Seismic landslide space scale prediction method and system based on graph deep learning

Abstract

The invention relates to a seismic landslide space scale prediction method and system based on graph deep learning, and the method comprises the steps: dividing a slope unit according to a digital elevation model of a research region, obtaining geographic elements related to a landslide disaster, integrating the data of each geographic element based on the slope unit, and carrying out the prediction of the landslide space scale. Constructing sub-graph data by utilizing spatial connection among the slope units in the grids, and forming a graph structure data set by all the sub-graph data; the slope unit in the subgraph is regarded as a node v, any two nodes are regarded as having a connection relationship, the connection relationship between the nodes is represented by using an edge matrix E, the element value of the edge matrix E is 1, and the landslide plane area in the slope unit without landslide occurrence is set as 0; and evaluating the plane area of the earthquake landslide in the area range by using the graph structure data set and a graph Transform model. And a practical framework and a technical path are provided for spatial distribution of landslide risks in an area range under the action of extreme events, especially earthquake landslide scale prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of geological disaster assessment and prediction, and in particular to a method for predicting the spatial scale of earthquake landslides based on graph deep learning. Background Art

[0002] Landslide is a common geological disaster, especially in areas with complex terrain such as mountains and hills. Extreme weather events (such as heavy rains and earthquakes) can easily induce landslides, causing serious loss of life and property. Therefore, accurately predicting the scope and spatial scale of earthquake landslides is of great significance for disaster prevention and mitigation.

[0003] The spatial scale of landslides is an important indicator of the degree of landslide hazard. However, it is difficult to assess the spatial scale of all earthquake landslides in a regional scale, especially for clustered landslides caused by extreme events. On the contrary, it is much easier to draw a list of polygonal landslides in a regional scale, so their shape index (such as width-to-length ratio) or other morphological characteristics (such as area) can be used to interpret the hazard of landslides. Traditional landslide hazard prediction models are mainly based on geological data, meteorological data, etc. to model the spatial or temporal frequency of landslides. However, these methods cannot explain the spatial scale of landslides. In addition, in general, the scale of landslides in a regional scale shows a significant long-tail distribution feature, that is, most slopes have no landslides or are small in scale, while the number of larger and larger landslides is relatively much smaller. Traditional methods exclude slopes with zero landslide area in the modeling process, because this may bias the prediction towards smaller values ​​and underestimate large-scale landslides. However, this method ignores the information provided by a large number of slopes with zero landslide area and cannot globally consider the information provided by all slopes. Summary of the invention

[0004] In view of the shortcomings of the prior art, the purpose of the present invention is to propose a method for spatial scale prediction of earthquake landslides based on graph deep learning, so as to provide a practical framework and technical path for the spatial distribution of landslide hazards in a region under the action of extreme events, especially for the scale prediction of earthquake landslides.

[0005] To achieve the above object, the technical solution of the present invention is as follows:

[0006] In a first aspect, the present invention provides a method for predicting the spatial scale of earthquake landslides based on graph deep learning, the prediction method comprising the following steps:

[0007] S1. Obtain the digital elevation model of the study area and divide the study area into slope units;

[0008] S2, obtain various geographical elements related to landslide hazards, and use different integration methods to integrate the geographical elements into each slope unit;

[0009] S3, sliding fixed-size grids in turn to extract slope units, extracting multiple slope units from each grid, all slope units extracted from the same grid constitute a slope unit subset, each slope unit subset is a subgraph data, the slope unit in the subgraph is regarded as a node v, any two nodes v i and v j are all considered to have a connection relationship, and the connection relationship between the nodes is represented by the edge matrix E∈R m×m Represented by, where m represents the number of nodes in the subgraph data, the element values ​​of the edge matrix E are all 1, the characteristics of the nodes in the subgraph data are the geographical elements integrated into the corresponding slope unit using different integration methods, and the label data of the subgraph data nodes are the size of the plane area of ​​the landslide in the corresponding slope unit. The plane area of ​​the landslide in the slope unit without landslide is set to 0; all subgraph data constitute a graph structure data set of landslide influencing factors, one subgraph corresponds to one edge matrix, and there is no connection relationship between adjacent subgraphs;

[0010] S4. Using a graph structured dataset, a graph Transformer model is used to evaluate the planar area of ​​earthquake landslides in the regional scope.

[0011] Furthermore, the slope unit is obtained by running r.slopunit code in Grass GIS environment.

[0012] Preferably, the digital elevation model is an SRTM model, an ASTER model, an ALOS model or a Copernicus model.

[0013] Furthermore, the geographical elements include: area of ​​slope unit, slope gradient, slope aspect, profile curvature, plane curvature, terrain position index, surface cover, stratum lithology, road density, earthquake intensity and peak ground acceleration.

[0014] The integration method includes: the mean, variance, mode, and maximum of the geographical elements within the slope unit;

[0015] For the elements of continuous variables, including slope gradient, slope aspect, profile curvature, plan curvature, and terrain position index, their mean and variance values ​​within the slope unit are used as element features and integrated into the slope unit;

[0016] For the elements of categorical variables, including surface cover and stratigraphic lithology elements, the mode of the number of element sub-categories within the slope unit is used as the element feature integrated into the slope unit;

[0017] For the road density, earthquake intensity, and peak ground acceleration elements, the maximum value of the element within the slope unit is used as the element feature and integrated into the slope unit.

[0018] Furthermore, the specific process of step 3 is:

[0019] Step S31, initialize the grid size, and slide the grid in sequence starting from the upper left corner of the studied area, with a sliding step length of 1 grid unit;

[0020] Step S32: Calculate the area of ​​the intersection of each slope unit and the first step grid, and use the area ratio formula to evaluate whether each slope unit can be extracted by the grid, and use the area ratio R to determine the slope unit subset extracted by the first step grid. The area ratio formula is:

[0021]

[0022] Among them, u i ∩K l represents the area where the i-th slope unit intersects with the l-th step grid in space, S(u i ∩K l ) represents the area of ​​the region, represents the area of ​​the i-th slope unit;

[0023] When R≥0.5, it indicates that the slope unit is extracted by the grid in step l, otherwise it will not be extracted by the grid in step l; all slope units extracted by the grid in step l constitute the slope unit subset;

[0024] Step S33: After the grid has traversed all slope units in the study area, L slope unit subsets are obtained and processed into L sub-graph data, where L is the number of sliding times. The slope unit is regarded as a node v in the sub-graph data. Any two nodes v in the sub-graph i and v j are considered as two nodes with a connection relationship.

[0025] Furthermore, the graph Transformer model includes a node embedding layer, a position embedding module and a multi-head attention module and a subsequent feedforward neural network.

[0026] Preferably, the graph structure data set sequentially passes through a node embedding layer, a position embedding module, a multi-head attention module and a subsequent feedforward neural network to output the predicted plane area of ​​the landslide.

[0027] Furthermore, the position embedding module uses the Weisfeiler-Lehman (WL) method to assign an absolute position node code to each node in the graph structure. i The node v j , whose node encoding is represented by WL(v j )∈N, the embedded node vector is represented as

[0028]

[0029] in, Represents node v j The position in the entire graph structure is determined by the node encoding WL(v j ) Use the sin() and cos() functions to update the position of the nodes in the subgraph in the entire graph structure; s is the number of all nodes in the graph structure dataset, d h The dimension of the node feature after the position update.

[0030] Furthermore, the graph Transformer model uses the squared error correlation area SERA function as the loss function of the graph Transformer model and defines a correlation function Use the correlation function to map both the true label and the predicted label value to [0,1], count the range of the true label value after mapping, determine the endpoint values ​​n1 and n2, n1 and n2 are between 0 and 1; use the plane area as the horizontal coordinate, the vertical coordinate is the square error between the true value and the predicted value after mapping, and set multiple discontinuity points t between n1 and n2, and the interval between adjacent t values ​​is 0.01-0.05.

[0031]

[0032] Among them, D t Represents y i The set of values ​​greater than the discontinuity point t; SER t is the square error at the discontinuity point t, and i are the predicted value and true value after mapping of sample i.

[0033] Furthermore, the correlation function The optional conversion type functions include: i) linear, ii) exponential, iii) logarithmic, iv) square root, and at least one of various roots; the various roots include v) square root, vi) cube root, vii) fifth root, viii) eighth root, ix) tenth root, x) twentieth root.

[0034] In a second aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which can implement the steps of the method when executed by a processor.

[0035] In a third aspect, the present invention provides a spatial scale prediction system for earthquake landslides based on graph deep learning, the system comprising a data processing module and a landslide area prediction module;

[0036] The data processing module divides the slope units according to the digital elevation model of the study area, obtains the geographical elements related to landslide hazards, integrates the data of various geographical elements based on the slope units, and constructs sub-graph data using the spatial connection between slope units in the grid. All sub-graph data constitute a graph structure data set;

[0037] The landslide area prediction module takes the sub-graph as input and the plane area of ​​the landslide in the slope unit as output. It uses the graph structure dataset to train the graph Transformer model for earthquake landslide area prediction.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] 1. The present invention uses slope units as the basic calculation units and takes all slope units into consideration in the model, i.e., slope units with a landslide area of ​​0 are not excluded. The sliding grid method is used to extract slope units in the study area as samples of graph structure data. This type of data structure can establish spatial dependencies within slope units and has richer spatial connection information than traditional structured data.

[0040] 2. The present invention uses a graph Transformer model to process graph data with spatial connection relationships to achieve landslide area modeling. The position embedding module and multi-head attention mechanism can effectively capture the global spatial correlation within the landslide occurrence area, making up for the defect that the interdependence between slopes cannot be considered in traditional landslide disaster spatial modeling, thereby improving the accuracy of earthquake landslide area prediction.

[0041] 3. The SERA function is used as the loss function for model training, and the label value is mapped to between 0 and 1 using the correlation function, which can minimize the error of the overall distribution of the sample under consideration, while taking into account the distribution characteristics of the population and the tail. Therefore, the model architecture of the present invention can estimate the planar area of ​​the landslide in the area of ​​interest without selecting a specific statistical distribution, and can still explain the problem of sample imbalance. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the specific implementation of the present invention, the following will briefly introduce the drawings required for use in the specific implementation. In all the drawings, each element or part is not necessarily drawn according to the actual scale.

[0043] Figure 1 A flowchart of a method for predicting spatial scale of earthquake landslides based on graph deep learning provided by an embodiment of the present invention;

[0044] Figure 2 A schematic diagram of a method for extracting slope units in a study area by sliding a fixed-size grid provided in an embodiment of the present invention.

[0045] In the figure, the white grid is a grid of fixed size. The initial grid size is set to 5*5km, and multiple slope units can be extracted from each grid.

[0046] Figure 3 A schematic diagram of an extracted subgraph structure provided in one embodiment of the present invention, wherein dots represent nodes, nodes represent slope units extracted from a grid, the color of the nodes is mapped to the area where landslides occur in the slope units, and there are edges connecting any nodes.

[0047] Figure 4 This is an earthquake landslide area prediction map provided by an embodiment of the present invention. The darker the color, the larger the predicted landslide area in the slope unit. White means that no landslide occurs in the slope unit, that is, the landslide area is 0. DETAILED DESCRIPTION

[0048] The following embodiments of the technical solution of the present invention are described in detail in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and are therefore only used as examples, and cannot be used to limit the protection scope of the present invention.

[0049] like Figure 1 The flowchart of the spatial scale prediction method of earthquake landslide based on graph deep learning is shown in FIG. The prediction method includes:

[0050] Step 1: Obtain the digital elevation model of the study area and divide the study area into slope units;

[0051] Step 2, obtaining geographical elements such as geological topography, hydrological environment and the like related to landslide hazards, and integrating the geographical elements into the slope unit using different integration methods;

[0052] Step 3: Slide the fixed-size grids in turn to extract slope units. Multiple slope units are extracted from each grid. All slope units extracted from the same grid constitute a slope unit subset. Each slope unit subset is a subgraph data. The slope unit in the subgraph is regarded as a node v. Any two nodes v i and v j are all considered to have a connection relationship, and the connection relationship between the nodes is represented by the edge matrix E∈R m×m Represented by, where m represents the number of nodes in the subgraph data, the element values ​​of the edge matrix E are all 1, the characteristics of the nodes in the subgraph data are the geographical elements integrated into the corresponding slope unit using different integration methods, the label data of the subgraph data nodes are the size of the plane area of ​​the landslide in the corresponding slope unit, and the plane area of ​​the landslide in the slope unit (node) where no landslide occurs is set to 0; all subgraph data constitute a graph structure dataset of landslide influencing factors, a subgraph Figure 1 There are no connections between adjacent subgraphs.

[0053] Step 4: According to the graph structure data set, the graph Transformer model is used to evaluate the plane area of ​​earthquake landslides within the regional range. The nonlinear mapping relationship between the influencing factors of landslides and the plane area of ​​landslides is evaluated by the trained graph Transformer model, so as to realize the prediction of the spatial distribution of landslide area. The input of the graph Transformer model is a subgraph (including the structure and corresponding features of the node), and the output is the predicted plane area of ​​the landslide in the node.

[0054] Specifically, first, a digital elevation model of the study area of ​​interest can be obtained, and based on this model, the slope unit division method is applied to divide the study area into multiple slope units. Subsequently, geographical elements such as geological topography, hydrological environment, and surface cover related to landslide hazards can be obtained, and various geographical elements can be integrated into the slope units using different integration methods. For the extracted slope units, different slope unit sets (i.e., slope unit subsets) are segmented using a grid with a fixed size, and these slope unit sets can then be constructed as a graph structured data set. Finally, the graph structured data set is input into the graph Transformer model for training and testing the model's performance in predicting the plane area of ​​landslides within a regional scope.

[0055] In this embodiment, optionally, in step 1, the digital elevation model is an SRTM model, an ASTER model, an ALOS model, a Copernicus model, etc., and the spatial resolution can be selected to be 12×12, 30×30, or 90×90 meters.

[0056] Optionally, in step 1, the slope unit is calculated by running the r.slopunit code in the Grass GIS environment.

[0057] In this embodiment, optionally, in step 2, the geographical elements such as geological topography, hydrological environment and the like related to landslide hazards include: the area of ​​slope unit, slope gradient, slope aspect, profile curvature, plane curvature, terrain position index, surface cover, stratum lithology, road density, earthquake intensity and peak ground acceleration. Among them, geological topography: the area of ​​slope unit, slope gradient and slope aspect; hydrological environment: profile curvature, plane curvature and terrain position index;

[0058] Optionally, in step 2, the different integration methods include: mean value, variance, mode, maximum value, etc. of geographical elements within the slope unit.

[0059] Specifically, for the elements of continuous variables, such as slope gradient, slope aspect, profile curvature, plan curvature, and terrain position index geographical elements, their mean and variance values ​​within the slope unit are used as element features and integrated into the slope unit;

[0060] For the elements of categorical variables, such as surface cover and stratum lithology, the mode of the number of element sub-categories in the slope unit is used as the element feature integrated into the slope unit;

[0061] For road density, earthquake intensity, and peak ground acceleration elements, the maximum value of the element within the slope unit is used as the element feature and integrated into the slope unit.

[0062] The specific process of step 3 is:

[0063] Step S31, the initialization grid size is set to 5-10km, in this embodiment, it is set to 5×5km, and the grid is slid from left to right, from top to bottom, or from top to bottom, from left to right, starting from the upper left corner of the studied area; the sliding step is 1 grid unit; the size of the grid is defined as 5×5km in length and width;

[0064] Step S32: Calculate the area of ​​the intersection of each slope unit and the first step grid, and use the area ratio formula to evaluate whether each slope unit can be extracted by the grid, and use the area ratio R to determine the slope unit subset extracted by the first step grid. The area ratio formula is:

[0065]

[0066] Among them, u i ∩K l represents the area where the i-th slope unit intersects with the l-th step grid in space, S(u i ∩K l ) represents the area of ​​the region, Represents the area of ​​the i-th slope unit. When R ≥ 0.5, it indicates that the slope unit is extracted by the l-step grid, otherwise it will not be extracted by the l-step grid. All slope units extracted by the l-step grid constitute the slope unit subset.

[0067] Step S33: After the grid has traversed all slope units in the study area, L slope unit subsets are obtained and processed into L sub-graph data, where L is the number of sliding times, and the slope unit is regarded as a node v in the sub-graph data. Any two nodes (slope units) v in the sub-graph i and v j Both are regarded as two nodes (slope units) with a connection relationship.

[0068] Specifically, the planar area of ​​the landslide within the slope unit can be realized in the GIS platform by calculating the area of ​​the polygon.

[0069] In this embodiment, optionally, in step 4, the graph Transformer model includes a node embedding layer, a position embedding module and a multi-head attention module and a subsequent feedforward neural network.

[0070] Specifically, the graph structure dataset passes through a node embedding layer, a position embedding module, a multi-head attention module, and a subsequent feedforward neural network in sequence.

[0071] Specifically, the position embedding module uses the Weisfeiler-Lehman (WL) method to assign an absolute position node code to each node in the graph structure. Specifically, for the subgraph node set (V i ) of the node v j , whose node encoding is represented by WL(v j )∈N, so the embedded node vector is expressed as

[0072]

[0073] in, Represents node v j The position in the entire graph structure is determined by the node encoding WL(v j ) Use the sin() and cos() functions to update the position of the node in the subgraph in the entire graph structure. When the position is updated, the characteristics of the node also change accordingly; s is the number of all nodes in the graph structure dataset, d h The dimension of the node feature after the position update.

[0074] Specifically, the calculation method of attention weight in the multi-head attention module is:

[0075]

[0076] Where Q, K, and V are query matrices, key matrices, and value matrices, respectively. k is the dimension of the key vector.

[0077] In this embodiment, optionally, in step 4, the graph Transformer model uses the squared error correlation area (SERA) function as the loss function of the graph Transformer model. The purpose is to eliminate the significant long-tail distribution characteristics of extreme landslide area labels, that is, there are many landslides with small areas and few landslides with large areas, that is, most slope units are displayed as no landslides (landslide area is 0), and the proportion of super-large landslides is very small, thereby avoiding modeling deviations.

[0078] Define a correlation function Use the correlation function to map both the true label and the predicted label value to [0,1], count the range of the true label value after mapping, determine the endpoint values ​​n1 and n2, n1 and n2 are between 0 and 1; use the plane area as the horizontal coordinate, the vertical coordinate is the square error between the true value and the predicted value after mapping, and set multiple discontinuity points t between n1 and n2, and the interval between adjacent t values ​​is 0.01-0.05.

[0079] The calculation method of SERA function is as follows:

[0080]

[0081] Among them, D t Represents y i The set of values ​​greater than the discontinuity point t; SER t is the square error at the discontinuity point t, and i are the predicted value and true value after mapping of sample i.

[0082] Specifically, the SERA function, the correlation function The optional conversion type functions include: i) linear, ii) exponential, iii) logarithmic, iv) square root, and various roots, including at least one of v) square root, vi) cube root, vii) fifth root, viii) eighth root, ix) tenth root, x) twentieth root, etc. Select a suitable correlation function according to the characteristics of the landslide, such as the fifth root, eighth root, tenth root and logarithmic transformation are preferred for the long-tail distribution landslide area.

[0083] In this embodiment, optionally, different correlation functions can be used in the process of training the graph Transformer model. Verify the trained graph Transformer model. The specific training process of the graph Transformer model includes:

[0084] Step S41, constructing a training set and a validation set through the graph structure data;

[0085] Step S42: using a standardization method to standardize the training set and the validation set respectively;

[0086] Step S43: According to different correlation functions Obtain different SERA loss functions, train the graph Transformer model, then input the training set into the graph Transformer model and train the model using error backpropagation and the SERA function as the loss function;

[0087] Step S44, using the root mean square (MSE) function to evaluate the performance of each graph Transformer model in the validation set, and taking the graph Transformer model with the smallest MSE value in the validation set and the corresponding SERA loss function as the optimal model of the plane area of ​​earthquake landslide;

[0088] Step S45, according to the performance of the subsequent modeling results (such as model prediction error or prediction accuracy, the smaller the better, adjust the grid size as needed), adjust the size of the initialization grid to (5+0.5c)×(5+0.5c)km, where c is the number of subsequent grid size adjustments, and repeat the above process to obtain a graph Transformer model that meets the error accuracy requirements for the prediction of the spatial distribution of landslide area.

[0089] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. 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 the present invention.

[0090] Any matters not described in the present invention are applicable to the prior art.

Claims

1. A spatial scale prediction method for earthquake landslides based on graph deep learning, characterized in that: The prediction method comprises the following steps: S1. Obtain the digital elevation model of the study area and divide the study area into slope units; S2, obtain various geographical elements related to landslide hazards, and use different integration methods to integrate the geographical elements into each slope unit; S3, sliding fixed-size grids in turn to extract slope units, extracting multiple slope units from each grid, all slope units extracted from the same grid constitute a slope unit subset, each slope unit subset is a subgraph data, the slope unit in the subgraph is regarded as a node v, any two nodes v i and v j are all considered to have a connection relationship, and the connection relationship between the nodes is represented by the edge matrix E∈R m×m Represented by, where m represents the number of nodes in the subgraph data, the element values ​​of the edge matrix E are all 1, the characteristics of the nodes in the subgraph data are the geographical elements integrated into the corresponding slope unit using different integration methods, and the label data of the subgraph data nodes are the size of the plane area of ​​the landslide in the corresponding slope unit. The plane area of ​​the landslide in the slope unit without landslide is set to 0; all subgraph data constitute a graph structure data set of landslide influencing factors, one subgraph corresponds to one edge matrix, and there is no connection relationship between adjacent subgraphs; S4. Using a graph structured dataset, a graph Transformer model is used to evaluate the planar area of ​​earthquake landslides in the regional scope.

2. The method for predicting the spatial scale of earthquake landslides according to claim 1, characterized in that: The slope unit is obtained by running r.slopunit code in Grass GIS environment; the digital elevation model is SRTM model, ASTER model, ALOS model or Copernicus model.

3. The method for predicting the spatial scale of earthquake landslides according to claim 1, characterized in that: The geographical elements include: area of ​​slope unit, slope gradient, slope aspect, profile curvature, plane curvature, terrain position index, surface cover, stratum lithology, road density, earthquake intensity and peak ground acceleration; The integration method includes: the mean, variance, mode, and maximum of the geographical elements within the slope unit; For the elements of continuous variables, including slope gradient, slope aspect, profile curvature, plan curvature, and terrain position index, their mean and variance values ​​within the slope unit are used as element features and integrated into the slope unit; For the elements of categorical variables, including surface cover and stratigraphic lithology elements, the mode of the number of element sub-categories within the slope unit is used as the element feature integrated into the slope unit; For the road density, earthquake intensity, and peak ground acceleration elements, the maximum value of the element within the slope unit is used as the element feature and integrated into the slope unit.

4. The method for predicting the spatial scale of earthquake landslides according to claim 1, characterized in that: The specific process of step 3 is: Step S31, initialize the grid size, and slide the grid in sequence starting from the upper left corner of the studied area, with a sliding step of 1 grid unit; Step S32: Calculate the area of ​​the intersection of each slope unit and the first step grid, and use the area ratio formula to evaluate whether each slope unit can be extracted by the grid, and use the area ratio R to determine the slope unit subset extracted by the first step grid. The area ratio formula is: Among them, u i ∩K l represents the area where the i-th slope unit intersects with the l-th step grid in space, S(u i ∩K l ) represents the area of ​​the region, represents the area of ​​the i-th slope unit; When R≥0.5, it indicates that the slope unit is extracted by the grid in step l, otherwise it will not be extracted by the grid in step l; all slope units extracted by the grid in step l constitute the slope unit subset; Step S33: After the grid has traversed all slope units in the study area, L slope unit subsets are obtained and processed into L sub-graph data, where L is the number of sliding times. The slope unit is regarded as a node v in the sub-graph data. Any two nodes v in the sub-graph i and v j are considered as two nodes with a connection relationship.

5. The method for predicting the spatial scale of earthquake landslides according to claim 1, characterized in that: The graph Transformer model includes a node embedding layer, a position embedding module, a multi-head attention module and a subsequent feedforward neural network; the graph structure data set passes through the node embedding layer, the position embedding module, the multi-head attention module and the subsequent feedforward neural network in sequence to output the predicted plane area of ​​the landslide.

6. The method for predicting the spatial scale of earthquake landslides according to claim 1, characterized in that: The position embedding module uses the Weisfeiler-Lehman (WL) method to assign an absolute position node code to each node in the graph structure. i The node v j , whose node encoding is represented by WL(v j )∈N, the embedded node vector is represented as in, Represents node v j The position in the entire graph structure is determined by the node encoding WL(v j ) Use the sin() and cos() functions to update the position of the nodes in the subgraph in the entire graph structure; s is the number of all nodes in the graph structure dataset, d h The dimension of the node feature after the position update.

7. The method for predicting the spatial scale of earthquake landslides according to claim 1, characterized in that: The graph Transformer model uses the squared error correlation area SERA function as the loss function of the graph Transformer model and defines a correlation function Use the correlation function to map both the true label and the predicted label value to [0,1], count the range of the true label value after mapping, determine the endpoint values ​​n1 and n2, n1 and n2 are between 0 and 1; use the plane area as the horizontal coordinate, the vertical coordinate is the square error between the true value and the predicted value after mapping, and set multiple discontinuity points t between n1 and n2, and the interval between adjacent t values ​​is 0.01-0.

05. Among them, D t Represents y i The set of values ​​greater than the discontinuity point t; SER t is the square error at the discontinuity point t, and i are the predicted value and true value after mapping of sample i.

8. The method for predicting the spatial scale of earthquake landslides according to claim 7, characterized in that: The correlation function The optional conversion type functions include: i) linear, ii) exponential, iii) logarithmic, iv) square root, and at least one of various roots; the various roots include v) square root, vi) cube root, vii) fifth root, viii) eighth root, ix) tenth root, x) twentieth root.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of any method described in claims 1-8 can be implemented.

10. A spatial scale prediction system for earthquake landslides based on graph deep learning, characterized in that: The system includes a data processing module and a landslide area prediction module; The data processing module divides the slope units according to the digital elevation model of the study area, obtains the geographical elements related to landslide hazards, integrates the data of various geographical elements based on the slope units, and constructs sub-graph data using the spatial connection between slope units in the grid. All sub-graph data constitute a graph structure data set; The landslide area prediction module takes the sub-graph as input and the plane area of ​​the landslide in the slope unit as output. It uses the graph structure dataset to train the graph Transformer model for earthquake landslide area prediction.

Images