A method for calculating spatial regional weight of landslide surface deformation

Abstract

The present invention provides a method for calculating the spatial regional weight of landslide surface deformation, and belongs to the field of engineering geological technology. The method first obtains deformation data of each monitoring point from a landslide monitoring system, processes the deformation data of the monitoring point into time series data of different channels according to its spatial distribution, and inputs them into a convolutional neural network (CNN) model for processing; the data processed by the convolutional neural network (CNN) model is used as input data, and input into an improved convolution attention module (ICBAM) for calculation, and finally outputs the displacement prediction value of each monitoring point of the landslide and the weight of the area where each monitoring point is located in the landslide surface deformation. The present invention calculates the weight of the area where the monitoring point is located in the landslide displacement prediction based on the monitoring data, judges the influence of the area on the landslide displacement prediction based on the weight, and finally can effectively predict the landslide displacement.

Description

Technical Field

[0001] The invention belongs to the field of engineering geology research, and in particular relates to a method for calculating the spatial regional weight of landslide surface deformation. Background Art

[0002] Landslides are one of the most common natural geomorphological processes, causing massive casualties and property losses around the world each year. Landslide prevention and control in my country still faces enormous challenges.

[0003] Over the past few decades, with the development and introduction of various sensors for landslide monitoring, numerous landslide monitoring systems have been established worldwide. These highly accurate and environmentally adaptable monitoring devices provide an essential foundation for analyzing landslide deformation characteristics and mechanisms. Currently, most landslide deformation research methods rely solely on single-point modeling, resulting in uncertainty in early landslide warning. Examples include artificial neural networks (ANNs), support vector machines (SVMs), extreme learning machines (ELMs), and the inverse velocity method. These methods generally offer good results. However, landslide deformation exhibits spatial significance, particularly for large landslides, where deformation characteristics at monitoring points vary with spatial location. Therefore, displacement predictions based on single-point modeling fail to reveal the true extent of large-scale landslide deformation. Limited research has considered the spatial correlation between different points on a slope, such as the seemingly uncorrelated model (SUR). However, this model only considers the spatial correlation between multiple monitoring points and ignores the contribution of the spatial region within each monitoring point to the overall landslide deformation, resulting in significant deviations in landslide displacement predictions.

[0004] In summary, while many models have been developed for landslide deformation research, most of these models ignore the overall correlation of landslides or the impact of the weights of various spatial regions on landslide displacement. To address this technical issue, the present invention provides a method for calculating the spatial regional weights of landslide surface deformation. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for calculating the spatial area weight of landslide surface deformation, which solves the problems of "uncertainty caused by relying on single-point modeling to predict landslide displacement" and "ignoring the weight of spatial area in landslide deformation causing inaccuracy in landslide displacement prediction". The weight of each spatial area in the landslide surface deformation is calculated through the monitoring data (displacement, velocity, acceleration) of the monitoring point.

[0006] To achieve the above object, the technical solution adopted by the present invention is:

[0007] A method for calculating the spatial regional weight of landslide surface deformation, characterized by comprising the following steps:

[0008] S1. Acquire monitoring data of each monitoring point from the landslide monitoring system, and obtain various deformation data related to landslide deformation characteristics based on the monitoring data;

[0009] S2. Processing the various deformation data of the monitoring points into time series data of different channels according to the spatial distribution of the monitoring points and inputting them into the convolutional neural network (CNN) model for processing, where each channel corresponds to one type of deformation data;

[0010] S3, taking the data processed by the convolutional neural network (CNN) model in step S2 as input data and inputting it into an improved convolutional attention module (ICBAM) for screening, wherein the ICBAM includes a channel attention module, a spatial attention module, and a residual block; the input of the ICBAM is used to determine the importance of different deformation features through the channel attention module, and the output of the channel attention module is used as the input of the spatial attention module to determine the importance of the deformation data of the spatial positions of different monitoring points for the final landslide prediction result; the residual block links the input of the ICBAM and the output of the spatial attention module;

[0011] S4. Finally, a fully connected layer outputs the spatial correlation of the deformation data of each monitoring point of the landslide, and at the same time outputs the spatial attention weight of the area where each monitoring point is located in the landslide surface deformation.

[0012] In step S2, the multiple deformation data of the monitoring points are processed into time series data of different channels according to the spatial distribution of the monitoring points and input into a convolutional neural network (CNN) model for processing, which includes the following steps:

[0013] S21. Input data

[0014] The deformation data of the monitoring points are processed into multi-channel images according to their spatial distribution. Different types of deformation data correspond to different channels and are input into the convolutional neural network (CNN) model.

[0015] S22. Mathematical operations

[0016] The convolutional neural network (CNN) model includes convolution and pooling operations, and extracts spatial feature maps by merging deformation information of different monitoring points. A convolutional neural network (CNN) model includes convolutional layers, pooling layers, and fully connected layers. In the convolutional layer, the output feature matrix M is generated by moving the filter k on the input feature F. The pooling operation is usually performed after the convolution layer. The adjacent elements of M are integrated by maximum pooling, average pooling, or other pooling operations. The convolution equation is as follows:

[0017]

[0018] Among them, m is an element in the matrix M, ∑ represents the sum of the matrix elements, c is the number of channels of the input feature F, and f i,j is a submatrix of F, with the same size as k, i, j are the row and column moving steps of filter k respectively; when the submatrix f i,j The filter k is applied repeatedly while moving across the input feature map.

[0019] The input of the improved convolutional attention module ICBAM in step S3 is the spatial feature map X output by the convolutional neural network (CNN) model, X∈R C×H×W , where C, H, and W correspond to the sizes of the ICBAM input data in three different dimensions. ICBAM infers the channel feature map X according to formulas (2)-(4) in turn. C ∈R C×H×W , spatial feature map X S ∈R C×H×W And the final refined feature map X'∈R after the introduction of the residual block C×H×W , during the attention process, the size of the original features does not change:

[0020] S31, channel attention module

[0021] The channel attention module in step S3 contains a fully connected layer, a sigmoid activation function and four different pooling operations: maximum pooling, mean pooling, median pooling and combined pooling; the output of step S2 is used as the input X∈R of the channel attention module C×H×W , input X and perform four pooling operations at the same time to obtain data A, M, D and C respectively. Then these four groups of data are processed by the fully connected layer and activation function with shared weights to obtain the channel attention weight A c ;The input of the channel attention module and A c Multiply to get the output X of the channel attention module c ;

[0022] S32, spatial attention module

[0023] The spatial attention module in step S3 includes a standard 7×7 convolutional layer, an activation function and two pooling layers, namely: mean pooling and maximum pooling; the input of the spatial attention module is the output X of the channel attention module. c , X c After two pooling layers, the spatial attention weights As, As and X are obtained after 7×7 convolution layer and activation function processing. c Multiply to get the output X of the spatial attention module s ;

[0024] S33, residual block

[0025] The residual block in step 3 contains a standard 1×1 convolution layer and a ReLU activation function. The input of the residual block is the output X of step S2 and the output X of the spatial attention module. s , X is processed by 1×1 convolution layer and X s The final feature map X' is obtained after adding and processing by the ReLU activation function, which is the output of IBCAM.

[0026] The specific calculation method of the improved convolutional attention module ICBAM is as follows:

[0027]

[0028]

[0029] X'=ReLU(conv 1×1 (X)+X s ) (4)

[0030] in Represents element-by-element multiplication, ReLU is the activation function, conv 1×1 is the convolution layer, the convolution kernel size is 1×1, Xc is the channel attention result, X s is the spatial attention result, X' is the final refined feature map of ICBAM, Ac is the channel attention weight, and As is the spatial attention weight;

[0031] Channel attention weight A in the channel attention module c The calculation process is:

[0032] A=AvgPool(X) (5)

[0033] M=MaxPool(X) (6)

[0034] D=MedPool(X) (7)

[0035]

[0036] Then, A, M, D, and C pass through a shared weight network MLP consisting of a fully connected layer and an activation function, and the output vectors are summed element-wise to obtain the final channel attention weight:

[0037] A c =σ(MLP(A)+MLP(M)+MLP(D)+MLP(C)) (9)

[0038] Where σ is the Sigmoid activation function, A c ∈R C×1×1The number of channels C of Ac is the same as that of the original feature map X, and each position of Ac is assigned a weight corresponding to the channel of the original feature map;

[0039] Channel feature map X c ∈R C×H×W As the input of the spatial attention module to calculate the spatial attention weight, the calculation method is as follows:

[0040] A s =σ(f 7×7 ([AvgPool(X c ):MaxPool(X c )])) (10)

[0041] Where σ is the Sigmod activation function, f 7×7 It is a convolution layer with a convolution kernel size of 7×7.

[0042] In step S4, the larger the weight calculation result of the area where the landslide monitoring point to be studied is located, the greater the influence of the deformation data of the area where the monitoring point is located on the landslide displacement prediction result; the smaller the weight calculation result of the area where the monitoring point is located, the smaller the influence of the deformation data of the area where the monitoring point is located on the landslide displacement prediction result.

[0043] The deformation data includes displacement, velocity, acceleration, displacement difference between adjacent monitoring points, velocity difference between adjacent monitoring points, average displacement between adjacent monitoring points, etc., and is selected based on the actual prediction effect. At least three types of deformation data are input.

[0044] The spatial area weight of each monitoring point ranges from 0 to 1, and the sum of the spatial area weights of all points is 1. The larger the spatial area weight, the greater the relative impact.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] (1) The present invention combines a landslide monitoring system, a convolutional neural network (CNN) model, and an ICBAM. The time series landslide deformation data are processed into time series data of different channels (corresponding to the above channels) according to their spatial distribution, and then input into the CNN model for landslide displacement prediction. The weight of the area where a specific monitoring point is located can be calculated based on the deformation data, and then the influence of the area on the landslide displacement prediction can be judged based on the weight. The greater the weight, the more likely the area is to have a landslide, and the greater the influence on the landslide displacement prediction. The spatial attention weight of each monitoring point ranges from 0 to 1, and the sum of the spatial attention weights of all points is 1. The larger the spatial attention weight, the greater the relative influence.

[0047] (2) The present invention collects the time series deformation data (displacement, velocity, acceleration, etc.) of each monitoring point at the same sampling time and sampling interval, and inputs the deformation data into the CNN model according to the spatial distribution of the landslide monitoring points, thereby maintaining the spatial distribution.

[0048] (3) The proposed method is used for large-scale landslide displacement prediction. The spatial attention mechanism in the CNN+ICBAM model is used to quantify the attention weights of spatial regions and reflect the spatial characteristics of landslide deformation, which makes the prediction model physically explainable. A large landslide with step deformation characteristics in the Three Gorges Reservoir area of ​​my country is used as an example to verify the feasibility and effectiveness of the proposed model.

[0049] (4) The present invention realizes multi-point landslide displacement prediction by considering the spatial deformation relationship of different locations of the landslide, overcoming the technical problem of uncertainty caused by single-point modeling of the traditional model, taking into account the spatial correlation of landslide deformation. Due to the introduction of the attention mechanism, the deformation correlation of different regions can be accurately captured, which increases the rationality and interpretability of landslide deformation prediction, and can calculate the weight of each spatial region in the landslide surface deformation, and determine the degree of influence of the monitoring point area on the landslide displacement prediction. The most important significance is that it can help people prevent landslides. The present invention aims to standardize the impact of each landslide area on landslide displacement prediction, improve the effectiveness and rationality of landslide early warning and prevention, and play a certain guiding role in landslide disaster prevention and mitigation. It solves the technical problems that the single-point modeling method for landslide displacement prediction cannot consider the spatial correlation of deformation between monitoring points, and the contribution of deformation data of different regions to the final prediction result to the uncertainty of landslide displacement prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0051] Figure 1 The figure is a flow chart of a method for calculating the spatial regional weight of landslide surface deformation according to the present invention.

[0052] Figure 2 is the presentation form of data in the CNN model, where (a) is the data representation of a three-channel digital image in the CNN model, and the data in each channel is the corresponding value of the ABC channel of the pixel point; (b) is the data representation of the spatial deformation data of the monitoring point in the CNN model, where the data in each channel is a kind of deformation data of the corresponding monitoring point.

[0053] Figure 3 Schematic diagram of the ICBAM structure.

[0054] Figure 4 Paulownia Bay landslide monitoring network. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. Of course, the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0056] The method for calculating the spatial regional weight of landslide surface deformation of the present invention comprises the following steps:

[0057] S1. First, monitoring data of each monitoring point is obtained from the landslide monitoring system. The monitoring data is processed to obtain various deformation data related to the deformation characteristics of the landslide. The deformation data is time series data, and the deformation data includes displacement, velocity, acceleration, etc., which can be determined according to the actual situation.

[0058] S2. The deformation data of the monitoring points are processed into time series data of different channels according to their spatial distribution and input into a multi-channel convolutional neural network (CNN) model for processing, with each type of deformation data corresponding to one channel;

[0059] S3. The data processed by the multi-channel convolutional neural network (CNN) model in step S2 is used as input data and input into an improved convolutional attention module (ICBAM) for screening, wherein the ICBAM includes a channel attention module, a spatial attention module and a residual block; the input of the ICBAM is used to determine the importance of different deformation features through the channel attention module, and the output of the channel attention module is used as the input of the spatial attention module to determine the importance of the deformation data of the spatial positions of different monitoring points for the final landslide prediction result. The residual block links the input of the ICBAM and the output of the spatial attention module to avoid the degradation of the learning ability of the deep network model.

[0060] S4. Finally, a fully connected layer outputs the spatial correlation of the deformation data of each monitoring point of the landslide, and at the same time outputs the spatial attention weight of the area where each monitoring point is located in the landslide surface deformation.

[0061] In step S2, inputting the monitoring data of the monitoring point into a multi-channel convolutional neural network (CNN) model for processing includes the following steps:

[0062] S21. Input data

[0063] Since the convolutional neural network is a forward-moving neural network that processes data in Euclidean space and is used for image recognition, local feature extraction, and time series prediction, the deformation data of the monitoring points containing spatial distribution is input into a multi-channel convolutional neural network (CNN) model of different digital images. For the landslide monitoring system, the deformation data is recorded by the monitoring points, and the monitoring points correspond to the positions of the pixel points. Different types of deformation data correspond to different channels. The deformation data of the monitoring points are used as input data of the convolutional neural network (CNN) model and processed into a multi-channel image. In this embodiment, there are three types of deformation data (displacement, velocity, and acceleration), and a three-channel image is formed. This is not a limitation to the method of the present invention, and other deformation data may also be included, such as: the displacement difference between adjacent monitoring points, the velocity difference between adjacent monitoring points, the average displacement between adjacent monitoring points, etc.

[0064] S22. Mathematical operations

[0065] The convolutional neural network (CNN) model includes convolution and pooling operations, and extracts spatial feature maps by merging deformation information of different monitoring points. A convolutional neural network (CNN) model includes convolutional layers, pooling layers, and fully connected layers. In the convolutional layer, the output feature matrix M is generated by moving the filter k on the input feature F. The pooling operation is usually performed after the convolution layer. The adjacent elements of M are integrated by maximum pooling, average pooling, or other pooling operations. The convolution equation is as follows:

[0066]

[0067] Where m is an element in the matrix M, ∑ represents the sum of the matrix elements, c is the number of channels in F, and f i,j It is a submatrix of F with the same size as k, i and j are the row and column moving steps of filter k respectively; when the submatrix f moves on the input feature map, filter k is repeatedly applied.

[0068] The improved convolutional attention module ICBAM in step S3 includes a channel attention module, a spatial attention module and a residual block:

[0069] S31, channel attention module

[0070] The channel attention module in step S3 contains a fully connected layer, a sigmoid activation function and four different pooling operations: maximum pooling, mean pooling, median pooling and combined pooling. The output of step S2 is used as the input X∈R of the channel attention module. C×H×W, where C, H, and W correspond to the sizes of the ICBAM input data in three different dimensions. Input X performs four pooling operations simultaneously to obtain data A, M, D, and C respectively. These four sets of data are then processed by the weight-sharing fully connected layer and activation function to obtain the channel attention weight A c The input of the channel attention module is the same as A c Multiply to get the output X of the channel attention module c .

[0071] S32, spatial attention module

[0072] The spatial attention module in step S3 contains a standard 7×7 convolutional layer, an activation function and two pooling layers. The input of the spatial module is the output X of the channel attention module. c , X c After passing through two pooling layers, the 7×7 convolution layer and activation function are processed with X c Multiply to get the output X of the spatial attention module s .

[0073] S33, residual block

[0074] The residual block in step 3 contains a standard 1×1 convolution layer and a ReLU activation function. The input of the residual block is the output X of step S2 and the output of the spatial attention module. After X is processed by the 1×1 convolution layer, it is combined with X s The output of IBCAM is obtained by adding them together and processing them with ReLU activation function.

[0075] The input data in step S3 is the output X of S2, and ICBAM infers the channel feature map X in turn. C ∈R C×H×W , spatial feature map X S ∈R C×H×W And the final feature map X'∈R after the introduction of the residual block C×H×W ( Figure 3 ), during the attention processing, the size of the original feature does not change. The specific ICBAM calculation method is as follows:

[0076]

[0077]

[0078] X'=ReLU(conv 1×1 (X)+X s ) (4)

[0079] in Represents element-by-element multiplication, ReLU is the activation function, conv 1×1is the convolution layer, the convolution kernel size is 1×1, Xc is the channel attention result, X s is the spatial attention result, X' is the final refined feature map of ICBAM, Ac is the channel attention weight, and As is the spatial attention weight;

[0080] Channel attention weight A in the channel attention module c The calculation method is:

[0081] A=AvgPool(X) (5)

[0082] M=MaxPool(X) (6)

[0083] D=MedPool(X) (7)

[0084]

[0085] Then, A, M, D and C pass through a shared weight network MLP consisting of a fully connected layer and an activation function, and the output vectors are summed element-wise to obtain the final channel weight, which is given by:

[0086] A c =σ(MLP(A)+MLP(M)+MLP(D)+MLP(C)) (9)

[0087] Where A c is the channel attention weight, σ is the Sigmoid activation function, A c ∈R C×1×1 The number of channels C of Ac is the same as that of the original feature map X, and each position of Ac is assigned a weight corresponding to the channel of the original feature map;

[0088] Channel feature map X c ∈R C×H×W As the input of the spatial attention module to generate spatial attention weights, the calculation method is as follows:

[0089] A s =σ(f 7×7 ([AvgPool(X c ):MaxPool(X c )])) (10)

[0090] Where σ is the Sigmod activation function, f 7×7 It is a convolution layer with a convolution kernel size of 7×7.

[0091] In step S4, the larger the weight calculation result of the area where the specific monitoring point is located, the greater the influence of the deformation data of the area where the specific monitoring point is located on the landslide displacement prediction result; the smaller the weight calculation result of the area where the specific monitoring point is located, the smaller the influence of the deformation data of the area where the specific monitoring point is located on the landslide displacement prediction result.

[0092] Example

[0093] A landslide in the Three Gorges Reservoir area of ​​China, the Paotongwan landslide, was selected as a case study. The Paotongwan landslide is located in Wushan County, Chongqing, east of the Daxi River, a tributary of the Yangtze River. The Paotongwan landslide is 310 m long and 350 m wide, respectively. Its total area is 10.85 million m², and its volume is 3.25 million m³. The landslide mass is primarily composed of silty clay and quartz sandstone gravels, with a thickness of 20 to 40 m and an average surface slope of 25°. The gravels account for approximately 66–71% of the total volume and are primarily sized between 10 and 50 cm. The landslide slides primarily at a 240° angle toward the Daxi River. The bedrock is the Middle Triassic Badong Formation (T2b), composed primarily of interbedded siltstone and sandy mudstone.

[0094] (1) Extract monitoring data from the Paotong Bay landslide area for analysis

[0095] The deformation of the Paotongwan landslide began in the rainy season of 1998, and a GPS monitoring system was implemented in September 2006 to measure the dynamic evolution of the landslide. Figure 4 As shown, six GPS monitoring stations are located on the sliding mass. GPS monitoring points in different areas of the landslide provide detailed deformation data. Specifically, cumulative displacement data measured by GPS stations around section Ⅰ-Ⅰ' in the southern part of the landslide show the greatest deformation, with an average displacement of 251.5 mm at two points (WS01 and WS02). WS03 and WS04 exhibit similar cumulative displacements in the central part of the landslide, with cumulative displacements of approximately 200 mm and an average rate of approximately 2.38 mm / month. Recorded data indicate that deformation is minimal in the northern part of the landslide, with two points (WS05 and WS06) exhibiting cumulative displacements of less than 170 mm. Total displacements at the monitoring points ranged from 157 to 280 mm, with an average rate of 1.92 to 3.43 mm / month. During the monitoring period, from May to September, the landslide also exhibited alternating periods of slow and accelerated deformation. Specifically, the acceleration changes at WS01 and WS02 were essentially identical, with the acceleration gradually increasing from zero to 8.86 mm / s from May to July. 2 From July to September, the acceleration was 8.86 mm / s 2 Reduced to -6.22 mm / s 2 For WS03 and WS04, the acceleration changes are large, with an average acceleration change of 20.75 mm / s 2, while WS05 and WS06 have an average acceleration of 7 mm / s 2 . Figure 4 The deformation data of the six monitoring points WS01-WS06 were obtained, including the cumulative displacement, velocity (monthly displacement), and acceleration of each point.

[0096] (2) Model input

[0097] The input spatial data is converted into a deformation signature matrix X based on the spatial distribution of monitoring points. Each element in the matrix represents a monitoring point. For the Paulownia Bay landslide, the spatial input is a three-row, two-column matrix representing the spatial distribution and storing the spatial deformation signature of the entire monitoring system. Each element of X records a time series of three types of deformation data: cumulative displacement, velocity, and acceleration.

[0098] The model input described in this application refers to the input of the entire CNN+ICBAM model, which is composed of a convolutional neural network (CNN) model, an improved convolutional attention module (ICBAM), and a fully connected layer in series. The output of the last fully connected layer and the spatial attention weights of the spatial attention module constitute the output of the entire CNN+ICBAM model. The CNN+ICBAM model is a dual-output model. Model training is performed using monitoring point data from the same landslide at different times. The trained CNN+ICBAM model is used to calculate the spatial area weights of the actual landslide surface deformation.

[0099] The three deformation data of the monitoring points are processed into time series data of different channels according to the spatial distribution of the monitoring points and input into the convolutional neural network (CNN) model for processing, with each channel corresponding to one deformation data;

[0100] The data processed by the convolutional neural network (CNN) model is used as input data and input into the improved convolutional attention module (ICBAM) for screening. The ICBAM consists of a channel attention module, a spatial attention module, and a residual block. The input of the ICBAM is used to determine the importance of different deformation features through the channel attention module. The output of the channel attention module is used as the input of the spatial attention module to determine the importance of deformation data at different spatial locations of monitoring points for the final landslide prediction results. The residual block links the input of the ICBAM and the output of the spatial attention module.

[0101] (3) Output the weight distribution of landslide monitoring points to verify the accuracy of the model

[0102] Using the landslide test sample, the spatial attention module of the improved convolutional attention module ICBAM is calculated (A S ) Extract the spatial attention weight. As shown in Table 5, the weight value A SThe deformation levels at different locations of the landslide are revealed. For the Paulownia Bay landslide: the minimum weight is 0.11 at the WS01 monitoring point and the maximum weight is 0.26 at the WS02 monitoring point, indicating that the weight values ​​of all sites are not very different. On the other hand, the average maximum displacement of the WS01 and WS02 points in the actual deformation of the landslide is 251.5mm; the cumulative displacement of WS03 and WS04 in the middle of the landslide is about 200mm; the minimum cumulative displacement values ​​of the two points WS05 and WS06 in the north of the landslide are both less than 170mm. It can be seen from this that the weight of the landslide A S The distribution is in good agreement with the relative magnitude of the actual deformation displacement at different monitoring points, indicating that the CNN+ICBAM model can accurately calculate the weight values ​​of the areas where different monitoring points are located in the landslide surface deformation.

[0103] Table 1 Spatial attention weight distribution of the Paotongwan landslide.

[0104] Monitoring points WS01 WS02 WS03 WS04 WS05 WS06 Weight value 0.11 0.26 0.15 0.195 0.18 0.13

[0105] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for calculating the spatial regional weight of landslide surface deformation, characterized in that: The following steps are involved: S1. Acquire monitoring data of each monitoring point from the landslide monitoring system, and obtain various deformation data related to the deformation characteristics of the landslide according to the monitoring data; S2, processing the various deformation data of the monitoring points into time series data of different channels according to the spatial distribution of the monitoring points and inputting them into the convolutional neural network (CNN) model for processing, where each channel corresponds to one deformation data; S3, taking the data processed by the convolutional neural network (CNN) model in step S2 as input data, and inputting it into the improved convolutional attention module ICBAM for screening, wherein ICBAM includes a channel attention module, a spatial attention module and a residual block; The input of ICBAM is used to determine the importance of different deformation features through the channel attention module. The output of the channel attention module is used as the input of the spatial attention module to determine the importance of deformation data at different monitoring points for the final landslide prediction results. The residual block links the input of ICBAM and the output of the spatial attention module. S4. Finally, a fully connected layer is used to output the spatial correlation of the deformation data of each monitoring point of the landslide, and at the same time, the spatial attention weight of the area where each monitoring point is located in the landslide surface deformation is output.

2. The method for calculating the spatial regional weight of landslide surface deformation according to claim 1, characterized in that: In step S2, the multiple deformation data of the monitoring points are processed into time series data of different channels according to the spatial distribution of the monitoring points and input into a convolutional neural network (CNN) model for processing, which includes the following steps: S21. Input data The deformation data of the monitoring points are processed into multi-channel images according to the spatial distribution of the monitoring points. Different types of deformation data correspond to different channels and are input into the convolutional neural network (CNN) model. S22. Mathematical operations The convolutional neural network (CNN) model contains convolution and pooling operations. The spatial feature map is extracted by merging the deformation information of different monitoring points. A convolutional neural network (CNN) model includes convolutional layers, pooling layers, and fully connected layers. In the convolutional layer, the output feature matrix M is generated by moving the filter k on the input feature F. The pooling operation is usually performed after the convolutional layer. The adjacent elements of M are integrated by maximum pooling, average pooling, or other pooling operations. The convolution equation is as follows: Where m is an element in the matrix M, ∑ represents the sum of the matrix elements, c is the number of channels of the input feature F, and f i,j is a submatrix of F, with the same size as k, i and j are the row and column moving steps of filter k respectively; when the submatrix f i,j The filter k is applied repeatedly while moving over the input feature map.

3. The method for calculating the spatial regional weight of landslide surface deformation according to claim 1, characterized in that: The input of the improved convolutional attention module ICBAM in step S3 is the spatial feature map X output by the convolutional neural network (CNN) model, X∈R C×H×W , where C, H, and W correspond to the sizes of the ICBAM input data in three different dimensions. ICBAM infers the channel feature map X according to formulas (2)-(4) in turn. C ∈R C×H×W , spatial feature map X S ∈R C×H×W And the final refined feature map X'∈R after the introduction of the residual block C×H×W， During the attention process, the size of the original features does not change: S31, channel attention module The channel attention module in step S3 includes a fully connected layer, a sigmoid activation function and four different pooling operations: maximum pooling, mean pooling, median pooling and combined pooling; the output of step S2 is used as the input X∈R of the channel attention module C×H×W , input X and perform four pooling operations at the same time to obtain data A, M, D and C respectively. Then these four sets of data are processed by the fully connected layer and activation function with shared weights to obtain the channel attention weight A c ; The input of the channel attention module and A c Multiply to get the output X of the channel attention module c ; S32, spatial attention module The spatial attention module in step S3 includes a standard 7×7 convolutional layer, an activation function and two pooling layers, namely: mean pooling and maximum pooling; the input of the spatial attention module is the output X of the channel attention module c , X c After two pooling layers, the spatial attention weights As, As and X are obtained after 7×7 convolution layer and activation function processing. c Multiply to get the output X of the spatial attention module s ; S33, residual block The residual block in step 3 contains a standard 1×1 convolution layer and a ReLU activation function. The input of the residual block is the output X of step S2 and the output X of the spatial attention module. s , X is processed by a 1×1 convolution layer and X s After adding and processing with ReLU activation function, the final feature map X' is obtained, which is the output of IBCAM; The specific calculation method of the improved convolutional attention module ICBAM is as follows: X'=ReLU(conv 1×1 (X)+X s ) (4) in Represents element-by-element multiplication, ReLU is the activation function, conv 1×1 is a convolutional layer with a kernel size of 1×1, Xc is the channel attention result, and X s is the spatial attention result, X' is the final refined feature map of ICBAM, Ac is the channel attention weight, and As is the spatial attention weight; Channel attention weight A in the channel attention module c The calculation process is: A=AvgPool(X) (5) M=MaxPool(X) (6) D=MedPool(X) (7) Then, A, M, D, and C pass through a shared weight network MLP consisting of a fully connected layer and an activation function, and the output vectors are summed element-wise to obtain the final channel attention weights: A c =σ(MLP(A)+MLP(M)+MLP(D)+MLP(C)) (9) In the formula, σ is the Sigmoid activation function, A c ∈R C×1×1 The number of channels C of the original feature map is the same as the number of channels of the original feature map X, and each position of Ac is assigned a weight corresponding to the channel of the original feature map; Channel feature map X c ∈R C×H×W As the input of the spatial attention module, the spatial attention weight is calculated as follows: A s =σ(f 7×7 ([AvgPool(X c ):MaxPool(X c )])) (10) Where σ is the Sigmod activation function, f 7×7 It is a convolution layer with a convolution kernel size of 7×7.

4. The method for calculating the spatial regional weight of landslide surface deformation according to claim 1, characterized in that: In step S4, the larger the weight calculation result of the area where the landslide monitoring point to be studied is located, the greater the influence of the deformation data of the area where the monitoring point is located on the landslide displacement prediction result; the smaller the weight calculation result of the area where the monitoring point is located, the smaller the influence of the deformation data of the area where the monitoring point is located on the landslide displacement prediction result.

5. The method for calculating the spatial regional weight of landslide surface deformation according to claim 1, characterized in that: The deformation data includes displacement, velocity, acceleration, displacement difference between adjacent monitoring points, velocity difference between adjacent monitoring points, and average displacement between adjacent monitoring points. The deformation data are selected according to the actual prediction effect, and at least three types of deformation data are input.

6. The method for calculating the spatial regional weight of landslide surface deformation according to claim 1, characterized in that: The spatial area weight of each monitoring point ranges from 0 to 1, and the sum of the spatial area weights of all points is 1. The larger the spatial area weight, the greater the relative impact.

Images