Landslide hazard prediction system and method

Abstract

The present application belongs to the field of geological disaster early warning, and provides a landslide danger prediction system and method, wherein the method comprises: fusing and normalizing processing of multi-source disaster-causing factors, obtaining an influence factor matrix and performing weighted correction; determining the spatial correlation strength and semantic correlation degree between nodes according to the correlation edges; obtaining the correlation strength and correlation characteristics of the nodes by using feature mapping, correlation weight calculation and information aggregation adaptive learning; learning by using a correlation graph model according to the correlation strength and correlation characteristics of the nodes, obtaining a global prediction model, and optimizing the global prediction model; predicting the target landslide area by using the optimized global prediction model, obtaining a prediction result, and dividing the prediction result into danger levels according to a preset probability threshold. The present application has the beneficial effect of improving the precision of landslide danger prediction.

Description

Technical Field

[0001] This invention relates to the field of geological disaster early warning technology, and in particular to a landslide hazard prediction system and method. Background Technology

[0002] The occurrence of landslide disasters is the result of the synergistic effect of multiple factors such as topography, strata, rainfall, and vegetation in space. Its core characteristics are spatial heterogeneity (the influence intensity of the same factor varies significantly in different regions) and complex correlation (nonlinear interaction exists between multiple factors).

[0003] Existing technologies have the following limitations: The spatial homogeneity assumption is flawed: Traditional models (such as statistical regression and ordinary machine learning) treat the influence of factors as globally consistent, failing to reflect spatial differences such as "the impact of slope on landslides is drastically different in fault zones and gentle areas," leading to biases in local hazard assessments. Landslide influencing factors include topographic features (slope, aspect, elevation), geological conditions (lithology, fault distribution), triggering factors (rainfall, vegetation cover), and disaster-bearing bodies (roads), etc. Existing methods often treat these as independent feature inputs, making it difficult to capture key combined effects such as "fault zone + heavy rainfall" and "steep slope + low vegetation." Some graphical models use a single type of node and edge, failing to distinguish the essential differences between "topographic units" and "geological bodies," and also struggling to model the "one-way threat relationship between triggering factors and disaster-bearing bodies," resulting in the loss of semantic connections between factors. Pixel-level models (such as CNNs) only focus on local window features, ignoring regional-scale geological patterns; while global statistical models struggle to capture the abrupt changes in micro-topography, and this separation limits prediction accuracy. Summary of the Invention

[0004] Aimed at at least in solving one of the technical problems existing in the prior art, the present invention provides a landslide hazard prediction system and method, which improves the accuracy of landslide hazard prediction.

[0005] One aspect of the present invention provides a landslide hazard prediction system, comprising:

[0006] The data acquisition unit is used to collect multi-source geological and environmental data of the target landslide area;

[0007] The feature construction unit is used to determine multi-source disaster-causing factors based on multi-source geological data and environmental data, and to fuse and normalize the multi-source disaster-causing factors to obtain the influence factor matrix.

[0008] A spatially heterogeneous weighted modeling unit is used to perform weighted correction on the influence factor matrix according to a dynamic weighting mechanism to obtain an optimized factor matrix;

[0009] Spatial and semantic association unit, used to determine association edges according to preset node types, and to determine the spatial association strength and semantic association degree between nodes according to the association edges;

[0010] The adaptive learning unit is used to obtain the association strength and association features of nodes by adaptive learning through feature mapping, association weight calculation and information aggregation based on the preset node type and preset association edges.

[0011] The multi-element interactive network computing unit is used to learn a global prediction model based on the association strength and association characteristics of nodes using an association graph model, and to apply classification bias correction, spatial morphology constraints and end-to-end optimization to the global prediction model to obtain an optimized global prediction model.

[0012] The landslide hazard prediction unit is used to predict the target landslide area using an optimized global prediction model, obtain the prediction results, and classify the hazard level of the prediction results according to a preset probability threshold.

[0013] According to the aforementioned landslide hazard prediction system, the data acquisition unit is used for:

[0014] Data is collected from the target landslide area using three-dimensional lidar, ground-penetrating radar, soil moisture sensors, automatic weather stations, GNSS monitoring points, and satellite receiving devices to obtain topographic data, geological structure data, hydrogeological monitoring data, meteorological data, GNSS monitoring data, and remote sensing images. The collected multi-source geological and environmental data are then transmitted to the central server via at least one of the following methods: 4G / 5G, LoRa, or satellite communication.

[0015] According to the landslide hazard prediction system, the feature construction unit is used for:

[0016] The target landslide area was divided into multiple assessment units. Based on multi-source geological and environmental data, the multi-source disaster-causing factors of each assessment unit were determined. These multi-source disaster-causing factors include topographic features, geological features, triggering features, and disaster-bearing features. Topographic features include slope s, slope aspect a, and relative elevation difference h; geological features include lithology type l and distance from fault zones f; triggering features include annual cumulative rainfall r and vegetation cover index v; and disaster-bearing features include distance from roads d.

[0017] The original feature matrix obtained after fusing topographic features, geological features, triggering features, and disaster-bearing features is as follows:

[0018] ;

[0019] in, ,in Indicates transpose. As an identifier for the evaluation unit number; The dimension represents the number of disaster-causing factors; To evaluate the number of units;

[0020] The original feature matrix is ​​mapped to the interval [0,1] using a normalization method, resulting in the normalized influence factor matrix:

[0021] ;

[0022] in, This represents the normalization result. Represents the original feature matrix. The first sample One eigenvalue; For the first The minimum value of each feature among all samples. No. The maximum value of a feature across all samples.

[0023] According to the aforementioned landslide hazard prediction system, the spatially heterogeneous weighted modeling unit is used for:

[0024] Based on the spatial distribution characteristics of disaster-causing factors in the assessment unit, the dynamic weighting mechanism is used to calculate the weighting of each assessment unit. The local weights are:

[0025] ;

[0026] ;

[0027] The dynamic weighting mechanism is determined by the first law of geography. As an evaluation unit With evaluation unit Spatial distance, and As an evaluation unit coordinates and As an evaluation unit coordinates For adaptive bandwidth, According to the evaluation unit The density of surrounding disaster-causing factors is dynamically adjusted;

[0028] According to the evaluation unit The original feature matrix is ​​used to determine the local linear regression model of the geographic weighted regression (GWR). for:

[0029] ;

[0030] in, For local coefficient vectors, This is the error vector;

[0031] Based on the local linear regression model of geographically weighted regression (GWR) Local coefficients are determined by weighted least squares. for:

[0032] ;

[0033] in, It is a diagonal weight matrix;

[0034] Based on local coefficients The corrected feature matrix is ​​determined as follows:

[0035] ;

[0036] in, The factor matrix is ​​a diagonal matrix of local coefficients. The influence factor matrix is ​​then weighted and corrected using the modified characteristic matrix to obtain the optimized factor matrix. for:

[0037] .

[0038] According to the aforementioned landslide hazard prediction system, the spatial and semantic association units are used for:

[0039] Encode the preset node type using one-hot vectors to obtain the node. Type encoding for:

[0040] ;

[0041] Optimize the factor matrix Fusion features with type encoding for:

[0042] ;

[0043] Based on the fusion characteristics, the associated edges are determined, and the spatial association strength and semantic association degree between nodes are determined based on the associated edges. The associated edges include intra-elements of the same type of elements, cross-element interaction edges, and trigger-disaster orientation edges.

[0044] In this context, inline edges of the same type are used for nodes of the same type, and the edge weight of inline edges of the same type is specified. To convey information about the continuity of similar elements:

[0045] ;

[0046] Among them, cross-category feature interaction edges are used to connect nodes of different types, and the edge weights of cross-category feature interaction edges are... For feature similarity:

[0047] ;

[0048] in, For nodes The fused feature vector, for transpose, The norm is used; the trigger-disaster directional edge indicates that it only points from the trigger node to the disaster-bearing node, and the edge weight of the trigger-disaster directional edge indicates the one-way threat of the triggering factor to the disaster-bearing body.

[0049] According to the aforementioned landslide hazard prediction system, the spatial and semantic association units are used for:

[0050] Map features to nodes and neighboring nodes Type-specific linear transformations:

[0051] ;

[0052] in, For type-specific weight matrices, , and For type-specific linear transformation results, For nodes The query vector, For nodes The key vector, For nodes Value vector, node The type is ,node The type of the neighbor node is ;

[0053] The association weight calculation includes fusing spatial distance and feature similarity to calculate the association strength between nodes. for:

[0054] ;

[0055] in, For attention parameter vectors, For nodes The neighborhood group, For activation function, For nodes With nodes Local weights; For nodes The The key vectors of the neighboring nodes, Node is With the Edge characteristics of each neighboring node, For nodes With the Local weights of each neighboring node;

[0056] Based on the strength of the association between nodes By aggregating information from surrounding nodes, the association features of the nodes are obtained. for:

[0057] ;

[0058] Among them, association features In the formula This is the activation function.

[0059] According to the aforementioned landslide hazard prediction system, the multi-factor interactive network computing unit is used for:

[0060] Get Nodes Association features The pre-activation value is obtained by transforming it through a fully connected layer. :

[0061] ;

[0062] in, This is the output layer weight matrix. For bias terms, pre-activation values In the formula Use the Sigmoid activation function;

[0063] The pre-activation values ​​are mapped to landslide probabilities using the Sigmoid activation function:

[0064] ;

[0065] in, Indicates the first The predicted probability of a landslide occurring in each assessment unit;

[0066] Based on the landslide probability, a global prediction model is obtained. This model is then optimized using classification bias correction and spatial morphology constraints. The classification bias correction uses cross-entropy loss to measure the difference between the predicted probability and the true label, and the formula is as follows:

[0067] ;

[0068] in, The label represents the actual landslide, where 1 represents a landslide and 0 represents a non-landslide; N represents the total number of spatial units.

[0069] The spatial morphology constraint is based on the spatial continuity assumption of the graph structure, constraining the difference in prediction results between adjacent nodes, as shown in the formula:

[0070] ;

[0071] in Let be the set of all edges in the associated graph. The total number of edges;

[0072] The total loss function is obtained by correcting for classification bias and spatial morphology constraints. This total loss function is then used to optimize the global prediction model. The total loss function is:

[0073] ;

[0074] in The regularization coefficient is , and >0;

[0075] Furthermore, an end-to-end optimization approach is used to calculate the gradient of the total loss function with respect to all parameters of the global prediction model, and the parameters are iteratively updated using the Adam optimizer.

[0076] According to the aforementioned landslide hazard prediction system, the landslide hazard prediction unit is used for:

[0077] According to the preset probability threshold Predict probability The risk levels are classified as follows:

[0078] ;

[0079] By mapping the hazard level to the spatial coordinates of the target landslide area, a visualized landslide hazard zoning map is obtained.

[0080] The landslide hazard prediction system further includes an early warning unit and an edge computing unit. The early warning unit performs alarm processing based on the prediction results of the prediction unit through a geographic information system platform, a mobile terminal, and an audible and visual alarm device set up in the target landslide area. The edge computing unit is used to collaboratively process multi-source geological data and environmental data according to the request of the central server.

[0081] Another aspect of the present invention provides a landslide hazard prediction method, comprising:

[0082] Collect multi-source geological and environmental data of the target landslide area;

[0083] Based on multi-source geological and environmental data, multi-source disaster-causing factors are identified, and the multi-source disaster-causing factors are fused and normalized to obtain an impact factor matrix.

[0084] The influence factor matrix is ​​weighted and corrected according to the dynamic weighting mechanism to obtain an optimized factor matrix;

[0085] Determine the associated edges based on the preset node types, and determine the spatial association strength and semantic association degree between nodes based on the associated edges;

[0086] Based on the preset node type and preset associated edges, adaptive learning is adopted using feature mapping, association weight calculation and information aggregation to obtain the association strength and association features of the nodes;

[0087] Based on the association strength and association characteristics of the nodes, an association graph model is used for learning to obtain a global prediction model. The global prediction model is then subjected to classification bias correction, spatial morphology constraints and end-to-end optimization to obtain an optimized global prediction model.

[0088] The target landslide area is predicted by an optimized global prediction model, and the prediction results are classified into risk levels according to a preset probability threshold.

[0089] The beneficial effects of this invention are as follows: Through a dynamic weighting mechanism, the adaptive adjustment of disaster-causing factors in different regions is achieved, overcoming the spatial uniformity defect of traditional models; the fusion and normalization of multi-source data quantifies spatial correlation and semantic relationship through nodes and edges, improving data utilization efficiency; the improved network model can adaptively learn the influence intensity between elements and capture implicit patterns; the constructed global assessment framework combines single-element and element-related hazards, and through optimization, achieves more accurate prediction, providing a refined landslide hazard zoning map for decision-making and management. Attached Figure Description

[0090] Figure 1 This is a block diagram of a landslide hazard prediction system according to an embodiment of the present invention;

[0091] Figure 2 This is a schematic diagram of the network architecture of the landslide hazard prediction system according to an embodiment of the present invention;

[0092] Figure 3 This is a schematic diagram of the landslide hazard prediction method according to an embodiment of the present invention;

[0093] Figure 4 This is a schematic diagram comparing the performance of the global prediction model in this invention with that of random guessing, logistic regression, support vector machine, and random forest. Detailed Implementation

[0094] The embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings. Throughout the description, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions. In the following description, suffixes such as "module," "part," or "unit" used to denote elements are used only for the purpose of illustrative purposes and have no specific meaning in themselves. Therefore, "module," "part," or "unit" can be used interchangeably. Terms such as "first," "second," etc., are used only to distinguish technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the sequential relationship of the indicated technical features. In the following description, the consecutive reference numerals for method steps are for ease of review and understanding. Adjusting the implementation order of steps, in conjunction with the overall technical solution of the present invention and the logical relationship between the various steps, will not affect the technical effect achieved by the technical solution of the present invention. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0095] Figure 1 This is a block diagram of a landslide hazard prediction system according to an embodiment of the present invention. It includes a data acquisition unit 101 for acquiring multi-source geological and environmental data of the target landslide area; a feature construction unit 102 for determining multi-source disaster-causing factors based on the multi-source geological and environmental data, fusing and normalizing the multi-source disaster-causing factors to obtain an influence factor matrix; a spatial heterogeneous weighted modeling unit 103 for weighting and correcting the influence factor matrix according to a dynamic weighting mechanism to obtain an optimized factor matrix; a spatial and semantic association unit 104 for determining association edges based on preset node types, and determining the spatial association strength and semantic association degree between nodes based on the association edges; and adaptive learning. Unit 105 is used to obtain the association strength and association features of nodes by using feature mapping, association weight calculation and information aggregation adaptive learning according to preset node types and preset association edges; Multi-element interactive network calculation unit 106 is used to learn a global prediction model by using an association graph model according to the association strength and association features of nodes, and to apply classification bias correction, spatial morphology constraints and end-to-end optimization to the global prediction model to obtain an optimized global prediction model; Landslide hazard prediction unit 107 is used to predict the target landslide area through the optimized global prediction model, obtain the prediction result, and classify the hazard level of the prediction result according to a preset probability threshold.

[0096] In some embodiments, adjacent data acquisition units 101, feature construction units 102, spatial heterogeneous weighted modeling units 103, spatial and semantic association units 104, adaptive learning units 105, and multi-element interactive network computing units 106 can interact with each other and transmit control signals through wired communication (such as Ethernet, optical fiber) or wireless communication (such as Wi-Fi, 4G / 5G, LoRa, or satellite communication).

[0097] In some embodiments, reference Figure 2 The diagram shows the network architecture of the landslide hazard prediction system. The data acquisition unit 101 includes a three-dimensional lidar, ground-penetrating radar, soil moisture sensor, automatic weather station, GNSS monitoring point and satellite receiving device set up in the target landslide area to collect data, such as topographic data, geological structure data, hydrogeological monitoring data, meteorological data, GNSS monitoring data and remote sensing images. The collected multi-source geological data and environmental data are sent to the central server through at least one of the following methods: 4G / 5G, LoRa or satellite communication.

[0098] in, Figure 2 The network architecture diagram shown also includes an early warning unit and an edge computing unit. The early warning unit performs alarm processing through the geographic information system platform, mobile terminals, and audible and visual alarm devices set up in the target landslide area based on the prediction results of the prediction unit. The edge computing unit is used to collaboratively process multi-source geological data and environmental data according to the requests of the central server.

[0099] It should be noted that the edge computing unit is used to replace the central server's function when the central server is short of resources or the central server is geographically distant and cannot respond in time, thereby improving the data collection speed and response efficiency.

[0100] In some embodiments, the feature construction unit 102 is used to standardize, spatially align, imput missing values, and remove noise from multi-source data, and construct an input feature matrix (such as multi-source disaster-causing factors) based on geological and environmental characteristics, providing standardized data input for subsequent modeling. This unit can be implemented using a GPU server or an embedded edge computing device and communicates with the modeling unit through an internal bus interface.

[0101] In some embodiments, the spatial heterogeneous weighted modeling unit 103, the spatial and semantic association unit 104, and the adaptive learning unit 105 are used to achieve local weighted estimation and non-stationary spatial modeling based on the geological and environmental heterogeneity of different spatial locations. Specifically:

[0102] The spatial weight matrix is ​​determined using geographic coordinates and variograms; the optimal estimation parameters are obtained based on local sample data; spatial estimates of each sampling point are output based on a heterogeneous weighting mechanism; this is used to construct graph structure relationships among multi-source geological elements and explore nonlinear interaction patterns; a graph association network is constructed with geological elements as nodes and spatial proximity or physical correlation as edges; the weighted relationships between nodes are calculated using the graph association model; and spatial weighting information and multi-element features are fused to output comprehensive risk features. The global prediction model of this invention can be deployed on cloud servers, GPU clusters, or edge nodes to achieve efficient parallel computing.

[0103] In some embodiments, the feature construction unit 102 is used to obtain multiple disaster-causing factors of the target landslide area, and to fuse and normalize the multiple disaster-causing factors to obtain an influence factor matrix.

[0104] In some embodiments, the feature building unit 102 is used for:

[0105] The target landslide area is divided into multiple assessment units (e.g., 30m x 30m). The multi-source disaster-causing factors of each assessment unit include topographic features, geological features, triggering features, and disaster-bearing features. Topographic features include slope s, slope aspect a, and relative elevation difference h; geological features include lithology type l and distance from the fault zone f; triggering features include annual cumulative rainfall r and vegetation cover index v; and disaster-bearing features include distance from the road d.

[0106] The original feature matrix obtained after fusing topographic features, geological features, triggering features, and disaster-bearing features is as follows:

[0107] ;

[0108] in, ,in Indicates transpose. As an identifier for the evaluation unit number; To assess the number of units, The dimension represents the number of disaster-causing factors;

[0109] The original feature matrix is ​​mapped to the interval [0,1] using the min-max normalization method, resulting in the normalized influence factor matrix:

[0110] ;

[0111] in, This represents the normalization result. Represents the original feature matrix. The first sample One eigenvalue; For the first The minimum value of each feature among all samples. No. The maximum value of a feature across all samples.

[0112] In some embodiments, the spatial heterogeneous weighted modeling unit 103 is used for:

[0113] The influence factor matrix is ​​weighted and corrected using a dynamic weighting mechanism to obtain an optimized factor matrix, including:

[0114] Based on the spatial distribution characteristics of disaster-causing factors in the assessment unit, the dynamic weighting mechanism is used to calculate the weighting of each assessment unit. The local weights are:

[0115] ;

[0116] ;

[0117] The dynamic weighting mechanism is determined by the first law of geography. As an evaluation unit With evaluation unit Spatial distance, and As an evaluation unit coordinates and As an evaluation unit coordinates For adaptive bandwidth, According to the evaluation unit The density of surrounding disaster-causing factors is dynamically adjusted;

[0118] According to the evaluation unit The original feature matrix is ​​used to determine the local linear regression model of the geographic weighted regression (GWR). for:

[0119] ;

[0120] in, For local coefficient vectors, This is the error vector;

[0121] Based on the local linear regression model of geographically weighted regression (GWR) Local coefficients are determined by weighted least squares. for:

[0122] ;

[0123] in, It is a diagonal weight matrix;

[0124] Based on local coefficients The corrected feature matrix is ​​determined as follows:

[0125] ;

[0126] in, The factor matrix is ​​a diagonal matrix of local coefficients. The influence factor matrix is ​​then weighted and corrected using the modified characteristic matrix to obtain the optimized factor matrix. for:

[0127] ;

[0128] In some embodiments, the spatial and semantic association unit 104 is used for:

[0129] Encode the preset node type using one-hot vectors to obtain the node. Type encoding for:

[0130] ;

[0131] Optimize the factor matrix Fusion features with type encoding for:

[0132] ;

[0133] Based on the fusion characteristics, the associated edges are determined, and the spatial association strength and semantic association degree between nodes are determined based on the associated edges. The associated edges include intra-elements of the same type of elements, cross-element interaction edges, and trigger-disaster orientation edges.

[0134] In this context, inline edges of the same type are used for nodes of the same type, and the edge weight of inline edges of the same type is specified. To convey information about the continuity of similar elements:

[0135] ;

[0136] Among them, cross-category feature interaction edges are used to connect nodes of different types, and the edge weights of cross-category feature interaction edges are... For feature similarity:

[0137] ;

[0138] in, For nodes The fused feature vector, for transpose, The norm is used; the trigger-disaster directional edge indicates that it only points from the trigger node to the disaster-bearing node, and the edge weight of the trigger-disaster directional edge indicates the one-way threat of the triggering factor to the disaster-bearing body.

[0139] In some embodiments, the adaptive learning unit 105 is used for:

[0140] Mapping features to nodes and neighboring nodes Type-specific linear transformations:

[0141] ;

[0142] in, For type-specific weight matrices, , and For type-specific linear transformation results, For nodes The query vector, For nodes The key vector, For nodes Value vector, node The type is ,node The type of the neighbor node is ;

[0143] The association weight calculation includes fusing spatial distance and feature similarity to calculate the association strength between nodes. for:

[0144] ;

[0145] in, For attention parameter vectors, For nodes The neighborhood group, For activation function, For nodes With nodes Local weights; For nodes The The key vectors of the neighboring nodes, Node is With the Edge characteristics of each neighboring node, For nodes With the Local weights of each neighboring node;

[0146] Based on the strength of the association between nodes By aggregating information from surrounding nodes, the association features of the nodes are obtained. for:

[0147] ;

[0148] Among them, association features In the formula This is the activation function.

[0149] In some embodiments, the multi-factor interactive network computing unit 106 is used for:

[0150] Get Nodes Association features The pre-activation value is obtained by transforming it through a fully connected layer. :

[0151] ;

[0152] in, This is the output layer weight matrix. For bias terms, pre-activation values In the formula Use the Sigmoid activation function;

[0153] The pre-activation values ​​are mapped to landslide probabilities using the Sigmoid activation function:

[0154] ;

[0155] in, Indicates the first The predicted probability of a landslide occurring in each assessment unit;

[0156] Based on the landslide probability, a global prediction model is obtained. This model is then optimized using classification bias correction and spatial morphology constraints. The classification bias correction uses cross-entropy loss to measure the difference between the predicted probability and the true label, and the formula is as follows:

[0157] ;

[0158] in, The label represents the actual landslide (1 represents a landslide, 0 represents a non-landslide), and N represents the total number of spatial units.

[0159] The spatial morphology constraint is based on the spatial continuity assumption of the graph structure, constraining the difference in prediction results between adjacent nodes, as shown in the formula:

[0160] ;

[0161] in Let be the set of all edges in the associated graph. The total number of edges;

[0162] The total loss function is obtained by correcting for classification bias and spatial morphology constraints. This total loss function is then used to optimize the global prediction model. The total loss function is:

[0163] ;

[0164] in The regularization coefficient is , and >0;

[0165] Furthermore, the gradient of the total loss function with respect to all parameters of the global prediction model is calculated using an end-to-end optimization approach, and the parameters are iteratively updated through the Adam optimizer to complete the end-to-end optimization.

[0166] In some embodiments, output layer weights For example, the detailed derivation is as follows:

[0167] According to the gradient chain rule, the total loss is... The gradient is:

[0168] ;

[0169] The gradient of the cross-entropy loss is decomposed layer by layer using the chain rule:

[0170] ;

[0171] Among them, the first item The second item The third item The result obtained by merging is: ;

[0172] Similarly, the gradient of each edge (i, j) is calculated and summed:

[0173] ;

[0174] The parameter update formula is:

[0175] ;

[0176] in The learning rate is used, and the gradient is calculated via backpropagation. This represents the total number of edges in the association graph. The optimization process described above also applies to all parameters of the model. Gradients are calculated and updated through backpropagation layer by layer, ultimately yielding the total loss. Minimize the value to ensure that the prediction results both fit the actual landslide distribution and meet the spatial continuity constraint.

[0177] In some embodiments, the landslide hazard prediction unit 107 is used for:

[0178] According to the preset probability threshold Predict probability The risk levels are classified as follows:

[0179] ;

[0180] By mapping the hazard level to the spatial coordinates of the target landslide area, a visualized landslide hazard zoning map is obtained.

[0181] In some embodiments, extremely high-risk areas (such as areas with overlapping "steep slopes + fault zones + high rainfall") can be designated as key prevention and control areas.

[0182] In some embodiments, reference Figure 3 The landslide hazard prediction flowchart shown includes, but is not limited to, steps S100~S700:

[0183] S100, collects multi-source geological and environmental data of the target landslide area;

[0184] S200: Based on multi-source geological and environmental data, multi-source disaster-causing factors are determined, and the multi-source disaster-causing factors are fused and normalized to obtain the influence factor matrix.

[0185] S300, the influence factor matrix is ​​weighted and corrected according to the dynamic weighting mechanism to obtain the optimized factor matrix;

[0186] S400: Determine the associated edges based on the preset node type, and determine the spatial association strength and semantic association degree between nodes based on the associated edges;

[0187] S500, based on preset node types and preset associated edges, uses feature mapping, associated weight calculation and information aggregation adaptive learning to obtain the association strength and associated features of nodes;

[0188] S600 uses a correlation graph model to learn based on the correlation strength and correlation characteristics of nodes, obtains a global prediction model, and applies classification bias correction, spatial morphology constraints and end-to-end optimization to the global prediction model to obtain an optimized global prediction model.

[0189] The S700 uses an optimized global prediction model to predict the target landslide area, obtains the prediction results, and classifies the risk level of the prediction results according to a preset probability threshold.

[0190] In some embodiments, geological hazard survey data and spatial distribution information of a landslide area were used to construct a two-dimensional raster model of the study area, ultimately generating 1120 two-dimensional spatial raster cells. These cells were determined to be landslide units based on historical landslide records, field verification results, and stability assessment criteria (i.e., a landslide occurrence probability greater than a critical threshold). Landslide influencing factors were then extracted, assigning landslide-sensitive factors such as topographic slope, aspect, soil and rock type, vegetation cover, rainfall, and distance from faults to each raster cell. The entire study area was divided into equally spaced square raster cells with a resolution of 30 meters, each cell having an area of ​​900 square meters. To simulate the spatially continuous distribution of landslide-sensitive factors, an inverse distance weighting method was used to interpolate the discrete survey data, and the interpolation results were accurately allocated to the spatial raster cells in the model. After processing, 1120 two-dimensional cells containing complete landslide influencing factors were finally obtained, including 560 landslide cells (determined based on historical landslide boundaries and grid space overlay analysis) and 560 non-landslide cells (selected from stable areas outside high-incidence landslide zones and typical stable grid cells without historical landslide records).

[0191] Figure 4 This diagram illustrates a comparison of the global prediction model of this invention with models such as random guessing, logistic regression, support vector machine, and random forest. In landslide prediction, the problem is modeled as a binary classification task: landslide units in the dataset are labeled as label 1, and non-landslide units are labeled as label 0. Based on the spatial location information of landslide units, topographic features, and various landslide-inducing factors, a feature dataset for model training was constructed. Subsequently, the dataset was divided into training and test sets in an 8:2 ratio for model training. To verify the performance advantage of the proposed model, logistic regression (LR), support vector machine (SVM), and random forest (RF) were selected as comparison models for accuracy evaluation. Experimental results show that the proposed model is more competitive in prediction performance, and a landslide prediction model (LPM) was constructed based on this.

[0192] The model implementation in this embodiment is based on the PyTorch-1.8 framework and trained using an NVIDIA GeForce GTX1060Ti GPU. During training, the inference algorithm of the spatial correlation layer undergoes 20 iterations to ensure the stability of the optimization effect. A batch size of 32 is used during training. To improve the model's generalization ability and stability, stochastic gradient descent (SGD) is selected as the optimizer, with a learning rate set to 1e-5. The overall training process lasts for 60 epochs. Furthermore, to reduce the impact of parameter initialization randomness on model performance, each network architecture is independently trained 5 times using different random seeds to enhance the reliability of the results. The final trained model possesses the ability to capture spatial correlations between grid cells; its specific performance can be found in [link to relevant documentation]. Figure 4 .

[0193] Furthermore, although the invention has been described in the context of functional modules, it should be understood that, unless otherwise stated, one or more of the described functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding the invention. Rather, considering the properties, functions, and internal relationships of the various functional modules in the apparatus disclosed in the embodiments of the invention, the actual implementation of the module will be understood within the scope of conventional skill of an engineer. Therefore, those skilled in the art can implement the invention as set forth in the claims using ordinary techniques without excessive experimentation. It is also understood that the specific concepts disclosed are merely illustrative and are not intended to limit the scope of the invention, which is determined by the full scope of the appended claims and their equivalents.

[0194] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0195] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can include, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0196] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0197] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0198] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0199] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

[0200] The above is a detailed description of the preferred embodiments of the present invention, but the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention, and these equivalent modifications or substitutions are all included within the scope defined by the claims of this application.

Claims

1. A landslide hazard prediction system, characterized in that, include: The data acquisition unit is used to collect multi-source geological and environmental data of the target landslide area; The feature construction unit is used to determine multi-source disaster-causing factors based on multi-source geological data and environmental data, and to fuse and normalize the multi-source disaster-causing factors to obtain the influence factor matrix. A spatially heterogeneous weighted modeling unit is used to perform weighted correction on the influence factor matrix according to a dynamic weighting mechanism to obtain an optimized factor matrix; Spatial and semantic association unit, used to determine association edges according to preset node types, and to determine the spatial association strength and semantic association degree between nodes according to the association edges; The adaptive learning unit is used to obtain the association strength and association features of nodes by adaptive learning through feature mapping, association weight calculation and information aggregation based on the preset node type and preset association edges. The multi-element interactive network computing unit is used to learn a global prediction model based on the association strength and association characteristics of nodes using an association graph model, and to apply classification bias correction, spatial morphology constraints and end-to-end optimization to the global prediction model to obtain an optimized global prediction model. The landslide hazard prediction unit is used to predict the target landslide area through an optimized global prediction model, obtain the prediction results, and classify the hazard level of the prediction results according to a preset probability threshold. The adaptive learning unit is used for: Map features to nodes and neighboring nodes Type-specific linear transformations: in, For type-specific weight matrices, , and For type-specific linear transformation results, For nodes The query vector, For nodes The key vector, For nodes Value vector, node The type is ,node The type of the neighbor node is ; The association weight calculation includes fusing spatial distance and feature similarity to calculate the association strength between nodes. for: in, For attention parameter vectors, For nodes The neighborhood group, For activation function, For nodes With nodes Local weights; For nodes The The key vectors of the neighboring nodes, Node is With the Edge characteristics of each neighboring node, For nodes With the Local weights of each neighboring node; Based on the strength of the association between nodes By aggregating information from surrounding nodes, the association features of the nodes are obtained. for: Among them, association features In the formula For activation functions; The multi-element interactive network computing unit is used for: Get Nodes Association features The pre-activation value is obtained by transforming it through a fully connected layer. : in, This is the output layer weight matrix. For bias terms, pre-activation values In the formula Use the Sigmoid activation function; The pre-activation values ​​are mapped to landslide probabilities using the Sigmoid activation function: in, Indicates the first The predicted probability of a landslide occurring in each assessment unit; Based on the landslide probability, a global prediction model is obtained. This model is then optimized using classification bias correction and spatial morphology constraints. The classification bias correction uses cross-entropy loss to measure the difference between the predicted probability and the true label, and the formula is as follows: in, The label represents the actual value, where 1 indicates a landslide and 0 indicates a non-landslide; N represents the total number of spatial units. The spatial morphology constraint is based on the spatial continuity assumption of the graph structure, constraining the difference in prediction results between adjacent nodes, as shown in the formula: in Let be the set of all edges in the associated graph. The total number of edges; The total loss function is obtained by correcting for classification bias and spatial morphology constraints. This total loss function is then used to optimize the global prediction model. The total loss function is: in The regularization coefficient is , and >0; Furthermore, an end-to-end optimization approach is used to calculate the gradient of the total loss function with respect to all parameters of the global prediction model, and the parameters are iteratively updated using the Adam optimizer.

2. The landslide hazard prediction system according to claim 1, characterized in that, The data acquisition unit is used for: Data is collected from the target landslide area using three-dimensional lidar, ground-penetrating radar, soil moisture sensors, automatic weather stations, GNSS monitoring points, and satellite receiving devices to obtain topographic data, geological structure data, hydrogeological monitoring data, meteorological data, GNSS monitoring data, and remote sensing images. The collected multi-source geological and environmental data are then transmitted to the central server via at least one of the following methods: 4G / 5G, LoRa, or satellite communication.

3. The landslide hazard prediction system according to claim 1, characterized in that, The feature construction unit is used for: The target landslide area is divided into multiple assessment units. Based on multi-source geological and environmental data, the multi-source disaster-causing factors of the assessment units are determined. These multi-source disaster-causing factors include topographic features, geological features, triggering features, and disaster-bearing features. Topographic features include slope s, slope aspect a, and relative elevation difference h. Geological features include lithology type l and fault zone distance f. Triggering characteristics include annual cumulative rainfall (r) and vegetation cover index (v); disaster-bearing characteristics include distance from roads. ; The original feature matrix obtained after fusing topographic features, geological features, triggering features, and disaster-bearing features is as follows: in, ,in Indicates transpose. As an identifier for the evaluation unit number; The dimension represents the number of disaster-causing factors; To evaluate the number of units; The original feature matrix is ​​mapped to the interval [0,1] using a normalization method, resulting in the normalized influence factor matrix: in, This represents the normalization result. Represents the original feature matrix. The first sample One eigenvalue; For the first The minimum value of each feature among all samples. No. The maximum value of a feature across all samples.

4. The landslide hazard prediction system according to claim 3, characterized in that, The spatial heterogeneous weighted modeling unit is used for: Based on the spatial distribution characteristics of disaster-causing factors in the assessment unit, the dynamic weighting mechanism is used to calculate the weighting of each assessment unit. The local weights are: The dynamic weighting mechanism is determined by the first law of geography. As an evaluation unit With evaluation unit Spatial distance, and As an evaluation unit coordinates and As an evaluation unit coordinates For adaptive bandwidth, According to the evaluation unit The density of surrounding disaster-causing factors is dynamically adjusted; According to the evaluation unit The original feature matrix is ​​used to determine the local linear regression model of the geographic weighted regression (GWR). for: in, For local coefficient vectors, This is the error vector; Based on the local linear regression model of geographically weighted regression (GWR) Local coefficients are determined by weighted least squares. for: in, It is a diagonal weight matrix; Based on local coefficients The corrected feature matrix is ​​determined as follows: in, The factor matrix is ​​a diagonal matrix of local coefficients. The influence factor matrix is ​​then weighted and corrected using the modified characteristic matrix to obtain the optimized factor matrix. for: 。 5. The landslide hazard prediction system according to claim 4, characterized in that, The spatial and semantic association unit is used for: Encode the preset node type using one-hot vectors to obtain the node. Type encoding for: Optimize the factor matrix Fusion features with type encoding for: Based on the fusion characteristics, the associated edges are determined, and the spatial association strength and semantic association degree between nodes are determined based on the associated edges. The associated edges include intra-elements of the same type of elements, cross-element interaction edges, and trigger-disaster orientation edges. In this context, inline edges of the same type are used for nodes of the same type, and the edge weight of inline edges of the same type is specified. To convey information about the continuity of similar elements: Among them, cross-category feature interaction edges are used to connect nodes of different types, and the edge weights of cross-category feature interaction edges are... For feature similarity in, For nodes The fused feature vector, for transpose, The norm is used; the trigger-disaster directional edge indicates that it only points from the trigger node to the disaster-bearing node, and the edge weight of the trigger-disaster directional edge indicates the one-way threat of the triggering factor to the disaster-bearing body.

6. The landslide hazard prediction system according to claim 1, characterized in that, The landslide hazard prediction unit is used for: According to the preset probability threshold Predict probability The risk levels are classified as follows: By mapping the hazard level to the spatial coordinates of the target landslide area, a visualized landslide hazard zoning map is obtained.

7. The landslide hazard prediction system according to claim 6, characterized in that, It also includes an early warning unit and an edge computing unit. The early warning unit performs alarm processing through a geographic information system platform, a mobile terminal, and an audible and visual alarm device set up in the target landslide area based on the prediction results of the prediction unit. The edge computing unit is used to collaboratively process multi-source geological data and environmental data according to the requests of the central server.

8. A landslide hazard prediction method for the system according to any one of claims 1-7, characterized in that, include: Collect multi-source geological and environmental data of the target landslide area; Based on multi-source geological and environmental data, multi-source disaster-causing factors are identified, and the multi-source disaster-causing factors are fused and normalized to obtain an impact factor matrix. The influence factor matrix is ​​weighted and corrected according to the dynamic weighting mechanism to obtain an optimized factor matrix; Determine the associated edges based on the preset node types, and determine the spatial association strength and semantic association degree between nodes based on the associated edges; Based on the preset node type and preset associated edges, adaptive learning is adopted using feature mapping, association weight calculation and information aggregation to obtain the association strength and association features of the nodes; Based on the association strength and association characteristics of the nodes, an association graph model is used for learning to obtain a global prediction model. The global prediction model is then subjected to classification bias correction, spatial morphology constraints and end-to-end optimization to obtain an optimized global prediction model. The target landslide area is predicted by an optimized global prediction model, and the prediction results are classified into risk levels according to a preset probability threshold.

Images