Deep learning-based pegmatite type lithium ore prediction method and device

Abstract

The embodiment of the invention provides a pegmatite type lithium ore prediction method and device based on deep learning, and the method comprises the steps: obtaining a plurality of different types of data of a to-be-collected region, converting the space coordinates of various types of data into a same space coordinate system, and generating a geoscience space database; rasterizing various types of data in the geoscience spatial database to generate a prediction element graph, a mask graph and a prediction label graph; generating a training data set and a verification data set based on the prediction element graph, the mask graph and the prediction label graph of the target area; adopting the training data set and the verification data set to train a preset full convolutional neural network to generate a prediction model; and based on the prediction model, finding out the distribution condition of the pegmatite type lithium ores in the selected to-be-collected area. By adopting the technical scheme, a batch of lithium-containing pegmatite veins with a certain form, scale and spatial distribution can be quickly and effectively identified.

Description

Technical Field

[0001] The embodiments of this specification relate to the technical field of mineral resource exploration, and in particular to a method and device for predicting 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. Because the distribution range of pegmatite lithium mines is not limited to a few basins like brine lithium mines, and they have higher grades and more mature mining and refining technologies, they have once again attracted widespread attention from the industry and academia.

[0003] In my country, lithium is a strategically scarce mineral. Increasing lithium reserves through mineral exploration is a key link in resolving the contradiction between supply and demand of lithium resources.

[0004] However, with the improvement of the exploration level, the chances of discovering large pegmatite lithium deposits in eastern China are getting smaller and smaller, and the focus of prospecting work has to be shifted to the western region with high altitude, deep cutting, difficult access and low working level. However, the western region is vast and comprehensive, systematic and in-depth field investigation is very difficult. How to invest limited manpower, material resources, financial resources and time into the most promising areas has become the first problem to be solved in the exploration of pegmatite lithium deposits, which urgently requires the emergence of efficient and low-cost prospecting technologies and methods.

[0005] Therefore, how to quickly and effectively identify a group of lithium-bearing pegmatite veins with a certain shape, scale and spatial distribution has become a difficult point in the research of lithium mineral prospecting and exploration technology. Summary of the invention

[0006] In view of this, the embodiments of this specification provide a method and device for predicting pegmatite-type lithium deposits based on deep learning, which can quickly and effectively identify a batch of lithium-containing pegmatite veins with a certain shape, scale and spatial distribution.

[0007] The embodiment of this specification provides a method for predicting pegmatite-type lithium deposits based on deep learning, including:

[0008] Acquire multiple different types of data in the area to be collected, and convert the spatial coordinates of various types of data into the same spatial coordinate system to generate a geospatial database;

[0009] rasterizing various types of data in the geospatial database to generate a prediction element map, a mask map, and a prediction label map;

[0010] Generate a training data set and a verification data set based on the prediction element map, the mask map and the prediction label map of the target area, wherein the target area is an area in which a pegmatite-type lithium ore body has been explored in the area to be collected;

[0011] Using the training data set and the validation data set to train a preset fully convolutional neural network to generate a prediction model;

[0012] Based on the prediction model, the distribution of pegmatite-type lithium deposits in the selected area to be mined is ascertained.

[0013] The embodiment of this specification also provides a pegmatite-type lithium ore prediction device based on deep learning, comprising:

[0014] A data acquisition unit acquires a plurality of different types of data in 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;

[0015] A 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;

[0016] A data generation unit generates a training data set and a verification data set based on the prediction element map, the mask map and the prediction label map of the target area, wherein the target area is an area in which a pegmatite-type lithium ore body has been explored in the area to be collected;

[0017] The processing unit 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, find out the distribution of pegmatite-type lithium ore in the selected area to be collected.

[0018] By adopting the deep learning-based pegmatite lithium ore prediction method provided in the embodiment of this specification, a geospatial database can be generated by performing spatial coordinate conversion on multiple different types of data in the area to be collected, and 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, and then a training data set and a verification data set can be generated based on the prediction element map, mask map and prediction label map of the target area. Since the target area is an area where a pegmatite lithium ore body has been explored in the area to be collected, the prediction model obtained by training the preset full convolutional neural network 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, 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 lithium ore in the selected area to be collected can be quickly and effectively ascertained, and then a batch of lithium-bearing pegmatite veins with a certain shape, scale and spatial distribution can be identified. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0020] Figure 1 A flowchart of a method for predicting pegmatite-type lithium deposits based on deep learning in an embodiment of this specification is shown;

[0021] Figure 2 A flowchart of a rasterization process in an embodiment of this specification is shown;

[0022] Figure 3 A schematic diagram of converting a surface into a grid in an embodiment of this specification is shown;

[0023] Figure 4 A schematic diagram of a multi-ring buffer analysis in an embodiment of this specification is shown;

[0024] Figure 5 A flow chart showing a method for generating a training data set and a verification data set in an embodiment of this specification is shown;

[0025] Figure 6 A schematic diagram of a generation process of a prediction model in an embodiment of this specification is shown;

[0026] Figure 7 A simplified diagram of a deep learning algorithm network structure in an embodiment of this specification is shown;

[0027] Figure 8 A schematic diagram of an attention mechanism in an embodiment of this specification is shown;

[0028] Fig. 9 A schematic diagram of a bilinear interpolation principle in an embodiment of this specification is shown;

[0029] Fig.10 A schematic diagram showing a principle of area division in an embodiment of this specification is shown;

[0030] Fig.11 A schematic diagram of the structure of a pegmatite-type lithium ore prediction device based on deep learning in an embodiment of this specification is shown. DETAILED DESCRIPTION

[0031] As described in the background technology, the current exploration schemes for pegmatite-type lithium deposits are limited and cannot effectively explore pegmatite-type lithium deposits.

[0032] With the advent of the era of big data in geoscience, based on the massive geological, mineral, geophysical, geochemical and remote sensing data accumulated in the early stage, the use of artificial intelligence methods to delineate a number of prospecting targets has become an important step in saving prospecting costs, focusing on prospecting targets and improving prospecting efficiency in mineral exploration. However, these methods often have obvious drawbacks. For example, the recognition accuracy of traditional machine learning methods (such as feature analysis, information volume, evidence weight, logistic regression, support vector machine and random forest, etc.) depends to a large extent on the designer's ability to manually extract prospecting information based on prior knowledge, and can only handle problems that are easy to classify in low-dimensional space; self-organizing feedforward neural networks (such as BP, etc.), although they have the ability to deeply mine and highly integrate prospecting information from the original input, often do not consider the location and structure of the research object when extracting deep prospecting information, and the number of parameters is large and the efficiency is low; convolutional neural networks (such as LeNet, AlexNet, VGG and ResNet, etc.) can take into account both spatial structure and feature information at the same time, so as to extract deep prospecting information, but they only judge whether there is a mine in the research scope, but cannot judge 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 for finely defining prospecting target areas of lithium-bearing pegmatite veins with specific shapes, scales and spatial distributions.

[0033] In order to solve the above technical problems, the embodiment of this specification provides a pegmatite-type lithium ore prediction method based on deep learning, which can generate a geospatial database by performing spatial coordinate conversion on multiple different types of data in the area to be collected, and then generate a prediction element map, a mask map and a prediction label map by rasterizing various types of data in the geospatial database, and then generate a training data set and a verification data set based on the prediction element map, mask map and prediction label map of the target area. Since the target area is an area where a pegmatite-type lithium ore body has been explored in the area to be collected, the prediction model obtained by training the preset full convolutional neural network with 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, so as to more comprehensively and objectively reflect the complex mapping relationship between the comprehensive information and the target, so that 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 a certain shape, scale and spatial distribution can be identified.

[0034] 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.

[0035] See also Figure 1 A flowchart of a method for predicting pegmatite-type lithium deposits based on deep learning in an embodiment of this specification is shown in FIG. Figure 1 As shown, the following steps can be performed:

[0036] S11, acquiring a plurality of different types of data in the area to be collected, and converting the spatial coordinates of the various types of data into the same spatial coordinate system to generate a geospatial database.

[0037] Among them, based on the geological mineralization laws and combined with the existing regional geological data, the exploration areas that are favorable to the target deposit type can be selected as the areas to be collected, which can increase the possibility of exploring pegmatite-type lithium mines.

[0038] In some examples, the screening criteria for selecting the area to be mined include: 1) the area to be mined is located within a known large-scale mineralization belt or mineral cluster; 2) there are known pegmatite-type lithium deposits in the surrounding area.

[0039] In this embodiment, the multiple different types of data refer to parameters used to characterize different parameter information of the area to be collected. By acquiring the multiple different types of data, the morphology of the area to be collected can be more realistically reflected.

[0040] In this embodiment, by converting the spatial coordinates of various types of data into the same spatial coordinate system, the format consistency of various types of data can be improved, which is conducive to reducing the difficulty of subsequent data processing in the geospatial database.

[0041] S12, performing 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.

[0042] In this embodiment, by rasterizing various types of data in the geospatial database, the information of various types of data can be retained to the greatest extent, thereby improving the accuracy of the obtained prediction element map, mask map and prediction label map.

[0043] In this embodiment, the prediction element map refers to a graphical expression generated by rasterizing and synthesizing various data such as geology, geophysics, geochemistry, and remote sensing, which can display the distribution pattern and prediction results of lithium-bearing pegmatite veins from different aspects.

[0044] A mask map is usually a two-dimensional graphic generated by rasterizing areas used to represent restricted areas for prospecting, areas with no mineralization potential, or areas with a large amount of data missing. It is used to represent areas in the area to be sampled that do not participate in the training and prediction of pegmatite-type lithium ore models.

[0045] The prediction label map is usually a two-dimensional graphic generated by rasterizing the lithium-bearing granite pegmatite veins that have been identified in the acquisition area. On the prediction label map, certain colors are used to represent the distribution of the lithium-bearing pegmatite veins that have been identified.

[0046] S13, generating a training data set and a verification data set based on the predicted feature map, the mask map and the predicted label map of the target area.

[0047] The target area is an area in the area to be collected where pegmatite-type lithium ore bodies have been explored.

[0048] In this embodiment, through steps S11 and S12, the prediction element map, mask map and prediction label map of the entire area to be collected can be obtained. By selecting the prediction element map, mask map and prediction label map of the target area, and the target area is the area where pegmatite-type lithium ore bodies have been explored in the area to be collected, the accuracy of the training data set and verification data set formed subsequently can be guaranteed.

[0049] S14, using the training data set and the verification data set to train a preset fully convolutional neural network to generate a prediction model.

[0050] In this embodiment, the training data set and the verification data set are obtained from an area where pegmatite-type lithium ore bodies have been explored. Therefore, by using the training data set and the verification data set to train the 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.

[0051] S15, based on the prediction model, finding out the distribution of pegmatite-type lithium ore in the selected area to be mined.

[0052] In this embodiment, through step S14, the prediction model can learn the relationship between the spatial structure and the target of multiple different types of data based on the acquired training data set and verification data set, so that based on the prediction model, the distribution of pegmatite-type lithium deposits in the selected area to be collected can be quickly and effectively ascertained, and then a batch of lithium-bearing pegmatite veins with a certain shape, scale and spatial distribution can be identified.

[0053] In other words, the deep learning-based pegmatite-type lithium ore prediction method provided by the present invention, instead of studying the traditional cause-and-effect relationship, studies the correlation between multiple different types of data (such as comprehensive information such as geology, geophysical exploration, geochemical exploration, and remote sensing) in the area to be collected and the target (background and lithium-bearing pegmatite veins) (often containing cause-and-effect relationships but not limited to cause-and-effect relationships); and not only studies the correlation between the attribute characteristics of the comprehensive information and the target, but also studies the relationship between the spatial structure of the comprehensive information and the target, which is in line with the characteristics of geological big data, thereby more comprehensively and objectively reflecting the complex mapping relationship between the comprehensive information and the target.

[0054] In this embodiment, the multiple different types of data of the area to be collected may include: geological data, mineral data, geophysical data, geochemical data, remote sensing data, and mask data, wherein:

[0055] Geological data mainly refers to regional geological maps. In general, in order to facilitate similar analogies, the collection of regional geological maps within the work area should adhere to the principle of scale equivalence, that is, the scale of the geological maps of the area to be collected needs to be consistent, and it is best to choose data with a larger scale.

[0056] For example, if the regional geological map with a scale of 1:200000 has achieved full coverage in the area to be collected, all the data of 1:200000 will be used, and the regional geological map with a scale of 1:250000 cannot be included; for another example, if the regional geological maps with scales of 1:200000 and 1:50000 have achieved full coverage in the study area, it is necessary to use a regional geological map with a larger scale of 1:50000 to ensure a more accurate prediction result.

[0057] Among them, regional geological maps are mainly stored in vector files, and some old data exist in paper form and need to be vectorized first.

[0058] Geophysical data mainly involve regional gravity, magnetic and radioactive measurement data obtained from geophysical scanning. Generally, in order to facilitate similar analogies, the collection of regional geophysical data in the area to be collected should adhere to the principle of scale equivalence, that is, the scale of regional geophysical data in the area to be collected needs to be consistent, and it is best to choose data with a larger scale.

[0059] For example: if the regional geophysical data with a scale of 1:200000 has achieved full coverage in the area to be collected, then all the data with a scale of 1:200000 will be used, and the regional geophysical data with a scale of 1:250000 cannot be included; for another example, if the regional geophysical data with scales of 1:200000 and 1:50000 have achieved full coverage in the study area, then the regional geophysical data with a larger scale of 1:50000 will need to be used to ensure a more accurate prediction result.

[0060] Among them, the geophysical data of the area to be collected are mainly stored in table files or vector files, and some old data exist in the form of paper media, which need to be vectorized first.

[0061] Geochemical exploration data mainly involve regional stream sediments, rock debris and soil measurement data obtained from geochemical exploration scanning. Different scales and different geochemical landscapes have different media. The scales are mainly 1:200000 and 1:50000.

[0062] Among them, the 1:200000 geochemical scanning measurement data 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, and Zr.

[0063] The mandatory data for the 1:50000 geochemical scanning measurement are 16 elements, including Au, Ag, As, Bi, Cd, Cr, Co, Cu, Hg, Mo, Ni, Pb, Sb, Sn, W, and Zn. Other elements or oxides can be added as appropriate based on actual conditions.

[0064] In general, in order to facilitate similar analogies, the collection of regional geochemical exploration data in the area to be collected should adhere to the principle of scale equivalence, that is, the scale of the regional geochemical exploration data in the area to be collected needs to be consistent, and it is best to choose data with a larger scale.

[0065] For example: if the regional geochemical exploration data with a scale of 1:200000 has achieved full coverage in the working area, then all the data of 1:200000 will be used, and the regional geochemical exploration data with a scale of 1:250000 cannot be included; for another example, if the regional geochemical exploration data with scales of 1:200000 and 1:50000 have achieved full coverage in the study area, then the regional geochemical exploration data with a larger scale of 1:50000 will need to be used to ensure more accurate prediction results.

[0066] In addition, although there are larger-scale geochemical exploration data, if the content of important target elements has not been measured (such as the Li element used to search for pegmatite-type lithium deposits, which has not been measured in some 1:50,000 regional geochemical exploration work), the 1:200,000 geochemical exploration data will also be used.

[0067] Among them, geochemical exploration data are mainly stored in table files or vector files, and some old data exist in paper form and need to be vectorized first.

[0068] There are many types of remote sensing data. As far as the search for pegmatite-type lithium deposits is concerned, remote sensing data should meet two requirements: high spatial resolution and high spectral resolution.

[0069] In this embodiment, since the lithium-bearing pegmatite vein is not large in scale, generally a few meters to tens of meters wide and tens to hundreds of meters long, it is necessary to select a remote sensing image with a high spatial resolution; secondly, the remote sensing image should contain a band that can reflect the spectral characteristics of lithium, so as to effectively extract relevant information about lithium. Studies have shown that lithium (lithium-bearing pegmatite veins) have multiple absorption valleys and reflection peaks in short-wave infrared (SWIR), and this spectral feature 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 because of their sub-meter high spatial resolution; in terms of high spectral resolution, Resource-1 02D satellite (ZY1E) or Gaofen-5 (GF-5) satellite images are also an ideal choice because they contain multiple bands corresponding to the characteristic spectra of lithium (lithium-bearing pegmatite veins) in the short-wave infrared range.

[0070] Currently, regional remote sensing data are mainly stored in raster data files.

[0071] The main purpose of mask data is to separate the restricted areas for prospecting, areas with no mineralization potential or areas with a large amount of data missing (for pegmatite-type lithium deposits, the three zones and three lines, basins, lakes, snow, etc. in the working area) in the study area according to actual conditions, so that these areas will not participate in training models and conducting predictions, thereby further limiting the scope of work.

[0072] Among them, the collected and produced mask data are often stored in vector file format.

[0073] Mineral data, as far as pegmatite-type lithium deposits are concerned, mainly involves collecting ore body distribution information in the area to be sampled, which is mainly manifested as lithium-bearing pegmatite veins with a certain shape, scale and spatial distribution.

[0074] Among them, mineral data is often stored in vector file format.

[0075] It should be noted that the multiple different types of data collected in the above examples are only examples for indicating the acquisition of multiple data and should not be construed as a limitation to the present invention.

[0076] In this embodiment, due to the different types, ages and sources of the collected data, there are also great differences in the geographic and projection coordinate systems. Among them, the projection coordinate system includes Gauss-Kruger (3 degree zone or 6 degree zone), UTM and Lambert projection, and the common geographic coordinate systems corresponding to the projection coordinate system include Beijing 54, Xi'an 80, CGCS2000 and WGS84.

[0077] In order to analyze data in a unified spatial coordinate system, the geographic coordinate system will uniformly adopt the current CGCS2000 geographic coordinate system in my country, and the projection coordinate system will depend on the specific situation. In the case of a small range, Gauss-Kruger (3-degree zone or 6-degree zone) projection is mainly used, and if the range is large, Lambert projection is used. Different spatial coordinate systems can be converted into a unified coordinate system using the three-parameter or seven-parameter "projection transformation" method in ArcGIS software.

[0078] In this embodiment, after the geospatial database is generated, the data in the geospatial database may be processed to form a prediction element map, a mask map, and a prediction label map.

[0079] In a specific example, see Figure 2 A flowchart of a rasterization process in an embodiment of the present disclosure is shown in FIG. Figure 2 As shown, the following steps can be performed:

[0080] S21, 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.

[0081] Specifically, different types of data have different formats, and when rasterized, the processing methods are not exactly the same. By rasterizing various types of data, corresponding raster data can be obtained, and by synthesizing these raster data, a prediction element 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.

[0082] Moreover, in 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 amount of calculation will be greatly increased and the operation efficiency will be affected. In general, in order not to lose data information, the grid size of the data with the highest accuracy is often used as the preset grid size.

[0083] For example, when geological, geophysical, geochemical and remote sensing data are available at the same time, the grid size of the remote sensing data with the largest scale (highest accuracy) is often used as a common grid to rasterize all data, which basically ensures that data information will not be lost.

[0084] In addition, the size of the ore body will also affect the size of the preset grid. If the ore body is relatively large, the grid size can be appropriately increased to reduce unnecessary calculations.

[0085] In this embodiment, when the data is the geological data, and when the geological data is determined to be a stratum and a rock mass, the surface-to-raster method is used to rasterize the vectorized stratum and 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 surface elements, and then the surface-to-raster method is used to rasterize the vectorized fold structure or fault structure to form raster data corresponding to the geological data.

[0086] Specifically, geological data is mainly processed for strata, rock mass and structure. The processing methods can be roughly divided into three categories: the first category is the processing of strata and rock mass, the second category is the processing of fold structure, and the third category is the processing of fault structure.

[0087] Specifically, 1) for strata and rock masses, since they are represented as surface elements on geological maps, the “surface to grid” method can be directly used (e.g. Figure 3 1) For folds (synclines and anticlines), the fold structures can be circled on the geological map according to the distribution of strata (synclines are often shown as new in the middle and old on both sides; anticlines are often shown as old in the middle and new on both sides) to form surface elements. Similarly, the vectorized folds can be directly rasterized (for example, in *.tiff format) using the "surface to grid" method. 2) For fault structures (normal faults, reverse faults, and strike-slip faults), the "multi-ring buffer zone" method can be used according to the scope of their mineralization (such as Figure 4 As shown in the figure, the fault structure is formed into surface elements. Similarly, the vectorized fault structure is directly rasterized (for example, *.tiff format) using the "surface to raster" method.

[0088] In some embodiments, after the geological data is rasterized, in order to unify the dimensions, improve the prediction effect and speed up the calculation, the grid attributes are processed using a "standardization" method, and the standardization process is shown in the following formula (1).

[0089]

[0090] Among them, x i is the i-th value, x mean is the mean value, σ is the standard deviation, ξ is a constant, x std is the standardized value.

[0091] When the data is geophysical data and the geophysical data is a table file, the data in the table file is standardized and then interpolated using the inverse distance weighted method; and when the geophysical data is a vectorized contour surface map, the surface-to-raster method is used to convert the vectorized geophysical map into a rasterized geophysical map, and then the standardization process is used to form raster data corresponding to the geophysical data.

[0092] Specifically, when the geophysical data is for a table file, the table data is first processed using the "standardization" method, and then the data is interpolated using the "inverse distance weighted" method (Formula 2) (for the preset first grid) to form geophysical gridded data (for example, *.tiff format); for geophysical vectorized contour maps, first, the "surface to raster" method is used to convert the vectorized geophysical map into a rasterized geophysical map (for example, *.tiff format), and then the "standardization" method (see Formula 1) is used for processing.

[0093]

[0094] Where 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.

[0095] When the data is geochemical data and the geochemical data is a table file, the data in the table file is standardized and then interpolated using the inverse distance weighted method; and when the geochemical data is a vectorized contour surface map, the surface-to-raster method is used to convert the vectorized geochemical map into a rasterized geochemical map, and then the standardization process (see formula 1) is used to form raster data corresponding to the geochemical data.

[0096] When the data is remote sensing data, the original image of the remote sensing data is calibrated (for example, radiometric calibration, FLAASH atmospheric correction, orthorectification), and then a standardization method is used to form raster data corresponding to the remote sensing data.

[0097] Therefore, according to different data types, different rasterization methods can be used to form corresponding raster data.

[0098] In this embodiment, after a plurality of raster data are formed, the plurality of raster data may be synthesized to form a prediction element map.

[0099] Specifically, band synthesis and common grid-based synthesis are used to fuse the multiple raster data to form a multi-factor prediction feature map. The so-called band synthesis is essentially to create a separate 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, that is, a multi-factor prediction feature map.

[0100] S22, rasterizing the mask space distribution map of the area to be collected according to a preset second grid to form the mask map.

[0101] Specifically, the method includes: according to the second grid, using a surface-to-grid method to convert the mask space distribution map into a mask map.

[0102] In this embodiment, the format of the mask image may be a *.tiff format.

[0103] 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.

[0104] In this embodiment, the target area, as the data collection area for training data and verification data, should meet at least the following conditions: 1) The exploration degree 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, so as to ensure 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 process of prospecting and prediction, so that the trained deep learning model has a high generalization ability, so that when conducting prospecting and prediction 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.

[0105] In this case, the generation process of the prediction label map may include: according to the third grid, using the surface-to-grid method, rasterizing the ore body spatial distribution map of the target area to form the prediction label map.

[0106] In this embodiment, the format of the predicted label image may be in *.tiff format.

[0107] In this embodiment, the first grid, the second grid and the third grid are of the same size.

[0108] After determining the predicted feature map, mask map, and predicted label map of the target area, a training data set and a validation data set can be generated.

[0109] In an alternative example, see Figure 5 A flowchart of a method for generating a training data set and a verification data set in an embodiment of this specification, such as Figure 5 As shown, specifically including:

[0110] S51, using sliding window technology, according to the set window size and step size, the predicted feature map, mask map and predicted label map of the target area are cropped and expanded to generate an expanded data set.

[0111] Specifically, the basic idea of ​​using the deep learning-based pegmatite lithium ore comprehensive information prospecting prediction method to carry out prospecting target area delineation is still similar analogy, and it requires the support of larger capacity training samples. In reality, there are not too many lithium-bearing pegmatite veins identified in the data collection area, so the data set is relatively small.

[0112] In this embodiment, in order to expand the samples and meet the needs of deep learning, a sliding window method is adopted, and a consistent window size and step size are used to crop and expand the prediction feature map, mask map and prediction label map of the data collection area to generate a data set for deep learning model training and verification.

[0113] In other words, a set of one-to-one corresponding sub-prediction feature maps, sub-mask maps, and sub-prediction label maps are generated.

[0114] In this embodiment, the method of generating the expansion data set is:

[0115]

[0116] Among them, H is the height of the prediction feature map of the acquisition area, W is the width of the prediction feature map of the acquisition area, h is the height of the sliding window, w is the width of the sliding window, and s h is the step length in the height direction, s w is the step length in the width direction, and N is the number of generated samples.

[0117] It should be noted that when using the sliding window technology, it is better to have a sliding window that can more completely cover the lithium-bearing pegmatite vein, and the step size is usually smaller than the window size to increase the sample size.

[0118] S52, randomly shuffle the data in the expanded data set, and use a portion of the data as the training data set and the remaining data as the verification data set according to a preset ratio.

[0119] Specifically, by randomly shuffling the data in the expanded data set, it is possible to ensure that the relationship between samples in the expanded data set is random and has no logic in order, thereby preventing similar samples from being too concentrated in the training set or validation set, resulting in the low generalization ability of the trained deep learning model; then, the randomly shuffled data set is divided into corresponding training set and validation set folders according to a certain ratio, for example, 7:3 or 8:2, to form training sets and validation sets for subsequent model training and verification work.

[0120] After obtaining the training set and validation set that meet the training requirements, the preset fully convolutional neural network can be trained to generate a prediction model.

[0121] See also Figure 6 A schematic diagram of a generation process of a prediction model in an embodiment of this specification is shown in FIG. Figure 6 As shown, the following generation steps can be performed:

[0122] S61, determining a loss function suitable for the fully convolutional neural network.

[0123] Specifically, the loss function is an indicator of 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 has reached the optimal level during the training stage.

[0124] In this embodiment, in the prospecting prediction of pegmatite-type lithium deposits, whether it is a training sample or a verification sample, the number of backgrounds is often much larger than the number of lithium-bearing pegmatite veins. For this reason, this model uses an improved weighted cross entropy function, that is, a smaller weight is given to the background and a larger weight is given to the lithium-bearing pegmatite veins. The so-called background refers to the distribution area without lithium-bearing pegmatite veins in the training samples. These areas can be distributed with any non-lithium-bearing pegmatite geological bodies, such as sedimentary rocks, metamorphic rocks, etc.

[0125] In this embodiment, the loss function is:

[0126]

[0127] Where m is the number of training samples, m0 is the number of background training samples, m1 is the number of lithium-bearing pegmatite vein training samples, and m = (m0) + (m1), the background weight w m0 =m1 / m, weight w of lithium-bearing pegmatite vein 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.

[0128] S62, determining an optimization function that is compatible with the fully convolutional neural network.

[0129] Specifically, the optimization function is used to optimize the weight parameters. Adam combines the advantages of the two methods, Momentum and RMSprop, to achieve efficient search of parameter space, thus being widely used. This deep learning model optimization adopts this method, and the optimization function is:

[0130]

[0131] Where w is the weight parameter, := means that the value on the left is updated or replaced by the value on the 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, and v dw is the momentum exponential weighted average of dw, s dw for dw 2 Momentum exponential weighted average, ε is a constant, and t represents the number of iterations;

[0132] S63, performing weight parameter initialization operation in the fully convolutional neural network.

[0133] Among them, the weight parameter initialization 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 all weight parameters are not allowed to be initialized to 0, so random initialization is required. Since this deep learning method uses PReLU as the activation function, in order to avoid the gradient disappearance problem caused by the deepening of the number of layers, the He random initialization method is adopted, which can ensure that information flows effectively during the forward propagation and back propagation process, so that the variance of the input signals of different layers is roughly equal.

[0134] Specifically, by adjusting the initial value of the weight, we ensure that the variance of the input signal remains consistent in each layer of the network, which helps improve the training efficiency and performance of the network. He is randomly initialized to a normal (Gaussian) distribution with a mean of 0 and a variance of 2 / n, that is,

[0135]

[0136] Among them, n is the number of weight parameters of each convolutional layer, and η is the learnable parameter in the PReLU activation function.

[0137] 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 back propagation to update the weight parameters in the fully convolutional neural network.

[0138] Specifically, step S64 is performed to input the training sample set into the model in batches and calculate the prediction result F through forward propagation. NC = 2, and then backpropagates and updates the weight parameters using the Adam gradient descent method.

[0139] S65, inputting the verification data set into the trained fully convolutional neural network, and using a forward propagation method to evaluate the quality of the formed prediction model according to pre-selected verification evaluation indicators.

[0140] Specifically, step S65 is performed to input all the validation sample sets into the model and calculate the prediction result F through forward propagation. NC =2 (e.g., GA, MA, MA of class i, and MIU) to evaluate the quality of the deep learning model.

[0141] In this embodiment, the quality of the deep learning model is mainly verified based on various evaluation indicators of the confusion matrix, among which the global accuracy GA, the average accuracy MA, and the single intersection-union ratio IU i , and average intersection-union ratio MIU as evaluation indicators.

[0142] in:

[0143]

[0144] Among them, n ii is the number of pixels of category i predicted as category i; n ij is the number of pixels of category i predicted to be category j; n ji is the number of pixels of category j predicted as category i; n cls is the number of target categories; GA is the global accuracy, MA is the average accuracy; IU i is the intersection-union ratio of class i; MIU is the average intersection-union ratio.

[0145] S66, repeat the iterative training and verification steps until the loss function value and the verification evaluation index meet the requirements.

[0146] For example, the above steps are repeated multiple times until the cost function and the verification evaluation index reach the required accuracy, which is mainly manifested in that the cost function value no longer decreases significantly and the verification evaluation index no longer increases significantly, and the optimal deep learning model is obtained.

[0147] Therefore, by adopting the above training method, on the one hand, the sliding window method is used to produce and expand the training and verification data sets, which lays the sample foundation for the deep learning model based on big data; on the other hand, the He method is used to randomly initialize the weights, and by adjusting the initial value of the weights, it is ensured that the variance of the input signal remains consistent in each layer of the network, thereby improving the training efficiency and performance of the network; in addition, for the training samples of lithium-bearing pegmatite veins that are far smaller than the background training samples, the loss function is improved, and a smaller weight is given to the background with a large number of samples, while 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) determining the main network architecture based on the spatial information and prospecting information of the prediction element graph, wherein: Figure 7 As shown, the main network architecture has a coding-decoding main architecture with a symmetrical structure.

[0150] In this way, it is possible to fully extract the deep prospecting information of the prediction element map while making full use of the spatial information of the prediction element map to identify lithium-bearing pegmatite veins and background at the pixel scale.

[0151] A2) setting a coding structure of the main network architecture, the coding structure comprising: a type I convolutional layer and a downsampling structure; and the execution logic of the coding structure is: performing convolution calculation through a type I convolutional layer, and sequentially performing batch normalization, a type I activation function and an attention mechanism module to obtain a first prediction feature map with different heights, widths and numbers of features.

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

[0153] A21) Set the height to H in , width is W in , the number of features is F in The predicted feature map is obtained by using F b Convolutional layers (type I) are used for convolution calculation, and batch normalization, activation function (type I) and attention mechanism modules are performed in sequence to obtain a height of H in , width is W in , the number of features is F b The first predicted feature map of .

[0154] A22) Repeat step A21) until the height H is obtained. in , width is W in , the number of features is Fb The first predicted feature map of .

[0155] A23) Set the height to H in , width is W in , the number of features is F b Predict the feature map and get a height of H by downsampling in / 2, width is W in / 2, the number of features is F b The first predicted feature map of .

[0156] A24) Set the height to H in / 2, width is W in / 2, the number of features is F b Predicting feature maps by using 2F b Convolutional layers (type I) are used for convolution calculation, and batch normalization, activation function (type I) and attention mechanism modules are performed in sequence to obtain a height of H in / 2, width is W in / 2, the number of features is 2F b The first predicted feature map of .

[0157] A25) Repeat step A24) and still get a height of H in / 2, width is W in / 2, the number of features is 2F b The first predicted feature map of .

[0158] A26) Set the height to H in / 2, width is W in / 2, the number of features is 2F b The predicted feature map is downsampled to a height of H in / 4, width is W in / 4, the number of features is 2F b The first predicted feature map of .

[0159] A27) Set the height to H in / 4, width is W in / 4, the number of features is 2F b Predict feature maps by using 4F b Convolutional layers (type I) are used for convolution calculation, and batch normalization, activation function (type I) and attention mechanism modules are performed in sequence to obtain a height of H in / 4, width is W in / 4, the number of features is 4F b The first predicted feature map of .

[0160] A28) Repeat step A27) and still get a height of H in / 4, width is Win / 4, the number of features is 4F b The first predicted feature map of .

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

[0162] A30) Set the height to H in / 8, width is W in / 8, the number of features is 4F b Predicting feature maps by using 8F b Convolutional layers (type I) are used for convolution calculation, and batch normalization, activation function (type I) and attention mechanism modules are performed in sequence to obtain a height of H in / 8, width is W in / 8, the number of features is 8F b The first predicted feature map of .

[0163] A31) Repeat step A30) and still get a height of H in / 8, width is W in / 8, the number of features is 8F b The first predicted feature map of .

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

[0165] A33) Set the height to H in / 16, width is W in / 16, the number of features is 8F b Predicting feature maps by using 8F b Convolutional layers (type I) are used for convolution calculation, and batch normalization, activation function (type I) and attention mechanism modules are performed in sequence to obtain a height of H in / 16, width is W in / 16, the number of features is 8F b The first predicted feature map of .

[0166] A34): Repeat step A33) and still get a height of H in / 16, width is W in / 16, the number of features is 8F b The first predicted feature map of .

[0167] A4) setting a decoding structure of the main network architecture, the decoding structure comprising: upsampling, skip connection and convolution layer, wherein the convolution layer comprises: type I convolution layer and type II convolution layer; and the execution logic of the decoding structure is: obtaining the second prediction element graph through upsampling, and integrating it with the corresponding first prediction element graph in the encoding structure through skip connection to obtain a third prediction element graph; then, performing convolution calculation through type I and II convolution layers, as well as batch normalization, type I activation function, attention mechanism module, type II activation function, and maximum likelihood transformation to obtain a prediction result graph.

[0168] Specifically, see Figure 7 , the execution logic of the decoding structure is:

[0169] A41) Set the height to H in / 16, width is W in / 16, the number of features is 8F b The first predicted feature map is obtained by upsampling to a height of H in / 8, width is W in / 8, the number of features is 8F b The second predicted feature map.

[0170] A42) The up-sampled height is H in / 8, width is W in / 8, the number of features is 8F b The second predicted element map and the corresponding height in the encoding structure are H in / 8, width is W in / 8, the number of features is 8F b The first predicted feature map of is integrated through skip connections to obtain a height of H in / 8, width is W in / 8, the number of features is 16F b The third prediction factor map of .

[0171] A43) Set the above height to H in / 8, width is W in / 8, the number of features is 16F b The third predicted feature map is obtained by using 8F b Convolutional layers (type I) are used for convolution calculation, and batch normalization, activation function (type I) and attention mechanism modules are performed in sequence to obtain a height of H in / 8, width is W in / 8, the number of features is 8F b The third prediction factor map of .

[0172] A44) Set the height H in A43) in / 8, width is W in / 8, the number of features is 8F b The third predicted feature map is obtained by using 4F b Convolutional layers (type I) are used for convolution calculation, and batch normalization, activation function (type I) and attention mechanism modules are performed in sequence to obtain a height of H in / 8, width is W in / 8, the number of features is 4F b The third prediction factor map of .

[0173] A45) Set the height H in A44) in / 8, width is W in / 8, the number of features is 4F b The third prediction feature map is obtained by upsampling to a height of H in / 4, width is W in / 4, the number of features is 4F b The second predicted feature map.

[0174] A46) The up-sampled height is H in / 4, width is W in / 4, the number of features is 4F b The second predicted element map and the corresponding height in the encoding structure are H in / 4, width is W in / 4, the number of features is 4F b The first predicted feature map of is integrated through skip connections to obtain a height of H in / 4, width is W in / 4, the number of features is 8F b The third prediction factor map of .

[0175] A47) Set the above height to H in / 4, width is W in / 4, the number of features is 8F b The third predicted feature map is obtained by using 4F b Convolutional layers (type I) are used for convolution calculation, and batch normalization, activation function (type I) and attention mechanism modules are performed in sequence to obtain a height of H in / 4, width is W in / 4, the number of features is 4F b The third prediction factor map of .

[0176] A48) Set the height in A47) to H in / 4, width is W in / 4, the number of features is 4F b Predicting feature maps by using 2F b Convolutional layers (type I) are used for convolution calculation, and batch normalization, activation function (type I) and attention mechanism modules are performed in sequence to obtain a height of H in / 4, width is W in / 4, the number of features is 2F b The third prediction factor map of .

[0177] A49) Set the height in A48) to H in / 4, width is W in / 4, the number of features is 2F b The third prediction feature map is obtained by upsampling to a height of H in / 2, width is W in / 2, the number of features is 2F b The second predicted feature map.

[0178] A50) The up-sampled height is H in / 2, width is W in / 2, the number of features is 2F b The second predicted element map and the corresponding height in the encoding structure are H in / 2, width is W in / 2, the number of features is 2F b The first predicted feature map of is integrated through skip connections to obtain a height of H in / 2, width is W in / 2, the number of features is 4F b The third prediction factor map of .

[0179] A51) Set the height H in A50) in / 2, width is W in / 2, the number of features is 4F b The third predicted feature map is obtained by using 2F b Convolutional layers (type I) are used for convolution calculation, and batch normalization, activation function (type I) and attention mechanism modules are performed in sequence to obtain a height of H in / 2, width is W in / 2, the number of features is 2F b The third prediction factor map of .

[0180] A52) will change the height of A51) to H in / 2, width is W in / 2, the number of features is 2F b The third predicted feature map is obtained by using Fb Convolutional layers (type I) are used for convolution calculation, and batch normalization, activation function (type I) and attention mechanism modules are performed in sequence to obtain a height of H in / 2, width is W in / 2, the number of features is F b The third prediction factor map of .

[0181] A53) Set the height in A52) to H in / 2, width is W in / 2, the number of features is F b The third prediction feature map is obtained by upsampling to a height of H in , width is W in , the number of features is F b The second predicted feature map.

[0182] A54) The up-sampled height is H in , width is W in , the number of features is F b The second predicted element map and the corresponding height in the encoding structure are H in , width is W in , the number of features is F b The first predicted feature map of is integrated through skip connections to obtain a height of H in , width is W in , the number of features is 2F b The third prediction factor map of .

[0183] A55) Set the height H in A54) in , width is W in , the number of features is 2F b The third predicted feature map uses F b Convolutional layers (type I) are used for convolution calculation, and batch normalization, activation function (type I) and attention mechanism modules are performed in sequence to obtain a height of H in , width is W in , the number of features is F b The third prediction factor map of .

[0184] A56) Set the height in A55) to H in , width is W in , the number of features is F b The third predicted feature map uses F b Convolutional layers (type I) are used for convolution calculation, and batch normalization, activation function (type I) and attention mechanism modules are performed in sequence to obtain a height of H in , width is W in , the number of features is F b The third prediction factor map of .

[0185] A57) Set the height in A56) to H in , width is W in , the number of features is F b The third predicted feature map uses F NC = 2 convolution layers (Type II) perform convolution calculations to obtain a height of H in , width is W in , the number of categories is F NC =2 Classification prediction graph.

[0186] A58) Set the above height to H in , width is W in , the number of features is F NC The classification prediction graph with =2 is transformed through the activation function (type II) and the maximum likelihood method to obtain the prediction result graph F1.

[0187] In order to better understand and illustrate the working logic of the fully convolutional neural network in the embodiments of this specification, examples are given for illustration.

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

[0189] Based on the relationship between the convolution process and the size 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, that is: H out =H in , W out =W in , when the convolution process step size of the 3×3 convolutional layer is 1 and the dilation coefficient is 1, the padding should be 1.

[0190] It should be pointed out 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] Among them, H out is the height of the output prediction feature map, Hin 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 expansion coefficient, P refers to padding, K refers to the convolutional layer size, and S refers to the step size.

[0194] Next, we perform batch normalization on the data. To simplify the search for hyperparameters, make the selection of hyperparameters less sensitive, and make the fully convolutional neural network more robust and easier to train, we need to normalize not only the original prediction feature map, but also the prediction feature map of each intermediate layer, as shown in Formula 13. This is batch normalization, and we get the value after quantity normalization.

[0195]

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

[0197] Then, define the activation function (type I). The activation function (type I) is introduced to enable the fully convolutional neural network to solve nonlinear problems. At present, the ReLU activation function is widely used in the hidden layer. It can solve the gradient vanishing problem caused by the saturated activation function and greatly speed up the learning speed. However, the disadvantage of this function is that when a very large gradient passes through the back propagation process, the back propagation update may cause the center of the weight distribution to be less than 0, so that the derivative at this point is always 0, and the back propagation cannot update the weight, that is, it enters the inactivated state.

[0198] PReLU is an improved version of the ReLU activation function. The principle is shown in Formula 14. It can adaptively learn the parameter η from the data. It can not only 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 of the activation function input after the prediction feature map undergoes convolution operation and batch normalization, and η is a learnable parameter.

[0201] Next, set the attention mechanism. The attention mechanism can be divided into feature attention mechanism and spatial attention mechanism. The feature attention mechanism is mainly to normalize each feature of the input prediction factor map through the maximum pooling layer, the fully connected layer and the Sigmoid activation function in turn to form the weight of each feature, and then multiply each weight by the corresponding feature of the input prediction factor map, and then assign a certain weight to each feature to strengthen the role of the importance of different features for target classification prediction; the spatial attention mechanism is to normalize each spatial position of the input prediction factor map through the maximum pooling layer, convolution operation and Sigmoid activation function to form the weight of each position, and then multiply each weight by the corresponding position of the input prediction factor map, and then assign a certain weight to each spatial position to strengthen the role of the importance of different spatial positions for target classification prediction. As far as the prospecting prediction of pegmatite-type lithium deposits is concerned, different features and different positions do have different effects on classification prediction.

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

[0203] Then, set the downsampling structure and its downsampling process: the purpose of downsampling is to sparsely process the prediction feature map, reduce the amount of data calculation, and increase the receptive field. Among them, the pooling layer is the most commonly used method, including the average pooling layer and the maximum pooling layer. In recent years, people have found that the maximum pooling layer can retain more important features than the average pooling layer. Therefore, the network downsampling structure uses a 2×2 maximum pooling layer with a step size of 2 in the pooling process.

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

[0205]

[0206] Among them, z(x,y) is the value to be determined at the unknown point (x,y), z(x 1,y1 ) is a known point (x 1,y1 ), z(x 2,y1 ) is a known point (x 2,y1 ), z(x 1,y2 ) is a known point (x 1,y2), z(x 2,y2 ) is a known point (x 2,y2 ) is a known value.

[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 prediction feature map input in any upsampling layer in the network structure.

[0208] Then, define the jump connection method. There are currently two types of jump connection methods. One is that the prediction feature map formed by upsampling during the decoding process is added to the prediction feature map in the corresponding coding structure. Since it has completed the linear fusion of the data, it is generally not convolved again, but directly upsampled; the other is that the prediction feature map formed by upsampling during the decoding process and the prediction feature map in the corresponding coding structure are collaged, and then convolved twice for nonlinear fusion. In view of the fusion method of collage + convolution combination, it can better cross-feature nonlinear fusion of shallow features and deep features, thereby making full use of spatial information and semantic information, and the effect will be better. This jump connection adopts the collage method.

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

[0210] Finally, the activation function (Type II) is set and the prediction result graph is generated.

[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 prediction result map, as shown in formula (16).

[0212]

[0213] Among them, 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 input value that makes Softmax(x i ) Get the maximum value of x i value.

[0214] Therefore, a fully convolutional neural network with the above-mentioned 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 pays more attention to combining shallow spatial information with deep semantic information through jump connections, which can not only determine whether there are lithium-bearing pegmatite veins in the area to be collected, but also describe its morphology, scale and spatial distribution in detail, thereby ensuring the prediction accuracy of lithium-bearing pegmatite veins, and being able to finely define the prospecting target area, providing detailed basis for further prospecting and exploration.

[0215] On the other hand, the comprehensive information prospecting prediction method based on deep learning provided by the present invention has improved the important components of the deep learning model and optimized the overall network structure. The use of PReLU instead of ReLU activation function prevents the network deactivation problem that may be caused when a large gradient passes through the back propagation process, and also speeds up the network convergence speed and reduces the error rate; the introduction of feature and spatial attention mechanism strengthens the role of different features and spatial position importance in the identification of lithium-bearing pegmatite veins, and improves the prediction accuracy; bilinear interpolation is used instead of transposed convolution for upsampling. Since there are no parameters and no iteration is required, the network operation efficiency is greatly accelerated, making the deep learning network more practical.

[0216] After forming the prediction model, the distribution of pegmatite-type lithium ore in the selected area to be collected can be ascertained. Specifically, it includes: dividing the area to be collected into various sub-areas; inputting each sub-area into the prediction model to generate a prediction result map of each sub-area; splicing the prediction result map to form a prospecting target area distribution map of the area to be collected; selecting a surface area containing pegmatite-type veins in 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 of the pegmatite-type veins, the degree of mineralization, and the surface occurrence information.

[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 edge of the sub-area have a good prediction effect, a small amount of overlap is often used.

[0218] Among them, Fig.10 As shown, the area to be collected QY is divided into various sub-areas (for example Fig.10 In 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 used, and the width of the overlap is at least twice the number of type I convolutions of the entire network.

[0219] In a specific 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] Next, 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 ultimately discover pegmatite-type lithium mines.

[0223] The embodiment of this specification also provides a device corresponding to the pegmatite lithium ore prediction method based on deep learning, such as Fig.11 A schematic diagram of the structure of a pegmatite-type lithium ore prediction device based on deep learning in an embodiment of this specification is shown in FIG. Fig.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 in 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] A 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] A data generation unit 130 generates a training data set and a verification data set based on a prediction element map, a mask map, and a prediction label map of a target area, wherein the target area is an area in which a pegmatite-type lithium ore body has been explored in the area to be collected;

[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 ore in the selected area to be collected.

[0228] The specific working process and principle 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 description of the aforementioned example.

[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 as 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, so the protection scope of the present invention shall be subject to the scope defined by the claims.

Claims

1. A method for predicting pegmatite-type lithium deposits 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 various types of data into the same spatial coordinate system to generate a geospatial database; rasterizing various types of data in the geospatial database to generate a prediction element map, a mask map, and a prediction label map; Generate a training data set and a verification data set based on a prediction element map, a mask map, and a prediction label map of a target area, wherein the target area is an area in which a pegmatite-type lithium ore body has been explored in the area to be collected; Using the training data set and the validation data set to train a preset fully convolutional neural network 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.

2. The pegmatite-type lithium ore prediction method based on deep learning according to claim 1 is characterized in that: The rasterization process of various types of data in the geospatial database to generate a prediction element map, a mask map and a prediction label map includes: According to a preset first grid, various types of data in the geospatial database are subjected to rasterization processing to form raster data corresponding to various types of data, and a plurality of the raster data are subjected to data synthesis to form the prediction element map; According to a preset second grid, rasterizing the mask space distribution map of the area to be collected to form the mask map; A target area is selected from the area to be collected, and the ore body spatial distribution map of the target area is rasterized according to a preset third grid to form a prediction label map.

3. The pegmatite-type lithium ore prediction method based on deep learning according to claim 2 is characterized in that: The multiple different types of data include: geological data, geophysical data, geochemical data, remote sensing data, and mask data; The method of rasterizing various types of data in the geospatial database according to the preset first grid to form raster data corresponding to 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, the vectorized stratum and the rock mass are rasterized by using a surface-to-raster method, 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 rasterize the vectorized fold structure or fault structure to form raster data corresponding to the geological data; When the data is geophysical data or geochemical data, and the geophysical data or geochemical data is a table file, the data in the table file is standardized, and then the inverse distance weighted method is used for interpolation processing; and when the geophysical data or geochemical data is a vectorized contour surface map, the vectorized geophysical map or geochemical map is converted into a rasterized geochemical map by using a surface-to-grid method, and then the standardized processing is used to form raster data corresponding to the geophysical data or geochemical data; When the data is remote sensing data, a calibration operation is performed on the original image of the remote sensing data, and then a standardization method is adopted to form raster data corresponding to the remote sensing data; The step of synthesizing the plurality of grid data to form the prediction element map comprises: Band synthesis and common grid-based synthesis are adopted to synthesize the plurality of raster data to form the prediction element map.

4. The method for predicting pegmatite-type lithium ore based on deep learning according to claim 2, characterized in that: The step of processing the mask space distribution map of the area to be collected according to the preset second grid to form the mask map includes: According to the second grid, the mask space distribution map is converted into a mask map by using a surface-to-grid conversion method.

5. The method for predicting pegmatite-type lithium ore based on deep learning according to claim 1, characterized in that: The generating of a training data set and a verification data set based on the predicted feature map, the mask map and the predicted label map of the target area includes: The sliding window technology is used to crop and expand the predicted element map, mask map and predicted label map of the target area according to the set window size and step size to generate an expanded data set, wherein the expanded data set is generated in the following manner: Among them, H is the height of the prediction feature map of the acquisition area, W is the width of the prediction feature map of the acquisition area, h is the height of the sliding window, w is the width of the sliding window, and s h is the step length in the height direction, s w is the step length in the width direction, and N is the number of generated samples; The data in the expanded data set are randomly shuffled, and according to a preset ratio, a portion of the data is used as the training data set, and the remaining data is used as the verification data set.

6. The method for predicting pegmatite-type lithium ore based on deep learning according to claim 1, characterized in that: The method of using the training data set and the verification data set to train a preset fully convolutional neural network to generate a prediction model includes: Determine a loss function suitable for the fully convolutional neural network, the loss function is: Where m is the number of training samples, m0 is the number of background training samples, m1 is the number of lithium-bearing pegmatite vein training samples, and m = (m0) + (m1), the background weight w m0 =m1 / m, weight w of lithium-bearing pegmatite vein 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 gt i ,y pre i ) is the loss function of training sample i; Determine an optimization function that is compatible with the fully convolutional neural network, the optimization function comprising: Where w is the weight parameter, := means that the value on the left is updated or replaced by the value on the 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, and v dw is the momentum exponential weighted average of dw, s dw for dw 2 Momentum exponential weighted average, ε is a constant, and t represents the number of iterations; Execute a weight parameter initialization operation in the fully convolutional neural network; 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 for back propagation to update the weight parameters in the fully convolutional neural network; Inputting the verification data set into the trained fully convolutional neural network, and using a forward propagation method to evaluate the quality of the formed prediction model according to pre-selected verification evaluation indicators; Repeat the iterative training and validation steps until the loss function value and validation evaluation index meet the requirements.

7. The pegmatite-type lithium ore prediction method based on deep learning according to claim 1 or 6, characterized in that: The fully convolutional neural network is obtained in the following manner: Based on the spatial information and prospecting information of the prediction element graph, determining the network main structure, wherein the network main structure has a coding-decoding main structure with a symmetrical structure; Setting a coding structure of the main network architecture, the coding structure includes a type I convolution layer and a downsampling structure; and the execution logic of the coding structure is: performing convolution calculation through a type I convolution layer, and sequentially performing batch normalization, a type I activation function, and an attention mechanism module to obtain a first prediction element graph with different heights, widths, and feature numbers; 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 II convolution layers, and batch normalization, type I activation function and attention mechanism module are performed in sequence to obtain a prediction result graph.

8. The method for predicting pegmatite-type lithium ore according to claim 7, characterized in that: The activation function of the type I convolutional layer is: PReLU(x)=max(0,x)+ηmin(0,x) Among them, x is the value of the activation function input after the prediction feature graph 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(x 1,y1 ) is a known point (x 1,y1 ), z(x 2,y1 ) is a known point (x 2,y1 ), z(x 1,y2 ) is a known point (x 1,y2 ), z(x 2,y2 ) is a known point (x 2,y2 ) is a known value.

9. The method for predicting pegmatite-type lithium ore based on deep learning according to claim 1, characterized in that: The method of finding out the distribution of pegmatite-type lithium ore in the selected area to be mined based on the prediction model includes: Dividing the area to be collected into various sub-areas; Input each sub-region into the prediction model to generate a prediction result map for each sub-region; The prediction result maps are spliced ​​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, mineralization degree and surface occurrence information of the pegmatite-type veins.

10. A pegmatite-type lithium ore prediction device based on deep learning, characterized in that: include: A data acquisition unit acquires a plurality of different types of data in 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; A 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; A data generation unit generates a training data set and a verification data set based on a prediction element map, a mask map, and a prediction label map of a target area, wherein the target area is an area in which a pegmatite-type lithium ore body has been explored in the area to be collected; A processing unit, using 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, the distribution of pegmatite-type lithium deposits in the selected area to be mined is determined.

Images