Method and system for identifying ore-bearing pegmatite based on remote sensing geochemical fusion data

Abstract

The present application belongs to the technical field of pegmatite identification, and discloses a method and system for identifying ore-bearing pegmatite based on remote sensing geochemical fusion data. The method is based on remote sensing, geochemical and geological map data, and based on a 2DCNN-LSTM model to complete the fusion of remote sensing and geochemical data, and based on a deep priori super-resolution model DIP to improve the resolution of the fused data. Finally, based on a composite field perception deep metric learning (DNM-DML) model, an ore-bearing pegmatite identification model is constructed to complete the identification of ore-bearing pegmatite and other lithologies. The present application can not only realize the identification of pegmatite and other lithologies, but also realize the identification of ore-bearing pegmatite due to the addition of geochemical element information. The method provides a new method for identifying ore-bearing pegmatite and plays an important role in accurately finding pegmatite type lithium mines.

Description

Technical Field

[0001] This invention belongs to the field of pegmatite identification technology, and particularly relates to a method and system for identifying mineral-bearing pegmatites based on remote sensing geochemical fusion data. Background Technology

[0002] Lithium (Li) is used in fields such as atomic energy, special alloys, special glasses, and storage batteries, and is especially known as an "energy metal" due to its application in lithium batteries. In February 2022, the United States released a list of 50 critical minerals, and lithium, a rare metal in global scarcity, was included. As a 21st-century energy metal, lithium supports the development of my country's new energy and clean energy industries and is one of my country's important strategic minerals, playing a vital role in addressing bottlenecks in my country's energy resources and environmental protection. However, my country's dependence on imported lithium is as high as 80%, making it highly susceptible to international supply events, which seriously affects my country's economic and national defense security. In 2016, my country's lithium consumption was 92,400 tons, and it is projected that by 2030, my country's lithium resource demand will reach 1.95 million tons, indicating a very severe supply and demand situation. In general, geophysical methods are rarely used to explore pegmatite deposits in China. This is mainly because the physical properties of pegmatites are not significantly different from those of granite, schist, and gneiss, especially from granite, making them difficult to distinguish. Furthermore, pegmatite bodies are generally small in scale, making it difficult to produce significant anomalies using conventional geophysical methods such as gravity, magnetic, and electrical methods. Therefore, using remote sensing technology to identify pegmatites and invert elemental content provides a new approach and means for lithium deposit detection. However, the identification of ore-bearing pegmatites requires geochemical indicators, but the sampling intervals for geochemical data are often large, making it difficult to meet the requirements for identifying ore-bearing pegmatites. Summary of the Invention

[0003] To overcome the problems existing in related technologies, the present invention discloses a method and system for identifying ore-bearing pegmatites based on remote sensing geochemical fusion data. The purpose of this invention is to provide a method based on the fusion of hyperspectral and geochemical data, improving the resolution of geochemical data. Furthermore, it applies super-resolution technology to enhance the detail information of the fused data and introduces a deep composite neighborhood sensing network to achieve ore-bearing pegmatite identification. Moreover, this invention uses a deep learning algorithm to fuse remote sensing data and geochemical data, completing element prediction and improving the resolution of geochemical data, providing high-resolution fused data for ore-bearing pegmatite identification. Based on this, it applies super-resolution technology to enhance the detail information of the fused data and introduces a deep composite neighborhood sensing network for ore-bearing pegmatite identification, providing a new approach and technology for the comprehensive identification of ore-bearing pegmatites based on multi-source data.

[0004] The technical solution is as follows: A method for identifying mineral-bearing pegmatites based on remote sensing geochemical fusion data, comprising:

[0005] S1. Collect geochemical data of stream sediments in the study area and send them for testing to obtain elemental content. Organize the latitude and longitude information of the sampling location, convert the two-dimensional table data of elemental content into vector data, and apply the inverse distance weighted interpolation method to convert the vector data into raster data.

[0006] S2, based on the standardized correction DS framework, applies a non-negative least squares model to correct remote sensing images;

[0007] S3. Based on the Pearson correlation coefficient, calculate the coefficient between the spectral transformation and the processed results and the elemental content, select the transformation method with the highest correlation coefficient, and determine the prediction order of major and trace elements.

[0008] S4. The remote sensing image is divided into low-frequency and high-frequency data by Gaussian filtering. A 2D CNN-LSTM network structure is applied to establish the relationship between low-frequency data and collected geochemical data of water system sediments. The trained CNN-LSTM model is applied to the high-frequency data to predict the content of anomalous elements. The low-frequency data is superimposed on the high-frequency data to complete the fusion of remote sensing geochemical data.

[0009] S5 downsamples remote sensing geochemical fusion data into low-resolution image data, optimizes the structure parameters of the deep prior network (DIP), learns image details, and obtains fusion data with more spatially resolved detail information.

[0010] S6 uses a combination of convolutional neural networks and graph neural networks to construct a composite neighborhood perception network. It employs deep metric learning to mine the characteristics of different lithological samples, reduce intra-class distance, increase inter-class distance, and complete the identification of mineral-bearing pegmatites and their related lithologies.

[0011] S7 uses user precision (UA), producer precision (PA), and overall precision (OA, AA, and kappa coefficients) to evaluate the model recognition performance of pegmatites and other lithologies.

[0012] In step S1, the element content is obtained, including: constructing vector data with coordinate information and geochemical content from the geochemical content data and corresponding coordinate data obtained from the test; converting the vector data into raster data using the inverse distance weighted interpolation method, wherein the raster size in the raster data takes into account the sampling interval of the geochemical data and adopts a size between the minimum sampling point spacing and the maximum sampling point spacing;

[0013] The specific steps of the inverse distance weighted interpolation method are as follows: ① Determine the coordinates of the unknown point and the known points; ② Calculate the distance between the unknown point and each known point; ③ Calculate the weight of each known point based on the distance. The formula for calculating the weight is: In the formula, w is the weight and d is the distance; ④ Calculate the value of the unknown point based on the weight and the values ​​of the known points. The calculation formula is: In the formula, z is the value of the unknown point, and w i Given the weight of point i, z i Let i be the value of a known point.

[0014] In step S2, the non-negative least squares model is applied to correct the remote sensing image, including: extracting hyperspectral image data corresponding to the sampling points on the opposite side and forming a two-dimensional list; resampling the ground indoor spectral data according to the image spectral data; using the resampled ground indoor spectral data as the target variable and inputting it into the non-negative least squares NNLS model to correct the hyperspectral image data, thereby obtaining the corrected hyperspectral image data.

[0015] Nonnegative least squares is used to calculate nonnegative spectral reflectance values. In a given linear system, it finds a nonnegative vector x such that ||Ax-b|| 2 The minimum value is given by formula (1) as follows: min x ||Ax-b|| 2 x≥0(1);

[0016] Where x≥0 constrains each element to be nonnegative, min x Let x be the minimum value of x, A be the coefficient matrix, x be the image spectral variable, and b be a set of constants.

[0017] In step S3, determining the prediction order of major and trace elements includes: performing first-order, second-order, standard normalization, and envelope removal processing on the spectral data of the corrected GF-5 remote sensing image; calculating the coefficient between the transformed and processed results and the elemental content based on the Pearson correlation coefficient; selecting the transformation method with the highest correlation coefficient; determining the prediction order of major elements based on the average correlation coefficient between elements with significant responses in the visible-shortwave infrared band and their spectra; and determining the prediction order of trace elements based on the average correlation coefficient between trace elements and major elements. The formula for calculating the Pearson correlation coefficient is: In the formula, Cov(X, Y) is the covariance of X and Y, Var[X] is the variance of X, and Var[Y] is the variance of Y.

[0018] (1) Spectral data preprocessing: The influence of linear or near-linear noise spectrum and background on the target spectrum is eliminated using the first-order differential processing formula (10); the baseline drift and background signal are eliminated using the second-order differential processing formula (11). The spectral differential transformation calculation formula is as follows: In the formula, R′(λ) i R″(λ) is the first derivative of the spectral reflectance. i R(λ) is the second derivative of the spectral reflectance. i+1 ) represents the wavelength λ i+1Reflectivity at point R(λ) i-1 ) represents the wavelength λ i-1 Reflectance at point R′(λ) i+1 ) represents the wavelength λ i+1 The first derivative of the reflectivity at point R′(λ) i-1 ) represents the wavelength λ i-1 The first derivative of reflectance at a given point, Δλ is the wavelength interval; using the envelope removal method CR formula (12), the absorption peak characteristics of the spectrum are highlighted to the maximum extent, and the differences between spectral curves are distinguished; the envelope removal transformation calculation formula is as follows: In the formula, R′ c (λ i ) represents the wavelength λ i The envelope value at point R(λ) i ) represents the wavelength λ i The reflectivity value, R c (λ i ) represents the corresponding wavelength λ i The value at the envelope;

[0019] (2) Correlation calculation: Analyze the correlation between the transformed data and the element content, calculate the Pearson correlation coefficient, and take the highest value of the Pearson correlation coefficient as the optimal transformation method;

[0020] (3) Ranking of principal and heavy metal elements: Select principal and heavy metal elements that have obvious responses in the visible to short-wave infrared bands, calculate the spectral correlation of these elements, and rank them.

[0021] (4) Calculation of the correlation coefficient of trace elements: Further analysis of the correlation between trace elements and the above major elements.

[0022] In step S4, remote sensing geochemical data fusion is completed, including: processing remote sensing images based on the optimal transformation method, dividing remote sensing images into low-frequency data and high-frequency data using Gaussian filtering; establishing the relationship between low-frequency data and collected stream sediment geochemical data using a 2DCNN-LSTM network structure; determining the element prediction order according to the Pearson correlation coefficient between spectra and elements from high to low, and predicting element content sequentially; adding the estimation results of major elements and heavy metals with high spectral response to the trace element estimation model; applying the trained model to high-frequency data to predict anomalous element content; and superimposing the background geochemical content data retrieved from low-frequency data with the anomalous geochemical content data retrieved from high-frequency data to complete the remote sensing geochemical data fusion.

[0023] In step S5, the structural parameters of the deep prior network (DIP) are optimized to learn image details and obtain fused data with more spatially resolved detail information, including:

[0024] This paper utilizes a generative network architecture to leverage image statistical priors, enabling the network to automatically reduce noise and prioritize the reconstruction of image details during the iterative process. After a certain number of iterations, optimization stops without overfitting the noise, restoring the original image resolution. The expression is as follows: In the formula, The optimized parameters are θ, which represents the network parameters, fθ, which represents the constructed deep network, x, which represents the low-resolution image, and z, which represents the input noise.

[0025] In step S6, the identification of ore-bearing pegmatites and their related lithologies is completed, including: rasterizing the vectorized lithology data to ensure that the geochemical raster data and the lithology raster data have the same range; constructing a composite neighborhood-aware network identification model using a convolutional neural network combined with a graph neural network, and setting the parameters; wherein, feature extraction is performed based on the composite neighborhood-aware CoConv, and mean calculation is performed based on standard image convolution to transform the features into a metric space, and a nonparametric mini-batch metric classifier is used to classify pegmatites and other lithologies; during the model training process, the standard Adam optimizer is used to train the composite neighborhood-aware deep metric learning model.

[0026] In step S7, the model recognition effect of pegmatite and other lithologies is evaluated, including: using the geological map as a reference map of the actual results, using the pegmatite and other lithology recognition results as the classification result map, establishing a confusion matrix of the classification results, and using the indicators through the confusion matrix [g ij ] k×k The calculation yields the following formula: In the formula, G is the confusion matrix, where each row represents the actual class and each column represents the predicted class; g ij The number of samples whose actual class is i and whose predicted class is j;

[0027] In the formula, OA represents the overall accuracy, and g aa The diagonal elements are k, the number of categories, and g. ab The number of samples that are actually class a but are predicted as class b is determined by the diagonal elements g of the confusion matrix G. aa The sum of these values ​​divided by the total number of samples is used to calculate the overall accuracy of the model.

[0028] UA (n) The accuracy rate for the nth class is calculated by dividing the number of correctly classified samples in the nth class by the number of predicted samples in the nth class; this represents the proportion of correct identifications for class n. nn Let g be the value on the diagonal of the nth type confusion matrix. na This represents the element in the nth row of the confusion matrix;

[0029] In the formula, AA is the average precision, which is used to measure the average performance of the classification model across all categories. The average precision of all categories is obtained by summing the precision of each category and then dividing by the number of categories K.

[0030] In the formula, PA (n) The user accuracy for class n is calculated by dividing the number of correctly classified samples in class n by the actual number of samples in class n; it represents the proportion of class n that is correctly identified. n The element in the nth column of the confusion matrix;

[0031] In the formula, Kappa is the evaluation index used to test the consistency of classification results, M is the total number of samples, and g is the weight of the samples. a+ Let g represent the sum of the a-th row of the matrix. +a This is the sum of column a.

[0032] Another objective of this invention is to provide a system for identifying mineral-bearing pegmatites based on remote sensing geochemical fusion data. This system implements the aforementioned method for identifying mineral-bearing pegmatites based on remote sensing geochemical fusion data. The system includes:

[0033] The geochemical data fusion module is used to fuse remote sensing, geochemical and geological map data based on a 2DCNN-LSTM model.

[0034] The fusion data resolution module improves the resolution of fused data based on the deep prior super-resolution model (DIP).

[0035] The module for identifying mineralized pegmatites and other lithologies is used to construct a mineralized pegmatite identification model based on the composite domain perception depth metric learning (DNM-DML) model to complete the identification of mineralized pegmatites and other lithologies.

[0036] Combining all the above technical solutions, the beneficial effects of this invention are as follows: This method collects geochemical data and geological maps of stream sediments in the study area, normalizes the geochemical data, and vectorizes the geological maps; constructs a dataset; fuses remote sensing data and geochemical data based on a two-dimensional convolutional neural network-long short-term memory network (2DCNN-LSTM) model to form remote sensing-geochemical fused data; improves the resolution of the remote sensing-geochemical fused data based on the depth image prior algorithm (DIP) to meet the identification requirements of pegmatite veins; and establishes a mineral-bearing pegmatite identification model based on a composite domain-aware depth metric learning (DNM-DML) model to classify and identify mineral-bearing pegmatites and other lithologies in the region, and evaluates the identification results. This invention can not only identify pegmatites and other lithologies, but also identify mineral-bearing pegmatites by incorporating geochemical element information. This method provides a new approach for identifying mineral-bearing pegmatites and plays an important role in further narrowing the prospecting range of pegmatite-type lithium deposits.

[0037] The technical solution of this invention, once transformed, is beneficial to the discovery of lithium deposits, saving prospecting companies significant manpower and resources and providing substantial economic benefits. It fills the technological gap in identifying ore-bearing pegmatites based on the fusion of remote sensing and geochemical data. The accurate discovery and identification of ore-bearing pegmatites has always been a challenge in mineral exploration, especially in high-altitude, deeply dissected areas inaccessible to human access. The discovery of ore-bearing pegmatites is a direct indicator of lithium deposit discovery. Due to the high altitude, many prospecting equipment cannot reach them, and their physical parameters are difficult to distinguish from other geological bodies. Geological prospectors have long sought to establish a method for identifying ore-bearing pegmatites using minimal sampling and non-contact methods. This invention aims to solve this problem by establishing a method for identifying ore-bearing pegmatites using only minimal sampling data and a combination of non-contact remote sensing data, providing an unprecedented technical solution for lithium deposit discovery. Attached Figure Description

[0038] 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;

[0039] Figure 1 This is a flowchart of the method for identifying mineral-bearing pegmatites based on remote sensing geochemical fusion data provided in this embodiment of the invention;

[0040] Figure 2 This is a schematic diagram of the method for identifying mineral-bearing pegmatites based on remote sensing geochemical fusion data provided in this embodiment of the invention.

[0041] Figure 3 This is a diagram of the 2D CNN-LSTM network structure provided in an embodiment of the present invention;

[0042] Figure 4This is a structural diagram of the composite domain-aware deep metric learning model provided in an embodiment of the present invention;

[0043] Figure 5 This is a simulation diagram illustrating the principle of the inverse distance interpolation example provided in this embodiment of the invention;

[0044] Figure 6 This is a flowchart of the spectral correction process provided in an embodiment of the present invention;

[0045] Figure 7 This is a flowchart of remote sensing geochemical data fusion provided in an embodiment of the present invention;

[0046] Figure 8 This is an IDW interpolation diagram of elemental content provided in an embodiment of the present invention;

[0047] Figure 9 Estimate element content fusion maps for 2D CNN-LSTM models;

[0048] Figure 10 This invention provides a super-resolution result false-color composite image;

[0049] Figure 11 Loss curves for the model training and validation sets;

[0050] Figure 12 An accuracy curve for the model training and validation sets;

[0051] Figure 13 This is an image showing the identification results of the mineral-bearing pegmatite and the main lithology provided in an embodiment of the present invention;

[0052] Figure 14 This is a confusion matrix diagram provided in an embodiment of the present invention. Detailed Implementation

[0053] 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.

[0054] The innovation of this invention lies in the fact that it innovatively combines deep learning models for mining spatial features (2DCNN) and sequence features (LSTM) to form remote sensing geochemical fusion data with spatial features and encrypted element content features, and successfully introduces a composite domain-aware deep metric learning (DNM-DML) model to achieve the extraction and identification of mineral-bearing pegmatites.

[0055] Example 1, as Figure 1 As shown in the embodiments of the present invention, the method for identifying ore-bearing pegmatites based on remote sensing, geochemical, and geological map data is based on remote sensing, geochemical, and geological map data. It uses a 2DCNN-LSTM model to fuse remote sensing and geochemical data, improves the resolution of the fused data using a depth prior super-resolution model (DIP), and finally constructs an ore-bearing pegmatite identification model based on a domain-aware depth metric learning (DNM-DML) model to identify ore-bearing pegmatites and other lithologies. Specifically, it includes the following steps:

[0056] S1. Collect geochemical data of stream sediments in the study area and send them for testing to obtain elemental content. Organize the latitude and longitude information of the sampling location, convert the two-dimensional table data of elemental content into vector data, and apply the inverse distance weighted interpolation method to convert the vector data into raster data.

[0057] S2, based on the standardized correction DS (standardized correction) framework, applies a non-negative least squares model to correct remote sensing images;

[0058] S3. Based on the Pearson correlation coefficient, calculate the coefficient between the spectral transformation and the processed results and the elemental content, select the transformation method with the highest correlation coefficient, and determine the prediction order of major and trace elements.

[0059] S4. The remote sensing image is divided into low-frequency and high-frequency data by Gaussian filtering. A 2D CNN-LSTM network structure is applied to establish the relationship between low-frequency data and collected geochemical data of water system sediments. The trained CNN-LSTM model is applied to the high-frequency data to predict the content of anomalous elements. The low-frequency data is superimposed on the high-frequency data to complete the fusion of remote sensing geochemical data.

[0060] S5 downsamples remote sensing geochemical fusion data into low-resolution image data, optimizes the structure parameters of the deep prior network (DIP), learns image details, and obtains fusion data with more spatially resolved detail information.

[0061] S6 uses a combination of convolutional neural networks and graph neural networks to construct a composite neighborhood perception network. It employs deep metric learning to mine the characteristics of different lithological samples, reduce intra-class distance, increase inter-class distance, and complete the identification of mineral-bearing pegmatites and their related lithologies.

[0062] S7 uses User's Accuracy (UA), Producer's Accuracy (PA), Overall Accuracy (OA), Average Accuracy (AA), and kappa coefficient to evaluate the model recognition performance of pegmatites and other lithologies.

[0063] In an embodiment of the present invention, Figure 2 This is a schematic diagram of the principle of the mineral-bearing pegmatite identification method based on remote sensing geochemical fusion data provided in this embodiment of the invention.

[0064] In step S1 of this embodiment of the invention, obtaining the element content includes: combining the geochemical content data obtained from the test and the corresponding coordinate data to form vector data with coordinate information and geochemical content; the grid size in the raster data should take into account the sampling interval of the geochemical data, and is generally between the minimum sampling point spacing and the maximum sampling point spacing.

[0065] In step S2 of this embodiment of the invention, correcting the remote sensing image includes: extracting hyperspectral image data corresponding to the sampling point on the opposite side and forming a two-dimensional list; resampling the ground indoor spectral data according to the image spectral data; using the resampled ground indoor spectral data as the target variable and inputting it into the non-negative least squares (NNLS) model to correct the hyperspectral image data, thereby obtaining the corrected hyperspectral image data.

[0066] Non-negative Least Squares (NNLS) is a special type of least squares algorithm designed to find the optimal non-negative solution. It is used to calculate the non-negativity of spectral reflectance values. In a given linear system, it finds a non-negative vector x such that ||Ax-b|| 2 The minimum value is given by formula (1) as follows:

[0067] min x ||Ax-b|| 2 x≥0 (1)

[0068] Where x≥0 constrains each element to be non-negative, minn x Let x be the minimum value of x, A be the coefficient matrix, x be the image spectral variable, and b be a set of constants.

[0069] In step S3, determining the prediction order of major and trace elements includes: performing first-order, second-order, standard normal variable, and envelope removal processing on the spectral data of the corrected GF-5 remote sensing image; calculating the coefficient between the transformed and processed results and the element content based on the Pearson correlation coefficient; selecting the transformation method with the highest correlation coefficient; determining the prediction order of major elements based on the average correlation coefficient between elements (major elements and heavy metal elements, etc.) that have significant responses in the visible-shortwave infrared band and the spectrum; and determining the prediction order of trace elements based on the average correlation coefficient between trace elements and major elements.

[0070] Pearson correlation coefficient calculation formula:

[0071]

[0072] In the formula, Cov(X, Y) is the covariance of X and Y, Var[X] is the variance of X, and Var[Y] is the variance of Y.

[0073] In step S4, the remote sensing geochemical data fusion includes: processing the GF-5 remote sensing image based on the optimal transformation method, and dividing the GF-5 remote sensing image into low-frequency and high-frequency data using Gaussian filtering; establishing the relationship between the low-frequency data and the collected hydrogeochemical data of river sediments using a 2D CNN-LSTM network structure; where 2D CNN is a convolutional neural network model with two-dimensional convolutional kernels, its advantage lies in extracting spatial features; Long Short-Term Memory (LSTM) is a special type of recurrent neural network (RNN) that can learn long-term dependencies and has achieved breakthrough results in many sequence modeling problems. This invention determines the element prediction order according to the Pearson correlation coefficient between spectra and elements from high to low, and predicts the element content sequentially. The estimation results of major elements and heavy metals with high spectral response are added to the trace element estimation model; the trained model is also applied to the high-frequency data to predict anomalous element content; and the background geochemical content data retrieved from low-frequency inversion is superimposed with the anomalous geochemical content data retrieved from high-frequency inversion to complete the remote sensing geochemical data fusion. Among them, the 2D CNN-LSTM network structure Figure 3 As shown in the figure, S represents the spectral dataset, and G... n This represents the nth element.

[0074] In step S5, the Deep Prior Network (DIP) algorithm utilizes image statistical priors through a generative network architecture. These priors are not derived from pre-trained samples but are obtained directly from the network structure. During iteration, the algorithm automatically denoises and prioritizes the reconstruction of image details. When the network has iterated a certain number of times, optimization stops without overfitting the noise, thus recovering a high-quality, high-resolution image.

[0075]

[0076] In the formula, The optimized parameters are θ, which represents the network parameters, fθ, which represents the constructed deep network, x, which represents the low-resolution image, and z, which represents the input noise.

[0077] In step S6, the identification of ore-bearing pegmatites and their related lithologies includes: rasterizing the vectorized lithology data to ensure consistency between the geochemical raster data and the lithology raster data; constructing a composite neighborhood-aware network identification model using a convolutional neural network combined with a graph neural network, setting parameters, where feature extraction is performed based on composite neighborhood awareness CoConv, and mean calculation is performed based on standard image convolution to transform the features into a metric space; finally, a non-parametric mini-batch metric classifier is used to classify pegmatites and other lithologies; during model training, the standard Adam optimizer is used to train the composite neighborhood-aware deep metric learning model, where approximately 30%, 0.5%, and 69.5% of samples are randomly selected from each category as training, validation, and test datasets, respectively; the number of iterations and accuracy are determined based on the premise that the loss and accuracy (acc) of the training and validation sets tend to stabilize; the structure of the composite neighborhood-aware deep metric learning model is as follows. Figure 4 As shown.

[0078] In step S7, the evaluation of the model recognition effect of pegmatite and other lithologies includes: using the geological map as a reference map of the actual results, using the recognition results of pegmatite and other lithologies as the classification result map, establishing a confusion matrix of the classification results, and using indicators such as UA, PA, OA, AA and kappa coefficient to evaluate the model recognition effect of pegmatite and other lithologies.

[0079] All indicators can be obtained through the confusion matrix [g] ij ] k×k The calculation yields the following formula:

[0080]

[0081] In the formula, G is the confusion matrix, where each row represents the actual class and each column represents the predicted class; g ij The number of samples whose actual class is i and whose predicted class is j;

[0082]

[0083] In the formula, OA represents the overall accuracy, and g aa The diagonal elements are k, the number of categories, and g. ab The number of samples that are actually class a but are predicted as class b is determined by the diagonal elements g of the confusion matrix G. aa The sum of these values ​​divided by the total number of samples is used to calculate the overall accuracy of the model.

[0084]

[0085] UA(n) represents the user accuracy for class n, calculated by dividing the number of correctly classified samples in class n by the number of predicted samples in class n; it is the proportion of correct identification for class n.nn Let g be the value on the diagonal of the nth type confusion matrix. na This represents the element in the nth row of the confusion matrix;

[0086]

[0087] In the formula, AA is the average precision, which is used to measure the average performance of the classification model across all categories. The average precision of all categories is obtained by summing the precision of each category and then dividing by the number of categories K.

[0088]

[0089] In the formula, PA(n) is the user accuracy rate for the nth class, calculated by dividing the number of correctly classified samples in the nth class by the actual number of samples in the nth class; it represents the proportion of class n that is correctly identified. n The element in the nth column of the confusion matrix;

[0090]

[0091] In the formula, Kappa is the evaluation index used to test the consistency of classification results, M is the total number of samples, and g is the weight of the sample. a+ Let g represent the sum of the a-th row of the matrix. +a This is the sum of column a.

[0092] Example 2: This embodiment of the invention provides a system for identifying mineral-bearing pegmatites based on remote sensing geochemical fusion data. The system includes:

[0093] The geochemical data fusion module is used to fuse remote sensing, geochemical and geological map data based on a 2DCNN-LSTM model.

[0094] The fusion data resolution module improves the resolution of fused data based on the deep prior super-resolution model (DIP).

[0095] The module for identifying mineralized pegmatites and other lithologies is used to construct a mineralized pegmatite identification model based on the composite domain perception depth metric learning (DNM-DML) model to complete the identification of mineralized pegmatites and other lithologies.

[0096] 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.

[0097] Experiment Example 1, Data Preparation and Preprocessing: The vector dataset of 383 geochemical points was converted into 100*100m raster data in ArcGIS using the inverse distance interpolation method, comprising 5410 rasters, with each sample containing 19 elements. The simulation diagram of the inverse distance interpolation case is shown below. Figure 5 As shown.

[0098] Experiment Example 2: Direct Standardization (DS) procedure.

[0099] (1) Spectral data extraction: Extract the spectrum of the ground sampling point location from the image data, pull the hyperspectral cube data into a two-dimensional hyperspectral data, with rows representing the number of samples and columns representing the number of bands in a two-dimensional list.

[0100] (2) Construction of non-negative least squares model: Hyperspectral extracted data is used as variables, and indoor spectral data resampled according to the image spectral resolution is used as target variables, and input into the non-negative least squares (NNLS) model.

[0101] (3) Model application: The established NNLS model is applied to two-dimensional hyperspectral data for spectral correction.

[0102] (4) Hyperspectral data output: Organize the data output by the NNLS model into a structured two-dimensional table and output the corrected hyperspectral data.

[0103] (5) Hyperspectral data reconstruction: The hyperspectral data is reconstructed into hyperspectral cube data to obtain the corrected image hyperspectral data. The spectral correction process is as follows: Figure 6 As shown.

[0104] Experiment Example 3: Element estimation sorting steps.

[0105] (1) Spectral Data Preprocessing: In hyperspectral data processing, spectral differentiation is a commonly used transformation method that extracts spectral parameters by calculating the differential value of the original spectral reflectance. Low-order differential processing of the spectrum is less sensitive to noise, thus the data after differential transformation is more effective in practical applications. First-order differential processing (D1) (Equation 10) can eliminate the influence of linear or near-linear noise spectra, background, etc., on the target spectrum; second-order differential processing (D2) (Equation 11) can effectively eliminate baseline drift and background signals, improve analysis accuracy, and enhance the correlation between reflectance in some bands and surface parameters. The formula for calculating spectral differential transformation is as follows:

[0106]

[0107] In the formula, R′(λ) i R″(λ) is the first derivative of the spectral reflectance. i R(λ) is the second derivative of the spectral reflectance. i+1 ) represents the wavelength λ i+1 Reflectivity at point R(λ) i-1) represents the wavelength λ i-1 Reflectance at point R′(λ) i+1 ) represents the wavelength λ i+1 The first derivative of the reflectivity at point R′(λ) i-1 ) represents the wavelength λ i-1 The first derivative of the reflectivity at a given point, where Δλ is the wavelength interval;

[0108] Standard Normal Variance (SNV) is a commonly used data preprocessing method in spectral analysis. Its purpose is to reduce the influence of factors such as variations in light source intensity, particle size differences, or sample surface roughness, allowing for clearer observation and analysis of the chemical information in the spectral data. It involves centering the mean of the spectral data at each wavelength by subtracting the average value for that wavelength. Variance scaling is then applied by dividing the spectral data at each wavelength by its standard deviation. Continuum Removal (CR), also known as the continuum removal method (Equation 12), is a common spectral processing method that maximizes the highlighting of absorption peak characteristics, facilitating the differentiation of differences between spectral curves. The formula for calculating the envelope removal transformation is as follows:

[0109]

[0110] In the formula, R c ′(λ i ) represents the wavelength λ i The envelope value at point R(λ) i ) represents the wavelength λ i The reflectivity value, R c (λ i ) represents the corresponding wavelength λ i The value at the envelope;

[0111] (2) Correlation calculation: The correlation between the transformed data and the element content is calculated by applying the Pearson correlation coefficient to select the optimal transformation method, that is, the transformation method with the largest Pearson correlation coefficient and the highest correlation.

[0112] (3) Ranking of major and heavy metal elements: Select major and heavy metal elements (such as Zr and Fe) that have significant responses in the visible to short-wave infrared bands, calculate the Pearson correlation coefficients between these elements and the spectrum, and rank them from high to low according to the correlation coefficient values. Calculate the spectral correlation of these elements and rank them accordingly.

[0113] (4) Calculation of correlation coefficients of trace elements: Further analyze the correlation between trace elements (such as Pb, Ag, etc.) and the above-mentioned major elements, calculate the Pearson correlation coefficients between trace elements and the above-mentioned major elements, and sort them from high to low according to the values.

[0114] Experiment Example 4: Steps for fusing remote sensing geochemical data.

[0115] (1) Image registration: crop single-element geochemical layers and hyperspectral remote sensing images (HSI) of the same size, and complete spatial correction and registration.

[0116] (2) Image filtering: The hyperspectral remote sensing image is decomposed using the Gaussian filtering method to obtain the high-frequency component (HSI_H) and low-frequency component (HSI_L) of the multispectral band, which represent the image detail information and background information, respectively.

[0117] (3) Scale transformation: For the geochemical layer of each element and the low-frequency component HSI_L of the multispectral remote sensing image, the nearest neighbor interpolation method is used to resample it to the same resolution as the high-frequency component HSIH, so that the two data can be analyzed at a unified scale.

[0118] (4) Construction of deep learning model estimation: Based on the element content sorting, deep learning models (2D CNN-LSTM) for low frequency component (HSI_L) and geochemical element content (Geo_L) are built in sequence to calculate the low frequency information fusion layer.

[0119] Geo_L i =F ij (HSI_L j )

[0120] In the formula, i is the index of the hyperspectral band, j is the index of the element, and F ij Nonlinear models built for deep learning, (HSI_L) j ) represents the hyperspectral low-frequency data of the j-th element.

[0121] (5) Image reconstruction: Based on the constructed deep learning model, geochemical information and high-frequency components (HSI_H) are reconstructed to obtain a high-frequency information fusion layer (Geo_H).

[0122] Geo_H i =F ij (HSI_H j )

[0123] In the formula, (HSI_H j () represents the hyperspectral high-frequency data of the j-th element;

[0124] (6) Image Fusion: The high-frequency information fusion layer representing spatial details is injected into the low-frequency information fusion layer to obtain the fused remote sensing geochemical data. The flowchart for remote sensing geochemical data fusion is as follows: Figure 7As shown. The root mean square error (RMSE) and the coefficient of determination (R-squared) are used. 2 The model's performance on the training and validation sets was evaluated using RMSE and Mean Absolute Error (MAE). The model exhibited good stability, with similar evaluation results for all elements on both the training and test sets, indicating no overfitting. In the experiment estimating elemental content scatter plots based on a 2D CNN-LSTM model for low-frequency hyperspectral data, Ta showed the best prediction performance, with an RMSE of 0.070 and R² on the training set. 2 =0.983, MAE=0.050, RMSE=0.072, R 2 =0.982, MAE=0.052. Except for the Ca element (training set RMSE=0.110, R... 2 =0.795, MAE=0.076, validation set RMSE=0.111, R²=0.794, MAE=0.078) and K elements (training set RMSE=0.079, R²=0.795, MAE=0.076, validation set RMSE=0.111, R²=0.794, MAE=0.078) and K elements (training set RMSE=0.079 2 =0.745, MAE=0.063, RMSE=0.081, R 2 =0.724, MAE=0.064) results are slightly worse. Other elements R 2 All values ​​were above 0.9. Among them, the prediction results for Li, the main ore-forming element in the area, were better, with RMSE = 0.086 and R0.9 on the training set. 2 =0.983, MAE=0.063, RMSE=0.089, R 2 =0.982, MAE=0.064.

[0125] The fused data retains both the spatial distribution trends of the refined elemental data and the texture details of the hyperspectral image data, with a spatial resolution of 30m. The following shows the inverse distance weighted interpolated image before fusion of 19 elements. Figure 8 (IDW interpolation plot of elemental content) and fusion results ( Figure 9 (For 2D CNN-LSTM models, the element content fusion map is estimated. When the overall distribution trend of element content is consistent, the details of the fused data are clearer.)

[0126] Experiment Example 5: Data Fusion Super-Resolution Step. Previous studies have applied deep prior super-resolution techniques, which do not rely on data learning but capture statistical information of images solely based on the network structure before learning. This invention applies this technique to further supplement the spatial detail information of remote sensing geochemical data. It includes the following steps:

[0127] (1) Network structure: A 3D convolution-based encoder-decoder "hourglass" network structure is used to process spectral and spatial information simultaneously. The encoder-decoder structure is used to extract and reduce image features and reconstruct the high-resolution morphology of the original image from the encoded features, respectively, restoring details and size layer by layer.

[0128] (2) Initialize Input: Input an initialized random noise image. This noise image can be generated by randomly sampling from a matrix of a given size, with the value of each pixel uniformly distributed between 0 and 1. Specifically, a three-dimensional noise matrix of the same size as the target image can be generated using the numpy.random.rand(height, width, depth) function from Python's NumPy library.

[0129] (3) Compilation Model: A mean squared error (MSE) loss function is defined to measure the difference between the network output and the high-resolution image. The high-resolution image is generated by fusing remote sensing geochemical data, while the low-resolution image is generated by downsampling the high-resolution image, simulating the real-world low-resolution image acquisition process. The Adam optimizer is used to calculate the loss function to minimize the loss value.

[0130] (4) Training the model: The high-resolution image is downsampled to generate a low-resolution image, and the low-resolution image is used as the input of the training data. The learning rate is set to 0.01, the number of iterations is 1200, and the noise level is 0.01.

[0131] (5) Result Evaluation: The peak signal-to-noise ratio (PSNR) was used to evaluate the model's performance in reconstructing high-resolution and low-resolution images, comparing it with traditional interpolation-based upsampling methods (Bicubic, Nearest) and the images reconstructed by the depth image prior algorithm. Figure 10 The peak signal-to-noise ratio (PSNR) of the super-resolution result false-color composite image (R:Be, G:Rb, B:Li) is higher (Table 1).

[0132] Table 1509 Land Cover Categories and Sample Count

[0133] PSNR 40.2331 47.2966 49.3666

[0134] Example 6: Steps for constructing a composite domain-aware deep metric learning network model

[0135] (1) Experimental Parameter Settings: The feature extraction module consists of 5 CoConv blocks. In each block, CoConv applies a 3×3 convolution to the composite neighborhood of each target and outputs a 64-dimensional feature for each neighborhood. The neighborhood linear transformation also maps each neighborhood to a 64-dimensional feature. The metric space embedding module contains 64 standard 3×3 convolutional kernels, followed by a nonparametric mini-batch metric classifier. The composite neighborhood of each target contains 9×9 pixel-level neighborhoods and 25 superpixel-level neighborhoods. During the training phase, the composite domain-aware deep metric learning model is trained using the Adam optimizer with a learning rate of 5×10⁻⁶. -4 The batch size is 64, and the epoch is 200. The code is based on Python 3.8.13 and PyTorch 1.9.1, and runs on an RTX A6000 graphics processor (GPU).

[0136] (2) Experimental Data: The dataset consists of fused remote sensing geochemical data from the 509 mining area, containing 19 bands and a spatial dimension of 512×217 pixels. There are 10 land cover classes. Approximately 30%, 0.5%, and 69.5% of the samples from each class were randomly selected as training, validation, and test datasets, respectively, as shown in Table 2.

[0137] Table 2. 509 Land Feature Categories and Sample Count

[0138]

[0139]

[0140] Experiment Example 7: Validation of the Results of the Composite Domain-Aware Deep Metric Learning Method. To validate the model's accuracy, the accuracy and loss curves of the composite domain-aware deep metric learning method model were calculated. The results are shown below. Figure 11 Loss curves in the model training and validation sets; Figure 12 The accuracy curves on the model training and validation sets show that ( Figures 11-12 The changes in loss and accuracy (acc) on the training and validation sets are shown. After approximately 100 iterations, the loss and accuracy tend to stabilize, with the loss on the training set around 0.26 and the loss on the validation set around 1.02. The accuracy on the training set is around 0.992, and the accuracy on the validation set is around 0.972.

[0141] Based on the above steps, the identification results of the ore-bearing pegmatite and the main lithology are obtained, see [link to relevant documentation]. Figure 13 ,in Figure 13 Figure (a) is a geological map, and Figure (b) is a lithology classification map obtained by the composite domain perception depth metric learning method. The composite domain perception depth metric learning method has a good recognition of the distribution morphology of exposed mineralized pegmatite veins.

[0142] Confusion matrix ( Figure 14 The depth metric classification matrix visually reflects the number of accurately identified lithological units and easily confused categories. The values ​​on the diagonal represent the number of correctly identified lithological units, serving as a direct indicator of model performance. Higher diagonal values ​​indicate a higher proportion of lithological units accurately identified by the model. Off-diagonal elements represent discrepancies between actual and predicted lithological units, i.e., the number of actual lithological units incorrectly classified as different lithological units by the model. In the prediction confusion matrix, the depth metric classification model has more correctly identified pixels. In the depth metric classification model, only the Upper Formation of the Triassic Bayan Har Mountains Group is misclassified as ore-bearing pegmatite. The number of samples misclassified as non-ore-bearing pegmatite, from highest to lowest, is: the Middle Formation, Upper Formation, and biotite monzogranite of the Bayan Har Mountains Group.

[0143] To more intuitively evaluate the model's classification performance, four metrics were calculated: Overall Accuracy (OA), Average Accuracy (AA), and Kappa to assess the model's overall classification performance. Additionally, User's Accuracy (UA) and Producer's Accuracy (PA) were used to evaluate the classification performance for each specific rock type (see Table 3). From Table 3, the composite domain-aware depth metric learning method showed good classification performance for all three metrics: AA (80.48%), OA (98.10%), and Kappa coefficient (97.32%). The model demonstrated good classification performance for six lithologies: Holocene alluvial-diluvial deposits, Pleistocene glacial deposits, Triassic Bayan Har Mountains Group, biotite granodiorite, biotite monzogranite, and fine-grained tonalite diorite, with UA and PA accuracies both exceeding 97.17%. For pegmatites with poor exposure and low labeling rates, the user accuracy for non-mineralized pegmatites reached 83.08%, and the producer accuracy was 72.71%. For mineralized pegmatites, the user accuracy was 64.55%, and the producer accuracy was 29.76%.

[0144] Table 3 Performance statistics of the composite neighborhood sensing network method

[0145]

[0146] 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 mineral-bearing pegmatites based on remote sensing geochemical fusion data, characterized in that, The method includes: S1. Collect geochemical data of stream sediments in the study area and send them for testing to obtain elemental content. Organize the latitude and longitude information of the sampling location, convert the two-dimensional table data of elemental content into vector data, and apply the inverse distance weighted interpolation method to convert the vector data into raster data. S2, based on the standardized correction DS framework, applies a non-negative least squares model to correct remote sensing images; S3. Based on the Pearson correlation coefficient, calculate the coefficient between the spectral transformation and the processed results and the elemental content, select the transformation method with the highest correlation coefficient, and determine the prediction order of major and trace elements. S4. The remote sensing image is divided into low-frequency and high-frequency data by Gaussian filtering. A 2D CNN-LSTM network structure is applied to establish the relationship between low-frequency data and collected geochemical data of water system sediments. The trained CNN-LSTM model is applied to the high-frequency data to predict the content of anomalous elements. The low-frequency data is superimposed on the high-frequency data to complete the fusion of remote sensing geochemical data. S5 downsamples remote sensing geochemical fusion data into low-resolution image data, optimizes the structure parameters of the deep prior network (DIP), learns image details, and obtains fusion data with more spatially resolved detail information. S6 uses a combination of convolutional neural networks and graph neural networks to construct a composite neighborhood perception network. It employs deep metric learning to mine the characteristics of different lithological samples, reduce intra-class distance, increase inter-class distance, and complete the identification of mineral-bearing pegmatites and their related lithologies. S7 uses user precision (UA), producer precision (PA), and overall precision (OA, AA, and kappa coefficients) to evaluate the model recognition performance of pegmatites and other lithologies.

2. The method for identifying mineral-bearing pegmatites based on remote sensing geochemical fusion data according to claim 1, characterized in that, In step S1, the element content is obtained, including: constructing vector data with coordinate information and geochemical content from the geochemical content data and corresponding coordinate data obtained from the test; converting the vector data into raster data using the inverse distance weighted interpolation method, wherein the raster size in the raster data takes into account the sampling interval of the geochemical data and adopts a size between the minimum sampling point spacing and the maximum sampling point spacing; The specific steps of the inverse distance weighted interpolation method are as follows: ① Determine the coordinates of the unknown and known points; ② Calculate the distance between the unknown point and each known point; ③ Calculate the weight of each known point based on the distance. The formula for calculating the weight is: In the formula, w is the weight and d is the distance; ④ Calculate the value of the unknown point based on the weights and the values ​​of the known points. The calculation formula is as follows: In the formula, z is the value of the unknown point, and w i Given the weight of point i, z i Let i be the value of a known point.

3. The method for identifying mineral-bearing pegmatites based on remote sensing geochemical fusion data according to claim 1, characterized in that, In step S2, a non-negative least squares model is applied to correct the remote sensing image, including: Extract hyperspectral image data corresponding to the sampling points on the opposite side and form a two-dimensional list; resample the ground indoor spectral data according to the image spectral data, use the resampled ground indoor spectral data as the target variable, input the non-negative least squares NNLS model to correct the hyperspectral image data, and obtain the corrected hyperspectral image data. Nonnegative least squares is used to calculate nonnegative spectral reflectance values. In a given linear system, it finds a nonnegative vector x such that ||Ax-b|| 2 The minimum value is given by formula (1) as follows: min x ||Ax-b|| 2 x≥0(1) Where x≥0 constrains each element to be nonnegative, min x Let x be the minimum value of x, A be the coefficient matrix, x be the image spectral variable, and b be a set of constants.

4. The method for identifying mineral-bearing pegmatites based on remote sensing geochemical fusion data according to claim 1, characterized in that, In step S3, determining the prediction order of major and minor elements includes: The GF-5 remote sensing image spectrum after correction was processed using first-order derivatives, second-order derivatives, standard normal variables, and envelope removal. Based on the Pearson correlation coefficient, the coefficients between the transformed and processed results and the elemental content were calculated. The transformation method with the highest correlation coefficient was selected. The prediction order of major elements was determined based on the average correlation coefficient between elements with significant responses in the visible-shortwave infrared band and the spectrum. Then, the prediction order of trace elements was determined based on the average correlation coefficient between trace elements and major elements. The formula for calculating the Pearson correlation coefficient is as follows: In the formula, Cov(X,Y) is the covariance of X and Y, Var[X] is the variance of X, and Var[Y] is the variance of Y.

5. The method for identifying mineral-bearing pegmatites based on remote sensing geochemical fusion data according to claim 4, characterized in that, Also includes: (1) Spectral data preprocessing: The influence of linear or near-linear noise spectra and background on the target spectrum is eliminated using the first-order differential processing formula (10); baseline drift and background signals are eliminated using the second-order differential processing formula (11). The formula for calculating the spectral differential transformation is as follows: In the formula, R'(λ) i R(λ) is the first derivative of the spectral reflectance. i R(λ) is the second derivative of the spectral reflectance. i+1 ) represents the wavelength λ i+1 Reflectivity at point R(λ) i-1 ) represents the wavelength λ i-1 Reflectivity at point R'(λ) i+1 ) represents the wavelength λ i+1 The first derivative of the reflectivity at point R'(λ) i-1 ) represents the wavelength λ i-1 The first derivative of the reflectivity at a given point, where Δλ is the wavelength interval; The envelope removal method, CR formula (12), is used to maximize the highlighting of the absorption peak characteristics of the spectrum and distinguish the differences between spectral curves. The envelope removal transformation calculation formula is as follows: In the formula, R' c (λ i ) represents the wavelength λ i The envelope value at point R(λ) i ) represents the wavelength λ i The reflectivity value, R c (λ i ) represents the corresponding wavelength λ i The value at the envelope; (2) Correlation calculation: Analyze the correlation between the transformed data and the element content, calculate the Pearson correlation coefficient, and take the highest Pearson correlation coefficient as the optimal transformation method; (3) Ranking of major and heavy metal elements: Major and heavy metal elements that have significant responses in the visible to short-wave infrared bands were selected, their spectral correlations were calculated, and they were ranked. (4) Calculation of the correlation coefficient of trace elements: Further analysis was conducted on the correlation between trace elements and the aforementioned major elements.

6. The method for identifying mineral-bearing pegmatites based on remote sensing geochemical fusion data according to claim 1, characterized in that, In step S4, the remote sensing geochemical data fusion is completed, including: Based on the optimal transformation method for processing remote sensing images, Gaussian filtering is used to divide the remote sensing images into low-frequency and high-frequency data. A 2D CNN-LSTM network structure is applied to establish the relationship between low-frequency data and collected stream sediment geochemical data. The element prediction order is determined according to the Pearson correlation coefficient between spectra and elements from high to low, and the element content is predicted sequentially. The estimation results of major elements and heavy metals with high spectral response are added to the trace element estimation model. The trained model is also applied to high-frequency data to predict anomalous element content. The background geochemical content data retrieved from low frequency is superimposed on the anomalous geochemical content data retrieved from high frequency to complete the fusion of remote sensing geochemical data.

7. The method for identifying mineral-bearing pegmatites based on remote sensing geochemical fusion data according to claim 1, characterized in that, In step S5, the structural parameters of the deep prior network (DIP) are optimized to learn image details and obtain fused data with more spatially resolved detail information, including: This paper utilizes a generative network architecture to leverage image statistical priors, enabling the network to automatically reduce noise and prioritize the reconstruction of image details during the iterative process. After a certain number of iterations, optimization stops without overfitting the noise, restoring the original image resolution. The expression is as follows: In the formula, The parameters obtained through optimization are θ, which represents the network parameters, and f. θ For the constructed deep network, x is a low-resolution image and z is the input noise.

8. The method for identifying mineral-bearing pegmatites based on remote sensing geochemical fusion data according to claim 1, characterized in that, In step S6, the identification of ore-bearing pegmatites and their related lithologies is completed, including: Vectorized lithological data were rasterized to ensure consistency between geochemical and lithological raster data ranges. A composite neighborhood-aware network recognition model was constructed using a convolutional neural network combined with a graph neural network, and parameters were set accordingly. Feature extraction was performed based on the composite neighborhood-aware CoConv model, and mean calculation was performed based on standard image convolution to transform the features into a metric space. A nonparametric mini-batch metric classifier was used to classify pegmatites and other lithologies. During model training, the standard Adam optimizer was used to train the composite neighborhood-aware deep metric learning model.

9. The method for identifying mineral-bearing pegmatites based on remote sensing geochemical fusion data according to claim 1, characterized in that, In step S7, the model identification effect of pegmatite and other lithologies is evaluated, including: Using the geological map as a reference for the actual results, and the identification results of pegmatite and other lithologies as the classification result map, a confusion matrix of the classification results is established. The indicators are then analyzed through the confusion matrix [g] ij ] k×k The calculation yields the following formula: In the formula, G is the confusion matrix, where each row represents the actual class and each column represents the predicted class; g ij The number of samples whose actual class is i and whose predicted class is j; In the formula, OA represents the overall accuracy, and g aa The diagonal elements are k, where k is the number of categories, and g is the number of categories. ab The number of samples that are actually class a but are predicted as class b is determined by the diagonal elements g of the confusion matrix G. aa The sum of these values ​​divided by the total number of samples is used to calculate the overall accuracy of the model. UA (n) The accuracy rate for the nth class is calculated by dividing the number of correctly classified samples in the nth class by the number of predicted samples in the nth class; this represents the proportion of correct identifications for class n. nn Let g be the value on the diagonal of the nth type confusion matrix. na This represents the element in the nth row of the confusion matrix; In the formula, AA is the average precision, which is used to measure the average performance of the classification model across all categories. The average precision of all categories is obtained by summing the precision of each category and then dividing by the number of categories K. In the formula, PA (n) The user accuracy for class n is calculated by dividing the number of correctly classified samples in class n by the actual number of samples in class n; it represents the proportion of class n that is correctly identified. an The element in the nth column of the confusion matrix; In the formula, Kappa is the evaluation index used to test the consistency of classification results, M is the total number of samples, and g is the weight of the samples. a+ Let g represent the sum of the a-th row of the matrix. +a This is the sum of column a.

10. A system for identifying mineral-bearing pegmatites based on remote sensing geochemical fusion data, characterized in that, The system implements the method for identifying mineral-bearing pegmatites based on remote sensing geochemical fusion data as described in any one of claims 1-9, and the system includes: The geochemical data fusion module is used to fuse remote sensing, geochemical and geological map data based on a 2DCNN-LSTM model. The fusion data resolution module improves the resolution of fused data based on the deep prior super-resolution model (DIP). The module for identifying mineralized pegmatites and other lithologies is used to construct a mineralized pegmatite identification model based on the composite domain perception depth metric learning (DNM-DML) model to complete the identification of mineralized pegmatites and other lithologies.

Images