Geochemical ore-prospecting method based on graph neural network and spatial features

Abstract

This invention relates to a geochemical prospecting method based on graph neural networks and spatial features, enabling geochemical prospecting under different geological environments based on accurately predicted background concentrations of different chemical elements. The method includes: preprocessing raw geochemical multi-element data and separating it from background data to obtain background area data; constructing spatial context statistical features for each sampling point of the geochemical multi-element data according to a preset spatial scale, and fusing them with the original environmental features of that sampling point to generate an enhanced environmental feature vector for each sampling point; constructing a graph topology based on the statistical correlation between chemical elements in the background area data using the maximum spanning tree algorithm; constructing a graph neural network based on the graph topology, inputting the enhanced environmental feature vector of each sampling point into the graph neural network, outputting the predicted background concentration of each chemical element at each sampling point, and calculating a mineral prediction map.

Description

Technical Field

[0001] This invention relates to the field of geological prediction and data processing technology, and in particular to a geochemical prospecting method based on graph neural networks and spatial features. Background Technology

[0002] Geochemical exploration is a core method for predicting and evaluating deep, concealed mineral resources. Its basic principle involves systematically collecting surface (such as soil and stream sediments) or subsurface rock samples, analyzing the spatial distribution patterns of trace chemical elements, and identifying geochemical anomalies formed by geological processes within deep ore bodies. With the decreasing availability of easily discoverable shallow deposits and the increasing depth of exploration, the surface geochemical background is becoming increasingly complex due to the influence of overburden, tectonic structures, and lithological variations. The traditional simple criterion of "high value equals anomaly" is no longer sufficient to meet the demands of refined deep mineral exploration.

[0003] In recent years, Convolutional Neural Networks (CNNs) and Graph Neural Networks (GNNs) have begun to be applied in this field. CNN-like methods typically require input data to be regular grid images. However, geochemical sampling data is inherently an irregularly distributed discrete point cloud. Forcing it to interpolate into a regular grid not only introduces artificial interpolation errors but also loses the microscopic distribution details of the original sampling points. Although conventional GNN methods can directly process point data in non-Euclidean space, most current research is based on constructing static geometric maps based on the Euclidean distance of sampling points, i.e., only considering spatial proximity. This approach ignores the inherent, dynamic genetic relationships between chemical elements (such as the symbiotic assemblage of ore-forming elements) and cannot truly reflect the internal mechanisms of geochemical systems.

[0004] A search of existing technical literature revealed a patent application (CN202511643985.4) entitled "Method for Predicting Mineralization in Deep Edges of Tungsten Deposits Based on Machine Learning." This patented method includes: acquiring multi-source geological data, constructing a multi-dimensional spatial grid, and dividing it into candidate prediction units; extracting quartz vein density, alteration index, geochemical assemblage anomaly weights, and high-frequency magnetic anomaly waveform features to construct a mineralization control feature tensor matrix; extracting spatial dependencies between candidate units based on graph neural networks to construct a mineralization potential prediction model, outputting a mineralization probability map and a mineralization hotspot map; and dynamically optimizing model parameters and feature weights through spatial overlay analysis and anomaly response identification mechanisms. While this patent can significantly improve the accuracy and efficiency of tungsten exploration in deep edges, it lacks the ability to dynamically adjust the correlation weights between elements according to the local geological environment, exhibiting a lack of explicit environmental adaptation mechanisms. Summary of the Invention

[0005] Therefore, it is necessary to provide a geochemical prospecting method based on graph neural networks and spatial features to address the above-mentioned technical problems, so as to realize geochemical prospecting based on the background content of different chemical elements in different geological environments.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a geochemical mineral exploration method based on graph neural networks and spatial features, the method comprising: S1: Preprocess the raw geochemical multi-element data to obtain geochemical multi-element data and separate the background data to obtain background area data; S2: Construct the spatial context statistical features of each sampling point of the geochemical multi-element data according to the preset spatial scale, and fuse them with the original environmental features of the sampling point to generate an enhanced environmental feature vector for each sampling point; S3: Based on the statistical correlation between chemical elements in the background area data, the maximum spanning tree algorithm is used to construct the graph topology; S4: Construct a graph neural network based on the graph topology, input the enhanced environmental feature vector of each sampling point into the graph neural network, and output the predicted background content of each chemical element in each sampling point. Calculate the mineral prediction map based on the predicted background content.

[0007] Preferably, the preprocessing includes: Data below the detection limit in the original geochemical multi-element data are filled to obtain the first geochemical multi-element data. The log1p function is used to perform a logarithmic transformation on the first geochemical multi-element data to obtain the second geochemical multi-element data; Robust normalization was used to standardize the second geochemical multi-element data to obtain the geochemical multi-element data.

[0008] Preferably, step S1 further obtains data on mineral-induced anomaly zones; the background data separation includes: If the Euclidean distance from the sampling point of the geochemical multi-element data to the known mineral deposit is greater than a preset distance threshold, the geochemical multi-element data is background area data; If the Euclidean distance from the sampling point of the geochemical multi-element data to the known mineral deposit is not greater than a preset distance threshold, the geochemical multi-element data is mineral-induced anomaly data.

[0009] Preferably, step S2 includes: The KD-Tree algorithm is used to retrieve and index the two-dimensional coordinates of all sampling points in the background area data. Based on the established retrieval index and the search radius of the preset spatial scale, a neighborhood point set search is performed on each sampling point of the geochemical multi-element data to obtain the neighborhood point set of each sampling point at different scales. Statistical moments are calculated for all chemical elements in the neighborhood point set of each sampling point at different scales to obtain the statistical characteristics of each sampling point at different scales. The statistical features of each sampling point at different scales and the original environmental features of that sampling point are concatenated into vectors to obtain the enhanced environmental feature vector of each sampling point.

[0010] Preferably, the statistical characteristics include spatial mean and spatial standard deviation.

[0011] Preferably, step S3 includes: Define a preset number of chemical elements as preset number of graph nodes, and construct a set of graph nodes; The Pearson correlation coefficients between all chemical elements in the background region data were calculated to obtain the Pearson correlation coefficient matrix and the corresponding statistical significance p-value matrix. Select the relevant edges in the statistical significance p-value matrix where the statistical significance p-value is less than a first preset threshold, and negative the absolute value of the corresponding Pearson correlation coefficient in the Pearson correlation coefficient matrix to obtain the statistically significant edges of the correlation. Based on the graph node set and the statistically significant edges, the maximum spanning tree algorithm is used to construct the basic skeleton of the graph, and the skeleton edges are obtained. Strongly correlated edges are selected based on the fact that the statistical significance p-values ​​in the statistical significance p-value matrix are less than a first preset threshold, and the Pearson correlation coefficients in the Pearson correlation coefficient matrix are greater than a second preset threshold; Add the strongly related edges to the skeleton edges, and add a self-loop to each graph node in the skeleton edges to obtain an unassigned graph topology; A two-dimensional attribute vector is assigned to each skeleton edge, each strongly correlated edge, and each self-loop in the unassigned graph topology to construct the graph topology. The attributes of the two-dimensional attribute vector include correlation strength and positive or negative correlation.

[0012] Preferably, the graph neural network includes an environment modulation module, a dynamic modulation module, and a main neighborhood aggregation convolution module; The environment modulation module uses a fully connected neural network to map the enhanced environment feature vector corresponding to each sampling point into a modulation vector, and the dimension of the modulation vector is the number of graph nodes in the graph node set; The dynamic modulation module multiplies the modulation component corresponding to each graph node in the modulation vector and the environmental feature vector corresponding to the enhanced environmental feature vector in each sampling point to obtain the modulated environmental feature corresponding to each graph node. The modulated environmental feature, element embedding and lithology embedding are vector concatenated to obtain the initial input feature corresponding to each graph node. The element embedding is the mapping vector of the chemical element corresponding to the graph node, and the lithology embedding is the mapping vector of the lithology code in the original environmental feature of the sampling point. The main neighborhood aggregation convolution module passes the initial input features corresponding to each graph node through several convolutional layers to obtain the final output features corresponding to each graph node. Then, it performs scalar mapping through the output linear layer to obtain the background content prediction value of each chemical element in the geochemical multi-element data sampled at each sampling point. Each convolutional layer passes the initial input features and aggregated features processed by the previous convolutional layer through linear transformation and nonlinear activation to obtain the initial input features to be processed by the next convolutional layer. The initial input features to be processed by the first convolutional layer are the initial input features processed by the dynamic modulation module. The initial input features processed by the last convolutional layer are the final output features corresponding to the graph node. The aggregated features are the features obtained by aggregating the mean, maximum, minimum, and standard deviation of the initial input features processed by the previous convolutional layer corresponding to all neighboring nodes in the graph topology.

[0013] Preferably, constructing a graph neural network based on the graph topology includes: constructing an initial graph neural network based on the graph topology; training the initial graph neural network to obtain a graph neural network; the training includes: Based on a preset spatial scale, the spatial context statistical features of each sampling point in the background area data are constructed, and then fused with the original environmental features of the sampling point to generate a background environmental feature vector for each sampling point. The background environment feature vector is divided into a training set and a validation set according to a preset ratio. The training set is input into the initial graph neural network for training using a preset loss function until the loss function converges, thus obtaining the graph neural network. During the training process, the validation set is used for hyperparameter tuning.

[0014] Preferably, the calculation expression of the preset loss function is as follows:

[0015] in, Indicates the loss value. This represents the set of sampling points corresponding to the training set after discarding a preset proportion, and SmoothL1 represents the SmoothL1 loss function. Indicates the first The sampling point of the nth sampling point Predicted background content values ​​for each chemical element Indicates the first The sampling point of the nth sampling point The true background content of each chemical element.

[0016] Preferably, the mineral prediction map is calculated based on the background content prediction value, including: The prediction residual vector is calculated based on the predicted and actual background content values ​​of each chemical element at each sampling point. The robust covariance matrix of the predicted residual vector is estimated using the minimum covariance determinant algorithm, and the Mahalanobis distance of each sampling point is calculated based on the robust covariance matrix. The node embedding vector of each sampling point in the graph neural network is extracted and spliced ​​with preset element embedding segments to obtain the fingerprint vector. Then, the cosine similarity between the fingerprint vector and the average fingerprint vector of the known mineral deposit samples is calculated to obtain the cosine similarity of each sampling point. The node embedding vector is composed of the final output features corresponding to all graph nodes in the sampling point. The Mahalanobis distance and cosine similarity of each sampling point are fused to obtain the mineral prediction map.

[0017] Compared with the prior art, the beneficial effects of the present invention are: This invention provides a geochemical prospecting method based on graph neural networks and spatial features. By searching the neighborhood point set of each sampling point, it captures spatial anomaly patterns from all angles. The graph neural network model is no longer limited to single-point values, but can "see" the geochemical field gradient around the sampling point. This enables the model to effectively identify anomalous halos with spatial continuity and significantly reduces false anomalies caused by single-point noise. Furthermore, the graph topology structure built based on Pearson correlation and statistical significance in the model can truly reflect the symbiotic combination of elements in nature. This makes anomaly detection no longer a black box operation, but a systematic deviation analysis with clear geochemical significance. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the geochemical mineral exploration method based on graph neural networks and spatial features in an embodiment of the present invention. Figure 2 This is a schematic diagram of the geochemical anomaly detection process in the application stage of the geochemical prospecting method based on graph neural networks and spatial features in an embodiment of the present invention. Figure 3 This is a schematic diagram of the geochemical anomaly detection process during the training phase of the geochemical prospecting method based on graph neural networks and spatial features in an embodiment of the present invention. Figure 4This is a schematic diagram of the AUC results of the gold and silver system in the geochemical prospecting method based on graph neural networks and spatial features in an embodiment of the present invention. Figure 5 This is a schematic diagram of the AUC results of the lead-zinc system based on graph neural networks and spatial features in an embodiment of the present invention. Figure 6 This is a schematic diagram of the baseline model A (PCA) results of the gold and silver system based on graph neural networks and spatial features in the geochemical prospecting method of the present invention. Figure 7 This is a schematic diagram of the baseline model A (PCA) results of the lead-zinc exploration system based on graph neural networks and spatial features in an embodiment of the present invention. Figure 8 This is a schematic diagram of the baseline model B (Vanilla GCN) of the gold and silver system based on graph neural networks and spatial features in an embodiment of the present invention. Figure 9 This is a schematic diagram of the baseline model B (Vanilla GCN) of the lead-zinc exploration system based on graph neural networks and spatial features in an embodiment of the present invention. Figure 10 This is a schematic diagram of the results of the model (EADM-GNN) of the gold and silver system based on graph neural networks and spatial features in the embodiment of the present invention. Figure 11 This is a schematic diagram of the model (EADM-GNN) of the lead-zinc mineral exploration system based on graph neural networks and spatial features in an embodiment of the present invention. Figure 12 This is a schematic diagram of the background area weight analysis results of a geochemical prospecting method for gold and silver deposits based on graph neural networks and spatial features in an embodiment of the present invention. Figure 13 This is a schematic diagram of the background area weight analysis results of a lead-zinc deposit based on a geochemical prospecting method using graph neural networks and spatial features, as described in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] Example 1 like Figure 1 As shown, this embodiment proposes a geochemical mineral exploration method based on graph neural networks and spatial features. The method includes: S1: Preprocess the raw geochemical multi-element data to obtain geochemical multi-element data and separate the background data to obtain background area data; The specific implementation of this step is as follows: In order for the model to accurately learn the geological and geochemical laws under the "mineral-free state", the original geochemical multi-element data is first subjected to strict background separation and preprocessing to suppress the interference of extreme outliers on the statistical distribution. Specifically, for the trace data below the limit of detection (LOD) in the original geochemical multi-element data, LOD√2 is used for filling to conform to the distribution law of geochemical data, resulting in the first geochemical multi-element data; the first geochemical data is transformed by Log1p(ln(x+1)) to eliminate the long-tail distribution effect of the data, resulting in the second geochemical multi-element data; the environmental characteristics (coordinates, fracture distance, magnetic values, spatial statistical characteristics) data in the second geochemical multi-element data are processed by robust normalization (Robust Scaler), and the calculation formula is X′=(X The median and interquartile range (IQR) are used to obtain geochemical multi-element data after preprocessing.

[0021] After preprocessing, calculate the Euclidean distance from all sampling points of the geochemical multi-element data to known ore deposits or mineral occurrences. Set a distance threshold Dth (e.g., 3000 meters or 5000 meters, depending on the actual situation), define data in areas with a distance greater than Dth as background data, and the remaining areas as mineral-induced anomaly data.

[0022] S2: Construct the spatial context statistical features of each sampling point of the geochemical multi-element data according to the preset spatial scale, and fuse them with the original environmental features of the sampling point to generate an enhanced environmental feature vector for each sampling point; The specific implementation of this step is as follows: This step constructs a multi-scale geochemical field gradient sensing encoder, aiming to endow discrete sampling points with spatial sensing capabilities and capture the gradient information of local geochemical fields. This encoder utilizes a spatial indexing algorithm to extract the geochemical field distribution characteristics of the sampling point's neighborhood at multiple preset spatial scales. By aggregating local first moments (spatial mean) and second moments (spatial standard deviation), it generates a high-dimensional sensing vector containing spatial gradient information (in this invention, "gradient" refers to the degree of dispersion of element content within the neighborhood, approximated by the standard deviation). Specifically: First, the KD-Tree algorithm is used to establish a fast retrieval index for the two-dimensional coordinates (X,Y) of all sampling points in the background area data. At the same time, a set of multi-level search radii R={r1,r2,…,rn} (e.g., 500m, 1000m, 2000m, depending on the actual situation) are set to represent the near-ore halo, the outer halo and the regional background scale, respectively. Based on the established retrieval index and the search radius of the preset spatial scale, a neighborhood point set search is performed on each sampling point of the geochemical multi-element data. For any sampling point Pi, its neighborhood point set N(Pi,rk) is searched under each radius rk. The statistical moments of each chemical element within the neighborhood point set are calculated to obtain the statistical characteristics of each sampling point at different scales. These statistical characteristics include the spatial mean and spatial standard deviation. The spatial mean reflects the local background level at that scale, while the spatial standard deviation reflects the degree of variation in the local geochemical field (mineralization is often accompanied by high variability). The statistical features of the n scales calculated above are concatenated with the original environmental features (coordinates, fracture distance, magnetic anomaly, lithological coding) of the point to generate a high-dimensional enhanced environmental feature vector E_aug for each sampling point.

[0023] S3: Based on the statistical correlation between chemical elements in the background area data, the maximum spanning tree algorithm is used to construct the graph topology; The specific implementation of this step is as follows: This step constructs a statistically-constrained adaptive topology generator, aiming to realistically reflect the causal relationships between elements and provide the underlying data structure for subsequent graph neural networks. Specifically: First, define the N chemical elements as N graph nodes, and construct a graph node set V; Based on the background area data generated in step S1, calculate the Pearson correlation coefficient matrix C between its elements and its corresponding statistical significance p-value matrix P; Skeleton edge construction: Only edges with statistically significant correlations (i.e., statistical significance p-value less than the first preset threshold (p<0.05)) are retained. The absolute value of their Pearson correlation coefficients is negative. Based on the graph node set and the statistically significant edges, the maximum spanning tree algorithm is used to construct the basic skeleton of the graph, ensuring that all element nodes are connected. Strong correlation enhancement: Based on the MST skeleton edges, additional strong correlation edges are added that satisfy the statistical significance p-value being less than the first preset threshold and the Pearson correlation coefficient being greater than the second preset threshold (|Cij|>Tcorr and Pij<0.05) to capture statistically significant strong symbiotic relationships; Adding self-loops: Add a self-loop to each graph node in the skeleton edge (the self-loop's property is fixed at [1.0, 1.0]), resulting in an unassigned graph topology; Edge attribute assignment: For each edge (including skeleton edges, strongly correlated edges, and self-loops) in the unassigned edge set E of the graph topology, a two-dimensional attribute vector is assigned: [|C_{ij}|,sign(C_{ij})], which represents the association strength and positive / negative correlation, respectively.

[0024] The resulting graph topology features edges with absolute weights and signs (positive / negative correlation), forming dynamic edge indices and edge attributes. The edge index defines which edges connect to which edges, while the edge attribute defines the correlation strength and sign. Together, they define how nodes exchange information.

[0025] S4: Construct a graph neural network based on the graph topology, input the enhanced environmental feature vector of each sampling point into the graph neural network, and output the predicted background content of each chemical element in each sampling point. Calculate the mineral prediction map based on the predicted background content.

[0026] The specific implementation of this step is as follows: This step constructs an Environment-Aware Dynamic Modulation Graph Neural Network (EADM-GNN). This graph neural network takes the enhanced environment feature vector E_aug generated in step S2 as input, uses the dynamic graph topology G=(V,E) generated in step S3 for information transmission, and finally outputs the predicted background content values ​​for each element. This network architecture mainly includes the following three core sub-modules: Environment Modulation Module: This module aims to simulate the nonlinear control of macroscopic geological environments (such as lithology and structure) on the geochemical behavior of microscopic elements. For each sampling point i, after inputting an enhanced environmental feature vector E_aug,i with dimension D_env, a fully connected multi-layer perceptron (MLP) is used to map the enhanced environmental feature vector E_aug,i into a modulation vector M_i with dimension N, where N is the number of chemical elements (graph nodes) in the graph node set V. The calculation formula is as follows: M_i=tanh(W_m·E_aug,i+b_m) W_m and b_m are learnable weight parameters, and the tanh function restricts the output to the (-1,1) interval to represent positive promotion or negative inhibition.

[0027] Dynamic Modulation Module: For the j-th element node (j=1...N) in the graph topology corresponding to each sampling point, the j-th component of the modulation vector M_i is used as a scaling factor and multiplied by the corresponding environmental feature vector in the enhanced environmental feature vector, thereby achieving differentiated regulation of the environment's response to elements. The modulated environmental feature is then concatenated with the learnable element embedding and lithology embedding to form the initial feature Hj(0) of the graph node. The calculation formula is as follows: Eenv(j)=Eenv(j)*Mi[j], Hj(0)=[Eenv(j);Embelem(j);Emblitho] Through this step, the network generates an initial input feature H^(0) with environment specificity for each graph node (chemical element) of each sampling point, wherein the element embedding is the mapping vector of the chemical element corresponding to the graph node, and the lithology embedding is the mapping vector of the lithology encoding in the original environmental features of the sampling point.

[0028] The Principal Neighbourhood Aggregation (PNA) Convolutional Module employs multiple stacked Principal Neighbourhood Aggregation (PNA) graph convolutional layers to extract non-linear correlation features between elements. Unlike traditional GCNs, PNA combines multiple aggregators to capture high-order statistical information of neighborhood distributions. For any node j, its initial input feature H^(0) needs to pass through L convolutional layers (L=4 in this embodiment). In each convolutional layer, the convolutional layer aggregates the initial features of its neighboring nodes k∈N(j) (the initial features H^(0) of all graph nodes (including neighboring nodes) are generated through the above dynamic modulation process, only the element embeddings and modulation coefficients are different). The mean, maximum, minimum, and standard deviation of the neighbor feature set N(j) are statistically analyzed. After concatenating the four statistics, the aggregated features are obtained, and the calculation expression is as follows: Agg_j=[Mean(N(j)),Max(N(j)),Min(N(j)),Std(N(j))] Combining the initial input features and edge attributes of the graph node itself, the initial input features H^(l+1) of the graph node in the next convolutional layer are updated through linear transformation and nonlinear activation (such as ELU). In subsequent convolutional layers, the initial input features of the graph node are continuously updated. H^(0) is gradually updated to H^(1), H^(2), …, H^(L) through multiple convolutions. After L layers of propagation, the final feature H^(L) of each graph node is obtained. The initial input features of neighboring nodes are also the result of the previous update. Taking the first convolutional layer as an example (l=0): each graph node j has an initial input feature Hj(0) (hidden shape). The initial features {Hk(0)} of all its neighboring nodes k∈(j) are collected. The aggregation result is calculated: Aggj(0)=[Mean({Hk(0)}),Max({Hk(0)}),Min({Hk(0)}),Std({Hk(0)})] This is a 4×hidden vector (four statistics concatenated; PNAConv has various internal combinations, typically including summation and concatenation, but the final output dimension remains hidden). The aggregation result, along with the initial feature Hj(0) of the current graph node (or edge attribute information), undergoes a learnable linear transformation and is then passed through an activation function to obtain new features: Hj(1)=σ(W·Comb(Hj(0),Aggj(0),edge_attr)) Repeat the above process for L layers to obtain the final node features Hj(L). Input the final node features Hj(L) into an output linear layer nn.Linear(hidden,1). This linear layer acts independently on each graph node, mapping the features of each node to a scalar. For a sampling point, the model will output N scalar prediction values, corresponding to the background content prediction values ​​of N chemical elements.

[0029] Example 2 This embodiment further illustrates the geochemical prospecting method based on graph neural networks and spatial features proposed in Embodiment 1.

[0030] like Figure 2 As shown, the mineral prediction map is calculated based on the background content prediction value, including: The prediction residual vector is calculated based on the predicted and actual background content values ​​of each chemical element at each sampling point. The robust covariance matrix of the predicted residual vector is estimated using the minimum covariance determinant algorithm, and the Mahalanobis distance of each sampling point is calculated based on the robust covariance matrix. The node embedding vector of each sampling point in the graph neural network is extracted and spliced ​​with preset element embedding segments to obtain the fingerprint vector. Then, the cosine similarity between the fingerprint vector and the average fingerprint vector of the known mineral deposit samples is calculated to obtain the cosine similarity of each sampling point. The node embedding vector is composed of the final output features corresponding to all graph nodes in the sampling point. The Mahalanobis distance and cosine similarity of each sampling point are fused to obtain the mineral prediction map.

[0031] In the specific implementation process, after using the trained EADM-GNN model to predict the geochemical multi-element data of the entire region (including background area data and mineral-induced anomaly area data), the following dual indicators are also used to determine anomalies: Robust Mahalanobis Score: Calculates the prediction residual vector Res=Y_true for each sampling point of the entire data area. Y_pred is calculated from the predicted and actual background concentrations of each chemical element at each sampling point. The robust covariance matrix Σ of the predicted residual vector is estimated using the Min CovDet algorithm, and the Mahalanobis distance for each sampling point is calculated. Collinearity redundancy among elements is eliminated (e.g., lead and zinc often co-occur), quantifying the degree to which the sample deviates from the background statistical regularity.

[0032] Fingerprint Similarity: Extract node embeddings from the penultimate layer (before the output linear layer) of the graph neural network model. The node embeddings are composed of the final output features corresponding to all graph nodes at the sampling point. For a specific mineralization system (e.g., the gold-silver system), select the embedding fragments corresponding to its core indicator elements (e.g., gold, arsenic, antimony) and concatenate them to form the fingerprint vector. Calculate the cosine similarity between the fingerprint vector of the test point and the average fingerprint vector of known ore deposit samples. This verifies whether the anomaly has "mineralization properties" and effectively eliminates non-mineralized false anomalies caused by high stratigraphic background.

[0033] The two indicators mentioned above are combined and calculated using the formula Score Final = Score Mahal × Similarity, which outputs the final mineralization prediction map.

[0034] Example 3 This embodiment further illustrates the geochemical mineral exploration method based on graph neural networks and spatial features proposed in the above embodiments.

[0035] Constructing a graph neural network based on the graph topology includes: constructing an initial graph neural network based on the graph topology; training the initial graph neural network to obtain a new graph neural network; as follows: Figure 3 As shown, the training includes: Based on a preset spatial scale, the spatial context statistical features of each sampling point in the background area data are constructed, and then fused with the original environmental features of the sampling point to generate a background environmental feature vector for each sampling point. The background environment feature vector is divided into a training set and a validation set according to a preset ratio. The training set is input into the initial graph neural network for training using a preset loss function until the loss function converges, thus obtaining the graph neural network. During the training process, the validation set is used for hyperparameter tuning.

[0036] In the specific implementation process, the model training task is background reconstruction, that is, using the input "environment + neighborhood" information to predict the element content that the sampling point should have in a mineral-free state. During the model training phase, only the background area data defined in step S1 is input into the model for forward propagation. Specifically, based on the retrieval index established using the KD-Tree algorithm for the two-dimensional coordinates of all sampling points in the background area data, a neighborhood point set search is performed on each sampling point in the background area data using a search radius of a preset spatial scale, obtaining the background neighborhood point set of each sampling point at different scales. Statistical moments are calculated for all chemical elements within the background neighborhood point set of each sampling point at different scales to obtain the background statistical features of each sampling point at different scales. The background statistical features of each sampling point at different scales and the original environmental features of that sampling point are concatenated as vectors to obtain the background environmental feature vector of each sampling point. The training strategy uses a spatial grid to divide the background environmental feature vector (8km grid, 20% as validation set, 80% as training set) to prevent spatial autocorrelation overfitting. The training set is input into the initial graph neural network using a preset loss function for training. During training, for each sampling point (graph), the SmoothL1 loss of all its element graph nodes is first calculated, and then the average value is taken as the loss of that sampling point. Within each batch, the 10% of sampling points with the largest loss are discarded (Trim Loss) to enhance robustness, and the average value of the losses of the remaining sampling points is taken as the batch loss. The loss of the entire training set is the average of the batch losses. The optimization objective is to minimize this loss. The expression for calculating the loss function is as follows:

[0037] in, Indicates the loss value. This represents the set of sampling points corresponding to the training set after discarding a preset proportion, and SmoothL1 represents the SmoothL1 loss function. Indicates the first The sampling point of the nth sampling point Predicted background content values ​​for each chemical element Indicates the first The sampling point of the nth sampling point The true background content of each chemical element.

[0038] To verify the effectiveness and universality of the method of this invention under different metallogenic systems and complex geological backgrounds, this embodiment selects the gold-silver and lead-zinc metallogenic systems in a typical mineralized area for dual verification. Experimental data include: Study area: Contains approximately N=7236 sampling points, covering an area of ​​approximately 1694 square kilometers; Sampling elements: 16 elements including gold, silver, copper, lead, zinc, arsenic, antimony, and mercury were analyzed; Background set partitioning: Set a distance threshold of 5000 meters and partition 80% of the data that are far from known mining sites into the training set.

[0039] To highlight the superiority of the technical solution of this invention, the following three sets of comparative experiments were set up: Baseline Model A (PCA-Mahalanobis): A traditional multivariate statistical method that uses only Principal Component Analysis (PCA) combined with Mahalanobis distance, without including spatial features or deep learning modules. This paper constructs an anomaly detection model based on Principal Component Analysis (PCA). First, the background dataset is subjected to PCA dimensionality reduction, retaining 90% of the variance contribution rate. Then, the Minimum Covariance Determinant (MCD) algorithm is used to estimate the robust mean and covariance matrix of the background in low-dimensional space. Finally, the Mahalanobis distance from all samples to the background center is calculated as the anomaly score. This model does not utilize any spatial neighborhood information and does not consider the nonlinear effects of environmental factors such as lithology.

[0040] Baseline Model B (Vanilla GCN): Uses a standard Graph Convolutional Network (GCN), with the graph structure built based on a fixed Euclidean distance (static graph), and lacks multi-scale spatial feature fusion and environment modulation mechanisms. Input Features: Only uses the original magnetic properties and coordinates of the sampling points, without concatenating the multi-scale spatial gradient features proposed in this method. Graph Structure: Employs a static fully connected graph structure, assuming equal-strength associations between all elements, and does not apply the MST topology optimization and p-value significance screening strategies of this invention. Network Architecture: Removes the environment modulation layer, directly inputting features into the standard GCN convolutional layer, and does not employ the PNA multi-aggregator mechanism. This model aims to verify the core contributions of the dynamic graph structure and environment modulation mechanism.

[0041] The method model of this invention (EADM-GNN) adopts the complete technical solution described in this case, including multi-scale gradient coding, dynamic MST graph structure and environmental modulation mechanism.

[0042] The final experimental results are compared below: (1) Comparison of anomaly detection accuracy (AUC-ROC) To quantitatively evaluate the ability of each model to identify known mineral deposits, this embodiment uses the Receiver Operating Characteristic (ROC) curve and its area under the curve (AUC) as the core evaluation indicators. Ground Truth Definition: First, the Euclidean distance from all sampling points within the study area to the nearest known mineral deposit / mineral point is calculated. A distance threshold Dhit is set (500 meters in this embodiment). Positive Sample (labeled 1): Sampling points less than Dhit from the known mineral point represent "mineral-induced anomalies" that the model should identify. Negative Sample (labeled 0): Sampling points greater than Dhit from the known mineral point represent unmineralized background or unknown areas. ROC Curve Construction: The anomaly score output by each model is used as the criterion. A series of True Positive Rate (TPR) and False Positive Rate (FPR) pairs are calculated using a sliding classification threshold. TPR (vertical axis): Measures how many genuine mining sites the model "captured" (recall rate). FPR (horizontal axis): Measures how many background points the model "falsely reported as mining sites" (false positive rate). Physical meaning of AUC value: AUC values ​​range from 0 to 1. A higher AUC value indicates that the model can more accurately identify known mining sites while maintaining a low false positive rate. If AUC = 0.5, it means the model's predictive ability is equivalent to random guessing; if AUC > 0.9, it means the model has extremely high anomaly detection accuracy. Figure 4 , Figure 5 As shown, the AUC of this model is significantly higher than that of traditional methods and ordinary GCN, proving that the introduction of multi-scale spatial features and environmental modulation mechanisms can effectively improve the identification accuracy of mineral-induced anomalies.

[0043] (2) Spatial morphology description and convergence verification like Figure 6 , Figure 7 As shown, in both the gold-silver and lead-zinc systems, the anomaly maps generated by the baseline model A (PCA) only utilize numerical statistical features, displaying only high-value clusters and failing to utilize spatial neighborhood information. The anomaly points near the deposits lack spatial continuity. For example... Figure 8 , Figure 9 As shown, after introducing graph convolution into the baseline model B (Vanilla GCN), the outliers near the ore deposit exhibit some clustering characteristics, but they remain relatively blurry and relatively noisy. Figure 10 , Figure 11As shown, thanks to multi-scale spatial gradient encoding, the anomaly patterns generated by this model (EADM-GNN) exhibit extremely high spatial convergence. Whether in the gold-silver or lead-zinc system, the top 5% of high anomalies are precisely focused within known deposits (hollow circles), forming dense and clear concentration centers, and the anomaly boundaries highly coincide with the trend of the geological structural zones.

[0044] (3) Verification of environmental background suppression and disambiguation capabilities Technical challenges: The southeastern part of the study area contains typical high geochemical background strata with high original lead and zinc abundance, which can easily interfere with prediction.

[0045] Comparative analysis: such as Figure 7 , Figure 9 As shown, in the lead-zinc system experiment, baseline model A (PCA) and baseline model B (Vanilla GCN) mistakenly identified the high background strata in the middle and lower parts as large-area, continuous strong anomalies, resulting in serious false positives. Our model, through an Environment Modulation Layer, successfully learned the lithological and environmental characteristics of this region as "high background, non-mineralized." In the EADM-GNN results map, the false high values ​​in the southeast were effectively decomposed and suppressed (transformed into low background values), and the model successfully distinguished between "high stratigraphic background" and "mineralized anomalies." This result strongly demonstrates the significant advantages of the method of this invention in handling geological background non-stationarity.

[0046] (4) Verification of the geological interpretability of the model To verify the effectiveness of the proposed "environment modulation layer" and "multi-scale spatial feature fusion" mechanism, this embodiment performs weight analysis on the trained gold-silver system model and lead-zinc system model. The weight matrix of the environment modulation module is extracted, and the sensitivity of each element to spatial environmental features at different scales is analyzed.

[0047] like Figure 12As shown, in the gold-silver system, the background content of gold is significantly controlled by SP_STD_HG_3000 (the volatility of mercury within 3000 meters, weight +0.1544) and SP_MEAN_SN_2000 (the mean of tin within 2000 meters, weight +0.1681). This aligns with geological principles: mercury, as a typical low-temperature frontal halo element, indicates the extent of hydrothermal activity; while tin, as a high-temperature element, indicates the center of deep magmatic heat sources. The model successfully captures the joint control of the "deep magma (tin) - shallow hydrothermal (mercury)" mineralization system on the gold background field. The model exhibits significant differences in spatial scale sensitivity to different elements. For example, antimony is mainly controlled by local autocorrelation characteristics at a small scale (500-1000 meters) (weight +0.1646), reflecting its local enrichment habit controlled by fault structures; while boron shows strong sensitivity to the arsenic background field at a large scale (3000 meters) (weight -0.1931). This result strongly demonstrates the necessity of introducing multi-scale spatial features in this method: single-scale convolution cannot simultaneously capture the distribution patterns of locally enriched elements (such as antimony) and regional background elements (such as boron), while this method achieves a panoramic perception of complex geochemical fields by fusing multi-level features from 500m to 3000m.

[0048] like Figure 13 As shown, in the lead-zinc system, the background zinc content exhibits a significant negative correlation weight (-0.1897) with SP_MEAN_ZN_1000 (the local average zinc concentration within a 1000-meter range), while being positively modulated by SP_MEAN_HG_5000 (the average mercury concentration within a 5000-meter range). This indicates that in the background region, the distribution of zinc has strong spatial autocorrelation and is controlled by the background field of large-scale low-temperature hydrothermal activity (mercury halo). The gold content in the background region is not independent and random, but is positively correlated with SP_MEAN_ZN_1000 (+0.2066). This implies that in a mineral-free background, gold and zinc have a certain symbiotic background relationship. The above weight ranking confirms that the EADM-GNN model of this invention does not perform simple numerical fitting, but successfully learns the control law of background geological environment (such as the range of hydrothermal activity and element symbiotic combination) on element abundance. When the model is applied to unknown regions, any combination of elements that violates the above-mentioned "background control law" (e.g., extremely high gold content but lack of corresponding zinc background support) will be identified by the model as a significant geochemical anomaly.

[0049] (5) Conclusion Experimental results from both gold-silver and lead-zinc systems demonstrate that the geochemical anomaly detection prospecting method based on dynamic graph neural networks and multi-scale spatial feature fusion proposed in this invention possesses excellent generalization capabilities. In terms of prospecting accuracy, it can more precisely characterize anomaly areas; in terms of interference resistance, it can effectively eliminate false anomalies caused by high background strata by utilizing environmental modulation mechanisms. Compared to existing technologies, this method significantly improves the signal-to-noise ratio, providing a more reliable and interpretable scientific basis for exploration.

Claims

1. A geochemical prospecting method based on graph neural networks and spatial features, characterized in that, include: S1: Preprocess the raw geochemical multi-element data to obtain geochemical multi-element data and separate the background data to obtain background area data; S2: Construct the spatial context statistical features of each sampling point of the geochemical multi-element data according to the preset spatial scale, and fuse them with the original environmental features of the sampling point to generate an enhanced environmental feature vector for each sampling point; S3: Based on the statistical correlation between chemical elements in the background area data, the maximum spanning tree algorithm is used to construct the graph topology; S4: Construct a graph neural network based on the graph topology, input the enhanced environmental feature vector of each sampling point into the graph neural network, and output the predicted background content of each chemical element in each sampling point. Calculate the mineral prediction map based on the predicted background content.

2. The geochemical prospecting method based on graph neural networks and spatial features according to claim 1, characterized in that, The preprocessing includes: Data below the detection limit in the original geochemical multi-element data are filled to obtain the first geochemical multi-element data. The log1p function is used to perform a logarithmic transformation on the first geochemical multi-element data to obtain the second geochemical multi-element data; Robust normalization was used to standardize the second geochemical multi-element data to obtain the geochemical multi-element data.

3. The geochemical prospecting method based on graph neural networks and spatial features according to claim 2, characterized in that, Step S1 also yields data on the mineral-induced anomaly zone; The background data separation includes: If the Euclidean distance from the sampling point of the geochemical multi-element data to the known mineral deposit is greater than a preset distance threshold, the geochemical multi-element data is background area data; If the Euclidean distance from the sampling point of the geochemical multi-element data to the known mineral deposit is not greater than a preset distance threshold, the geochemical multi-element data is mineral-induced anomaly data.

4. The geochemical prospecting method based on graph neural networks and spatial features according to claim 1, characterized in that, Step S2 includes: The KD-Tree algorithm is used to retrieve and index the two-dimensional coordinates of all sampling points in the background area data. Based on the established retrieval index and the search radius of the preset spatial scale, a neighborhood point set search is performed on each sampling point of the geochemical multi-element data to obtain the neighborhood point set of each sampling point at different scales. Statistical moments are calculated for all chemical elements in the neighborhood point set of each sampling point at different scales to obtain the statistical characteristics of each sampling point at different scales. The statistical features of each sampling point at different scales and the original environmental features of that sampling point are concatenated into vectors to obtain the enhanced environmental feature vector of each sampling point.

5. The geochemical prospecting method based on graph neural networks and spatial features according to claim 4, characterized in that, The statistical characteristics include spatial mean and spatial standard deviation.

6. The geochemical prospecting method based on graph neural networks and spatial features according to claim 1, characterized in that, Step S3 includes: Define a preset number of chemical elements as preset number of graph nodes, and construct a set of graph nodes; The Pearson correlation coefficients between all chemical elements in the background region data were calculated to obtain the Pearson correlation coefficient matrix and the corresponding statistical significance p-value matrix. Select the relevant edges in the statistical significance p-value matrix where the statistical significance p-value is less than a first preset threshold, and negative the absolute value of their corresponding Pearson correlation coefficients in the Pearson correlation coefficient matrix to obtain the statistically significant edges of the correlation. Based on the graph node set and the statistically significant edges, the maximum spanning tree algorithm is used to construct the basic skeleton of the graph, and the skeleton edges are obtained. Strongly correlated edges are selected based on the fact that the statistical significance p-values ​​in the statistical significance p-value matrix are less than a first preset threshold, and the Pearson correlation coefficients in the Pearson correlation coefficient matrix are greater than a second preset threshold; Add the strongly related edges to the skeleton edges, and add a self-loop to each graph node in the skeleton edges to obtain an unassigned graph topology; A two-dimensional attribute vector is assigned to each skeleton edge, each strongly correlated edge, and each self-loop in the unassigned graph topology to construct the graph topology. The attributes of the two-dimensional attribute vector include correlation strength and positive or negative correlation.

7. The geochemical prospecting method based on graph neural networks and spatial features according to claim 6, characterized in that, The graph neural network includes an environment modulation module, a dynamic modulation module, and a main neighborhood aggregation convolution module; The environment modulation module uses a fully connected neural network to map the enhanced environment feature vector corresponding to each sampling point into a modulation vector, and the dimension of the modulation vector is the number of graph nodes in the graph node set; The dynamic modulation module multiplies the modulation component corresponding to each graph node in the modulation vector and the environmental feature vector corresponding to the enhanced environmental feature vector in each sampling point to obtain the modulated environmental feature corresponding to each graph node. The modulated environmental feature, element embedding and lithology embedding are vector concatenated to obtain the initial input feature corresponding to each graph node. The element embedding is the mapping vector of the chemical element corresponding to the graph node, and the lithology embedding is the mapping vector of the lithology code in the original environmental feature of the sampling point. The main neighborhood aggregation convolution module passes the initial input features corresponding to each graph node through several convolutional layers to obtain the final output features corresponding to each graph node. Then, it performs scalar mapping through the output linear layer to obtain the background content prediction value of each chemical element in the geochemical multi-element data sampled at each sampling point. Each convolutional layer passes the initial input features and aggregated features processed by the previous convolutional layer through linear transformation and nonlinear activation to obtain the initial input features to be processed by the next convolutional layer. The initial input features to be processed by the first convolutional layer are the initial input features processed by the dynamic modulation module. The initial input features processed by the last convolutional layer are the final output features corresponding to the graph node. The aggregated features are the features obtained by aggregating the mean, maximum, minimum, and standard deviation of the initial input features processed by the previous convolutional layer corresponding to all neighboring nodes in the graph topology.

8. The geochemical prospecting method based on graph neural networks and spatial features according to claim 1, characterized in that, The process of constructing a graph neural network based on the graph topology includes: constructing an initial graph neural network based on the graph topology; and training the initial graph neural network to obtain a graph neural network; the training includes: Based on a preset spatial scale, the spatial context statistical features of each sampling point in the background area data are constructed, and then fused with the original environmental features of the sampling point to generate a background environmental feature vector for each sampling point. The background environment feature vector is divided into a training set and a validation set according to a preset ratio. The training set is input into the initial graph neural network for training using a preset loss function until the loss function converges, thus obtaining the graph neural network. During the training process, the validation set is used for hyperparameter tuning.

9. The geochemical prospecting method based on graph neural networks and spatial features according to claim 8, characterized in that, The calculation expression for the preset loss function is as follows: in, Indicates the loss value. This represents the set of sampling points corresponding to the training set after discarding a preset proportion, and SmoothL1 represents the SmoothL1 loss function. Indicates the first The sampling point of the nth sampling point Predicted background content values ​​for each chemical element Indicates the first The sampling point of the nth sampling point The true background content of each chemical element.

10. The geochemical prospecting method based on graph neural networks and spatial features according to claim 7, characterized in that, The mineral prediction map is calculated based on the background content prediction value, including: The prediction residual vector is calculated based on the predicted and actual background content values ​​of each chemical element at each sampling point. The robust covariance matrix of the predicted residual vector is estimated using the minimum covariance determinant algorithm, and the Mahalanobis distance of each sampling point is calculated based on the robust covariance matrix. The node embedding vector of each sampling point in the graph neural network is extracted and spliced ​​with preset element embedding segments to obtain the fingerprint vector. Then, the cosine similarity between the fingerprint vector and the average fingerprint vector of the known mineral deposit samples is calculated to obtain the cosine similarity of each sampling point. The node embedding vector is composed of the final output features corresponding to all graph nodes in the sampling point. The Mahalanobis distance and cosine similarity of each sampling point are fused to obtain the mineral prediction map.

Images