Geochemical field reconstruction methods and apparatus based on ore-forming self-organization processes

Abstract

This application provides a geochemical field reconstruction method and apparatus based on the self-organizing process of mineralization. An improved void space pyramid pooling module is embedded in the target convolutional neural network. This module includes deformable convolutional branches with multiple void ratios and renormalization group constraint branches. The renormalization group constraint branches incorporate fractal power law constraints. The reconstruction process follows the self-organizing behavior of the mineralization process through these multiple deformable convolutional branches with different void ratios and fractal power law constraints. Furthermore, multi-scale feature sampling is achieved through these multiple deformable convolutional branches with different void ratios, improving the matching degree between the reconstructed geochemical data and the actual mineral deposits. Narrowing the target area improves the removal of "false anomalies," ultimately enhancing the mineral exploration indication value of the reconstructed geochemical data.

Description

Technical Field

[0001] This application relates to the field of mineral exploration technology, and in particular to a geochemical field reconstruction method and apparatus based on the self-organizing process of mineralization. Background Technology

[0002] Geochemical surveying is one of the most direct and effective means of obtaining subsurface information in mineral exploration. Its core task is to identify anomalies directly related to mineralization, or "mineralized anomalies," from a complex geochemical background field. Areas with high values ​​of mineralized chemical elements are considered to have a higher probability of being mineral deposits, and these locations are identified as potential mineral deposits. However, the spatial distribution of geochemical elements is not only controlled by primary geological mineralization processes but also strongly influenced by surface geochemical processes (such as weathering, erosion, transportation, and deposition).

[0003] In actual exploration work, especially in vegetated areas, loess-covered areas, or Quaternary sedimentary areas, surface soil or stream sediment samples often contain a large number of "false anomalies" formed by secondary enrichment at the surface. These false anomalies may have high numerical values, but they have no direct connection with deep ore bodies and can easily mislead the exploration direction, resulting in huge waste of funds. Therefore, it is necessary to reconstruct geochemical data to remove "false anomalies".

[0004] Currently, with the development of artificial intelligence (AI) technology, using deep learning to process geochemical data has become a cutting-edge trend in mineral prediction. However, conventional CNNs require a large amount of sample data for training and to obtain the distribution characteristics of geochemical data from a large amount of data. However, geological prospecting is a typical "small sample" or even "zero sample" scenario. In an exploration area, the number of known mineral deposits (positive samples) is extremely small, or even zero. Therefore, due to the small number of samples, the model training effect is poor. Furthermore, since the geochemical data used as sample data itself contains "false anomalies" caused by surface geochemical processes, using these geochemical data as samples will further reduce the model's performance, resulting in poor reconstruction of geochemical data and thus reducing the accuracy of geochemical field reconstruction based on mineralization self-organization processes. Summary of the Invention

[0005] This application provides a geochemical field reconstruction method and apparatus based on the self-organizing process of mineralization, in order to solve the technical problems mentioned in the background art.

[0006] In a first aspect, this application provides a geochemical field reconstruction method based on ore-forming self-organization processes, including: Acquire the geochemical data to be reconstructed; The geochemical data to be reconstructed is input into the target convolutional neural network, and the target convolutional neural network is controlled to extract features from the geochemical data to be reconstructed to obtain a first feature map. The target convolutional neural network includes at least three decoding layers, and any one of the intermediate decoding layers is provided with an improved void space pyramid pooling module. The improved void space pyramid pooling module includes multiple deformable convolutional branches with different void ratios. A second feature map is obtained based on the first feature map and the portion between the last encoding layer and the decoding layer in the target convolutional neural network, which is equipped with the improved hollow spatial pyramid pooling module. The second feature map is processed by multiple deformable convolutional branches with different dilation rates to obtain a feature map corresponding to each deformable convolutional branch. Based on the feature map corresponding to each of the deformable convolutional branches, and the structure after the intermediate decoding layer in the target convolutional neural network with an improved hollow spatial pyramid pooling module, the reconstructed geochemical data is obtained.

[0007] Secondly, this application provides a geochemical field reconstruction device based on a self-organizing ore-forming process, comprising: The acquisition module is used to acquire the geochemical data to be reconstructed. The model processing module is used to input the geochemical data to be reconstructed as input data into a target convolutional neural network, control the target convolutional neural network to extract features from the geochemical data to be reconstructed, and obtain a first feature map. The target convolutional neural network includes at least three decoding layers, and any one of the intermediate decoding layers is provided with an improved voided spatial pyramid pooling module. The improved voided spatial pyramid pooling module includes multiple deformable convolutional branches with different void ratios. A second feature map is obtained based on the first feature map and the portion between the last encoding layer in the target convolutional neural network and the decoding layer provided with the improved voided spatial pyramid pooling module. The second feature map is processed by multiple deformable convolutional branches with different porosity to obtain a feature map corresponding to each deformable convolutional branch; and the reconstructed geochemical data is obtained based on the feature map corresponding to each deformable convolutional branch and the structure after the intermediate decoding layer of the target convolutional neural network with an improved porosity spatial pyramid pooling module.

[0008] Thirdly, this application provides an electronic device, including: a processor and a memory; The memory stores instructions that the computer executes; The processor executes computer execution instructions stored in memory, causing the processor to perform the method as described in any of the first aspects.

[0009] Fourthly, embodiments of this application provide a readable storage medium including a program or instructions that, when run on a computer, execute the method described in any of the first aspects above.

[0010] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the method described in any of the first aspects.

[0011] The geochemical field reconstruction method and apparatus based on the self-organizing process of mineralization provided in this application have the following beneficial effects: An improved void space pyramid pooling module is embedded in the target convolutional neural network. This module includes deformable convolutional branches with multiple void ratios and renormalization group constraint branches. The renormalization group constraint branches incorporate fractal power-law constraints. During the reconstruction of the geochemical data to be reconstructed, these branches ensure that the reconstruction process follows the self-organizing behavior of the mineralization process. Furthermore, the multiple deformable convolutional branches with varying void ratios enable multi-scale feature sampling, improving the matching degree between the reconstructed geochemical data and the actual mineral deposits, thus narrowing the target area. Moreover, the inclusion of multiple deformable convolutional branches with different void ratios and fractal power-law constraints reduces the dependence on sample data when training the target convolutional neural network. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 A flowchart of a geochemical field reconstruction method based on ore-forming self-organizing processes provided in an embodiment of this application; Figure 2 A schematic diagram of the structure of an improved U-Net model provided in an embodiment of this application; Figure 3 A schematic diagram illustrating the working principle of an improved void space pyramid pooling module provided in an embodiment of this application; Figure 4 A schematic diagram of the structure of a geochemical field reconstruction device based on a self-organizing mineralization process provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application; Figure 6 A gold mine prediction map; Figure 7 This is a map showing the prediction of arsenic deposits. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions in the embodiments of this application are described clearly and completely below. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are also within the scope of protection of this application.

[0015] When using geochemical data for mineral exploration, the goal is to identify anomalies directly related to mineralization—the "mineralized anomalies"—from a complex geochemical background field. However, the spatial distribution of geochemical elements is not only controlled by primary geological mineralization processes but also strongly influenced by surface geochemical processes (such as weathering, erosion, transportation, and deposition). For example, water flow can carry chemical elements to low-lying areas, but these areas may not actually contain mineral deposits. In other words, geochemical data can contain "false anomalies." Similarly, geophysical and remote sensing data can also contain "false anomalies" caused by other factors. Therefore, using geochemical, geophysical, or remote sensing data containing "false anomalies" for mineral exploration prediction is inefficient and inaccurate.

[0016] Therefore, when using geochemical data for mineral exploration, it is necessary to first eliminate "false anomalies". In the current technology, the way to eliminate "false anomalies" is through the experience of the staff, but this method is highly subjective and inefficient.

[0017] With the development of artificial intelligence, some researchers have also used AI models to process geochemical and geophysical data. However, conventional CNNs require a large amount of sample data for training and to obtain the distribution characteristics of geochemical data from a large amount of data. However, geological prospecting is a typical "small sample" or even "zero sample" scenario. In an exploration area, the number of known mineral deposits (positive samples) is extremely small, or even zero. Therefore, due to the small number of samples, the model training effect is poor. Furthermore, since the geochemical data used as sample data itself contains "false anomalies" caused by surface geochemical processes, using these geochemical data as samples will further reduce the model's performance.

[0018] In addition, for mineralization, the formation of mineral deposits is a complex self-organizing process. The spatial distribution of chemical elements often follows fractal or multifractal laws (power-law distribution). However, current "black box" AI models lack such mathematical constraints. Geochemical data after removing "false anomalies" often loses the original geological texture details or produces smooth transitions that violate geological laws.

[0019] Furthermore, during AI model training, existing self-supervised autoencoders struggle to balance the challenge of removing noise—specifically, "false anomalies"—caused by sampling errors, analysis errors, and geological impurities in geochemical data while preserving the weak, localized heterogeneous signals crucial for mineral exploration.

[0020] Therefore, to address the technical problems existing in the prior art, this application proposes a geochemical field reconstruction method and apparatus based on the self-organizing process of mineralization. An improved void space pyramid pooling module is set in the decoding layer of the neural network model. This module includes multiple deformable convolutional branches with different void ratios. When reconstructing geochemical data by removing "false anomalies," the complex boundary contours of anomalous patches are extracted using these deformable convolutional branches with different void ratios in the decoding layer. This more accurately maintains the spatial continuity and true geometric structure of the anomalous region during the reconstruction process, making the boundaries of the reconstructed geochemical data more consistent with the self-organizing process during mineralization, thus improving the accuracy of mineral exploration using the reconstructed geochemical data.

[0021] Figure 1 This is a flowchart illustrating a geochemical field reconstruction method based on a self-organizing mineralization process, provided as an embodiment of this application. The method described in this embodiment can be executed by an electronic device such as a laptop with logical operation capabilities, for example... Figure 1 As shown, the method includes: S101. Obtain the geochemical data to be reconstructed.

[0022] In this step, the geochemical data to be reconstructed refers to the original geochemical data of the area to be surveyed. This data can be obtained through publicly available databases, such as the China Geological Survey's "Geological Cloud" and the Geochemical Research Database of the Chinese Academy of Sciences, or through on-site surveys and investigations to ensure the authenticity of the geochemical data.

[0023] S102. Input the geochemical data to be reconstructed into the target convolutional neural network as input data, and control the target convolutional neural network to extract features from the geochemical data to be reconstructed to obtain the first feature map.

[0024] The target convolutional neural network includes at least three decoding layers, and any one of the intermediate decoding layers is provided with an improved dilated spatial pyramid pooling module, which includes multiple deformable convolutional branches with different dilation rates.

[0025] In this step, the target convolutional neural network is generally selected from image processing network models. This embodiment uses U-Net as an example. Based on the traditional U-Net structure, this embodiment improves the traditional U-Net structure to obtain an improved U-Net model, which is the target convolutional neural network of this application. Figure 2 As shown, the improved U-Net model includes: 4 encoding layers, a bottleneck layer, and 4 decoding layers.

[0026] Among them, for Figure 2 The improved U-Net model shown extracts features from the geochemical data to be reconstructed through four coding layers to obtain the first feature map. The specific process by which the improved U-Net model obtains the first feature map can be found in existing techniques and will not be elaborated here.

[0027] For four decoding layers, in order to enable the deformable convolutional branches to capture features that better match the contours of real abnormal patches, an improved dilated spatial pyramid pooling module is typically set in the decoding layer preceding the last decoding layer. Therefore, as... Figure 2 As shown, an improved void space pyramid pooling module is set in the third decoding layer.

[0028] For multiple deformable convolution branches with different void ratios, for example, such as Figure 3 As shown, four parallel deformable convolutional branches with different dilatancy rates are set, where the dilatancy rates are set to [1, 12, 24, 36].

[0029] S103. Obtain a second feature map based on the first feature map and the portion between the last encoding layer and the decoding layer with an improved hollow spatial pyramid pooling module in the target convolutional neural network.

[0030] In this step, after obtaining the first feature map, the second feature map is obtained by passing through the bottleneck layer, the first decoding layer, and the second decoding layer. The specific process of obtaining the second feature map through the bottleneck layer, the first decoding layer, and the second decoding layer can be referred to the existing technology, and will not be repeated here.

[0031] S104. Perform feature processing on the second feature map through multiple deformable convolution branches with different dilation rates to obtain the feature map corresponding to each deformable convolution branch.

[0032] In this step, the second feature map is input into the third decoding layer. Each deformable convolutional branch extracts features from the second feature map to obtain the feature map corresponding to each deformable convolutional branch. The learnable offset of each deformable convolutional branch is obtained through model training.

[0033] S105. Based on the feature map corresponding to each deformable convolutional branch, and the structure after the intermediate decoding layer in the target convolutional neural network with the improved hollow spatial pyramid pooling module, the reconstructed geochemical data is obtained.

[0034] In this step, feature fusion is performed on the four feature maps and the feature map obtained by the fully connected layer. For example, channel stitching is used to obtain the third feature map. The reconstructed geochemical data is obtained through the third feature map and the fourth decoding layer. Compared with the original geochemical data, the reconstructed geochemical data removes interference caused by impurities and noise, and has fewer "false anomalies".

[0035] The process of processing the third feature map by the fourth coding layer and subsequent structures can be referred to existing technologies, and will not be elaborated here.

[0036] After obtaining the reconstructed geochemical data via S105, the method further includes: S106. Mineral resource prediction based on reconstructed geochemical data.

[0037] In this step, the reconstructed geochemical data is applied to mineral prediction to identify areas where minerals may exist, i.e., to determine the target area and conduct mineral prediction.

[0038] Optionally, one specific implementation of S106 is as follows: S1061. The reconstructed geochemical data is upsampled using fractal interpolation to obtain upsampled geochemical data. S1062. Mineral resource prediction based on upsampled geochemical data.

[0039] Specifically, the reconstructed geochemical data is upsampled using fractal interpolation from 1:200,000 to 1:50,000, resulting in a more refined anomalous structure in the self-organized pseudo-segmentation mineralization prediction map and enhancing the contrast between the target and the background.

[0040] In this embodiment, the acquired geochemical data to be reconstructed is input into a target convolutional neural network (CNN). The CNN includes at least three decoding layers, and any one of the intermediate decoding layers is equipped with an improved voided spatial pyramid pooling module. The improved voided spatial pyramid pooling module includes multiple deformable convolutional branches with different void ratios. The CNN is controlled to extract features from the geochemical data to be reconstructed, obtaining a first feature map. The first feature map is used as input data for the portion between the last encoding layer and the decoding layer equipped with the improved voided spatial pyramid pooling module in the CNN, obtaining a second feature map. Anisotropic features are extracted from the second feature map through multiple deformable convolutional branches with different void ratios, obtaining a feature map corresponding to each deformable convolutional branch. Based on the feature maps corresponding to each deformable convolutional branch and the structure after the intermediate decoding layer equipped with the improved voided spatial pyramid pooling module in the CNN, the reconstructed geochemical data is obtained. This embodiment incorporates an improved void spatial pyramid pooling module with multiple deformable convolutional branches of varying void ratios embedded in the decoding layer. Utilizing the dynamic adjustment of sampling point positions of the convolutional kernels by these deformable convolutional branches, the branches with different void ratios can actively "fit" the complex and varied boundaries and morphologies of geochemical anomaly patches, such as curved fault zones or irregular alteration halos. This effectively avoids the fragmentation and blurring of continuous geological bodies during feature extraction, thus maintaining the spatial continuity and true geometric structure of the anomaly region more accurately during reconstruction. The reconstructed geochemical data better reflects the self-organizing process of mineralization, improving the removal of "false anomalies" and ultimately enhancing the mineral exploration indication value of the reconstructed geochemical data.

[0041] Figure 1 The illustrated embodiment, by setting deformable convolution branches with different void ratios, makes the reconstructed geochemical data more consistent with the self-organization of mineralization, thus addressing the issue of further simulating self-organizing behavior during the mineralization process, such as... Figure 2 As shown, a renormalization group constraint branch can be set in the improved hollow space pyramid pooling module to embed fractal power law constraints into the target convolutional neural network.

[0042] For the k-th deformable convolutional branch, its feature extraction scale is s k The data in the feature map corresponding to that branch is averaged to obtain the mean of the feature map. m k Then the feature extraction scale s k and the mean of the feature map m k The relationships between them must satisfy fractal power law constraints:

[0043] Here, α is the fractal dimension, and C is a constant obtained through model training. During training, the goodness of fit Gof of this linear relationship is calculated and used as part of the loss function, forcing the features learned by the convolutional neural network to conform to the fractal growth law of geological processes, so that the reconstructed geochemical data can more realistically reflect the self-organizing behavior in the mineralization process.

[0044] When a renormalization group constraint branch is set in the improved hollow space pyramid pooling module, and fractal power law constraints are embedded into the target convolutional neural network, the method further includes, based on the above embodiments, the target convolutional neural network with this structure: S201. Using the superpixel segmentation method, superpixel clustering is performed on the geochemical data to be reconstructed to obtain a segmentation map containing multiple superpixel regions.

[0045] In this step, the geochemical data to be reconstructed, which is regularly gridded, is combined with spatial coordinates to construct a feature vector. The SLIC algorithm is used to initialize the cluster centers, the composite distance between each pixel and the center is calculated iteratively, and the element similarity and spatial proximity are measured. The pixel allocation and center update are completed by balancing the weights through the compactness parameter, and a label map composed of "geochemical units" is obtained, which is a segmentation map containing multiple superpixel regions.

[0046] S202. Based on the fractal power law constraint between the feature maps corresponding to multiple deformable convolution branches with different dilatancy rates, obtain the position label of each convolution kernel at each sampling point when any deformable convolution branch extracts features from the second feature map.

[0047] Among them, the location label is used to identify the superpixel region where the convolution kernel of any deformable convolution branch is located.

[0048] In this step, we will take a deformable convolutional branch with a dilation rate of 12 as an example. When a 3×3 convolutional kernel has a dilation rate of 12, its receptive field (the actual input area it covers) becomes very large, equivalent to a standard convolutional kernel of about 25×25. However, the actual number of learnable weight parameters is still 3×3=9.

[0049] The central convolutional kernel is a convolutional kernel located at the center. When the central convolutional kernel is located at the target sampling point, due to the deformable convolution, each convolutional kernel is offset under the action of a learnable offset to determine the position of each convolutional kernel after the offset, that is, the label of the superpixel region corresponding to each convolutional kernel.

[0050] For the target sampling point, the output of the central convolutional kernel at each deformation sampling point (i.e., each convolutional kernel) is obtained by the following formula:

[0051] in, p 0 represents the target sampling point, Δ p n This represents a learnable offset that helps capture geometric deformations outside of a fixed mesh. p n This represents the sampling point when the central convolutional kernel is located at the target sampling point, and other convolutional kernels have not undergone positional offset.

[0052] The learnable offset of each convolutional kernel is obtained through model training.

[0053] S203. Based on the position label of the convolution kernel of any deformable convolution branch, perform self-organized pseudo-segmentation on the map to be segmented to obtain a mineralization prediction map after self-organized pseudo-segmentation of geochemical patches.

[0054] In this step, based on the labels of the superpixel regions corresponding to each convolution kernel, the label of the superpixel region with the most labels is determined. The frequency of occurrence of this label is compared with the preset number of labels. When the frequency of occurrence of this label is greater than the preset number of labels, the superpixel region corresponding to this label is determined as a geochemical patch with self-organizing characteristics, and a mineralization prediction map after self-organized pseudo-segmentation of the geochemical patch is obtained.

[0055] Therefore, based on the geochemical field reconstruction method based on the self-organizing process of mineralization, which includes S201-S203, one implementation of S106 is: to predict mineral resources based on the reconstructed geochemical data and the mineralization prediction map.

[0056] Specifically, the reconstructed geochemical data is obtained by extracting features from the anisotropic characteristics of the geochemical data, while the mineralization prediction map is obtained based on the self-organizing behavior of the mineralization process under the condition of fractal power law constraints. Therefore, mineral prediction based on the reconstructed geochemical data and the mineralization prediction map can narrow the target area and increase the probability of mineralization within the target area.

[0057] In this embodiment, superpixel clustering is combined with variable convolution. By leveraging the characteristic of variable convolution that allows for dynamic adjustment of the sampling point position of the convolution kernel, self-organized pseudo-segmentation is performed on the image to be segmented, which contains multiple different superpixel regions. This avoids the excessive sensitivity of traditional superpixel segmentation algorithms to parameters and also compensates for the sparsity and boundary loss of deformable convolution offset points.

[0058] Optional, such as Figure 2As shown, a variational inference module is set in the bottleneck layer of the improved U-Net model. The variational inference module performs convolution operation on the features output by the 4th layer encoder to obtain the corresponding mean μ and log-variance logσ². Then, a reparameterization technique is introduced to sample random noise ε ~ N(0,1) from the standard normal distribution. The standard deviation σ is obtained according to the formula σ = exp(0.5 × logσ²). Then, the latent variable z is generated by the formula z = μ + σ⊙ε, where ⊙ represents element-wise multiplication. This latent variable z is the probabilistic latent feature map. Therefore, even if the same original DEM data is input, the z will be slightly different each time the random noise ε is different, which introduces uncertainty. Finally, the latent variable z is input to the decoder to participate in the decoding operation and obtain the reconstructed geochemical data.

[0059] By setting a variational inference module in the bottleneck layer, the encoder output is smoothed. Since the noise of the variational inference module is sampled from a standard normal distribution, it is equivalent to introducing a probabilistic model into the improved U-Net model. Random noise and local fluctuations are unique and not shared, and are therefore suppressed. As a result, the improved U-Net model has stronger robustness to noise and local fluctuations in the geochemical data to be reconstructed, and the output reconstructed geochemical data is smoother. This widens the gap with the geochemical data to be reconstructed, increasing the residual data. The larger the residual data, the greater the probability of the existence of mineral deposits, thus avoiding the omission of mineral deposits during mineral exploration prediction.

[0060] Optional, based on Figure 2 The structure of the target convolutional neural network shown is obtained by training the initial convolutional neural network with the geochemical data to be reconstructed. The loss function used needs to consider structural similarity index, mean squared error loss, and the fractal power law constraint fit, etc. Specifically, its loss function formula is as follows: Loss=0.2 ×Color +SSIM(y, Geol ×y')+0.7*L2(y, Geol ×y')+ φ (1- gof ,0)+(-ELBO) Where y and y' represent the geochemical data to be reconstructed and the reconstructed geochemical data, respectively.

[0061] Color represents the color histogram loss, which uses differentiable histogram matching to maintain color consistency before and after reconstruction.

[0062] Geol: Geological Weighted Map. Created using geomorphic anomaly masks, known faults, and rock mass distributions, and normalized to 1.0-1.5, it serves as a weighted mask to weight the Structural Similarity Index (SSIM) and the mean square error (L2) loss, thereby enhancing the reconstruction accuracy of favorable mineralization areas.

[0063] Gof: The goodness of fit of the fractal power law constraint in the renormalization group constraint branch, where φ is the penalty coefficient to ensure the fractal properties of the features.

[0064] ELBO: Lower bound of evidence for variational inference. Negative values ​​are used to maximize the lower bound and optimize the latent distribution.

[0065] SSIM: Actually 1-SSIM, representing feature optimization for reconstruction details.

[0066] Therefore, through Figure 2 The structure of the target convolutional neural network shown, and the loss function used during training, enable the target convolutional neural network to reconstruct the geochemistry to be reconstructed, achieving the following results: By adding structural similarity index, mean squared error loss, fractal power law constraint fitting degree to the loss function, and setting an improved void space pyramid pooling module in the intermediate decoding layer, and embedding a renormalization group constraint branch in the improved void space pyramid pooling module, the fractal power law constraint is embedded into the target convolutional neural network. This enables the target neural network to reconstruct the geochemical field under geological constraints, and utilizes multiple deformable convolutional branches with different void ratios to divide the geochemical field into self-organized patch units. This achieves noise reduction and interference elimination of the geochemical data to be reconstructed, as well as enriching the local details of the reconstructed data.

[0067] It should be noted that when applying the geochemical data to be reconstructed, for example, when performing step S101, and when training the initial convolutional neural network to obtain the target convolutional neural network, it is necessary to obtain the geochemical data to be reconstructed. Specifically, the original geochemical data (such as 20 elements such as Cu, Pb, and Zn) is kriging-interpolated to 670×1000 pixels, normalized to [0, 1], and cropped into 512×512 image blocks (10% overlap) to obtain the geochemical data to be reconstructed.

[0068] In training the initial convolutional neural network to obtain the target convolutional neural network, the ReLU activation function is used to avoid gradient vanishing. Edge processing is also required, employing a "clone-replacement" strategy to eliminate edge effects introduced by the convolution operation.

[0069] To verify the effectiveness of the method described in this application, mineral prediction for arsenic and gold deposits was conducted using the Baren Zhelimu mining area in Inner Mongolia as an example. Figure 6This is a gold mine prediction map. Figure 7 For arsenic ore prediction map, through Figure 6 and Figure 7 It is evident that, using the scheme presented in this application, the extracted anomalous morphologies (i.e., the mineral deposits determined based on the reconstructed geochemical data) highly match the spatial distribution of known mineral deposits. Furthermore, through... Figure 6 The comparison between the four figures in the text, and Figure 7 The comparison between the four maps shows that by upsampling the reconstructed geochemical data, a more refined anomalous structure can be displayed in the mineralization prediction map after self-organized pseudo-segmentation.

[0070] Figure 4 A schematic diagram of a geochemical field reconstruction device based on a self-organizing mineralization process provided in an embodiment of this application is shown below. Figure 4 As shown, the geochemical field reconstruction device based on the self-organizing process of mineralization includes: an acquisition module 401 and a model processing module 402. Optionally, such as... Figure 4 As shown, the geochemical field reconstruction device based on the ore-forming self-organizing process also includes at least one of the following: prediction module 403, segmentation module 404, and training module 405.

[0071] Acquisition module 401 is used to acquire the geochemical data to be reconstructed; The model processing module 402 is used to input the geochemical data to be reconstructed as input data into a target convolutional neural network, control the target convolutional neural network to extract features from the geochemical data to be reconstructed, and obtain a first feature map. The target convolutional neural network includes at least three decoding layers, and any intermediate decoding layer among the at least three decoding layers is provided with an improved voided spatial pyramid pooling module. The improved voided spatial pyramid pooling module includes multiple deformable convolutional branches with different void ratios. A second feature map is obtained based on the first feature map and the portion between the last encoding layer in the target convolutional neural network and the decoding layer provided with the improved voided spatial pyramid pooling module. Feature processing is performed on the second feature map through multiple deformable convolutional branches with different void ratios to obtain a feature map corresponding to each deformable convolutional branch. Reconstructed geochemical data is obtained based on the feature map corresponding to each deformable convolutional branch and the structure after the intermediate decoding layer in the target convolutional neural network provided with the improved voided spatial pyramid pooling module.

[0072] Optionally, the improved void space pyramid pooling module further includes: a renormalization group constraint branch, wherein the renormalization group constraint branch is provided with fractal power law constraints; The model processing module 402 performs feature processing on the second feature map through multiple deformable convolutional branches with different dilation rates to obtain a feature map corresponding to each deformable convolutional branch, specifically for: Based on the renormalization group constraint branch, feature extraction is performed on the second feature map through multiple deformable convolution branches with different dilation rates to obtain a feature map corresponding to each deformable convolution branch. The feature maps corresponding to the multiple deformable convolution branches with different dilation rates satisfy the fractal power law constraint.

[0073] Optionally, the segmentation module 404 is used to perform superpixel clustering on the geochemical data to be reconstructed using a superpixel segmentation method to obtain a segmentation map containing multiple superpixel regions; The model processing module 402 is further configured to, based on satisfying the fractal power law constraint among the feature maps corresponding to multiple deformable convolution branches with different void ratios, obtain the position label of each convolution kernel at each sampling point when any deformable convolution branch extracts features from the second feature map, wherein the position label is used to identify the superpixel region where the convolution kernel of any deformable convolution branch is located, and perform self-organized pseudo-segmentation on the image to be segmented according to the position label of the convolution kernel of any deformable convolution branch to obtain a mineralization prediction map after self-organized pseudo-segmentation of geochemical patches.

[0074] Optionally, after the model processing module 402 obtains the reconstructed geochemical data, the prediction module 403 is used for: Mineral resource prediction is performed based on the reconstructed geochemical data.

[0075] Optionally, the prediction module 403 performs mineral resource prediction based on the reconstructed geochemical data, specifically for: Mineral resources are predicted based on the reconstructed geochemical data and the mineralization prediction map.

[0076] Optionally, before the model processing module 402 inputs the geochemical data to be reconstructed as input data into the target convolutional neural network, the training module 405 is used to: The initial convolutional neural network is trained based on the geochemical data to be reconstructed to obtain the target convolutional neural network. The loss function includes at least: structural similarity index, mean squared error loss, and fractal power law constraint fit. When calculating the structural similarity index and the mean squared error loss, a weight mask is added to the reconstructed geochemical data. The weight mask is generated by geological maps or geomorphological analysis.

[0077] Optionally, the prediction module 403 performs mineral resource prediction based on the reconstructed geochemical data, specifically for: The reconstructed geochemical data were upsampled using fractal interpolation to obtain upsampled geochemical data. Mineral resources are predicted based on the upsampled geochemical data.

[0078] The geochemical field reconstruction device based on the ore-forming self-organizing process provided in this application embodiment can be referred to the above method embodiment for its specific implementation process. The implementation principle and technical effect are similar, and will not be repeated here.

[0079] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device may be a server, such as... Figure 5 As shown, the electronic device includes a processor 501 and a memory 502.

[0080] The memory 502 stores computer-executed instructions.

[0081] The processor 501 executes the computer execution instructions stored in the memory 502, causing the processor 501 to perform the method described in any of the above embodiments.

[0082] The electronic device provided in this application embodiment can be referred to the above method embodiment for its specific implementation process. The implementation principle and technical effect are similar, and will not be repeated here.

[0083] In the above Figure 5 In the illustrated embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0084] The memory may include high-speed RAM, and may also include non-volatile storage (NVM), such as at least one disk storage.

[0085] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0086] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the method shown in the above-described method embodiments.

[0087] The aforementioned computer-readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0088] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an application-specific integrated circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0089] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0090] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A geochemical field reconstruction method based on ore-forming self-organization processes, characterized in that, include: Acquire the geochemical data to be reconstructed; The geochemical data to be reconstructed is input into the target convolutional neural network, and the target convolutional neural network is controlled to extract features from the geochemical data to be reconstructed to obtain a first feature map. The target convolutional neural network includes at least three decoding layers, and any one of the intermediate decoding layers is provided with an improved void space pyramid pooling module. The improved void space pyramid pooling module includes multiple deformable convolutional branches with different void ratios. A second feature map is obtained based on the first feature map and the portion between the last encoding layer and the decoding layer in the target convolutional neural network, which is equipped with the improved hollow spatial pyramid pooling module. The second feature map is processed by multiple deformable convolutional branches with different dilation rates to obtain a feature map corresponding to each deformable convolutional branch. Based on the feature map corresponding to each of the deformable convolutional branches, and the structure after the intermediate decoding layer in the target convolutional neural network with an improved hollow spatial pyramid pooling module, the reconstructed geochemical data is obtained.

2. The method according to claim 1, characterized in that, The improved void space pyramid pooling module further includes: a renormalization group constraint branch, wherein the renormalization group constraint branch is provided with fractal power law constraints; The step of performing feature processing on the second feature map through multiple deformable convolutional branches with different dilation rates to obtain a feature map corresponding to each deformable convolutional branch includes: Based on the renormalization group constraint branch, feature extraction is performed on the second feature map through multiple deformable convolution branches with different dilation rates to obtain a feature map corresponding to each deformable convolution branch. The feature maps corresponding to the multiple deformable convolution branches with different dilation rates satisfy the fractal power law constraint.

3. The method according to claim 2, characterized in that, Also includes: The superpixel segmentation method is used to perform superpixel clustering on the geochemical data to be reconstructed, resulting in a segmentation map containing multiple superpixel regions. Based on satisfying the fractal power law constraint among the feature maps corresponding to multiple deformable convolution branches with different dilation rates, when any deformable convolution branch extracts features from the second feature map, the position label of each convolution kernel of the center convolution kernel of the any deformable convolution branch at each sampling point is obtained. The position label is used to identify the superpixel region where the convolution kernel of the any deformable convolution branch is located. Based on the position label of the convolution kernel of any deformable convolution branch, the map to be segmented is subjected to self-organized pseudo-segmentation to obtain a mineralization prediction map of the geochemical patch after self-organized pseudo-segmentation.

4. The method according to claim 3, characterized in that, Also includes: Mineral resource prediction is performed based on the reconstructed geochemical data.

5. The method according to claim 4, characterized in that, The mineral prediction based on the reconstructed geochemical data includes: Mineral resources are predicted based on the reconstructed geochemical data and the mineralization prediction map.

6. The method according to any one of claims 2-5, characterized in that, Before inputting the geochemical data to be reconstructed into the target convolutional neural network, the process further includes: The initial convolutional neural network is trained based on the geochemical data to be reconstructed to obtain the target convolutional neural network. The loss function includes at least: structural similarity index, mean squared error loss, and fractal power law constraint fit. When calculating the structural similarity index and the mean squared error loss, a weight mask is added to the reconstructed geochemical data. The weight mask is generated by geological maps or geomorphological analysis.

7. The method according to claim 4 or 5, characterized in that, The mineral prediction based on the reconstructed geochemical data includes: The reconstructed geochemical data were upsampled using fractal interpolation to obtain upsampled geochemical data. Mineral resources are predicted based on the upsampled geochemical data.

8. An electronic device, characterized in that, include: Processor and memory; The memory stores instructions that the computer executes; The processor executes computer execution instructions stored in memory, causing the processor to perform the method described in any one of claims 1-7.

9. A readable storage medium, characterized in that, Includes a program or instructions that, when run on a computer, execute the method as described in any one of claims 1-7.

10. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method as described in any one of claims 1-7.

Images