Black square table landslide displacement prediction method based on GCN-MHA-GRU

Abstract

The invention relates to the technical field of landslide displacement data prediction, in particular to a GCN-MHA-GRU-based landslide displacement prediction method for a black square platform. Acquiring synthetic aperture radar monitoring data of the black square table area; extracting landslide surface deformation data of the black square table area from the synthetic aperture radar monitoring data; decomposing the landslide earth surface deformation data by adopting a three-dimensional deformation decomposition method according to a satellite orbit; establishing a landslide prediction model based on a graph neural network, a multi-head attention mechanism and a gating circulation unit; inputting the decomposed landslide surface deformation data into the landslide prediction model for training; and carrying out landslide displacement prediction on the black square table area according to the trained landslide prediction model. According to the method, an efficient, accurate and robust landslide displacement prediction model is constructed, and powerful technical support can be provided for scientific prevention and control of geological disasters.

Description

Technical Field

[0001] The present invention relates to the technical field of landslide displacement data prediction, and in particular to a Heifangtai landslide displacement prediction method based on GCN-MHA-GRU. Background Art

[0002] Landslides, as a serious geological disaster, pose a huge threat to people's lives and property, especially in the loess tableland area of ​​Northwest China. The severity of landslide disasters is aggravated by its complex geological structure and the special physical and mechanical properties and water sensitivity of loess. The Heifangtai area of ​​Yongjing County, Gansu Province is a typical area. The area is frequently attacked by landslides, which not only seriously threatens the lives of local residents, but also causes immeasurable losses to daily production and life. Therefore, how to predict landslide displacement through historical data and control it within an acceptable error range to provide technical support for local landslide early warning is a research of great significance.

[0003] Based on whether the correlation between spatial and temporal features is taken into account, the current research on landslide displacement is divided into two categories. The first is to build a landslide displacement prediction model based on temporal features. For example, a gated recurrent network (GRU) model was constructed using historical deformation data from GPS monitoring stations. The results showed that GRU has better prediction effect than static models in predicting short-term cumulative landslide displacement. A multi-head attention gated recurrent network (MultiHeadAttention) was proposed based on the deformation results obtained by small baseline subset interferometric synthetic aperture radar technology. The results showed that the MHA-GRU prediction model is superior to the Multivariate-L The STM and GRU models have long-term landslide displacement prediction results that are more consistent with the original InSAR cumulative deformation results. A lightweight model, LiteTransNet, was constructed using the current cutting-edge Transformer, which showed better prediction results compared with the GRU and LSTM models. However, the above studies did not consider the spatial characteristics between landslide displacement data, making the prediction results less reliable and effective. In addition, without considering the impact of spatial characteristics on the relationships between individual landslide points, the geological conditions of the precipitation and ground subsidence areas in the entire region remained unchanged during the landslide displacement prediction process, resulting in low prediction accuracy.

[0004] The second type of prediction model takes into account the spatial relationship between each landslide displacement point. For example, a graph neural network is established based on the location of GNSS monitoring stations to improve the spatial correlation between each monitoring station and realize the prediction of landslides. The spatial correlation between different monitoring points is aggregated by graph convolution to successfully predict future land displacement and deposition. However, the above landslide displacement prediction model does not consider the temporal correlation between the historical data of each monitoring station. Summary of the Invention

[0005] In order to solve the problems existing in the prior art, the present invention provides a Heifangtai landslide displacement prediction method based on GCN-MHA-GRU.

[0006] The present invention adopts the following technical solutions, which are based on the Heifangtai landslide displacement prediction method of GCN-MHA-GRU, including: Obtain synthetic aperture radar monitoring data of the Heifangtai area; The landslide surface deformation data of Heifangtai area were extracted from synthetic aperture radar monitoring data; Decomposing the landslide surface deformation data using a three-dimensional deformation decomposition method according to satellite orbits; A landslide prediction model was established based on graph neural networks, multi-head attention mechanism, and gated recurrent units; Inputting the decomposed landslide surface deformation data into the landslide prediction model for training; The landslide displacement in the Heifangtai area is predicted based on the trained landslide prediction model.

[0007] Furthermore, synthetic aperture radar monitoring data of the Heifangtai area was obtained, including: Synthetic aperture radar monitoring data in each flight mode is obtained according to the satellite orbit of the polar-orbiting satellite; the flight mode includes an ascending orbit mode and a descending orbit mode.

[0008] Furthermore, the landslide surface deformation data of the Heifangtai area were extracted from the synthetic aperture radar monitoring data. Specifically, the landslide surface deformation data of the Heifangtai area were obtained by using small baseline synthetic aperture radar interferometry measurement technology.

[0009] Furthermore, the landslide surface deformation data is decomposed and processed using a three-dimensional deformation decomposition method according to the satellite orbit, including: According to the beam incidence angle of the synthetic aperture radar in the ascending orbit mode and the descending orbit mode and the satellite flight azimuth, the landslide surface deformation data is three-dimensionally decomposed to obtain the landslide surface deformation data in the north-south, east-west and vertical directions respectively.

[0010] Furthermore, a landslide prediction model is established based on graph neural networks, multi-head attention mechanism, and gated recurrent units, including: The landslide prediction model uses a community partitioning algorithm to modularize the landslide surface deformation data to obtain a landslide deformation graph structure, and inputs the landslide deformation graph structure into a graph neural network; The graph neural network performs a graph convolution operation on the landslide deformation graph structure to output a landslide deformation feature graph; The landslide prediction model uses spatiotemporal correlation analysis to extract feature vectors from the landslide deformation feature map and establishes a corresponding graph structure feature matrix; The landslide prediction model uses a multi-head attention mechanism to calculate the correlation of the graph structure feature matrix and output the attention weight of the graph structure feature matrix; The gated recurrent unit processes according to the feature structure matrix and the corresponding attention weights and outputs the landslide displacement prediction result.

[0011] Furthermore, the graph neural network performs graph convolution operations on the landslide deformation graph structure, specifically: The graph neural network includes a propagation module and an output module; The output module includes an aggregator and an updater; the output module is used to update the node features in the landslide deformation graph structure, specifically: The aggregator is used to collect features of nodes and their adjacent nodes in the landslide deformation graph structure; The updater performs graph convolution operations based on the features collected by the aggregator to update the features of the nodes; The output module is used to output the graph structure feature matrix corresponding to the landslide deformation graph structure according to the node features updated by the propagation module.

[0012] Furthermore, the Heifangtai landslide displacement prediction method based on GCN-MHA-GRU is characterized in that the landslide prediction model uses a multi-head attention mechanism to calculate the correlation of the graph structure feature matrix, specifically: Performing a linear transformation on the graph structure feature matrix, and dividing the graph structure feature matrix into a plurality of subspaces under linear projection according to the linear transformation result; The dot product attention method is used to calculate the correlation of each subspace, and the attention score of each subspace is obtained according to the correlation; The attention scores of each subspace are scaled according to the dimension of the hidden layer, and the scaled attention scores in each subspace are concatenated to obtain the attention weights of the graph structure feature matrix.

[0013] Furthermore, the gated recurrent unit performs processing according to the graph structure feature matrix and the corresponding attention weights, including: The gated recurrent unit includes an input layer and an output layer, and the output layer includes an update gate and a reset gate; The input layer receives the graph structure feature matrix and attention weight of the current time step as current information, and obtains the graph structure feature matrix and attention weight of the previous time step as hidden information; The output layer performs weighted calculations on the graph structure feature matrix and attention weights in the hidden information through the update gate and the reset gate, determines the ratio of transmitting the hidden information to the current information, and outputs the prediction result until all the information is transmitted.

[0014] The beneficial effects of the present invention are as follows: the present invention establishes a landslide prediction model based on a graph neural network, a multi-head attention mechanism and a gated recurrent unit. The model comprehensively considers the time series characteristics in the InSAR data and the spatial distribution characteristics of the study area, and can effectively cope with the challenge of accurately predicting landslide displacement while simultaneously capturing information in the time and space dimensions. Among them, the landslide prediction model adopts the method of spatiotemporal correlation analysis to extract the dynamic time characteristics in the InSAR time series data on the one hand; on the other hand, it can show the spatial correlation and dependency between nodes in the InSAR data; the introduction of the multi-head self-attention mechanism further enhances the model's ability to capture global key information; finally, the gated recurrent unit is used to achieve accurate prediction of landslide displacement. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0016] Figure 1 Schematic diagram of the flow of the Heifangtai landslide displacement prediction method based on GCN-MHA-GRU according to an embodiment of the present invention; Figure 2 Schematic diagram of the GCN-MHA-GRU prediction model structure according to an embodiment of the present invention; Figure 3 This is a schematic diagram of node network partitioning results of a community partitioning algorithm according to an embodiment of the present invention; Figure 4 This is a schematic diagram of a graph neural network propagation rule structure according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of a multi-head attention mechanism according to an embodiment of the present invention; Figure 6 This is a schematic diagram of deformation results of SBAS-InSAR orbit data according to an embodiment of the present invention; Figure 7 2 is a schematic diagram for comparing model prediction results according to an embodiment of the present invention. DETAILED DESCRIPTION

[0017] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0018] The Heifangtai area is a loess tableland formed by the Yellow River IV terrace. It is connected to the mountains in the north and the Yellow River II terrace in the south. There is no transition to the III terrace. The altitude is between 1560.4 and 1759.5 meters. It has a temperate semi-arid climate. The annual precipitation is mainly concentrated in the summer, and the rainfall is small. Due to its unique local geological structure and natural environmental factors, as well as long-term agricultural irrigation, the Heifangtai area has become the most serious and frequent landslide area in Gansu Province, posing a major potential threat to the safety of life and property of local residents. Therefore, the present invention proposes a Heifangtai landslide displacement prediction method based on GCN-MHA-GRU, and its flow chart is shown as follows: Figure 1 Shown, including: Obtain synthetic aperture radar monitoring data of the Heifangtai area; With the rapid development of aerospace and remote sensing technology, Interferometric Synthetic Aperture Radar (InSAR) technology has become the preferred tool for early identification and monitoring and early warning of landslide disasters due to its unique advantages in monitoring tiny surface deformations. This technology can achieve continuous monitoring over a large area, all day and all weather, accurately capture tiny ground deformations, and is relatively independent of the influence of atmospheric changes and seasonal factors. The working principle of InSAR is ingenious: first, two SAR antennas with interferometric capabilities are used to obtain two coherent single-view complex images with a certain perspective difference in the same area. Subsequently, by accurately subtracting the terrain phase simulated by an external digital elevation model (DEM), the deformation information of the ground target can be accurately extracted. This process not only reflects the high precision of InSAR technology, but also highlights its strong application potential in geological disaster monitoring.

[0019] In an embodiment of the present invention, for the same area, synthetic aperture radar (SAR) images can be acquired by polar-orbiting satellites in two different flight modes: an ascending orbit mode (the satellite flies from south to north) and a descending orbit mode (the satellite flies from north to south). These two modes can extract surface information from different perspectives. InSAR observations from a single orbit can only provide one-dimensional deformation information along the line of sight (LOS) direction, and are susceptible to various geometric deformation effects during this process, including but not limited to the dwelling phenomenon, perspective zoom effect, and terrain shadowing. These geometric factors can significantly interfere with and distort deformation measurement results in the LOS direction. Given the limitations of single-orbit InSAR observations, to overcome the impact of these geometric deformation effects on deformation measurement accuracy and further more intuitively reflect deformation information in the study area, the embodiment of the present invention adopts a multi-orbit observation strategy, specifically covering data from two ascending orbits and one descending orbit. Based on this, a comprehensive analysis function model for multi-source heterogeneous InSAR datasets is constructed, aiming to comprehensively and accurately obtain monitoring data for the Heifangtai area.

[0020] The landslide surface deformation data of Heifangtai area were extracted from synthetic aperture radar monitoring data; To further enhance the monitoring capabilities of InSAR technology, SBAS-InSAR technology has emerged. This is the small-baseline synthetic aperture radar interferometry technique used in the embodiments of the present invention. The acquired monitoring data were processed using SARscape 5.6.2 on the ENVI (Environment for Visualizing Image) platform using SBAS-InSAR on 150 descending, 167 ascending I, and 182 ascending II Sentinel-1A images. The main process includes connection map generation, interferometry processing, SBAS-InSAR inversion, and geocoding. The maximum normal baseline (%) was set to 100%, and the maximum time baseline (days) was set to 120 days to generate the connection map. The multi-view ratio was set to 4:1 to improve the signal-to-noise ratio of the interferogram. This allowed the extraction of landslide surface deformation data in the Heifangtai area from the acquired synthetic aperture radar monitoring data.

[0021] The landslide surface deformation data were decomposed and processed using a three-dimensional deformation decomposition method according to satellite orbits. 3D deformation decomposition is a high-precision method that uses multi-orbit radar data to comprehensively obtain multi-dimensional information on landslide deformation. According to the difference between the radar beam incident angle and the satellite flight azimuth in ascending and descending orbit imaging, the ascending orbit data and the descending orbit data are decomposed to obtain accurate landslide deformation information in the north-south, east-west and vertical directions. This method is based on the difference between the radar beam incident angle and the satellite flight azimuth in ascending and descending orbit imaging. By accurately analyzing the data collected by the ascending orbit and the descending orbit, the landslide deformation is accurately decomposed and extracted in the north-south, east-west and vertical directions. The synthetic aperture radar (InSAR) observation technology projects the actual surface deformation onto the radar's line-of-sight (LOS) direction. It should be noted that the deformation in the LOS direction can be further decomposed into components in the vertical direction (i.e., vertically downward deformation), the east-west direction (i.e., east-west deformation in the horizontal plane), and the north-south direction (i.e., north-south deformation in the horizontal plane). This decomposition method not only helps to gain a deeper understanding of the three-dimensional characteristics of surface deformation, but also provides the possibility of analyzing and assessing the impact of surface deformation from multiple perspectives. The expression is: Where: The true deformation is projected to the LOS deformation value; , and Respectively represent vertical, north-south, and east-west deformations (vertical upward, north, and east are considered positive); is the SAR sensor incident angle; is the satellite flight azimuth; the time-series InSAR observation results from multiple orbits are combined to solve the three-dimensional equations, which can be expressed in matrix form.

[0022] A landslide prediction model was established based on graph neural networks, multi-head attention mechanism, and gated recurrent units; In the embodiment of the present invention, a landslide prediction model combining graph neural network (GCN), multi-head attention mechanism (MHA) and gated recurrent unit (GRU) is proposed, which is called GCN-MHA-GRU model. The model structure is as follows: Figure 2 As shown in the figure, the community partitioning algorithm is used to present the time series data in the form of node groups. During the data acquisition process, the Align layer is used to process the dimensional features of the data so that it can be used as the input of the CGN module. The MHA module can process time series data in parallel and deeply mine the geographical and functional information between nodes to improve the accuracy of landslide displacement prediction. The GRU module captures and utilizes the long-term dependencies in time series data to improve the generalization ability of the model. The main components of the GCN-MHA-GRU model are as follows: Community Partitioning Algorithm (Fast unfolding algorithm): This algorithm presents the structural information of internal links in the form of node groups and uses them as meta-nodes in the network connection process. By deeply identifying the network community structure, it obtains more network topological relationships that affect community functions. By comparing the initial community structure information with the newly obtained community structure information, the node weights in the network community are calculated. Specifically, nodes belonging to the same area have strong correlations. For example, adjacent roads in a transportation network, proteins with similar functions in biological sciences, and landslides mentioned in this invention all have the same essential connotations. The Fast unfolding algorithm measures the closeness of the community through modularity (Q), which is defined as follows: in represents the edge weight between node i and node j, represents the sum of the weights of the edges connected to node i, represents the sum of the weights of the edges connected to node j, represents the community to which node i belongs, represents the community to which node j belongs, Represents a judgment node With node The Fast unfolding algorithm is divided into two stages. In the first stage, each node in the network community link has an independent community label. Each node traverses the relationship with all adjacent nodes and updates the community label with the largest modularity increment until the community label no longer changes. In the second stage, all nodes in the same community are regarded as new super nodes. The weight of the edge between nodes is the sum of the edge weights of all nodes in the initial community, forming a new graph, such as Figure 3 A simple network with 12 nodes is shown as an example to represent the calculations of the first and second stages. In the embodiment of the present invention, the algorithm is applied to the landslide surface deformation data processing in the Heifangtai area to facilitate better application of the data in the prediction model proposed in the present invention.

[0023] Graph Neural Networks (GCNs): Graph neural networks have emerged as a powerful tool specifically designed to learn representations of non-Euclidean data and have demonstrated significant effectiveness in the field of graph modeling. Given that many time series data contain complex interactions between time and variables, graph neural networks cleverly treat time points or variables as nodes, and convert the correlations between them into edges to construct a network or graph structured model. This model can accurately capture and characterize the diverse and complex relationships between nodes, thereby demonstrating higher efficiency in handling practical tasks. Taking the analysis of connections between regions and the prediction of landslide displacement as an example, the application of graph neural networks is particularly prominent. In this scenario, the points in each region are abstracted as nodes in the graph, and the distance relationship between them is intuitively displayed in the form of edges. By continuously updating the correlation between each node and other nodes, graph neural networks can accurately predict the cumulative deformation.

[0024] In areas where landslides are common, many important datasets acquired through InSAR (Infrared SAR) do not have regular Euclidean spatial structures but are presented in the form of graphs, such as protein molecular structures and transportation networks. Integrating such structured datasets into neural network models and applying them to real-world applications is a significant challenge. Each node in a graph neural network should possess its own characteristics and structural information. GCN utilizes a clever transformation of the convolution operator and extends its application to the graph domain. This not only achieves the translation invariance, local perception, and weight sharing of CNNs in the graph domain, but also enables end-to-end learning of node features and structural information. GCN takes a graph G = (V, A, X) as input, where V represents the set of information about all nodes in the graph, X is the feature matrix, and A is the adjacency matrix describing the graph neural network structure. Based on spectral decomposition and graph convolution, Graph GCN leverages the convolution theorem to first transform the signal from the original space to the frequency domain using the Fourier transform, then multiply the signal in the frequency domain, and finally transform it back to the original space using the inverse Fourier transform, thus achieving graph convolution. Specifically, GCN consists of two parts: a propagation module and an output module. The propagation module consists of an aggregator and an updater. Based on the information of each node in the graph, it can obtain more potential relationships in the source node through the information of its adjacent nodes. The updater constantly updates the state of the newly learned node information. In addition to the obvious information and structure of the nodes obtained by the propagation module, the output module deeply mines the potential information to complete the corresponding task. Each GCN layer generates a node-level output feature matrix H (the scale of the feature matrix is ​​N*F, where N is the number of all nodes in the graph and F is the dimension of the potential information obtained by each node). The propagation rules are as follows: in, is the normalized adjacency matrix, A is the initial adjacency matrix, is the adjacency matrix bias term; For the The node feature matrix of the layer, is the activation function, is the normalized degree matrix, For the The node feature matrix of the layer, Representative The weight matrix of the layer.

[0025] Spatial and temporal correlation analysis: For landslide areas, landslide displacement generally shows obvious temporal correlation. On the one hand, landslide displacement is affected by recent agricultural irrigation; on the other hand, landslide displacement has a strong periodicity and is easily affected by summer rainfall. For example, in the Heifangtai area, landslide disasters are frequent during the season when farmland requires a large amount of irrigation. For the landslide displacement to be predicted, historical observation data with a 12-day period is obtained through InSAR to predict possible landslide displacement in the future and avoid loss of life and property. This embodiment of the present invention extracts the eigenvector of each node area, then superimposes all the eigenvectors to obtain a feature matrix X, and then transmits the obtained feature matrix to the GRU layer to extract the temporal characteristics of each node area.

[0026] Due to geographical factors, the spatial correlation of landslide susceptibility in different regions can also vary significantly. For example, the occurrence of landslide geological hazards in the Heifangtai area is mainly affected by neighboring areas and the distance from nearby reservoirs and cultivated land. To fully simulate the spatial correlation of landslide displacement in the region, the embodiment of the present invention constructs two graphs with different topological structures to encode the correlation between neighboring nodes and functional correlation: (1) Geographic Adjacency Graph (GAG) : In this embodiment of the present invention, in addition to the self-connection of the node area itself, all other areas are connected, so that the node area can learn more features from the surrounding areas, which is expressed as: in: Represented in a geographic adjacency graph In, nodes and The relationship between the nodes, ga is the name of the adjacency graph ga.

[0027] (2) Functional Similarity Graph (FSG) :By learning two node areas with similar landslide displacements, a deeper relationship between them is explored, and full connection has been performed in the first step, considering both the influencing factors of distance and the influencing factors of other node areas with similar functions; for example, in a certain area, land irrigation, different farmers have different land ownership, different irrigation levels, different irrigation areas, and different crops planted on the land, so the landslide displacements in the corresponding node areas are also different. In the embodiment of the present invention, POI (Point-of-interest) similarity is used to quantify the functional similarity between different areas. By calculating the POI similarity between any two areas, it is determined whether the two have similar trajectories based on the obtained value. The similarity threshold is set to 0.8, and the specific setting value can also be set according to the actual usage. It is reasonable. The present invention does not impose any restrictions. The POI similarity calculation process is specifically expressed as follows: in , Represent the POI vectors of region i and region j respectively, sim(*) is the calculation function of the Pearson coefficient, Represented in a geographic adjacency graph In, nodes and The relationship between nodes, fs represents the name of the functional similarity graph fs.

[0028] After constructing the geographic adjacency graph and functional similarity graph, we use a graph convolutional neural network to extract spatial correlations between the features of the two to obtain the spatial features of each node area. We then combine the spatial features obtained from the two to obtain the final spatial feature information: The final graph structure feature matrix Z and adjacency matrix A are obtained, where , , Indicates integrating the results into G.

[0029] Multi-head self-attention machine (Multi-head Attention): In the embodiment of the present invention, a multi-head self-attention mechanism is used after spatiotemporal graph convolution to model the correlation between labels, thereby integrating information from different node spaces and extracting more geographical and functional information between nodes. This method can not only extract node features more comprehensively, but also enhance the ability to capture the relationship between labels from multiple perspectives; by utilizing the parallel processing characteristics of multi-head attention, a more fine-grained data representation can be obtained, thereby significantly improving the accuracy of the model and the ability to model label dependencies and node features, such as Figure 4As shown, the embodiment of the present invention uses InSAR time series data as Query (Q), linearly transforms the initialized label vector matrix and divides it into Key (K) and Value (V), and divides it into different subspaces with different linear projections. Subsequently, in each subspace, the scaled dot product attention is used to obtain the correlation between each time step and the subspace, and the attention of each subspace is calculated to obtain the attention score. At the same time, in order to solve the problems of gradient explosion and gradient disappearance, the attention score is further scaled by the square root of the dimension K of the hidden layer. Then, the attention scores obtained in different subspaces are spliced ​​and linearly transformed. Finally, the output of each attention head is connected, and the final output is the result of all attention weights.

[0030] Gated Recurrent Unit (GRU): The Gated Recurrent Unit (GRU) is a recurrent neural network (RNN) specifically designed for time series prediction tasks. This model can more effectively capture and utilize long-term dependencies in time series data. This improvement enhances the generalization ability of landslide displacement prediction. The uniqueness of the GRU lies in controlling the updating and transmission of information through a gated recurrent mechanism, obtaining important information in the time series and reducing the propagation of useless information. It can effectively prevent the network from losing key information when processing new input data and can utilize long-term dependencies in the sequence for modeling. In the embodiment of the present invention, the GRU obtains surface time series deformation information obtained using small baseline integrated interferometric synthetic aperture radar technology, which can predict the cumulative deformation caused by the landslide. It also learns more temporal features from the time series data to make future deformation predictions. Such predictions can serve as an important reference for decision-making to mitigate the risks posed by landslides.

[0031] As an improved RNN, GRU solves the problem of gradient explosion through the gating mechanism, such as Figure 5 As shown, each GRU unit will and the hidden state at time t-1 As input, the output part consists of an update gate and a reset gate. The update gate determines how much historical information is passed to the current step, and the reset gate determines how much historical information can be passed to the future. Both of these gated units are controlled by and Calculate the weighted sum and determine the hidden layer state at time t after the activation function. There are four main mathematical expressions involved in GRU, which are expressed as follows: in, is the sigmoid activation function, Represents the weight matrix from the input state to the reset gate, represents the weight matrix from the hidden state to the reset gate, 、 and are three different bias values, with subscripts r, z, and h as distinguishing marks. is the input at the current moment, is the hidden state of the previous moment, tanh is the hyperbolic tangent activation function, ⊙ represents the Hadamard product, Represents the candidate hidden state at the current time step.

[0032] The decomposed landslide surface deformation data is input into the landslide prediction model for training; and the landslide displacement of the Heifangtai area is predicted according to the trained landslide prediction model.

[0033] The accuracy evaluation of landslide displacement prediction is an important part of landslide displacement model establishment. In the embodiment of the present invention, the root mean square error (RMSE), mean absolute percentage error (MAPE) and coefficient of determination (R 2 ) is used as the accuracy evaluation index to evaluate the prediction performance, RMSE and MAPE are used to quantify the difference between the true value and the predicted value. The lower the score, the higher the consistency with the true value. At the same time, R 2 Evaluates the strength of the linear correlation between the true and predicted values. The closer the index is to 1, the better the model is able to capture the trends and patterns in landslide displacement. The following is the definition of the metric: Where, and are the true value and predicted value of the landslide displacement of the i-th data point, respectively, and m is the number of data points.

[0034] Selecting appropriate hyperparameters can effectively improve model performance and robustness and avoid overfitting and underfitting. The embodiment of the present invention further adopts four data enhancement techniques: time shift, time scaling, noise addition, and data repetition, which increase the diversity and quantity of samples, thereby improving the performance and generalization ability of the model. After multiple experiments, the overall hyperparameter settings of the model were finally determined. For the GCN-MHA-GRU model, 12 is used as the historical time step, "GLU" or Gated Linear Unit is used as the activation function, [3, 2] is the spatial convolution kernel size, Chebyshev graph convolution is used as the graph convolution type, Dropout is set to 0.3, adam is used as the optimizer, 20 is the step size of the learning rate decay, and the learning rate decay coefficient is 0.5.

[0035] In another experimental embodiment of the present invention: like Figure 6 As shown in the deformation results from the ascending and descending SBAS-InSAR data, the deformation locations of the three data types are essentially consistent, with the Heifangtai landslide exhibiting significant deformation. The overall annual average deformation rate for the descending track is -2.82 mm / year, with a local maximum of 6.07 mm / year and a local minimum of -41.24 mm / year. For ascending track I, the overall average deformation rate is -2.13 mm / year, with a local maximum of 12.55 mm / year and a local minimum of -26.19 mm / year. For descending track II, the overall average deformation rate is -3.61 mm / year, with a local maximum of 16.86 mm / year and a local minimum of -36.25 mm / year. The SBAS-InSAR results show that the overall spatial deformation of the significant landslide area at Heifangtai is slow, while the local deformation is severe. Due to the limitations of single-track InSAR data, the embodiment of the present invention uses three-dimensional InSAR decomposition to decompose the deformation monitoring data into east-west, north-south, and vertical displacement monitoring data, directly obtaining the three-dimensional surface deformation field of the significant area. By comparing the displacement monitoring data before and after the three-dimensional deformation decomposition, the directional information of the slope movement can be clearly obtained.

[0036] The three-dimensional deformation results are highly correlated with the direction of the slopes in the prominent areas of Heifangtai. These areas have experienced multiple landslides, and the upper slopes of the source area show relatively obvious deformation results. The north-south deformation rate is -20.83~59.69mm / year, the east-west deformation rate is -106.39~130.45mm / year, and the vertical deformation rate is -135.43~27.85mm / year. The Heifangtai area shows a relatively obvious downward displacement trend along the downslope of the source area, which is consistent with the The frequent occurrence of landslides in the region corresponds to this. Furthermore, the vertical displacement rate gradually decreases from the upslope to the downslope of the tableland. The overall horizontal displacement of the tableland is significantly greater than the vertical displacement. The Heifangtai landslide-prone area is primarily characterized by slow movement along the source slope. Landslides in Heifangtai are primarily divided into four areas: Jiaojia, Moshigou, Fangtai, and Dangchuan. All areas exhibit downward displacement along the upslope of the tableland. Combined with regional topographic analysis, it was found that multiple landslides have formed multiple distinct landslide zones on the downslope of the tableland due to multiple landslides. Under the influence of gravity, landslides have the characteristic of moving along the upslope to the downslope of the tableland. Comparison of three-dimensional deformation in east-west, north-south, and vertical directions reveals that the regional extent of landslides is roughly consistent.

[0037] The embodiment of the present invention further compares and analyzes the displacement results of four typical areas on June 25, 2023 with the GCN-MHA-GRU prediction results, and substitutes the difference between them into the same standard for comparison with the true value. The comparison result diagram is shown in the figure below. Figure 7 As shown in the figure, from the analysis of the predicted values ​​and the true values ​​of the four typical regions, it can be concluded that the spatial distribution of the deformation zone within the region is consistent, and according to the errors of the four typical regions, it can be concluded that the errors are basically kept within a small error range. The above results show that the GCN-MHA-GRU model proposed in this invention has good accuracy in predicting the spatial distribution and temporal characteristics of different regions.

[0038] To verify the prediction accuracy of GCN-MHA-GRU, the embodiment of the present invention selected several typical landslide displacement models, namely GCN, GRU, and GCN-GRU, for comparison. Among the models that only consider InSAR landslide displacement prediction, GCN-MHA-GRU outperforms other typical models in all evaluation indicators in the four study areas A, B, C, and D. The GCN, GRU, and GCN-GRU models only consider temporal features and ignore the closely related spatial connections between regions. The GCN-MHA-GRU better considers the spatial correlation of the entire region and extracts more detailed regional divisions before model training. By analyzing the spatial and temporal features between each node, it can accurately capture the more complex spatial relationships in the study area.

[0039] Using three-dimensional data collected by drones, significant signs of landslides can be clearly identified in the area, and these landslides are located in close proximity to multiple residential areas. This is of great practical significance for assessing the risk of local landslide displacement. Further observations revealed that these landslide phenomena exhibit close spatial connectivity, indicating the possibility of complex interactions between them. To more accurately analyze and predict the dynamic behavior of these landslides, this paper uses a graph neural network method, treating the fishing net points within the landslide as nodes in the network and analyzing the spatiotemporal correlations between these nodes to capture the interaction patterns between them. This method not only takes into account the characteristics of the individual landslides themselves, but also comprehensively considers the possible mutual influences between landslides, thereby providing more accurate and comprehensive landslide prediction results, which is of great significance for protecting the lives and property of local residents.

[0040] 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. The Heifangtai landslide displacement prediction method based on GCN-MHA-GRU is characterized by: include: Obtain synthetic aperture radar monitoring data of the Heifangtai area; The landslide surface deformation data of Heifangtai area were extracted from synthetic aperture radar monitoring data; Decomposing the landslide surface deformation data using a three-dimensional deformation decomposition method according to satellite orbits; A landslide prediction model was established based on graph neural networks, multi-head attention mechanism, and gated recurrent units; Inputting the decomposed landslide surface deformation data into the landslide prediction model for training; The landslide displacement in the Heifangtai area is predicted based on the trained landslide prediction model.

2. The Heifangtai landslide displacement prediction method based on GCN-MHA-GRU according to claim 1 is characterized by: Acquire synthetic aperture radar monitoring data for the Heifangtai area, including: Synthetic aperture radar monitoring data in each flight mode is obtained according to the satellite orbit of the polar-orbiting satellite; the flight mode includes an ascending orbit mode and a descending orbit mode.

3. The Heifangtai landslide displacement prediction method based on GCN-MHA-GRU according to claim 1 is characterized by: The landslide surface deformation data of the Heifangtai area were extracted from the synthetic aperture radar monitoring data. Specifically, the landslide surface deformation data of the Heifangtai area were obtained using small baseline synthetic aperture radar interferometry measurement technology.

4. The Heifangtai landslide displacement prediction method based on GCN-MHA-GRU according to claim 2 is characterized by: Decomposing and processing the landslide surface deformation data using a three-dimensional deformation decomposition method according to satellite orbits includes: According to the beam incidence angle of the synthetic aperture radar in the ascending orbit mode and the descending orbit mode and the satellite flight azimuth, the landslide surface deformation data is three-dimensionally decomposed to obtain the landslide surface deformation data in the north-south, east-west and vertical directions respectively.

5. The Heifangtai landslide displacement prediction method based on GCN-MHA-GRU according to claim 1 is characterized by: A landslide prediction model is established based on graph neural networks, multi-head attention mechanism, and gated recurrent units, including: The landslide prediction model uses a community partitioning algorithm to modularize the landslide surface deformation data to obtain a landslide deformation graph structure, and inputs the landslide deformation graph structure into a graph neural network; The graph neural network performs a graph convolution operation on the landslide deformation graph structure to output a landslide deformation feature graph; The landslide prediction model uses spatiotemporal correlation analysis to extract feature vectors from the landslide deformation feature map and establishes a graph structure feature matrix corresponding to the feature vectors; The landslide prediction model uses a multi-head attention mechanism to calculate the correlation of the graph structure feature matrix and output the attention weight of the graph structure feature matrix; The gated recurrent unit processes the graph structure feature matrix and the corresponding attention weights and outputs the landslide displacement prediction result.

6. The Heifangtai landslide displacement prediction method based on GCN-MHA-GRU according to claim 5 is characterized by: The graph neural network performs graph convolution operations on the landslide deformation graph structure, specifically: The graph neural network includes a propagation module and an output module; The output module includes an aggregator and an updater; the output module is used to update the node features in the landslide deformation graph structure, specifically: The aggregator is used to collect features of nodes and their adjacent nodes in the landslide deformation graph structure; The updater performs graph convolution operations based on the features collected by the aggregator to update the features of the nodes; The output module is used to output a graph structure feature matrix corresponding to the landslide deformation graph structure according to the node features updated by the propagation module.

7. The Heifangtai landslide displacement prediction method based on GCN-MHA-GRU according to claim 5 is characterized by: The landslide prediction model uses a multi-head attention mechanism to calculate the correlation of the graph structure feature matrix, specifically: Performing a linear transformation on the graph structure feature matrix, and dividing the graph structure feature matrix into a plurality of subspaces under linear projection according to the linear transformation result; The dot product attention method is used to calculate the correlation of each subspace, and the attention score of each subspace is obtained according to the correlation; The attention scores of each subspace are scaled according to the dimension of the hidden layer, and the scaled attention scores in each subspace are concatenated to obtain the attention weights of the graph structure feature matrix.

8. The Heifangtai landslide displacement prediction method based on GCN-MHA-GRU according to claim 5 is characterized by: The gated recurrent unit performs processing according to the feature structure matrix and the corresponding attention weights, including: The gated recurrent unit includes an input layer and an output layer, and the output layer includes an update gate and a reset gate; The input layer receives the feature structure matrix and attention weight of the current time step as current information, and obtains the feature structure matrix and attention weight of the previous time step as hidden information; The output layer performs weighted calculations on the feature structure matrix and attention weights in the hidden information through the update gate and the reset gate, determines the ratio of the hidden information to the current information, and outputs the prediction result until all the information is transferred.

Images