A pegmatite identification method based on geophysical data and deep learning algorithm

Abstract

The present application belongs to the technical field of pegmatite identification, and discloses a pegmatite identification method based on geochemical data and deep learning algorithm. The method collects the stream sediment geochemical data and geological map of the study area, and performs normalization processing on the geochemical data and vectorization processing on the geological map; constructs a data set; performs data oversampling processing, establishes a pegmatite and main lithology classification and identification model based on a deep learning integrated algorithm; evaluates the performance of the pegmatite and main lithology classification and identification model; performs regional pegmatite and main lithology classification and identification; and evaluates the pegmatite and main lithology identification result. The present application can improve the identification accuracy of small sample data such as pegmatite, and provides a new type of identification and mapping technology for quickly and accurately identifying the main lithology and small target lithology such as pegmatite in high mountainous areas or areas with insufficient geological data.

Description

Technical Field

[0001] This invention belongs to the field of pegmatite identification technology, and particularly relates to a pegmatite identification method based on geochemical data and deep learning algorithms. Background Technology

[0002] Lithium resources are crucial raw materials for industries such as electronics, biomedicine, new energy vehicles, photovoltaics, aerospace equipment, and emerging functional metal materials. Therefore, addressing the limited supply of lithium resources requires extensive research and exploration. Currently, global rare metal resources can be categorized into three main deposit types: hard rock, lacustrine, and others. Among these, hard rock lithium deposits, primarily hosted in granitic pegmatites, have attracted significant attention in exploration activities. These pegmatites are attractive due to their widespread distribution, high-grade ore, and the availability of advanced exploration and extraction technologies. Much geological evidence also indicates that certain areas possess substantial lithium resource development potential. However, these areas are characterized by high altitude, harsh ecological environments, relatively low research levels, and limited data accumulation. Therefore, conducting lithological identification, including pegmatite identification, under these challenging conditions, and improving the accuracy of pegmatite identification, is undoubtedly of great significance for finding regional hard rock lithium deposits in similar areas and under similar conditions. However, pegmatites, due to their similarity to granite lithology and small exposed areas, present difficulties in interpretation based on conventional geophysical methods and identification using remote sensing images.

[0003] Based on the above analysis, the problems and defects of the existing technology are as follows: When identifying small targets, the existing technology suffers from an imbalance between pegmatite samples and main lithology samples, resulting in low accuracy in classifying and identifying pegmatite and main lithology in the target area. Summary of the Invention

[0004] To overcome the problems existing in related technologies, this invention discloses an embodiment of a pegmatite identification method based on geochemical data and deep learning algorithms. Specifically, it relates to a technology that utilizes geochemical data and deep ensemble learning algorithms to achieve rapid identification of pegmatites and lithology, and to improve the accuracy of pegmatite identification. The purpose of this invention is to perform geological lithology identification and improve the accuracy of pegmatite identification, providing a new technology for lithology mapping, pegmatite identification, and the search for hard-rock lithium deposits in high-altitude areas, areas with harsh natural environments, and areas with limited data.

[0005] The technical solution is as follows: A pegmatite identification method based on geochemical data and deep learning algorithms, using geochemical data as a foundation, combined with ensemble learning algorithms in deep learning, employing an oversampling data balancing method, constructing 1DCNN, 2D-3D CNN, and DC CNN models, and using a mean ensemble method to classify and identify small target pegmatites and the main lithology; specifically including the following steps:

[0006] A pegmatite identification method based on geochemical data and deep learning algorithms is characterized by the following steps: The method uses geochemical data as a foundation, combines ensemble learning algorithms from deep learning, employs an oversampling data balancing method to construct 1DCNN, 2D-3D CNN, and DC CNN models, and uses a mean ensemble method to classify and identify small target pegmatites and the main lithology; specifically, it includes the following steps:

[0007] S1. Collect geochemical data and geological maps of stream sediments in the study area, normalize the geochemical data, vectorize the geological maps, and label the attributes of each geological body.

[0008] S2, construct the dataset, using geochemical data as feature data and lithological categories as label data;

[0009] S3, data oversampling is performed on the pegmatite samples corresponding to the geochemical data to achieve sample data balance;

[0010] S4. Using the balanced dataset, complete the training of 1D-CNN, 2D-3D CNN, and three different classifiers of dual-channel CNN to establish a pegmatite and main lithology classification and identification model.

[0011] S5 calculates the loss function value, number of iterations, and accuracy to obtain the loss curve and accuracy curve, and evaluates the performance of the pegmatite and main lithology classification and identification model.

[0012] S6 generates three classification results for pegmatite and main lithology identification based on three different recognition models: 1D CNN, 3D-2D CNN and DC-CNN. The mean ensemble method is used to form the regional pegmatite and main lithology classification and identification results of the fourth model.

[0013] S7. Compare the overall accuracy (OA), user accuracy (UA), average accuracy (AA), and Kappa coefficient of the four results to evaluate the identification results of pegmatite and main lithology.

[0014] In step S1, normalization processing is performed, including:

[0015] Geochemical vector data was converted to raster data using inverse distance interpolation. The geometric mean of elemental abundance was calculated, and a CLR transformation was performed on the geochemical data. The geometric mean of elemental abundance was calculated for each sample containing D elements. The expression for the geometric mean g(x) of all elements is:

[0016]

[0017] In the formula, g(x) is the geometric mean of all elements, D is the number of element types, and x j Let be the content value of the j-th element;

[0018] Apply the CLR transform to each element x of each sample, as shown in the expression:

[0019]

[0020] In the formula, CLR(x) is the transformation function of the central logarithmic ratio of element x, x i Let be the content value of the i-th element.

[0021] In step S2, the dataset is constructed, including:

[0022] The normalized geochemical element values ​​at each location are used as feature data. The lithological attributes of the corresponding locations on the geological maps are labeled from 1 to n according to the total number of categories and used as label data. Thus, the label data and feature data together constitute the dataset S for pegmatite and main lithology identification, expressed as:

[0023]

[0024] In the formula, S is the dataset, x i Let y be the feature vector of the i-th sample. i For x i Related category labels, Let Y be an n-dimensional feature data vector space, and Y be a set of lithological attribute labels for geological maps, Y = {1, ..., n}.

[0025] In step S3, data oversampling processing is performed, including:

[0026] Obtain the unique lithology label and corresponding sample count `labelCounts`, calculate the maximum sample count `maxCount` and the inverse ratio `labelInverseRatios`; based on the ratio, duplicate the corresponding number of samples for each lithology category with fewer samples, randomly permutate the newly generated dataset, and return the oversampled and randomly permuted dataset. The expression is:

[0027] labelInverseRatios=maxCount / labelCounts

[0028] In the formula, labelInverseRatios is the reciprocal of the ratio, maxCount is the maximum sample count, and labelCounts is the sample count corresponding to the unique lithology label.

[0029] In step S4, the 1D-CNN recognition model construction method includes:

[0030] Geochemical data were converted into raster data with consistent geochemical element quantities, and lithological vectorized data were rasterized to ensure consistency in range with the geochemical data. A 1D-CNN recognition model was constructed, and its parameters were set and optimized using the Adam optimizer model. 30% of the data set was randomly selected as training data for training the 1D-CNN recognition model, and the model was evaluated based on its accuracy and loss on the training and validation data. The trained data was then used as prediction data for lithological identification and classification on the remaining 70% of the data. The 30% training data covered all lithologies in the study area.

[0031] The 3D-2D CNN recognition model construction method includes: optimizing the model through parameter settings and the Adam optimizer model, and evaluating the model performance by calculating accuracy and loss on training and validation data; then performing lithology identification and classification on the remaining 70% of the data, extracting mixed features of elements and space through 3D convolutional layers; continuing to extract spatial features through deformation and 2D convolutional layers, and applying the Softmax function to the fully connected layers for lithology classification, thus completing the identification of pegmatite and the main lithology.

[0032] Furthermore, by training three different classifiers in a dual-channel CNN, a classification and identification model for pegmatite and main lithology was established, including:

[0033] The dual-channel CNN combining two architectures includes: 1D-CNN channel one and a hybrid CNN channel two composed of 3D-CNN and 2D-CNN; the Softmax function is used to calculate the predicted probability of different categories for each sample, the arg max function is used to find the category index with the highest probability, and the category of the pixel is determined as the category corresponding to the index position with the highest probability, thus completing the identification of pegmatite and main lithology;

[0034] The formula for arg max is:

[0035]

[0036] In the formula, f(x) is the function f applied to the variable x, x is a variable, and S is a set. For each element y in the set, f(y) is the function f applied to the variable y;

[0037] For a function f(x) or a mapping f: X→Y, when the range of x is S, the set of x points for which the result of arg max makes f(x) reach its maximum value indicates that the maximum result value is not searched within the range of all input variables of f(x). The Softmax function is applied to map the feature information output by the fully connected layer to class probabilities that sum to 1.

[0038] In step S5, the performance of the pegmatite and main lithology classification and identification model is evaluated, including:

[0039] The evaluation is based on three metrics: the calculated loss function value, the number of iterations, and the accuracy, forming loss function curves and accuracy curves. The formula for calculating the logarithmic loss value is as follows:

[0040]

[0041] In the formula, L represents the loss function, Y is the true label, P(Y|x) is the probability that the model predicts sample Y belongs to class X, N is the number of samples, and M is the number of classes; p ij It serves as an indicator; if sample i belongs to category j, it is 1, otherwise it is 0.

[0042] In step S6, the mean integration method includes:

[0043] Select the pixel at the corresponding location, and choose the class with the average probability from the three classification results as the result for that pixel to generate a fourth lithological classification map of the study area; the mean integration formula is:

[0044]

[0045] In the formula, N is the number of samples, C is the number of lithological categories, and x i Let ω be the eigenvector. k P(ω) represents the probability value for class discrimination. j |x i Let x be the i-th sample, and let x be its feature vector. i In the case of category ω k The conditional probability; mean is the average conditional probability of the three classification results, and max is the class ω with the highest average conditional probability. j .

[0046] In step S7, the results of pegmatite and main lithology identification are evaluated, including:

[0047] All metrics are analyzed using the confusion matrix [g] ij ] k×k The calculation yields the following formula:

[0048]

[0049]

[0050]

[0051]

[0052]

[0053] In the formula, UA represents user precision, OA represents overall classification precision, Kappa represents the classification coefficient for a specific category, AA represents average precision, and g aa For the number of correctly predicted lithologies of type a, g ab The number of lithologies predicted as class b for class a; M is the total number of pixels, g a+ To sum the columns of a specific category in a confusion matrix, g +a Let G be the row sum of a specific category, and G be the confusion matrix. KK g represents the number of lithologies predicted to be of type k. ak The number of lithologies predicted as class k for class a, UA (N) This represents the accuracy rate for users in the Nth class, where K is the number of sample classes.

[0054] Combining all the above technical solutions, the advantages and positive effects of this invention are as follows: This invention addresses the core problem of how to quickly identify pegmatites in high-altitude areas and improve their identification accuracy by combining geochemical data with deep integrated learning, thereby achieving the identification of pegmatites and the main lithology and improving the identification accuracy of small targets such as pegmatites. The first step involves collecting geochemical data for the study area and normalizing the geochemical data. The second step involves oversampling the pegmatite samples corresponding to the geochemical data to achieve data balance. The third step uses the balanced dataset to train three different classifiers: 1D-CNN, 3D-2D hybrid CNN, and a dual-channel CNN combining two architectures, to establish a lithology classification and identification model. The fifth step involves performance evaluation based on loss and accuracy curves. The sixth step uses the mean ensemble method to obtain the fourth type of pegmatite identification and other lithology identification results. The seventh step compares and analyzes the overall accuracy (OA), user accuracy (UA), average accuracy (AA), and Kappa coefficient of the four models, and finally generates the optimal identification map of pegmatite and other lithologies in the study area based on the mean ensemble method, i.e., the fourth model. This invention can improve the identification accuracy of small sample data such as pegmatite, providing a novel identification and mapping technology for quickly and accurately identifying the main lithology and small target lithologies such as pegmatite in high-altitude areas or areas with insufficient geological data. Attached Figure Description

[0055] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure;

[0056] Figure 1 This is a flowchart of the pegmatite identification method based on geochemical data and deep learning algorithms provided in an embodiment of the present invention;

[0057] Figure 2This is a schematic diagram of a pegmatite identification system based on geochemical data and deep learning algorithms provided in an embodiment of the present invention;

[0058] Figure 3 This is an example diagram of inverse distance interpolation provided in an embodiment of the present invention;

[0059] Figure 4 This is an example diagram of the CLR transformation provided in the embodiments of the present invention;

[0060] Figure 5 This is a data balancing example diagram provided in an embodiment of the present invention;

[0061] Figure 6 This is an example diagram of band synthesis provided in an embodiment of the present invention;

[0062] Figure 7 This is a structural diagram of the 1D-CNN recognition model provided in an embodiment of the present invention;

[0063] Figure 8 This is a structural diagram of the 3D-2D hybrid CNN recognition model provided in an embodiment of the present invention;

[0064] Figure 9 This is a flowchart illustrating the construction process of the dual-channel CNN recognition model (DC-CNN model) provided in this embodiment of the invention.

[0065] Figure 10 These are loss curves from the training and validation sets of the three models provided in this embodiment of the invention;

[0066] Figure 11 These are accuracy curves from the training and validation sets of the three models provided in this embodiment of the invention.

[0067] The diagram shows: 1. Geochemical data processing module; 2. Dataset construction module; 3. Sample data balancing module; 4. Pegmatite and main lithology classification and identification model establishment module; 5. Pegmatite and main lithology classification and identification model performance evaluation module; 6. Pegmatite and main lithology classification and identification module; 7. Pegmatite and main lithology identification result evaluation module. Detailed Implementation

[0068] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0069] The innovation of the pegmatite identification method based on geochemical data and deep learning algorithms provided in this invention is that it opens up a new perspective for pegmatite identification and lithology identification, changes the domestic and foreign pegmatite identification methods that mainly rely on remote sensing images or manual field surveys, and innovatively proposes a pegmatite identification approach and technical method based on deep learning algorithms and combining geochemistry and geological maps.

[0070] This invention departs from the traditional approach of using geophysical interpretation and remote sensing to identify pegmatites. Instead, it innovatively uses geochemical data as a foundation and combines it with ensemble learning algorithms from deep learning to identify small targets. This breaks down the imbalance between pegmatite samples and main lithology samples, enabling the identification of pegmatites and main lithology in the target area. Furthermore, by employing a mean ensemble method, it improves the accuracy of pegmatite and main lithology classification and identification in the region. This provides a new technology for finding pegmatite-type lithium deposits in areas with small targets, weak data foundations, and harsh natural environments.

[0071] Specifically, based on geochemical data and combined with ensemble learning algorithms in deep learning, this study addresses the problem of pegmatite exposure areas being small and the imbalance between its samples and the main lithology samples, which makes identification difficult. An oversampling data balancing method is adopted, and 1D CNN, 2D-3D CNN, DC CNN models and mean ensemble methods are established to classify and identify pegmatite and main lithology in areas with harsh natural environments and weak data foundations.

[0072] Example 1, as Figure 1 As shown, this embodiment of the invention provides a method for identifying pegmatite based on geochemical data and deep learning algorithms, including the following steps:

[0073] S1. Collect geochemical data and geological maps of stream sediments in the study area, normalize the geochemical data, vectorize the geological maps, and label the attributes of each geological body.

[0074] S2, construct the dataset, using geochemical data as feature data and lithological categories as label data;

[0075] S3, data oversampling is performed on the pegmatite samples corresponding to the geochemical data to achieve sample data balance;

[0076] S4. Using the balanced dataset, complete the training of 1D-CNN, 2D-3D CNN, and three different classifiers of dual-channel CNN to establish a pegmatite and main lithology classification and identification model.

[0077] S5 calculates the loss function value, number of iterations, and accuracy to obtain the loss curve and accuracy curve, and evaluates the performance of the pegmatite and main lithology classification and identification model.

[0078] S6 generates three classification results for pegmatite and main lithology identification based on three different recognition models: 1D CNN, 3D-2D CNN and DC-CNN. The mean ensemble method is used to form the regional pegmatite and main lithology classification and identification results of the fourth model.

[0079] S7. Compare the overall accuracy (OA), user accuracy (UA), average accuracy (AA), and Kappa coefficient of the four results to evaluate the identification results of pegmatite and main lithology.

[0080] In step S1 of this embodiment of the invention, normalization is performed as follows: the geochemical vector data is converted into raster data based on inverse distance interpolation, the geometric mean of element content is calculated, and the geochemical data is subjected to CenteredLog-Ratio (CLR) transformation to solve the non-negativity and closure problem of geochemical data, so that the geochemical data is normalized and can be used for subsequent calculations.

[0081] In step S2 of this embodiment of the invention, the dataset is constructed according to the following method:

[0082] The normalized geochemical element values ​​of each point are used as feature data, and the lithological attributes of the corresponding geological map are labeled from 1 to n according to the total number of categories, and used as label data. Thus, the label data and feature data together constitute the dataset S for pegmatite and main lithology identification.

[0083]

[0084] In the formula, S is the dataset, x i Let y be the feature vector of the i-th sample. i For x i Related category labels, Let Y be an n-dimensional feature data vector space, and Y be a set of lithological attribute labels for geological maps, Y = {1, ..., n}.

[0085] In step S3 of this embodiment, oversampling is performed as follows: Addressing the issue of the small surface exposure area of ​​pegmatite as the identification target, oversampling is necessary for this limited sample data. First, unique lithological labels and corresponding sample counts are obtained, and the maximum sample count and the reciprocal of the ratio are calculated. Based on the ratio, a corresponding number of samples are copied for each lithological category with fewer samples, and the newly generated dataset is randomly arranged. This ensures that the distribution of samples from different lithological categories within the dataset is random. Finally, the oversampled and randomly arranged dataset is returned.

[0086] It is understood that the calculation of the maximum sample count and the reciprocal of the ratio; the replication of a corresponding number of samples for each lithology category with fewer samples according to the ratio, and the random permutation of the newly generated dataset are intended to address the problem of data sample imbalance.

[0087] Specifically, the algorithm obtains the unique lithology label and the corresponding sample count (labelCounts), calculates the maximum sample count (maxCount) and the inverse ratio (labelInverseRatios), replicates the corresponding number of samples for each lithology category with fewer samples according to the ratio, and randomly permutes the newly generated dataset. Finally, it returns the oversampled and randomly permuted dataset.

[0088] labelInverseRatios=maxCount / labelCounts

[0089] In the formula, labelInverseRatios is the reciprocal of the ratio, maxCount is the maximum sample count, and labelCounts is the sample count corresponding to the unique lithology label.

[0090] In step S4 of this embodiment, the 1D-CNN recognition model construction method is as follows: Geochemical data is converted into raster data with consistent numbers of geochemical elements, and lithological vectorized data is rasterized to ensure consistency in range with the geochemical data; a 1D-CNN recognition model is constructed, and parameters are set and optimized using the Adam optimizer model; 30% of the data set is randomly selected as training data for training the 1D-CNN recognition model, and the model is evaluated based on its accuracy and loss on the training and validation data; the trained data is used as prediction data to perform lithological identification and classification on the remaining 70% of the data. The 30% training data covers all lithologies in the study area.

[0091] In step S4 of this embodiment of the invention, the 3D-2D hybrid CNN recognition model construction method is as follows: 3D-2D model construction is carried out through parameter setting, Adam optimizer model optimization, etc., and the model performance is evaluated by accuracy and loss calculation on training data and validation data. After the performance is good, the remaining 70% of the data is used for lithological identification and classification. Overall, the hybrid features of elements and space are extracted through 3D convolutional layers, and then spatial features are extracted through deformation and 2D convolutional layers. Finally, the Softmax function is applied to the fully connected layers for lithological classification, completing the identification of pegmatite and main lithology.

[0092] In step S4 of this embodiment, the method for constructing a pegmatite and main lithology classification and identification model using two architectures of dual-channel CNN (DC-CNN model) is as follows: Two architectures of dual-channel CNN (DC-CNN model) are combined, i.e., channel one is 1D-CNN. Channel two is a hybrid CNN composed of 3D-CNN and 2D-CNN. Finally, the Softmax function is used to calculate the predicted probability of different categories for each sample, and then the argmax function is used to find the category index with the highest probability. The category of that pixel is then determined as the category corresponding to the index position with the highest probability. This completes the identification of pegmatite and main lithology. The formula for argmax is as follows:

[0093]

[0094] In the formula, f(x) is the function f applied to the variable x, x is a variable, and S is a set. For each element y in the set, f(y) is the function f applied to the variable y;

[0095] For a function f(x) or a mapping f: X→Y, when the range of x is S, the set of x points for which the result of arg max makes f(x) reach its maximum value indicates that the maximum result value is not searched within the range of all input variables of f(x). The Softmax function is applied to map the feature information output by the fully connected layer to class probabilities that sum to 1.

[0096] In step S5 of this embodiment of the invention, the performance evaluation method of the pegmatite and main lithology classification and identification model is as follows: Based on the calculation of three indicators, namely the loss function value, the number of iterations, and the accuracy, a loss function curve and an accuracy curve are formed. The entire training dataset is fed into the model each time. With each iteration, the model weights are updated. As the number of iterations increases, the model continuously learns the data features, the loss function value decreases, the curve becomes flatter, and the model performance gradually improves. The performance of the pegmatite and main lithology classification and identification model is evaluated based on the curve features. Among them, the loss function value mainly selects the log loss value as the calculation content, and the accuracy mainly refers to the total number of correctly identified individuals / the total number of identified individuals.

[0097] The formula for calculating the logarithmic loss value is as follows:

[0098]

[0099] In the formula, L represents the loss function, Y is the true label, P(Y|X) is the probability that the model predicts sample Y belongs to class X, N is the number of samples, and M is the number of classes; p ij This is an indicator; it is 1 if sample i belongs to category j, and 0 otherwise.

[0100] It is understandable that the purpose of the logarithmic loss value calculation formula is to improve the accuracy of evaluation and identification.

[0101] In step S6 of this embodiment, the mean integration method is as follows: The mean integration method is used to integrate the results of the three classification models. A pixel at a corresponding location is selected, and the class with the average probability among the three classification results is chosen as the result for that pixel, as shown in the following formula. This process generates a fourth lithological classification map of the study area. The expression for selecting the class with the average probability among the three classification results as the result for that pixel is:

[0102]

[0103] In the formula, N is the number of samples, C is the number of lithological categories, and x i Let ω be the eigenvector. k P(ω) represents the probability value for class discrimination. j |x i Let x be the i-th sample, and let x be its feature vector. i In the case of category ω k The conditional probability; mean is the average conditional probability of the three classification results, and max is the class ω with the highest average conditional probability. j .

[0104] The three classification models are the 1D CNN recognition model, the 2D-3D CNN recognition model, and the dual-channel CNN hybrid model that combines the two architectures, namely the DC CNN recognition model.

[0105] It is understood that this invention innovatively proposes the above-mentioned calculation formula and applies geochemical data for lithological classification. The role of this mean integration is to average the identification probability results of the three classification models, thereby improving the identification of pegmatites and the identification accuracy.

[0106] In step S7 of this embodiment of the invention, the method for evaluating the pegmatite and main lithology identification results is as follows:

[0107] All metrics, including User Accuracy (UA), Overall Accuracy (OA), Kappa (classification coefficient) for a specific class, and Average Accuracy (AA), can be obtained using the confusion matrix [g]. ij ] k×k The calculation yields the following formula:

[0108] All metrics are analyzed using the confusion matrix [g] ij ] k×k The calculation yields the following formula:

[0109]

[0110]

[0111]

[0112]

[0113]

[0114] In the formula, UA represents user precision, OA represents overall classification precision, Kappa represents the classification coefficient for a specific category, AA represents average precision, and g aa For the number of correctly predicted lithologies of type a, g ab The number of lithologies predicted as class b for class a; M is the total number of pixels, g a+ To sum the columns of a specific category in a confusion matrix, g +a Let G be the row sum of a specific category, and G be the confusion matrix. KK g represents the number of lithologies predicted to be of type k. ak The number of lithologies predicted as class k for class a, UA( N) Let K represent the User's Accuracy for the Nth class, where K is the number of sample classes.

[0115] It is understandable that the above calculation formula improves the identification of pegmatite and its accuracy.

[0116] As can be seen from the above embodiments, this invention changes the approach of using geophysical interpretation and remote sensing to identify pegmatites. Instead, it innovatively uses geochemical data as a basis and combines ensemble learning algorithms in deep learning to identify small targets. This breaks the imbalance between pegmatite samples and main lithology samples, enabling the identification of pegmatites and main lithology in the target area. Furthermore, by using the mean ensemble method, it improves the accuracy of pegmatite and main lithology classification and identification in the region. This provides a new technology for finding pegmatite-type lithium deposits in areas with small targets, weak data foundations, and harsh natural environments.

[0117] This invention provides an intelligent method for lithium ore exploration and lithology identification in areas with harsh natural environments and weak data foundations, which can greatly reduce the working costs of geological mapping and lithium ore exploration in high-altitude and deeply dissected areas, save a lot of manpower and material resources, and improve the efficiency of geological mapping and mineral exploration.

[0118] This invention fills a gap in the domestic and international methods for finding lithium deposits in high-altitude and cold regions by combining geochemistry, geological maps, and deep learning.

[0119] Efficiently conducting lithium ore exploration in high-altitude, deeply dissected areas has always been a challenging technical problem for geological prospectors. Pegmatites, a key indicator of hard-rock lithium deposits, are particularly problematic due to their small outcrops and difficulty in identification. This invention offers a new perspective on pegmatite and lithological identification, moving away from the current reliance on remote sensing images or manual field surveys. It innovatively proposes a pegmatite identification approach and methodology based on deep learning algorithms, combining geochemistry and geological maps.

[0120] Example 2, as Figure 2 As shown, this embodiment of the invention provides a pegmatite identification system based on geochemical data and deep learning algorithms, comprising:

[0121] Geochemical data processing module 1 is used to collect geochemical data and geological maps of stream sediments in the study area, normalize the geochemical data, vectorize the geological maps, and label the attributes of each geological body.

[0122] Dataset building module 2 is used to build datasets, using geochemical data as feature data and lithological categories as label data;

[0123] Sample data balancing module 3 is used to perform data oversampling on the pegmatite samples corresponding to the geochemical data to achieve sample data balancing.

[0124] Module 4, which establishes a classification and identification model for pegmatite and main lithology, uses a balanced dataset to complete the training of 1D-CNN, 2D-3D CNN, and three different classifiers combined with dual-channel CNN to establish a classification and identification model for pegmatite and main lithology.

[0125] The performance evaluation module 5 for the pegmatite and main lithology classification and identification model is used to evaluate the performance of the pegmatite and main lithology classification and identification model by calculating the loss function value, the number of iterations, and the accuracy, and obtaining the loss curve and accuracy curve.

[0126] The pegmatite and main lithology classification and identification module 6 is used to generate three classification results for pegmatite and main lithology identification based on three different identification models: 1D CNN, 3D-2D CNN and DC-CNN. The mean ensemble method is used to form the regional pegmatite and main lithology classification and identification results of the fourth model.

[0127] The evaluation module 7 for pegmatite and main lithology identification results is used to compare the overall accuracy OA, user accuracy UA, average accuracy AA, and Kappa coefficient of the four results to evaluate the pegmatite and main lithology identification results.

[0128] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0129] The information interaction and execution process between the above-mentioned devices / units are based on the same concept as the method embodiments of the present invention. For details on their specific functions and technical effects, please refer to the method embodiments section, and they will not be repeated here.

[0130] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this invention. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments.

[0131] This invention also provides a computer device comprising: at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor, wherein the processor executes the computer program to implement the steps in any of the above method embodiments.

[0132] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps described in the various method embodiments above.

[0133] This invention also provides an information data processing terminal, which, when executed on an electronic device, provides a user input interface to implement the steps described in the above method embodiments. The information data processing terminal is not limited to mobile phones, computers, or switches.

[0134] This invention also provides a server that, when executed on an electronic device, provides a user input interface to implement the steps described in the above method embodiments.

[0135] This invention provides a computer program product that, when run on an electronic device, enables the electronic device to implement the steps described in the various method embodiments.

[0136] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying the computer program code to a photographing device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks.

[0137] To further illustrate the effects of the embodiments of the present invention, the following experiments were conducted.

[0138] Example 1: Normalization steps for geochemical data:

[0139] In geochemistry, the Centered Log-Ratio (CLR) transformation is a commonly used method for processing component data. Component data refers to polymorphic data where the sum of all components is a constant (e.g., 100% or 1). The CLR transformation can address the non-negativity and closure issues of component data, making the data suitable for statistical analysis. The calculation steps of the CLR transformation are as follows:

[0140] (1) Data preprocessing preparation:

[0141] The 2915 geochemical point vector datasets were converted into 100*100m raster data in ArcGIS using inverse distance interpolation, comprising 83025 raster cells, with each sample containing 19 elements. An example of inverse distance interpolation is shown below. Figure 3 As shown.

[0142] (2) Calculate the geometric mean of element content:

[0143] For each sample, there are D types of elements. The geometric mean g(x) of all elements is calculated as follows:

[0144]

[0145] In the formula, g(x) is the geometric mean of all elements, D is the number of element types, and x jLet be the content value of the j-th element;

[0146] (3) Perform CLR transformation:

[0147] Apply the CLR transform to each element x of each sample, as shown in the expression:

[0148]

[0149] In the formula, CLR(x) is the transformation function of the central log-ratio of element x, where x is the central log-ratio. i Let be the content value of the i-th element.

[0150] The aforementioned calculation formula serves as a preprocessing technique for geochemical data, avoiding nonnegativity and closure issues.

[0151] Suppose there are three samples (sample 1, sample 2, sample 3), each sample contains three elements (x1 + x2 + x3 = 100%). A simplified diagram before and after the CLR transformation is shown below. Figure 4 The diagram shows an example of CLR transformation.

[0152] Experimental Example 2: Oversampling Processing Steps.

[0153] The main purpose of oversampling is to address class imbalance through sampling techniques, especially oversampling of weaker (minority) classes. The specific steps are described below:

[0154] (1) Count the number of samples for each type of lithology: First, count the number of samples for each type of lithology by traversing the dataset and counting the samples for each category.

[0155] (2) Sort and find the lithology category with the most samples: Sort the number of samples for each lithology category to find the lithology category with the most samples.

[0156] (3) Calculate the ratio of the maximum number of samples to the number of samples per lithology category: By calculating the ratio of the maximum number of samples to the number of samples per lithology category, it can be determined how many samples need to be replicated for each lithology category with fewer samples to achieve balance.

[0157] (4) Replicate samples based on ratios: Based on the ratios calculated in the previous step, replicate the corresponding number of samples for each lithological category with fewer samples. This includes replicating geochemical variables and their corresponding lithological labels.

[0158] (5) Randomly arrange the samples: To ensure the randomness of the data and the generalization ability of the model, the newly generated dataset is randomly arranged. This ensures that the distribution of samples of different lithological categories in the dataset is random.

[0159] (6) Return the balanced dataset: Returns the oversampled and randomly permuted dataset, which now has a balanced number of samples for each lithology class. Such a dataset is better suited for training machine learning models because it reduces the likelihood that the model will be biased towards the lithology class with more samples.

[0160] By following the steps above, a new dataset with an equal number of samples for each lithological category can be obtained, thus providing a useful data foundation for further geochemical analysis and lithological classification. This method is simple and effective, and is a commonly used technique for dealing with imbalanced geological data samples. A simple schematic diagram illustrating the process before and after sample balancing, using pegmatite and other lithologies as examples, is shown below. Figure 5 Example diagram of data balancing.

[0161] Experiment Example 3: Steps for building a 1D-CNN recognition model.

[0162] (1) Data preparation.

[0163] In ArcGIS, the Composite Bands tool was used to synthesize 19 geochemical raster data into a single 19-channel raster. The lithological vector surface data was rasterized, with the raster size and extent consistent with the geochemical data. An example of band synthesis is shown below. Figure 6 As shown.

[0164] (2) Constructing a model.

[0165] ① Create a Sequential model in Python. Create an input layer with a tensor shape of 19*1. ② Add a one-dimensional convolutional layer (Conv1D) to extract geochemical features from the input data. It has 32 filters, a kernel size of 3, and the activation function is ReLU. ③ Add a max-pooling layer (MaxPooling1D) to reduce the dimensionality of the geochemical features and retain the most important information. The pooling size is 2. ④ Create a second convolutional layer with the same parameters as the first. ⑤ Create a second max-pooling layer with the same parameters as the first. ⑥ Create a third convolutional layer with similar parameters to the first two, but with a stride of 2. ⑦ Add a flattening layer to convert the multidimensional features into one dimension so that they can be passed to the fully connected layer. ⑧ Create a dropout layer to randomly drop 40% of the nodes during training to prevent overfitting. ⑨ Add a fully connected layer (Dense) to learn the relationships between the features. ⑩ A fully connected layer with 150 nodes was created, using ReLU as the activation function. The 1D-CNN recognition model structure is as follows: Figure 7 As shown.

[0166] (3) Model Compilation: The model compilation specifies the optimizer (such as Adam or SGD), loss function (such as binary_crossentropy or categorical_crossentropy), and evaluation metric (such as accuracy). Define the Adam optimizer to compile the model, using the classification cross-entropy as the loss function. The Adam optimizer performs optimization, and the evaluation metric is accuracy.

[0167] (4) Training the model: Randomly select 30% of the samples in the region as training data and labels to train the model. Specify the training period (epochs is 120) and batch size (batch_size is 32).

[0168] (5) Evaluation and use of the model: Evaluate the performance of the model to determine its accuracy and loss on training and validation data, and use the trained model for lithological identification in the study area.

[0169] Experiment Example 4: Steps for constructing a 3D-2D hybrid CNN recognition model.

[0170] (1) Model Construction: ① Create a Sequential model, stacking layers sequentially. Create an input layer with a tensor shape of 9*9*19*1. ② Create the first 3D convolutional layer (Conv3D), setting 32 filters with a kernel size of (3,3,5) and the activation function ReLU. ③ Create the first 3D max-pooling layer (MaxPooling3D) with a pooling size of (1,1,2). ④ Create the second 3D convolutional layer with the same parameters as the first. ⑤ Create a dropout layer with a dropout rate of 0.4. ⑥ Reshape the output of the 3D convolutional layer into the shape required for the 2D convolutional layer. ⑦ Create a 2D convolutional layer (Conv2D), setting 128 filters with a kernel size of (3,3), a stride of (2,2), and the activation function ReLU. ⑧ Create a dropout layer with a dropout rate of 0.4. ⑨ Flatten the output of the 2D convolutional layer into one dimension. ⑩ Create a dropout layer with a dropout rate of 0.4. An output layer is created with nodes corresponding to the number of lithology categories, and the activation function is Softmax, for lithology classification tasks. The structure of the 3D-2D hybrid CNN recognition model is as follows: Figure 8 As shown.

[0171] (2) Compilation model: The classification cross-entropy is used as the loss function, and the Adam optimizer is used for optimization. The evaluation index is accuracy.

[0172] (3) Training the model: Randomly select 30% of the samples in the region as training data and labels to train the model. Specify the training period (epochs is 120) and batch size (batch_size is 32).

[0173] (4) Evaluation and use of the model: Evaluate the performance of the model to determine its accuracy and loss on the training and validation data, and apply the constructed model to predict the lithology of the study area.

[0174] The constructed hybrid convolutional neural network model first processes three-dimensional data through 3D convolutional layers to extract hybrid features of elements and space, then continues to extract spatial features through deformation and 2D convolutional layers, and finally performs lithological classification through fully connected layers.

[0175] Experiment Example 5: Steps for constructing a dual-channel CNN recognition model (DC-CNN model). For example... Figure 9 As shown.

[0176] A dual-channel CNN (DC-CNN model) consists of a 1D CNN as channel one and a hybrid CNN as channel two, composed of 3D CNN and 2D CNN. Finally, the Softmax function is applied to map the feature information output by the fully connected layers to class probabilities that sum to 1. The output layer feature vector is represented as v, and the i-th feature is x. i The Softmax value can be calculated using the following formula:

[0177]

[0178] In the formula, x is the input vector, i is the index of an element in the vector, and x i Let be the i-th feature input to the fully connected layer, and e be the base of the natural logarithm.

[0179] Experiment Example 6: Three model verification steps.

[0180] To verify the accuracy of the model, the accuracy and loss curves of three recognition models—1D CNN, 2D-3D CNN, and DC CNN—were calculated. The results are as follows: Figure 10 Loss curves for the three models in the training and validation sets, and Figure 11 Accuracy curves for the three models in the training and validation sets.

[0181] It can be seen that the loss functions of the training and validation sets decrease rapidly with increasing iteration count, and then tend to stabilize. The 3D-2D CNN and DC-CNN models converge to a stable state around 40 iterations, with an accuracy of approximately 90%. The DC-CNN model has slightly higher accuracy than the 3D-2D CNN model and exhibits stronger stability. The 1D CNN converges around the 80th training iteration, with an accuracy of approximately 75%. Figure 11Through performance comparison experiments of all the above classification methods, it can be observed that the classification accuracy is improved compared with 3D-2DCNN, DC-CN, and 1D CNN. This proves the effectiveness of the model proposed in this patent for classifying and identifying small samples of pegmatite.

[0182] Experiment Example 7: Validation of results from three models and the mean ensemble method.

[0183] (1) Based on the above steps, four types of pegmatites and main lithology identification results are obtained, and the address map obtained by 1D CNN, 2D-3D CNN, DC CNN and Mean ensemble can be obtained.

[0184] (2) To more intuitively evaluate the model's classification performance, three metrics were calculated: Overall Accuracy (OA), Average Accuracy (AA), and Kappa to assess the model's overall classification performance. Additionally, User's Accuracy (UA) was used to evaluate the classification performance for each specific rock type (see Table 1).

[0185] (2) The 1D CNN model, utilizing only the geochemical information contained in each pixel, had lower classification accuracy than other models (OA = 75.711%, AA = 75.594%, Kappa = 73.584%). In contrast, the 3D-2D CNN model, combining geochemical information with spatial information, showed significantly improved accuracy (OA = 89.446%, AA = 89.541%, Kappa = 88.541%). The use of spatial information is a valuable asset for rock type classification. The DC-CNN model, combining the two methods, showed a slight improvement in accuracy (OA = 89.479%, AA = 89.189%, Kappa = 88.574%). Finally, the ensemble strategy, combining the three models, yielded the best rock type identification results (OA = 90.422%, AA = 90.502%, Kappa = 89.643%).

[0186] For ChSL b2 The method is highly accurate for specific rock types with large areas and simple boundaries, such as γδβT3. In identifying pegmatites (ρ), the integration of the three classification results increased the user accuracy of pegmatite classification to 65.53%, demonstrating that this method significantly improves the accuracy of pegmatite identification.

[0187] Through performance comparison experiments of all the above classification methods, it can be observed that the classification accuracy based on the 3D-2D CNN model is higher than that of the 1D CNN model and the dimensionality reduction hybrid convolutional neural network. The classification performance is stable and excellent, and the combined results of the three classification methods further improve the classification accuracy.

[0188] Table 1 Performance statistics of four classification methods

[0189]

[0190]

[0191] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention and within the spirit and principles of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for identifying pegmatite based on geochemical data and deep learning algorithms, characterized in that, This method, based on geochemical data and combined with ensemble learning algorithms from deep learning, employs an oversampling data balancing method to construct 1DCNN, 2D-3D CNN, and DC CNN models. It then uses a mean ensemble method to classify and identify small-target pegmatites and the main lithology. Specifically, it includes the following steps: S1. Collect geochemical data and geological maps of stream sediments in the study area, normalize the geochemical data, vectorize the geological maps, and label the attributes of each geological body. S2, construct the dataset, using geochemical data as feature data and lithological categories as label data; S3, data oversampling is performed on the pegmatite samples corresponding to the geochemical data to achieve sample data balance; S4. Using the balanced dataset, complete the training of 1D-CNN, 2D-3D CNN, and three different classifiers of dual-channel CNN to establish a pegmatite and main lithology classification and identification model. S5 calculates the loss function value, number of iterations, and accuracy to obtain the loss curve and accuracy curve, and evaluates the performance of the pegmatite and main lithology classification and identification model. S6 generates three classification results for pegmatite and main lithology identification based on three different recognition models: 1D CNN, 3D-2D CNN and DC-CNN. The mean ensemble method is used to form the regional pegmatite and main lithology classification and identification results of the fourth model. S7. Compare the overall accuracy (OA), user accuracy (UA), average accuracy (AA), and Kappa coefficient of the four results to evaluate the identification results of pegmatite and main lithology.

2. The pegmatite identification method based on geochemical data and deep learning algorithms according to claim 1, characterized in that, In step S1, normalization processing is performed, including: Geochemical vector data was converted to raster data using inverse distance interpolation. The geometric mean of elemental abundance was calculated, and a CLR transformation was performed on the geochemical data. The geometric mean of elemental abundance was calculated for each sample containing D elements. The expression for the geometric mean g(x) of all elements is: In the formula, g(x) is the geometric mean of all elements, D is the number of element types, and x j Let be the content value of the j-th element; Apply the CLR transform to each element x of each sample, as shown in the expression: In the formula, CLR(x) is the transformation function of the central logarithmic ratio of element x, x i Let be the content value of the i-th element.

3. The pegmatite identification method based on geochemical data and deep learning algorithms according to claim 1, characterized in that, In step S2, the dataset is constructed, including: The normalized geochemical element values ​​at each location are used as feature data. The lithological attributes of the corresponding locations on the geological maps are labeled from 1 to n according to the total number of categories and used as label data. Thus, the label data and feature data together constitute the dataset S for pegmatite and main lithology identification, expressed as: In the formula, S is the dataset, x i Let y be the feature vector of the i-th sample. i For x i Related category labels, Let Y be an n-dimensional feature data vector space, and Y be a set of lithological attribute labels for geological maps, Y = {1, ..., n}.

4. The pegmatite identification method based on geochemical data and deep learning algorithms according to claim 1, characterized in that, In step S3, data oversampling processing is performed, including: Obtain the unique lithology label and corresponding sample count `labelCounts`, calculate the maximum sample count `maxCount` and the inverse ratio `labelInverseRatios`; based on the ratio, duplicate the corresponding number of samples for each lithology category with fewer samples, randomly permutate the newly generated dataset, and return the oversampled and randomly permuted dataset. The expression is: labelInverseRatios=maxCount / labelCounts In the formula, labelInverseRatios is the reciprocal of the ratio, maxCount is the maximum sample count, and labelCounts is the sample count corresponding to the unique lithology label.

5. The pegmatite identification method based on geochemical data and deep learning algorithms according to claim 1, characterized in that, In step S4, the 1D-CNN recognition model construction method includes: Geochemical data were converted into raster data with consistent geochemical element quantities, and lithological vectorized data were rasterized to ensure consistency in range with the geochemical data. A 1D-CNN recognition model was constructed and its parameters were set, and optimized using the Adam optimizer model. 30% of the data set was randomly selected as training data for training the 1D-CNN recognition model, and the model was evaluated based on its accuracy and loss on the training and validation data. The trained data was then used as prediction data for lithological identification and classification on the remaining 70% of the data. The 30% training data covered all lithologies in the study area.

6. The pegmatite identification method based on geochemical data and deep learning algorithms according to claim 1, characterized in that, In step S4, the 3D-2D CNN recognition model construction method includes: optimizing the model through parameter settings and the Adam optimizer model, and evaluating the model performance by calculating accuracy and loss on training and validation data; performing lithology identification and classification on the remaining 70% of the data, extracting mixed features of elements and space through 3D convolutional layers; continuing to extract spatial features through deformation and 2D convolutional layers, and applying the Softmax function to classify the fully connected layers for lithology classification, thus completing the identification of pegmatite and the main lithology.

7. The pegmatite identification method based on geochemical data and deep learning algorithms according to claim 1, characterized in that, By training three different classifiers in a dual-channel CNN, a classification and identification model for pegmatite and main lithology is established, including: The dual-channel CNN combining two architectures includes: 1D-CNN channel one and a hybrid CNN channel two composed of 3D-CNN and 2D-CNN; the Softmax function is used to calculate the predicted probability of different categories for each sample, the arg max function is used to find the category index with the highest probability, and the category of the pixel is determined as the category corresponding to the index position with the highest probability, thus completing the identification of pegmatite and main lithology; The formula for arg max is: In the formula, f(x) is the function f applied to the variable x, x is a variable, and S is a set. For each element y in the set, f(y) is the function f applied to the variable y; For a function f(x) or a mapping f:X→Y, when the range of x is S, the set of x points for which the result of arg max makes f(x) reach its maximum value indicates that the maximum result value is not searched within the range of all input variables of f(x). The Softmax function is applied to map the feature information output by the fully connected layer to class probabilities that sum to 1.

8. The pegmatite identification method based on geochemical data and deep learning algorithms according to claim 1, characterized in that, In step S5, the performance of the pegmatite and main lithology classification and identification model is evaluated, including: The evaluation is based on three metrics: the calculated loss function value, the number of iterations, and the accuracy, forming loss function curves and accuracy curves. The formula for calculating the logarithmic loss value is as follows: In the formula, L represents the loss function, Y is the true label, P(Y|X) is the probability that the model predicts sample Y belongs to class X, N is the number of samples, and M is the number of classes; p ij It serves as an indicator; if sample i belongs to category j, it is 1, otherwise it is 0.

9. The pegmatite identification method based on geochemical data and deep learning algorithms according to claim 1, characterized in that, In step S6, the mean integration method includes: Select the pixel at the corresponding location, and choose the class with the average probability from the three classification results as the result for that pixel to generate a fourth lithological classification map of the study area; the mean integration formula is: In the formula, N is the number of samples, C is the number of lithological categories, and x i Let ω be the eigenvector. k P(ω) represents the probability value for class discrimination. j |x i Let x be the i-th sample, and let x be its feature vector. i In the case of category ω k The conditional probability; mean is the average conditional probability of the three classification results, and max is the class ω with the highest average conditional probability. j .

10. The pegmatite identification method based on geochemical data and deep learning algorithms according to claim 1, characterized in that, In step S7, the results of pegmatite and main lithology identification are evaluated, including: All metrics are analyzed using the confusion matrix [g] ij ] k×k The calculation yields the following formula: In the formula, UA represents user precision, OA represents overall classification precision, Kappa represents the classification coefficient for a specific category, AA represents average precision, and g aa g represents the number of correctly predicted lithologies of type a. ab The number of lithologies predicted as class b for class a; M is the total number of pixels, g a+ To sum the columns of a specific category in a confusion matrix, g +a Let G be the row sum of a specific category, and G be the confusion matrix. KK g represents the number of lithologies predicted to be of type k. ak The number of lithologies predicted as class k for class a, UA (N) This represents the accuracy rate for users in the Nth class, where K is the number of sample classes.

Images