Slope disaster risk assessment method and system based on spatiotemporal monitoring data

Abstract

The present application relates to the technical field of slope monitoring, in particular to a slope disaster risk assessment method and system based on space-time monitoring data. The method comprises: obtaining category type geological environment, numerical type hydro-meteorology and time series type inducing factor data of the slope; mapping the hydro-meteorology into a modulation vector by a multilayer perception machine and fusing it with a geological environment embedding vector to generate a category feature code; constructing a monitoring point adjacency graph, and constructing a terrain influence direction matrix according to relative displacement, elevation difference and three-dimensional landslide dominant potential energy direction vector; extracting the inducing factor node hidden state by a gated recurrent unit, and generating a space-time coupling feature code after asymmetric mask modulation is applied to the initial aggregation weight of graph attention; and splicing and fusing into a risk representation vector by a hierarchical attention network to output a risk level by a classifier. The scheme improves the accuracy, stability and interpretability of slope disaster risk assessment.

Description

Technical Field

[0001] This invention relates to the field of slope monitoring technology, specifically to a method and system for assessing slope disaster risks based on spatiotemporal monitoring data. Background Technology

[0002] Slope hazards are a type of geological disaster characterized by high destructiveness and suddenness, and are frequently found in mountainous areas, along transportation routes, and around large-scale construction projects. Conducting scientific and accurate slope hazard risk assessments to achieve early identification, warning, and risk prevention is crucial for improving disaster prevention and mitigation capabilities. With the development of monitoring technology, routine monitoring methods for slope areas are constantly being enriched, enabling relevant departments to continuously acquire a large amount of multi-source heterogeneous monitoring data. This data typically includes categorical static geological environmental characteristics such as stratigraphy, vegetation cover, and slope structure; numerical hydrological and meteorological information such as rainfall and groundwater levels; and time-series inducing factors such as surface displacement and deformation rates.

[0003] Existing slope hazard assessment models often struggle to effectively characterize the deep interactions between categorical static geological data and numerical dynamic hydrological and meteorological data when processing multi-source heterogeneous data. They frequently employ simple splicing or parallel input methods for feature fusion, failing to fully exploit the synergistic modulation effects between different data types, resulting in insufficient multimodal feature representation. Simultaneously, slope hazard evolution exhibits significant spatial correlations and topographic orientation dependence. Factors such as relative elevation differences, horizontal spatial relationships between nodes, slope orientation, and landslide potential energy direction exert significant asymmetric guiding effects on material migration paths and hazard expansion directions. Current mainstream spatiotemporal prediction models largely rely on planar node adjacency relationships for spatial feature extraction, lacking a refined expression of the influence direction and potential energy transfer patterns of real three-dimensional terrain, making it difficult to reveal the coupling relationship between time-series triggering factors and complex three-dimensional terrain topology. Therefore, existing technologies still struggle to fully extract and integrate effective risk characteristics from multi-source heterogeneous monitoring data, and are also unable to accurately represent the spatiotemporal asymmetric influence patterns dominated by three-dimensional topography. This restricts the accuracy and reliability of slope disaster risk assessment results, and there is an urgent need to propose a risk assessment method that can coordinate multi-dimensional heterogeneous data and deeply mine the spatiotemporal coupling characteristics of topography. Summary of the Invention

[0004] In a first aspect, the present invention provides the following technical solution: a slope disaster risk assessment method based on spatiotemporal monitoring data, comprising the following steps: acquiring monitoring data from each monitoring point within a slope area, wherein the monitoring data includes categorical geological environment data, numerical hydrological and meteorological data, and time-series inducing factor data; mapping the numerical hydrological and meteorological data in the neighborhood of the monitoring point into modulation vectors using a multilayer perceptron; fusing the modulation vectors with the embedding vectors of the categorical geological environment data to generate categorical feature codes; constructing a slope spatial adjacency map with the monitoring points as nodes; constructing a topographic influence direction matrix based on the horizontal relative displacement vectors, relative elevation differences, and a preset three-dimensional landslide dominant potential energy direction vector between the nodes; and using gated loop units to process each... The temporal-series induced factor data corresponding to the monitoring points are encoded, and the node-level hidden states of each monitoring point are extracted. The node-level hidden states of the target node are used as query vectors. The initial aggregation weights from adjacent source nodes to the target node are calculated based on a graph attention network. The initial aggregation weights are subjected to asymmetric mask modulation processing using the terrain influence direction matrix. Based on the modulated weights, graph information is aggregated to generate spatiotemporally coupled feature codes. The category feature codes and the spatiotemporally coupled feature codes are concatenated to generate a hybrid coding feature vector. The hybrid coding feature vector is weighted and fused using a hierarchical attention network to generate a risk representation vector. The risk representation vector is input into a classifier to output the probability of each disaster risk level to determine the slope disaster risk level.

[0005] By using a multilayer perceptron, numerical hydrological and meteorological data within the neighborhood of monitoring points are mapped into modulation vectors and fused with the embedding vectors of categorical geological environmental data. This allows meteorological and hydrological conditions to dynamically adjust the expression weights of geological static features, avoiding the disconnect between environmental conditions and geological background caused by simple splicing of categorical and numerical features. Furthermore, the initial aggregation weights of the graph attention network are asymmetrically masked using the topographic influence direction matrix, and combined with the temporal variation patterns extracted by the gated recurrent unit, ensuring that spatial information aggregation aligns with the landslide potential energy transmission trend. Finally, a hierarchical attention network is used to weight and fuse the hybrid encoded feature vectors, and a classifier outputs the probability of each disaster risk level. This coordinates static geological attributes and temporal excitation features in risk representation, improving the accuracy and reliability of slope disaster risk level determination.

[0006] Further, generating category feature codes includes: extracting hydrological and meteorological element sequence data within a preset distance threshold and time window, centered on the target monitoring point, and converting it into a numerical feature vector; inputting the numerical feature vector into the multilayer perceptron, and outputting a context modulation vector through linear transformation and activation function mapping; performing element-wise multiplication of the context modulation vector with the initial embedding vector of the category geological environment data of the target monitoring point to generate a category feature code coupled with the local environmental state.

[0007] Furthermore, the preset distance threshold ranges from 50 to 200 meters, and the time window ranges from 3 to 14 days. Within the time window, three core meteorological elements—rainfall, soil moisture content, and groundwater level—are extracted. The sequences within different time windows are uniformly resampled into seven representative time points, forming a 21-dimensional numerical feature vector for the three core meteorological elements. The multilayer perceptron includes an input layer, two hidden layers with 64 and 32 neurons respectively, and an output layer. The internal nodes use the ReLU activation function, and an intermediate modulation value in the range of 0 to 1 is obtained through a Sigmoid activation layer. Then, a linear transformation is performed to generate a 64-dimensional context modulation vector with a value range of 0.5 to 1.5.

[0008] By limiting the distance threshold and time window to reasonable ranges and extracting three core meteorological elements—rainfall, soil moisture content, and groundwater level—the characterization of local environmental conditions focuses on key disaster-causing factors. By uniformly resampling sequences within different time windows into seven representative time points to form a fixed 21-dimensional numerical feature vector, inconsistent sequence lengths prevent inconsistencies in input dimensions that hinder stable processing. By setting hidden layers with 64 and 32 neurons respectively and applying ReLU activation, followed by a Sigmoid activation layer to obtain intermediate modulation values ​​in the 0-1 range, and then linearly transforming these values ​​to generate modulation vectors ranging from 0.5 to 1.5, the modulation coefficients can weaken non-critical dimensions while enhancing key disaster-causing dimensions, thereby improving the accuracy of categorical feature encoding in representing the geological susceptibility background under meteorological and hydrological changes.

[0009] Further, constructing the terrain influence direction matrix includes: calculating the Euclidean distance between the spatial coordinates of each monitoring point; performing dimensionless processing on the Euclidean distance and calculating a negative exponential function value as the initial connection weight between nodes in the slope spatial adjacency map; obtaining the horizontal relative displacement vector from the source node to the target node by subtracting the source node coordinates from the target node coordinates, and calculating the relative elevation difference based on the elevations of the source node and the target node; combining the horizontal relative displacement vector and the relative elevation difference into a three-dimensional spatial relative vector; performing dimensionless processing on the three-dimensional spatial relative vector to obtain a dimensionless three-dimensional spatial relative vector; calculating the inner product of the dimensionless three-dimensional spatial relative vector and the three-dimensional landslide dominant potential energy direction vector to obtain a direction consistency score; and performing nonnegative mapping on the direction consistency score to generate the terrain influence direction matrix.

[0010] By dimensionlessly processing the Euclidean distance between the spatial coordinates of monitoring points and calculating the negative exponential function value as the initial connection weight, spatial proximity relationships are integrated into the slope spatial adjacency map in a distance-prior manner. By combining the horizontal relative displacement vector and the relative elevation difference into a three-dimensional spatial relative vector and performing an inner product with the three-dimensional landslide dominant potential energy direction vector, a directional consistency score is obtained, enabling the connection between nodes to simultaneously carry spatial orientation and potential energy transmission trend information. By nonnegating the directional consistency score and mapping it to generate a terrain influence direction matrix, subsequent map information aggregation obtains directional guidance from the spatial dynamics level, avoiding ignoring the terrain energy transmission law by simply relying on feature similarity.

[0011] Furthermore, the Euclidean distance between the source node and the target node is calculated using spatial three-dimensional coordinates. The Euclidean distance is then subjected to dimensionless division using a preset slope influence radius. The dimensionless distance is substituted into a negative exponential function to obtain the initial connection weight. The slope influence radius ranges from 100 meters to 300 meters. A truncation threshold is set, and long-distance nodes with initial connection weights lower than the truncation threshold are disconnected to sparsify the slope spatial adjacency graph.

[0012] Furthermore, the time-series precipitating factor data is encoded using a gated recurrent unit to extract node-level hidden states. This includes: dividing the time-series precipitating factor data into sequence segments according to a preset time step and inputting them into the gated recurrent unit; at each time step, the gated recurrent unit reads the precipitating factor data at the current moment, combines it with the hidden state at the previous moment, and controls the retention and forgetting of historical time-series features through update gates and reset gates; and uses a nonlinear activation function to map the updated hidden state features to extract node-level hidden states representing the time-series change patterns of precipitating factors. The preset time step ranges from 1 to 12 hours, the sequence segment length is 24 time steps, the input feature dimension of the gated recurrent unit is 5-dimensional, and the hidden layer dimension is 128-dimensional.

[0013] By dividing the time-series precipitating factor data into sequence segments according to a preset time step and inputting them into a gated recurrent unit, and using update and reset gates at each time step to control the retention and forgetting of historical time-series features, the extraction of the temporal change pattern of precipitating factors is avoided due to invalid or abnormal perturbation information. By setting the input feature dimension to 5 dimensions, the hidden layer dimension to 128 dimensions, and using sequence segments with a length of 24 time steps, the node-level hidden state can condense the cumulative effect and mutation precursor of proximal precipitating conditions while ensuring the temporal representation capacity, thereby providing stable and discriminative temporal features for subsequent spatial aggregation.

[0014] Further, the initial aggregation weights are subjected to asymmetric mask modulation processing using the terrain influence direction matrix, including: determining the relative altitude of adjacent source nodes and target nodes; when the source node altitude is higher than the target node, and the inner product of the dimensionless three-dimensional spatial relative vector and the three-dimensional landslide dominant potential energy direction vector is greater than zero, the dimensionless relative elevation difference or the direction consistency score is substituted into a preset exponential gain function to calculate the enhancement coefficient, and the enhancement coefficient is multiplied by the initial aggregation weights of the corresponding transmission path to obtain the enhanced modulation aggregation weights; when the source node altitude is lower than the target node, or the inner product is less than or equal to zero, the absolute value of the dimensionless relative elevation difference or the direction deviation determined according to the direction consistency score is substituted into a preset exponential decay function to calculate the penalty coefficient, and the penalty coefficient is multiplied by the initial aggregation weights to obtain the suppressed modulation aggregation weights.

[0015] By first determining the relative elevation of adjacent source nodes and target nodes, and when the elevation of the source node is higher than that of the target node and the inner product of the three-dimensional relative vector and the three-dimensional landslide dominant potential energy direction vector is greater than zero, the relative elevation difference or direction consistency score is substituted into the exponential gain function to calculate the enhancement coefficient to increase the aggregation weight of the propagation path with the trend. When the elevation of the source node is lower than that of the target node or the inner product is less than or equal to zero, the absolute value of the relative elevation difference or the direction deviation is substituted into the exponential decay function to calculate the penalty coefficient to suppress the propagation path against the trend or deviating from the dominant potential energy direction. This allows the asymmetric mask modulation to fit the asymmetric mechanical characteristics of the upper landslide body strongly pushing the lower layer while the lower layer has limited traction on the upper layer, avoiding the participation of misleading information from the reverse slope direction in the aggregation, and improving the directional rationality of the spatiotemporal coupling characteristics.

[0016] Furthermore, a risk representation vector is generated by weighted fusion of the hybrid encoded feature vector using a hierarchical attention network, including: dividing the hybrid encoded feature vector into categorical feature segments and spatiotemporally coupled feature segments according to feature attributes; within each feature segment, calculating dimensional attention weights for different feature dimensions using a dimensional attention network, and performing element-wise weighting to highlight key features; evaluating the combined impact of the categorical feature segments and the spatiotemporally coupled feature segments on risk assessment using a segment-level attention network, and calculating segment-level attention weights; and weighted fusion of all feature segments based on the dimensional attention weights and the segment-level attention weights to generate the risk representation vector.

[0017] Further, the risk representation vector is input into a classifier to output the probability of each disaster risk level to determine the slope disaster risk level. This includes: inputting the risk representation vector into a fully connected layer for feature dimensionality reduction and high-order feature extraction. The fully connected layer consists of two layers: the first layer reduces the risk representation vector from 64 dimensions to 32 dimensions and inserts a regularization layer with a dropout rate of 0.3; the second layer maps the 32-dimensional features to output features equal to the number of preset disaster risk level categories. The extracted features are input into a Softmax classifier to calculate the exponential ratio corresponding to each preset disaster risk level. The exponential ratio is normalized to output a probability distribution vector with a sum of 1. The category corresponding to the highest probability value is selected as the slope disaster risk level. The preset disaster risk levels include low risk, medium risk, high risk, and extremely high risk.

[0018] In a second aspect, the present invention provides a slope disaster risk assessment system based on spatiotemporal monitoring data, including a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the slope disaster risk assessment method based on spatiotemporal monitoring data of the present invention is implemented.

[0019] The beneficial effects of this invention are as follows: By using a multilayer perceptron to map neighborhood hydrological and meteorological data into modulation vectors and fusing them element-by-element with the geological environment embedding vector, the underlying geological susceptibility under alternating meteorological and hydrological conditions is accurately characterized, avoiding environmental disconnect caused by simple splicing of categorical and numerical features; a topographic influence direction matrix is ​​constructed by using Euclidean distance negative exponential initial connection weights and consistency scores of the three-dimensional landslide dominant potential energy direction, and asymmetric mask modulation is applied to the initial aggregation weights of graph attention, ensuring that information aggregation aligns with the top-down movement characteristics of landslide soil, filtering out misleading propagation from the reverse slope direction or deviating from the dominant potential energy direction; the temporal variation patterns of inducing factors are extracted using gated recurrent units, and then the categorical feature segments and spatiotemporally coupled feature segments are first dimensionally refined and then balanced overall through a hierarchical attention network, resolving the expression conflict between static stratigraphic attributes and temporal excitation; finally, risk level probabilities are output through fully connected dimensionality reduction and normalized classification, improving the accuracy, stability, and interpretability of slope disaster risk assessment. Attached Figure Description

[0020] Figure 1 This is a flowchart of the slope disaster risk assessment method based on spatiotemporal monitoring data according to an embodiment of the present invention;

[0021] Figure 2 This is a schematic diagram of the feature modulation process in an embodiment of the present invention; Figure 3 This is a schematic diagram comparing the ablation test performance in the embodiments of the present invention; Figure 4 This is a structural block diagram of a slope disaster risk assessment system based on spatiotemporal monitoring data, according to an embodiment of the present invention. Detailed Implementation

[0022] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0023] Example of a slope disaster risk assessment method based on spatiotemporal monitoring data: like Figure 1 As shown, the slope disaster risk assessment method based on spatiotemporal monitoring data of the present invention includes: S1. Acquire monitoring data of the slope area. The monitoring data includes categorical geological environment data, numerical hydrological and meteorological data, and time-series inducing factor data. The numerical hydrological and meteorological data in the neighborhood of the monitoring point is mapped into modulation vectors through a multilayer perceptron. The modulation vectors are fused with the embedding vectors of the categorical geological environment data to generate categorical feature codes.

[0024] Monitoring data from various monitoring points within the slope area were acquired using sensor networks and geological exploration equipment. Categorical geological environmental data, including lithology distribution, vegetation cover type, and slope structure type, were collected. The StandardScaler module was used to perform zero-mean normalization on numerical hydrological and meteorological data, including soil moisture content, rainfall, and groundwater level. Time-series induced factor data, including historical displacement and rainfall time series, were converted into tensor format. A multilayer perceptron containing input, hidden, and output layers was constructed using the torch.nn.Linear module. The hidden layer employed the ReLU activation function. A low-dimensional dense vector of the numerical hydrological and meteorological data was calculated using the forward propagation algorithm and then nonlinearly mapped through a fully connected layer with a Sigmoid activation function, outputting an intermediate modulation value. This intermediate modulation value was further mapped to a preset modulation coefficient range, such as 0.5 to 1.5, through a linear interval transformation to generate a modulation vector. This ensured that the modulation coefficient could be less than 11 to weaken non-critical dimensions or greater than 11 to enhance critical disaster-causing dimensions. The torch.nn.Embedding module is used to map discrete categorical geological environment data into continuous embedding vectors. The torch.mul function is used to calculate the Hadamard product of the modulation vector and the embedding vector, and output the normalized categorical feature code.

[0025] In one embodiment, a multilayer perceptron maps numerical hydrological and meteorological data in the neighborhood of a monitoring point into modulation vectors. The modulation vectors are then fused with the embedding vectors of categorical geological environmental data to generate categorical feature codes, including: Centered on the target monitoring point, extract hydrological and meteorological element sequence data within a preset distance threshold and time window and convert them into numerical feature vectors; Numerical feature vectors are input into a multilayer perceptron, and the context modulation vector is output through linear transformation and activation function mapping. The context modulation vector is multiplied element-wise with the initial embedding vector of the categorical geological environment data of the target monitoring point to generate a categorical feature code that couples the local environmental state.

[0026] When extracting hydrometeorological elements, the preferred range for the preset distance threshold is 50 to 200 meters, for example, 100 meters; the preferred range for the time window is 3 to 14 days, for example, the past 7 days. Three core meteorological elements—rainfall, soil moisture content, and groundwater level—are extracted. The sequences within different time windows are uniformly resampled into 7 representative time points through diurnal aggregation, linear interpolation, truncation, or zero-mask padding, ensuring that the three hydrometeorological elements form a fixed 21-dimensional numerical feature vector before being input into the multilayer perceptron. Specifically, when the time window is shorter than 7 days, missing representative time points are padded with masks or interpolated; when the time window is longer than 7 days, it is compressed into 7 representative time points using equal-interval sampling or average pooling. This numerical feature vector is then input into the multilayer perceptron, whose network structure includes at least an input layer, two hidden layers with 64 and 32 neurons respectively, and an output layer. The internal nodes use the ReLU activation function for nonlinear feature extraction. An intermediate modulation value in the range of 0 to 1 is obtained through a Sigmoid activation layer, and then a 64-dimensional context modulation vector with a value range of 0.5 to 1.5 is generated through a linear transformation.

[0027] In the feature fusion stage, for categorical geological environmental data such as lithology distribution, vegetation cover type, and slope structure type, this categorical geological environmental data is pre-input into the embedding layer network, converted into an initial embedding vector with 64-dimensional features, and then output. Element-wise multiplication is performed between the obtained context modulation vector and the initial embedding vector, allowing meteorological features to dynamically adjust the expression weights of geostatic features. For example, if a certain dimension of the modulation vector has a value of 1.3, the feature of the corresponding dimension of the geological embedding vector is enhanced; conversely, if it is 0.7, the feature of that dimension is weakened. This outputs a 64-dimensional categorical feature code, representing the background state of geological susceptibility under alternating meteorological and hydrological environments with high fidelity. The feature modulation process is illustrated below. Figure 2 As shown.

[0028] S2. Construct a slope spatial adjacency graph with monitoring points as nodes. Construct a topographic influence direction matrix based on the horizontal relative displacement vector, relative elevation difference, and preset three-dimensional landslide dominant potential energy direction vector between nodes. Encode the time-series inducing factor data corresponding to each monitoring point using gated cyclic units. Extract the node-level hidden state of each monitoring point. Use the node-level hidden state of the target node as the query vector. Calculate the initial aggregation weights from adjacent source nodes to the target node based on a graph attention network. Perform asymmetric mask modulation processing on the initial aggregation weights using the topographic influence direction matrix. Generate spatiotemporal coupled feature encoding based on the modulated weights through graph information aggregation.

[0029] An undirected graph object is initialized using the NetworkX library. Monitoring points are added as nodes to the graph, and the Euclidean distance between nodes is calculated. When the distance is less than a preset threshold, the `add_edge` function is used to connect the two nodes, generating a spatial adjacency graph. The undirected graph is only used to determine whether spatial adjacency exists between monitoring points. During the graph information aggregation stage, each undirected adjacency edge is expanded into two directed message passing paths in opposite directions, and different modulation coefficients are assigned to each path based on the terrain influence direction matrix. The relative elevation difference is calculated using the elevation data of the source and target nodes, and the horizontal relative displacement vector from the source node to the target node is obtained based on the horizontal spatial coordinate difference between the source and target nodes. The horizontal relative displacement vector and the relative elevation difference are concatenated to form a three-dimensional spatial relative vector. This three-dimensional spatial relative vector is then multiplied by a preset three-dimensional landslide dominant potential energy direction vector to obtain a direction consistency score. The direction consistency score is nonnegated to obtain the terrain influence direction matrix used to modulate the aggregation weights. Each element of the terrain influence direction matrix is ​​a nonnegative modulation coefficient. The torch.nn.GRU module is called to input time-series induced factor data into the gated loop unit in time step order.

[0030] By controlling the forgetting and retention of information through reset and update gates, gated recurrent unit encoding is performed on the temporal inducing factor sequence of each monitoring point, and the output state of the last time step of each monitoring point is extracted as the node-level hidden state. During graph attention calculation, the node-level hidden state of the target node is mapped to a query vector through a linear layer, and the node-level hidden states or node features of adjacent source nodes are mapped to key and value vectors. The inner product is calculated based on the query vector and the corresponding key vectors of adjacent source nodes, and processed using the LeakyReLU activation function. Normalization is performed in the neighborhood of the target node using the Softmax function to obtain the initial aggregation weights from adjacent source nodes to the target node. The torch.mul function is used to multiply the corresponding elements of the nonnegative terrain influence direction matrix with the initial aggregation weights element-wise to complete asymmetric mask modulation processing. The modulated weights are then normalized by dividing by the sum of all modulated weights in the neighborhood of the same target node, ensuring that the final aggregation weights remain nonnegative and sum to 1. The torch.matmul function is used to perform matrix multiplication of the features of adjacent source nodes with the modulated weights to achieve weighted summation. The ELU activation function is used to perform nonlinear transformation to output a spatiotemporal aggregated representation that integrates spatial and temporal characteristics. Then, a linear mapping layer is used to reduce the 128-dimensional spatiotemporal aggregated representation to 64-dimensional, generating a 64-dimensional spatiotemporal coupled feature code.

[0031] In one embodiment, a slope spatial adjacency map is constructed with monitoring points as nodes. A terrain influence direction matrix is ​​constructed based on the horizontal relative displacement vectors, relative elevation differences, and a preset three-dimensional landslide dominant potential energy direction vector between nodes, including: Calculate the Euclidean distance between the spatial coordinates of each monitoring point, and after dimensionless processing of the Euclidean distance, calculate the negative exponential function value as the initial connection weight between nodes in the slope spatial adjacency diagram. Obtain the horizontal spatial coordinates and elevation data of adjacent source nodes and target nodes. Subtract the source node coordinates from the target node coordinates to obtain the horizontal relative displacement vector from the source node to the target node. Calculate the relative elevation difference based on the elevations of the source node and the target node. The horizontal relative displacement vector and the relative elevation difference are combined into a three-dimensional spatial relative vector. The three-dimensional spatial relative vector is then dimensionless to obtain a dimensionless three-dimensional spatial relative vector. The inner product of the dimensionless three-dimensional spatial relative vector and the preset three-dimensional landslide dominant potential energy direction vector is calculated to obtain the direction consistency score. The direction consistency score is then non-negatively mapped to generate a terrain influence direction matrix that represents the trend of terrain energy transmission between nodes.

[0032] When constructing spatial topological connections, spatial three-dimensional coordinates are used. Compute source node With the target node Euclidean distance between And using a preset slope influence radius R, preferably within the range of 100 meters to 300 meters, for example, taking R = 200 m, a dimensionless division process is performed to obtain... Substituting the dimensionless distance into the negative exponential function The initial connection weights are obtained from the equation, for example, when... At this point, the corresponding weight is approximately 0.606. The initial connection weights are stored as distance prior edge weights in the slope spatial adjacency graph and participate in the calculation of the unnormalized attention coefficients in the graph attention network along with the query-key similarity, or they are multiplied by the initial aggregation weights as adjacency mask weights before neighborhood normalization. A truncation threshold, such as 0.1, is set; for long-distance nodes with weights below this threshold, the connections are cut off to sparsify the slope spatial adjacency graph.

[0033] Obtain the source node respectively With the target node Horizontal spatial coordinates and elevation data, define source node Point to target node The three-dimensional spatial displacement vector is And define the elevation difference of descent. Among them, when When the value is greater than 0, it indicates the source node. The altitude is higher than the target node .right Dimensionless processing is performed to obtain a dimensionless three-dimensional spatial displacement vector. R is the preset radius of influence of the slope. This represents the maximum elevation difference in the region. In this embodiment, the "three-dimensional spatial relative vector" refers to the dimensionless three-dimensional spatial displacement vector formed after the three-dimensional spatial displacement vector from the source node to the target node has been dimensionlessized. The phrase "the source node's altitude is higher than the target node's" corresponds to the elevation difference during descent. The case where the value is greater than 0. Based on the historical main sliding motion pattern of this slope, a three-dimensional landslide dominant potential energy direction vector is pre-set. The three-dimensional landslide dominant potential energy direction vector is a unit vector. For example, when the original candidate direction vector is (0.707, -0.707, -1.0), it is first normalized to obtain a unit direction vector of approximately The dimensionless three-dimensional spatial displacement vector The unit vector of the pre-defined three-dimensional landslide dominant potential energy direction Perform the inner product operation to obtain ;when When the value is greater than 0, it indicates the source node. Point to target node The direction of spatial displacement is consistent with the direction of the dominant potential energy of the landslide.

[0034] Furthermore, As a score for directional consistency, combined with the descent elevation difference Mapped to nonnegative modulation coefficients via exponential gain or exponential decay functions. ;when Greater than 0 and When greater than 0, A value greater than or equal to 1 is used to enhance the propagation path; when Less than or equal to 0, or When less than or equal to 0, Located between 0 and 1, these are used to suppress counter-current propagation or paths deviating from the dominant potential energy direction. All The terrain influence direction matrix is ​​formed, while the original inner product value... It serves only as an intermediate quantity in generating nonnegative modulation coefficients and is not directly multiplied by the initial aggregation weights. This provides strong prior directional guidance for information transmission from a spatial dynamics perspective.

[0035] In one embodiment, a gated loop unit is used to encode time-series precipitating factor data to extract hidden states, including: The time-series inducing factor data are divided into sequence segments according to the time step and input into the gated loop unit; At each time step, the gated loop unit reads the triggering factor data at the current moment, combines it with the hidden state at the previous moment, and controls the retention and forgetting of historical time sequence features by updating the gate and resetting the gate. By using a nonlinear activation function to map the updated hidden state features, the hidden state representing the temporal variation pattern of the inducing factors is extracted.

[0036] For multidimensional time-series data on inducing factors such as surface displacement deformation, pore water pressure, and dip angle change rate, the data is divided into 24 time steps, each representing a sequence segment of the past 12 days, according to a preset time step, preferably 1 to 12 hours (e.g., 12 hours). These segments are then input into a gated recurrent unit. The network structure of this gated recurrent unit includes a reset gate, an update gate, and a hidden layer. The input feature dimension corresponds to the number of categories of inducing factors, for example, set to 5 dimensions. The hidden layer dimension is preferably set to 128 dimensions to ensure the time-series representation capacity.

[0037] At each time step of the model's propagation computation In the process, the network simultaneously reads the 5-dimensional inducing factor input vector at the current moment. And the 128-dimensional hidden state transmitted from the previous moment. The network first calculates the discard coefficient using a reset gate with a sigmoid activation function to control the discarding of invalid or abnormal perturbation information from the early stages. Then, an update gate with the same sigmoid activation function is used to evaluate the importance of the previous hidden state to the current hidden state, determining how many long-term temporal dependencies should be retained. Candidate hidden states for the current time step are generated based on a tanh nonlinear activation function mapping, and the update gate merges these candidate hidden states with the old hidden states to calculate the current hidden state. The aforementioned nonlinear mapping object represents the feature information of the hidden state and candidate hidden states, rather than the network weight parameters of the gated recurrent unit. After the entire time window loop ends, the output state vector of the last time step is extracted and output as the query vector. The 128-dimensional hidden state output highly condenses the cumulative effect and mutation precursors of the near-end catastrophe induction conditions.

[0038] In one embodiment, the initial aggregation weights are subjected to asymmetric mask modulation using the terrain influence direction matrix, including: Determine the relative altitude between adjacent source nodes and target nodes; When the source node's elevation is higher than the target node's, and the inner product of the three-dimensional spatial relative vector and the three-dimensional landslide dominant potential energy direction vector is greater than zero, it is determined that the message transmission direction is consistent with the ground potential energy transmission trend. The dimensionless relative elevation difference or direction consistency score is substituted into the preset exponential gain function to calculate the enhancement coefficient. The enhancement coefficient is multiplied by the initial aggregation weight of the corresponding transmission path to obtain the enhanced modulation aggregation weight. When the source node's elevation is lower than the target node's, or when the inner product of the three-dimensional spatial relative vector and the three-dimensional landslide dominant potential energy direction vector is less than or equal to zero, it is determined that the message transmission is in reverse or deviates from the dominant potential energy direction. The absolute value of the relative elevation difference or the direction deviation after dimensionless processing is substituted into the preset exponential decay function to calculate the penalty coefficient. The penalty coefficient is multiplied by the initial aggregation weight to obtain the suppressed modulation aggregation weight.

[0039] In the information transmission phase of graph aggregation, by decreasing the elevation difference Determine the direction of potential energy transfer, and Maximum elevation difference in the region The division is performed to obtain a dimensionless normalized descent height difference. Simultaneously, combining the dimensionless three-dimensional spatial displacement vector That is, the inner product score of the dimensionless three-dimensional spatial relative vector and the three-dimensional landslide dominant potential energy direction vector. Determine whether the propagation path aligns with the dominant landslide potential energy direction to ensure consistency between the asymmetric mask modulation and the construction basis of the topographic influence direction matrix. When the elevation difference decreases... A dimensionless three-dimensional spatial displacement vector that is greater than 0. The inner product score with the dominant potential energy direction vector of the three-dimensional landslide When the value is greater than 0, it indicates that the distance from the source node i to the target node is... When the information transmission direction simultaneously satisfies two conditions—from high to low and consistent with the direction of the dominant landslide potential energy—it is determined to be a propagation path following the landslide's momentum. At this point, Substitute into form In the exponential gain function, the gain control factor The preferred range is 0.5 to 2.0, for example, taking... Suppose a pair of nodes Then the enhancement coefficient is calculated. It is approximately equal to 1.49. If the initial aggregation weight is 0.3, the enhanced modulation aggregation weight after multiplying the two will increase to 0.447.

[0040] When the elevation difference decreases Less than or equal to 0, or a dimensionless three-dimensional spatial displacement vector The inner product score with the dominant potential energy direction vector of the three-dimensional landslide When the value is less than or equal to 0, it indicates that the path does not meet the conditions for propagation with the trend, and is judged as a propagation path against the trend or deviating from the direction of the dominant potential energy, violating the dominant characteristic of the landslide soil moving from top to bottom. The deviation in direction can be defined as... This represents the degree of deviation of the spatial displacement direction from the source node to the target node relative to the direction of the dominant landslide potential energy. In this case, the elevation difference is taken. The absolute value is dimensionless to obtain And will penalize the input amount Substitute into form In the exponential decay function, λ is a preset weighting coefficient between 0 and 1. The decay penalty factor... The preferred range is 1.0 to 3.0, for example, taking... Assume the penalty input on the backpropagation path. =0.3, the calculated penalty coefficient Approximately 0.55, the corresponding initial weight of 0.3 is suppressed and weakened to 0.165. Since both the enhancement coefficient and the penalty coefficient are non-negative, the modulated aggregation weight will not have negative values; further normalization within the neighborhood ensures that the graph information aggregation process is a non-negative weighted aggregation. This asymmetric masking process simulates the asymmetric mechanical characteristics of the strong pushing and squeezing of the upper landslide body on the lower layer and the limited traction feedback of the lower layer on the upper layer in a real three-dimensional landslide space.

[0041] S3. The category feature encoding and the spatiotemporal coupling feature encoding are concatenated to generate a hybrid encoding feature vector. The hybrid encoding feature vector is weighted and fused using a hierarchical attention network to generate a risk representation vector. The risk representation vector is input into the classifier to output the probability of each disaster risk level to determine the slope disaster risk level.

[0042] The `torch.cat` function is used to concatenate categorical feature encoding and spatiotemporally coupled feature encoding into a one-dimensional feature tensor to generate a hybrid encoded feature vector. A first-layer dimensional attention module is constructed using a fully connected layer combined with a Softmax operation to generate attention weights for each feature dimension within both the categorical and spatiotemporally coupled feature segments. Element-wise weighting is then applied to the corresponding feature segments based on these dimensional attention weights to output an enhanced feature representation. In the second-layer segment-level attention module, a globally learnable context vector is constructed. The dot product similarity between the enhanced feature representation and the context vector is calculated, and the attention distribution is obtained using the Softmax function. Based on this attention distribution, a weighted fusion operation is used to aggregate global features to generate a risk representation vector. A classifier is constructed consisting of two fully connected layers and a batch normalization layer based on the `torch.nn.BatchNorm1d` module. The risk representation vector is input into the classifier for feature dimensionality reduction and normalization. The final fully connected layer uses the Softmax function to output probability distribution vectors representing four risk levels: low, medium, high, and very high. The index of the maximum value in the probability distribution vector is extracted using the torch.argmax function, and the category corresponding to the maximum probability is used as the determined slope disaster risk level.

[0043] In one embodiment, a hybrid encoded feature vector is generated by concatenating category feature encoding with spatiotemporally coupled feature encoding. A hierarchical attention network is then used to perform a weighted fusion operation on the hybrid encoded feature vector to generate a risk representation vector, including: The hybrid coding feature vector is divided into categorical feature segments and spatiotemporally coupled feature segments according to feature attributes; Within each feature segment, attention weights for different feature dimensions are calculated using a dimensional attention network, and element-wise weighting is performed to highlight key features. Between each feature segment, the combined impact of category feature segments and spatiotemporally coupled feature segments on risk assessment is evaluated through a segment-level attention network, and segment-level attention weights are calculated. All feature segments are weighted and fused based on dimension-level attention weights and segment-level attention weights to generate a risk representation vector.

[0044] The concatenated hybrid encoded feature vector is divided into two parts according to attribute definitions: the first 64 dimensions correspond to the category feature segment mapping static attributes to the environment, and the latter 64 dimensions correspond to the spatiotemporal coupled feature segment output by the graph network aggregation after dimensionality reduction by a linear mapping layer. Specifically, the gated recurrent unit corresponding to each monitoring point outputs a 128-dimensional node-level hidden state. The node-level hidden state of the target node is first used as a query vector in graph attention calculation. The 128-dimensional spatiotemporal aggregated representation obtained by the graph network is then reduced to a 64-dimensional spatiotemporal coupled feature code by a linear mapping layer. Therefore, concatenating this with the 64-dimensional category feature code yields a 128-dimensional hybrid encoded feature vector. For processing within the feature segments, a dimensional attention network is constructed. The input to this dimensional attention network is either the category feature segment or the spatiotemporal coupled feature segment. The network structure consists of a fully connected layer combined with a Softmax operation, outputting 64-dimensional weight vectors for both the category feature segment and the spatiotemporal coupled feature segment. The weight values ​​range from 0 to 1. For example, a certain dimension representing key geological fracture zone information could have a weight value of 0.85. Based on this, a dot product weighting is performed with the original features to extract and highlight the core local causal factors that cause geological disasters. In this embodiment, dimensional attention does not directly sum and compress the 64-dimensional feature segments. Instead, it first performs element-wise weighting to keep the feature segment length at 64 dimensions. If an intra-segment summary vector is needed, weighted summation or weighted superposition can be performed in the subsequent segment-level fusion stage.

[0045] After internal feature extraction, a segment-level attention network is used to represent the overall contribution of the two major feature categories. The input to the segment-level attention network is two refined feature segments. The network structure is a shared perceptron layer based on a learnable query vector. The network is calculated using a hyperbolic tangent function combined with a softmax function, outputting normalized segment-level attention weights. For example, the categorical feature segment receives a weight of 0.4, and the spatiotemporally coupled feature segment receives a weight of 0.6, with a total weight of 1.0. The respective output segment-level attention weights are multiplied and added together with the corresponding refined feature segments to produce a one-dimensional risk representation vector with a length of 64. This mechanism of first refining dimensions and then balancing the overall structure resolves the information interference conflict between the inherent geological attributes of the slope and the external temporal excitations in the representation space.

[0046] In one embodiment, the risk representation vector is input into a classifier to output the probability of each hazard risk level to determine the slope hazard risk level, including: The risk representation vector is input into a fully connected layer for feature dimensionality reduction and high-order feature extraction; The extracted features are input into the Softmax classifier to calculate the index ratio corresponding to each preset disaster risk level; The exponential ratio is normalized to output a probability distribution vector with a sum of 1. The category corresponding to the highest probability value is selected as the slope disaster risk level.

[0047] The generated risk representation vector, such as a 64-dimensional dense vector, is input into a fully connected layer network composed of multilayer perceptrons for dimensionality reduction. The input to the fully connected layer network is the 64-dimensional risk representation vector, and the network structure consists of two layers: the first layer converts the representation vector... The dimensions are reduced from 64 to 32, with a regularization layer with a dropout rate preferably set to 0.3 inserted in between to prevent parameter overfitting. The second layer maps these 32 high-order features to output features equal to the number of preset slope risk level categories. For example, if the preset slope risk level includes four levels: low risk, medium risk, high risk, and extremely high risk, then the output features are 4-dimensional, and the obtained feature vectors are 1.2, 0.5, 3.1, and -0.8.

[0048] The 4-dimensional output feature is then input into a Softmax classifier layer. This Softmax classifier layer is structured as an exponential and normalization network, using the natural logarithm base e to calculate the exponent of each output element, for example... Approximately 22.20, this category's index value is divided by the sum of the indices of all categories, for example, a sum of 27.62. This division process performs global normalization of the index ratio, outputting a probability distribution vector distributed between 0 and 1 with a sum of 1, such as 0.12, 0.06, 0.80, and 0.02. The highest probability value in this vector, 0.80, is detected, and the corresponding index position is determined to represent a high-risk level, thereby determining the disaster warning risk level of the target slope at the current time point. This mechanism ensures that the classification results have both a clear maximum likelihood tendency and quantitative interpretability of confidence levels.

[0049] The experiment was based on a five-year monitoring dataset from a key geological disaster-prone area, encompassing over 100,000 samples from multiple sources including rainfall, pore water pressure, surface displacement, and geological lithology. This dataset was divided into training, validation, and test sets in an 8:1:1 ratio. All experiments were conducted on GPU servers with equivalent computing power. Three sets of control conditions were set: the first set used the complete model of this invention with the full algorithm logic; the second set removed the terrain-affected direction matrix and asymmetric mask modulation, replacing it with symmetric aggregation of the base map; and the third set removed the hierarchical attention network, replacing it with multilayer perceptron fusion.

[0050] The complete model of this invention has a prediction accuracy of [percentage missing] on the test set. The macro F1 score is The area under the receiver operating characteristic curve is 0.962. In contrast, the prediction accuracy of the model with the terrain mask removed drops to... Macro F1 scores fell back to The area under the curve decreased to 0.915. The model that removed hierarchical attention performed the worst, with a prediction accuracy of only [missing data]. The macro F1 score is The area under the curve is 0.898. A comparison of ablation test performance is shown in the diagram below. Figure 3 As shown.

[0051] The complete model achieved a 5.2 percentage point improvement in accuracy compared to the model without terrain masks, demonstrating that asymmetric mask modulation processing, based on spatial relative elevation differences and the consistency score of the dominant landslide potential energy direction in 3D, can highly restore the true gravitational potential energy migration direction of the landslide soil and filter out misleading information propagated by graph nodes on the opposite slope or deviating from the dominant potential energy direction. The complete model also achieved a 6.9 percentage point improvement in accuracy compared to the model without hierarchical attention, indicating that using a weighted fusion operation with dimensional and segment-level dual-layer networks can identify and amplify core catastrophic causative factors, resolving the expression conflicts arising from the splicing of static categorical features and spatiotemporally coupled features.

[0052] On the other hand, this invention also provides a slope disaster risk assessment system based on spatiotemporal monitoring data. For example... Figure 4 The slope disaster risk assessment system based on spatiotemporal monitoring data shown includes a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the slope disaster risk assessment method based on spatiotemporal monitoring data according to the first aspect of the present invention.

[0053] The slope disaster risk assessment system based on spatiotemporal monitoring data also includes other components well known to those skilled in the art, such as communication interfaces. Their settings and functions are known in the art and will not be described in detail here.

Claims

1. A slope disaster risk assessment method based on spatiotemporal monitoring data, characterized in that, Includes the following steps: Acquire monitoring data from each monitoring point within the slope area. The monitoring data includes categorical geological environment data, numerical hydro-meteorological data, and time-series inducing factor data. Use a multilayer perceptron to map the numerical hydro-meteorological data in the neighborhood of the monitoring point into modulation vectors. Then, fuse the modulation vectors with the embedding vectors of the categorical geological environment data to generate categorical feature codes. A slope spatial adjacency graph is constructed with the monitoring points as nodes. A terrain influence direction matrix is ​​constructed based on the horizontal relative displacement vector, relative elevation difference, and preset three-dimensional landslide dominant potential energy direction vector between the nodes. The time-series inducing factor data corresponding to each monitoring point is encoded using gated loop units. The node-level hidden state of each monitoring point is extracted. The node-level hidden state of the target node is used as the query vector. The initial aggregation weight from the adjacent source node to the target node is calculated based on the graph attention network. The initial aggregation weight is subjected to asymmetric mask modulation processing using the terrain influence direction matrix. The spatiotemporal coupled feature encoding is generated by graph information aggregation based on the modulated weight. The category feature encoding and the spatiotemporal coupling feature encoding are concatenated to generate a hybrid encoding feature vector. A hierarchical attention network is used to perform a weighted fusion operation on the hybrid encoding feature vector to generate a risk representation vector. The risk representation vector is input into a classifier to output the probability of each disaster risk level to determine the slope disaster risk level.

2. The slope disaster risk assessment method based on spatiotemporal monitoring data according to claim 1, characterized in that, Generate categorical feature codes, including: Centered on the target monitoring point, extract hydrological and meteorological element sequence data within a preset distance threshold and time window and convert them into numerical feature vectors; The numerical feature vector is input into the multilayer perceptron, and the context modulation vector is output through linear transformation and activation function mapping. The context modulation vector is multiplied element-wise with the initial embedding vector of the categorical geological environment data of the target monitoring point to generate a categorical feature code coupled with the local environmental state.

3. The slope disaster risk assessment method based on spatiotemporal monitoring data according to claim 2, characterized in that, The preset distance threshold ranges from 50 to 200 meters, and the time window ranges from 3 to 14 days. Within the time window, three core meteorological elements—rainfall, soil moisture content, and groundwater level—are extracted. The sequences within different time windows are uniformly resampled into seven representative time points, so that the three core meteorological elements form a 21-dimensional numerical feature vector. The multilayer perceptron includes an input layer, two hidden layers with 64 and 32 neurons respectively, and an output layer. The internal nodes use the ReLU activation function, and the intermediate modulation value in the range of 0 to 1 is obtained through the Sigmoid activation layer. Then, a 64-dimensional context modulation vector with a value range of 0.5 to 1.5 is generated by linear transformation.

4. The slope disaster risk assessment method based on spatiotemporal monitoring data according to claim 1, characterized in that, Constructing the terrain influence direction matrix includes: Calculate the Euclidean distance between the spatial coordinates of each monitoring point, and after dimensionless processing of the Euclidean distance, calculate the negative exponential function value as the initial connection weight between the nodes of the slope spatial adjacency graph. The horizontal relative displacement vector from the source node to the target node is obtained by subtracting the source node coordinates from the target node coordinates, and the relative elevation difference is calculated based on the source node elevation and the target node elevation. The horizontal relative displacement vector and the relative elevation difference are combined into a three-dimensional spatial relative vector. The three-dimensional spatial relative vector is then dimensionless to obtain a dimensionless three-dimensional spatial relative vector. The inner product of the dimensionless three-dimensional spatial relative vector and the three-dimensional landslide dominant potential energy direction vector is calculated to obtain a direction consistency score. The direction consistency score is then non-negatively mapped to generate the terrain influence direction matrix.

5. The slope disaster risk assessment method based on spatiotemporal monitoring data according to claim 4, characterized in that, The Euclidean distance between the source node and the target node is calculated using spatial three-dimensional coordinates. The Euclidean distance is then subjected to dimensionless division using a preset slope influence radius. The dimensionless distance is then substituted into a negative exponential function to obtain the initial connection weight. The value range of the slope influence radius is 100 meters to 300 meters. A cutoff threshold is set, and long-distance nodes with initial connection weights lower than the cutoff threshold are disconnected to sparsify the slope spatial adjacency graph.

6. The slope disaster risk assessment method based on spatiotemporal monitoring data according to claim 1, characterized in that, The time-series triggering factor data is encoded using a gated loop unit to extract node-level hidden states, including: The time-series inducing factor data is divided into sequence segments according to a preset time step and input into the gated loop unit; At each time step, the gated loop unit reads the triggering factor data at the current moment, combines it with the hidden state at the previous moment, and controls the retention and forgetting of historical time sequence features by updating the gate and resetting the gate. By using a nonlinear activation function to map the updated hidden state features, node-level hidden states representing the temporal variation patterns of inducing factors are extracted. The preset time step ranges from 1 to 12 hours, the sequence segment has a length of 24 time steps, the input feature dimension of the gated recurrent unit is 5-dimensional, and the hidden layer dimension is 128-dimensional.

7. The slope disaster risk assessment method based on spatiotemporal monitoring data according to claim 4, characterized in that, The initial aggregation weights are subjected to asymmetric mask modulation processing using the terrain influence direction matrix, including: Determine the relative elevation between adjacent source nodes and target nodes; when the elevation of the source node is higher than that of the target node, and the inner product of the dimensionless three-dimensional spatial relative vector and the three-dimensional landslide dominant potential energy direction vector is greater than zero, substitute the dimensionless relative elevation difference or the direction consistency score into the preset exponential gain function to calculate the enhancement coefficient, and multiply the enhancement coefficient with the initial aggregation weight of the corresponding transmission path to obtain the enhanced modulation aggregation weight; When the source node's altitude is lower than the target node's altitude, or when the inner product result is less than or equal to zero, the absolute value of the dimensionless relative elevation difference or the directional deviation determined based on the directional consistency score is substituted into a preset exponential decay function to calculate a penalty coefficient. The penalty coefficient is then multiplied by the initial aggregation weight to obtain the suppressed modulation aggregation weight.

8. The slope disaster risk assessment method based on spatiotemporal monitoring data according to claim 1, characterized in that, A risk representation vector is generated by weighting and fusing the hybrid encoded feature vectors using a hierarchical attention network, including: The hybrid encoded feature vector is divided into category feature segments and spatiotemporal coupled feature segments according to feature attributes; Within each feature segment, a dimensional attention network is used to calculate the dimensional attention weights for different feature dimensions, and element-wise weighting is performed to highlight key features. Between the category feature segment and the spatiotemporal coupling feature segment, a segment-level attention network is used to evaluate the combined impact of both on risk assessment and to calculate the segment-level attention weights. The risk representation vector is generated by weighting and fusing all feature segments according to the dimension-level attention weight and the segment-level attention weight.

9. The slope disaster risk assessment method based on spatiotemporal monitoring data according to claim 8, characterized in that, The risk representation vector is input into a classifier to output the probability of each disaster risk level to determine the slope disaster risk level, including: The risk representation vector is input into a fully connected layer for feature dimensionality reduction and high-order feature extraction. The fully connected layer consists of two layers. The first layer reduces the risk representation vector from 64 dimensions to 32 dimensions and inserts a regularization layer with a dropout rate of 0.

3. The second layer maps the 32-dimensional features to output features equal to the number of preset disaster risk level categories. The extracted features are input into the Softmax classifier to calculate the index ratio corresponding to each preset disaster risk level; The index ratio is normalized to output a probability distribution vector with a sum of 1. The category corresponding to the maximum probability value is selected as the slope disaster risk level. The preset disaster risk levels include low risk, medium risk, high risk and extremely high risk.

10. A slope disaster risk assessment system based on spatiotemporal monitoring data, characterized in that, It includes a processor and a memory, the memory storing computer program instructions, which, when executed by the processor, implement the slope disaster risk assessment method based on spatiotemporal monitoring data as described in any one of claims 1 to 9.

Images