A quantitative prediction method for seafloor sulfide mineralization based on machine learning

Abstract

The present invention discloses a method for quantitative prediction of seabed sulfide mineralization based on machine learning. The method first determines seabed sulfide prospecting indicators, constructs a spatial data set based on the prospecting indicators, then establishes a machine learning model, combines the prediction results of two machine learning models, random forest and isolation forest, using fuzzy operators, and finally delineates the prospective mineralization area. Compared with traditional statistical techniques and empirical exploration models, the method of the present invention is applicable to small sample data and can fully utilize the mineralization information in the characteristic variables, and is a method for quantitative prediction of seabed sulfide mineralization.

Description

Technical Field

[0001] The present invention belongs to the technical field of seabed sulfide mineralization prediction, and in particular relates to a quantitative prediction method for seabed sulfide mineralization based on machine learning. Background Art

[0002] The submarine sulfide deposits are rich in copper, zinc, gold, silver and other elements. The total amount of submarine sulfide deposits in the global submarine newborn volcanic belt can reach 6×10 8 Tons, of which the sulfide resources on slow and ultra-slow spreading mid-ocean ridges (spreading rate <40mm / yr) account for 86% of the total sulfide resources on the mid-ocean ridges. Seafloor sulfide, an important potential mineral resource, has attracted widespread attention from scientists at home and abroad. The use of traditional deep-sea exploration technology to conduct seafloor sulfide exploration is not only difficult but also costly, which leads to the discovery of hydrothermal areas and the collection of data related to hydrothermal mineralization. In order to narrow the exploration scope and improve the efficiency of mineral exploration, a mineralization prediction method suitable for small samples and that can fully reflect the mineralization information of existing data is needed to delineate the prospecting area. Nowadays, some traditional data-driven or knowledge-driven methods have been applied in the prediction of seafloor sulfide mineralization, but these methods cannot meet the requirements of narrowing the exploration scope. Therefore, it is urgent to propose a method suitable for quantitative prediction of seafloor sulfide mineralization to guide further resource evaluation and prediction. Summary of the invention

[0003] The present invention aims to solve the problem that the traditional data-driven or knowledge-driven mineralization prediction methods cannot simultaneously meet the requirements of being applicable to small sample data and enabling the existing data to fully reflect mineralization information. A machine learning method suitable for quantitative prediction of seabed sulfide mineralization is proposed. To achieve the above purpose, the present invention adopts the following technical solution: A method for quantitative prediction of seabed sulfide mineralization based on machine learning, comprising the following steps:

[0004] Step 1: Determine deep seabed sulfide prospecting indicators from three aspects: topography, geology, and water bodies: Collect literature related to seabed sulfide mineralization, understand its geological background and mineralization process, and thus determine deep seabed sulfide prospecting indicators, including three categories: topography, geology, and water bodies.

[0005] Step 2: Construct spatial data set: including the division of study area, establishment of target variables and characteristic variables; the specific steps are as follows:

[0006] Step 2.1: Divide the study area into a grid with a pixel size of, for example, 1 km × 1 km.

[0007] Step 2.2: The target variable is a binary variable, with 1 and 0 representing mining areas and non-mining areas respectively. In the training set, the cells where the known hydrothermal areas are located are used as mining areas; when selecting non-mining areas, the evidence weight method is first used to delineate non-prospective areas, and then selection is made within the non-prospective areas according to the selection criteria for non-mining areas in mineralization prediction. In the validation set, the cells corresponding to the MSI < 0.3 area are used as mining areas, where MSI is calculated from the content of Al, Fe, and Mn elements according to the formula Al / (Al+Fe+Mn), and is an indicator for judging polymetallic sediments and non-polymetallic sediments. MSI < 0.3 indicates that the sediments are polymetallic sediments, indicating that there is a high probability of a hydrothermal area here; the selection method for non-mining areas is the same as that for the training set.

[0008] Step 2.3: Generate characteristic variables based on the prospecting indicators in step 1.

[0009] Step 3: construct random forest and isolation forest models, train and determine model parameters based on the spatial data set constructed in step 2; and use fuzzy operators in fuzzy logic principles to integrate the mineralization prediction results of random forest and isolation forest models. Fuzzy logic is a mineralization prediction method that normalizes characteristic variables according to their contribution to mineral exploration to generate a membership degree (0-1), and then integrates the membership degrees of all characteristic variables through fuzzy operators to form a mineralization potential map. In the present invention, fuzzy operators are not used to integrate prediction variables, but fuzzy operators are used to integrate the prediction results of random forest models and isolation forest models. Since the prediction results based on random forest and isolation forest models are between 0-1, normalization is not required before integration. Commonly used fuzzy operators in mineralization prediction include fuzzy AND, fuzzy OR, fuzzy algebraic sum, fuzzy algebraic product and gamma (γ) operator: fuzzy AND takes the minimum membership among all variables; fuzzy OR takes the maximum membership among all variables; the integration result of fuzzy algebraic product is less than the minimum membership of the variable; the result of fuzzy algebraic sum is greater than the maximum membership of the variable; the integration result of γ operator is between the maximum membership and the minimum membership of the variable. When the membership of one or more variables is very high and the membership of other variables is very low, the γ operator can make the final membership between the two extreme values. Therefore, the present invention selects γ operator as the fuzzy operator for integrating the mineralization prediction results of random forest and isolation forest models. The integration formula is:

[0010] μ C =[1-(1-μ A )(1-μ B )] γ ×[μ A μ B ] 1-γ

[0011] μ A , μ Bare the mineralization potential values ​​predicted by the random forest model and the isolation forest model, μ C is the final result after integration, γ is the value of the γ operator (0≤μ A，B，C ≤1, 0≤γ≤1).

[0012] Step 4: Delineate prospective areas for the seabed sulfide mineralization potential map drawn based on the machine learning model. Draw a capture rate curve based on the mineralization probability. The horizontal axis of the curve is the ratio of the cell area to the study area when it is greater than the mineralization probability threshold, and the vertical axis is the ratio of the number of hydrothermal areas contained when it is greater than the mineralization probability threshold to the total number of hydrothermal areas. The mineralization potential value corresponding to the first inflection point of the capture rate curve with a vertical axis greater than 60% is selected as the threshold for dividing prospective areas and non-prospective areas. The cells with a mineralization probability greater than the threshold are prospective mineralization areas.

[0013] Furthermore, the selection criteria for non-mining areas in the mineralization prediction in step 2.2 further include:

[0014] (1) The number of non-mining areas should be equal to the number of mining areas.

[0015] (2) Non-mining areas should be far away from known hydrothermal areas. The method for determining “far away” is as follows: count the distance between each hydrothermal area and the nearest hydrothermal area among all hydrothermal areas; sort the above proximity distances from large to small and draw a scatter plot, remove abnormally large values ​​and select the maximum proximity distance as the threshold for non-mining areas to be “far away” from known hydrothermal areas. When the distance between non-mining areas and known hydrothermal areas is greater than the threshold, the non-mining area is considered to be “far away” from the known hydrothermal areas.

[0016] (3) The location of non-mining areas is random in space. First, the Monte Carlo method is used to generate 10 groups of randomly distributed non-mining areas under the conditions of (1) and (2). Then, the randomness of the locations of the 10 groups of non-mining areas is evaluated through nearest neighbor analysis, and the group with the neighbor ratio (ANN) closest to 1 is selected as the non-mining area.

[0017] Among them, the method of using nearest neighbor analysis to determine whether the spatial position of points in the study area is "random" is as follows:

[0018] 1) Calculate the observed average distance: measure the distance between each point and its nearest point, and calculate the mean of these distances,

[0019]

[0020] where d i is the distance between the i-th point and its nearest point;

[0021] 2) Calculate the expected average distance: Assuming that all points in the study area are randomly distributed, the expected average distance of these points spread over the entire study area is,

[0022]

[0023] Where n represents the number of points in the study area, and A represents the area of ​​the study area;

[0024] 3) Calculate the neighbor ratio (ANN),

[0025]

[0026] The closer the ANN value is to 1, the more random the spatial distribution of the points is.

[0027] Furthermore, in step 3, constructing a random forest model further includes:

[0028] (1) Construct a dataset D, which contains sample set D s , sample set D s From the training set D t and validation set D v Assume that the total number of samples with labels (target variables and predictor variables) in the data set D is S, and the decision tree sample set D is constructed by bagging method. s ; Bootstrap sampling method is used to select samples from the sample set D s Select about 2 / 3 of the training samples as the training set D t To train a decision tree; the remaining 1 / 3 is called out-of-bag data and can be used as a validation set D v , which is used to unbiasedly estimate the model error.

[0029] (2) Determine the number of predictor variables selected by the decision tree node branch (m try ).

[0030] (3) Determine the number of decision trees (k) in the random forest model.

[0031] (4) K decision trees that do not need to be pruned are constructed into a random forest model and applied to the data set D for prediction. The arithmetic mean of the classification results of all decision trees is the final prediction result.

[0032] Furthermore, in step 3, constructing an isolation forest model further includes:

[0033] (1) Randomly select p points from the training data as subsamples and put them into the root node of an isolated tree.

[0034] (2) Random cuts are made between the maximum and minimum values ​​of randomly selected predictor variables until all samples are isolated or the binary tree grows to full height.

[0035] (3) Repeat the isolation process of a tree t times to form an isolation forest. The average path length of the sample on the isolated tree set is the anomaly score, that is, the mineralization potential value:

[0036]

[0037] Where c(n) is the average path length when n samples are given, which is used to standardize the path length h(x) of sample x. E(h(x)) is the expected path length of sample x in a batch of isolated trees.

[0038] Furthermore, in step 3, the method for determining the value of the γ operator further includes: drawing a prediction area (PA) map according to the prediction results generated by different γ values, and there are two curves in the PA map, which are the percentage of known mineral points contained when it is greater than a certain prediction value, that is, the prediction rate curve; and the percentage curve of the area occupied by the study area when it is greater than the prediction value. The intersection of the two curves is the criterion for evaluating the quality of the prediction results. The higher the position of the intersection in the PA map, the more known hydrothermal areas are contained in a relatively small prediction area, and the more reliable the selected γ operator is. Therefore, the γ value when the intersection of the PA map is the highest is selected as the best γ operator, and the prediction result is the final mineralization potential map.

[0039] The beneficial technical effect of the present invention is that the quantitative prediction method of seabed sulfide mineralization based on machine learning proposed in the present invention is not only suitable for small samples, but also allows the use of binary, multi-classification and continuous data as predictive variables according to the different mineralization indication methods of various indicators, so that it can reflect the mineralization information to the greatest extent. The identified prospective areas can provide a basis for further narrowing the exploration scope. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a flow chart of the implementation process of the present invention;

[0041] Figure 2 It is a comparison diagram of the prospective area delineated by the present invention and the prospective area delineated by the evidence weight model; DETAILED DESCRIPTION

[0042] The present invention is further described in detail below with reference to the accompanying drawings and specific examples.

[0043] First, a study area is determined. Taking the 49°E-50.25°E ocean ridge section in the southwest Indian Ocean as an example, the implementation process of the quantitative prediction method of seafloor sulfide mineralization based on machine learning is described in detail below (e.g. Figure 1 shown):

[0044] Step 1: Determine the deep-sea sulfide prospecting indicators: Collect literature related to seabed sulfide mineralization, understand its geological background and mineralization process, and thus determine the deep-sea sulfide prospecting indicators, including 14 types in 3 categories: topography, geology, and water bodies, as shown in Table 1.

[0045] Table 1 is the prospecting index (i.e., characteristic variables) during the implementation of the present invention;

[0046]

[0047]

[0048] Step 2: Construct a spatial data set, including the division of the study area, the establishment of target variables and characteristic variables. The specific steps are as follows:

[0049] Step 2.1: Divide the study area into a grid with a pixel size of 1 km × 1 km.

[0050] Step 2.2: The target variable is a binary variable, with 1 and 0 representing mining areas and non-mining areas respectively. In the training set, the cells where 11 known hydrothermal areas are located are used as mining areas. When selecting non-mining areas, the evidence weight method is first used to delineate non-prospective areas, and then selection is made within the non-prospective areas according to the selection criteria for non-mining areas in mineralization prediction. In the validation set, 132 cells corresponding to the MSI < 0.3 area are used as mining areas, and the selection method for non-mining areas is the same as that of the training set.

[0051] The selection criteria for non-ore areas in the mineralization prediction in step 2.2 are as follows:

[0052] (1) The number of non-mining areas should be equal to the number of mining areas, which is 11.

[0053] (2) Non-mining areas should be far away from known hydrothermal areas. Statistics on the nearest distances between hydrothermal areas in the study area show that after eliminating abnormally large values ​​of the nearest distances, the maximum nearest distance is 16.5 km. Therefore, non-mining areas should be more than 16.5 km away from all hydrothermal areas.

[0054] (3) The location of non-mining areas is random in space. First, the Monte Carlo method is used to generate 10 groups of randomly distributed non-mining areas under the conditions of (1) and (2). Then, the randomness of the locations of the 10 groups of non-mining areas is evaluated by nearest neighbor analysis. The ANN value of the second group is 0.97, which is closest to 1 among the 10 groups. Therefore, the second group is selected as the non-mining area.

[0055] Step 2.3: Generate characteristic variables based on the prospecting indicators in step 1.

[0056] Step 3: Construct random forest and isolation forest models, and select the best γ operator to integrate the mineralization prediction results of random forest and isolation forest models. The formula is:

[0057] μ C =[1-(1-μ A )(1-μ B )] γ ×[μ A μ B ] 1-γ

[0058] μ A , μ B are the mineralization potential values ​​predicted by the random forest model and the isolation forest model, μ C is the final result after integration, γ is the value of the γ operator (0≤μ A，B，C ≤1, 0≤γ≤1).

[0059] The specific operation process of step 3 is as follows:

[0060] Step 3.1 builds a random forest model, further including:

[0061] (1) Assume that in data set D, the total number of samples with labels (target variables and predictor variables) is 11. Use bagging to construct a decision tree sample set. Use Bootstrap sampling to select the sample set D. s Select 7 (about 2 / 3) training samples as training set D t , used to train a decision tree. The remaining 1 / 3 is called out-of-bag data and can be used as the test set D v , which is used to unbiasedly estimate the model error.

[0062] (2) Determine the three predictor variables selected for the decision tree node branch.

[0063] (3) Determine the 200 decision trees in the random forest model.

[0064] (4) 200 decision trees that do not need to be pruned are constructed into a random forest model and applied to data set D for prediction. The arithmetic mean of the classification results of all decision trees is the final prediction result.

[0065] Step 3.2 builds the isolation forest model, further including:

[0066] (1) Randomly select 256 points from the training data as subsamples and put them into the root node of an isolated tree.

[0067] (2) Random cuts are made between the maximum and minimum values ​​of randomly selected predictor variables until all samples are isolated or the binary tree grows to full height.

[0068] (3) Repeat the isolation process of a tree 100 times to form an isolation forest. The average path length of the sample on the isolated tree set is the anomaly score, that is, the mineralization potential value:

[0069]

[0070] Where c(n) is the average path length when n samples are given, which is used to standardize the path length h(x) of sample x. E(h(x)) is the expected path length of sample x in a batch of isolated trees.

[0071] Step 3.3 The method for selecting the best γ operator further includes: drawing a prediction area (PA) map according to the prediction results generated when γ = 0.95, 0.93, 0.80, and 0.85, respectively. When γ = 0.95, the intersection of the two curves is the criterion for evaluating the quality of the prediction results. The higher the intersection appears in the PA map, the more known mineral points are contained in a relatively small prediction area. Therefore, γ = 0.95 is selected as the fuzzy operator for integrating the prediction results of the random forest and isolation forest models.

[0072] Step 4: Delineate prospective areas based on the seafloor sulfide mineralization potential map drawn based on the machine learning model. The mineralization potential value of 0.43 corresponding to the first inflection point when the proportion of mineral points on the vertical coordinate of the capture rate curve is greater than 60% is selected as the threshold for dividing prospective areas and non-prospective areas. The mineralization probability is greater than the threshold and the continuous area is greater than 10km 2 The cells are the mineralization prospective areas, and finally 8 prospective areas were identified, which were named ABCDEFGH from west to east. The identified prospective areas cover 28% of the study area and include all hydrothermal areas.

[0073] First, the model integrated by the γ operator was evaluated by sensitivity, specificity, accuracy, Kappa coefficient, and AUC value. The results showed that the evaluation indicators on the training set and the validation set were greater than 90%, among which the specificity and AUC values ​​were close to 100% (Table 2). The results after the integration of the γ operator were then compared with the prediction results of using random forest and isolation forest alone, and it was found that the integrated model had the best prediction performance (Table 2). The above evaluation shows that the machine learning model proposed in the present invention not only retains the advantages of the random forest model and the isolation forest model in being suitable for small samples (the number of samples in the study area is 11, which is a small sample), but also has better prediction performance than the case where the random forest model and the isolation forest model are used alone.

[0074] Table 2. Evaluation results of different models

[0075]

[0076] Comparing the above 8 prospective areas with the prospective areas delineated by the traditional data-driven evidence weight method, it is found that except for the edge parts of prospective areas C and E, the other 6 prospective areas are all within the prospective areas delineated by the evidence weight method, and the overall area is significantly smaller than the prospective areas delineated by the evidence weight method, indicating that the prospective areas delineated by the present invention are more accurate than those delineated by the evidence weight method ( Figure 2 ). This is because compared with the weight of evidence method that only uses binary data as feature variables, the machine learning model of the present invention is not only applicable to small samples, but also uses continuous data as feature variables to make fuller use of data. The prospecting area delineated by the present invention helps to further narrow the exploration scope.

[0077] The above embodiments are used to illustrate the present invention rather than to limit the present invention. Any modification and change made to the present invention within the spirit of the present invention and the protection scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A quantitative prediction method for seafloor sulfide mineralization based on machine learning, characterized in that: The steps include: Step 1: Determine the seabed sulfide prospecting indicators from the three aspects of topography, geology and water body; Step 2: Construct spatial data set: including the division of study area, establishment of target variables and characteristic variables; specifically: Step 2.1: Divide the study area using a grid; Step 2.2: The target variable is a binary variable, with 1 and 0 representing mining areas and non-mining areas respectively; in the training set, the cells where the known hydrothermal areas are located are used as mining areas; when selecting non-mining areas, the evidence weight method is first used to delineate non-prospective areas, and then selection is made within the non-prospective areas according to the selection criteria for non-mining areas in mineralization prediction; in the validation set, the cells corresponding to the MSI < 0.3 area are used as mining areas, and the selection method for non-mining areas is the same as that of the training set; Step 2.3: Generate characteristic variables based on the prospecting indicators in step 1; Step 3: Construct random forest and isolation forest models, train and determine model parameters based on the spatial data set constructed in step 2; use fuzzy operators in fuzzy logic principles to integrate the mineralization prediction results of random forest and isolation forest models; Step 4: Delineate prospective areas on the seafloor sulfide mineralization potential map obtained based on the results of step 3; draw a capture rate curve based on the mineralization probability, the abscissa of the curve is the ratio of the cell area to the study area when it is greater than the mineralization probability threshold, and the ordinate is the ratio of the number of hydrothermal areas contained when it is greater than the mineralization probability threshold to the total number of hydrothermal areas. The mineralization potential value corresponding to the first inflection point of the capture rate curve with a ordinate greater than 60% is selected as the threshold for dividing prospective areas and non-prospective areas. The cell with a mineralization probability greater than the threshold is the mineralization prospective area; The selection criteria for non-ore areas in the mineralization prediction in step 2.2 are: (1) The number of non-mining areas should be equal to the number of mining areas; (2) Non-mining areas should be far away from known hydrothermal areas. The method for determining "far away" is as follows: count the distance between each hydrothermal area and the nearest hydrothermal area among all hydrothermal areas; sort the above proximity distances from large to small and draw a scatter plot. After eliminating abnormally large values, select the maximum proximity distance as the threshold for non-mining areas to be "far away" from known hydrothermal areas. When the distance between non-mining areas and known hydrothermal areas is greater than the threshold, the non-mining area is considered to be "far away" from the known hydrothermal areas. (3) The location of non-mining areas is random in space. First, the Monte Carlo method is used to generate 10 groups of randomly distributed non-mining areas under the conditions of (1) and (2). Then, the randomness of the locations of the 10 groups of non-mining areas is evaluated by nearest neighbor analysis, and the group with the neighbor ratio (ANN) closest to 1 is selected as the non-mining area. In step 3, the prediction results of the random forest model and the isolation forest model are integrated using a fuzzy operator. Since the prediction results based on the random forest model and the isolation forest model are between 0 and 1, normalization is not required before integration. Specifically: m C =[1-(1-μ A )(1-m B )] γ ×[μ A m B ] 1-γ μ A , μ B are the mineralization potential values ​​predicted by the random forest model and the isolation forest model, μ C is the final result after integration, γ is the value of the γ operator, 0≤γ≤1.

2. The method for quantitative prediction of seabed sulfide mineralization based on machine learning according to claim 1, characterized in that: The method of using nearest neighbor analysis to evaluate randomness is as follows: 1) Calculate the observed average distance: measure the distance between each point and its nearest point, and calculate the mean of these distances, where d i is the distance between the ith point and its nearest point; n represents the number of points in the study area; 2) Calculate the expected average distance: Assuming that all points in the study area are randomly distributed, the expected average distance of these points spread over the entire study area is, Where A represents the area of ​​the study area; 3) Calculate the neighbor ratio (ANN), The closer the ANN value is to 1, the more random the spatial distribution of the points is.

3. The method for quantitative prediction of seabed sulfide mineralization based on machine learning according to claim 1, characterized in that: In step 3, the method for determining the value of the γ operator is: draw a prediction area (PA) map based on the prediction results generated by different γ values. There are two curves in the PA map, which are the percentage of known mineral points contained when it is greater than a certain prediction value, that is, the prediction rate curve; and the percentage curve of the area of ​​the study area occupied when it is greater than the prediction value. The intersection of the two curves is the criterion for evaluating the quality of the prediction results. The higher the position of the intersection in the PA map, the more known hydrothermal areas are contained in a relatively small prediction area, and the more reliable the selected γ operator is. Therefore, the γ value when the intersection of the PA map is the highest is selected as the optimal γ operator, and the prediction result is the final mineralization potential map.

Images