Landslide pattern recognition method based on virtual-real domain adaptation

Abstract

The application discloses a landslide mode recognition method based on virtual-real domain self-adaption, relates to the technical field of landslide mode recognition, determines a virtual domain marking data set based on selecting a typical landslide geological structure, and determines real domain landslide radar data based on real domain landslides; the virtual domain marking data set and the real domain landslide radar data are combined to recognize a target landslide mode of a current real domain landslide, the two kinds of data are used as samples in the process, and the problem of overfitting and poor generalization ability caused by insufficient sample quantity in the prior art is overcome; meanwhile, an updated feature extractor obtained through training can guarantee performance in a real scene; and then, the obtained target landslide mode is more accurate, and early warning can be effectively made for landslide disaster-prone areas.

Description

Technical Field

[0001] This invention relates to the field of landslide pattern recognition technology, and in particular to a landslide pattern recognition method based on virtual-real domain adaptation. Background Technology

[0002] my country has a vast mountainous area and is prone to landslides, which pose a serious threat to people's lives and property and the operation of major infrastructure. Landslides, as a typical geological disaster, are characterized by their sudden occurrence and wide-ranging impact. They often disrupt transportation arteries, damage buildings, and encroach on farmland, seriously threatening the lives and property of the people. Simultaneously, they cause irreversible damage to the ecological environment. Accurately and quickly acquiring landslide patterns is a core prerequisite for conducting disaster loss assessments, developing emergency rescue plans, and implementing subsequent remediation projects. Traditional landslide pattern identification methods have many limitations, failing to meet the efficiency and accuracy requirements of modern disaster prevention and mitigation work. Early identification often relies on manual visual interpretation of remote sensing images, requiring professionals to manually delineate landslide boundaries using geological knowledge to determine the landslide pattern within the landslide area. This not only consumes significant manpower and time but also has extremely low identification efficiency, making it unsuitable for large-scale area investigations or emergency response to sudden landslide disasters. Furthermore, it is susceptible to the influence of personal experience and subjective judgment, leading to biases and affecting the scientific rigor of subsequent assessments and remediation work. While some automated identification methods have been applied, they are mostly based on traditional machine learning algorithms. However, deep learning models based on real data often suffer from overfitting due to insufficient sample size, resulting in poor generalization ability. Numerical simulation can generate three-dimensional deformation vectors for the entire landslide process, solving the problem of "no data available." However, there are significant distribution differences (i.e., "domain gaps") between the generated simulation data and real slope radar data, leading to a severe performance degradation of classifiers trained directly on simulation data in real-world scenarios. Furthermore, current landslide early warning technologies mainly focus on predicting instability time (such as landslide dating) or displacement, lacking real-time intelligent identification capabilities for landslide sliding patterns (circular sliding, planar sliding, wedge-shaped sliding, toppling failure, etc.). Summary of the Invention

[0003] Therefore, it is necessary to propose a landslide pattern recognition method based on virtual-real domain adaptation to address the above problems.

[0004] A landslide pattern recognition method based on virtual-real domain adaptation, the method comprising:

[0005] S1: Select typical landslide geological structures and collect corresponding geological parameters. Based on the geological parameters, establish a three-dimensional discrete element numerical model for each preset landslide mode. The preset landslide modes include: circular arc sliding, planar sliding, wedge sliding, and toppling failure. The geological parameters include: rock layer type, structural plane orientation, and strength parameters.

[0006] S2: Multiple virtual monitoring points are uniformly distributed across the entire slope surface of each three-dimensional discrete element numerical model;

[0007] S3: Obtain the three-dimensional displacement vector of each virtual monitoring point under each landslide mode;

[0008] S4: Determine the spatiotemporal raster and virtual domain labeled dataset based on each 3D displacement vector in each landslide mode;

[0009] S5: Based on the spatiotemporal grid and by introducing a gating mechanism, the multi-head self-attention mechanism of the feature extractor is updated, thereby obtaining the updated feature extractor;

[0010] S6: By updating the feature extractor, the spatiotemporal grid is used to extract features to obtain the virtual domain depth feature vector. The depth feature vector is composed of the depth temporal feature vector and the depth spatial feature vector. The virtual domain depth feature vectors corresponding to the four landslide modes constitute the feature library.

[0011] S7: Supervised training is performed on the virtual domain labeled dataset using the landslide pattern label classifier, virtual domain deep feature vector, and labels corresponding to different landslide patterns in the Transformer encoder model to determine the classification loss;

[0012] S8: Obtain the real-domain landslide radar data corresponding to each landslide mode. The real-domain landslide radar data of the four landslide modes constitute the real-domain landslide sample. The sample is preprocessed according to the data structure consistent with the virtual domain labeled dataset to form the real-domain labeled dataset. At the same time, the obtained real-domain raster corresponds to the spatiotemporal raster, and the first real-domain depth feature vector is extracted.

[0013] S9: Input the virtual domain depth feature vector and the first real domain depth feature vector into the domain discriminator for domain classification and discrimination, and the domain discriminator outputs the domain discrimination loss; the domain discrimination loss is combined with the classification loss to construct a joint loss;

[0014] S10: The updated feature extractor and the domain discriminator utilize a joint loss and introduce a gradient reversal layer. This gradient reversal layer maintains an identity mapping during forward propagation and multiplies the gradient by a negative coefficient during backward propagation, thereby achieving minimax adversarial optimization between the feature extractor and the domain discriminator and obtaining the updated joint loss.

[0015] S11: Train the updated feature extractor by updating the joint loss to obtain a domain-invariant domain-aligned feature extractor;

[0016] S12: Input the real-domain raster into the domain-aligned feature extractor to obtain the second real-domain depth feature vector corresponding to each landslide pattern. Then, input the second real-domain depth feature vector and the feature library into the KNN nearest neighbor model. The KNN nearest neighbor model searches the feature library for the second real-domain depth feature vector based on the depth feature vector similarity metric. The K nearest neighbors form the neighborhood set;

[0017] S13: Calculate the frequency of each landslide pattern label in the neighborhood set, and smooth the frequency distribution corresponding to each landslide pattern by combining the kernel density estimation model to obtain the posterior probability of the real domain landslide sample belonging to each landslide pattern, and retain the landslide pattern corresponding to the largest posterior probability. When the largest posterior probability is greater than the matching degree threshold, the landslide pattern corresponding to the largest posterior probability is taken as the target landslide pattern of the real domain landslide.

[0018] In one embodiment, the determination of the spatiotemporal raster and virtual domain labeled dataset based on each three-dimensional displacement vector under each landslide mode includes:

[0019] Based on the principle of slope radar interferometry, each three-dimensional displacement vector under each landslide mode is projected onto the radar line of sight to generate the corresponding virtual radar LOS displacement time series.

[0020] Each virtual radar LOS displacement time series is converted into the original radar polar coordinate format, and then reconstructed into LOS displacement values ​​with a spatial width and height two-dimensional grid and a three-dimensional structure of time frame sequence after interpolation.

[0021] The LOS displacement values ​​of all virtual monitoring points under each landslide mode at a fixed time are arranged into a two-dimensional grid diagram.

[0022] The two-dimensional grid images of each landslide mode at different times are stacked along the time axis to form a spatiotemporal grid.

[0023] The spatiotemporal grid corresponding to each landslide pattern Landslide pattern labeling was performed, and four types of labeled spatiotemporal rasters constituted a virtual domain labeled dataset.

[0024] In one embodiment, updating the multi-head self-attention mechanism of the feature extractor based on the spatiotemporal grid and introducing a gating mechanism to obtain an updated feature extractor includes:

[0025] Each spatiotemporal grid is divided into several time sub-blocks, and each time sub-block is mapped to a corresponding feature vector. Multiple feature vectors form a spatiotemporal token.

[0026] The relative spatiotemporal offset is determined based on the relative offset in spatial height direction, relative offset in spatial width direction, and relative time offset between any two spatiotemporal tokens.

[0027] A learnable relative position offset is constructed by mapping relative spatiotemporal offset based on a relative position offset mapping function.

[0028] The gating coefficients of the token pair are obtained by operating on the relative spatiotemporal offset based on the relative position feature mapping function and the sigmoid activation function;

[0029] The multi-head self-attention magnitude is determined by the gating coefficient, relative position bias, query vector obtained by linear mapping of the a-th spatiotemporal token, and key vector obtained by linear mapping of the b-th spatiotemporal token, in order to update the multi-head self-attention mechanism of the feature extractor and thus obtain the updated feature extractor.

[0030] In one embodiment,

[0031] The method utilizes the principle of slope radar interferometry to project each three-dimensional displacement vector under each landslide mode onto the radar line-of-sight direction, generating a corresponding virtual radar LOS displacement time series, which is achieved through the following expression:

[0032]

[0033] in, For each landslide pattern Virtual radar LOS displacement timing at any given moment; It is oriented east-west; It is oriented north-south; θ is the radar incident angle; φ is the radar azimuth angle; i is the irregularly distributed virtual monitoring point; t is time;

[0034] The process of converting each virtual radar LOS displacement time series into the original radar polar coordinate format, and then reconstructing it into a LOS displacement value with a two-dimensional grid of spatial width and height and a three-dimensional structure of time frame sequence after interpolation, is achieved by the following expression:

[0035]

[0036]

[0037] in, The LOS displacement value of a new grid point (W,H) obtained after interpolation of a virtual monitoring point at time t. For virtual radar LOS displacement timing; For weights; d i is the distance from the grid point to the monitoring point; p is the weight parameter.

[0038] In one embodiment, the step of stacking the two-dimensional grid images of each landslide mode at different times along the time axis to form a spatiotemporal grid is achieved by the following expression:

[0039]

[0040]

[0041]

[0042] in, Here is a two-dimensional grid diagram for the i-th landslide mode; H is the spatial height; W is the spatial width; T is the time frame; For the Mth landslide mode, a spatiotemporal grid is used.

[0043] The virtual domain labeled dataset expression is as follows:

[0044]

[0045] Where Ds is the virtual domain labeled dataset; For spatiotemporal grids; The label corresponding to the Mth landslide pattern; n is the landslide pattern index; represents the total number of landslide patterns; {} represents the set.

[0046] In one embodiment,

[0047] The relative position offset is achieved by the following expression:

[0048]

[0049]

[0050]

[0051] in, For spatiotemporal grids; This is the a-th spacetime token; For feature vectors; This refers to relative spatiotemporal offset; This refers to the relative offset in the spatial height direction; This refers to the relative offset in the direction of spatial width. This is a relative time offset; For relative position offset; This represents a relative position offset mapping function;

[0052] The relative position feature mapping function and sigmoid activation function Regarding the relative spatiotemporal offset Perform the operation to obtain the gating coefficient of the token pair This can be achieved using the following expression:

[0053]

[0054] in: This is the gating coefficient for the corresponding token pair; It is the sigmoid activation function; , These are learnable parameters; It is a relative position feature mapping function;

[0055] The multi-head self-attention magnitude is achieved through the following expression:

[0056]

[0057] in, Represents the a-th spacetime token With the b-th spacetime token The size of the multi-head self-attention between them; Represents the a-th spacetime token The query vector obtained after linear mapping; This represents the b-th spacetime token. The key vector obtained after linear mapping; Represents the key vector transpose; The feature dimension representing the query vector or key vector; Indicates relative position offset; This represents the gating coefficient.

[0058] In one embodiment, the virtual domain deep feature vector and the classification loss are implemented using the following expression:

[0059]

[0060]

[0061] in, This indicates an update to the feature extractor. Spatiotemporal grid for each landslide pattern; This represents the depth feature vector of the virtual domain. For classification loss; A landslide pattern label classifier; The label corresponding to the Mth landslide pattern; Here, n is the loss function; n is the landslide mode index. This represents the total number of landslide patterns.

[0062] In one embodiment, the updated joint loss, the first ground truth deep feature vector, and the second ground truth deep feature vector are implemented using the following expression:

[0063]

[0064]

[0065]

[0066]

[0067] in, The first real-domain depth feature vector; Indicates updating the feature extractor For real-domain raster; L is the joint loss; For classification loss; Loss is determined by the domain; To update the joint loss; The second real-domain depth feature vector; This is a domain-aligned feature extractor.

[0068] In one embodiment, the landslide pattern recognition method based on virtual-real domain adaptation further includes:

[0069] S20: Define the target landslide pattern of the real-domain landslide as the dominant landslide pattern, and the other three landslide patterns as candidate landslide patterns; determine the competitive intensity of the dominant landslide pattern relative to each candidate landslide pattern;

[0070] S21: Construct a typical spatial structure template for each landslide mode in the dominant landslide mode and candidate landslide modes. The typical spatial structure template is used to characterize the relevant anomaly features of each landslide mode in real-domain landslide radar data. The relevant anomaly features include: anomaly area distribution characteristics, expansion direction and geometric structure.

[0071] S22: Match the abnormal area distribution characteristics corresponding to the dominant landslide pattern with the typical spatial structure template of each candidate landslide pattern to obtain the spatial template matching degree;

[0072] S23: Determine the abnormal region for each candidate landslide pattern, and determine the area ratio of the abnormal region, the degree of consistency between the expansion direction of the abnormal region and the expansion direction of the candidate landslide pattern, and the connectivity of the abnormal region. Based on the area ratio, the degree of consistency, and the connectivity, determine the activity level of the candidate landslide pattern.

[0073] S24: By combining the weighted difference in posterior probability between the current dominant landslide mode and each candidate landslide mode, the competition intensity, spatial template matching degree and activity level, the conversion driving force of the dominant landslide mode for each candidate landslide mode is constructed.

[0074] S25: Using the Softmax normalization principle, the transformation driving quantity is mapped to the transformation probability, that is, the transformation probability of the dominant landslide mode to each candidate landslide mode; when the transformation probability is greater than or equal to the transformation probability threshold, and the spatial template matching degree is greater than or equal to the spatial template matching degree threshold, the corresponding candidate landslide mode is the landslide mode to be transformed by the dominant landslide mode.

[0075] In one embodiment, the process for determining the conversion probability threshold and the spatial template matching degree threshold is as follows:

[0076] Based on dividing historical monitoring samples into a set of samples showing occurrence pattern transformation and a set of samples showing no occurrence pattern transformation;

[0077] Analyze the dominant landslide patterns in historical monitoring samples separately. Relative to each candidate landslide pattern The transformation drivers are the first mean, second mean, first standard deviation, and second standard deviation in the sample set of transformations in the occurrence mode and the sample set of transformations in the non-occurrence mode, respectively.

[0078] Each candidate landslide pattern in the historical monitoring sample The spatial template matching degree is the third mean, fourth mean, third standard deviation, and fourth standard deviation in the sample set where pattern transformation occurred and the sample set where pattern transformation did not occur, respectively;

[0079] The conversion probability threshold is determined based on the first mean, the second mean, the first standard deviation, and the second standard deviation;

[0080] The spatial template matching threshold is determined based on the third mean, fourth mean, third standard deviation, and fourth standard deviation.

[0081] This invention identifies target landslide patterns in the current real-world landslide by combining virtual domain labeled datasets and real-world landslide radar data. The use of these two types of data as samples overcomes the problems of overfitting and poor generalization ability caused by insufficient sample size in existing technologies. At the same time, the trained and updated feature extractor can ensure performance in real-world scenarios, thereby making the acquired target landslide patterns more accurate and providing effective early warning for areas prone to landslide disasters. Attached Figure Description

[0082] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0083] in:

[0084] Figure 1 This is a flowchart of a landslide pattern recognition method based on virtual-real domain adaptation in one embodiment;

[0085] Figure 2 A flowchart illustrating the conversion of a dominant landslide mode into candidate landslide modes in one embodiment;

[0086] Figure 3 This is a structural block diagram of a computer device in one embodiment. Detailed Implementation

[0087] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0088] To address the technical problems in the background art, this application provides a landslide pattern recognition method based on virtual-real domain adaptation.

[0089] like Figure 1 As shown, in one embodiment, a landslide pattern recognition method based on virtual-real domain adaptation is provided. This method can be applied to both terminals and servers; this embodiment illustrates its application to a terminal. The landslide pattern recognition method based on virtual-real domain adaptation specifically includes the following steps:

[0090] S1: Select typical landslide geological structures and collect corresponding geological parameters. Based on the geological parameters, establish a three-dimensional discrete element numerical model for each preset landslide mode. The preset landslide modes include: circular arc sliding, planar sliding, wedge sliding, and toppling failure. The geological parameters include: rock layer type, structural plane orientation, and strength parameters.

[0091] S2: Multiple virtual monitoring points are uniformly distributed across the entire slope surface of each three-dimensional discrete element numerical model; multiple virtual monitoring points are uniformly distributed across the entire slope surface of each three-dimensional discrete element numerical model to monitor slope surface displacement and simulate the landslide change process of the corresponding landslide mode of the three-dimensional discrete element numerical model under different triggering conditions (rainfall, excavation, water level change) (e.g., deformation-instability-sliding).

[0092] S3: Obtain the three-dimensional displacement vector (east-west, north-south, and sky) of each virtual monitoring point under each landslide mode;

[0093] S4: Determine the spatiotemporal raster and virtual domain labeled dataset Ds based on each 3D displacement vector under each landslide pattern;

[0094] S5: Based on the spatiotemporal grid and by introducing a gating mechanism, the multi-head self-attention mechanism of the feature extractor is updated, thereby obtaining the updated feature extractor. ;

[0095] S6: By updating the feature extractor spatiotemporal grid Feature extraction is performed to obtain the virtual domain deep feature vector. Deep feature vectors Composed of depth temporal feature vectors and depth spatial feature vectors; virtual domain depth feature vectors corresponding to the four landslide modes. Constructing a feature library ;

[0096] S7: Landslide Pattern Label Classifier in Transformer Encoder Model Virtual domain deep feature vector Virtual domain labeled dataset with labels corresponding to different landslide patterns Supervised training is performed to determine classification loss. ;

[0097] S8: Obtain the real-domain landslide radar data corresponding to each landslide mode. The real-domain landslide radar data of the four landslide modes constitute the real-domain landslide sample, and are then compared with the virtual domain labeled dataset. A consistent data structure is used for preprocessing to form a real-world labeled dataset. Simultaneously obtained real-domain raster With the spatiotemporal grid Correspondingly, the first ground-domain depth feature vector is extracted. ;

[0098] S9: Transfer the virtual domain deep feature vector With the first real-domain depth feature vector Common input domain discriminator Gd Perform domain classification and discrimination; domain discriminator G d Output domain discrimination loss The domain discrimination loss Combined with the classification loss Construct the joint loss L;

[0099] S10: The updated feature extractor Domain discriminator G d The joint loss L is utilized, and a gradient reversal layer (GRL) is introduced. This gradient reversal layer maintains the identity mapping during forward propagation and multiplies the gradient by a negative coefficient -λ during backward propagation, thereby achieving minimax adversarial optimization between the feature extractor and the domain discriminator, and obtaining an updated joint loss.

[0100] S11: By updating the joint loss Update the feature extractor Training yields a domain-invariant domain-aligned feature extractor. ;

[0101] S12: Convert the real-domain raster Input to domain-aligned feature extractor The second real-domain depth feature vector corresponding to each landslide pattern is obtained. and the second real-domain depth feature vector and the feature library The input is fed into the KNN nearest neighbor model, which is based on a deep feature vector similarity metric from a feature library. Searching for and second real-domain deep feature vectors The K nearest neighbors form the neighborhood set. ;

[0102] S13: Calculate the frequency of each landslide pattern label in the neighborhood set, and smooth the frequency distribution corresponding to each landslide pattern using a kernel density (KDE) estimation model to obtain the posterior probability that the real-domain landslide sample belongs to each landslide pattern. The landslide pattern corresponding to the largest posterior probability is retained. When the largest posterior probability is greater than the matching degree threshold τ, the landslide pattern corresponding to the largest posterior probability is used as the target landslide pattern of the real domain landslide.

[0103] In one embodiment, determining the spatiotemporal raster and virtual domain labeled dataset Ds based on each three-dimensional displacement vector under each landslide mode includes:

[0104] Using the principle of slope radar interferometry, each three-dimensional displacement vector under each landslide mode is projected onto the radar line-of-sight direction to generate the corresponding virtual radar LOS displacement time series. ;

[0105] Each virtual radar LOS displacement time series is converted into the original radar polar coordinate format (slant range × azimuth angle), and then reconstructed into LOS displacement values ​​with a spatial width and height two-dimensional grid and a time frame sequence three-dimensional structure after interpolation.

[0106] The LOS displacement values ​​of all virtual monitoring points under each landslide mode at a fixed time t are arranged into a two-dimensional grid diagram.

[0107] The two-dimensional grid diagrams at different times for each landslide pattern are... ,Will Stacking along the time axis forms a spacetime grid. ;

[0108] The spatiotemporal grid corresponding to each landslide pattern Landslide pattern labeling was performed, with four types of spatiotemporal grids being labeled. The virtual domain labeled dataset Ds is constructed.

[0109] In one embodiment, the multi-head self-attention mechanism of the feature extractor is updated based on the spatiotemporal grid and a gating mechanism is introduced, thereby obtaining an updated feature extractor. include:

[0110] Each spatiotemporal grid The time is divided into several time sub-blocks, and each time sub-block is mapped to a corresponding feature vector. Multiple feature vectors form a spatiotemporal token.

[0111] Based on any two spacetime tokens and The relative spatiotemporal offset is determined by the relative offset in the spatial height direction, the relative offset in the spatial width direction, and the relative time offset.

[0112] Relative spatiotemporal offset based on relative position offset mapping function Perform mapping to construct learnable relative position biases. ;

[0113] Based on relative position feature mapping function and sigmoid activation function Regarding the relative spatiotemporal offset Perform the operation to obtain the gating coefficient of the token pair ;

[0114] By gating coefficient Relative position offset The a-th spacetime token The query vector obtained after linear mapping and the b-th spacetime token The key vector obtained after linear mapping Determine the size of the multi-head self-attention This is used to update the multi-head self-attention mechanism of the feature extractor, thereby obtaining an updated feature extractor. .

[0115] In one embodiment,

[0116] The method utilizes the principle of slope radar interferometry to project each three-dimensional displacement vector under each landslide mode onto the radar line-of-sight direction, generating a corresponding virtual radar LOS displacement time series. This can be achieved using the following expression:

[0117]

[0118] in, For each landslide pattern Virtual radar LOS displacement timing at any given moment; It is oriented east-west; It is oriented north-south; θ is the radar incident angle; φ is the radar azimuth angle; i is the irregularly distributed virtual monitoring point; t is time;

[0119] The process of converting each virtual radar LOS displacement time series into the original radar polar coordinate format (slant range × azimuth angle), and then reconstructing it into a LOS displacement value with a two-dimensional grid of spatial width and height and a three-dimensional structure of time frame sequence after interpolation, is achieved by the following expression:

[0120]

[0121]

[0122] in, The LOS displacement value of a new grid point (W,H) obtained after interpolation of a virtual monitoring point at time t. For virtual radar LOS displacement timing; For weights; d i is the distance from the grid point to the monitoring point; p is the weight parameter (usually 2).

[0123] In one embodiment, two-dimensional grid images of each landslide pattern at different times are stacked along the time axis to form a spatiotemporal grid. This can be achieved using the following expression:

[0124]

[0125]

[0126]

[0127] in, Here is a two-dimensional grid diagram for the i-th landslide mode; H is the spatial height; W is the spatial width; T is the time frame; For the Mth landslide mode, a spatiotemporal grid is used.

[0128] The virtual domain labeled dataset Ds expression is as follows:

[0129]

[0130] Where Ds is the virtual domain labeled dataset; For spatiotemporal grids; The label corresponding to the Mth landslide pattern; n is the landslide pattern index; represents the total number of landslide patterns; {} represents the set.

[0131] In one embodiment,

[0132] The relative position offset This can be achieved using the following expression:

[0133]

[0134]

[0135]

[0136] in, For spatiotemporal grids; This is the a-th spacetime token; For feature vectors; This refers to relative spatiotemporal offset; This refers to the relative offset in the spatial height direction; This refers to the relative offset in the direction of spatial width. This is a relative time offset; For relative position offset; This represents a relative position offset mapping function;

[0137] The relative position feature mapping function and sigmoid activation function Regarding the relative spatiotemporal offset Perform the operation to obtain the gating coefficient of the token pair This can be achieved using the following expression:

[0138]

[0139] in: The gating coefficient for the corresponding token pair is adaptively generated by the gating function and is used to adjust the relative position offset. The degree of contribution to the magnitude of multi-head self-attention; It is the sigmoid activation function; , These are learnable parameters; It is a relative position feature mapping function;

[0140] The size of the multi-head self-attention This can be achieved using the following expression:

[0141]

[0142] in, Represents the a-th spacetime token With the b-th spacetime token The size of the multi-head self-attention between them; Represents the a-th spacetime token The query vector obtained after linear mapping; This represents the b-th spacetime token. The key vector obtained after linear mapping; Represents the key vector transpose; The feature dimension representing the query vector or key vector; Indicates relative position offset; This represents the gating coefficient.

[0143] In one embodiment, the virtual domain deep feature vector and the classification loss This can be achieved using the following expression:

[0144]

[0145]

[0146] in, This indicates an update to the feature extractor. Spatiotemporal grid for each landslide pattern; This represents the depth feature vector of the virtual domain. For classification loss; A landslide pattern label classifier; The label corresponding to the Mth landslide pattern; Here, n is the loss function; n is the landslide mode index. This represents the total number of landslide patterns.

[0147] In one embodiment, the updated joint loss The first real-domain depth feature vector and the two real-domain depth feature vectors This can be achieved using the following expression:

[0148]

[0149]

[0150]

[0151]

[0152] in, The first real-domain depth feature vector; Indicates updating the feature extractor For real-domain raster; L is the joint loss; For classification loss; The domain discrimination loss is used to measure the domain discriminator G. d The ability to distinguish between virtual and real domain landslide samples; λ is a weighting coefficient used to adjust the relative weight between the classification task and the domain alignment task; To update the joint loss; The second real-domain depth feature vector; This is a domain-aligned feature extractor.

[0153] In one embodiment, the landslide pattern recognition method based on virtual-real domain adaptation, such as Figure 2 As shown, it also includes:

[0154] S20: Define the target landslide pattern of the real-domain landslide as the dominant landslide pattern. The other three landslide models were selected as candidate landslide models. Determine the competitive intensity of the dominant landslide mode relative to each candidate landslide mode. Both the dominant landslide mode and the candidate landslide mode are preset landslide modes, which include: circular arc sliding, planar sliding, wedge sliding, and toppling failure.

[0155] S21: Constructing the dominant landslide model Compared with candidate landslide patterns Typical spatial structure templates for each landslide pattern The typical spatial structure template The relevant anomalous features used to characterize each landslide pattern in real-domain landslide radar data include: anomalous region distribution characteristics R, expansion direction, and geometric structure.

[0156] S22: Will dominate landslide pattern The corresponding abnormal area distribution characteristics R and the typical spatial structure template of each candidate landslide mode Perform matching to obtain the spatial template matching degree. ;

[0157] S23: Determine each candidate landslide pattern Identify the abnormal regions and determine the area percentage of the abnormal regions. The degree of consistency between the expansion direction of the anomalous region and the expansion direction of the candidate landslide pattern. Connectivity with abnormal regions Based on the area ratio Consistency and the connectivity Identify candidate landslide models activity level ;

[0158] S24: Weighted average of the posterior probability differences between the current dominant landslide model and each candidate landslide model, and competition intensity. Spatial template matching degree and activity level Construct the transformation driver for each candidate landslide mode corresponding to the dominant landslide mode. ;

[0159] S25: Utilize the Softmax normalization principle to transform the driving force. This is mapped to a transformation probability, i.e., the probability of the dominant landslide mode transforming into each candidate landslide mode; when the transformation probability... Greater than or equal to the conversion probability threshold And the spatial template matching degree Greater than or equal to the spatial template matching threshold When the corresponding candidate landslide mode is the landslide mode to be converted from the dominant landslide mode;

[0160] In one embodiment, the conversion probability threshold Matching degree threshold with the spatial template The determination process is as follows:

[0161] Based on dividing historical monitoring samples into a set of occurrence pattern transformation samples Set of samples that did not undergo mode transformation ;

[0162] Analyze the dominant landslide patterns in historical monitoring samples separately. Relative to each candidate landslide pattern Conversion driver In the sample set of occurrence mode transformation respectively Set of samples that did not undergo mode transformation First mean Second mean First standard deviation Second standard deviation ;

[0163] Each candidate landslide pattern in the historical monitoring sample Spatial template matching degree In the sample set of occurrence mode transformation respectively Set of samples that did not undergo mode transformation The third mean Fourth mean Third standard deviation and the fourth standard deviation ;

[0164] According to the first mean Second mean First standard deviation Second standard deviation Determine the conversion probability threshold ;

[0165] According to the third mean Fourth mean Third standard deviation and the fourth standard deviation Determine the spatial template matching threshold .

[0166] In one embodiment,

[0167] The dominant landslide pattern The corresponding abnormal area distribution characteristics R and the typical spatial structure template of each candidate landslide mode Perform matching to obtain the spatial template matching degree. This can be achieved using the following expression:

[0168]

[0169] in, Indicates the degree of spatial template matching; This represents the similarity calculation function; Indicates the dominant landslide mode The corresponding abnormal region distribution characteristics R and; This represents the typical spatial structure template corresponding to the Mth candidate landslide pattern; the maximum spatial template matching degree. The corresponding candidate landslide pattern best matches the distribution feature R of the abnormal area;

[0170] Based on the area ratio Consistency and the connectivity Identify candidate landslide models activity level This can be achieved using the following expression:

[0171]

[0172] in, Indicates candidate landslide patterns The level of activity; Indicates the area percentage of the abnormal region; This indicates the degree of consistency between the expansion direction of the anomalous region and the expansion direction of the candidate landslide pattern; Indicates the connectivity of abnormal regions; , , These are the weighting coefficients.

[0173] In one embodiment, the construction of the dominant landslide pattern corresponds to the transformation driving force of each candidate landslide pattern. This can be achieved using the following expression:

[0174]

[0175] in, Indicates the dominant landslide mode To candidate landslide models The conversion driver; These are the weighting coefficients; Weight the posterior probability differences between candidate landslide patterns and each dominant landslide pattern; dominant landslide mode Compared to candidate landslide models The intensity of competition is weighted; Weight the spatial template matching degree; Weighting the activity level of candidate landslide patterns; Spatial template matching degree;

[0176] In one embodiment, the transformation driving quantity is achieved by utilizing the Softmax normalization principle. The mapping to transformation probabilities, i.e., the transformation probabilities from the dominant landslide mode to each candidate landslide mode, are realized by the following expression:

[0177]

[0178] in, Indicates the dominant landslide model To candidate landslide models The probability of transformation. m represents the Softmax normalization calculation performed on all candidate landslide modes other than the current dominant landslide mode; Indicates the dominant landslide model To candidate landslide models The conversion driver; It is an exponential mapping function.

[0179] In one embodiment, the first mean Second mean First standard deviation Second standard deviation This can be achieved using the following expression:

[0180]

[0181]

[0182]

[0183]

[0184] in, This is the first mean; It is the second mean; The first standard deviation; The second standard deviation; This is a sample set for the transformation of occurrence patterns; This is the set of samples for which no pattern transformation has occurred. The dominant landslide pattern in historical monitoring samples Relative to each candidate landslide pattern The conversion driver; Historical monitoring samples;

[0185] In one embodiment, the third mean Fourth mean Third standard deviation and the fourth standard deviation This can be achieved using the following expression:

[0186]

[0187]

[0188]

[0189]

[0190] in, It is the third mean; It is the fourth mean; The third standard deviation; The fourth standard deviation; This is a sample set for the transformation of occurrence patterns; This is the set of samples for which no pattern transformation has occurred. For each candidate landslide pattern in the historical monitoring sample Spatial template matching degree; Historical monitoring samples;

[0191] In one embodiment,

[0192] The determination of the conversion probability threshold based on the first mean, the second mean, the first standard deviation, and the second standard deviation is achieved through the following expression:

[0193]

[0194] in, This is the conversion probability threshold; This is the first mean; It is the second mean; The first standard deviation; The second standard deviation;

[0195] The determination of the spatial template matching threshold based on the third mean, fourth mean, third standard deviation, and fourth standard deviation is achieved through the following expression:

[0196]

[0197] in, This is the spatial template matching threshold; It is the third mean; It is the fourth mean; The third standard deviation; This is the fourth standard deviation.

[0198] This invention identifies target landslide patterns in the current real-world landslide by combining virtual domain labeled datasets and real-world landslide radar data. The use of these two types of data as samples overcomes the problems of overfitting and poor generalization ability caused by insufficient sample size in existing technologies. At the same time, the trained and updated feature extractor can ensure performance in real-world scenarios, thereby making the obtained target landslide patterns more accurate and providing effective early warning for areas prone to landslide disasters.

[0199] Figure 3 An internal structural diagram of a computer device in one embodiment is shown. This computer device can specifically be a terminal or a server. Figure 3As shown, the computer device includes a processor, memory, and network interface connected via a system bus. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and may also store a computer program. When executed by the processor, this computer program enables the processor to implement a landslide pattern recognition method based on virtual-real domain adaptation. The internal memory may also store a computer program, which, when executed by the processor, enables the processor to implement the landslide pattern recognition method based on virtual-real domain adaptation. Those skilled in the art will understand that... Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0200] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0201] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0202] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A landslide pattern recognition method based on virtual-real domain adaptation, characterized in that, The method includes: S1: Select typical landslide geological structures and collect corresponding geological parameters. Based on the geological parameters, establish a three-dimensional discrete element numerical model for each preset landslide mode. The preset landslide modes include: circular arc sliding, planar sliding, wedge sliding, and toppling failure. The geological parameters include: rock layer type, structural plane orientation, and strength parameters. S2: Multiple virtual monitoring points are uniformly distributed across the entire slope surface of each three-dimensional discrete element numerical model; S3: Obtain the three-dimensional displacement vector of each virtual monitoring point under each landslide mode; S4: Determine the spatiotemporal raster and virtual domain labeled dataset based on each 3D displacement vector in each landslide mode; S5: Based on the spatiotemporal grid and by introducing a gating mechanism, the multi-head self-attention mechanism of the feature extractor is updated, thereby obtaining the updated feature extractor; S6: By updating the feature extractor, the spatiotemporal grid is used to extract features to obtain the virtual domain depth feature vector. The depth feature vector is composed of the depth temporal feature vector and the depth spatial feature vector. The virtual domain depth feature vectors corresponding to the four landslide modes constitute the feature library. S7: Supervised training is performed on the virtual domain labeled dataset using the landslide pattern label classifier, virtual domain deep feature vector, and labels corresponding to different landslide patterns in the Transformer encoder model to determine the classification loss; S8: Obtain the real-domain landslide radar data corresponding to each landslide mode. The real-domain landslide radar data of the four landslide modes constitute the real-domain landslide sample. The sample is preprocessed according to the data structure consistent with the virtual domain labeled dataset to form the real-domain labeled dataset. At the same time, the obtained real-domain raster corresponds to the spatiotemporal raster, and the first real-domain depth feature vector is extracted. S9: Input the virtual domain depth feature vector and the first real domain depth feature vector into the domain discriminator for domain classification and discrimination, and the domain discriminator outputs the domain discrimination loss; the domain discrimination loss is combined with the classification loss to construct a joint loss; S10: The updated feature extractor and the domain discriminator utilize a joint loss and introduce a gradient reversal layer. This gradient reversal layer maintains an identity mapping during forward propagation and multiplies the gradient by a negative coefficient during backward propagation, thereby achieving minimax adversarial optimization between the feature extractor and the domain discriminator and obtaining the updated joint loss. S11: Train the updated feature extractor by updating the joint loss to obtain a domain-invariant domain-aligned feature extractor; S12: Input the real-domain raster into the domain-aligned feature extractor to obtain the second real-domain depth feature vector corresponding to each landslide mode, and input the second real-domain depth feature vector and the feature library into the KNN nearest neighbor model. The KNN nearest neighbor model searches the feature library for the K nearest neighbor samples that are closest to the second real-domain depth feature vector based on the depth feature vector similarity measure to form a neighborhood set. S13: Calculate the frequency of each landslide pattern label in the neighborhood set, and smooth the frequency distribution corresponding to each landslide pattern by combining the kernel density estimation model to obtain the posterior probability of the real domain landslide sample belonging to each landslide pattern, and retain the landslide pattern corresponding to the largest posterior probability. When the largest posterior probability is greater than the matching degree threshold, the landslide pattern corresponding to the largest posterior probability is taken as the target landslide pattern of the real domain landslide.

2. The landslide pattern recognition method based on virtual-real domain adaptation according to claim 1, characterized in that, The dataset for determining the spatiotemporal raster and virtual domain annotation based on each three-dimensional displacement vector under each landslide pattern includes: Based on the principle of slope radar interferometry, each three-dimensional displacement vector under each landslide mode is projected onto the radar line of sight to generate the corresponding virtual radar LOS displacement time series. Each virtual radar LOS displacement time series is converted into the original radar polar coordinate format, and then reconstructed into LOS displacement values ​​with a spatial width and height two-dimensional grid and a three-dimensional structure of time frame sequence after interpolation. The LOS displacement values ​​of all virtual monitoring points under each landslide mode at a fixed time are arranged into a two-dimensional grid diagram. The two-dimensional grid images of each landslide mode at different times are stacked along the time axis to form a spatiotemporal grid. The spatiotemporal grid corresponding to each landslide pattern Landslide pattern labeling was performed, and four types of labeled spatiotemporal rasters constituted a virtual domain labeled dataset.

3. The landslide pattern recognition method based on virtual-real domain adaptation according to claim 1, characterized in that, The step of updating the multi-head self-attention mechanism of the feature extractor based on the spatiotemporal grid and introducing a gating mechanism to obtain the updated feature extractor includes: Each spatiotemporal grid is divided into several time sub-blocks, and each time sub-block is mapped to a corresponding feature vector. Multiple feature vectors form a spatiotemporal token. The relative spatiotemporal offset is determined based on the relative offset in spatial height direction, relative offset in spatial width direction, and relative time offset between any two spatiotemporal tokens. A learnable relative position offset is constructed by mapping relative spatiotemporal offset based on a relative position offset mapping function. The gating coefficients of the token pair are obtained by operating on the relative spatiotemporal offset based on the relative position feature mapping function and the sigmoid activation function; The multi-head self-attention magnitude is determined by the gating coefficient, relative position bias, query vector obtained by linear mapping of the a-th spatiotemporal token, and key vector obtained by linear mapping of the b-th spatiotemporal token, in order to update the multi-head self-attention mechanism of the feature extractor and thus obtain the updated feature extractor.

4. The landslide pattern recognition method based on virtual-real domain adaptation according to claim 2, characterized in that, The method utilizes the principle of slope radar interferometry to project each three-dimensional displacement vector under each landslide mode onto the radar line-of-sight direction, generating a corresponding virtual radar LOS displacement time series, which is achieved through the following expression: in, For each landslide pattern Virtual radar LOS displacement timing at any given moment; It is oriented east-west; It is oriented north-south; θ is the radar incident angle; φ is the radar azimuth angle; i is the irregularly distributed virtual monitoring point; t is time; The process of converting each virtual radar LOS displacement time series into the original radar polar coordinate format, and then reconstructing it into a LOS displacement value with a two-dimensional grid of spatial width and height and a three-dimensional structure of time frame sequence after interpolation, is achieved by the following expression: in, The LOS displacement value of a new grid point (W,H) obtained after interpolation of a virtual monitoring point at time t. For virtual radar LOS displacement timing; For weights; d i is the distance from the grid point to the monitoring point; p is the weight parameter.

5. The landslide pattern recognition method based on virtual-real domain adaptation according to claim 2, characterized in that, The process of stacking the two-dimensional grid images of each landslide mode at different times along the time axis to form a spatiotemporal grid is achieved by the following expression: in, Here is a two-dimensional grid diagram for the i-th landslide mode; H is the spatial height; W is the spatial width; T is the time frame; For the Mth landslide mode, a spatiotemporal grid is used. The virtual domain labeled dataset expression is as follows: Where Ds is the virtual domain labeled dataset; For spacetime grids; The label corresponding to the Mth landslide pattern; n is the landslide pattern index; represents the total number of landslide patterns; {} represents the set.

6. The landslide pattern recognition method based on virtual-real domain adaptation according to claim 3, characterized in that, The relative position offset is achieved by the following expression: in, For spacetime grids; This is the a-th spacetime token; For feature vectors; This refers to relative spatiotemporal offset; This refers to the relative offset in the spatial height direction; This refers to the relative offset in the direction of spatial width. This is a relative time offset; For relative position offset; This represents a relative position offset mapping function; The relative position feature mapping function and sigmoid activation function Regarding the relative spatiotemporal offset Perform the operation to obtain the gating coefficient of the token pair This can be achieved using the following expression: in: This is the gating coefficient for the corresponding token pair; It is the sigmoid activation function; , These are learnable parameters; It is a relative position feature mapping function; The multi-head self-attention magnitude is achieved through the following expression: in, Represents the a-th spacetime token With the b-th spacetime token The size of the multi-head self-attention between them; Represents the a-th spacetime token The query vector obtained after linear mapping; This represents the b-th spacetime token. The key vector obtained after linear mapping; Represents the key vector Transpose of; The feature dimension representing the query vector or key vector; Indicates relative position offset; This represents the gating coefficient.

7. The landslide pattern recognition method based on virtual-real domain adaptation according to claim 3, characterized in that, The virtual domain depth feature vector and the classification loss are implemented through the following expressions: in, This indicates an update to the feature extractor. Spatiotemporal grid for each landslide pattern; This represents the depth feature vector of the virtual domain. For classification loss; A landslide pattern label classifier; The label corresponding to the Mth landslide pattern; Here, n is the loss function; n is the landslide mode index. This represents the total number of landslide patterns.

8. The landslide pattern recognition method based on virtual-real domain adaptation according to claim 3, characterized in that, The updated joint loss, the first ground truth depth feature vector, and the second ground truth depth feature vector are implemented through the following expression: in, The first real-domain depth feature vector; Indicates updating the feature extractor For real-domain raster; L is the joint loss; For classification loss; Loss is determined by the domain; To update the joint loss; This is the second real-domain depth feature vector; This is a domain-aligned feature extractor.

9. The landslide pattern recognition method based on virtual-real domain adaptation according to claim 1, characterized in that, Also includes: S20: Define the target landslide mode of the real domain landslide as the dominant landslide mode, and the other three landslide modes as candidate landslide modes. Determine the competitive intensity of the dominant landslide mode relative to each candidate landslide mode; S21: Construct a typical spatial structure template for each landslide mode in the dominant landslide mode and candidate landslide modes. The typical spatial structure template is used to characterize the relevant anomaly features of each landslide mode in real-domain landslide radar data. The relevant anomaly features include: anomaly area distribution characteristics, expansion direction and geometric structure. S22: Match the abnormal area distribution characteristics corresponding to the dominant landslide pattern with the typical spatial structure template of each candidate landslide pattern to obtain the spatial template matching degree; S23: Determine the abnormal region for each candidate landslide pattern, and determine the area ratio of the abnormal region, the degree of consistency between the expansion direction of the abnormal region and the expansion direction of the candidate landslide pattern, and the connectivity of the abnormal region. Based on the area ratio, the degree of consistency, and the connectivity, determine the activity level of the candidate landslide pattern. S24: By combining the weighted difference in posterior probability between the current dominant landslide mode and each candidate landslide mode, the competition intensity, spatial template matching degree and activity level, the conversion driving force of the dominant landslide mode for each candidate landslide mode is constructed. S25: Using the Softmax normalization principle, the transformation driving quantity is mapped to the transformation probability, that is, the transformation probability of the dominant landslide mode to each candidate landslide mode; when the transformation probability is greater than or equal to the transformation probability threshold, and the spatial template matching degree is greater than or equal to the spatial template matching degree threshold, the corresponding candidate landslide mode is the landslide mode to be transformed by the dominant landslide mode.

10. The landslide pattern recognition method based on virtual-real domain adaptation according to claim 9, characterized in that, The process for determining the conversion probability threshold and the spatial template matching degree threshold is as follows: Based on dividing historical monitoring samples into a set of samples showing occurrence pattern transformation and a set of samples showing no occurrence pattern transformation; Analyze the dominant landslide patterns in historical monitoring samples separately. Relative to each candidate landslide pattern The transformation drivers are the first mean, second mean, first standard deviation, and second standard deviation in the sample set of transformations in the occurrence mode and the sample set of transformations in the non-occurrence mode, respectively. Each candidate landslide pattern in the historical monitoring sample The spatial template matching degree is the third mean, fourth mean, third standard deviation, and fourth standard deviation in the sample set where pattern transformation occurred and the sample set where pattern transformation did not occur, respectively; The conversion probability threshold is determined based on the first mean, the second mean, the first standard deviation, and the second standard deviation; The spatial template matching threshold is determined based on the third mean, fourth mean, third standard deviation, and fourth standard deviation.

Images