Wide-area landslide susceptibility prediction method

Abstract

The invention discloses a wide-area landslide susceptibility prediction method, which comprises the following steps of: constructing a directed connection graph of a target wide area, and inputting node data into a residual feature conversion module in a landslide risk prediction model to obtain a first latent feature matrix; inputting the first latent feature matrix into a first linear self-attention Transform module to obtain a query matrix, a key matrix and a value matrix; further obtaining a diagonal matrix; obtaining a second latent feature matrix through the diagonal matrix, the value matrix, the normalized query matrix and the normalized key matrix; a third latent feature matrix is obtained through a second latent feature matrix and a second linear self-attention Transform module; inputting the directed connection graph into a graph filter layer for normalization and power iteration processing to obtain a graph filtering operator, and performing hierarchical neighbor aggregation processing on node data to obtain an aggregation feature matrix; a fused feature matrix is obtained through the aggregation feature matrix, the third latent feature matrix and the first multi-layer perceptron module, and the landslide risk probability is further obtained.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of landslide susceptibility prediction, and in particular to a wide-area landslide susceptibility prediction method. BACKGROUND

[0002] Under the influence of seasonal heavy rainfall and large-scale engineering construction, many landslides occur every year, often causing serious losses to local residents, building facilities and the environment, and therefore it is necessary to carry out efficient landslide prediction to reduce unnecessary losses and ensure the safety of residents.

[0003] In related technologies, support vector machines (SVM), SVM optimized by genetic algorithms (GA-SVM), generalized regression neural networks and other models are usually used for prediction, but single models have limited generalization ability, are strongly dependent on data quality and parameter accuracy, are difficult to fully capture the coupling effect of multiple factors, are prone to overfitting or underfitting, and have poor dynamic adaptability, resulting in prediction results that may deviate from reality or be difficult to apply to complex and variable landslide scenarios.

[0004] Therefore, how to quickly and accurately predict whether a region in a large-scale complex heterogeneous geographical environment will have a landslide has become a problem to be solved. SUMMARY

[0005] Therefore, the wide-area landslide susceptibility prediction method provided by the embodiments of the present application at least solves the problem that related technologies cannot quickly and accurately predict whether a region in a large-scale complex heterogeneous geographical environment will have a landslide.

[0006] According to a first aspect of the embodiments of the present application, a wide-area landslide susceptibility prediction method is provided, comprising: dividing the geographical environment of a target wide area to obtain a plurality of slope units; and constructing a directed connection graph of the target region using the plurality of slope units, wherein the node data of the directed connection graph is the plurality of slope units, and the edge data of the directed connection graph is the influence and dependence relationship between the plurality of slope units; inputting the node data into a residual feature conversion module in a landslide risk prediction model to obtain a first latent feature matrix; the landslide risk prediction model further comprises two identical linear self-attention Transformer modules, a plurality of multilayer perception machine modules and a graph filter layer; inputting the first latent feature matrix into the first linear self-attention Transformer module to obtain a query matrix, a key matrix and a value matrix; and obtaining a diagonal matrix based on the normalized query matrix and the normalized key matrix; obtaining a second latent feature matrix based on the diagonal matrix, the value matrix, the normalized query matrix and the normalized key matrix; and obtaining a third latent feature matrix based on the second latent feature matrix and a second linear self-attention Transformer module; inputting the directed connection graph into the graph filter layer for normalization and power iteration processing to obtain a graph filter operator, and performing hierarchical neighbor aggregation processing on the node data by using the graph filter operator to obtain an aggregated feature matrix; obtaining a fused feature matrix based on the aggregated feature matrix, the third latent feature matrix and a first multi-layer perception module, and obtaining a predicted landslide risk probability based on the fused feature matrix.

[0007] Optionally, the diagonal matrix is obtained based on the normalized query matrix and the normalized key matrix, including: obtaining the diagonal matrix based on the normalized query matrix, the normalized key matrix, an all-1 unit column matrix and a diagonal matrix formula, the diagonal matrix formula being as follows: ; In the above formula, is a diagonal matrix, is a normalized query matrix, is a normalized key matrix, is an all-1 unit column matrix, N is a dimension, and T is a transpose.

[0008] Optionally, the second latent feature matrix is obtained based on the diagonal matrix, the value matrix, the normalized query matrix and the normalized key matrix, including: obtaining the second latent feature matrix based on the diagonal matrix, the value matrix, the normalized query matrix, the normalized key matrix, a first latent feature matrix and a latent feature matrix formula, the latent feature matrix formula being as follows: ; In the above formula, Z is the second latent feature matrix, is an inverse matrix of , V is a value matrix, is a hyperparameter of residual connection, is the first latent feature matrix.

[0009] Optionally, the hierarchical neighbor aggregation processing on the node data by using the graph filter operator to obtain the aggregated feature matrix, including: obtaining the aggregated feature matrix based on the graph filter operator, the node data and a feature aggregation formula, the feature aggregation formula being as follows: ; In the above formula, X is the node data, is the graph filter operator, and r is a maximum number of neighbor hops allowed to be considered. The graph filter operator is as follows: ; In the above formula, A is an adjacency matrix, and D is a degree matrix.

[0010] Optionally, the obtaining of the fused feature matrix based on the aggregated feature matrix, the third latent feature matrix and the first multi-layer perception module, and the obtaining of the predicted landslide risk probability based on the fused feature matrix comprise: obtaining a fourth latent feature matrix based on the aggregated feature matrix and the third multi-layer perception module, and obtaining the fused feature matrix based on the third latent feature matrix, the fourth latent feature matrix and a feature fusion formula in the first multi-layer perception module, the feature fusion formula being as follows: ; In the above formula, is the fused feature matrix, is a weight matrix, is the third latent feature matrix, is the fourth latent feature matrix.

[0011] Optionally, the residual feature conversion module comprises a second multi-layer perception module and a layer normalization layer. The inputting of the node data into the residual feature conversion module in the landslide risk prediction model to obtain the first latent feature matrix comprises: inputting the node data into the second multi-layer perception module to obtain a linearly transformed latent feature matrix, and inputting the linearly transformed latent feature matrix into the layer normalization layer to obtain a processed latent feature matrix; performing activation function processing on the processed latent feature matrix to obtain a non-linearly transformed latent feature matrix, and fusing the non-linearly transformed latent feature matrix and the node data through residual connection to obtain the first latent feature matrix.

[0012] According to a second aspect of the embodiment of the present application, an electronic device is provided, comprising: a processor, a memory, a communication interface and a communication bus, the processor, the memory and the communication interface complete communication with each other through the communication bus; the memory is used to store at least one executable instruction, and the executable instruction causes the processor to perform operations corresponding to the method of the first aspect.

[0013] According to a third aspect of the embodiment of the present application, a computer storage medium is provided, which stores a computer program, and the program is executed by a processor to implement the method of the first aspect.

[0014] According to the scheme provided by the embodiment of the present application, a target wide-area geographical environment is divided to obtain a plurality of slope units; a directed connection graph of a target region is constructed using the plurality of slope units, node data of the directed connection graph being the plurality of slope units, and edge data of the directed connection graph being an influence and dependence relationship between the plurality of slope units; the node data is input into a residual feature conversion module in a landslide risk prediction model to obtain a first latent feature matrix; the landslide risk prediction model further includes two identical linear self-attention Transformer modules, a plurality of multilayer perceptron modules, and a graph filter layer; the first latent feature matrix is input into a first linear self-attention Transformer module to obtain a query matrix, a key matrix, and a value matrix; a diagonal matrix is obtained based on the normalized query matrix and the normalized key matrix; a second latent feature matrix is obtained based on the diagonal matrix, the value matrix, the normalized query matrix, and the normalized key matrix; a third latent feature matrix is obtained based on the second latent feature matrix and a second linear self-attention Transformer module; the directed connection graph is input into the graph filter layer for normalization and power iteration processing to obtain a graph filtering operator, and the node data is subjected to hierarchical neighbor aggregation processing using the graph filtering operator to obtain an aggregated feature matrix; a fused feature matrix is obtained based on the aggregated feature matrix, the third latent feature matrix, and a first multilayer perceptron module, and a predicted landslide risk probability is obtained based on the fused feature matrix. In this process, two branches are set, including: a global feature extraction branch, which is a graph Transformer module with linear attention, reduces the computational complexity through an improved key-value decomposition strategy, and realizes efficient modeling of the global dependence relationship between nodes; and a local feature extraction branch, which uses a graph filter layer to realize efficient feature aggregation, capture local spatial features in a heterogeneous geographical environment, and integrate the features extracted by the two branches through weighted fusion. Finally, a consistency regularization loss function is combined to alleviate the overfitting problem caused by sample sparseness and improve the accuracy of landslide risk probability prediction. BRIEF DESCRIPTION OF DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor based on these drawings. Figure 1 A flowchart of a wide-area landslide susceptibility prediction method provided by an embodiment of the present application; Figure 2 An effect diagram of a main framework provided by an embodiment of the present application; Figure 3 An effect schematic diagram of the first linear self-attention Transformer module operation provided by the embodiment of the present application is shown in FIG. 6. Figure 4 An effect schematic diagram of the graph filtering operator operation process provided by the embodiment of the present application is shown in FIG. 7. Figure 5 An effect schematic diagram of the feature fusion provided by the embodiment of the present application is shown in FIG. 8. Figure 6 A structural schematic diagram of an electronic device of the embodiment of the present application is shown in FIG. 9. DETAILED DESCRIPTION

[0016] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some of the embodiments of the present application, but not all the embodiments of the present application. The following embodiments are used to illustrate the present application, but not to limit the scope of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present application.

[0017] In the following description, “some embodiments” are described, which describe a subset of all possible embodiments, but it can be understood that “some embodiments” can be the same subset or different subsets of all possible embodiments, and can be combined with each other without conflict.

[0018] It should be noted that the terms “first\second\third” involved in the embodiments of the present application are only to distinguish similar objects, and do not represent a specific order of the objects. It can be understood that “first\second\third” can be interchanged with a specific order or sequence as allowed, so that the embodiments of the present application described here can be implemented in an order other than that illustrated or described here.

[0019] Those skilled in the art can understand that, unless otherwise defined, all terms (including technical terms and scientific terms) used herein have the same meaning as generally understood by those skilled in the art to which the embodiments of the present application belong. It should also be understood that terms such as those defined in general dictionaries should be understood as having meanings consistent with those in the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as such.

[0020] Figure 1 A flowchart of a wide-area landslide susceptibility prediction method provided by the embodiment of the present application is shown in FIG. 10. The wide-area landslide susceptibility prediction method provided by the embodiment of the present application can be executed by an electronic device, which can be a computer, a server, etc.

[0021] As shown in Figure 1 A wide-area landslide susceptibility prediction method, comprising: S101, divide the geographical environment of the target area to obtain a plurality of slope units; and use the plurality of slope units to construct a directed connection graph of the target area, the node data of the directed connection graph being the plurality of slope units, and the edge data of the directed connection graph being the influence and dependence relationship between the plurality of slope units.

[0022] In an embodiment of the present application, the wide area is a vast area or network connected by long-distance communication technology across geographical or administrative boundaries. The basic data corresponding to the geographical environment of the target wide area is processed and analyzed using professional geographic information analysis algorithms and models, so that the areas with unique topographic properties can be accurately identified and extracted from the geographical environment. Finally, these areas are divided into independent slope units. The slope units are used as the node data of the directed connection graph, and the edge data of the directed connection graph is the influence and dependence relationship between the slope units. The directed connection graph is constructed by the node data and the edge data.

[0023] S102, input the node data into the residual feature conversion module in the landslide risk prediction model to obtain a first latent feature matrix; the landslide risk prediction model further comprises two identical linear self-attention Transformer modules, a plurality of multilayer perceptron modules and a graph filter layer.

[0024] In an embodiment of the present application, the landslide risk prediction model is used to predict the probability of subsequent landslide risk in the target area. The landslide risk prediction model includes a residual feature conversion module, two identical linear self-attention Transformer modules, a plurality of multilayer perceptron modules and a graph filter layer. The residual feature conversion module is used for feature extraction, transformation and residual connection of data. The node data of the directed connection graph is input into the residual feature conversion module in the landslide risk prediction model for feature extraction, transformation and connection to obtain a first latent feature matrix.

[0025] Among them, the plurality of multilayer perceptron modules includes a first multilayer perceptron module, a second multilayer perceptron module and a third multilayer perceptron module, which are used in subsequent steps.

[0026] S103, input the first latent feature matrix into the first linear self-attention Transformer module to obtain a query matrix, a key matrix and a value matrix; and obtain a diagonal matrix based on the normalized query matrix and the normalized key matrix.

[0027] In the embodiments of the present application, the linear self-attention Transformer module decomposes the traditional attention into two steps of normalization mapping and linear interaction. First, after inputting the first latent feature matrix into the first linear self-attention Transformer module, the first latent feature matrix is linearly converted through a linear feedforward layer to obtain a query matrix (Q), a key matrix (K) and a value matrix (V), and Q and K are normalized. The linear conversion formulas are as follows: ; ; ; In the above formulas, are linear feedforward layers, is the first latent feature matrix, is a set of real numbers, N is the number of rows, and u is the number of columns.

[0028] The normalization formula is as follows: ; ; In the above formulas, denotes the Frobenius norm, is the normalized query matrix, is the normalized key matrix, and F is the matrix Euclidean norm.

[0029] Then linear interaction is performed. By using the first-order Taylor expansion of the Softmax function for approximation, the original operation is decomposed, and the calculation process is rearranged from the standard form to the simplified form , and the complexity is reduced by combining the diagonal matrix The calculation of the diagonal matrix is mainly based on the normalized query matrix and the normalized key matrix to obtain the diagonal matrix.

[0030] S104, based on the diagonal matrix, the value matrix, the normalized query matrix and the normalized key matrix, a second latent feature matrix is obtained; and based on the second latent feature matrix and the second linear self-attention Transformer module, a third latent feature matrix is obtained.

[0031] In the embodiment of the present application, the second latent feature matrix is obtained based on the calculation of the diagonal matrix, the value matrix, the normalized query matrix and the normalized key matrix. The architecture and function of the second linear self-attention Transformer module and the first linear self-attention Transformer module are the same, the first latent feature matrix and the second latent feature matrix are fused through the residual connection to obtain the connected latent feature matrix, and the connected latent feature matrix is input into the second linear self-attention Transformer module to perform linear conversion on the connected latent feature matrix through the linear feedforward layer to obtain the query matrix (Q), the key matrix (K) and the value matrix (V), and then the Q and K are normalized, then the diagonal matrix is calculated, and finally the third latent feature matrix is obtained based on the diagonal matrix, the value matrix, the normalized query matrix and the normalized key matrix.

[0032] In the embodiment of the present application, the directed connection graph is input into the graph filter layer for normalization and power iteration processing to obtain a graph filter operator, and the graph filter operator is used for hierarchical neighbor aggregation processing of node data to obtain an aggregated feature matrix.

[0033] In the embodiment of the present application, the directed connection graph is input into the graph filter layer, and the adjacency matrix corresponding to the directed connection graph is normalized and processed by power iteration to obtain a graph filter operator. The formula of the graph filter operator is as follows: ; In the above formula, A is the adjacency matrix, and D is the degree matrix.

[0034] Further, the information of the neighbor nodes is aggregated layer by layer using the graph filter operator and the node features to generate an enhanced representation of each node, and a plurality of different aggregated feature matrices are obtained.

[0035] S106, based on the aggregated feature matrix, the third latent feature matrix and the first multilayer perceptron module, a fused feature is obtained, and based on the fused feature matrix, a landslide risk probability is obtained.

[0036] In the embodiment of the present application, the aggregated feature matrix is a local matrix, the third latent feature matrix is a global matrix, and based on the local matrix, the global matrix and the first multilayer perceptron module, a fused feature matrix is obtained, and based on the fused feature matrix, a landslide risk probability is obtained.

[0037] Figure 2 An effect schematic diagram of a main body frame provided by the embodiment of the present application is shown in FIG. 1. Figure 2In the middle, the node data enters the global information extraction branch processing, specifically: input the node data into the second multilayer perception machine module composed of 48 neurons in the hidden layer, get the linear transformed latent feature matrix, then through the layer normalization layer, activation function and residual connection, get the first latent feature matrix. Input the first latent feature matrix into the first linear self-attention Transformer module to get the second latent feature matrix, and perform residual connection processing on the second latent feature matrix and the first latent feature matrix to get the connected latent feature matrix, then input the connected latent feature matrix into the second linear self-attention Transformer module for the same processing to get the third latent feature matrix. At the same time, the directed connected graph enters the local information extraction branch in parallel, specifically: input the directed connected graph into the graph filter layer for normalization and power iteration processing to get the graph filter operator, and use the graph filter operator to perform hierarchical neighbor aggregation processing on the node data to get the aggregated feature matrix (X1, X2, X3,..., Xn), set the hidden layer to be composed of 48 neurons multiple third multilayer perception modules, input each aggregated feature matrix into the corresponding third multilayer perception module to get the fourth latent feature matrix, and after the fourth latent feature matrix is processed by the activation function, the fourth multilayer perception module and the Softmax function, the latent feature matrix (probability matrix) with the same dimension as the third latent feature matrix is obtained, the two latent feature matrices with the same dimension are weighted and fused, and the fused feature matrix is processed by the first multilayer perception machine module, and finally the predicted landslide risk probability is output.

[0038] It can be understood that in the embodiments of the present application, the target wide-area geographical environment is divided to obtain a plurality of slope units; and a directed connection graph of the target area is constructed using the plurality of slope units, the node data of the directed connection graph being the plurality of slope units, and the edge data of the directed connection graph being the influence and dependence relationship between the plurality of slope units; the node data is input into a residual feature conversion module in the landslide risk prediction model to obtain a first latent feature matrix; the landslide risk prediction model further includes two identical linear self-attention Transformer modules, a plurality of multilayer perceptron modules, and a graph filter layer; the first latent feature matrix is input into the first linear self-attention Transformer module to obtain a query matrix, a key matrix, and a value matrix; and a diagonal matrix is obtained based on the normalized query matrix and the normalized key matrix; a second latent feature matrix is obtained based on the diagonal matrix, the value matrix, the normalized query matrix, and the normalized key matrix; and a third latent feature matrix is obtained based on the second latent feature matrix and the second linear self-attention Transformer module; the directed connection graph is input into the graph filter layer for normalization and power iteration processing to obtain a graph filtering operator, and the graph filtering operator is used for hierarchical neighbor aggregation processing on the node data to obtain an aggregated feature matrix; a fused feature matrix is obtained based on the aggregated feature matrix, the third latent feature matrix, and the first multilayer perceptron module, and a predicted landslide risk probability is obtained based on the fused feature matrix. In this process, two branches are set, including: a global feature extraction branch, which is a graph Transformer model with linear attention, reduces the computational complexity through an improved key-value decomposition strategy, and realizes efficient modeling of the global dependence relationship between nodes; a local feature extraction branch, which uses a graph filtering operator to realize efficient feature aggregation, capture local spatial features in heterogeneous geographical environments, and the features extracted by the two branches are integrated through weighted fusion. Finally, a consistency regularization loss function is combined to alleviate the overfitting problem caused by sample sparseness and improve the accuracy of landslide risk probability prediction.

[0039] In some embodiments of the present application, obtaining the diagonal matrix based on the normalized query matrix and the normalized key matrix in S103 can be implemented by S1031, which is described by the following steps.

[0040] S1031, obtaining the diagonal matrix based on the normalized query matrix, the normalized key matrix, the all-1 unit column matrix, and the diagonal matrix formula.

[0041] In some embodiments of the present application, the all-1 unit column matrix is an N-dimensional all-1 column matrix, and the diag function is used to convert the N-dimensional matrix into an N*N diagonal matrix. The diagonal matrix formula is as follows: ; In the above formula, is a diagonal matrix, is a normalized query matrix, is a normalized key matrix, is a full 1 unit column matrix, N is a dimension, and T is a transpose, is a function for processing diagonal elements of a matrix.

[0042] In some embodiments of the present application, S104 can be implemented by S1041, which is described by the following steps.

[0043] S1041, based on the diagonal matrix, the value matrix, the normalized query matrix, the normalized key matrix, the first latent feature matrix, and the latent feature matrix formula, obtains the second latent feature matrix.

[0044] In some embodiments of the present application, after the diagonal matrix, the value matrix, the normalized query matrix, the normalized key matrix, and the first latent feature matrix, the second latent feature matrix is calculated in combination with the hyperparameter of the residual connection and the latent feature matrix formula. The latent feature matrix formula is as follows: ; In the above formula, Z is the second latent feature matrix, is the first latent feature matrix, and β is the hyperparameter of the residual connection, is the inverse matrix of , and V is the value matrix.

[0045] Figure 3 is a schematic diagram of the effect of the first linear self-attention Transformer module operation provided by the embodiments of the present application, in which Figure 3 the first latent feature matrix is input into the linear attention layer, converted first to obtain Q, K, and V, and Q and K are normalized to obtain normalized Q and K, the size of Q and K is N, u is a feature dimension, is an N-dimensional full 1 column matrix, that is, The column matrix has N rows and 1 column, and the value of each element in each row is 1. The normalized K is transposed, and the transposed K is multiplied by to obtain a u-dimensional full 1 column feature matrix. The u-dimensional full 1 column feature matrix is multiplied by the normalized Q to obtain an N-dimensional full 1 column feature matrix. After the N-dimensional full 1 column feature matrix and the processed matrix (that is, the element values in the S feature matrix are multiplied by N) are added, the N-dimensional feature matrix is converted into an N×N diagonal matrix by diag, and the inverse matrix of is further obtained. At the same time, the transposed normalized K is multiplied by V to obtain a feature matrix (this step is a key operation of the present application, which can reduce the complexity from the original , and the normalized Q is multiplied by the feature matrix to obtain a feature matrix with a size of N and a feature dimension of u, the feature matrix with a size of N and a feature dimension of u is added to the processed V (that is, the element values in the V feature matrix are multiplied by N), and the added feature matrix is multiplied by the inverse matrix of , and the calculated feature matrix is processed through a residual connection, a normalization layer and a Relu activation function, and finally a second latent feature matrix is obtained.

[0046] In some embodiments of the application, the hierarchical neighbor aggregation processing of the node data in S105 using the graph filtering operator can be implemented by S1051, which is described as follows.

[0047] S1051, obtaining an aggregated feature matrix based on the graph filtering operator, the node data and the feature aggregation formula.

[0048] In some embodiments of the application, the feature aggregation formula is as follows: ; In the above formula, X is the node data, is the graph filtering operator, and r is the maximum number of neighbor hops allowed to be considered.

[0049] Figure 4 is an effect diagram of a graph filtering operator operation process provided by an embodiment of the application, in which Figure 4 , i is n, r is the maximum number of neighbor hops allowed to be considered, and the value range of n is obtained according to the value of r. When n is 1, the degree matrix D is obtained according to the adjacency matrix A, and then is calculated according to the graph filtering operator. and X and the feature aggregation formula are used to output When n is 2, similarly, is obtained, and and X and the feature aggregation formula are used to calculate , until n is equal to r, and the output is obtained.

[0050] In some embodiments of the application, S106 can be implemented by S1061, which is described as follows.

[0051] S1061, obtaining a fourth latent feature matrix based on the aggregated feature matrix and the third multi-layer perception, and obtaining a fused feature matrix based on the third latent feature matrix, the fourth latent feature matrix and the feature fusion formula in the first multi-layer perception module.

[0052] In some embodiments of the present application, each aggregated feature matrix corresponds to a third multi-layer perception module, different aggregated feature matrices are input into the respective corresponding third multi-layer perception modules, a plurality of processed feature matrices are obtained, the plurality of processed latent feature matrices are fused, and the fourth latent feature matrix (probability matrix) is obtained through the activation function, another multi-layer perception module and the Softmax function. Then the third latent feature matrix and the fourth latent feature matrix are input into the first multi-layer perception module, fused by using the fusion formula, and the fused feature matrix is obtained, and the predicted landslide risk probability is obtained based on the fused feature matrix and the fully connected layer in the first multi-layer perception module. Wherein, the feature fusion formula is as follows: ; In the above formula, is the fused feature matrix, is the weight matrix, is the third latent feature matrix, is the fourth latent feature matrix.

[0053] Figure 5 is a feature fusion effect diagram provided by an embodiment of the present application, in which Figure 5 the length of is N, and the feature dimension is u, Similarly, the weight of is set to , the weight of is set to , and the fused feature is obtained by adding the respective weights multiplied by and

[0054] After the fused feature is linearly fed back, the predicted landslide risk probability is obtained.

[0054] In some embodiments of the present application, S102 can be implemented by S1021 to S1022, which is described by the following steps.

[0055] S1021, input the node data into the second multi-layer perception module to obtain the linearly transformed latent feature matrix, and input the linearly transformed latent feature matrix into the layer normalization layer to obtain the processed latent feature matrix.

[0056] S1022, the processed latent feature matrix is processed by the activation function to obtain the non-linearly transformed latent feature matrix, and the non-linearly transformed latent feature matrix and the node data are fused through the residual connection to obtain the first latent feature matrix.

[0057] In some embodiments of the present application, the residual feature conversion module comprises a second multi-layer perception module and a layer normalization layer, the node data is input into the second multi-layer perception module for linear transformation, a linearly transformed latent feature matrix is obtained, then the linearly transformed latent feature matrix is input into the layer normalization layer for normalization processing, a processed latent feature matrix is obtained, the processed latent feature matrix is processed through an activation function, a nonlinearly transformed latent feature matrix is obtained, and the nonlinearly transformed latent feature matrix and the node data are fused through a residual connection to obtain a first latent feature matrix.

[0058] In the embodiment of the present application, when the landslide risk prediction model is trained, the training data is training data containing labeled samples and unlabeled samples, the total loss is composed of two parts: a standard supervised loss (such as cross-entropy) is used on the labeled samples to ensure that the model fits the true label; a consistency regularization term is introduced on the unlabeled samples to constrain the model to maintain the stability of the prediction output on input perturbations (such as data augmentation, noise injection or Dropout). This regularization mechanism implicitly uses the data manifold assumption, that is, similar or adjacent samples should have similar prediction distributions, thereby improving the model generalization ability. The parameters of the landslide risk prediction model are optimized through the total loss until the trained landslide risk prediction model is obtained.

[0059] Referring to Figure 6 , a structural schematic diagram of an electronic device according to an embodiment of the present application is shown, and the specific implementation of the electronic device is not limited in the specific embodiments of the present application.

[0060] As Figure 6 shown, the electronic device can include a processor 502, a communications interface 504, a memory 506, and a communications bus 508.

[0061] Among them: The processor 502, the communications interface 504, and the memory 506 complete mutual communication through the communications bus 508.

[0062] The communications interface 504 is configured to communicate with other electronic devices or servers.

[0063] The processor 502 is configured to execute the program 510, and specifically can execute the related steps in the above method embodiments.

[0064] Specifically, the program 510 can include program code, and the program code includes computer operation instructions.

[0065] The processor 502 can be a central processing unit (CPU), or an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present application. The one or more processors included in the smart device can be of the same type or different types, such as one or more CPUs and one or more ASICs.

[0066] The memory 506 is configured to store a program 510. The memory 506 can include a high-speed RAM memory, and can further include a non-volatile memory, such as at least one disk memory.

[0067] The program 510 can be specifically configured to enable the processor 502 to perform operations corresponding to the methods described in the above method embodiments.

[0068] The specific implementation of each step in the program 510 can refer to the corresponding description in the above method embodiments, and will not be described herein. It can be clearly understood by those skilled in the art that, for the convenience and brevity of description, the specific working processes of the devices and modules described above can refer to the corresponding process descriptions in the above method embodiments, and will not be described herein.

[0069] It should be noted that, according to the needs of implementation, each component / step described in the embodiments of the present application can be split into more components / steps, or two or more components / steps or part of the operations of the components / steps can be combined into a new component / step, to achieve the purpose of the embodiments of the present application.

[0070] The above method according to the embodiments of the present application can be implemented in hardware, firmware, or as software or computer code that can be stored in a recording medium such as a CD ROM, a RAM, a floppy disk, a hard disk or an optical disk, or downloaded through a network and originally stored in a remote recording medium or a non-transitory machine readable medium and then stored in a local recording medium, so that the method described herein can be processed by such software on a recording medium using a general purpose computer, a special purpose processor or programmable or special purpose hardware such as an ASIC or an FPGA. It can be understood that the computer, the processor, the microprocessor controller or the programmable hardware includes a storage component (for example, a RAM, a ROM, a flash memory, etc.) that can store or receive software or computer code, when the software or computer code is accessed and executed by the computer, the processor or the hardware, the method described herein is implemented. In addition, when a general purpose computer accesses the code for implementing the method shown herein, the execution of the code will convert the general purpose computer into a special purpose computer for executing the method shown herein.

[0071] Those skilled in the art can understand that the units and method steps of the examples described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software manner depends on the specific application and design constraints of the technical solution. The skilled person can use different methods to realize the described functions for each specific application, but such implementation should not be considered as beyond the scope of the embodiments of the present application.

[0072] The above embodiments are only used to illustrate but not to limit the embodiments of the present application. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the embodiments of the present application. Therefore, all equivalent technical solutions belong to the scope of the embodiments of the present application, and the patent protection scope of the embodiments of the present application should be defined by the claims.

Claims

1. A method for predicting the susceptibility of landslides over a wide area, characterized in that, include: The geographical environment of the target area is divided into multiple slope units; and a directed connection graph of the target area is constructed using the multiple slope units. The node data of the directed connection graph is the multiple slope units, and the edge data of the directed connection graph is the influence and dependency relationship between the multiple slope units. The node data is input into the residual feature transformation module in the landslide risk prediction model to obtain the first latent feature matrix; the landslide risk prediction model also includes two identical linear self-attention Transformer modules, multiple multilayer perceptron modules, and a graph filter layer; The first latent feature matrix is ​​input into the first linear self-attention Transformer module to obtain the query matrix, key matrix, and value matrix; and a diagonal matrix is ​​obtained based on the normalized query matrix and the normalized key matrix. The second latent feature matrix is ​​obtained based on the diagonal matrix, value matrix, normalized query matrix, and normalized key matrix. And based on the second latent feature matrix and the second linear self-attention Transformer module, the third latent feature matrix is ​​obtained; The directed connection graph is input into the graph filter layer for normalization and exponential iteration to obtain a graph filter operator. The graph filter operator is then used to perform hierarchical neighbor aggregation on the node data to obtain an aggregated feature matrix. The fused feature matrix is ​​obtained based on the aggregated feature matrix, the third latent feature matrix, and the first multilayer perceptron module, and the predicted landslide risk probability is obtained based on the fused feature matrix.

2. The method according to claim 1, characterized in that, The process of obtaining a diagonal matrix based on the normalized query matrix and the normalized key matrix includes: The diagonal matrix is ​​obtained based on the normalized query matrix, the normalized key matrix, the all-one unit column matrix, and the formula for the diagonal matrix, as follows: ； In the above formula, It is a diagonal matrix. The normalized query matrix, The normalized key matrix, It is an all-one unit column matrix, where N is the dimension and T is the transpose.

3. The method according to claim 2, characterized in that, The process of obtaining the second latent feature matrix based on the diagonal matrix, value matrix, normalized query matrix, and normalized key matrix includes: The second latent feature matrix is ​​obtained based on the diagonal matrix, value matrix, normalized query matrix, normalized key matrix, first latent feature matrix, and latent feature matrix formula. The latent feature matrix formula is as follows: ； In the above formula, Z is the second latent feature matrix. for The inverse matrix of V, where V is the value matrix. For the hyperparameters of the residual connection, This is the first latent feature matrix.

4. The method according to claim 1, characterized in that, The process of using the graph filtering operator to perform hierarchical neighbor aggregation on the node data to obtain an aggregated feature matrix includes: The aggregated feature matrix is ​​obtained based on graph filtering operators, node data, and feature aggregation formulas. The feature aggregation formulas are as follows: ； In the above formula, X represents the node data. Here, r is the graph filtering operator, and r is the maximum number of neighbor hops allowed to be considered. The graph filtering operator is as follows: ； In the above formula, A is the adjacency matrix and D is the degree matrix.

5. The method according to claim 1, characterized in that, The step of obtaining a fused feature matrix based on the aggregated feature matrix, the third latent feature matrix, and the first multilayer perceptron module, and obtaining the predicted landslide risk probability based on the fused feature matrix, includes: The fourth latent feature matrix is ​​obtained based on the aggregated feature matrix and the third multilayer perceptron module. The fused feature matrix is ​​then obtained based on the third latent feature matrix, the fourth latent feature matrix, and the feature fusion formula in the first multilayer perceptron module. The feature fusion formula is as follows: ； The above formula, The fused feature matrix, This is the weight matrix. This is the third latent feature matrix. This is the fourth latent feature matrix.

6. The method according to claim 1, characterized in that, The residual feature transformation module includes a second multilayer perceptron module and a layer normalization layer; The step of inputting node data into the residual feature transformation module of the landslide risk prediction model to obtain the first latent feature matrix includes: The node data is input into the second multilayer perceptron module to obtain the linearly transformed latent feature matrix, and the linearly transformed latent feature matrix is ​​input into the layer normalization layer to obtain the processed latent feature matrix. The processed latent feature matrix is ​​processed by an activation function to obtain a nonlinearly transformed latent feature matrix. The nonlinearly transformed latent feature matrix and the node data are then fused through residual connection to obtain the first latent feature matrix.

Images