Mineral prediction method and system based on multiscale sample heterogeneity

Abstract

The application discloses a mineral prediction method and system based on multiscale sample unevenness, and the method comprises the following steps: converting collected geological data into image data to obtain geochemical maps of different scales; dividing a grid according to a certain pixel ratio for the geochemical maps, and forming training samples according to expert knowledge; performing sample data enhancement; establishing a joint model of an autoencoder and a generative adversarial network, and using small-scale geochemical maps to find geochemical exploration anomaly areas; and constructing a convolutional neural network model with an attention mechanism to classify mineral products in the anomaly areas of large-scale data. The joint model can learn the internal relations and characteristics among multivariate geochemical data, extract geochemical exploration anomalies on small-scale data, and effectively avoid the influence of noise in the geochemical data. The attention mechanism is used on large-scale data to make the model pay more attention to key information in the image, and the image classification accuracy can be effectively improved.

Description

Technical Field

[0001] This invention relates to a mineral prediction method and system based on multi-scale sample heterogeneity, belonging to the fields of geoinformation science and computer vision. Background Technology

[0002] With the rapid development of my country's economy, mineral resources have increasingly become an important material foundation for economic and social development. How to better develop and utilize mineral resources has become a crucial topic in the field of geology. As artificial intelligence technology matures, compared to the previous reliance solely on expert knowledge and related geological software for mineralization prediction, computer technology and artificial intelligence are gradually permeating the field of earth sciences. Through the use of machine learning and deep learning techniques, mineralization prediction has evolved from traditional methods to a combination of traditional and artificial intelligence approaches, resulting in more accurate and reliable results.

[0003] In recent years, artificial intelligence technology has gradually achieved certain results in mineral resource prediction. However, due to the scarcity of samples and the fact that the mineralized area is relatively small compared to the background area, the existing processing methods have not utilized sample data of different scales and cannot more accurately predict the mineralized area. Summary of the Invention

[0004] Purpose of the invention: In view of the shortcomings of the existing technology, the purpose of this invention is to propose a mineral prediction method and system based on multi-scale sample heterogeneity, so as to find the relationship between geochemical data and mineralization more comprehensively and objectively, find mineralization patterns in geochemical data, and predict mineralization probability in new geochemical data.

[0005] Technical Solution: To achieve the above-mentioned objectives, the present invention provides a mineral resource prediction method based on multi-scale sample heterogeneity, comprising the following steps:

[0006] Step 1: Convert the collected geological data into image data to obtain geochemical maps at different scales, namely, geochemical maps at 1:P1 and 1:P2, where 1:P1 and 1:P2 represent the ratio of distance on the map to the actual distance, and P1 is greater than P2;

[0007] Step 2: Divide the geochemical map into grids according to a certain pixel ratio and form training samples based on expert knowledge; divide each prediction map into grids, and if a grid contains a mineral deposit, select it as a metallogenic area; otherwise, select it as a background area.

[0008] Step 3: Perform data augmentation on the obtained training samples;

[0009] Step 4: Introduce the Generative Adversarial Network (WGAN) into the deep convolutional autoencoder, where the generator in the WGAN serves as the decoder in the autoencoder. The encoder maps grid data of sample size 1:P1 from the image space to the latent space. Then, train the WGAN to obtain the output and calculate the Euclidean distance between the input and output data as the geochemical anomaly score.

[0010] Step 5: Based on the geochemical anomaly score map obtained from the 1:P1 geochemical map, sample labeling is performed in the 1:P2 grid data. Geochemical anomaly scores higher than a set threshold are labeled as positive samples, otherwise they are labeled as negative samples. Based on this, specific geochemical data classification is performed; a convolutional neural network model is constructed, and an attention mechanism is added to the model so that the model focuses on the information of the region of interest, thereby learning the pattern between mineralization and geochemical data.

[0011] Step 6: Input the 1:P2 training sample data into the convolutional neural network model for iteration and update the parameters of the neural network;

[0012] Step 7: Use the trained convolutional neural network model to predict the mineralization probability of the geochemical data of the area to be predicted, and generate a probability prediction distribution map of mineral resources in the area.

[0013] Furthermore, the geological data collected in step 1 includes geographic location information and geochemical element information. The geographic location information includes the latitude and longitude coordinates of the known mining area, and the geochemical element information includes the content of various geochemical elements in the known mining area. The geographic location information and geochemical element information are input into ArcGIS software, and the geochemical factors are converted into geochemical spatial data through inverse distance weighted interpolation to obtain geochemical maps at different scales.

[0014] Furthermore, in step 3, data augmentation is performed by adding noise and using generative adversarial networks to generate new samples.

[0015] Furthermore, in step 4

[0016] Assumption 1: The background region data of P1 follows a probability distribution in a certain hidden space. The WAGN network is used to learn the latent representation of the background region data, thereby generating the generator model G and discriminator model D in WGAN. Specifically, the steps are as follows:

[0017] Step 41: Construct an encoder E using a convolutional neural network. Input the data samples of the 1:P1 background region after data augmentation into the encoder, and perform feature extraction through the encoder to map them to the latent space z.

[0018] Step 42: A generative adversarial network (GAN) is used after the encoder network, where the generator G in the GAN acts as the decoder in the autoencoder to complete the mapping from the latent space to the image space; in order to obtain the feature distribution of the latent space where the background region is located, the loss function of the input image x and the reconstructed image G(E(x)) is minimized:

[0019]

[0020] Where n is the number of pixels in the image;

[0021] Step 43: During encoder training, statistical information from the discriminator, calculated for both real and generated images, is used to guide the training of encoder E. The loss function used by discriminator D to guide encoder training is as follows:

[0022]

[0023] Where, n d is the dimension of the intermediate feature representation, k is the weighting factor, and f(·) is the feature generated by the discriminator D during the operation, with the purpose of statistically analyzing the information of the given input;

[0024] Step 44: Train encoder E using background region data with the aim of minimizing loss function L2 to generate an image that is more similar to the feature distribution of the original image;

[0025] Step 45: After training, calculate the Euclidean distance between the original input image xx and the output image y generated by the trained generative adversarial network as the geochemical anomaly score. Set an anomaly threshold; scores above the threshold are considered anomalies, and scores below the threshold are considered non-anomalies. The Euclidean distance formula is:

[0026]

[0027] Furthermore, step 5 includes the following steps:

[0028] Step 51: Based on the anomaly score, obtain the cluster of anomalies in the current 1:P1 geochemical map, and then obtain the areas with a high probability of mineralization; based on the existing data, obtain the geochemical data of the corresponding 1:P2 location.

[0029] Step 52: Construct a convolutional neural network model. The input data of the model is a c×h×w matrix, where c represents the number of image channels of geochemical data. Each channel contains information on one geochemical element, for a total of c geochemical information. h and w represent the height and width of the image. The data output format of the convolutional neural network is a probability vector representing each mineral type or background area.

[0030] Step 53: In constructing the convolutional model, the channel attention module is integrated to apply a weight to the feature map on each channel. That is, attention is focused on the mineral area to be identified by the following formula, and feature extraction processing is performed on these areas to capture the correlation in the channel data of the geochemical exploration data.

[0031] g(x) = Sigmoid(W2ReLU(W1x))

[0032] Sigmoid and ReLU are activation functions, and W1 and W2 are weight parameters of the fully connected layer. After calculating the channel attention, an attention weight matrix g(x) can be obtained. Multiplying the obtained weight matrix g(x) with the original feature map x yields the feature information required for the next layer of the network, i.e., the input of the next layer.

[0033] Furthermore, the loss function for training the network model in step 6 is:

[0034] Where L CE (p i ,y i ) represents the cross-entropy of multi-class classification, y i It's a real tag, p i It predicts the label classification probability, where N is the total number of samples; GD(g i ) is the gradient density;

[0035]

[0036]

[0037]

[0038] g i Gm represents the gradient. i ε represents the gradient magnitude, and ε is a hyperparameter.

[0039] This invention provides a mineral prediction system based on multi-scale sample heterogeneity, comprising:

[0040] The preprocessing module converts the collected geological data into image data, resulting in 1:P1 and 1:P2 geochemical maps, where 1:P1 and 1:P2 represent the ratio of distance on the map to the actual distance, with P1 being greater than P2; it also divides the geochemical maps into grids according to a certain pixel ratio and forms training samples based on expert knowledge; and it divides each prediction map into grids, selecting the metallogenic area if the grid contains a mineral deposit, otherwise selecting it as a background area.

[0041] An anomaly detection module is used to introduce a generative adversarial network into a deep convolutional autoencoder. The generator in the generative adversarial network acts as the decoder in the autoencoder. The encoder maps grid data of sample size 1:P1 from the image space to the latent space. Then, the generative adversarial network is trained to obtain the output. The Euclidean distance between the input data and the output data is calculated as the geochemical anomaly score.

[0042] The data augmentation module is used to augment the data of the mining area and the obtained geochemical anomaly areas;

[0043] The classification model module is used to label samples in the 1:P2 grid data based on the geochemical anomaly score map obtained in the 1:P1 geochemical map. Geochemical anomaly scores above a set threshold are labeled as positive samples, and otherwise as negative samples. Based on this, specific geochemical data classification is performed. The module also constructs a convolutional neural network model and adds an attention mechanism to the model so that it focuses on the information of the region of interest, thereby learning the pattern between mineralization and geochemical data.

[0044] The model training module is used to input 1:P2 training sample data into the convolutional neural network model for iteration and update the parameters of the neural network.

[0045] Additionally, the classification and prediction module uses the trained convolutional neural network model to predict the mineralization probability of the geochemical data for the area to be predicted, generating a probability prediction distribution map of mineral resources in the area.

[0046] The present invention also provides a computer system, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when loaded onto the processor, implements the aforementioned mineral prediction method based on multi-scale sample heterogeneity.

[0047] Beneficial Effects: This invention provides a mineral resource prediction method and system based on multi-scale sample heterogeneity. To more accurately and objectively determine the relationship between geochemical data and mineralization probability, this invention first analyzes grid data at a smaller scale and extracts geochemical anomalies. Compared to traditional anomaly extraction methods, the proposed method more effectively reduces the impact of noise and identifies anomalies in complex geological environments. Furthermore, after identifying anomaly regions in the small-scale buffer map, specific mineralization data are classified for these regions using large-scale data. This invention utilizes data at these two different scales for mineralization prediction and employs a gradient-based loss function to address the problem of uneven mineral resource data samples, resulting in more accurate and reliable final results. Attached Figure Description

[0048] Figure 1 This is an overall flowchart of an embodiment of the present invention;

[0049] Figure 2 This is a schematic diagram of the geochemical data anomaly analysis process in an embodiment of the present invention;

[0050] Figure 3 This is a diagram of the channel attention network structure CA_Block in an embodiment of the present invention;

[0051] Figure 4 This is a diagram of the ResNet34CA_Block network structure with channel attention mechanism added in an embodiment of the present invention. Detailed Implementation

[0052] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading the present invention, any modifications of the present invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0053] like Figure 1 As shown in the figure, the mineral prediction method based on multi-scale sample heterogeneity disclosed in this invention mainly includes the following steps:

[0054] Step 1: Convert the collected geological data into image data to obtain two geochemical maps at different scales.

[0055] This process involves collecting relevant geological and geochemical data, including geographic location information and geochemical element information. Geographic location information includes the latitude and longitude coordinates of known mining areas, while geochemical element information includes the content of various geochemical elements in known mining areas. The geographic location information and geochemical element information are input into ArcGIS software, and the geochemical factors are converted into geochemical spatial data using inverse distance weighted interpolation, ultimately obtaining geochemical maps at different scales. Two different scale geochemical maps are used: a 1:P1 geochemical map and a 1:P2 geochemical map, where 1:P1 and 1:P2 represent the ratio of the distance on the map to the actual distance, with P1 being greater than P2. This embodiment uses a 1:200,000 geochemical map and a 1:50,000 geochemical map.

[0056] Step 2: Divide the geochemical map into grids according to a certain pixel ratio and form training samples based on expert knowledge; divide each prediction map into grids, and if a grid contains a mineral deposit, select it as a metallogenic area; otherwise, select it as a background area. In this embodiment, the grids are divided into grids of equal size, and the already divided grid data are further divided into metallogenic areas and background areas based on expert knowledge.

[0057] Step 3: Perform data augmentation on the obtained training samples. In this embodiment, data augmentation is performed by adding noise and using a generative adversarial network to generate new samples. This mainly includes adding noise to the original data, setting the random noise to 0.03, 0.04, and 0.05, and adding noise 20 times per sample; then, the image reconstruction function of a deep convolutional autoencoder is used to increase the number of samples.

[0058] Step 4: Anomaly identification on 1:200,000 grid data. A generative adversarial network (GAN) is introduced into a deep convolutional autoencoder (DEC). The DEC maps the 1:200,000 grid data from the image space to the latent space. The GAN is then trained to obtain the output, and the Euclidean distance between the input and output data is calculated as the geochemical anomaly score.

[0059] In this embodiment, it is assumed that the 1:200,000 background region data satisfies a certain probability distribution in the hidden space. The WAGN network is used to learn the latent representation of the background region data, thereby generating the generator model G and discriminator model D in the WGAN. For example... Figure 2 As shown, the specific steps include the following:

[0060] Step 41: Construct an encoder E using a convolutional neural network. Input the data-enhanced sample into the encoder and extract features through the encoder to obtain the feature distribution of the background area data, i.e., the latent space z.

[0061] Step 42: A generative adversarial network (GAN) is used after the encoder network, where the generator G in the GAN can be regarded as the decoder in the autoencoder. The previous step mapped the samples from the real image to the latent space z, and the decoder completes the mapping from the latent space to the image space. In order to obtain the feature distribution of the latent space where the background region is located, the loss function of the input image x and the reconstructed image G(E(x)) needs to be minimized.

[0062]

[0063] Where n is the number of pixels in the image;

[0064] Step 43: Simply minimizing the pixel size difference may result in generated images that do not perfectly conform to the distribution of the original images, because the true target locations in the hidden space of the given original image data are unknown; that is, the locations that ultimately affect the probability of mineralization in each sample are uncertain. This will cause the discriminator D to fail to assign high residual values ​​to anomalous images. Therefore, when training the encoder, the statistical information of the real and generated images calculated in the discriminator is used to guide the training of the encoder E. The loss function of the discriminator D guiding the encoder training is as follows:

[0065]

[0066] Where, n d is the dimension of the intermediate feature representation, k is the weighting factor, and f(·) is the feature generated by the discriminator D during the operation, with the purpose of statistically analyzing the information of the given input;

[0067] Step 44: Train encoder E using background region data with the aim of minimizing loss function L2 to generate an image that is more similar to the feature distribution of the original image.

[0068] Step 45: After training, calculate the Euclidean distance between the original input image x and the output image y generated by the trained generative adversarial network as the anomaly score. Set an anomaly threshold; scores above the threshold are considered anomalies, and scores below the threshold are considered non-anomalies. The Euclidean distance formula is:

[0069]

[0070] Step 5: For anomalous areas at a scale of 1:200,000, mineralization prediction is performed using data at a scale of 1:50,000. Specifically, based on the geochemical anomaly score map obtained from the 1:200,000 geochemical map, samples are labeled in the 1:50,000 grid data. Geochemical anomaly scores higher than a set threshold are labeled as positive samples (higher probability of mineralization), otherwise as negative samples (lower probability of mineralization). Based on this, specific geochemical data classification is performed. A convolutional neural network model is constructed, and an attention mechanism is added to the model. This mechanism allows the model to focus on information in areas of interest, thereby learning the patterns between mineralization and geochemical data. The specific steps include the following:

[0071] Step 51: Based on the anomaly score, the areas where anomalies are clustered on the current 1:200000 geochemical map can be identified, thus revealing areas with a high probability of mineralization; based on the existing data, the corresponding 1:50000 geochemical data can be obtained.

[0072] Step 52: Construct a convolutional neural network model. The input data of the model is a c×h×w matrix, where c represents the number of image channels of geochemical data. Each channel contains information on one geochemical element, for a total of c geochemical information. h and w represent the height and width of the image. The data output format of the convolutional neural network is a probability vector representing each mineral type or background area.

[0073] Step 53: In constructing the convolutional model, the channel attention module is integrated. Since each channel in geochemical data represents a type of element, different channels contain different information, which significantly affects their final mineralization probability. The channel attention network structure is as follows: Figure 3As shown, a weight is applied to the feature map of each channel to represent the similarity between that channel and key information. That is, the following formula focuses attention on the mineral deposits to be identified, quickly extracts features from these areas, and captures the correlations in the channel data of the geochemical exploration data;

[0074] g(x) = Sigmoid(W2ReLU(W1x))

[0075] Sigmoid and ReLU are activation functions, and W1 and W2 are the weight parameters of the fully connected layer. After calculating the channel attention, an attention weight matrix g(x) is obtained. This matrix contains the weight of each channel during network operation. Multiplying the obtained weight matrix g(x) by the original feature map x yields the feature information needed for the next layer, i.e., the input to the next layer. Figure 4 As shown.

[0076] Step 6: Input the training sample data at a ratio of 1:50000 into the convolutional neural network model for iteration and update the parameters of the neural network.

[0077] The number of samples from the ore-forming area is far less than the number of samples from the background area, indicating an imbalance in sample size; assuming the sample size is x... i After processing through a neural network, the output P = [p1, p2, ..., p] is obtained. N ] represents the classification probability corresponding to each sample, and the classification probability p corresponding to sample i is extracted. i ,make This term represents the cross-entropy of multi-class classification. The smaller the cross-entropy, the more accurate the neural network classification. To address the imbalance problem of easy and difficult classification samples, a gradient magnitude gm is introduced. i and gradient density GD(g i ), where the gradient magnitude gm i =|p i -y i |, where y i It's a real tag, p i It predicts the classification probability of the label of the i-th sample;

[0078]

[0079] Where g i This represents the gradient of the i-th sample.

[0080]

[0081] ε is a hyperparameter, resulting in the following loss function:

[0082]

[0083] N is the total number of samples. This way, during training, the weights of simple negative samples (i.e., most background images in mineral data) and very difficult samples (i.e., excessively anomalous mineralized area samples) will be reduced, meaning the loss will decrease and its impact on the model will be significantly reduced. The weights of normal difficult samples (i.e., ordinary mineralized area samples) will be increased, allowing the model to focus more on those effective normal difficult samples, i.e., generally mineralizable samples.

[0084] Training a neural network with the goal of minimizing the loss function can effectively prevent the network model from overfitting to the spatial data features of the background area and optimize the network model for learning the mineralization features of complex geochemical data.

[0085] Step 7: Use the trained convolutional neural network model to predict the mineralization probability of the geochemical data for the area to be predicted, generating a probability prediction distribution map of mineral resources in the area. Specifically, after obtaining the trained network model, input the geochemical data of the mining area to be predicted into the system, and use the sliding window algorithm to obtain the mineralization probability of each window area, ultimately forming a mineralization probability prediction distribution map of the entire mining area.

[0086] Based on the same inventive concept, this invention discloses a mineral prediction system based on multi-scale sample heterogeneity. A preprocessing module is used to convert collected geological data into image data, obtaining 1:P1 geochemical maps and 1:P2 geochemical maps. The geochemical maps are then divided into grids according to a certain pixel ratio, and training samples are formed based on expert knowledge. Each prediction map is divided into grids; if a grid contains a mineral deposit, it is selected as a mineralized area; otherwise, it is selected as a background area.

[0087] An anomaly detection module is used to introduce a generative adversarial network into a deep convolutional autoencoder. The generator in the generative adversarial network acts as the decoder in the autoencoder. The encoder maps grid data of sample size 1:P1 from the image space to the latent space. Then, the generative adversarial network is trained to obtain the output. The Euclidean distance between the input data and the output data is calculated as the geochemical anomaly score.

[0088] The data augmentation module is used to augment the data of the mining area and the obtained geochemical anomaly areas;

[0089] The classification model module is used to label samples in the 1:P2 grid data based on the geochemical anomaly score map obtained in the 1:P1 geochemical map. Geochemical anomaly scores above a set threshold are labeled as positive samples, and otherwise as negative samples. Based on this, specific geochemical data classification is performed. The module also constructs a convolutional neural network model and adds an attention mechanism to the model so that it focuses on the information of the region of interest, thereby learning the pattern between mineralization and geochemical data.

[0090] The model training module is used to input 1:P2 training sample data into the convolutional neural network model for iteration and update the parameters of the neural network.

[0091] Additionally, the classification and prediction module uses the trained convolutional neural network model to predict the mineralization probability of the geochemical data for the area to be predicted, generating a probability prediction distribution map of mineral resources in the area.

[0092] The specific working processes of each module described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. The division of modules is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple modules may be combined or integrated into another system.

[0093] Based on the same inventive concept, an embodiment of the present invention discloses a computer system including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded onto the processor, it implements the aforementioned mineral prediction method based on multi-scale sample heterogeneity.

[0094] Those skilled in the art will understand that the technical solution of this invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer system (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in the embodiments of this invention. The storage medium includes various media capable of storing computer programs, such as a USB flash drive, portable hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

Claims

1. A mineral resource prediction method based on multi-scale sample heterogeneity, characterized in that, Includes the following steps: Step 1: Convert the collected geological data into image data to obtain a 1:P1 geochemical map and a 1:P2 geochemical map, where 1:P1 and 1:P2 represent the ratio of the distance on the map to the actual distance, and P1 is greater than P2. Step 2: Divide the geochemical map into a grid according to a certain pixel ratio, and form training samples based on expert knowledge; Each predicted distribution map is divided into grids. If a grid contains a mineral deposit, it is selected as a metallogenic area; otherwise, it is selected as a background area. Step 3: Perform data augmentation on the obtained training samples; Step 4: Introduce the Generative Adversarial Network (GAN) into the deep convolutional autoencoder, where the generator in the GAN serves as the decoder in the autoencoder. The encoder maps grid data of sample size 1:P1 from the image space to the latent space. Then, train the GAN to obtain the output, and calculate the Euclidean distance between the input data and the output data as the geochemical anomaly score. The encoder is trained using background region data with the aim of minimizing the loss function in order to generate an image that is more similar to the feature distribution of the original image. Step 5: Based on the geochemical map obtained from the 1:P1 geochemical map, the samples are labeled in the 1:P2 grid data. Geochemical anomaly scores higher than a set threshold are labeled as positive samples, otherwise they are labeled as negative samples. Based on this, specific geochemical data are classified. A convolutional neural network model is constructed, and an attention mechanism is added to the model so that the model focuses on the information of the region of interest, thereby learning the pattern between mineralization and geochemical data. Step 6: Input the training sample data 1:P2 into the convolutional neural network model for iteration, and update the parameters of the neural network; the loss function for training the network model is: Where L CE (p i ,y i ) represents the cross-entropy of multi-class classification, y i It's a real tag, p i It predicts the label classification probability, where N is the total number of samples; GD(g i ) is the gradient density; g i Gm represents the gradient. i This represents the gradient magnitude, and ε is a hyperparameter. Step 7: Use the trained convolutional neural network model to predict the mineralization probability of the geochemical data of the area to be predicted, and generate a probability prediction distribution map of mineral resources in the area.

2. The mineral resource prediction method based on multi-scale sample heterogeneity according to claim 1, characterized in that, The geological data collected in step 1 includes geographic location information and geochemical element information. The geographic location information includes the latitude and longitude coordinates of the known mining area, and the geochemical element information includes the content of various geochemical elements in the known mining area. The geographic location information and geochemical element information are input into ArcGIS software, and the geochemical factors are converted into geochemical spatial data through inverse distance weighted interpolation to obtain geochemical maps at different scales.

3. The mineral resource prediction method based on multi-scale sample heterogeneity according to claim 1, characterized in that, In step 3, data augmentation is performed by adding noise and using generative adversarial networks to generate new samples.

4. The mineral resource prediction method based on multi-scale sample heterogeneity according to claim 1, characterized in that, Step 4 includes the following steps: Step 41: Construct an encoder E using a convolutional neural network. Input the 1:P1 background region data sample after data augmentation into the encoder, and perform feature extraction through the encoder to map it to the latent space. Step 42: A generative adversarial network (GAN) is used after the encoder network, where the generator G in the GAN acts as the decoder in the autoencoder to complete the mapping from the latent space to the image space; in order to obtain the feature distribution of the latent space where the background region is located, the loss function of the input image x and the reconstructed image G(E(x)) is minimized: Where n is the number of pixels in the image; Step 43: During encoder training, the statistical information of real and generated images computed in the generative adversarial network discriminator is used to guide the training of encoder E. The loss function of discriminator D guiding encoder training is as follows: Where, n d The dimension of the intermediate feature representation, k is the weighting factor, ||·|| 2 f(·) represents the square of the pixel difference, and f(·) is the feature generated by the discriminator D during the operation, the purpose of which is to statistically analyze the information of the given input. Step 44: Train encoder E using background region data with the aim of minimizing loss function L2 to generate an image that is more similar to the feature distribution of the original image; Step 45: After training, calculate the Euclidean distance between the input image x and the output image y generated by the trained generative adversarial network as the geochemical anomaly score.

5. A mineral resource prediction method based on multi-scale sample heterogeneity according to claim 1, characterized in that, Step 5 includes the following steps: Step 51: Based on the anomaly score, obtain the cluster of anomalies in the 1:P1 geochemical map, and obtain the geochemical data of the corresponding 1:P2 locations based on the existing data; Step 52: Construct a convolutional neural network model. The input data of the model is a c×h×w matrix, where c represents the number of image channels of the geochemical data, each channel contains information of a geochemical element, and h and w represent the height and width of the image. The data output format of the convolutional neural network is a probability vector representing each mineral type or background area. Step 53: In constructing the convolutional neural network model, an attention mechanism is integrated to apply a weight to the feature map on each channel. That is, attention is focused on the mineral area to be identified through the following formula, and feature extraction processing is performed on these areas to capture the correlation in the data of each channel in the geochemical exploration data. g(x) = Sigmoid(W2ReLU(W1x)) Sigmoid and ReLU are activation functions, and W1 and W2 are weight parameters of the fully connected layer. After the above channel attention calculation, an attention weight matrix g(x) is obtained. The weight matrix g(x) is multiplied by the original feature map x to obtain the feature information required for the next layer of the network.

6. A mineral resource prediction system based on multi-scale sample heterogeneity, characterized in that, include: The preprocessing module is used to convert the collected geological data into image data and obtain geochemical maps of 1:P1 and 1:P2, where 1:P1 and 1:P2 represent the ratio of the distance on the map to the actual distance, and P1 is greater than P2. And to divide the geochemical map into grids according to a certain pixel ratio, and to form training samples based on expert knowledge; Each predicted distribution map is divided into grids. If a grid contains a mineral deposit, it is selected as a mineralized area; otherwise, it is selected as a background area. An anomaly detection module is used to introduce a generative adversarial network into a deep convolutional autoencoder. The generator in the generative adversarial network acts as the decoder in the autoencoder. The encoder maps grid data of sample size 1:P1 from the image space to the latent space. Then, the generative adversarial network is trained to obtain the output. The Euclidean distance between the input data and the output data is calculated as the geochemical anomaly score. The encoder is trained using background region data with the aim of minimizing the loss function in order to generate an image that is more similar to the feature distribution of the original image. The data augmentation module is used to augment the data of the mining area and the obtained geochemical anomaly areas; The classification model module is used to label samples in the 1:P2 grid data based on the geochemical anomaly score map obtained in the 1:P1 geochemical map. Geochemical anomaly scores above a set threshold are labeled as positive samples, and otherwise as negative samples. Based on this, specific geochemical data classification is performed. The module also constructs a convolutional neural network model and adds an attention mechanism to the model so that it focuses on the information of the region of interest, thereby learning the pattern between mineralization and geochemical data. The model training module is used to input 1:P2 training sample data into the convolutional neural network model for iteration and to update the parameters of the neural network; the loss function for training the network model is: Where L CE (p i ,y i ) represents the cross-entropy of multi-class classification, y i It's a real tag, p i It predicts the label classification probability, where N is the total number of samples; GD(g i ) is the gradient density; g i Gm represents the gradient. i This represents the gradient magnitude, and ε is a hyperparameter. Additionally, the classification and prediction module uses the trained convolutional neural network model to predict the mineralization probability of the geochemical data for the area to be predicted, generating a probability prediction distribution map of mineral resources in the area.

7. A mineral prediction system based on multi-scale sample heterogeneity according to claim 6, characterized in that, In the anomaly detection module, the encoder is trained using background region data. The loss function for encoder training is as follows: Where n is the number of pixels in the image, n d The dimension of the intermediate feature representation, k is the weighting factor, ||·|| 2 f(·) represents the square of the pixel difference, and f(·) is the feature generated by the discriminator D during the operation, with the purpose of statistically analyzing the information of the given input.

8. A computer system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is loaded into the processor, it implements a mineral prediction method based on multi-scale sample heterogeneity according to any one of claims 1-5.

Images