example, random sampling can be performed only in areas with a slope greater than a certain threshold (e.g., 5 degrees) but where no landslides have been recorded, excluding areas that are flat or obviously less prone to landslides, thus creating a more challenging negative sample set. Each sample contains multi-channel spatial feature data corresponding to the raster cell, as well as rainfall time-series data for a period of time before the landslide (e.g., the previous 168 hours).

[0032] Please refer to the appendix. Figure 2 , attached Figure 2 This is a schematic diagram of geological knowledge graph construction and inference feature extraction according to an embodiment of the present invention. This part corresponds to step S1 in the overall process, the purpose of which is to formally express prior knowledge in the geological field and extract quantitative inference features from it that can be used to guide subsequent model training.

[0033] First, a geological knowledge graph is constructed. In this embodiment, the geological knowledge graph is defined as a directed anisomorphic graph. ,in, It is a collection of entities. It is a set of relations. It is a collection of relation types. The definition of an entity is based on the constituent elements of a landslide hazard system, including at least: static hazard-inducing factor entities, such as slope, aspect, elevation, lithology, and land use type; dynamic triggering factor entities, such as hourly rainfall and cumulative rainfall; and hazard event entities, such as landslide events and collapse events. The definition of a relation describes the interaction between entities, such as causes, is_located_in (spatial location relationship), has_property (attribute relationship), and positively_correlates_with (positive correlation relationship).

[0034] 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.

[0035] 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: 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. This vector encodes the semantic information of the entity and its associations with other entities in the graph in a low-dimensional space. To learn meaningful entity embedding vectors, the graph neural network model is trained on an unsupervised or self-supervised task. In this embodiment, the model is trained using a link prediction task. Specifically, a batch of positive triples (i.e., real links) and a batch of negative triples (non-existent links constructed by randomly replacing the head or tail entities) are sampled from the graph, and a loss function (e.g., cross-entropy loss) is defined. The goal is to minimize this loss function so that the model can distinguish between real links and non-existent links. All learnable parameters in the model, such as the weight matrix... All are updated using optimization algorithms such as gradient descent.

[0036] Finally, based on the generated entity embedding vectors, inference feature vectors are constructed. This vector consists of two parts. The first part is the correlation strength between entities. This is used to quantify the semantic similarity between key entity pairs (e.g., specific lithology and landslide events) in a map. In this embodiment, the association strength is obtained by calculating the cosine similarity between the two entity embedding vectors: in, and They are entities and The final embedding vector. The second part is the state transition path probability. This probability is applied to predefined, geologically significant disaster-causing paths within the map, such as the path from "continuous heavy rainfall" to "soil moisture saturation" and then to "high-risk landslide state." This probability value can be assigned based on historical data statistics or domain expert knowledge. Concatenating the calculated key correlation strength values ​​and state transition path probability values ​​constitutes the final inference feature vector. This vector will serve as knowledge-guiding information and will be input into the subsequent feature fusion module.

[0037] The function of the spatial feature extraction module is to automatically extract high-dimensional spatial features that can characterize the static disaster-causing environment of landslides from multi-source spatial data using convolutional neural networks (CNN), which corresponds to step S2 in the overall process.

[0038] First, a spatial data matrix for model input is constructed. For each raster cell sample in the study area, a cutoff matrix of size [size missing] is extracted centered on that cell. A neighborhood window is defined. All preprocessed multi-source spatial data layers within this window are stacked to form a three-dimensional multi-channel input matrix. In this matrix, and These represent the height and width of the window, for example, 32; This refers to the number of spatial data sources, i.e., the number of channels. For example, when using DEM, slope, aspect, and lithology data (which are 10 dimensions after uniquely heat-coded), =1+1+1+10=13.

[0039] Subsequently, the multi-channel input matrix The input is fed into a pre-defined convolutional neural network. This network consists of multiple convolutional blocks and a fully connected part connected in series.

[0040] A typical convolutional block contains, in sequence, a convolutional layer, an activation layer, and a pooling layer. Within the convolutional layer, a set of learnable convolutional kernels are used to convolve the output feature map of the previous layer to extract local spatial patterns. For the Output feature map of each convolutional block The calculation process can be expressed as follows: ; in, It is the first The output feature map of the layer, It is the first The set of convolutional kernels for the layer, with a size of, for example, 3×3; Represents convolution operation; It is the corresponding bias term; It is a non-linear activation function. In this embodiment, the activation function used is the Modified Linear Unit (ReLU), which is defined as ReLU. .

[0041] In the pooling layer, the activated feature maps are downsampled to reduce the data size, retain the most salient features, and increase the receptive field of subsequent convolutional layers. In this embodiment, max-pooling is used, for example, using a 2×2 pooling window that slides with a stride of 2, and taking the maximum value within each window as the output.

[0042] Input matrix The network passes through multiple convolutional blocks sequentially. As the network deepens, the spatial dimensions (H, W) of the feature maps gradually decrease, while the number of channels (i.e., the dimension of the features) typically increases. For example, the network can be configured to contain three convolutional blocks with output channels of 32, 64, and 128, respectively.

[0043] After the final convolutional block, the resulting high-dimensional feature map is input to the fully connected layer. First, a flattening operation reshapes the 3D feature map (e.g., 4×4×128) into a one-dimensional long vector. Then, this one-dimensional vector is input to at least one fully connected layer. The fully connected layer performs non-linear combination and mapping on the flattened features.

[0044] Finally, the output of the last fully connected layer of the network is the spatial feature vector extracted by the spatial feature extraction module. .in, It is a preset spatial feature vector dimension, for example, 128. This vector contains deep abstract features extracted from the original multi-source spatial data to characterize the static disaster-causing environment of landslides.

[0045] The function of the time series feature extraction module is to use the Transformer model to extract depth time series features from rainfall time series data that can reflect the dynamic induction process of landslides, corresponding to step S3 in the overall process.

[0046] First, a time-series data sequence is constructed for the model input. For each sample, its corresponding preprocessed rainfall time-series data is organized into an input matrix. .in, It is the length of the time series, for example, 168 time steps (representing hourly data for the 7 days prior to the landslide). These are the feature dimensions for each time step. For example, when each time step includes two features: hourly rainfall and 24-hour cumulative rainfall, =2.

[0047] Before inputting into the Transformer model, each time step is first processed through an input embedding layer (a linear transformation layer). Mapping 1D feature vectors to the working dimension of the model (For example, =128). Since the Transformer model itself does not have the ability to process sequence order, it is necessary to add positional encoding vectors to the embedded sequence. This is to incorporate absolute or relative position information for each time step. The position encoding vector is calculated as follows: ; ; 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.

[0048] 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.

[0049] 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: ; 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.

[0050] In the position-forward neural network sublayer, the same nonlinear transformation is performed on each position vector in the sequence of outputs from the self-attention sublayer. This network typically consists of two linear layers and a ReLU activation function. Its output is also residually connected to its input and normalized.

[0051] The input sequence passes through this sequentially After several encoder layers, the final output sequence is obtained. In order to obtain a single, fixed-dimensional feature vector representing the entire time series, in this embodiment, mean pooling is performed on the vectors of all time steps output by the last encoder layer, that is, the element-wise average of all vectors is calculated.

[0052] The vector obtained after average pooling is the temporal feature vector extracted by this module. ,in This vector deeply encodes the long-term dependencies and dynamic patterns of change in the rainfall time series.

[0053] Please refer to the appendix. Figure 3 , attached Figure 3 This is a schematic diagram of the structure of a knowledge-guided adaptive feature fusion module according to an embodiment of the present invention. The knowledge-guided adaptive feature fusion module is responsible for executing step S4 in the overall process. Its function is to receive the spatial feature vector and temporal feature vector generated by the preceding module, and use the inference feature vector for guidance to effectively fuse the three, and finally generate a high-dimensional fused feature vector containing multi-source information.

[0054] In this embodiment, the knowledge-guided adaptive feature fusion module receives three input vectors: a spatial feature vector generated by the spatial feature extraction module, and a 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 inference feature vector generated by the geographic knowledge graph construction and inference feature extraction module. .

[0055] The module fusion process first involves concatenating spatial feature vectors and temporal feature vectors. Specifically, it involves... and Perform a concatenation operation along its feature dimensions to form a concatenation array with dimension 1. spliced ​​feature vectors This vector integrates the original depth features from the static spatial environment and dynamic triggering factors.

[0056] Next, attention weights are generated in parallel. The inference feature vectors are then processed. The input is fed into a fully connected network dedicated to generating attention weights, which in this embodiment can be referred to as an attention generation network. This network contains a linear layer and a softmax activation function. The linear layer performs a dimensionality transformation on the input inference feature vector, and its output dimension is set to be the same as the dimension 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. The calculation formula is as follows: ; In this formula, and These are the weight matrix and bias term of the linear layer in the attention generation network, respectively, both of which are learnable parameters during model training. The application of the softmax function ensures that the output attention weight vector... All elements in the vector are non-negative and sum to 1. Each element of this vector corresponds to a concatenated feature vector. A weight value for the corresponding dimension.

[0057] Obtaining the attention weight vector Then, a weighting operation is performed. The Hadamard product (element-wise product) is used, which multiplies the attention weight vector element by element. Applied to concatenating feature vectors The output of this operation is a weighted feature vector, where each dimension of the concatenated feature vector is scaled by its corresponding attention weight.

[0058] Finally, the weighted feature vectors are non-linearly fused. This vector is then input into another fully connected layer (referred to here as the fusion layer) for final feature integration and mapping. This fusion layer performs a non-linear transformation on the weighted features to capture the complex interactions between different feature dimensions. The output of this layer is the final fused feature vector generated by this module. ,in This is the preset final fusion feature dimension. Its calculation formula is: ; in, This represents the Hadamard product operation. This represents the nonlinear transformation operation of the fusion layer. The output vector... The integrated spatial and temporal information, guided by knowledge and differentiated by weighting, will serve as the final input for the next stage of the prediction model.

[0059] This module receives the fused feature vector generated by the previous module and calculates the final probability of landslide occurrence 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 to perform this probability prediction.

[0060] Random forest is an ensemble learning model consisting of multiple independent decision trees. During model building and training, multiple subsets of training data are generated from the original training set using a bootstrap method, and each decision tree is trained based on one of these subsets. Furthermore, when splitting at a node of each decision tree, 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 a node to split. .

[0061] During the prediction phase, the fused feature vector generated by the knowledge-guided adaptive feature fusion module will be... The vector is fed into the already trained random forest model. Each decision tree consists of several nodes. Within each decision tree, the feature vector starts from the root node and proceeds layer by layer according to the feature splitting rules determined during the training phase (i.e., which feature dimension to choose and what threshold to use), following a specific path until a leaf node is reached. Each leaf node corresponds to an initial predicted probability derived from the statistics of the training samples it contains.

[0062] The model's final output is an aggregation of the predictions from all decision trees. Specifically, it aggregates all... The final probability of landslide occurrence is obtained by arithmetically averaging the initial predicted probabilities output by each decision tree. This value is a continuous value between 0 and 1. Its calculation formula is: ; In the formula, It is the first The predicted probabilities output by each decision tree.

[0063] Finally, this embodiment also includes a risk level classification step. Based on one or more preset probability thresholds, the calculated landslide occurrence probability is... This is mapped to discrete risk levels. For example, two thresholds can be set. and (For example, =0.3, =0.7). When When the grid cell is deemed low-risk, it is considered low-risk. When, it is judged as medium risk; when If the risk level is high, the grid cell is identified as such. The final risk level of the grid cell is then output as the final result of this prediction.

[0064] To further illustrate the technical solution of the present invention, the complete execution flow of this method will be described below using a specific application scenario as an example.

[0065] Please refer to the appendix. Figure 1 This embodiment aims to predict the location of a specific grid cell A within the study area at a specific time point. Taking landslide susceptibility as an example, this raster cell A has its own geographical attributes and historical environmental data.

[0066] First, data preparation and input construction are performed. Multi-source spatial data related to raster cell A are acquired. A 32×32 neighborhood window is constructed centered on raster cell A, and data from 13 channels, including DEM, slope, aspect, and lithology (after unique heat encoding), are stacked to form a 32×32×13 multi-channel spatial input matrix. Simultaneously, data from raster cell A at time points are acquired. The previous 168 hours of rainfall data were used to construct a time series input matrix with a length of 168 and a feature dimension of 2 (e.g., hourly rainfall, 24-hour cumulative rainfall).

[0067] Next, the three feature extraction modules work in parallel. In step S1, the inference feature vector is extracted. Based on the attributes of grid cell A (e.g., its lithology is "silty mudstone" and its slope 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, combined with regional rainfall characteristics, the preset disaster-causing path probability is obtained (e.g., the path probability from "continuous heavy rainfall" to "high risk of landslide"). These calculated and obtained values ​​are concatenated to form the inference feature vector corresponding to grid cell A. .

[0068] In step S2, spatial feature extraction is performed. The 32×32×13 multi-channel spatial input matrix constructed for grid cell A is input into a pre-trained convolutional neural network model. The data passes through multiple convolutional and pooling layers sequentially, and finally outputs a 128-dimensional spatial feature vector through a fully connected layer. .

[0069] In step S3, temporal feature extraction is performed. The 168-dimensional time series input matrix associated with grid cell A is processed by position encoding and then input into the pre-trained Transformer model. The data is processed in the model's multi-head self-attention mechanism and feedforward network, and finally, through average pooling, a 128-dimensional temporal feature vector is output. .

[0070] After feature extraction is complete, in step S4, the three feature vectors are fed into the knowledge-guided adaptive feature fusion module. First, the 128-dimensional... And 128-dimensional These are concatenated into a 256-dimensional concatenated feature vector. Simultaneously, Input the attention generation network to generate a 256-dimensional attention weight vector. Subsequently, through Hadama accumulation... This is applied to concatenate feature vectors, and the result is then input into a fusion layer, ultimately outputting, for example, a 64-dimensional fused feature vector. .

[0071] Finally, in step S5, probability prediction is performed. The generated 64-dimensional fused feature vector is then used. The input vector is fed into a trained random forest prediction model. This vector is evaluated across 100 decision trees within the model, each producing its own predicted probability. The arithmetic mean of the output probabilities from all decision trees is then calculated to obtain the final probability of a landslide occurring. For example, the calculated result is... =0.82. Based on the preset risk level classification threshold (e.g., a high-risk threshold of 0.7), since 0.82 ≥ 0.7, grid cell A at time point... The landslide risk was classified as "high risk". This result serves as the final output of this method for that 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; 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 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.

3. 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.

4. 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.

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 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.

6. 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.

7. 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.

8. 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.

9. 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.

10. 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