Regional-level landslide prediction method and device based on multi-mode decomposition and physical mechanism

Abstract

The invention provides a regional-level landslide prediction method and device based on multi-modal decomposition and a physical mechanism, and belongs to the technical field of landslide prediction.The method comprises the steps that a trend item, a periodic item and a burst item in historical landslide deformation data of a landslide area are combined, and a predicted value of landslide deformation in the landslide area at the next moment is obtained; according to the method, landslide prediction under the combined action of long-term events, periodic events and emergencies is realized, and under the condition that the landslide area is unstable, a predicted value of landslide deformation of the landslide area at the next moment and an adjacent landslide area with potential linkage instability are used as landslide prediction information of the landslide area. Therefore, the sudden landslide phenomenon and the linkage instability landslide phenomenon can be predicted, and the prediction effect of landslide prediction is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of landslide prediction, and in particular to a regional landslide prediction method and device based on multi-modal decomposition and physical mechanism. Background Art

[0002] Landslides, as a common geological disaster, are characterized by suddenness and cascading instability. Cascading instability occurs when an unstable landslide exerts stress on adjacent areas, causing landslides in adjacent areas within a short period of time and generating secondary disasters. This cascading instability further amplifies the landslide's destructiveness.

[0003] However, the existing landslide prediction methods are insufficient in predicting sudden landslides and cascading unstable landslides, resulting in poor prediction results. Summary of the Invention

[0004] The present invention proposes a regional landslide prediction method and device based on multimodal decomposition and physical mechanism. The method combines the trend item, period item and sudden item in the historical landslide deformation data of the landslide area to obtain the predicted value of the landslide deformation in the landslide area at the next moment, thereby realizing the prediction of landslides under the combined action of long-term events, periodic events and sudden events. In the event of instability in the landslide area, the predicted value of the landslide deformation in the landslide area at the next moment and the adjacent landslide areas with potential chain instability are used as landslide prediction information of the landslide area, thereby realizing the prediction of sudden landslide phenomena and chain instability landslide phenomena, thereby improving the prediction effect of landslide prediction.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] In a first aspect, the present invention provides a regional landslide prediction method based on multimodal decomposition and physical mechanisms, comprising: clustering the target region based on geological data, topographic data, and environmental factors of the target region to obtain multiple landslide regions within the target region; each of the multiple landslide regions belongs to only one geological unit; a geological unit is a stratum or rock mass region with consistent geotechnical properties and geological history. For each of the multiple landslide regions, based on the trend term, periodic term, and sudden term in the historical landslide deformation data of the landslide region, the trend landslide deformation variable, periodic landslide deformation variable, and sudden landslide deformation variable of the landslide region at the next moment are predicted; wherein the trend term indicates a long-term event within the landslide region; the periodic term indicates a periodic event within the landslide region; and the sudden term indicates a sudden event within the landslide region. The weighted sum of the trend landslide deformation variable, periodic landslide deformation variable, and sudden landslide deformation variable of the landslide region at the next moment is used as the predicted value of the landslide deformation within the landslide region at the next moment; the weighting coefficients of the trend landslide deformation variable, periodic landslide deformation variable, and sudden landslide deformation variable are determined by the type of landslide region. When a landslide area becomes unstable, adjacent landslide areas with potential cascading instability among the multiple adjacent landslide areas are determined based on the stress increment and pore water pressure coupling between the landslide area and each of the multiple adjacent landslide areas of the landslide area. The predicted value of the landslide deformation of the landslide area at the next moment and the adjacent landslide areas with potential cascading instability are used as the landslide prediction information of the landslide area. The instability of the landslide area is determined based on the landslide deformation acceleration of the landslide area within the time period at the current moment being greater than the instability threshold.

[0007] The present invention provides a regional landslide prediction method based on multimodal decomposition and physical mechanisms. The target area is first divided into multiple landslide regions based on the geological data, topographic data, and environmental factors of the target area. Each landslide region obtained after the division belongs to only one geological unit. For each landslide region, the trend term, periodic term, and sudden term in the historical landslide deformation data of the landslide region are combined to obtain a predicted value of the landslide deformation in the landslide region at the next moment, thereby achieving a prediction of landslides in the event of a combined effect of long-term events, periodic events, and sudden events. Furthermore, in the event of instability in the landslide region, the predicted value of the landslide deformation in the landslide region at the next moment and the adjacent landslide regions with potential cascading instability are used as landslide prediction information for the landslide region. In the process of dividing the target area into multiple landslide regions, the above method ensures that each landslide region has uniform geological conditions, thereby increasing the reliability of the stress increment and pore water pressure coupling between the landslide region and multiple adjacent landslide regions, and making the subsequent judgment of potential cascading instability more accurate, thereby achieving the prediction of sudden landslide phenomena and cascading instability landslide phenomena, thereby improving the predictive effect of landslide prediction.

[0008] In one implementation of the first aspect, clustering is performed on the target area to obtain multiple landslide areas within the target area, including:

[0009] Construct a characteristic matrix of the target area; wherein the target area includes multiple grid cells; the characteristic matrix of the target area includes a characteristic vector of each grid cell in the multiple grid cells; the characteristic vector of each grid cell includes geological data, topographic data and environmental factors corresponding to the grid cell; the characteristic vector satisfies: F = [P geo ,T topo ,M eteo ], F represents the eigenvector, P geo Represents geological data, T topo Represents terrain data, M eteo represents environmental factors;

[0010] The characteristic matrix of the target area is used to construct a constrained similarity matrix; the constrained similarity matrix satisfies: Among them, S represents the constraint similarity matrix; s 1,1 Indicates the similarity between the first grid unit and the first grid unit, s n,1 Indicates the similarity between the nth grid unit and the first grid unit, s 1,n Indicates the similarity between the first grid unit and the nth grid unit, s n,n It represents the similarity between the nth grid cell and the nth grid cell, where n represents the total number of grid cells in the target area. The similarity satisfies: Among them, s i,j represents the similarity between the i-th grid unit and the j-th grid unit, and 1≤i≤n and 1≤j≤n; F i represents the feature vector of the i-th grid cell, F j represents the eigenvector of the jth grid cell; σ represents the width of the kernel function;

[0011] By constraining the similarity matrix and the degree matrix, a Laplace matrix is ​​constructed; the Laplace matrix satisfies: J = DS, where J represents the Laplace matrix, D represents the degree matrix, and the degree matrix is ​​a diagonal matrix and

[0012] Perform eigendecomposition on the Laplace matrix to obtain the cluster eigenvector matrix;

[0013] K-means clustering is performed on the clustering feature vector matrix to obtain multiple clusters, and the area in the target area corresponding to each cluster in the multiple clusters is determined as the landslide area, thereby obtaining multiple landslide areas in the target area.

[0014] In an implementation of the first aspect, determining adjacent landslide areas with potential cascading instability among a plurality of adjacent landslide areas includes:

[0015] For each of the multiple adjacent landslide areas in the landslide area, the safety factor of the adjacent landslide area is calculated by coupling the stress increment and pore water pressure between the landslide area and the adjacent landslide area; the safety factor satisfies the following formula:

[0016]

[0017] Where c represents the soil cohesion, σ n represents the normal stress generated by the landslide area on the adjacent landslide area, Δσ B It represents the stress increment generated by the landslide area on the adjacent landslide area at the current moment, p new represents the pore water pressure between the landslide area and the adjacent landslide area at the current moment, and p new =p+γ w Δh w , p represents the pore water pressure Δh between the landslide area and the adjacent landslide area before the current moment w represents the groundwater level rise in the adjacent landslide area, γ w represents water density; φ represents the internal friction angle of the landslide, γ represents the soil density, h represents the thickness of the saturated zone adjacent to the landslide area, θ represents the slope of the landslide, τ transfer It represents the shear stress transferred from the landslide area to the adjacent landslide area;

[0018] Among multiple adjacent landslide areas, the adjacent landslide areas with safety factors greater than the safety threshold are identified as adjacent landslide areas with potential cascading instability.

[0019] In an implementation manner of the first aspect, the method further includes: for each landslide area among the multiple landslide areas, extracting a trend item, a period item, and a burst item from historical landslide deformation data of the landslide area.

[0020] In an implementation of the first aspect, extracting trend items, period items, and burst items from historical landslide deformation data of a landslide area includes:

[0021] After injecting noise into the historical landslide deformation data of the landslide area using an adaptive noise injection method, the first original landslide signal is obtained;

[0022] A modal decomposition method is used to extract a trend term from the first original landslide signal, and the remaining signal in the first original landslide signal is used as the second original landslide signal;

[0023] extracting a periodic term from the second original landslide signal by continuous wavelet transform, and taking the remaining signal in the second original landslide signal as the third original landslide signal;

[0024] Determine a signal in the third original landslide signal whose rainfall index is greater than a rainfall threshold or / and whose cumulative displacement is greater than an acceleration threshold as a sudden item;

[0025] The rainfall threshold satisfies the following formula;

[0026]

[0027] Where t represents time, I r (t) represents the rainfall threshold, c(t) represents the soil cohesion in the landslide area during the current time period, γ w represents the water density, h(t) represents the saturated zone thickness of the landslide area in the time period at the current moment, φ represents the internal friction angle of the landslide, γ represents the soil density, and θ(t) represents the internal friction angle of the landslide in the landslide area in the time period at the current moment. The rainfall index is calculated based on the environmental factors.

[0028] The acceleration threshold satisfies the following formula;

[0029]

[0030] Among them, k c (t) represents the acceleration threshold, and g represents the acceleration due to gravity.

[0031] In one implementation of the first aspect, predicting the trend landslide deformation variable, the periodic landslide deformation variable, and the sudden landslide deformation variable of the landslide area at the next moment includes:

[0032] The trend item is input into the trend item prediction model, and the trend item prediction model outputs the trend landslide deformation of the landslide area at the next moment; the trend item prediction model is a Transformer model, and the position adaptive attention of the trend item prediction model is calculated based on the time-related weight coefficient;

[0033] The periodic term is input into the periodic term prediction model, and the periodic term prediction model outputs the periodic landslide deformation variable of the landslide area at the next moment; the periodic term prediction model is the manba model; the residual block in the periodic term prediction model outputs the environmental factor weight;

[0034] The sudden item is input into the sudden item prediction model, and the sudden item prediction model outputs the sudden landslide deformation variable of the landslide area at the next moment; the sudden item prediction model is a Temporal Fusion Transformer model, and the physical constraint attention of the sudden item prediction model is calculated based on the rainfall threshold and acceleration threshold.

[0035] In one implementation of the first aspect, a calculation formula for the position-adaptive attention of the trend item prediction model is as follows:

[0036]

[0037] Where Q1 = X l W Q , W Q represents the first weight matrix, X l represents the input of the trend term prediction model, and X l =[L(t),E,ν,v creep ]∈R T×4 , t represents time, L(t) represents trend term, E represents elastic modulus, ν represents Poisson's ratio, v creep represents the creep rate, R T×4 Represents X l It consists of four vectors of dimension T, where T represents the time series; K1 = X l W K , W K Represents the second weight matrix; V1 = X l W V , W V represents the third weight matrix; Attention1(Q1,K1,V1) represents position adaptive attention; w pos (t) represents the time-dependent weight coefficient, and t pred represents the time point of the next moment; softmax(·) represents the normalization function; d k represents the data dimension, t represents time, w pos (t) represents the time-related weight coefficient, M time represents the lower triangular mask matrix, and

[0038] The calculation formula of the physical constraint attention of the burst item prediction model is as follows;

[0039]

[0040] Where Q2 = Linear([B(t), E(t)]), Linear(·) represents the linear transformation function, B(t) represents the burst term, E(t) represents the event marker, E(t)∈{0,1}, and E(t)=1 means R(t)≥I r (t) or / and a(t)≥k c (t), R(t) represents rainfall, I r (t) represents the rainfall threshold, a(t) represents the deformation acceleration, k c (t) represents the acceleration threshold; E(t) = 0 represents R(t) <Ir (t) and a(t) <k c (t); Attention2(Q,K,V) represents physical constraint attention.

[0041] In an implementation of the first aspect, the method further comprises: adjusting the trend landslide deformation variable, the periodic landslide deformation variable, and the sudden landslide deformation variable of the landslide area with potential chain landslides in the multiple landslide areas at the next moment based on the synergistic relationship between the multiple landslide areas.

[0042] In an implementation of the first aspect, the type of landslide area is rainfall-dominated, freeze-thaw-sensitive, clay creep, debris flow potential, steep slope unloading, gentle slope potential sliding, rapid response to sudden impact, seasonal periodic, or composite.

[0043] In a second aspect, the present invention provides a regional landslide prediction device based on multimodal decomposition and physical mechanisms, comprising a clustering module, a first prediction module, a second prediction module, and a third prediction module. The clustering module is used to cluster the target area based on the geological data, topographic data, and environmental factors of the target area to obtain multiple landslide areas within the target area; each of the multiple landslide areas belongs to only one geological unit; a geological unit is a stratum or rock mass area with consistent geotechnical properties and geological history. The first prediction module is used to, for each of the multiple landslide areas, predict the trend landslide deformation variable, periodic landslide deformation variable, and sudden landslide deformation variable of the landslide area at the next moment based on the trend item, periodic item, and sudden item in the historical landslide deformation data of the landslide area; wherein the trend item indicates a long-term event in the landslide area; the periodic item indicates a periodic event in the landslide area; and the sudden item indicates a sudden event in the landslide area. The second prediction module is configured to use the weighted sum of the trend, periodic, and sudden landslide deformation variables of the landslide region at the next moment as the predicted value of the landslide deformation within the landslide region at the next moment; the weighting coefficients of the trend, periodic, and sudden landslide deformation variables are determined by the type of landslide region. The third prediction module is configured to, when the landslide region becomes unstable, determine adjacent landslide regions with potential cascading instability within the multiple adjacent landslide regions based on the stress increment and pore water pressure coupling between the landslide region and each of the multiple adjacent landslide regions within the landslide region; and use the predicted value of the landslide deformation of the landslide region at the next moment and the adjacent landslide regions with potential cascading instability as landslide prediction information for the landslide region; wherein, the instability of the landslide region is determined based on the landslide deformation acceleration of the landslide region within the time period at the current moment being greater than the instability threshold.

[0044] In a third aspect, the present invention provides an electronic device comprising a processor and a memory coupled to the processor; the memory is used to store computer instructions, and when the electronic device is running, the processor executes the computer instructions stored in the memory, so that the electronic device performs the method described in the first aspect or any one of its implementations.

[0045] In a fourth aspect, the present invention provides a computer-readable storage medium comprising computer program instructions, which, when executed by a computer, enable the computer to execute the method as described in the first aspect or any one of its implementations.

[0046] In a fifth aspect, the present invention provides a computer program product, comprising computer program instructions, which, when executed on a computer, enable the computer to execute the method as described in the first aspect or any one of its implementations.

[0047] The technical effects corresponding to the above-mentioned second to fifth aspects and their possible implementation methods can refer to the above-mentioned description of the technical effects of the first aspect and its possible implementation methods, and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is one of the schematic diagrams of the regional landslide prediction method based on multimodal decomposition and physical mechanism provided in the embodiments of the present application;

[0049] Figure 2 This is the second schematic diagram of the regional landslide prediction method based on multimodal decomposition and physical mechanism provided in the embodiment of the present application;

[0050] Figure 3 This is the third schematic diagram of the regional landslide prediction method based on multimodal decomposition and physical mechanism provided in the embodiment of the present application;

[0051] Figure 4 This is the fourth schematic diagram of the regional landslide prediction method based on multimodal decomposition and physical mechanism provided in the embodiment of the present application;

[0052] Figure 5 This is the fifth schematic diagram of the regional landslide prediction method based on multimodal decomposition and physical mechanism provided in the embodiment of the present application;

[0053] Figure 6 It is a structural diagram of a regional landslide prediction device based on multimodal decomposition and physical mechanism provided in an embodiment of the present application. DETAILED DESCRIPTION

[0054] In the description and claims of the present invention, the terms "first" and "second" are used to distinguish different objects rather than to describe a specific order of objects.

[0055] In the embodiments of the present application, “and / or” represents the relationship between objects. For example, A and / or B can represent the following three situations: A exists alone, B exists alone, and A and B exist at the same time.

[0056] In the embodiments of this application, words such as "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in the embodiments of this application should not be interpreted as being preferred or advantageous over other embodiments or designs. Rather, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0057] In the description of the present invention, unless otherwise specified, "multiple" means two or more. For example, "multiple landslide areas" means two or more landslide areas; "multiple adjacent landslide areas" means two or more adjacent landslide areas.

[0058] The method and apparatus provided in the embodiments of the present application relate to a regional landslide prediction method and apparatus based on multimodal decomposition and physical mechanisms, which can be used for landslide prediction, especially for the prediction of sudden landslides and cascading unstable landslides.

[0059] To address the problem in prior art that existing landslide prediction methods are insufficient in predicting sudden landslides and cascading landslides, resulting in poor prediction results, the present invention provides a regional landslide prediction method and device based on multimodal decomposition and physical mechanisms. By combining the trend, periodic, and sudden terms in the historical landslide deformation data of the landslide region, a predicted value of the landslide deformation within the landslide region at the next moment is obtained, enabling the prediction of landslides caused by the combined effects of long-term events, periodic events, and sudden events. Furthermore, in the event of instability in the landslide region, the predicted value of the landslide deformation within the landslide region at the next moment and the adjacent landslide regions with potential cascading instability are used as landslide prediction information for the landslide region, thereby enabling the prediction of sudden landslides and cascading landslides, thereby improving the prediction results of landslide prediction.

[0060] For example, the regional landslide prediction method based on multimodal decomposition and physical mechanisms provided in embodiments of the present invention can be executed by an electronic device with processing capabilities, such as a computer or server. For example, the hardware components of a computer may include a processor, memory, a network interface, a user interface, a communication bus, and the like.

[0061] The processor is used to control the electronic device to perform related processing and computing tasks, such as clustering target areas, predicting the trend, periodic, and sudden landslide deformation of the landslide area at the next moment, and determining landslide prediction information for the landslide area. The processor may include a central processing unit (CPU) or other processors. The processor may be single-core or multi-core, for example, the processor may include multiple CPUs.

[0062] The memory is used to store computer instructions and related data, such as geological data, topographic data, and environmental factors of the target area, trend items, period items, and sudden items in historical landslide deformation data of the landslide area, predicted values ​​of landslide deformation in the landslide area at the next moment, and landslide prediction information for the landslide area. The memory can be a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM), a flash memory, an optical storage device, a magnetic disk storage medium, or other magnetic storage device, or any other medium capable of storing program code or data accessible by a computer. Optionally, the memory can be integrated into the processor, or the memory can be independent of the processor.

[0063] The network interface is used for the computer to communicate with other devices or communication networks. The network interface can be a transceiver with transceiver functions. Optionally, the network interface can include a standard wired interface or a wireless interface (such as a Wi-Fi interface, a Bluetooth interface, or a 5G interface).

[0064] The communication bus is used to achieve connection and communication between different components. For example, the processor, memory, network interface, and user interface mentioned above can be interconnected through the communication bus.

[0065] The user interface may include a display screen and an input unit (such as a keyboard). Optionally, the user interface may also include a standard wired interface and a wireless interface.

[0066] Those skilled in the art will appreciate that the above-mentioned computer may also include more or fewer components, or a combination of certain components, or different arrangements of components, which is not limited in the embodiments of the present application.

[0067] For example, Figure 1 As shown, the regional landslide prediction method based on multimodal decomposition and physical mechanism provided by the embodiment of the present application includes S101-S104.

[0068] S101. Clustering the target area based on geological data, topographic data, and environmental factors of the target area to obtain multiple landslide areas within the target area.

[0069] In the embodiments of the present application, each of the multiple landslide areas described above belongs to only one geological unit. A geological unit is a stratum or rock mass region with consistent geotechnical properties (e.g., lithology, strength, and permeability) and geological history. Types of geological units include homogeneous sandstone, layered mudstone, fault fracture zone, and residual soil.

[0070] In one implementation, combining Figure 1 ,like Figure 2 As shown, the above S101 includes S1011-S1015.

[0071] S1011. Construct a feature matrix of the target area.

[0072] The target area includes multiple grid cells. The characteristic matrix of the target area includes the characteristic vector of each grid cell in the multiple grid cells. The characteristic vector of each grid cell includes the geological data, topographic data and environmental factors corresponding to the grid cell; the characteristic vector satisfies: F = [P geo ,T topo ,M eteo ], F represents the eigenvector, P geo Represents geological data, T topo Represents terrain data, M eteo Represents environmental factors.

[0073] The following is a detailed introduction to the process of constructing the feature matrix of the target area.

[0074] Step 1: Obtain multi-source data of the target area.

[0075] Wherein, the multi-source data include historical landslide deformation data, geological data, topographic data and environmental factors. The geological data include fault location, stratum distribution, rock and soil density, internal friction angle and cohesion, etc. The topographic data include slope, slope aspect and terrain curvature, etc. The environmental factors include precipitation, soil surface temperature, soil moisture, snow thickness and surface water flow, etc. Exemplarily, the historical landslide deformation data can be obtained by spacecraft satellites (such as ESA Sentinel-1 satellite, etc.) InSAR (Interferometric Synthetic Aperture Radar) landslide displacement data, the geological data and topographic data can be obtained through geological experiments and geological surveys, and the environmental factors can be extracted from ERA5 (the fifth generation global atmospheric reanalysis data set developed by the European Centre for Medium-Range Weather Forecasts (ECMWF)). The acquisition method of the geological data, topographic data and environmental factors of the target area is only a feasible acquisition method, and the embodiment of the present application does not limit the acquisition method of the geological data, topographic data and environmental factors of the target area.

[0076] Step 2: Preprocess the historical landslide deformation data, geological data, topographic data and environmental factors of the target area.

[0077] The above-mentioned preprocessing process includes data standardization, data interpolation and data synchronization. For the historical landslide deformation data, geological data, topographic data and environmental factors of the target area, data standardization refers to processing the data so that each type of data has the same dimension and scale. Data interpolation refers to organizing each type of data into time series data and interpolating the missing data in the time series data. Data refers to synchronizing different types of data with different time resolutions so that the time resolution of each type of data is the same. Since data standardization, data interpolation and data synchronization are commonly used technical means in this technical field, the specific process of the above-mentioned data standardization, data interpolation and data synchronization is not described in detail in the embodiment of the present application.

[0078] Step 3: Divide the target area into grids and construct the feature matrix of the target area.

[0079] The target area is divided into grids to obtain multiple grid cells within the target area. The historical landslide deformation data, geological data, topographic data, and environmental factors of the target area obtained after the preprocessing in step 3 are then mapped to the multiple grid cells within the target area to obtain the historical landslide deformation data, geological data, topographic data, and environmental factors corresponding to each grid cell. The characteristic matrix of the target area is then constructed using the geological data, topographic data, and environmental factors corresponding to each grid cell.

[0080] Optionally, the size of the grid unit may be 100m×100m or 200m×200m. The embodiment of the present application does not limit the size of the grid unit.

[0081] S1012. Use the feature matrix of the target area to construct a constrained similarity matrix.

[0082] Optionally, the above constraint similarity matrix satisfies: Among them, S represents the constraint similarity matrix; s 1,1 Indicates the similarity between the first grid unit and the first grid unit, s n,1 Indicates the similarity between the nth grid unit and the first grid unit, s 1,n Indicates the similarity between the first grid unit and the nth grid unit, s n,n It represents the similarity between the nth grid cell and the nth grid cell, and n represents the total number of grid cells in the target area.

[0083] The above similarity satisfies: Among them, s i,j represents the similarity between the i-th grid unit and the j-th grid unit, and 1≤i≤n and 1≤j≤n; F i represents the feature vector of the i-th grid cell, F j represents the eigenvector of the jth grid cell; σ represents the width of the kernel function; I is a judgment function used to determine whether two grid cells belong to the same geological unit.

[0084] Furthermore, if there is a known landslide linkage record between the i-th grid unit and the j-th grid unit (for example, a landslide in the i-th grid unit is linked to a landslide in the j-th grid unit), then the above s i,j ←s i,j ×m, where the value of m ranges from 1.1 to 1.5.

[0085] S1013. Construct a Laplace matrix by constraining the similarity matrix and the degree matrix.

[0086] The Laplace matrix satisfies: J = DS, where J represents the Laplace matrix, D represents the degree matrix, and the degree matrix is ​​a diagonal matrix and

[0087] S1014. Perform eigendecomposition on the Laplace matrix to obtain a clustering eigenvector matrix.

[0088] Since the above-mentioned S1014 is a commonly used technical means in this technical field, the embodiment of the present application does not elaborate on the specific process of S1014.

[0089] S1015. Perform K-means clustering on the clustering feature vector matrix to obtain multiple clusters, determine the area in the target area corresponding to each cluster in the multiple clusters as a landslide area, and obtain multiple landslide areas in the target area.

[0090] The embodiment of the present application adopts the elbow method and the silhouette coefficient dual-indicator verification: calculate the intra-cluster error (WSS) and the average silhouette coefficient when K = 2 to 10, select the inflection point K (for example, when K = 5, the silhouette coefficient reaches a peak value of 0.65, and the WSS decreases slowly, then K = 5 is determined), where is the number of multiple clusters obtained by clustering. By continuously iteratively updating the cluster center, the similarity of landslide bodies in the same landslide area is made as high as possible, and the similarity between different landslide areas is made as low as possible. Finally, the entire target area is divided into K landslide area types (such as rainfall-dominated type, freeze-thaw sensitive area, clay soil creep type, debris flow potential type, steep slope unloading type, gentle slope potential sliding type, rapid response type of sudden impact, seasonal periodic type, and composite type).

[0091] It should be noted that, since K-means clustering is a commonly used technical means in this technical field, the specific operation process of K-means clustering is not described in detail in the embodiment of the present application.

[0092] Furthermore, multiple landslide areas in the target area can be dynamically updated. For example, when the addition of multi-source data to the target area causes a change in the characteristic distribution of the target area (e.g., KS test p < 0.05) or an environmental mutation event (24-hour rainfall > 100 mm), incremental clustering is re-performed and the K value is updated. Specifically, the similarity of multiple grid cells contained in the affected landslide area is recalculated (e.g., only RD-class areas are updated in the case of heavy rain); then, the cluster center is locally adjusted based on the sliding window mean of the new data (window size = 30 days); finally, the divided landslide areas are split or merged. If the silhouette coefficient of a landslide area is < 0.3, the landslide area is split into two new landslide areas; if the centroid distance between the clusters corresponding to the two landslide areas is less than a threshold, the two landslide areas are merged into a new landslide area.

[0093] In one application scenario, combined with Figure 2 ,like Figure 3 As shown, after the above S101, the above method further includes S105.

[0094] S105 . For each of the multiple landslide areas, extract a trend item, a period item, and a burst item from historical landslide deformation data of the landslide area.

[0095] Optionally, the above S105 includes S1051-S1054.

[0096] S1051. After injecting noise into historical landslide deformation data of a landslide area using an adaptive noise injection method, a first original landslide signal is obtained.

[0097] It should be understood that for each of the multiple landslide regions, the historical landslide deformation data for that landslide region refers to the historical landslide data represented by the multiple grid cells contained in that landslide region, and the historical landslide deformation data for that landslide region is time series data. Based on this, the historical landslide deformation data D(t) (also referred to as the original signal) for that landslide region should satisfy the following formula.

[0098] D(t)=L(t)+C(t)+B(t)+∈

[0099] Where t represents time, L(t) represents the trend term (low-frequency signals such as tectonic movement and long-term creep), C(t) represents the periodic term (medium-frequency signals such as seasonal rainfall and freeze-thaw cycles), B(y) represents the burst term (high-frequency signals such as earthquakes and step displacements caused by heavy rain), and ∈ represents noise data. It can be understood that the trend term L(t) indicates long-term events within the landslide area, the periodic term C(t) indicates periodic events within the landslide area, and the burst term B(t) indicates sudden events within the landslide area. L(t), C(t), and B(t) are all time series data.

[0100] Specifically, after adding N groups of Gaussian white noise to the original signal D(t) using the adaptive noise injection method, N groups of first original landslide signals are obtained. The first original landslide signals satisfy the following formula.

[0101] D i (t) = D(t) + βn i (t),i=1,…,N

[0102] Among them, D i (t) represents the addition of the i-th group of Gaussian white noise n to the original signal D(t) i (t) is the first original landslide signal after t, and β represents the control noise intensity (β is usually taken as 0.2 times the signal standard deviation).

[0103] S1052: extract a trend term from the first original landslide signal using a modal decomposition method, and use the remaining signal in the first original landslide signal as a second original landslide signal.

[0104] The specific process of the above S1052 is described in detail below.

[0105] Step 1: After performing EMD decomposition (Empirical Mode Decomposition) on the N groups of first original landslide signals, extract the first mode of the N groups of first original landslide signals. The average modal IMF of the first mode of the N groups of first original landslide signals is j (t) satisfies the following formula.

[0106]

[0107] Where t represents time and j represents the number of times step 1 is executed.

[0108] Step 2: Calculate the average modal IMF of the first mode obtained in step 1 j When the square sum is greater than or equal to the modal threshold, repeat step 1 above and continue to perform modal decomposition on the remaining signal after decomposition until the average modal IMF of the first mode obtained by decomposition is j The sum of the squares of (t) is less than the modal threshold, then the trend term L(t) = IMF1(t) + IMF2(t) + ... + IMF M (t), where M represents the total number of times step 1 is repeated.

[0109] Optionally, the modal threshold can be set based on the signal noise level, or the trend item accuracy under different thresholds can be tested on historical landslide deformation data of the landslide area, and the optimal trend item accuracy is used as the modal threshold. The embodiment of the present application does not limit the method for setting the modal threshold.

[0110] S1053: extracting periodic terms from the second original landslide signal through continuous wavelet transform, and taking the remaining signal in the second original landslide signal as the third original landslide signal.

[0111] Exemplarily, the third original landslide signal R(t) satisfies: R(t)=D(t)-L(t). Then, a continuous wavelet transform (CWT) is performed on the third original landslide signal R(t) with a trend term. The transformation formula of the continuous wavelet transform (CWT) is as follows.

[0112]

[0113] Among them, a is the scale (frequency), b is the translation (time), W(a,b) is the wavelet transform coefficient, ψ a,b (t) is the Morlet wavelet basis function, t represents time, R′(t) is the first-order derivative of the trend line signal over time, is ψ a,b The conjugate function of (t).

[0114] Specifically, in the process of performing continuous wavelet transform on the third original landslide signal R(t), the frequency a and time b are adjusted to generate a plurality of wavelet transform coefficients W(a, b) corresponding to different a and b, respectively. The wavelet transform coefficients W(a, b) that change with the time period are screened out from the plurality of wavelet transform coefficients W(a, b) to form a periodic term C(t), and D(t)-L(t)-C(t) is used as the third original landslide signal.

[0115] S1054: Determine, in the third original landslide signal, a signal in which the rainfall index is greater than the rainfall threshold and / or the cumulative displacement is greater than the acceleration threshold as a sudden item.

[0116] It should be noted that the embodiment of the present application targets two typical scenarios that may trigger landslides (rainfall and earthquake), and calculates the critical threshold value through the mechanical formula of the physical model to determine whether the landslide area has entered the sudden deformation stage, and thereby extracts the sudden term from the third original landslide signal D(t)-L(t)-C(t).

[0117] In the embodiment of the present application, for a typical scenario of landslide caused by rainfall, the rainfall threshold is calculated using infinite slope stability, and the rainfall threshold satisfies the following formula.

[0118]

[0119] Where t represents time, I r (t) represents the rainfall threshold, c(t) represents the soil cohesion in the landslide area during the current time period, γ w represents the water density, h(t) represents the thickness of the saturated zone in the landslide area during the time period at the current moment, φ represents the internal friction angle of the landslide, γ represents the soil density, and θ(t) represents the internal friction angle of the landslide in the landslide area during the time period at the current moment. The rainfall index is calculated based on environmental factors.

[0120] For the historical landslide deformation data corresponding to a certain time point in the third original landslide signal, when the rainfall index of the landslide area at that time point is greater than the rainfall threshold I r (t), it is determined that the landslide deformation state at this time has entered a high-risk state, and the historical landslide deformation data corresponding to this time point is extracted from the third original landslide signal as a sudden item.

[0121] For typical rainfall-induced landslide scenarios, the Newmark cumulative displacement method is used to determine the acceleration threshold. The acceleration threshold satisfies the following formula.

[0122]

[0123] Among them, k c (t) represents the acceleration threshold, k c(t) represents the acceleration threshold corresponding to the time period at time point t, and g represents the acceleration due to gravity.

[0124] For the historical landslide deformation data corresponding to a certain time point in the third original landslide signal, calculate the landslide deformation acceleration a(t) at that time point. The landslide deformation acceleration a(t) at that time point satisfies: Δv(t) represents the difference between the landslide deformation rate at this time point and the landslide deformation rate at the previous time point, and Δt represents the time difference between this time point and the previous time point.

[0125] When the landslide deformation acceleration a(t)>k c Calculate the cumulative displacement D at that time point N The cumulative displacement D at this time point N Satisfied: D N =∫∫[a(t)-k c (t)]dt 2 .

[0126] Furthermore, when D N When >3σ1, the landslide deformation state is judged to be in a high-risk state, and the historical landslide deformation data corresponding to this time point is extracted from the third original landslide signal as the sudden item. Among them, σ1 represents the cumulative deformation standard deviation, P represents the number of time points contained in the time period of the time point (also known as the size of the sliding window before the current time step), d i (t) represents the landslide deformation at the i-th time point in the time period of this time point.

[0127] S102 . For each of the multiple landslide areas, based on the trend item, period item, and sudden item in the historical landslide deformation data of the landslide area, predict the trend landslide deformation variable, periodic landslide deformation variable, and sudden landslide deformation variable of the landslide area at the next moment.

[0128] The trend deformation of the landslide area at the next moment refers to the deformation that will occur in the landslide area at the next moment under the influence of long-term events. The periodic deformation of the landslide area at the next moment refers to the deformation that will occur in the landslide area at the next moment under the influence of periodic events. The sudden deformation of the landslide area at the next moment refers to the deformation that will occur in the landslide area at the next moment under the influence of sudden events.

[0129] In one implementation, the above S102 includes S1021-S1023.

[0130] S1021. Input the trend item into the trend item prediction model, and the trend item prediction model outputs the trend landslide deformation variable of the landslide area at the next moment.

[0131] The trend item prediction model mentioned above is an improved Transformer model. It should be understood that the Transformer model group is an existing deep learning framework whose core idea is to achieve efficient parallel computing and long-range dependency modeling through an adaptive attention mechanism (Self-Attention).

[0132] The input content X of the trend item prediction model provided in the embodiment of the present application l Satisfied: X l =[L(t),E,v,v creep ]∈R T×4 , where E represents the elastic modulus; v represents the Poisson's ratio; v creep represents the creep rate; R T×4 Represents X l It consists of four vectors of dimension T; T represents the time series.

[0133] The trend item prediction model of the embodiment of the present application improves the attention mechanism of the existing Transformer model by adding a time-related weight coefficient and a lower triangular mask matrix to the attention mechanism of the Transformer model. On this basis, the position-adaptive attention of the trend item prediction model is calculated based on the time-related weight coefficient and the lower triangular mask matrix. The calculation formula of the position-adaptive attention of the above trend item prediction model is as follows.

[0134]

[0135] Where Q1 = X l W Q , W Q represents the first weight matrix, X l Represents the input of the trend item prediction model; K1 = X l W K , W K Represents the second weight matrix; V1 = X l W V , W V represents the third weight matrix; Attention1(Q1,K1,V1) represents position adaptive attention; w pos (t) represents the time-dependent weight coefficient, and t pred represents the time point of the next moment; softmax(·) represents the normalization function; d k represents the data dimension, t represents time, w pos (t) represents the time-related weight coefficient, Mtime represents the lower triangular mask matrix, and

[0136] It can be understood that the trend item prediction model provided in the embodiment of the present application adopts a sparse matrix (ie, the above-mentioned lower triangular mask matrix M time ) replaces the full matrix in the existing Transformer model, adjusting the existing Transformer model's attention mechanism from full attention to sparse attention. This adjustment improves the Transformer model's input processing efficiency and focuses attention on input data corresponding to time points that have a significant impact on landslide deformation, making the prediction results of the trend prediction model more accurate.

[0137] In the training process of the trend item prediction model provided in the embodiment of the present application, a basic prediction data is given in each layer of training process. The residual e of this layer is formed by subtracting the true observation value L(t) of the current time part from it i Then it is input to the next layer of calculation.

[0138] And the predicted value obtained after multiple stacking training The correction is performed using the Burgers creep model. The correction process satisfies the following formula.

[0139]

[0140] Among them, L pred (t) represents the corrected predicted value, η represents the viscosity coefficient (given by v creep Inversion), τ represents the relaxation time.

[0141] Then use the weighted sum of RMSE (root mean square error) and MAPE (mean absolute percentage error) (L trend =αRMSE+(1-α)MAPE, α is the weight coefficient) as the loss function of the trend item prediction model, and the AdamW optimizer adjusts the parameters. Finally, the trained trend item prediction model is obtained, and the trained trend item prediction model is used to predict the trend landslide deformation D of the landslide area at the next moment. trend-pred .

[0142] It should be understood that since the training process of the Transformer model is a common technical means in this technical field, the embodiments of this application will not further elaborate on the specific operations of the above training process.

[0143] S1022: Input the periodic term into a periodic term prediction model, and the periodic term prediction model outputs the periodic landslide deformation variable of the landslide area at the next moment.

[0144] The above-mentioned periodic item prediction model is the manba model. The input content of the above-mentioned periodic item prediction model is X c Satisfied: X c =[C(t),R(t),T(t),h w (t),θ]∈R T×4 , where C(t) represents the periodic term, R(t) represents the rainfall, T(t) represents the soil surface temperature, and h w (t) represents the groundwater level, and θ represents the slope gradient.

[0145] During the training process of the above-mentioned periodic term prediction model, the following operations are performed on each of the K residual blocks included in the above-mentioned periodic term prediction model: First, X c Generate the prediction value of the current residual block through the fully connected layer Synchronously output environmental factor weight w R in It is about the characteristic function of environmental factors. It is h k (h k Represents the hidden state of the current residual block) takes the average along the time dimension to obtain the global feature, which is mapped to the environmental factor weight through the fully connected layer, M i is the i-th environmental factor. The error e of the residual block k satisfy The residual block converts e k Enter the next residual block for further decomposition.

[0146] Then the predicted value C of multiple residual blocks after training pred (t) The following rock mass elastic deformation range must be met.

[0147]

[0148] Among them, Clip(·) represents the slicing function, E represents the elastic modulus, K represents the number of residual blocks, α k =(1+w R ΔR), is the average value of environmental factors.

[0149] In the training process of the above-mentioned periodic term prediction model, the cross entropy loss function is combined with the L1 regularization term (L cycle =L cross +λΣ w |w|, λ is the regularization coefficient, w is the model parameter, L cycle is a periodic term, L crossSince the training process of the manba model belongs to the common technical means in this technical field, the specific operation of the above training process will not be further described in the embodiment of the present application.

[0150] The residual block in the periodic term prediction model outputs the environmental factor weight. Specifically, for each environmental factor (such as rainfall), the environmental factor weight w of the above K residual blocks for the environmental factor is R The average value of is used as the final weight to analyze the importance of the environmental factor to landslide deformation. The above final weight satisfies the following formula.

[0151]

[0152] Among them, M R,i represents the i-th environmental factor among multiple environmental factors, Represents the environmental factor weight of the i-th environmental factor output by the k-th residual block among the K residual blocks.

[0153] It should be noted that the above-mentioned periodic item prediction model provided in the embodiment of the present application introduces environmental factors. On the one hand, it can quantify climate anomalies as deformation adjustment coefficients to improve the long-term prediction accuracy of the periodic item prediction model. On the other hand, it can clarify the contribution of environmental factors in the periodic item prediction model to improve the interpretability of the periodic item prediction model.

[0154] S1023: Input the sudden item into the sudden item prediction model, and the sudden item prediction model outputs the sudden landslide deformation variable of the landslide area at the next moment.

[0155] In the embodiment of the present application, the above-mentioned burst item prediction model is an improved Temporal Fusion Transformer model (hereinafter referred to as TFT model). The input content of the above-mentioned burst item prediction model is X B Satisfied: X B =[B(t),R(t),a(t),E(t),c,φ,θ], where B(t) represents the burst term, E(t) represents the event marker, R(t) represents the rainfall amount, E(t)∈{0,1}, and E(t)=1 means R(t)≥i r (t) or / and a(t)≥k c (t), I r (t) represents the rainfall threshold, a(t) represents the deformation acceleration, k c (t) represents the acceleration threshold; E(t) = 0 represents R(t) r (t) and a(t) <k c (t).

[0156] ​The physical constraint attention of the above-mentioned burst prediction model is calculated based on the rainfall threshold and acceleration threshold. The calculation formula of the physical constraint attention of the burst prediction model is as follows:

[0157]

[0158] Among them, Q2=Linear([B(t),E(t)]), K2=Linear([R(t),a(t),tanφ,sinθ]), V2=Linear([B(t),min(B(t),0.8D N )]), Linear(·) represents the linear transformation function, D N represents the cumulative displacement, which is calculated by the Newmark model, and Attention2(Q2,k2,V2) represents the physical constraint attention.

[0159] In the calculation formula of the physical constraint attention of the above-mentioned sudden item prediction model, Q2 is obtained by linear mapping after concatenating the historical landslide deformation data and event markers corresponding to the sudden item, which is used to capture when the sudden event is triggered; K2 is obtained by linear mapping after concatenating the environmental factor data and geological parameters, which is used to capture the conditions that trigger the sudden event.

[0160] It should be noted that the physical constraint attention Attention2 (Q2, k2, V2) of the above-mentioned sudden item prediction model has added a judgment threshold (i.e., I r (t) and k c (t)), when the input data corresponding to a certain time point is greater than the judgment threshold (ie R(t)≥I r (t) or / and a(t)≥k c (t)), the weight of the physical constraint attention Attention2(Q2, K2, V2) increases significantly, and the burst item prediction model performs incremental processing on the input data corresponding to this time point.

[0161] In the training process of the above-mentioned burst prediction model, the dynamic features of the deformation data are fused with the attention features, and the hidden state output H of each layer of the burst prediction model is obtained after processing. out The hidden state output H of each layer of the above burst prediction model is out The following formula is satisfied.

[0162] H out =LayerNorm(H dyn +Dropout(Attention2(Q2,K2,V2)))

[0163]

[0164] Among them, LayerNorm(·) represents the layer normalization function, and Dropout(·) represents the regularization function, which is used to randomly discard (set to zero) the output of neurons during training.

[0165] After obtaining the hidden state output H of each layer of the burst prediction model out Finally, dilated convolution is used to capture the multi-scale temporal pattern H tcn Multi-scale temporal pattern H tcn The following formula is satisfied.

[0166] H tcn =TCN(H out )

[0167] Where TCN(·) represents the dilated convolution function.

[0168] During the training process of the above-mentioned burst item prediction model, if the prediction value B of the burst item prediction model pred (t) exceeds the upper limit of the theoretical displacement, the cumulative displacement D calculated by Newmark N Perform alignment correction. The above process satisfies the following formula.

[0169] B pred (t)←min(B pred (t),D N )

[0170] D N =∫∫[a(y)-k c (t)]dt 2

[0171] On this basis, the loss function of the above-mentioned sudden item prediction model is the physical constraint loss function Physical Constraint Loss Function satisfy: D phys Indicates the upper limit of displacement calculated by the physical model.

[0172] It should be understood that since the training process of the TFT model is a common technical means in this technical field, the embodiment of the present application will not further elaborate on the specific operations of the above training process.

[0173] The model formula of the trained burst item prediction model is as follows.

[0174]

[0175] Wherein, σ2 is a Sigmoid function, which is used to ensure that the output of the burst item prediction model is non-zero when the input data corresponding to a certain time point is greater than the above-mentioned determination threshold.

[0176] S103 , taking the weighted sum of the trend landslide deformation variable, the periodic landslide deformation variable, and the sudden landslide deformation variable of the landslide area at the next moment as the predicted value of the landslide deformation in the landslide area at the next moment.

[0177] In the embodiment of the present application, the predicted value D of the landslide deformation in the landslide area at the next moment is pred The following formula is satisfied.

[0178] D pred =ω trend D trend-pred +ω cycle D cycle-pred +ω burst D burst-pred

[0179] Among them, D trend-pred Indicates the trend of landslide deformation in the landslide area at the next moment, D cycle-pred represents the periodic landslide deformation of the landslide area at the next moment, D burst-pred Indicates the sudden landslide deformation of the landslide area at the next moment; ω trend The weighted coefficient of the trend landslide deformation variable in the landslide area at the next moment, ω cycle The weighted coefficient of the periodic landslide deformation variable of the landslide area at the next moment, ω burst The weighted coefficient representing the sudden landslide deformation in the landslide area at the next moment.

[0180] It is understandable that due to the trend of landslide deformation D trend-pred Indicates the long-term events in the landslide area, the periodic landslide deformation D cycle-pred Indicates the periodic events in the landslide area, the sudden landslide deformation D burst-pred Indicates the sudden event in the landslide area. Then the predicted value of the landslide deformation in the landslide area at the next moment is D pred Landslide deformation under the combined effects of long-term events, periodic events and sudden events.

[0181] Furthermore, the weighting coefficients of the trend landslide deformation variable, the periodic landslide deformation variable and the sudden landslide deformation variable are determined by the type of the landslide area.

[0182] Optionally, the type of landslide area can be rainfall-dominated, freeze-thaw-sensitive, clay creep, debris flow potential, steep slope unloading, gentle slope potential sliding, rapid response to sudden impact, seasonal periodic or composite type (a total of nine types), or other types. The embodiments of the present application do not specifically limit the type of landslide area.

[0183] First, the initial weighting coefficients for the trend, periodic, and sudden landslide deformation variables are determined directly based on the landslide region type. When the landslide region is any of the nine types listed above, the weighting coefficients for the trend, periodic, and sudden landslide deformation variables are as follows.

[0184] In the first case, when the landslide area is rainfall-dominated, the weighting coefficient of the trend landslide deformation variable is 0.3, the weighting coefficient of the periodic landslide deformation variable is 0.6, and the weighting coefficient of the sudden landslide deformation variable is 0.1.

[0185] In the second case, when the landslide area is freeze-thaw sensitive, the weighting coefficient of the trend landslide deformation variable is 0.2, the weighting coefficient of the periodic landslide deformation variable is 0.7, and the weighting coefficient of the sudden landslide deformation variable is 0.1.

[0186] In the third case, when the landslide area is of clay creep type, the weighted coefficient of the trend landslide deformation variable is 0.7, the weighted coefficient of the periodic landslide deformation variable is 0.2, and the weighted coefficient of the sudden landslide deformation variable is 0.1.

[0187] In the fourth case, when the landslide area is of the potential type of debris flow, the weighting coefficient of the trend landslide deformation variable is 0.2, the weighting coefficient of the periodic landslide deformation variable is 0.2, and the weighting coefficient of the sudden landslide deformation variable is 0.6.

[0188] In the fifth case, when the landslide area is of steep slope unloading type, the weighted coefficient of the above trend landslide deformation variable is w trend =0.7, the weighting coefficient of the periodic landslide deformation variable is 0.1, and the weighting coefficient of the sudden landslide deformation variable is 0.2.

[0189] In the sixth case, when the landslide area is of the gentle slope potential landslide type, the weighting coefficient of the trend landslide deformation variable is 0.4, the weighting coefficient of the periodic landslide deformation variable is 0.5, and the weighting coefficient of the sudden landslide deformation variable is 0.1.

[0190] In the seventh case, when the landslide area is of the rapid response type with sudden impact, the weighting coefficient of the trend landslide deformation variable is 0.05, the weighting coefficient of the periodic landslide deformation variable is 0.1, and the weighting coefficient of the sudden landslide deformation variable is 0.85.

[0191] In the eighth case, when the landslide area is of seasonal periodic type, the weighting coefficient of the trend landslide deformation variable is 0.1, the weighting coefficient of the periodic landslide deformation variable is 0.8, and the weighting coefficient of the sudden landslide deformation variable is 0.1.

[0192] In the ninth case, when the landslide area is of a composite type, the weighting coefficient of the trend landslide deformation variable is 0.4, the weighting coefficient of the periodic landslide deformation variable is 0.3, and the weighting coefficient of the sudden landslide deformation variable is 0.3.

[0193] The three weighting coefficients in the above nine cases are just initial weights, which can be expressed as The three weighted coefficients of the landslide area at the next moment Based on environmental factors (such as rainfall intensity, soil temperature), the historical prediction error of the prediction model history Dynamic adjustment. The weight of the landslide deformation variable of the above landslide area at the next moment period The adjustment is calculated as follows.

[0194]

[0195] Among them, R(t), T(t) are the rainfall intensity and soil temperature at different time points, R ref is the maximum rainfall in history in the landslide area, T freeze is the freeze-thaw temperature threshold, T range Indicates the temperature range, and T range =T(t) max -T(t) min , T(t) max is the maximum temperature, T(t) min is the minimum temperature; e threshold is to set the error threshold, The weight coefficient of the initial periodic landslide deformation variable in the region is set. The weight coefficient of the sudden landslide deformation variable of all the above regions at the next moment is set. The calculation formula for dynamic changes is as follows

[0196]

[0197] Where a(t) is the landslide deformation acceleration, I r represents the historical rainfall threshold within the historical sliding time window, k c Represents the historical acceleration threshold within the historical sliding time window, The weight coefficient of the initial sudden landslide deformation variable is set for the landslide area. The weight coefficient of the trend landslide deformation variable of the landslide area at the next moment is set. satisfy:

[0198]

[0199] The other two terms are also normalized to obtain the final weight coefficient.

[0200]

[0201] S104. When the landslide area becomes unstable, adjacent landslide areas with potential cascading instability among the multiple adjacent landslide areas are determined based on the stress increment and pore water pressure coupling between the landslide area and each of the multiple adjacent landslide areas of the landslide area; and the predicted value of the landslide deformation of the landslide area at the next moment and the adjacent landslide areas with potential cascading instability are used as landslide prediction information of the landslide area.

[0202] The grounds for determining whether a landslide region is unstable are that the deformation acceleration of the landslide region during the current time period is greater than an instability threshold. The aforementioned adjacent landslide region refers to a landslide region that is directly adjacent to the landslide region within the multiple landslide regions. The present embodiments do not limit the value of the aforementioned instability threshold.

[0203] Illustratively, the above S104 includes S1041-S1042.

[0204] S1041. For each of the multiple adjacent landslide areas in the landslide area, calculate the safety factor of the adjacent landslide area by coupling the stress increment and pore water pressure between the landslide area and the adjacent landslide area.

[0205] Optionally, the above safety factor satisfies the following formula.

[0206]

[0207] Where c represents the soil cohesion, σ n represents the normal stress generated by the landslide area on the adjacent landslide area, Δσ B It represents the stress increment generated by the landslide area on the adjacent landslide area at the current moment, p new represents the pore water pressure between the landslide area and the adjacent landslide area at the current moment, and p new =p+γ w Δh w , p represents the pore water pressure Δh between the landslide area and the adjacent landslide area before the current moment w represents the groundwater level rise in the adjacent landslide area, γ w represents water density; φ represents the internal friction angle of the landslide, γ represents the soil density, h represents the thickness of the saturated zone adjacent to the landslide area, θ represents the slope of the landslide, τ transfer It represents the shear stress transferred from the landslide area to the adjacent landslide area.

[0208] S1042. Determine, among multiple adjacent landslide areas, adjacent landslide areas with safety factors greater than a safety threshold as adjacent landslide areas with potential cascading instability.

[0209] Optionally, combined Figure 3 ,like Figure 4 As shown, after the above S104, the above method further includes S106.

[0210] S106. Otherwise, the predicted value of the landslide deformation in the landslide area at the next moment is used as the landslide prediction information of the landslide area.

[0211] In one implementation, combining Figure 4 ,like Figure 5 As shown, before S103, the above method further includes S107.

[0212] S107 , based on the synergistic relationship among the multiple landslide areas, adjusting the trend landslide deformation variable, the periodic landslide deformation variable, and the sudden landslide deformation variable of the landslide area with potential chain landslide in the multiple landslide areas at the next moment.

[0213] It is understandable that, since the embodiment of the present application independently predicts the trend landslide deformation variable, periodic landslide deformation variable, and sudden landslide deformation variable of each landslide area in the target area at the next moment, without considering the physical connection between multiple landslide areas, it results in a situation where the landslide displacement of one landslide area is predicted to change suddenly, but the landslide displacement of the landslide area adjacent to the landslide area is predicted to be very small (such as landslide area A + 5mm, landslide area B + 0.1mm). This situation violates the mechanical interaction law between adjacent landslide areas and cannot identify hidden chain risks (such as the instability of area C will affect area D through faults). Therefore, to address these problems, a graph structure is constructed for all sub-areas to correct the prediction results of adjacent landslide areas. This design ensures that the prediction results conform to both data laws and mechanical principles through mandatory collaborative correction and physical constraint iteration.

[0214] The specific process of the above S107 is described in detail below.

[0215] Step 1: Based on the synergistic relationship between multiple landslide areas, each of the multiple landslide areas in the target area is taken as a node to construct a spatial map of the target area.

[0216] The spatial graph of the target area satisfies: G = (V, E), where G represents the spatial graph, V is the node set, and E is the edge set. The spatial graph G is determined based on the synergistic relationship between multiple landslide areas. The weight w of the edge connecting the nodes in the spatial graph G is ij It is the spatial distance d between the two landslide areas corresponding to the two nodes connected by the edge ij and geological correlation r ij Sure.

[0217] Specifically, for an edge connecting two nodes (corresponding to two landslide areas), the spatial distance d ij is the distance between the centroids of the two landslide areas and edge distance The weighted sum of the spatial distance d ij The calculation formula is as follows.

[0218]

[0219] Where i is one of the two landslide areas corresponding to the two nodes connected by an edge, and j is the other landslide area of ​​the two landslide areas corresponding to the two nodes connected by an edge; x i is the coordinate component of the geographical centroid of a landslide area on the x-axis of the UTM projection coordinate system; j is the coordinate component of the geographical centroid of another landslide area on the x-axis of the UTM projection coordinate system; i is the coordinate component of the geographical centroid of a landslide area on the y-axis of the UTM projection coordinate system; j is the coordinate component of the geographic centroid of another landslide area on the y-axis of the UTM projection coordinate system. Since the UTM projection coordinate system is a commonly used technical means in this technical field, the embodiment of the present application will not further describe the UTM projection coordinate system.

[0220] Optionally, the above edge distance The boundary between two regions is directly obtained by calculating it through GIS tools (such as across valley terrain). The spatial distance is dominated by the centroid distance, but the boundary distance weight is strengthened for adjacent regions (such as shared boundaries).

[0221] If there is a ridge or valley between two landslide areas, the terrain correction coefficient is introduced to adjust the spatial distance d. ij Correction is performed to obtain the final spatial distance The following formula is satisfied.

[0222]

[0223] where Δh is the height difference between the two landslide areas.

[0224] The above geological correlation r ij According to the geological structure P of the two landslide areas corresponding to the two nodes ego,i , Geotechnical Type P ego,j The geological correlation r ij Satisfies the following formula.

[0225] r ij =Sim(P ego,i ,Pego,j )

[0226] Where Sim(·) is the cosine similarity function.

[0227] Then the weight w of the above edge is ij The following formula is satisfied.

[0228]

[0229] Where E is the elastic modulus, h is the thickness of the sliding body, v is the Poisson's ratio, σ d is the standard deviation of the distance between the two landslide areas.

[0230] Step 2: Construct a spatiotemporal dynamic map of the target area.

[0231] The spatiotemporal dynamic graph of the target area mentioned above satisfies: G t =(V,E t ), where t is time, G t Represents the spatiotemporal dynamic map of the target area, E t is the set of edges that changes over time.

[0232] Step 3: Spatiotemporal dynamic map G of the target area t During the change process, when the displacement speed of a landslide area exceeds the speed threshold (for example, 6 mm / s 2 ), the trend landslide deformation amount of the landslide area adjacent to the landslide area at the next moment is adjusted. After adjustment, the trend landslide deformation amount of the landslide area adjacent to the landslide area at the next moment is as shown in the following formula.

[0233]

[0234] Where i is one of the two landslide areas corresponding to the two nodes connected by an edge, and j is the other landslide area of ​​the two landslide areas corresponding to the two nodes connected by an edge; in the above case, j refers to the landslide area, and i refers to the landslide area adjacent to the landslide area; D trend-pred,i D is the trend landslide deformation of the landslide area adjacent to the landslide area at the next moment after adjustment. trend-pred,i ′ is the trend landslide deformation of the landslide area adjacent to the landslide area at the next moment before adjustment, λ L is the rock mass force transmission efficiency coefficient, and λ L ∈[0,1]; is the landslide area adjacent to the landslide area, D burst-pred,i is the sudden landslide deformation of the landslide area adjacent to the landslide area at the next moment after adjustment, D burst-pred,i′ is the sudden landslide deformation of the landslide area adjacent to the landslide area at the next moment before adjustment, σ2(·) is the Sigmoid function; E j Indicates the event status, I(E j =1) indicates that the event status is active; D cycle-pred,i is the periodic landslide deformation of the landslide area adjacent to the landslide area at the next moment after adjustment, D cycle-pred,i ′ is the periodic landslide deformation of the landslide area adjacent to the landslide area at the next moment before adjustment, γ is the environmental impact coefficient, M j is the environmental factor value of the landslide area (including rainfall, soil temperature, snow thickness, soil moisture, surface water runoff), M i It is the environmental factor value of the landslide area adjacent to the landslide area (it may also include rainfall, soil temperature, snow thickness, soil moisture, and surface water runoff).

[0235] Step 5: Update the weights of multiple edges in the spatiotemporal dynamic graph of the target area.

[0236] The weights of multiple edges in the updated spatiotemporal dynamic graph of the target area satisfy the following formula.

[0237]

[0238] in, To update the weight behind, To update the corresponding spatial distance behind, To update the corresponding geological correlation.

[0239] In summary, in the regional landslide prediction method based on multimodal decomposition and physical mechanism provided by the embodiment of the present application, the target area is first divided into multiple landslide areas based on the geological data, topographic data and environmental factors of the target area, and each landslide area obtained after the division belongs to only one geological unit. For each landslide area, the trend item, periodic item and sudden item in the historical landslide deformation data of the landslide area are combined to obtain the predicted value of the landslide deformation in the landslide area at the next moment, so as to realize the prediction of the landslide in the combined effect of long-term events, periodic events and sudden events. And in the case of instability of the landslide area, the predicted value of the landslide deformation of the landslide area at the next moment and the adjacent landslide area with potential chain instability are used as the landslide prediction information of the landslide area. In the process of dividing the target area into multiple landslide areas, the above method ensures that each landslide area has uniform geological conditions, making the stress increment and pore water pressure coupling between the landslide area and multiple adjacent landslide areas more reliable, and making the subsequent judgment of potential cascading instability more accurate, thereby realizing the prediction of sudden landslide phenomena and cascading instability landslide phenomena, thereby improving the prediction effect of landslide prediction.

[0240] Accordingly, the embodiment of the present application provides a regional landslide prediction device based on multimodal decomposition and physical mechanism, such as Figure 5 As shown, it includes a clustering module 501 , a first prediction module 502 , a second prediction module 503 and a third prediction module 504 .

[0241] Clustering module 501 is used to cluster the target region based on its geological data, topographic data, and environmental factors, thereby obtaining multiple landslide regions within the target region. Each of the multiple landslide regions belongs to a single geological unit, which is a stratum or rock mass region with consistent geotechnical properties and geological history. For example, clustering module 501 is used to implement S101 of the aforementioned regional landslide prediction method based on multimodal decomposition and physical mechanisms.

[0242] The first prediction module 502 is configured to, for each of the multiple landslide regions, predict the trend landslide deformation variable, the periodic landslide deformation variable, and the sudden landslide deformation variable of the landslide region at the next moment based on the trend item, the periodic item, and the sudden item in the historical landslide deformation data of the landslide region. The trend item indicates a long-term event within the landslide region; the periodic item indicates a periodic event within the landslide region; and the sudden item indicates a sudden event within the landslide region. For example, the first prediction module 502 is configured to implement S102 of the aforementioned regional landslide prediction method based on multimodal decomposition and physical mechanisms.

[0243] The second prediction module 503 is configured to use the weighted sum of the trend, periodic, and sudden landslide deformation variables of the landslide region at the next moment as the predicted landslide deformation value within the landslide region at the next moment. The weighting coefficients of the trend, periodic, and sudden landslide deformation variables are determined by the type of landslide region. For example, the second prediction module 503 is configured to implement S103 of the aforementioned regional landslide prediction method based on multimodal decomposition and physical mechanisms.

[0244] The third prediction module 504 is configured to, when a landslide region becomes unstable, determine adjacent landslide regions with potential cascading instability within the multiple adjacent landslide regions based on the stress increment and pore water pressure coupling between the landslide region and each of the multiple adjacent landslide regions within the landslide region. The predicted landslide deformation of the landslide region at the next moment and the adjacent landslide regions with potential cascading instability are used as landslide prediction information for the landslide region. The instability of the landslide region is determined based on the landslide deformation acceleration of the landslide region within the time period at the current moment being greater than an instability threshold. For example, the third prediction module 504 is configured to implement the aforementioned step S104 of regional landslide prediction based on multimodal decomposition and physical mechanisms.

[0245] Optionally, clustering module 501 is specifically configured to: construct a characteristic matrix for the target region. Using the characteristic matrix for the target region, a constrained similarity matrix is ​​constructed. A Laplace matrix is ​​constructed using the constrained similarity matrix and the degree matrix. Eigendecomposition is performed on the Laplace matrix to obtain a clustering eigenvector matrix. K-means clustering is performed on the clustering eigenvector matrix to obtain multiple clusters. Regions in the target region corresponding to each of the multiple clusters are identified as landslide regions, thereby obtaining multiple landslide regions within the target region. For example, clustering module 501 is specifically configured to implement S1011-S1015 of the aforementioned regional landslide prediction method based on multimodal decomposition and physical mechanisms.

[0246] Exemplarily, the apparatus further includes an extraction module 505. The extraction module 505 is configured to extract, for each of the multiple landslide regions, a trend term, a period term, and a burst term from the historical landslide deformation data for the landslide region. For example, the extraction module 505 is configured to implement the aforementioned step S105 of regional landslide prediction based on multimodal decomposition and physical mechanisms.

[0247] In one application scenario, the extraction module 505 is specifically used to: inject noise into the historical landslide deformation data of the landslide area using an adaptive noise injection method to obtain a first original landslide signal. A modal decomposition method is used to extract a trend term from the first original landslide signal, and the remaining signal in the first original landslide signal is used as a second original landslide signal. A continuous wavelet transform is used to extract a periodic term from the second original landslide signal, and the remaining signal in the second original landslide signal is used as a third original landslide signal. A signal in the third original landslide signal having a rainfall index greater than a rainfall threshold and / or a cumulative displacement greater than an acceleration threshold is determined as a burst term. For example, the extraction module 505 is specifically used to implement S1051-S1054 of the above-mentioned regional landslide prediction method based on multimodal decomposition and physical mechanisms.

[0248] The various modules of the above-mentioned regional landslide prediction device based on multimodal decomposition and physical mechanism can also be used to execute other steps in the above-mentioned method embodiment. All relevant contents involved in the above-mentioned method embodiment can be referred to the functional description of the corresponding functional module and will not be repeated here.

[0249] The present application also provides an electronic device, comprising: a processor and a memory coupled to the processor; the memory is configured to store computer instructions, and when the electronic device is running, the processor executes the computer instructions stored in the memory, so that the electronic device performs the method of the above embodiment. The processor can implement the above clustering module 501, the first prediction module 502, the second prediction module 503, the third prediction module 504, and the extraction module 505; the memory can also be configured to store geological data, topographic data, and environmental factors of the target area, trend items, period items, and sudden items in historical landslide deformation data of the landslide area, predicted values ​​of landslide deformation in the landslide area at the next moment, and landslide prediction information for the landslide area.

[0250] An embodiment of the present application further provides a computer-readable storage medium, which includes a computer program. When the computer program runs on a computer, the method described in the above embodiment is executed.

[0251] An embodiment of the present application further provides a computer program product, which includes computer program instructions. When the computer program instructions are run on a computer, the method described in the above embodiment is executed.

[0252] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments.

[0253] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A regional landslide prediction method based on multimodal decomposition and physical mechanisms, characterized by: include: Clustering the target area based on geological data, topographic data, and environmental factors of the target area to obtain multiple landslide areas within the target area; Each of the plurality of landslide areas belongs to only one geological unit; the geological unit is a stratum or rock mass area with consistent geotechnical properties and geological history; For each of the plurality of landslide areas, based on a trend item, a period item, and a sudden item in historical landslide deformation data of the landslide area, a trend landslide deformation amount, a periodic landslide deformation amount, and a sudden landslide deformation amount of the landslide area at a next moment are predicted; wherein the trend item indicates a long-term event in the landslide area; the period item indicates a periodic event in the landslide area; and the sudden item indicates a sudden event in the landslide area; taking a weighted sum of the trend landslide deformation variable, the periodic landslide deformation variable and the sudden landslide deformation variable of the landslide area at the next moment as a predicted value of the landslide deformation in the landslide area at the next moment; the weighting coefficients of the trend landslide deformation variable, the periodic landslide deformation variable and the sudden landslide deformation variable are determined by the type of the landslide area; When the landslide area becomes unstable, adjacent landslide areas with potential chain instability among the multiple adjacent landslide areas are determined based on the stress increment and pore water pressure coupling between the landslide area and each of the multiple adjacent landslide areas of the landslide area; and the predicted value of the landslide deformation of the landslide area at the next moment and the adjacent landslide areas with potential chain instability are used as landslide prediction information of the landslide area; wherein, the basis for judging the instability of the landslide area is that the landslide deformation acceleration of the landslide area in the time period at the current moment is greater than the instability threshold.

2. The method according to claim 1, wherein The target area is clustered to obtain a plurality of landslide areas within the target area, including: Constructing a characteristic matrix of the target area; wherein the target area includes a plurality of grid cells; the characteristic matrix of the target area includes a characteristic vector of each grid cell in the plurality of grid cells; the characteristic vector of each grid cell includes geological data, topographic data and environmental factors corresponding to the grid cell; the characteristic vector satisfies: F = [P geo ,T topo ,M eteo ], F represents the eigenvector, P geo Represents geological data, T topo Represents terrain data, M eteo represents environmental factors; The characteristic matrix of the target region is used to construct a constraint similarity matrix; the constraint similarity matrix satisfies: Among them, S represents the constraint similarity matrix; s 1,1 Indicates the similarity between the first grid unit and the first grid unit, s n,1 Indicates the similarity between the nth grid unit and the first grid unit, s 1,n Indicates the similarity between the first grid unit and the nth grid unit, s n,n represents the similarity between the nth grid cell and the nth grid cell, where n represents the total number of grid cells in the target area; the similarity satisfies: Among them, s i,j represents the similarity between the i-th grid unit and the j-th grid unit, and 1≤i≤n and 1≤j≤n; F i represents the feature vector of the i-th grid cell, F j represents the eigenvector of the jth grid cell; σ represents the width of the kernel function; A Laplace matrix is ​​constructed by using the constrained similarity matrix and the degree matrix; the Laplace matrix satisfies: J=DS, wherein J represents the Laplace matrix, D represents the degree matrix, and the degree matrix is ​​a diagonal matrix and Performing eigendecomposition on the Laplace matrix to obtain a clustering eigenvector matrix; K-means clustering is performed on the clustering feature vector matrix to obtain multiple clusters, and the area in the target area corresponding to each cluster in the multiple clusters is determined as a landslide area to obtain multiple landslide areas in the target area.

3. The method according to claim 1, wherein The determining of adjacent landslide areas with potential cascading instability among the plurality of adjacent landslide areas comprises: For each of the plurality of adjacent landslide areas of the landslide area, a safety factor of the adjacent landslide area is calculated by coupling the stress increment and pore water pressure between the landslide area and the adjacent landslide area; the safety factor satisfies the following formula; Where c represents the soil cohesion, σ n represents the normal stress generated by the landslide area on the adjacent landslide area, Δσ B represents the stress increment generated by the landslide area on the adjacent landslide area at the current moment, p new represents the pore water pressure between the landslide area and the adjacent landslide area at the current moment, and p new =p+γ w Δh w , p represents the pore water pressure Δh between the landslide area and the adjacent landslide area before the current moment w represents the groundwater level rise in the adjacent landslide area, γ w represents water density; φ represents the internal friction angle of the landslide, γ represents the soil density, h represents the thickness of the saturated zone adjacent to the landslide area, θ represents the slope of the landslide, τ transfer represents the shear stress transmitted from the landslide area to the adjacent landslide area; The adjacent landslide areas with the safety factors greater than the safety threshold among the multiple adjacent landslide areas are determined as the adjacent landslide areas with potential cascading instability.

4. The method according to claim 1, wherein The method further comprises: For each landslide area among the plurality of landslide areas, a trend item, a period item and a sudden item are extracted from the historical landslide deformation data of the landslide area.

5. The method according to claim 4, wherein The extracting of trend items, period items and sudden items from the historical landslide deformation data of the landslide area includes: After injecting noise into the historical landslide deformation data of the landslide area using an adaptive noise injection method, a first original landslide signal is obtained; extracting the trend term from the first original landslide signal using a modal decomposition method, and using the remaining signal in the first original landslide signal as a second original landslide signal; extracting the periodic term from the second original landslide signal by continuous wavelet transform, and using the remaining signal in the second original landslide signal as the third original landslide signal; Determining a signal in the third original landslide signal whose rainfall index is greater than a rainfall threshold and / or whose cumulative displacement is greater than an acceleration threshold as the sudden item; The rainfall threshold satisfies the following formula: Where t represents time, I r (t) represents the rainfall threshold, c(t) represents the soil cohesion of the landslide area in the time period at the current moment, γ w represents the water density, h(t) represents the thickness of the saturated zone in the landslide area during the time period at the current moment, φ represents the internal friction angle of the landslide, γ represents the soil density, and θ(t) represents the internal friction angle of the landslide in the landslide area during the time period at the current moment; the rainfall index is calculated based on environmental factors; The acceleration threshold satisfies the following formula: Among them, k c (t) represents the acceleration threshold, and g represents the acceleration due to gravity.

6. The method according to claim 1, wherein The predicting of the trend landslide deformation variable, the periodic landslide deformation variable and the sudden landslide deformation variable of the landslide area at the next moment includes: The trend item is input into a trend item prediction model, and the trend item prediction model outputs the trend landslide deformation amount of the landslide area at the next moment; the trend item prediction model is a Transformer model, and the position adaptive attention of the trend item prediction model is calculated based on the time-related weight coefficient; The periodic term is input into a periodic term prediction model, and the periodic term prediction model outputs the periodic landslide deformation of the landslide area at the next moment; the periodic term prediction model is a manba model; the residual block in the periodic term prediction model outputs the environmental factor weight; The sudden item is input into a sudden item prediction model, and the sudden item prediction model outputs the sudden landslide deformation amount of the landslide area at the next moment; the sudden item prediction model is a Temporal Fusion Transformer model, and the physical constraint attention of the sudden item prediction model is calculated based on a rainfall threshold and an acceleration threshold.

7. The method according to claim 6, wherein The calculation formula of the position adaptive attention of the trend item prediction model is as follows: Where Q1 = X l W Q , W Q represents the first weight matrix, X l represents the input of the trend item prediction model, and X l =[L(t),E,v,v creep ]∈R T×4 , t represents time, L(t) represents the trend term, E represents elastic modulus, v represents Poisson's ratio, v creep represents the creep rate, R T×4 Represents X l It consists of four vectors of dimension T, where T represents the time series; K1 = X l WK ， W K Represents the second weight matrix; V1 = X l W V , W V represents the third weight matrix; Attention1(Q1, K1, V1) represents the position adaptive attention; w pos (t) represents the time-dependent weight coefficient, and t pred represents the time point of the next moment; softmax(·) represents the normalization function; d k represents the data dimension, t represents time, w pos (t) represents the time-dependent weight coefficient, M time represents the lower triangular mask matrix, and The calculation formula of the physical constraint attention of the burst item prediction model is as follows: Wherein, Q2=Linear([B(t),E(t)]), Linear(·) represents a linear transformation function, B(t) represents the burst term, E(t) represents an event marker, E(t)∈{0,1}, and E(t)=1 represents R(t)≥I r (t) or / and a(t)≥k c (t), R(t) represents rainfall, I r (t) represents the rainfall threshold, a(t) represents the deformation acceleration, k c (t) represents the acceleration threshold; E(t) = 0 represents R(t) r (t) and a(t) <k c (t); Attention2(Q,K,V) represents the physical constraint attention.​ 8. The method according to claim 1, wherein The method further comprises: Based on the cooperative relationship between the multiple landslide areas, the trend landslide deformation variable, the periodic landslide deformation variable and the sudden landslide deformation variable of the landslide area with potential chain landslide in the multiple landslide areas at the next moment are adjusted.

9. The method according to claim 1, wherein The types of landslide areas are rainfall-dominated, freeze-thaw-sensitive, clay soil creep, debris flow potential, steep slope unloading, gentle slope potential sliding, rapid response to sudden impact, seasonal periodic or composite.

10. A regional landslide prediction device based on multimodal decomposition and physical mechanism, characterized by: It includes a clustering module, a first prediction module, a second prediction module and a third prediction module; The clustering module is used to cluster the target area based on the geological data, topographic data and environmental factors of the target area to obtain a plurality of landslide areas within the target area; each of the plurality of landslide areas belongs to only one geological unit; the geological unit is a stratum or rock mass area with consistent geotechnical properties and geological history; The first prediction module is configured to, for each of the plurality of landslide areas, predict a trend landslide deformation amount, a periodic landslide deformation amount, and a sudden landslide deformation amount of the landslide area at a next moment based on a trend item, a periodic item, and a sudden item in historical landslide deformation data of the landslide area; wherein the trend item indicates a long-term event in the landslide area; the periodic item indicates a periodic event in the landslide area; and the sudden item indicates a sudden event in the landslide area; The second prediction module is configured to use a weighted sum of the trend landslide deformation variable, the periodic landslide deformation variable, and the sudden landslide deformation variable of the landslide area at the next moment as a predicted value of the landslide deformation in the landslide area at the next moment; the weighting coefficients of the trend landslide deformation variable, the periodic landslide deformation variable, and the sudden landslide deformation variable are determined by the type of the landslide area; The third prediction module is used to, when the landslide area becomes unstable, determine adjacent landslide areas with potential chain instability among the multiple adjacent landslide areas based on the stress increment and pore water pressure coupling between the landslide area and each of the multiple adjacent landslide areas of the landslide area; and use the predicted value of the landslide deformation of the landslide area at the next moment and the adjacent landslide areas with potential chain instability as landslide prediction information of the landslide area; wherein, the basis for judging the instability of the landslide area is that the landslide deformation acceleration of the landslide area in the time period at the current moment is greater than the instability threshold.

Images