A green land metallogenic prediction method based on a dynamic confusion matrix

Abstract

The present application belongs to the technical field of mineral exploration, and relates to a green land metallogenic prediction method based on a dynamic confusion matrix, comprising: preprocessing multi-source data to generate evidence layers; performing data enhancement through a synthetic minority over-sampling technique (SMOTE) algorithm to obtain multiple training sets and multiple validation sets; training an AI prediction model according to the multiple training sets; obtaining a metallogenic prediction result; constructing a dynamic confusion matrix corresponding to each classification threshold according to the metallogenic prediction result, a preset classification threshold interval and a gradient interval; identifying an optimal classification threshold of the AI prediction model based on an evaluation index dynamic curve and an area under the evaluation index dynamic curve; and obtaining a high-potential target area according to the optimal classification threshold of the AI prediction model. The present application solves the problem of low accuracy of existing green land metallogenic prediction methods through multi-source data, the SMOTE algorithm, a dynamic confusion matrix, a concentration area model and the like.

Description

Technical Field

[0001] This invention relates to the field of mineral exploration technology, and specifically discloses a greenfield mineralization prediction method based on a dynamic confusion matrix. Background Technology

[0002] Greenland mineralization prediction is mostly designed for small sample scenarios in greenlands, and usually faces problems such as few known deposits and scarce data. Traditional greenland mineralization prediction methods often use static confusion matrices composed of fixed classification thresholds (such as a classification threshold of 0.5) for mineralization prediction. This can easily lead to overestimation of the performance of the prediction model in small sample scenarios, and it cannot capture the continuous changes of the prediction model under different classification thresholds, resulting in low accuracy of mineralization prediction. Summary of the Invention

[0003] The purpose of this invention is to provide a greenland mineralization prediction method based on a dynamic confusion matrix, thereby solving the problem of low accuracy in existing greenland mineralization prediction methods.

[0004] The specific solution of the present invention is as follows: A greenfield mineralization prediction method based on a dynamic confusion matrix includes: Collect multi-source data and preprocess the multi-source data to generate evidence layers; Multiple known mining locations were obtained based on the evidence layer; based on the multiple known mining locations, data augmentation was performed using a synthetic minority class oversampling algorithm to obtain multiple training sets and multiple validation sets; The AI ​​prediction model is trained using multiple training sets; Based on multiple validation sets, mineralization prediction results are obtained through a trained AI prediction model. Based on the mineralization prediction results, a dynamic confusion matrix corresponding to each classification threshold is constructed through a preset classification threshold range and gradient interval. Each dynamic confusion matrix is ​​evaluated to generate dynamic curves and areas under the dynamic curves of the evaluation indicators under the AI ​​prediction model; the optimal classification threshold of the AI ​​prediction model is identified based on the dynamic curves and areas under the dynamic curves of the evaluation indicators under the AI ​​prediction model. High-potential target areas are obtained based on the optimal classification threshold of the AI ​​prediction model, and the rationality of the high-potential target areas is verified.

[0005] Preferably, multiple training sets and multiple validation sets are obtained, including: S1. Based on a known mining point, obtain k nearest neighbor samples using Euclidean distance; S2. Starting from a known mineral point, move a random proportion of the distance along the direction pointing to a nearest neighbor sample to generate a synthetic mineral point. S3. Repeat S1-S2 until all neighboring samples of all known mineral points are traversed to obtain multiple synthetic mineral points. Generate positive samples based on the multiple synthetic mineral points and multiple known mineral points. S4. Based on the positive samples, randomly select the same number of non-mining points from the evidence layer as the positive samples to generate negative samples; S5. Generate a balanced dataset based on positive and negative samples, and divide the balanced dataset into a training set and a validation set. S6. Repeat S1-S5 until the preset number of times is reached to obtain multiple training sets and multiple validation sets.

[0006] Preferred AI prediction models include: deep forest, random forest, convolutional neural network, and support vector machine.

[0007] Preferably, the AI ​​prediction model is trained based on multiple training sets, including: Based on multiple training sets, the parameters of the AI ​​prediction model are optimized using random grid search and ten-fold cross-validation to obtain a well-trained AI prediction model.

[0008] Preferably, the classification threshold range is 1-0.5, and the gradient interval range is 0.05-0.1.

[0009] Preferably, the dynamic curve of the evaluation index and the area under the dynamic curve of the evaluation index under the AI ​​prediction model include: Dynamic curves of accuracy and area under the dynamic curve, dynamic curves of recall and area under the dynamic curve, dynamic curves of F1 score and area under the dynamic curve, dynamic curves of Kappa coefficient and area under the dynamic curve under the dynamic curve of Kappa coefficient in AI prediction models.

[0010] Preferably, high-potential target regions are obtained based on the optimal classification threshold of the AI ​​prediction model, including: Based on the optimal classification threshold of the AI ​​prediction model, the dynamic confusion matrix corresponding to the optimal classification threshold is obtained; Based on the dynamic confusion matrix corresponding to the optimal classification threshold, the region with high prediction probability is obtained. Based on the high prediction probability region, the high potential target region is obtained by eliminating the uncertain region through the concentration-area model.

[0011] Preferably, verifying the rationale for high-potential target regions includes: By using information gain and SHAP diagrams, combined with regional metallogenic regularities, the characteristic importance of high-potential target areas is analyzed to eliminate misjudgments by AI prediction models. Then, exploration costs are combined to verify the exploration of high-potential target areas.

[0012] Preferably, multi-source data includes: geological information data, geochemical data, and remote sensing data.

[0013] Preferably, preprocessing multi-source data to generate an evidence layer includes: By using GIS spatial analysis, geological information data is transformed from qualitative information into numerical representation, quantitative geological data is obtained, and geological evidence layers are generated. The KMO test is used to determine whether geochemical data is suitable for principal component analysis. If the result is negative, the geochemical data is invalid. If the result is positive, principal component analysis is used to perform standardization, covariance matrix calculation, eigenvalue and eigenvector decomposition, principal component screening, element combination determination, and spatial value processing on the geochemical data in sequence to obtain the geochemical evidence layer. The remote sensing data is sequentially processed through atmospheric correction, image mosaicking, interference area masking, principal component analysis, and alteration data extraction to generate an alteration evidence layer. An evidence layer is generated based on the geological evidence layer, the geochemical evidence layer, and the alteration evidence layer.

[0014] Compared with the prior art, the present invention has the following advantages and beneficial effects: This invention enriches the feature dimensions of data through multi-source data; based on multi-source data, it augments scarce mineral deposit samples using a synthetic minority class oversampling algorithm, achieving class balance in the dataset and alleviating the sample scarcity problem; it trains the AI ​​prediction model using the augmented training set, improving the model's performance; by iteratively adjusting the classification threshold to construct a dynamic confusion matrix, it can capture the continuous changes of the AI ​​prediction model under different classification thresholds in real time, improving the accuracy and reliability of the AI ​​prediction model in small sample scenarios; by evaluating the dynamic confusion matrix, it generates dynamic curves and areas under the dynamic curves of evaluation indicators, identifies the optimal classification threshold based on these curves, and then combines the optimal classification threshold with concentration... The area model eliminates uncertain regions and obtains high-potential target areas, thereby improving the accuracy of mineralization prediction in green areas and small sample scenarios, and enhancing the scientific and economical nature of exploration decisions. Attached Figure Description

[0015] Figure 1 This is a flowchart of a greenfield mineralization prediction method based on a dynamic confusion matrix in an embodiment of the present invention.

[0016] Figure 2 These are the accuracy dynamic curves and the area under the accuracy dynamic curves in the embodiments of the present invention for deep forest, random forest, convolutional neural network and support vector machine.

[0017] Figure 3The dynamic curves of recall and the area under the dynamic curves of recall are shown in the embodiments of the present invention for deep forest, random forest, convolutional neural network and support vector machine.

[0018] Figure 4 The dynamic curves of F1 scores and the area under the dynamic curves of F1 scores in the embodiments of the present invention are shown.

[0019] Figure 5 The dynamic curves of the Kappa coefficients and the area under the dynamic curves of the Kappa coefficients are shown in the embodiments of the present invention for deep forest, random forest, convolutional neural network and support vector machine.

[0020] Figure 6 This refers to the classification accuracy and cumulative accuracy area corresponding to each classification threshold in the dynamic accuracy curve generated in this embodiment of the invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0022] A method for predicting greenfield mineralization based on a dynamic confusion matrix, such as Figure 1 As shown, it includes the following steps: S1. Collect multi-source data and preprocess the multi-source data to generate evidence layers; Since mineralization is a complex process influenced by various conditions, a single data source cannot fully reflect the complex characteristics of mineralization. Furthermore, when making mineralization predictions on small samples of greenfields, the limited number of known deposits and scarce label data mean that features obtained from a single data source may result in insufficient feature dimensions, leading to underlearning of the AI ​​prediction model and making it difficult to capture stable and accurate mineralization. Therefore, it is necessary to explore the mineralization process of greenfields under the combined influence of multiple data sources.

[0023] Multi-source data includes geological information data, geochemical data, and remote sensing data.

[0024] The collection of multi-source data includes: The study area's metallogenic geological background and mineralization factors were collected through literature review, online retrieval, and historical geological maps. This involved obtaining geological information data for the study area, including regional geological maps, tectonic system data, and basic information on known mineral deposits. Tectonic system data included the distribution of regional fault zones and the location of suture zones. Basic information on known mineral deposits included the location, scale, and mineralization type of known mineral deposits.

[0025] Download official remote sensing data for the research area from the relevant satellite websites. This data includes OLI sensor data and satellite sensor spectral data. Extract alteration information related to the target mineral from the remote sensing data. This alteration information includes shallow alteration and mid-to-deep hydrothermal alteration. Shallow alteration information includes surface iron oxides and hydroxyl groups, while mid-to-deep hydrothermal alteration information includes sericitization, kaolinization, and chloritization. For example, download OLI sensor data from the Landsat 8 satellite website and extract shallow alteration information such as surface iron oxides and hydroxyl groups; download satellite sensor spectral data from the Terra satellite website and extract mid-to-deep hydrothermal alteration information such as sericitization, kaolinization, and chloritization.

[0026] By searching or reviewing published geochemical survey results within the study area, the focus is on collecting multi-element anomaly data related to mineralization, i.e., obtaining geochemical data for the study area, and ensuring that the collected geochemical data covers the study area and surrounding metallogenic belts.

[0027] Preprocessing of multi-source data includes: Geological information data is transformed from qualitative information into numerical representations using GIS spatial analysis methods, resulting in quantitative geological data and generating geological evidence layers. For example, lithology, strata, intrusive rocks, and fault attributes are quantitatively classified according to their contribution to mineralization, facilitating subsequent calculations.

[0028] Geochemical data are subjected to the KMO test to determine whether all their features are suitable for principal component analysis. If the test is negative, the geochemical data is invalid. If the determination is yes, then the geochemical data are standardized sequentially using principal component analysis to obtain standardized geochemical data, eliminating the influence of dimensional differences; based on the standardized geochemical data, a covariance matrix is ​​derived by constructing a standard matrix to quantify the intrinsic relationship between variables; the covariance matrix is ​​decomposed to obtain eigenvalues ​​and corresponding eigenvectors; based on the eigenvalues ​​and corresponding eigenvectors, principal components are selected using the cumulative variance explained rate to obtain n principal components; Based on the principal components, multi-element combinations related to mineralization were obtained through orthogonal rotation, clustering and correlation-assisted methods. These multi-element combinations include As-Sb-Bi, Cu-Mo-Pb-Zn-Au-Ag-Cd, and W-Sn. Based on multi-element combinations, spatial interpolation is performed using the inverse distance weighting method to obtain a geochemical evidence layer.

[0029] Principal component analysis includes processes such as data standardization, covariance matrix calculation, eigenvalue and eigenvector decomposition, and principal component screening, which are used to determine the n components that contribute the most to mineralization prediction; the cumulative variance explained rate refers to the proportion of the variable in the total variance.

[0030] The dimensionless values ​​of remote sensing data are converted into physically meaningful radiance or surface reflectance to eliminate sensor errors. Based on the radiance or surface reflectance, atmospheric correction is performed using the FLAASH module of the ENVI model to remove scattering interference from water vapor and aerosols, restoring the true surface reflectance signal and improving the accuracy of alteration mineral identification. After atmospheric correction, multiple images are cropped according to the study area boundary and mosaicked to remove invalid areas. After mosaicking, interference area masking is performed using segmentation thresholds to obtain the true surface reflectance image, removing shadows and cloud interference, retaining only exposed rock areas. Based on the spectral characteristics of the reflectance image, multiple bands that highlight the target alteration are selected. Principal component analysis is performed on these bands to obtain principal components. Each principal component reflects the spectral differences of alteration minerals. Alteration data is extracted from the principal components, and an alteration evidence layer is generated. The alteration data represents the principal component features that significantly influence mineralization.

[0031] An evidence layer is generated based on the geological evidence layer, the geochemical evidence layer, and the alteration evidence layer.

[0032] The collected multi-source data is preprocessed to transform qualitative data into quantitative data, providing effective and accurate data support for subsequent AI prediction models to predict mineral deposits.

[0033] S2. Obtain multiple known mining locations based on the evidence layer; based on the multiple known mining locations, perform data augmentation through a synthetic minority class oversampling algorithm to obtain multiple training sets and multiple validation sets; Positive samples refer to mineral deposit samples, while negative samples refer to non-mineral deposit samples.

[0034] Due to the extremely limited number of mineral occurrences in the study area, mineral deposit samples are scarce, meaning there are very few positive samples and a large number of non-mineral occurrences, resulting in a very high proportion of non-mineral deposit samples and a large number of negative samples. This leads to a significant difference between the number of positive and negative samples. To balance the number of positive and negative samples, a synthetic minority oversampling algorithm is used to generate synthetic mineral occurrences that conform to the geological patterns of mineralization, thereby enhancing the positive sample data and achieving class balance in the dataset. The synthetic minority oversampling algorithm (SMOTE) is an algorithm that balances the dataset by interpolating to generate new samples, and it is particularly suitable for handling class imbalance problems.

[0035] S21. Based on a known mining point, obtain k nearest neighbor samples using Euclidean distance; Calculate a known mineral deposit using Euclidean distance. k nearest neighbor samples in feature space (j=1,2,…k); the value of k is determined by cross-validation to ensure that the selected nearest neighbor samples have similar metallogenic geological backgrounds.

[0036] S22. Starting from a known mineral point, move a random proportion of the distance along the direction pointing to a nearest neighbor sample to generate a synthetic mineral point. Ensure synthetic ore deposits The eigenvalues ​​are within a reasonable range of positive sample characteristics, which is consistent with the continuity of mineralization characteristics.

[0037] S23. Repeat S21-S22 until all neighboring samples of all known mineral points are traversed to obtain multiple synthetic mineral points. Generate positive samples based on the multiple synthetic mineral points and multiple known mineral points. S24. Based on the positive samples, randomly select the same number of non-mining points from the evidence layer as the positive samples to generate negative samples; To ensure that the number of positive samples is consistent with the number of negative samples, i.e., the number of mining sites equals the number of non-mining sites, a class balance in the dataset is achieved. Synthetic mining sites are generated using a synthetic minority class oversampling algorithm to increase the number of positive samples, ensuring a balance between positive samples (minority class samples) and negative samples (majority class samples). This effectively solves the class imbalance problem in small-sample scenarios in greenfields, providing sufficient feature learning basis for subsequent AI prediction models to predict mineralization.

[0038] S25. Generate a balanced dataset based on positive and negative samples, and divide the balanced dataset into a training set and a validation set. The balanced dataset is divided into a training set and a validation set according to a preset ratio, which can be 7:3. The ratio of positive samples to negative samples in the training set and the validation set remains unchanged, that is, the ratio of positive samples to negative samples is 1:1.

[0039] To minimize the impact of randomness, the balanced dataset can be randomly selected multiple times according to a preset ratio to form different training and validation sets.

[0040] S26. Repeat S21-S25 until the preset number of times is reached to obtain multiple training sets and multiple validation sets.

[0041] The preset number of attempts can be 5, which will yield 5 training sets and 5 validation sets for subsequent uncertainty analysis of the AI ​​prediction model.

[0042] S3. Train the AI ​​prediction model based on multiple training sets; AI prediction models include deep forest, random forest, convolutional neural network, and support vector machine; AI prediction models are suitable for mineralization prediction scenarios with small samples and high feature dimensions.

[0043] Based on multiple training sets, the parameters of the AI ​​prediction model are optimized using the random grid search method and the ten-fold cross-validation method to obtain a well-trained AI prediction model, thereby improving the prediction ability of the AI ​​prediction model and enhancing the accuracy, reliability, and stability of mineralization prediction.

[0044] S4. Based on multiple validation sets, obtain mineralization prediction results through the trained AI prediction model; based on the mineralization prediction results, construct a dynamic confusion matrix corresponding to each classification threshold through preset classification threshold intervals and gradient intervals. The validation set is the predicted sample.

[0045] Confusion matrices are a core tool for evaluating the performance of classification models in machine learning, especially suitable for binary or multi-class classification tasks. They present the correspondence between model predictions and true labels in matrix form, intuitively reflecting the accuracy and error type of the model's classification.

[0046] The confusion matrix is ​​a 2×2 matrix, with the horizontal axis representing the model's predicted class and the vertical axis representing the true class of the data. It contains four key metrics: True Positive (TP), False Positive (FP), True Negative (TN), and False Negative (FN). A True Positive means that both the true and predicted classes are positive; a False Positive means that both the true and predicted classes are negative; a True Negative means that both the true and predicted classes are negative; and a False Negative means that both the true and predicted classes are negative.

[0047] The core of dynamic confusion matrix lies in analyzing the dynamic changes of the confusion matrix under different thresholds, so as to adaptively adjust the classification threshold and thus optimize the performance of AI prediction model.

[0048] Based on practical experience in mineralization prediction, a classification threshold range and gradient interval range are set for the dynamic confusion matrix. The classification threshold range can be 1-0.5, and the gradient interval range can be 0.05-0.1. In the scenario of small sample data for mineralization prediction, based on the confidence of the mineral point category output by the AI ​​prediction model, a gradient classification threshold sequence from high to low is constructed. Typically, 0.9 is selected as the initial classification threshold, and 0.1 is selected as the gradient interval, so that the classification threshold gradually decreases from 0.9 to 0.5 with a gradient interval of 0.1. The setting of the classification threshold range is based on practical experience in mineralization prediction. High classification thresholds above 0.9 are screened as high-risk mineralization areas (i.e., mineral points), while 0.5 is used as the conventional threshold for binary classification tasks (i.e., the classification threshold of the static confusion matrix). The classification threshold range between 0.5 and 0.9 is regarded as potential mineralization areas (i.e., potential mineral points), providing core data support for subsequent dynamic analysis.

[0049] For example, the classification threshold range is 1-0.5, and the gradient interval is 0.1. When the classification threshold is 0.9, samples with a mineralization confidence score greater than 0.9 in the mineralization prediction results are classified as mineral deposits, and samples with a mineralization confidence score less than or equal to 0.9 in the mineralization prediction results are classified as non-mineral deposits, generating a dynamic confusion matrix under this classification threshold. Based on the gradient interval of 0.1, the classification threshold is reduced from 0.9 to 0.8. When the classification threshold is updated to 0.8, samples with a mineralization confidence score greater than 0.8 in the mineralization prediction results are classified as mineral deposits, and samples with a mineralization confidence score less than or equal to 0.8 in the mineralization prediction results are classified as non-mineral deposits, generating a dynamic confusion matrix under this classification threshold. This process continues until the classification threshold is updated to 0.5, generating a dynamic confusion matrix under a classification threshold of 0.5.

[0050] Traditional static confusion matrices classify samples using only a single classification threshold (typically 0.5). Therefore, this application reclassifies samples within the corresponding mineralization confidence interval each time the classification threshold is updated, generating a dynamic confusion matrix corresponding to that threshold. This allows the sample's category to be dynamically adjusted as the classification threshold changes, enabling some samples to uncover potential mineralization information that is ignored by traditional methods. The dynamic confusion matrix can more accurately determine the contribution of different features to mineralization prediction, overcoming the limitation of geochemical features dominating traditional mineralization prediction.

[0051] S5. Evaluate each dynamic confusion matrix separately to generate dynamic curves of evaluation indicators and areas under the dynamic curves of evaluation indicators under the AI ​​prediction model; identify the optimal classification threshold of the AI ​​prediction model based on the dynamic curves of evaluation indicators and areas under the dynamic curves of evaluation indicators under the AI ​​prediction model. The evaluation metrics include accuracy, recall, F1 score, and Kappa coefficient. Accuracy reflects the overall correctness of classification; recall reflects the completeness of mineral point identification, corresponding to the mineralization area coverage; the F1 score is the harmonic mean, balancing accuracy and recall; and the Kappa coefficient is a measure of consistency after eliminating random factors, improving the reliability of the evaluation.

[0052] The dynamic curve of the evaluation index visualizes the correspondence between the evaluation index and the classification threshold, i.e., the threshold-index dynamic image. The threshold-index dynamic image is the core basis for the division of high-potential target areas.

[0053] Based on each dynamic confusion matrix, dynamic curves of multi-dimensional evaluation metrics and the area under the dynamic curves of the evaluation metrics (i.e., the AUC value under the dynamic curves of the evaluation metrics) are generated for the AI ​​prediction model, such as... Figures 2-5 The figures show the dynamic curves of accuracy and area under the dynamic curve of accuracy, recall and area under the dynamic curve of recall, F1 score and area under the dynamic curve of F1 score, and Kappa coefficient and area under the dynamic curve of Kappa coefficient for Deep Forest, Random Forest, Convolutional Neural Network, and Support Vector Machine, respectively.

[0054] In generating accuracy dynamic curves, such as Figure 6 As shown, based on the dynamic confusion matrix corresponding to multiple classification thresholds, the classification accuracy and cumulative area of ​​accuracy (CAA) corresponding to each classification threshold need to be calculated. The accuracy plateau and curve slope are key indicators in the accuracy dynamic curve. The accuracy plateau refers to the region where the classification accuracy remains stable within a certain classification threshold range. When the curve slope is high, it indicates high effective mineralization information in the predicted samples, and fine-tuning the classification threshold will significantly improve the performance of the AI ​​prediction model. When the accuracy plateau is reached, it indicates a balanced ratio of mineralized and non-mineralized points in the predicted samples, and adjusting the classification threshold will not affect the change in recognition accuracy. Furthermore, when the highest accuracy plateau is reached, it corresponds to the classification threshold of a high-potential target area. Therefore, the accuracy plateau and curve slope are key criteria for selecting the optimal classification threshold.

[0055] The area under the dynamic curve of accuracy reflects the stability of global accuracy; the area under the dynamic curve of recall reflects the quality of the AI ​​prediction model's ability to capture mineral deposits; the area under the dynamic curve of F1 score reflects the comprehensive balance between accuracy and coverage; and the area under the dynamic curve of Kappa coefficient reflects the strength of consistency between prediction and actual distribution.

[0056] The optimal classification threshold for AI prediction models is identified based on the dynamic curves and areas under the dynamic curves of evaluation metrics. Specifically, the optimal classification thresholds for Deep Forest, Random Forest, Convolutional Neural Network, and Support Vector Machine are identified based on their respective dynamic curves and areas under the curves for accuracy, recall, F1 score, and Kappa coefficient. The criteria for determining the optimal classification threshold are: when the classification threshold decreases to a certain value, the dynamic curve of the evaluation metric enters a stable plateau period; if the classification threshold continues to decrease, it will lead to a sharp increase in false positives. The optimal classification threshold is the threshold for delineating high-potential target areas.

[0057] By comprehensively evaluating the performance of different models (deep forest, random forest, convolutional neural network, and support vector machine) under the optimal classification threshold, we can determine their mineralization prediction capabilities.

[0058] S6. Obtain high-potential target areas based on the optimal classification threshold of the AI ​​prediction model, and verify the rationality of the high-potential target areas.

[0059] The dynamic confusion matrix corresponding to the optimal classification threshold is obtained based on the optimal classification threshold of the AI ​​prediction model; high prediction probability regions are obtained based on the dynamic confusion matrix corresponding to the optimal classification threshold; based on the high prediction probability regions, units with large fluctuations in prediction probability and low consensus of AI prediction models, i.e., uncertain regions, are excluded by the concentration-area model, and reliable high-potential target areas are delineated.

[0060] By constraining both high prediction probability and low uncertainty, high potential target areas are obtained. These high potential target areas satisfy the rationality of the superposition of multiple geological features and are located in areas with high geochemical anomalies, dense linear structures, and overlapping areas of high remote sensing alteration anomalies.

[0061] The concentration-area model (CA model) quantifies mineral-induced anomalies and background noise by analyzing the power-law relationship between elemental concentration / anomaly concentration and the corresponding distribution area, thereby identifying areas of uncertainty. For geochemical anomalies or remotely sensed alteration anomalies, the concentration C and area A must satisfy the following power-law relationship: A(C)∝C D Where A(C) represents the area of ​​the region where the concentration is greater than the concentration threshold; C represents characteristic indicators such as elemental concentration and remote sensing alteration intensity; and D represents the fractal dimension. Mineral-induced anomalies are typically characterized by high concentration and small area; background noise is typically characterized by low concentration and large area; the intersection of mineral-induced anomalies and background noise is the threshold for dividing the anomaly background.

[0062] By using information gain and SHAP diagrams, combined with regional metallogenic regularities, we conduct feature importance analysis on high-potential target areas, analyze whether high-potential target areas conform to qualitative and quantitative models, eliminate misjudgments by AI prediction models, and then conduct exploration verification of high-potential target areas in combination with exploration costs. SHAP plots are visualization tools based on SHAP values ​​used to interpret the prediction results of machine learning models; they support both global feature importance analysis and local single-sample interpretation.

[0063] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting greenland mineralization based on a dynamic confusion matrix, characterized in that, include: Collect multi-source data and preprocess the multi-source data to generate evidence layers; Multiple known mineral deposits were obtained based on the evidence layer; Based on multiple known mining sites, data augmentation was performed using a synthetic minority class oversampling algorithm to obtain multiple training sets and multiple validation sets. The AI ​​prediction model is trained using multiple training sets; Based on multiple validation sets, mineralization prediction results are obtained through a trained AI prediction model. Based on the mineralization prediction results, a dynamic confusion matrix corresponding to each classification threshold is constructed through a preset classification threshold range and gradient interval. Each dynamic confusion matrix is ​​evaluated to generate dynamic curves and areas under the dynamic curves of the evaluation indicators under the AI ​​prediction model; the optimal classification threshold of the AI ​​prediction model is identified based on the dynamic curves and areas under the dynamic curves of the evaluation indicators under the AI ​​prediction model. High-potential target areas are obtained based on the optimal classification threshold of the AI ​​prediction model, and the rationality of the high-potential target areas is verified.

2. The greenfield mineralization prediction method based on a dynamic confusion matrix according to claim 1, characterized in that, The process of obtaining multiple training sets and multiple validation sets includes: S1. Based on a known mining point, obtain k nearest neighbor samples using Euclidean distance; S2. Starting from a known mineral point, move a random proportion of the distance along the direction pointing to a nearest neighbor sample to generate a synthetic mineral point. S3. Repeat S1-S2 until all neighboring samples of all known mineral points are traversed to obtain multiple synthetic mineral points. Generate positive samples based on the multiple synthetic mineral points and multiple known mineral points. S4. Based on the positive samples, randomly select the same number of non-mining points from the evidence layer as the positive samples to generate negative samples; S5. Generate a balanced dataset based on positive and negative samples, and divide the balanced dataset into a training set and a validation set. S6. Repeat S1-S5 until the preset number of times is reached to obtain multiple training sets and multiple validation sets.

3. The greenfield mineralization prediction method based on a dynamic confusion matrix according to claim 1, characterized in that, The AI ​​prediction models include: deep forest, random forest, convolutional neural network, and support vector machine.

4. The greenfield mineralization prediction method based on a dynamic confusion matrix according to claim 1, characterized in that, The process of training the AI ​​prediction model based on multiple training sets includes: Based on multiple training sets, the parameters of the AI ​​prediction model are optimized using random grid search and ten-fold cross-validation to obtain a well-trained AI prediction model.

5. The greenfield mineralization prediction method based on a dynamic confusion matrix according to claim 1, characterized in that: The classification threshold range is 1-0.5, and the gradient interval range is 0.05-0.

1.

6. The greenfield mineralization prediction method based on a dynamic confusion matrix according to claim 1, characterized in that, The dynamic curves and areas under the dynamic curves of the evaluation indicators under the AI ​​prediction model include: the dynamic curve of accuracy and its area under the dynamic curve, the dynamic curve of recall and its area under the dynamic curve, the dynamic curve of F1 score and its area under the dynamic curve, and the dynamic curve of Kappa coefficient and its area under the dynamic curve.

7. The greenfield mineralization prediction method based on a dynamic confusion matrix according to claim 1, characterized in that, The process of obtaining high-potential target regions based on the optimal classification threshold of the AI ​​prediction model includes: Based on the optimal classification threshold of the AI ​​prediction model, the dynamic confusion matrix corresponding to the optimal classification threshold is obtained; Based on the dynamic confusion matrix corresponding to the optimal classification threshold, the region with high prediction probability is obtained. Based on the high prediction probability region, the high potential target region is obtained by eliminating the uncertain region through the concentration-area model.

8. The greenfield mineralization prediction method based on a dynamic confusion matrix according to claim 1, characterized in that, The verification of the rationale for high-potential target regions includes: By using information gain and SHAP diagrams, combined with regional metallogenic regularities, the characteristic importance of high-potential target areas is analyzed to eliminate misjudgments by AI prediction models. Then, exploration costs are combined to verify the exploration of high-potential target areas.

9. The greenfield mineralization prediction method based on a dynamic confusion matrix according to claim 1, characterized in that, The multi-source data includes: geological information data, geochemical data, and remote sensing data.

10. The greenland mineralization prediction method based on a dynamic confusion matrix according to claim 9, characterized in that, The process of preprocessing multi-source data to generate an evidence layer includes: By using GIS spatial analysis, geological information data is transformed from qualitative information into numerical representation, quantitative geological data is obtained, and geological evidence layers are generated. The KMO test is used to determine whether geochemical data is suitable for principal component analysis. If the result is negative, the geochemical data is invalid. If the result is positive, principal component analysis is used to perform standardization, covariance matrix calculation, eigenvalue and eigenvector decomposition, principal component screening, element combination determination, and spatial value processing on the geochemical data in sequence to obtain the geochemical evidence layer. The remote sensing data is sequentially processed through atmospheric correction, image mosaicking, interference area masking, principal component analysis, and alteration data extraction to generate an alteration evidence layer. An evidence layer is generated based on the geological evidence layer, the geochemical evidence layer, and the alteration evidence layer.

Images