Pegmatite-Type Lithium Ore Prediction Method and Device Based on Deep Learning

Abstract

The embodiments of this specification provide a prediction method and device for pegmatite-type lithium ore based on deep learning. Among them, the method includes: obtaining multiple different types of data in the area to be collected, and converting the spatial coordinates of various types of data to the same spatial coordinate system to generate a geoscience spatial database; rasterizing various types of data in the geoscience spatial database to generate a prediction factor map, a mask map, and a prediction label map; generating a training dataset and a validation dataset based on the prediction factor map, the mask map, and the prediction label map of the target area; using the training dataset and the validation dataset to train a preset fully convolutional neural network to generate a prediction model; based on the prediction model, identify the distribution of pegmatite-type lithium ore in the selected area to be collected. By adopting the above technical solution, it is possible to quickly and effectively identify a batch of lithium-bearing pegmatite veins with a certain shape, scale, and spatial distribution.

Description

Technical Field

[0001] The embodiments of this specification relate to the technical field of mineral resource exploration, and particularly to a prediction method and device for pegmatite-type lithium ore based on deep learning. Background Art

[0002] In recent years, the development of the global green and low-carbon economy has greatly accelerated the growth of the international community's demand for lithium resources. Since the distribution range of pegmatite-type lithium ore is not as limited as that of brine-type lithium ore to only a few basins, and it has a relatively high grade and relatively mature mining and refining technologies, it has thus received extensive attention from the industrial and academic circles again.

[0003] However, with the improvement of exploration degree, the chance of discovering large-scale pegmatite-type lithium ore in eastern China is getting less and less. The focus of prospecting work has to shift to the western region with high altitude, deep cutting, difficult access, and low work degree. However, the western region is vast, and it is very difficult to conduct a comprehensive, systematic, and in-depth field survey. How to invest limited human, material, financial, and time resources into the most promising areas has become the first problem to be solved in the exploration of pegmatite-type lithium ore, which urgently requires the emergence of efficient and low-cost prospecting technical methods.

[0004] Therefore, how to quickly and effectively identify a batch of lithium-bearing pegmatite veins with certain shapes, scales, and spatial distributions has become a difficult point in the research of lithium ore prospecting exploration technology. Summary of the Invention

[0005] In view of this, the embodiments of this specification provide a prediction method and device for pegmatite-type lithium ore based on deep learning, which can quickly and effectively identify a batch of lithium-bearing pegmatite veins with certain shapes, scales, and spatial distributions.

[0006] The embodiments of this specification provide a prediction method for pegmatite-type lithium ore based on deep learning, including:

[0007] Obtain multiple different types of data in the area to be collected, and convert the spatial coordinates of various types of data to the same spatial coordinate system to generate a geoscience spatial database;

[0008] Perform rasterization processing on various types of data in the geoscience spatial database to generate a prediction feature map, a mask map, and a prediction label map;

[0009] Based on the prediction feature map, the mask map, and the prediction label map of the target area, generate a training data set and a validation data set, where the target area is the area in the area to be collected where pegmatite-type lithium ore bodies have been explored;

[0010] Use the training data set and the validation data set to train a preset fully convolutional neural network to generate a prediction model;

[0011] Based on the prediction model, identify the distribution of pegmatite-type lithium ore in the selected area to be collected.

[0012] The embodiment of this specification also provides a pegmatite-type lithium ore prediction device based on deep learning, including:

[0013] A data acquisition unit, which acquires multiple different types of data of the area to be collected, converts the spatial coordinates of various types of data into the same spatial coordinate system, and generates a geoscience spatial database;

[0014] A data processing unit, which performs rasterization processing on various types of data in the geoscience spatial database to generate a prediction feature map, a mask map, and a prediction label map;

[0015] A data generation unit, which generates a training data set and a validation data set based on the prediction feature map, the mask map, and the prediction label map of the target area, where the target area is the area in the area to be collected where pegmatite-type lithium ore bodies have been explored;

[0016] A processing unit, which uses the training data set and the validation data set to train a preset fully convolutional neural network to generate a prediction model; and based on the prediction model, identify the distribution of pegmatite-type lithium ore in the selected area to be collected.

[0017] By using the pegmatite-type lithium ore prediction method based on deep learning provided in the embodiment of this specification, through converting the spatial coordinates of multiple different types of data of the area to be collected, a geoscience spatial database can be generated. Then, by performing rasterization processing on various types of data in the geoscience spatial database, a prediction feature map, a mask map, and a prediction label map can be generated. Furthermore, based on the prediction feature map, the mask map, and the prediction label map of the target area, a training data set and a validation data set can be generated. Since the target area is the area in the area to be collected where pegmatite-type lithium ore bodies have been explored, when using the training data set and the validation data set to train a preset fully convolutional neural network, the obtained prediction model has better generalization ability, enabling the prediction model to better learn the relationship between the spatial structure of multiple different types of data and the target, so as to more comprehensively and objectively reflect the complex mapping relationship between the comprehensive information and the target. In this way, based on the prediction model, the distribution of pegmatite-type lithium ore in the selected area to be collected can be quickly and effectively identified, and then a batch of lithium-bearing pegmatite veins with certain shapes, scales, and spatial distributions can be identified. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only those of the embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on the provided drawings.

[0019] Figure 1 It shows a flowchart of a method for predicting pegmatite-type lithium ore based on deep learning in the embodiments of this specification;

[0020] Figure 2 It shows a flowchart of a rasterization process in the embodiments of this specification;

[0021] Figure 3 It shows a schematic diagram of converting a surface to a raster in the embodiments of this specification;

[0022] Figure 4 It shows a schematic diagram of a multi-ring buffer analysis in the embodiments of this specification;

[0023] Figure 5 It shows a flowchart of a generation method for a training dataset and a validation dataset in the embodiments of this specification;

[0024] Figure 6 It shows a schematic diagram of the generation process of a prediction model in the embodiments of this specification;

[0025] Figure 7 It shows a schematic diagram of a deep learning algorithm network structure in the embodiments of this specification;

[0026] Figure 8 It shows a schematic diagram of an attention mechanism in the embodiments of this specification;

[0027] Figure 9 It shows a schematic diagram of the principle of bilinear interpolation in the embodiments of this specification;

[0028] Figure 10 It shows a schematic diagram of the principle of area division in the embodiments of this specification;

[0029] Figure 11 It shows a schematic diagram of the structure of a device for predicting pegmatite-type lithium ore based on deep learning in the embodiments of this specification. Detailed implementation manners

[0030] As described in the background art, the current exploration schemes for pegmatite-type lithium ore are limited and cannot effectively explore pegmatite-type lithium ore.

[0031] With the advent of the era of big data in geoscience, the use of artificial intelligence (AI) to identify prospecting targets, drawing on the vast amounts of geological, mineral, geophysical, geochemical, and remote sensing data accumulated in previous studies, has become a crucial step in reducing prospecting costs, focusing on prospecting targets, and improving prospecting efficiency. However, these methods often suffer from significant drawbacks. For example, the recognition accuracy of traditional machine learning methods (such as feature analysis, information content, weight of evidence, logistic regression, support vector machines, and random forests) depends largely on the designer's ability to manually extract prospecting information based on prior knowledge, and they can only handle problems that are easy to classify in low-dimensional spaces. Self-organizing feedforward neural networks (such as BP) have the ability to deeply mine and highly integrate prospecting information from raw input, but they often do not consider the location and structure of the research object when extracting deep prospecting information, resulting in large number of parameters and low efficiency. Convolutional neural networks (such as LeNet, AlexNet, VGG, and ResNet) can simultaneously consider spatial structure and feature information, thereby extracting deep prospecting information, but they only determine whether there is a mineral within the study area, but cannot determine its specific location and spatial distribution. It is often difficult to systematically and comprehensively discover the complex and hidden mapping relationship between massive prospecting information and lithium-bearing pegmatite veins using the above existing methods. This is not enough to finely define prospecting target areas with lithium-bearing pegmatite veins with specific shapes, scales and spatial distributions.

[0032] To solve the above technical problems, the embodiment of this specification provides a pegmatite-type lithium ore prediction method based on deep learning. By performing spatial coordinate conversion on multiple different types of data in the area to be collected, a geospatial database can be generated. Then, by rasterizing various types of data in the geospatial database, a prediction element map, a mask map, and a prediction label map can be generated. Then, based on the prediction element map, mask map, and prediction label map of the target area, a training data set and a verification data set can be generated. Since the target area is the area where a pegmatite-type lithium ore body has been explored in the area to be collected, the prediction model obtained when the preset full convolutional neural network is trained using the training data set and the verification data set has better generalization ability, so that the prediction model can better learn the relationship between the spatial structure of multiple different types of data and the target, thereby being able to more comprehensively and objectively reflect the complex mapping relationship between the comprehensive information and the target. In this way, based on the prediction model, the distribution of pegmatite-type lithium ore in the selected area to be collected can be quickly and effectively ascertained, and a group of lithium-bearing pegmatite veins with a certain shape, scale, and spatial distribution can be identified.

[0033] In order to enable those skilled in the art to better understand the inventive concept, working principle and advantages of the embodiments of this specification, the pegmatite-type lithium ore prediction method based on deep learning in the embodiments of this specification is described in detail below.

[0034] Refer to Figure 1 the flowchart of a spodumene-type lithium ore prediction method based on deep learning in the embodiments of this specification shown in Figure 1 shown in the figure. The following steps can be executed:

[0035] S11. Obtain multiple different types of data for the area to be collected, and convert the spatial coordinates of various types of data to the same spatial coordinate system to generate a geoscience spatial database.

[0036] Among them, based on the geological metallogenic law and combined with the existing regional geological data, the area to be explored that is favorable for the target ore deposit type can be selected as the area to be collected, which can improve the possibility of discovering spodumene-type lithium ore.

[0037] In some examples, the screening criteria for the area to be collected include: 1) The area to be collected is within the range of a known large metallogenic belt or ore concentration area; 2) There is a known spodumene-type lithium ore deposit nearby.

[0038] In this embodiment, the multiple different types of data refer to the parameters used to characterize different parameter information of the area to be collected. By obtaining multiple different types of data, the topography of the area to be collected can be more truly reflected.

[0039] In this embodiment, by converting the spatial coordinates of various types of data to the same spatial coordinate system, the format consistency of various types of data can be improved, which is beneficial to reducing the difficulty of subsequent data processing in the geoscience spatial database.

[0040] S12. Perform rasterization processing on various types of data in the geoscience spatial database to generate a prediction feature map, a mask map, and a prediction label map.

[0041] In this embodiment, by performing rasterization processing on various types of data in the geoscience spatial database, the information of various types of data can be retained to the greatest extent, thereby improving the accuracy of the obtained prediction feature map, mask map, and prediction label map.

[0042] In this embodiment, the prediction feature map refers to a graphical expression form generated through rasterization and synthesis processing of various data such as geology, geophysics, geochemistry, and remote sensing, which can show the distribution law and prediction results of lithium-bearing pegmatite veins from different aspects.

[0043] The mask map is usually a two-dimensional graph generated by rasterizing the areas representing the no-go areas for prospecting work, areas without metallogenic potential, or areas with a large amount of missing data, and is used to represent the areas within the area to be collected that do not participate in the training and prediction of the spodumene-type lithium ore model.

[0044] The predicted label map is usually a two-dimensional graph generated by rasterizing the identified lithium-bearing granite pegmatite veins in the area to be collected. On the predicted label map, a certain color is used to represent the distribution of the identified lithium-bearing pegmatite veins.

[0045] S13. Generate a training data set and a validation data set based on the prediction factor map, the mask map, and the predicted label map of the target area.

[0046] Wherein, the target area is the area in the area to be collected where pegmatite-type lithium ore bodies have been explored.

[0047] In this embodiment, through steps S11 and S12, the prediction factor map, the mask map, and the predicted label map of the entire area to be collected can be obtained. By selecting the prediction factor map, the mask map, and the predicted label map of the target area, and the target area is the area in the area to be collected where pegmatite-type lithium ore bodies have been explored, this can ensure the accuracy of the subsequent formed training data set and validation data set.

[0048] S14. Use the training data set and the validation data set to train a preset fully convolutional neural network to generate a prediction model.

[0049] In this embodiment, the training data set and the validation data set are obtained from the area where pegmatite-type lithium ore bodies have been explored. Therefore, by using the training data set and the validation data set to train a preset fully convolutional neural network, the prediction model can have better generalization ability, and the prediction model can better learn the relationship between the spatial structure of multiple different types of data and the target, so as to more comprehensively and objectively reflect the complex mapping relationship between the comprehensive information and the target.

[0050] S15. Based on the prediction model, identify the distribution of pegmatite-type lithium ore in the selected area to be collected.

[0051] In this embodiment, through step S14, the prediction model can learn the relationship between the spatial structure of multiple different types of data and the target based on the obtained training data set and validation data set. Therefore, based on the prediction model, the distribution of pegmatite-type lithium ore in the selected area to be collected can be quickly and effectively identified, and then a batch of lithium-bearing pegmatite veins with certain morphology, scale, and spatial distribution can be identified.

[0052] In other words, the pegmatite-type lithium ore prediction method based on deep learning provided by the present invention replaces the study of traditional causal relationships and instead studies the correlation relationships (often including causal relationships but not limited to causal relationships) between multiple different types of data (such as comprehensive information on geology, geophysical exploration, geochemical exploration, and remote sensing, etc.) in the area to be collected and the targets (background and lithium-bearing pegmatite veins); moreover, it not only studies the association relationships between the attribute characteristics of the comprehensive information and the targets, but also studies the relationships between the spatial structures of the comprehensive information and the targets, which conforms to the characteristics of big earth science data, thus more comprehensively and objectively reflecting the complex mapping relationships between the comprehensive information and the targets.

[0053] In this embodiment, the multiple different types of data in the area to be collected may include: geological data, mineral data, geophysical exploration data, geochemical exploration data, remote sensing data, and mask data, where:

[0054] Geological data mainly refers to the regional geological map. Generally, for the convenience of similar analogy, the collection of regional geological maps in the work area should adhere to the principle of scale equivalence, that is, the scales of the regional geological maps in the area to be collected need to be consistent, and it is best to select materials with a larger scale.

[0055] For example, if the regional geological map with a scale of 1:200,000 in the area to be collected has achieved full coverage, then all materials with a scale of 1:200,000 are used, and the regional geological map with a scale of 1:250,000 cannot be included; also, for example, if the regional geological maps with scales of 1:200,000 and 1:50,000 have both achieved full coverage in the study area, then the regional geological map with a larger scale of 1:50,000 needs to be used to ensure a higher-precision prediction result.

[0056] Among them, the regional geological map is mainly stored in vector files, and some old materials exist in paper media and need to be vectorized first.

[0057] Geophysical exploration data mainly involves the regional gravity, magnetic method, and radioactive measurement data obtained from geophysical exploration surveys. Generally, for the convenience of similar analogy, the collection of regional geophysical exploration data in the area to be collected should adhere to the principle of scale equivalence, that is, the scales of the regional geophysical exploration data in the area to be collected need to be consistent, and it is best to select data with a larger scale.

[0058] For example: if the regional geophysical exploration data with a scale of 1:200,000 in the area to be collected has achieved full coverage, then all data with a scale of 1:200,000 are used, and the regional geophysical exploration data with a scale of 1:250,000 cannot be included; also, for example, if the regional geophysical exploration data with scales of 1:200,000 and 1:50,000 have both achieved full coverage in the study area, then the regional geophysical exploration data with a larger scale of 1:50,000 needs to be used to ensure a higher-precision prediction result.

[0059] Among them, the geophysical exploration data in the area to be collected are mainly stored in tabular files or vector files, and some old data exist in paper media and need to be vectorized first.

[0060] The geochemical exploration data mainly involve the regional stream sediment, rock debris and soil survey data obtained from geochemical prospecting surveys. The media vary for different scales and different geochemical landscapes. The scales are mainly two categories: 1:200,000 and 1:50,000.

[0061] Among them, the geochemical exploration survey data of 1:200,000 include 39 elements or oxides such as SiO2, Al2O3, K2O, Na2O, CaO, MgO, Fe2O3, Ag, As, Au, B, Ba, Be, Bi, Cd, Co, Cr, Cu, F, Hg, La, Li, Mn, Mo, Nb, Ni, P, Pb, Sb, Sn, Sr, Th, Ti, U, V, W, Y, Zn, Zr.

[0062] The required measurement data in the geochemical exploration survey data of 1:50,000 are 16 elements including Au, Ag, As, Bi, Cd, Cr, Co, Cu, Hg, Mo, Ni, Pb, Sb, Sn, W, Zn, and other elements or oxides can be added as appropriate according to the actual situation.

[0063] Generally, for the convenience of similar analogy, the collection of regional geochemical exploration data in the area to be collected should adhere to the principle of scale equivalence, that is, the scales of the regional geochemical exploration data in the area to be collected need to be consistent, and it is best to select data with a larger scale.

[0064] For example: If the regional geochemical exploration data with a scale of 1:200,000 in the work area is fully covered, then all the data of 1:200,000 are used, and the regional geochemical exploration data with a scale of 1:250,000 cannot be included; also, for example, if the regional geochemical exploration data with scales of 1:200,000 and 1:50,000 are both fully covered in the study area, then the regional geochemical exploration data with a larger scale of 1:50,000 needs to be used to ensure a higher-precision prediction result.

[0065] In addition, although there are geochemical exploration data with a larger scale, if the content of important target elements has not been measured (such as the Li element for finding pegmatite-type lithium ore, and some 1:50,000 regional geochemical exploration work has not measured it), then the geochemical exploration data of 1:200,000 will also be used.

[0066] Among them, the geochemical exploration data are mainly stored in tabular files or vector files, and some old data exist in paper media and need to be vectorized first.

[0067] There is a wide variety of remote sensing data. When it comes to searching for pegmatite-type lithium deposits, the remote sensing data should meet two requirements: high spatial resolution and high spectral resolution.

[0068] In this embodiment, since the scale of the lithium-bearing pegmatite veins is not large, generally a few meters to dozens of meters wide and dozens to hundreds of meters long, it is necessary to select remote sensing images with relatively high spatial resolution. Secondly, the remote sensing images should contain bands that can reflect the spectral characteristics of lithium in order to effectively extract relevant information about lithium. Existing research has pointed out that lithium (lithium-bearing pegmatite veins) has multiple absorption valleys and reflection peaks in the short-wave infrared (SWIR). This spectral characteristic can be used to select remote sensing images containing short-wave infrared. In terms of high spatial resolution, WorldView3, Jilin-1 or Gaofen-2 (GF-2) images are an ideal choice due to their sub-meter high spatial resolution. In terms of high spectral resolution, Ziyuan-1 02D satellite (ZY1E) or Gaofen-5 (GF-5) satellite images are also an ideal choice because they contain bands corresponding to multiple spectral characteristics of lithium (lithium-bearing pegmatite veins) in the short-wave infrared range.

[0069] Currently, regional remote sensing data is mainly stored in raster data files.

[0070] The main purpose of the mask data is to separate, according to the actual situation, the prospecting restricted areas, areas without mineralization potential or areas with a large amount of missing data (for pegmatite-type lithium deposits, the three zones and three lines, basins, lakes, snow cover, etc. within the working area) in the study area, so that these areas do not participate in training the model and carrying out predictions, further limiting the working scope.

[0071] Among them, the collected and produced mask data is often stored in vector file format.

[0072] For pegmatite-type lithium deposits, the mineral data mainly refers to collecting ore body distribution data for the area to be collected, mainly manifested as lithium-bearing pegmatite veins with a certain shape, scale and spatial distribution.

[0073] Among them, the mineral data is often stored in vector file format.

[0074] It should be noted that the multiple different types of data in the above examples for the collection area are only for illustrative purposes to represent obtaining multiple data and should not be construed as a limitation of the present invention.

[0075] In this embodiment, due to the different types, ages and sources of the collected data, there are also significant differences in the geographic and projection coordinate systems. Among them, the projection coordinate systems include Gauss-Kruger (3-degree zone or 6-degree zone), UTM and Lambert projection, and the common geographic coordinate systems corresponding to the projection coordinate systems include Beijing 54, Xi'an 80, CGCS2000 and WGS84.

[0076] For data analysis within a unified spatial coordinate system, the geographic coordinate system will uniformly adopt the current CGCS2000 geographic coordinate system in China, while the projection coordinate system will depend on the specific situation. In the case of a relatively small area, the Gauss-Kruger (3-degree zone or 6-degree zone) projection is mainly used. If the area is very large, the Lambert projection is used. The conversion of different spatial coordinate systems to a unified coordinate system can be carried out by using the three-parameter or seven-parameter "projection transformation" method within the ArcGIS software.

[0077] In this embodiment, after generating the geospatial database, the data in the geospatial database can be processed to form a prediction feature map, a mask map, and a prediction label map.

[0078] In a specific example, referring to Figure 2 the flowchart of a rasterization process in the embodiment of the present disclosure shown in Figure 2 as shown, the following steps can be executed:

[0079] S21, perform rasterization processing on various types of data in the geospatial database according to a preset first grid, form raster data corresponding to various types of data, and perform data synthesis on the multiple raster data to form the prediction feature map.

[0080] Specifically, different types of data have different formats, and when performing rasterization processing, the processing methods are not exactly the same. By performing rasterization processing on various types of data, corresponding raster data can be obtained, and by synthesizing these raster data, a prediction feature map can be obtained. Optionally, various types of data in the geospatial database include, but are not limited to, geological, geophysical, geochemical, and remote sensing data.

[0081] Moreover, during the rasterization process, if the grid is too large, the data will be thinned out and useful information will be lost; if the grid is too small, the computational amount will be greatly increased, affecting the operation efficiency. Generally, in order not to lose data information, the grid size of the data with the highest precision is often used as the preset grid size.

[0082] For example, when geological, geophysical, geochemical, and remote sensing data can be used simultaneously, the grid size of the remote sensing data with the largest scale (highest precision) is often used as the common grid for rasterizing all data, so as to basically ensure that the data information will not be lost.

[0083] In addition, the size of the ore body scale will also affect the size of the preset grid. If the ore body scale is relatively large, the grid size can also be appropriately increased to reduce unnecessary computational amount.

[0084] In this embodiment, when the data is the geological data, and when it is determined that the geological data is the strata and rock masses, the vectorized strata and rock masses are rasterized respectively by using the surface-to-raster method, and when it is determined that the geological data is a fold structure or a fault structure, the fold structure or the fault structure is converted into a polygon, and then the surface-to-raster method is used to rasterize the vectorized fold structure or fault structure to form the raster data corresponding to the geological data.

[0085] Specifically, the geological data mainly deals with strata, rock masses and structures. The processing methods can be roughly divided into three categories. The first category is the processing of strata and rock masses, the second category is the processing of fold structures, and the third category is the processing of fault structures.

[0086] Specifically, 1) For strata and rock masses, since they themselves appear as polygons on the geological map, the "surface-to-raster" method (as Figure 3 shown) can be directly used to rasterize the vectorized strata and rock masses (for example, in the *.tiff format); 2) For folds (synclines and anticlines), the fold structures can be delineated on the geological map according to the distribution of strata (synclines often show new in the middle and old on both sides; anticlines often show old in the middle and new on both sides) to form polygons, and similarly, the "surface-to-raster" method can be directly used to rasterize the vectorized folds (for example, in the *.tiff format); 3) For fault structures (normal faults, reverse faults and strike-slip faults), the "multi-ring buffer" method (as Figure 4 shown) can be used according to their ore-forming action range to form polygons of the fault structures, and similarly, the "surface-to-raster" method can be directly used to rasterize the vectorized fault structures (for example, in the *.tiff format).

[0087] In some embodiments, after the geological data is rasterized, in order to unify the dimension, improve the prediction effect and speed up the calculation speed, the "standardization" method is used to process the grid attributes, and the standardization process is shown in the following formula (1).

[0088]

[0089] Among them, x i is the i-th value, x mean is the average value, σ is the standard deviation, ξ is a fixed value, and x std is the value after standardization.

[0090] When the data is geophysical exploration data and the geophysical exploration data is a tabular file, the data in the tabular file is standardized and then interpolation is performed using the inverse distance weighting method; and when the geophysical exploration data is a vectorized isosurface map, the vectorized geophysical exploration map is converted into a rasterized geophysical exploration map using the surface-to-raster method, and then the standardized processing is used to form the raster data corresponding to the geophysical exploration data.

[0091] Specifically, when the geophysical exploration data is for a tabular file, first the tabular data is processed using the "standardization" method, and then the data is interpolated (for a preset first grid) using the "inverse distance weighting" method (Formula 2) to form the gridded data of geophysical exploration (for example, in the format of *.tiff); for the vectorized isosurface map of geophysical exploration, first, the vectorized geophysical exploration map is converted into a rasterized geophysical exploration map (for example, in the format of *.tiff) using the "surface-to-raster" method, and then the "standardization" method (see Formula 1) is used for processing.

[0092]

[0093] In the formula, v(x, y) is the estimated value of the interpolation point at point (x, y), v i is the true value of the i-th known point, d i is the distance between the i-th known point and the interpolation point, and p is the exponential parameter of the distance.

[0094] When the data is geochemical exploration data and the geochemical exploration data is a tabular file, the data in the tabular file is standardized and then interpolation is performed using the inverse distance weighting method; and when the geochemical exploration data is a vectorized isosurface map, the vectorized geochemical exploration map is converted into a rasterized geochemical exploration map using the surface-to-raster method, and then the standardized processing (see Formula 1) is used to form the raster data corresponding to the geochemical exploration data.

[0095] When the data is remote sensing data, calibration operations (such as radiometric calibration, FLAASH atmospheric correction, orthorectification) are performed on the original image of the remote sensing data, and then the standardized method is used to form the raster data corresponding to the remote sensing data.

[0096] Thus, according to different data types, different rasterization methods can be used to form their respective corresponding raster data.

[0097] In this embodiment, after forming multiple raster data, the multiple raster data can be synthesized to form a prediction factor map.

[0098] Specifically, band synthesis and common grid-based synthesis are used to fuse multiple raster data to form a multi-element prediction feature map. Band synthesis essentially creates a single raster dataset based on multiple bands. This patent includes each raster dataset as a separate band in a raster dataset to form a multi-band raster dataset, i.e., a multi-element prediction feature map.

[0099] S22 , performing rasterization processing on the mask space distribution map of the area to be collected according to a preset second grid to form the mask map.

[0100] Specifically, the method includes: converting the mask space distribution map into a mask map by using a surface-to-grid conversion method according to the second grid.

[0101] In this embodiment, the format of the mask image may be *.tiff format.

[0102] S23 , selecting a target area from the area to be collected, and rasterizing the ore body spatial distribution map of the target area according to a preset third grid to form a prediction label map.

[0103] In this embodiment, the target area serves as the data collection area for training data and verification data, and should meet at least the following conditions: 1) The exploration level of the target area should be high, and the occurrence, morphology, scale, spatial distribution, mineral composition, grade, resource reserves, and associated components of the ore bodies in the area should have been basically ascertained, thereby ensuring the accuracy of the training and verification samples; 2) The target area should be representative and diverse, and can be used as a reference for similar analogies in the prospecting prediction process, so that the trained deep learning model has a high generalization ability, so that when prospecting prediction is carried out in unknown areas, the identified prospecting target area has a high degree of credibility; 3) When there are few known areas in the target area, the pegmatite-type lithium ore distribution area with a high degree of exploration outside the working area can be fully utilized as the training and verification data set collection area through transfer learning.

[0104] In this case, the generation process of the prediction label map may include: according to the third grid, using a surface-to-grid method to rasterize the ore body spatial distribution map of the target area to form the prediction label map.

[0105] In this embodiment, the format of the predicted label image may be *.tiff format.

[0106] In this embodiment, the first grid, the second grid, and the third grid are of the same size.

[0107] After determining the predicted feature map, mask map, and predicted label map of the target area, a training dataset and a validation dataset can be generated.

[0108] In an optional example, refer to Figure 5 the flowchart of a method for generating a training dataset and a validation dataset in the embodiments of the present specification, as Figure 5 shown, specifically including:

[0109] S51. Using the sliding window technique, according to the set window size and step size, crop and expand the prediction feature map, mask map, and prediction label map of the target area to generate an expanded data set.

[0110] Specifically, the basic idea of using the comprehensive information prospecting prediction method for pegmatite-type lithium deposits based on deep learning to delineate the prospecting target area is still similar analogy, and it requires the support of a larger-capacity training sample. In actual situations, there are not many identified lithium-bearing pegmatite veins in the data collection area, so the data set is small.

[0111] In this embodiment, in order to expand the samples to meet the needs of deep learning, the sliding window (SlidingWindow) method is used to crop and expand the prediction feature map, mask map, and prediction label map of the data collection area with a consistent window size and step size to generate a data set for deep learning model training and validation.

[0112] In other words, it is to generate a corresponding set of sub-prediction feature maps, sub-mask maps, and sub-prediction label maps.

[0113] In this embodiment, the method for generating the expanded data set is as follows:

[0114]

[0115] where H is the height of the prediction feature map of the collection area, W is the width of the prediction feature map of the collection area, h is the height of the sliding window, w is the width of the sliding window, s h is the step size in the height direction, s w is the step size in the width direction, and N is the number of generated samples.

[0116] It should be noted that when using the sliding window technique, it is advisable that the sliding window can relatively completely cover the lithium-bearing pegmatite vein, and the step size is usually smaller than the window size to increase the sample size.

[0117] S52. Randomly shuffle the data in the expanded data set, and according to a preset ratio, take a part of the data as the training data set and the remaining part of the data as the validation data set.

[0118] Specifically, by randomly shuffling the data in the expanded data set, it is possible to ensure that the relationship between the samples in the expanded data set is random and there is no logic in the order, thereby preventing similar samples from being too concentrated in the training set or the validation set, resulting in a low generalization ability of the trained deep learning model. Then, the data set with the randomly shuffled order is divided into the corresponding training set and validation set folders according to a certain ratio, such as 7:3 or 8:2, to form a training set and a validation set for subsequent model training and validation work.

[0119] After obtaining the training set and the validation set that meet the training requirements, the preset fully convolutional neural network can be trained to generate a prediction model.

[0120] See Figure 6 The schematic diagram of the generation process of a prediction model in the embodiment of this specification shown in Figure 6 As shown, the following generation steps can be executed:

[0121] S61, determine the loss function adapted to the fully convolutional neural network.

[0122] Specifically, the loss function is an indicator representing the "badness" of the deep learning model, which is used to measure the effect of the weight parameter w on the training set. When the loss function value converges to a suitable accuracy, it indicates that the model reaches the optimal state in the training stage.

[0123] In this embodiment, in the prospecting prediction of pegmatite-type lithium deposits, whether it is training samples or validation samples, the number of backgrounds is often much larger than the number of lithium-bearing pegmatite veins. Therefore, this model uses an improved weighted cross-entropy function, that is, giving a smaller weight to the background and a larger weight to the lithium-bearing pegmatite veins. Among them, the so-called background refers to the distribution areas in the training samples where there are no lithium-bearing pegmatite veins, and these areas can distribute any non-lithium-bearing pegmatite geological bodies, such as sedimentary rocks, metamorphic rocks, etc.

[0124] In this embodiment, the loss function is:

[0125]

[0126] Among them, m is the number of training samples, m0 represents the number of training samples of the background, m1 represents the number of training samples of the lithium-bearing pegmatite veins, and m = (m0) + (m1). The weight w of the background m0 = m1 / m, the weight w of the lithium-bearing pegmatite veins m1 = m0 / m; y gt i is the true value of the i-th training sample, y pre i is the predicted value of the i-th training sample, L(y gti , y pre i is the loss function of training sample i.

[0127] S62. Determine an optimization function adapted to the fully convolutional neural network.

[0128] Specifically, the optimization function is used to optimize the weight parameters. Among them, Adam combines the advantages of the Momentum and RMSprop methods, realizes efficient search in the parameter space, and thus is widely used. This deep learning model optimization adopts this method, and the optimization function is:

[0129]

[0130] where w is the weight parameter, := means that the value on its left is updated or replaced by the value on its right, α is the learning rate, β1 is the first momentum, β2 is the second momentum, dw is the abbreviation of the derivative of the cost function with respect to the weight parameter, v dw is the exponentially weighted average of the momentum of dw, s dw is the exponentially weighted average of the momentum of dw 2 ε is a constant value, t represents the number of iterations;

[0131] S63. Perform the initialization operation of the weight parameters in the fully convolutional neural network.

[0132] Among them, the initialization of the weight parameters in the fully convolutional neural network is to prevent the so-called "symmetry" problem, that is, to prevent all convolutional layers in the same hidden layer from having the same function, and it is not allowed to initialize all weight parameters to 0. Therefore, random initialization is required. Since this deep learning method uses PReLU as the activation function, in order to avoid the problem of gradient disappearance caused by the deepening of the number of layers, the He random initialization method is adopted, which can ensure the effective flow of information in the forward propagation and backward propagation processes, and make the variances of the input signals of different layers roughly equal.

[0133] Specifically, by adjusting the initial value of the weight, it is ensured that in each layer of the network, the variance of the input signal remains consistent, which helps to improve the training efficiency and performance of the network. The He random initialization makes the weight W satisfy the normal (Gaussian) distribution with a mean of 0 and a variance of 2 / n, that is, the formula:

[0134]

[0135] where n is the number of weight parameters of each convolutional layer, and η is the learnable parameter in the PReLU activation function.

[0136] S64. Input the training data set into the fully convolutional neural network in batches, and use the forward propagation method to obtain the corresponding loss function value, and use the gradient descent method (Adam) for backpropagation to update the weight parameters in the fully convolutional neural network.

[0137] Specifically, perform step S64. By inputting the training sample set into the model in batches, through forward propagation, calculate the prediction result F NC =2 loss function, and then perform backpropagation and update the weight parameters by using the Adam gradient descent method.

[0138] S65. Input the validation data set into the trained fully convolutional neural network, and use the forward propagation method to evaluate the quality of the formed prediction model according to the pre-selected validation evaluation metrics.

[0139] Specifically, perform step S65. By inputting the entire validation sample set into the model, through forward propagation, it is possible to calculate the prediction result F NC =2 validation evaluation metrics (such as GA, MA, MA of class i, and MIU) to evaluate the quality of the deep learning model.

[0140] In this embodiment, the advantages and disadvantages of the deep learning model are mainly based on various evaluation metrics of the confusion matrix, among which the global accuracy GA, the average accuracy MA, and the single intersection over union IU i and the mean intersection over union MIU are used as evaluation metrics.

[0141] Among them:

[0142]

[0143]

[0144] where n ii is the number of pixels in which class i is predicted as class i; n ij is the number of pixels in which class i is predicted as class j; n ji is the number of pixels in which class j is predicted as class i; n cls is the number of target classes; GA is the global accuracy, MA is the average accuracy; IU i is the intersection over union of class i; MIU is the mean intersection over union.

[0145] S66. Repeat the training and validation steps iteratively until the loss function value and the validation evaluation metrics meet the requirements.

[0146] For example, repeat the above steps iteratively multiple times until the cost function and the validation evaluation metrics reach the required accuracy, mainly manifested as the cost function value no longer decreases significantly and the validation evaluation metrics no longer increase significantly, that is, the optimal deep learning model is obtained.

[0147] Thus, by adopting the above training method, on the one hand, the sliding window method is used to make and expand the training and validation data sets, laying a sample foundation for the deep learning model based on big data; on the other hand, the He method is used for random weight initialization, and by adjusting the initial value of the weight, it is ensured that the variance of the input signal remains consistent in each layer of the network, improving the training efficiency and performance of the network; in addition, considering that the training samples of lithium-bearing pegmatite veins are much smaller than the background training samples, the loss function is improved, a smaller weight is given to the background with a large number of samples, and a larger weight is given to the lithium-bearing pegmatite veins with a small number of samples, which alleviates the problem of sample imbalance to a certain extent, thereby improving the generalization ability of the prediction model.

[0148] In this embodiment, the fully convolutional neural network is obtained in the following manner:

[0149] A1) Based on the spatial information and prospecting information of the prediction feature map, determine the network main architecture, where, as Figure 7 shown, the network main architecture has an encoder-decoder main architecture with a symmetric structure.

[0150] In this way, while being able to fully extract the deep prospecting information of the prediction feature map, it can also make full use of the spatial information of the prediction feature map, so as to identify lithium-bearing pegmatite veins and the background at the pixel scale.

[0151] A2) Set the encoding structure of the network main architecture, and the encoding structure includes: type-I convolutional layer and downsampling structure; and the execution logic of the encoding structure is: perform convolution calculation through the type-I convolutional layer, and then perform batch normalization, type-I activation function, and attention mechanism module in sequence to obtain the first prediction feature map with different heights, widths, and numbers of features.

[0152] Specifically, referring to Figure 7 , the execution logic of the encoding structure is:

[0153] A21) For the prediction feature map with height H in , width W in , and number of features F in , perform convolution calculation through F b convolutional layers (type-I), and then perform batch normalization, activation function (type-I), and attention mechanism module in sequence to obtain a prediction feature map with height H in , width W in , and number of features F bThe first prediction element map.

[0154] A22) Repeat step A21) until a first prediction element map with a height of H in , a width of W in , and a feature count of F b is obtained.

[0155] A23) Downsample the first prediction element map with a height of H in , a width of W in , and a feature count of F b to obtain a first prediction element map with a height of H in / 2, a width of W in / 2, and a feature count of F b .

[0156] A24) Convolve the first prediction element map with a height of H in / 2, a width of W in / 2, and a feature count of F b using 2F b convolution layers (type I), followed by batch normalization, an activation function (type I), and an attention mechanism module, to obtain a first prediction element map with a height of H in / 2, a width of W in / 2, and a feature count of 2F b .

[0157] A25) Repeat step A24) until a first prediction element map with a height of H in / 2, a width of W in / 2, and a feature count of 2F b is obtained.

[0158] A26) Downsample the first prediction element map with a height of H in / 2, a width of W in / 2, and a feature count of 2F b to obtain a first prediction element map with a height of H in / 4, a width of W in / 4, and a feature count of 2F b .

[0159] A27) Convolve the first prediction element map with a height of H in / 4, a width of W in / 4, and a feature count of 2F b using 4F b convolution layers (type I), followed by batch normalization, an activation function (type I), and an attention mechanism module, to obtain a first prediction element map with a height of H in / 4, a width of W in / 4, the number of features is 4F b of the first prediction element map.

[0160] A28) Repeat step A27), still obtaining a height of H in / 4, a width of W in / 4, the number of features is 4F b of the first prediction element map.

[0161] A29) The prediction element map with a height of H in / 4, a width of W in / 4, the number of features is 4F b The prediction element map is downsampled to obtain a height of H in / 8, a width of W in / 8, the number of features is 4F b of the first prediction element map.

[0162] A30) The prediction element map with a height of H in / 8, a width of W in / 8, the number of features is 4F b The prediction element map, through convolution calculation using 8F b convolutional layers (type I), and then successively performing batch normalization, activation function (type I), and attention mechanism module, to obtain a height of H in / 8, a width of W in / 8, the number of features is 8F b of the first prediction element map.

[0163] A31) Repeat step A30), still obtaining a height of H in / 8, a width of W in / 8, the number of features is 8F b of the first prediction element map.

[0164] A32) The prediction element map with a height of H in / 8, a width of W in / 8, the number of features is 8F b The prediction element map is downsampled to obtain a height of H in / 16, a width of W in / 16, the number of features is 8F b of the first prediction element map.

[0165] A33) The prediction element map with a height of H in / 16, a width of W in / 16, the number of features is 8F b The prediction element map, through convolution calculation using 8F b convolutional layers (type I), and then successively performing batch normalization, activation function (type I), and attention mechanism module, to obtain a height of Hin / 16, with a width of W in / 16, with 8F feature numbers b of the first prediction element map.

[0166] A34): Repeat step A33), still obtaining a height of H in / 16, with a width of W in / 16, with 8F feature numbers b of the first prediction element map.

[0167] A4) Set the decoding structure of the network main architecture. The decoding structure includes: upsampling, skip connection, and convolutional layer, where the convolutional layer includes: type I convolutional layer and type II convolutional layer; and the execution logic of the decoding structure is: obtain the second prediction element map through upsampling, and integrate it with the corresponding first prediction element map in the encoding structure through skip connection to obtain the third prediction element map; then, perform convolutional calculation through type I and type II convolutional layers, as well as batch normalization, type I activation function, attention mechanism module, type II activation function, and maximum likelihood conversion to obtain the prediction result map.

[0168] Specifically, refer to Figure 7 , and the execution logic of the decoding structure is:

[0169] A41) Upsample the first prediction element map with a height of H in / 16, with a width of W in / 16, with 8F feature numbers b to obtain a second prediction element map with a height of H in / 8, with a width of W in / 8, with 8F feature numbers b of the second prediction element map.

[0170] A42) Integrate the second prediction element map with a height of H in / 8, with a width of W in / 8, with 8F feature numbers b obtained by upsampling with the corresponding first prediction element map with a height of H in / 8, with a width of W in / 8, with 8F feature numbers b through skip connection to obtain a third prediction element map with a height of H in / 8, with a width of W in / 8, with 16F feature numbers b of the third prediction element map.

[0171] A43) For the above-mentioned height of H in / 8, with a width of W in / 8, with 16F feature numbers bThe third prediction factor map is obtained by using 8F b convolutional layers (Type I) for convolution calculation, followed by batch normalization, activation function (Type I), and attention mechanism module, to obtain a third prediction factor map with a height of H in / 8, a width of W in / 8, and 8F feature numbers b .

[0172] A44) The third prediction factor map with a height of H in / 8, a width of W in / 8, and 8F feature numbers b is convolved using 4F b convolutional layers (Type I), followed by batch normalization, activation function (Type I), and attention mechanism module, to obtain a third prediction factor map with a height of H in / 8, a width of W in / 8, and 4F feature numbers b .

[0173] A45) The third prediction factor map with a height of H in / 8, a width of W in / 8, and 4F feature numbers b is upsampled to obtain a second prediction factor map with a height of H in / 4, a width of W in / 4, and 4F feature numbers b .

[0174] A46) The second prediction factor map with a height of H in / 4, a width of W in / 4, and 4F feature numbers b obtained by upsampling is integrated with the corresponding first prediction factor map in the encoding structure with a height of H in / 4, a width of W in / 4, and 4F feature numbers b through skip connection to obtain a third prediction factor map with a height of H in / 4, a width of W in / 4, and 8F feature numbers b .

[0175] A47) The above-mentioned third prediction factor map with a height of H in / 4, a width of W in / 4, and 8F feature numbers b is convolved using 4F b convolutional layers (Type I), followed by batch normalization, activation function (Type I), and attention mechanism module, to obtain a height of Hin / 4, with a width of W in / 4, with the number of features being 4F b of the third prediction element map.

[0176] A48) The height in A47) is H in / 4, with a width of W in / 4, with the number of features being 4F b The prediction element map is convolved using 2F b convolution layers (type I), followed by batch normalization, activation function (type I), and the attention mechanism module, to obtain a height of H in / 4, with a width of W in / 4, with the number of features being 2F b of the third prediction element map.

[0177] A49) The height in A48) is H in / 4, with a width of W in / 4, with the number of features being 2F b The third prediction element map is upsampled to obtain a height of H in / 2, with a width of W in / 2, with the number of features being 2F[[ID=3,6]] b of the second prediction element map.

[0178] A50) The upsampled height is H in / 2, with a width of W in / 2, with the number of features being 2F b The second prediction element map is integrated with the corresponding height in the encoding structure, which is H in / 2, with a width of W in / 2, with the number of features being 2F b The first prediction element map is integrated through skip connection to obtain a height of H in / 2, with a width of W in / 2, with the number of features being 4F b of the third prediction element map.

[0179] A51) The height in A50) is H in / 2, with a width of W in / 2, with the number of features being 4F b The third prediction element map is convolved using 2F b convolution layers (type I), followed by batch normalization, activation function (type I), and the attention mechanism module, to obtain a height of H in / 2, with a width of W in / 2, with the number of features being 2F b of the third prediction element map.

[0180] A52) The third prediction feature map with a height of H in / 2, a width of W in / 2, and 2F feature numbers b is convolved through the use of F b convolution layers (Type I), followed by batch normalization, activation function (Type I), and an attention mechanism module, to obtain a third prediction feature map with a height of H in / 2, a width of W in / 2, and F feature numbers b .

[0181] A53) The third prediction feature map with a height of H in / 2, a width of W in / 2, and F feature numbers b is upsampled to obtain a second prediction feature map with a height of H in , a width of W in , and F feature numbers b .

[0182] A54) The second prediction feature map with a height of H in , a width of W in , and F feature numbers b obtained by upsampling is integrated with the corresponding first prediction feature map in the encoding structure with a height of H in , a width of W in , and F feature numbers b through skip connection to obtain a third prediction feature map with a height of H in , a width of W in , and 2F feature numbers b .

[0183] A55) The third prediction feature map with a height of H in , a width of W in , and 2F feature numbers b is convolved through the use of F b convolution layers (Type I), followed by batch normalization, activation function (Type I), and an attention mechanism module, to obtain a third prediction feature map with a height of H in , a width of W in , and F feature numbers b .

[0184] A56) The third prediction feature map with a height of H in , a width of W in , and F feature numbers b is convolved through the use of F bA convolutional layer (Type I) performs convolutional calculations, followed by batch normalization, an activation function (Type I), and an attention mechanism module to obtain a third prediction feature map with a height of H in and a width of W in and a feature count of F b .

[0185] A57) The third prediction feature map with a height of H in , a width of W in , and a feature count of F b in A56) is subjected to convolutional calculations using F NC = 2 convolutional layers (Type II) to obtain a prediction map with a height of H in , a width of W in , and a classification count of F NC = 2 classification prediction maps.

[0186] A58) The above classification prediction map with a height of H in , a width of W in , and a feature count of F NC = 2 is transformed through an activation function (Type II) and the maximum likelihood method to obtain a prediction result map F1.

[0187] For a better understanding and illustration of the working logic of the fully convolutional neural network in the embodiments of this specification, an example is used for illustration.

[0188] Set the size of the convolutional layer (Type I) and the convolutional process: In the entire network structure, the role of the convolutional layer (Type I) is mainly to extract deep prospecting information. The larger the convolutional layer, the larger its receptive field, but the greater the amount of information loss. Therefore, setting a reasonable size of the convolutional layer is crucial. After years of practice, it is found that a 3×3 convolutional layer is reasonable for the fully convolutional neural network. Therefore, this network uses a 3×3 convolutional layer.

[0189] Based on the relationship between the convolutional process and the sizes of the input and output prediction feature maps, as shown in Formulas 11 and 12, in order to ensure that the size of the output prediction feature map remains unchanged, i.e., H out = H in , W out = W in , when the stride of the convolutional process in the 3×3 convolutional layer is 1 and the dilation factor is 1, the padding should be 1.

[0190] It should be noted that the number of channels of the convolutional layer (Type I) is the same as the number of channels of the input prediction feature map;

[0191] H out = (H in - D×(K - 1)+2P - 1) / S - 1 (11)

[0192] W out = (W in - D×(K - 1)+2P - 1) / S - 1 (12)

[0193] Where H out is the height of the output prediction feature map, H in is the height of the input prediction feature map, W out refers to the width of the output prediction feature map, W in refers to the width of the input prediction feature map, D refers to the dilation coefficient, P refers to padding, K refers to the convolutional layer size, and S refers to the stride.

[0194] Next, batch normalization processing is performed. To make the search for hyperparameters simpler, less sensitive to the selection of hyperparameters, and make the fully convolutional neural network more robust and easier to train, we not only need to normalize the original prediction feature map, but also need to normalize the prediction feature maps of each intermediate layer, as shown in Equation 13. This is batch normalization processing, and the value after batch normalization is obtained.

[0195]

[0196] Among them, x i is the i-th value, x mean is the average value, σ is the standard deviation, ξ is a constant value, γ and β respectively represent scaling and translation of the normalized data as required, and these two parameters are learnable parameters. is the value after normalization.

[0197] Then, the activation function (Type I) is defined. Introducing the activation function (Type I) is to enable the fully convolutional neural network to have the ability to solve non-linear problems. Currently, the ReLU activation function is mostly used in the hidden layer. It can solve the problem of gradient disappearance caused by the saturated activation function and greatly accelerate the learning speed. However, the disadvantage of this function is that when there is a very large gradient during the backpropagation process, the backpropagation update may cause the center of the weight distribution to be less than 0, so that the derivative at that place is always 0, and the backpropagation cannot update the weights, that is, it enters the inactive state.

[0198] PReLU is an improved version of the ReLU activation function. The principle is shown in Equation 14. It can adaptively learn the parameter η from the data, not only can solve the above problems, but also has the characteristics of fast convergence speed and low error rate.

[0199] PReLU(x) = max(0, x)+ηmin(0, x) (14)

[0200] Among them, \(x\) is the value input to the activation function after the predicted feature map undergoes convolution operation and batch normalization, and \(\eta\) is the learnable parameter.

[0201] Next, the attention mechanism is set. The attention mechanism can be divided into the feature attention mechanism and the spatial attention mechanism. The feature attention mechanism mainly normalizes each feature of the input predicted feature map through the max pooling layer, fully connected layer, and Sigmoid activation function in sequence to form the weight of each feature, and then multiplies each weight with the corresponding feature of the input predicted feature map respectively, thereby assigning a certain weight to each feature to strengthen the role of different feature importances in target classification prediction; the spatial attention mechanism normalizes each spatial position of the input predicted feature map through the max pooling layer, convolution operation, and Sigmoid activation function to form the weight of each position, and then multiplies each weight with the corresponding position of the input predicted feature map respectively, thereby assigning a certain weight to each spatial position to strengthen the role of different spatial position importances in target classification prediction. For the prospecting prediction of pegmatite-type lithium deposits, the roles of different features and different positions in classification prediction are indeed different.

[0202] Therefore, this method incorporates the feature and spatial attention mechanisms into the network to enhance the prediction effect. The principles of the feature attention mechanism and the spatial attention mechanism are as Figure 8 shown.

[0203] Subsequently, the downsampling structure and its downsampling process are set: The purpose of downsampling is to sparsely process the predicted feature map, reduce the data computation volume, and increase the receptive field. Among them, the pooling layer is the most commonly used method, including the average pooling layer and the max pooling layer. In recent years, it has been found that the max pooling layer can retain important features better than the average pooling layer. Therefore, the downsampling structure of this network uses a \(2\times2\) max pooling layer, and the stride of the pooling process is 2.

[0204] Next, the upsampling structure is set: Upsampling is mainly used to gradually restore the deep features to the original image size. The upsampling methods include transposed convolution and bilinear interpolation. According to practical experience, using bilinear interpolation in the upsampling process has a similar effect to using transposed convolution. However, compared with transposed convolution, bilinear interpolation has no parameters and there is no need for repeated iteration. Therefore, the efficiency is much higher than that of transposed convolution. Therefore, this method uses bilinear interpolation to implement upsampling. The principle of using bilinear interpolation is as shown in Equation 15, Figure 9 shown.

[0205]

[0206] Among them, z(x, y) is the value to be evaluated at the unknown point (x, y), z(x1, y1) is the known value at the known point (x1, y1), z(x2, y1) is the known value at the known point (x2, y1), z(x1, y2) is the known value at the known point (x1, y2), and z(x2, y2) is the known value at the known point (x2, y2).

[0207] It should be noted that the unknown points and known points in the formula refer to the known points and unknown points of the predicted feature map input in any upsampling layer in the network structure.

[0208] Then, define the skip connection method. There are currently two skip connection methods. One is that the predicted feature map formed by upsampling in the decoding process and the predicted feature map in the corresponding encoding structure are added. Since the linear fusion of data has been completed, generally no convolution will be performed again, but directly upsampling; the other is that the predicted feature map formed by upsampling in the decoding process and the predicted feature map in the corresponding encoding structure are tiled, and then two convolutions are performed for non-linear fusion. In view of the fusion method of the tile + convolution combination, which can better perform cross-feature non-linear fusion of shallow features and deep features, so as to make full use of spatial information and semantic information, and the effect will be better. The skip connection in this time adopts the tiling method.

[0209] Secondly, set the size of the convolutional layer (Type II) and the convolution process: Using the convolutional layer (Type II) is the last convolution process in the entire convolutional neural network structure. Its role is to generate predicted values of different categories. To achieve this goal, the size of the convolutional layer is 1×1, and the stride is 1. It should be noted that the number of features of the convolutional layer (Type II) is the same as the number of features of the input predicted feature map, and the number of output features is equal to the number of classifications of the prediction target (F NC = 2).

[0210] Finally, set the activation function (Type II) and generate the predicted result map.

[0211] In order to solve the multi-classification problem, the output layer uses Softmax as the activation function (Type II) to achieve the classification target with probabilistic significance, and uses the Argmax function in the maximum likelihood method to generate the predicted result map. See Equation (16) for details.

[0212]

[0213] Among them, the Argmax function is used to return the maximum value of the objective function, Softmax is a multi-classification activation function (also applicable to binary classification problems), x i is the input value, y pre is the x in the input value that makes Softmax(x i ) obtain the maximum valuei value.

[0214] Therefore, a fully convolutional neural network with the above characteristics is adopted. On the one hand, the deep learning model with a symmetrical structure adopted not only extracts deep semantic information through convolution calculation and downsampling, but also focuses on combining shallow spatial information with deep semantic information through jump connections. It can not only determine whether there are lithium-bearing pegmatite veins in the area to be collected, but also describe their shape, scale and spatial distribution in detail, thereby ensuring the prediction accuracy of lithium-bearing pegmatite veins, and can finely define the prospecting target area, providing a detailed basis for further prospecting and exploration.

[0215] On the other hand, the comprehensive information prospecting and prediction method based on deep learning provided by the present invention improves key components of the deep learning model and optimizes the overall network structure. Using PReLU instead of ReLU activation function prevents network deactivation that can occur when a large gradient passes through the backpropagation process, accelerates network convergence, and reduces error rates. The introduction of feature and spatial attention mechanisms strengthens the importance of different features and spatial positions in identifying lithium-bearing pegmatite veins, improving prediction accuracy. Using bilinear interpolation instead of transposed convolution for upsampling greatly accelerates network efficiency due to the lack of parameters and the need for iteration, making deep learning networks more practical.

[0216] After the prediction model is formed, the distribution of pegmatite-type lithium deposits in the selected area to be mined can be ascertained. This specifically includes: dividing the area to be mined into sub-areas; inputting each sub-area into the prediction model to generate a prediction result map for each sub-area; splicing the prediction result maps to form a prospecting target area distribution map for the area to be mined; selecting a surface area containing pegmatite-type veins from the prospecting target area distribution map, and ascertaining the morphology, scale, spatial distribution, grade, co-existing elements, and resource quantity of the lithium pegmatite vein ore body based on the exposure, mineralization degree, and surface occurrence information of the pegmatite-type veins.

[0217] Specifically, since the area of the selected area to be collected is relatively large, it is usually difficult to process all areas within the area to be collected at one time. Therefore, a sliding window method can be used to divide the area to be collected into sub-areas. In order to ensure that each sub-area makes full use of the surrounding data and makes the edges of the sub-areas have a good prediction effect, a small amount of overlapping is often used.

[0218] Among them, such as Figure 10 As shown, the area to be collected QY is divided into various sub-areas (for example Figure 10In order to ensure that each sub-region makes full use of the surrounding data, a small amount of overlap (as shown in the figure YY) is often adopted, and the width of the overlap is at least twice the number of type I convolutions of the entire network.

[0219] In one embodiment, the number of convolutions of the convolution layer (type I) is 18, so the width of YY is at least 36.

[0220] Then, the trained optimal deep learning model is loaded to carry out mineral exploration prediction. Each sub-area is input into the model one by one, and forward propagation is performed to output the prediction result map F1 of each sub-area. The edge part is removed and then collage is performed to form a mineral exploration target area distribution map of the unknown area.

[0221] Then, within the identified pegmatite-type lithium mineral prospecting target area, large-scale geological mapping was carried out mainly using the tracing method to delineate the distribution of lithium-bearing pegmatite veins on the surface.

[0222] Finally, for the lithium-bearing pegmatite veins delineated on the surface, based on their exposure, degree of mineralization, surface occurrence and other information, a certain degree of trench exploration + drilling (including chemical sample analysis) is used to determine the shape, scale, spatial distribution, grade, co-existing elements and resource quantity of the lithium-bearing pegmatite vein ore body, so as to finally discover pegmatite-type lithium mines.

[0223] The embodiment of this specification also provides a device corresponding to the pegmatite-type lithium ore prediction method based on deep learning, such as Figure 11 A schematic diagram of the structure of a pegmatite-type lithium ore prediction device based on deep learning in the embodiment of this specification is shown in FIG. Figure 11 As shown, the pegmatite-type lithium ore prediction device 100 based on deep learning may include:

[0224] The data acquisition unit 110 acquires a plurality of different types of data from the area to be collected, and converts the spatial coordinates of the various types of data into the same spatial coordinate system to generate a geospatial database;

[0225] The data processing unit 120 performs rasterization processing on various types of data in the geospatial database to generate a prediction element map, a mask map, and a prediction label map;

[0226] The data generation unit 130 generates a training data set and a validation data set based on the prediction factor map, the mask map, and the prediction label map of the target area, wherein the target area is an area in the area to be sampled where a pegmatite-type lithium ore body has been discovered;

[0227] The processing unit 140 uses the training data set and the verification data set to train a preset fully convolutional neural network to generate a prediction model; and based on the prediction model, finds out the distribution of pegmatite-type lithium deposits in the selected area to be mined.

[0228] The specific working processes and principles of the data acquisition unit 110 , the data processing unit 120 , the data generation unit 130 and the processing unit 140 may refer to the relevant descriptions of the aforementioned examples.

[0229] It is understandable that the division of the above units is only a division of logical functions, and in actual implementation, they can be fully or partially integrated into one physical entity, or physically separated. In addition, the above units can be implemented in the form of a processor calling software.

[0230] Although the embodiments of this specification are disclosed above, the present invention is not limited thereto. Any person skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention should be based on the scope defined by the claims.

Claims

1. A prediction method for pegmatite-type lithium ore based on deep learning, characterized in that, include: Acquire multiple different types of data in the area to be collected, and convert the spatial coordinates of the various types of data into the same spatial coordinate system to generate a geospatial database, wherein the multiple different types of data include: geological data, geophysical data, geochemical data, remote sensing data, and mask data; Various types of data in the geospatial database are rasterized to generate a prediction element map, a mask map, and a prediction label map, including: rasterizing various types of data in the geospatial database according to a preset first grid to form raster data corresponding to various types of data, and synthesizing a plurality of the raster data to form the prediction element map; rasterizing the mask space distribution map of the area to be collected according to a preset second grid to form the mask map; selecting a target area from the area to be collected, and rasterizing the ore body space distribution map of the target area according to a preset third grid to form a prediction label map, wherein the first grid, the second grid, and the third grid are of the same size; wherein, when the geological data, the geophysical data, the geochemical data, and the remote sensing data are used simultaneously, during the rasterization process, the grid size of the remote sensing data with the largest scale is used as the size of the preset grid; Generate a training data set and a validation data set based on the predicted element map, the mask map, and the predicted label map of the target area, wherein the target area is an area in the area to be sampled where a pegmatite-type lithium ore body has been discovered; Using the training data set and the validation data set, a preset fully convolutional neural network is trained to generate a prediction model; Based on the prediction model, the distribution of pegmatite-type lithium deposits in the selected area to be mined is ascertained; Among them, the loss function of the fully convolutional neural network is: Among them, m is the number of training samples, m0 represents the number of training samples of the background, m1 represents the number of training samples of the lithium-bearing pegmatite vein, and m = m0 + m1. The weight w of the background m0 = m1 / m, and the weight w of the lithium-bearing pegmatite vein m1 = m0 / m; y gt i is the true value of the i-th training sample, and y pre i is the predicted value of the i-th training sample. L(y gt i , y pre i ) is the loss function of the training sample i.

2. The method for predicting pegmatite-type lithium deposits based on deep learning according to claim 1, characterized in that: The rasterization processing of various types of data in the geospatial database according to the preset first grid to form raster data corresponding to the various types of data includes: When the data is the geological data, and when the geological data is determined to be a stratum and a rock mass, a surface-to-raster method is used to perform rasterization processing on the vectorized stratum and the rock mass, and when the geological data is determined to be a fold structure or a fault structure, the fold structure or the fault structure is converted into a surface element, and then the surface-to-raster method is used to perform rasterization processing on the vectorized fold structure or fault structure to form raster data corresponding to the geological data; When the data is geophysical exploration data or geochemical exploration data, and the geophysical exploration data or the geochemical exploration data is a table file, the data in the table file is standardized, and then inverse distance weighting interpolation is performed; and when the geophysical exploration data or the geochemical exploration data is a vectorized isosurface map, the vectorized geophysical exploration map or geochemical exploration map is converted into a rasterized geochemical exploration map by the surface-to-raster method, and then the standardized processing is adopted to form the raster data corresponding to the geophysical exploration data or the geochemical exploration data; When the data is remote sensing data, the original image of the remote sensing data is calibrated, and then the standardized method is adopted to form the raster data corresponding to the remote sensing data; The data synthesis of the multiple raster data to form the prediction factor map includes: The multiple raster data are synthesized by band synthesis and synthesis based on a preset grid to form the prediction factor map.

3. The pegmatite-type lithium ore prediction method based on deep learning according to claim 1, characterized in that, The processing of the mask spatial distribution map of the to-be-acquired area according to a preset second grid to form the mask map includes: According to the second grid, the mask spatial distribution map is converted into a mask map by the surface-to-raster method.

4. The pegmatite-type lithium ore prediction method based on deep learning according to claim 1, characterized in that The generation of the training data set and the validation data set based on the prediction factor map, the mask map and the prediction label map of the target area includes: The sliding window technique is adopted to crop and expand the prediction factor map, the mask map and the prediction label map of the target area according to the set window size and step length to generate an expanded data set, wherein the way to generate the expanded data set is: Among them, H is the height of the predicted element map of the acquisition area, W is the width of the predicted element map of the acquisition area, h is the height of the sliding window, w is the width of the sliding window, s h is the step size in the height direction, s w is the step size in the width direction, and N is the number of generated samples; The data in the expanded data set is randomly shuffled, and according to a preset ratio, a part of the data is used as the training data set, and the remaining part of the data is used as the validation data set.

5. The pegmatite-type lithium ore prediction method based on deep learning according to claim 1, wherein The preset fully convolutional neural network is trained by using the training data set and the validation data set to generate a prediction model, including: Determine an optimization function adapted to the fully convolutional neural network, and the optimization function includes: Among them, w is a weight parameter, := indicates that the value on its left is updated or replaced by the value on its right, α is the learning rate, β1 is the first momentum, β2 is the second momentum, dw is a shorthand for the derivative of the cost function with respect to the weight parameter, v dw is the exponentially weighted average of the momentum of dw, s dw is the exponentially weighted average of the momentum of dw 2 ε is a constant value, t represents the number of iterations; Perform the initialization operation of the weight parameters in the fully convolutional neural network; The training data set is input into the fully convolutional neural network in batches, and the corresponding loss function value is obtained by the forward propagation method, and the gradient descent method is used for backpropagation to update the weight parameters in the fully convolutional neural network; The validation data set is input into the trained fully convolutional neural network, and the quality of the formed prediction model is evaluated by the forward propagation method according to the pre-selected validation evaluation index; Repeat the training and validation steps until the loss function value and the validation evaluation index meet the requirements.

6. The pegmatite-type lithium ore prediction method based on deep learning according to claim 1 or 5, characterized in that The fully convolutional neural network is obtained in the following way: Based on the spatial information and prospecting information of the prediction factor map, determine the network main architecture, wherein the network main architecture has an encoder-decoder main architecture with a symmetric structure; Setting an encoding structure of the network main architecture, the encoding structure including a type I convolutional layer and a downsampling structure; and the execution logic of the encoding structure is: performing convolution calculation through the type I convolutional layer, and sequentially performing batch normalization, type I activation function, and attention mechanism module to obtain a first prediction feature map with different heights, widths, and number of features; A decoding structure of the main network architecture is set, and the decoding structure includes: upsampling, skip connection and convolution layer, wherein the convolution layer includes: type I convolution layer and type II convolution layer; and the execution logic of the decoding structure is: the second prediction element graph is obtained by upsampling, and is integrated with the corresponding first prediction element graph in the encoding structure by skip connection to obtain a third prediction element graph; then, convolution calculation is performed through type I and type II convolution layers, and batch normalization, type I activation function and attention mechanism module are performed in sequence to obtain a prediction result graph.

7. The pegmatite-type lithium ore prediction method according to claim 6, wherein The activation function of the type I convolutional layer is: PReLU(x)=max(0,x)+ηmin(0,x) Where x is the value of the activation function after the predicted feature map undergoes convolution operation and batch normalization, and η is a learnable parameter; The upsampling is achieved by bilinear interpolation, wherein the principle of bilinear interpolation includes: Among them, z(x,y) is the value to be determined at the unknown point (x,y), z(x1,y1) is the known value of the known point (x1,y1), z(x2,y1) is the known value of the known point (x2,y1), z(x1,y2) is the known value of the known point (x1,y2), and z(x2,y2) is the known value of the known point (x2,y2).

8. The pegmatite-type lithium ore prediction method based on deep learning according to claim 1, wherein The method of determining the distribution of pegmatite-type lithium deposits in the selected area to be mined based on the prediction model includes: Dividing the area to be collected into various sub-areas; Inputting each sub-region into the prediction model to generate a prediction result map for each sub-region; Splicing the prediction result maps to form a prospecting target area distribution map of the area to be sampled; A surface area containing pegmatite-type veins is selected in the prospecting target area distribution map, and the morphology, scale, spatial distribution, grade, co-existing elements and resource quantity of the lithium pegmatite vein ore body are ascertained based on the exposure, degree of mineralization and surface occurrence information of the pegmatite-type veins.

9. A prediction device for pegmatite-type lithium ore based on deep learning, characterized in that, include: a data acquisition unit, which acquires a plurality of different types of data from the area to be collected and converts the spatial coordinates of the various types of data into a common spatial coordinate system to generate a geospatial database, wherein the plurality of different types of data include geological data, geophysical data, geochemical data, remote sensing data, and mask data; The data processing unit performs rasterization processing on various types of data in the geospatial database to generate a prediction element map, a mask map and a prediction label map, including: rasterizing various types of data in the geospatial database according to a preset first grid to form raster data corresponding to various types of data, and synthesizing multiple raster data to form the prediction element map; rasterizing the mask space distribution map of the area to be collected according to a preset second grid to form the mask map; selecting a target area from the area to be collected, and rasterizing the ore body space distribution map of the target area according to a preset third grid to form a prediction label map, and the first grid, the second grid and the third grid are the same size; wherein, when the geological data, the geophysical data, the geochemical data and the remote sensing data are used simultaneously, in the rasterization process, the grid size of the remote sensing data with the largest scale is used as the size of the preset grid; a data generation unit for generating a training data set and a validation data set based on a prediction factor map, a mask map, and a prediction label map of a target area, wherein the target area is an area in the area to be sampled where a pegmatite-type lithium ore body has been discovered; The processing unit trains a preset fully convolutional neural network using the training data set and the validation data set to generate a prediction model; and based on the prediction model, ascertains the distribution of pegmatite-type lithium deposits in the selected area to be mined; Among them, the loss function of the fully convolutional neural network is: Among them, m is the number of training samples, m0 represents the number of training samples of the background, m1 represents the number of training samples of the lithium-bearing pegmatite vein, and m = m0 + m1. The weight w of the background m0 = m1 / m, and the weight w of the lithium-bearing pegmatite vein m1 = m0 / m; y gt i is the true value of the i-th training sample, and y pre i is the predicted value of the i-th training sample. L(y gt i , y pre i ) is the loss function of the training sample i.

Images