A Kriging gravity anomaly interpolation estimation method based on graph convolutional network

Through the graph convolution network-based method, the traditional Kriging method has solved the problems of high complexity and low accuracy in gravity anomaly interpolation calculation, and achieved more efficient interpolation prediction, which can handle complex spatial structures and nonlinear relationships, and improve the Computational efficiency.

CN118839153BActive Publication Date: 2024-11-26HENAN SONGSHAN LAB IND RES INST CO LTD LUOYANG BRANCH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411311069.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-20
Publication Date
2024-11-26
Estimated Expiration
2044-09-20

AI Technical Summary

Technical Problem

The traditional Kriging method has high complexity and low prediction accuracy in gravity anomaly interpolation calculation, making it difficult to effectively deal with complex spatial structures and nonlinear relationships, especially in large-scale data.

Method used

Using a graph-convolution network-based method, by constructing a spatial graph, the feature data of the observation points are input into the graph-convolution network for multi-layer graph convolution, extracting features and updating Kriging's weight, combining the nonlinear modeling ability of deep learning , generate more accurate interpolated prediction values.

Benefits of technology

It improves the accuracy and robustness of interpolation, reduces the computational complexity, better captures complex nonlinear spatial relationships and local global dependencies, and simplifies the covariance estimation process.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118839153B_ABST
    Figure CN118839153B_ABST
Patent Text Reader

Abstract

A Kriging gravity anomaly interpolation estimation method based on graph convolutional network collects feature data of gravity anomaly observation points to form feature vectors, and constructs a spatial graph with each observation point as a node; establishes an interpolation estimation model, uses graph convolutional network to perform multi-layer graph convolution to extract features, and uses the extracted features to construct a covariance function to update the Kriging weights; uses the collected feature data to train the interpolation estimation model, adjusts the convolution kernel parameters of the graph convolutional network by minimizing the loss function of Kriging interpolation, and obtains the recalculated Kriging weights after model training and adjustment; uses the recalculated Kriging weights to perform Kriging interpolation on unknown points to obtain the prediction results of gravity anomaly interpolation. While maintaining the spatial statistical advantages of Kriging interpolation, the hierarchical structure of the neural network is used to capture and express complex nonlinear relationships, thereby improving the accuracy of interpolation and reducing the computational complexity and amount of Kriging interpolation.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of geodetic gravity anomaly solution, and in particular to a Kriging gravity anomaly interpolation estimation method based on a graph convolutional network. Background Art

[0002] Gravity anomaly is the change in the strength of the earth's gravity field. When the density of material under the earth's surface changes, gravity anomaly will occur. However, the measurement of gravity anomaly can only be single-point measurement. In practical applications, interpolation calculations are often required. Especially in the calculation of geoid models, it is greatly affected by topography, crust density, calculation methods, etc., and cannot be accurately quantified using mathematical models. A better interpolation algorithm has better practicality in predicting and generalizing the measurement data.

[0003] Interpolation of geospatial data is a key technology in fields such as geographic information systems and environmental science. Kriging interpolation, as a widely used geostatistical method, predicts the value of unknown points by considering spatial autocorrelation. However, traditional Kriging interpolation methods are mainly based on linear assumptions, such as assuming that points of the same distance have the same correlation. This assumption may not fully capture complex spatial nonlinear relationships or spatial dependence of data when dealing with complex spatial structures or in the presence of non-uniform distributions. Gravity anomaly data has anisotropy and nonlinear relationships, which may lead to a decrease in prediction accuracy, and traditional Kriging interpolation involves solving a linear system whose scale is proportional to the number of data points, which will increase computational complexity and lead to low efficiency in large-scale data sets. Although co-Kriging can handle multivariate data, it requires modeling the covariance between each variable, has high computational complexity, and is difficult to scale. Summary of the invention

[0004] The technical problem to be solved by the present invention is to overcome the defects of the traditional Kriging method, such as complex gravity anomaly interpolation calculation and low prediction accuracy, and to provide a Kriging gravity anomaly interpolation estimation method based on a graph convolutional network.

[0005] The technical solution adopted by the present invention to solve the above technical problems is: a Kriging gravity anomaly interpolation estimation method based on a graph convolutional network, comprising the following steps:

[0006] (1) Collect characteristic data of gravity anomaly observation points. The characteristic data include spatial position characteristics of the observation points and observation characteristics that characterize the gravity anomaly of the observation points. The spatial position characteristics and observation characteristics constitute the characteristic vector of the observation point.

[0007] (2) Each observation point is regarded as a node, the feature vector of the observation point is regarded as the node feature, and the correlation between nodes is regarded as the edge feature; the spatial graph of the graph convolutional network is constructed with the node features and edge features of each node;

[0008] (3) Establishing an interpolation estimation model, wherein the interpolation estimation model takes the node features and edge features of the spatial graph as input, uses a graph convolution network to perform multi-layer graph convolution on the input to extract features, and uses the extracted features to construct a covariance function to update the Kriging weights. The updated Kriging weights are used to generate interpolation prediction values ​​through the Kriging interpolation method;

[0009] (4) Use the collected feature data to train the interpolation estimation model, adjust the convolution kernel parameters of the graph convolution network by minimizing the loss function of the Kriging interpolation, and obtain the recalculated Kriging weights after model training and adjustment;

[0010] (5) Use the Kriging weights recalculated after model training and adjustment to perform Kriging interpolation on the unknown points to obtain the prediction results of gravity anomaly interpolation.

[0011] The graph convolution network takes the initial feature vector of the node as input and uses multi-layer graph convolution to update the node features. The feature update formula of each layer is:

[0012] ;

[0013] in, Is a node In the The characteristics of the layer, Is a node Updated features, It is The weight matrix of the layer, is the adjacency matrix after adding self-loops, is the adjacency matrix, is the identity matrix; yes The degree matrix of which the diagonal elements are the degrees of each node; is the activation function.

[0014] In the graph convolutional network, a pooling layer is provided after each layer of graph convolution to perform global pooling on the updated node features.

[0015] The features extracted by the graph convolutional network are used to construct the covariance function represented by the radial basis function:

[0016] ;

[0017] in, is the covariance function, , Respectively represent nodes and , , Respectively represent nodes and No. Features extracted after layer graph convolution; Represents the distance between features.

[0018] The update of the Kriging weights is solved using the following equation:

[0019] ;

[0020] in, is the covariance matrix between the known observation points constructed, is the covariance vector between the known observation points and the predicted points, is the Kriging weight.

[0021] The graph convolution network adopts weighted convolution, and the features extracted after each layer of graph convolution operation are used to update the Kriging weights.

[0022] In the step (1), the spatial position features include the latitude and longitude coordinates and elevation of the observation point, and the observation features include the east-west vertical line deviation value, the north-south vertical line deviation value and the crust density value of the observation point.

[0023] In the step (2), the edge feature adopts distance, weight or edge weight based on covariance matrix.

[0024] In the step (4), during the training of the interpolation estimation model, the input data is forward propagated through the graph convolutional network to calculate the predicted value of the Kriging gravity anomaly interpolation, and the loss value of the model is calculated according to the loss function, and the convolution kernel parameters of the graph convolutional network are updated through back propagation.

[0025] The Kriging gravity anomaly interpolation prediction value is calculated using a weighted linear combination of known observation points. The calculation formula is:

[0026] ;

[0027] in, represents the interpolation prediction value of the unknown point, Represents known observation points The characteristic vector of Represents known observation points The kriging weights are Indicates the number of observation points.

[0028] The beneficial effects of the present invention are as follows: the complex spatial features in the data can be effectively extracted through the graph convolutional network, and the graph convolutional neural network can be combined with Kriging interpolation to introduce the nonlinear modeling capabilities of deep learning while maintaining the spatial statistical advantages of Kriging interpolation. The complex nonlinear relationships can be captured and expressed through the hierarchical structure of the neural network, and nonlinear spatial data can be better handled, thereby improving the accuracy of interpolation. Moreover, the graph convolutional network can effectively capture local and global spatial dependencies, map high-dimensional data into a lower-dimensional latent space, thereby simplifying the covariance estimation process and reducing the computational complexity and amount of Kriging interpolation.

[0029] Traditional Kriging assumes that the data is stationary, that is, the statistical characteristics are spatially invariant, but in many practical applications, this assumption is often not true. Graph convolutional networks can learn the complex spatial structure of data during training and have better generalization capabilities. This combination can make Kriging interpolation more robust to non-stationary data. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a flow chart of the graph convolution network Kriging gravity anomaly interpolation calculation of the present invention.

[0031] Figure 2 is the covariance function fitting curve of the gravity anomaly.

[0032] Figure 3 It is the Kriging gravity anomaly interpolation result based on graph convolutional network. DETAILED DESCRIPTION

[0033] The technical solution of the present invention is described clearly and completely below in conjunction with the accompanying drawings and specific implementation methods. The specific contents listed in the following embodiments are not limited to the technical features necessary for the technical problems to be solved by the technical solutions recorded in the claims. At the same time, the enumerated embodiments are only part of the present invention, not all embodiments.

[0034] The Kriging gravity anomaly interpolation estimation method based on graph convolutional network of the present invention uses graph convolutional network to perform multi-layer graph convolution on the collected feature data of gravity anomaly observation points to extract features, and uses the extracted features to update the Kriging weights. The convolution kernel parameters are dynamically calculated or adjusted through model training to reflect the Kriging weights. The recalculated weights are used to perform Kriging interpolation on the unknown points to obtain a more accurate prediction of the gravity anomaly value. The method includes the steps of data collection and collation, model construction, training, interpolation calculation output, etc. The specific method is as follows:

[0035] 1. Data collection and collation: Collect characteristic data of gravity anomaly observation points. The characteristic data include spatial location characteristics of the observation points and observation characteristics that characterize the gravity anomaly of the observation points. Spatial location characteristics include the latitude and longitude coordinates and elevation of the observation points. Observation characteristics include the deviation value of the east-west vertical line, the deviation value of the north-south vertical line, and the crust density value of the observation points. Spatial location characteristics and observation characteristics constitute the characteristic vector of the observation point. The collected data can be divided into training set, validation set and test set, which are used for training, optimizing or testing model performance respectively.

[0036] 2. Spatial graph construction and feature representation: Each observation point is regarded as a node, and the edges between nodes represent spatial correlation. The feature vector of the observation point is used as the node feature, such as the spatial position coordinates, vertical deviation and other observation values ​​and other related features. The correlation between nodes is used as the edge feature, such as distance, weight or edge weight based on the covariance matrix. The node features and edge features can be standardized through standardization and normalization to improve the stability and effect of model training. The spatial graph of the graph convolutional network is constructed with the node features and edge features of each node, and the spatial data is represented as a graph. , where the node Corresponding to the known sampling points, the edge Represents the spatial relationship between adjacent points, which is represented by an adjacency matrix. The initial feature vector of the node can characterize the measured gravity anomaly value of the observation point and is used as the input of the graph convolutional network.

[0037] 3. Establish an interpolation estimation model: The input layer of the interpolation estimation model receives the node features and edge features of the spatial graph. Design a graph convolution network with multi-layer graph convolution, update the node features through the message passing mechanism, and perform multi-layer graph convolution on the input node features to extract features. The covariance function used to calculate the Kriging weights is constructed based on the extracted features, and the convolution kernel parameters are dynamically calculated or adjusted to reflect the Kriging weights. The Kriging weights updated by the multi-layer graph convolution network are used to generate interpolation prediction values ​​through the Kriging interpolation method.

[0038] 3.1. Feature extraction of graph convolutional network: The graph convolutional network is based on the initial feature vector of the node As input, multi-layer graph convolution is used to update the node features. The feature update formula of each layer is:

[0039] ;

[0040] in, Is a node In the The characteristics of the layer, Is a node Updated features, It is The weight matrix of the layer, is the adjacency matrix after adding self-loops, is the adjacency matrix, is the identity matrix; yes The degree matrix of which the diagonal elements are the degrees of each node; is an activation function, for example, ReLU is used.

[0041] After multiple layers of graph convolution, the final feature of the node is ,in Describing graph convolutional networks, Representing the graph structure, these features capture high-order correlations and local nonlinear characteristics in spatial data.

[0042] In the graph convolution network, a pooling layer can be set after each layer of graph convolution to update the node features of the graph convolution operation. Perform global pooling to pool node features Aggregate to get global feature output In the present invention, maximum pooling is selected to perform global pooling operation.

[0043] 3.2 Adaptive estimation of covariance function: The covariance function is constructed using the features extracted by the graph convolutional network. The covariance function is represented by the radial basis function as follows:

[0044] ;

[0045] in, is the covariance function, , Respectively represent nodes and , , Respectively represent nodes and No. Features extracted after layer graph convolution; Representation Node and The distance between features. The covariance function defined in this way not only depends on the distance, but also can capture the nonlinear relationship between node features. Figure 2 shows the covariance function fitting curve of regional gravity anomalies using non-parametric methods.

[0046] 3.3. Kriging weight update: interpolation prediction value is a weighted linear combination of observations, expressed as:

[0047] ;

[0048] in, represents the interpolation prediction value of the unknown point, Represents known observation points The characteristic vector of Represents known observation points The kriging weights are Indicates the number of observation points.

[0049] Weight Determined by the minimum interpolation variance, satisfying the Kriging equation: ; Where C is the covariance matrix between known points, and c is the covariance vector between known points and predicted points.

[0050] The present invention uses graph convolutional networks to extract features and construct covariance functions. Based on the new covariance function , recalculate the weights in the kriging interpolation , the update of the Kriging weights is solved by the following equation: ; among them, among them, is the covariance matrix between the known observation points constructed, is the covariance vector between the known observation points and the predicted points, is the Kriging weight.

[0051] The convolution kernel parameters of the graph convolutional network can reflect the Kriging weights, use the Kriging weights to perform Kriging interpolation, and dynamically calculate or adjust the convolution kernel parameters to update the Kriging weights based on the principle of minimizing the loss function.

[0052] The graph convolution network uses weighted convolution. In each convolution operation, the output features of the graph convolution are Used to correct the Kriging interpolation weights, update the Kriging weights, and update the convolution kernel parameters through back propagation.

[0053] The weight update formula can be expressed as:

[0054] ;

[0055] in Is a linear mapping function used to adjust and update the Kriging weights , For Node No. The feature output after layer graph convolution, is the covariance vector between the known observation points and the predicted points. Features are obtained after layer graph convolution and global pooling aggregation , thereby constructing the covariance function to update the Kriging weights , by minimizing the loss function of Kriging interpolation, update the The convolution kernel parameters of the layer graph convolution are updated layer by layer. After the last layer graph convolution operation, the updated node features are Perform a global maximum pooling to obtain global features , calculate the final Kriging weights Used to generate interpolation results. This method updates the Kriging weights to the convolution kernel parameters to optimize the extraction of spatial features.

[0056] In the interpolation estimation model, the layer structure can be designed as needed, such as a fully connected layer or other layers suitable for the task.

[0057] 4. Model training: Use the collected feature data to train the interpolation estimation model. During the interpolation estimation model training process, forward propagate the input data through the graph convolutional network, calculate the predicted value of the Kriging gravity anomaly interpolation, and use the loss function to calculate the loss value of the model. According to the principle of minimizing the loss function, calculate the gradient through back propagation, and adjust and update the model parameters, including the convolution kernel parameters and Kriging weights, using the optimization algorithm. Adjust hyperparameters such as the learning rate, number of layers, and number of hidden units, and perform hyperparameter optimization to improve model performance.

[0058] 5. Prediction of gravity anomaly: Use the Kriging weights recalculated after model training optimization and adjustment to perform Kriging interpolation on unknown points to obtain the prediction results of gravity anomaly interpolation. The calculation formula is:

[0059] ;

[0060] in, represents the interpolation prediction value of the unknown point, Represents known observation points The characteristic vector of Represents known observation points The kriging weights are Indicates the number of observation points.

[0061] Figure 3 The gravity anomaly interpolation results of Kriging gravity anomaly based on graph convolutional network for the shown area.

[0062] The comparison of interpolation accuracy of gravity anomaly values ​​in irregular terrain areas is shown in the following table:

[0063]

[0064] As shown in the table above, the interpolation algorithm of the graph convolutional network is significantly better than ordinary Kriging interpolation and simple Kriging interpolation in the interpolation prediction of gravity anomalies, which proves that this method has a good effect in gravity anomaly interpolation in irregular terrain areas.

[0065] The above description of the specific implementation mode is only used to help understand the technical concept and core idea of ​​the present invention. Although the technical solution is described and illustrated using a specific preferred embodiment, it should not be understood as a limitation of the present invention itself. Those skilled in the art may make various changes in form and details without departing from the technical concept of the present invention. These easily conceived changes or substitutions should all be included in the protection scope of the present invention.

Claims

1. A Kriging gravity anomaly interpolation estimation method based on graph convolutional network, Features: The following steps are involved: (1) Collect characteristic data of gravity anomaly observation points. The characteristic data include spatial position characteristics of the observation points and observation characteristics that characterize the gravity anomaly of the observation points. The spatial position characteristics and observation characteristics constitute the characteristic vector of the observation point. (2) Each observation point is regarded as a node, the feature vector of the observation point is regarded as the node feature, and the correlation between nodes is regarded as the edge feature; the spatial graph of the graph convolutional network is constructed with the node features and edge features of each node; (3) Establishing an interpolation estimation model, wherein the interpolation estimation model takes the node features and edge features of the spatial graph as input, uses a graph convolution network to perform multi-layer graph convolution on the input to extract features, and uses the extracted features to construct a covariance function to update the Kriging weights. The updated Kriging weights are used to generate interpolation prediction values ​​through the Kriging interpolation method; (4) Use the collected feature data to train the interpolation estimation model, adjust the convolution kernel parameters of the graph convolution network by minimizing the loss function of the Kriging interpolation, and obtain the recalculated Kriging weights after model training and adjustment; (5) Use the Kriging weights recalculated after model training and adjustment to perform Kriging interpolation on the unknown points to obtain the prediction results of gravity anomaly interpolation.

2. A Kriging gravity anomaly interpolation estimation method based on graph convolutional network as claimed in claim 1, Features: The graph convolution network takes the initial feature vector of the node as input and uses multi-layer graph convolution to update the node features. The feature update formula of each layer is: ; in, Is a node In the The characteristics of the layer, Is a node Updated features, It is The weight matrix of the layer, is the adjacency matrix after adding self-loops, is the adjacency matrix, is the identity matrix; yes The degree matrix of which the diagonal elements are the degrees of each node; is the activation function.

3. The Kriging gravity anomaly interpolation estimation method based on graph convolutional network as claimed in claim 1, Features: In the graph convolutional network, a pooling layer is provided after each layer of graph convolution to perform global pooling on the updated node features.

4. The Kriging gravity anomaly interpolation estimation method based on graph convolutional network as claimed in claim 1, Features: The features extracted by the graph convolutional network are used to construct the covariance function represented by the radial basis function: ; in, is the covariance function, , Respectively represent nodes and , , Respectively represent nodes and No. Features extracted after layer graph convolution; Represents the distance between features.

5. The Kriging gravity anomaly interpolation estimation method based on graph convolutional network as claimed in claim 1, Features: The update of the kriging weights is solved using the following equation: ; in, is the covariance matrix between the known observation points constructed, is the covariance vector between the known observation points and the predicted points, is the Kriging weight.

6. The Kriging gravity anomaly interpolation estimation method based on graph convolutional network as claimed in claim 1, Features: The graph convolution network adopts weighted convolution, and the features extracted after each layer of graph convolution operation are used to update the Kriging weights.

7. The Kriging gravity anomaly interpolation estimation method based on graph convolutional network as claimed in claim 1, Features: In the step (1), the spatial position features include the latitude and longitude coordinates and elevation of the observation point, and the observation features include the east-west vertical line deviation value, the north-south vertical line deviation value and the crust density value of the observation point.

8. The Kriging gravity anomaly interpolation estimation method based on graph convolutional network as claimed in claim 1, Features: In the step (2), the edge feature adopts distance, weight or edge weight based on covariance matrix.

9. A Kriging gravity anomaly interpolation estimation method based on graph convolutional network as claimed in claim 1 or 5, Features: In the step (4), during the training of the interpolation estimation model, the input data is forward propagated through the graph convolutional network to calculate the predicted value of the Kriging gravity anomaly interpolation, and the loss value of the model is calculated according to the loss function, and the convolution kernel parameters of the graph convolutional network are updated through back propagation.

10. The Kriging gravity anomaly interpolation estimation method based on graph convolutional network as claimed in claim 1, Features: The Kriging gravity anomaly interpolation prediction value is calculated using a weighted linear combination of known observation points. The calculation formula is: ; in, represents the interpolation prediction value of the unknown point, Represents known observation points The characteristic vector of Represents known observation points The kriging weights are Represents the number of observation points.