Mineral resource potential assessment method based on remote sensing and deep learning

Abstract

The present invention discloses a method for evaluating the potential of mineral resources based on remote sensing and deep learning, comprising the following steps: drawing a mineral distribution layer, a lithology structure layer and an alteration layer; using geological metallogenic theory to determine the weight of each layer, using random forest to adjust the weight of each layer, and superimposing each layer to obtain a comprehensive map; based on the comprehensive map, using a self-developed deep learning model to perform image classification to obtain a prospecting target area. The method for evaluating the potential of mineral resources based on remote sensing and deep learning disclosed by the present invention can successfully identify multiple areas with potential mineral resources and is consistent with the results of on-site exploration; it can effectively reduce the blindness of exploration, optimize the positioning of prospecting targets, and greatly improve the success rate of exploration work.

Description

Technical Field

[0001] The present invention relates to a method for evaluating mineral resource potential based on remote sensing and deep learning, and belongs to the technical field of mineral exploration. Background Art

[0002] Traditional mineral resource assessment methods mainly rely on geological surveys and geophysical exploration, but these methods are often costly and inefficient.

[0003] Remote sensing technology can obtain large-scale, high-resolution surface information, providing a data basis for complex exploration, and can be combined with deep learning technology for mineral resource exploration.

[0004] Although existing technologies have proposed the use of remote sensing technology and deep learning technology for mineral resource exploration, these methods still have problems such as low precision and low accuracy in mineral resource assessment.

[0005] Therefore, it is necessary to conduct more in-depth research on mineral resource potential assessment methods to solve the above problems. Summary of the invention

[0006] In order to overcome the above problems, in-depth research was conducted and a method for evaluating mineral resource potential based on remote sensing and deep learning was designed, which includes the following steps:

[0007] Draw mineral distribution layers, lithology and structure layers, and alteration layers;

[0008] Determine the preliminary weights of each layer, and optimize and adjust the preliminary weights using a random forest algorithm;

[0009] Treat each layer as a channel and fuse the data of each layer to generate a comprehensive map by multi-channel superposition;

[0010] Based on the comprehensive map, a deep learning model is used to delineate the target area.

[0011] In a preferred embodiment, the mineral distribution layer, lithology structure layer and alteration layer are all obtained based on remote sensing images.

[0012] In a preferred embodiment, the remote sensing image is preprocessed before being used.

[0013] In a preferred embodiment, the pretreatment comprises:

[0014] Image correction: geometric correction and radiation correction of remote sensing images;

[0015] Image fusion: fusion of images with different resolutions;

[0016] Noise removal: filtering to remove noise from remote sensing images;

[0017] Feature extraction: Extract geological features related to mineral resources through image processing technology.

[0018] In a preferred embodiment, before determining the preliminary weights of each layer, the process also includes drawing a hydroxyl dyeing layer.

[0019] In a preferred embodiment, during the superposition process, different layers are determined to have different weights according to metallogenic theory.

[0020] In a preferred embodiment, the comprehensive graph is obtained by the following steps:

[0021] According to the mineralization theory, set the preliminary weight of each layer;

[0022] The random forest method is used to adjust the weights of each layer;

[0023] The layers are superimposed with the adjusted weights to obtain a comprehensive picture.

[0024] The deep learning model is a MINER_CNNET model, which is based on a convolutional neural network.

[0025] In a preferred embodiment, the input of the MINER_CNNET model also includes remote sensing images.

[0026] In a preferred embodiment, the MINER_CNNET model includes an autoencoder and a convolutional neural network. The autoencoder is used to extract features from remote sensing images, obtain remote sensing low-latitude features as the output of the autoencoder, superimpose the autoencoder output with the comprehensive map, and use the superimposed result as the input of the CNN network, which outputs potential mineralization areas.

[0027] The beneficial effects of the present invention include:

[0028] (1) Ability to successfully identify multiple areas with potential mineral resources and be consistent with field exploration results;

[0029] (2) It can automatically and precisely locate possible mineralization zones, especially in concealed ore bodies or newly discovered areas;

[0030] (3) It can effectively reduce the blindness of exploration, optimize the positioning of prospecting targets, and greatly improve the success rate of exploration work. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 A flow chart of a method for evaluating mineral resource potential based on remote sensing and deep learning according to a preferred embodiment of the present invention is shown;

[0032] Figure 2 The results of identifying potential mineral resource areas in Example 1 are shown. DETAILED DESCRIPTION

[0033] The present invention will be further described in detail below through the accompanying drawings and embodiments. Through these descriptions, the characteristics and advantages of the present invention will become more clear and distinct.

[0034] The word "exemplary" is used exclusively herein to mean "serving as an example, embodiment, or illustration." Any embodiment described herein as "exemplary" is not necessarily to be construed as preferred or advantageous over other embodiments. Although various aspects of the embodiments are shown in the drawings, the drawings are not necessarily drawn to scale unless otherwise noted.

[0035] According to the method for evaluating mineral resource potential based on remote sensing and deep learning provided by the present invention, Figure 1 As shown, the following steps are included:

[0036] Draw mineral distribution layers, lithology and structure layers, and alteration layers;

[0037] Determine the preliminary weights of each layer, and optimize and adjust the preliminary weights using a random forest algorithm;

[0038] Treat each layer as a channel and fuse the data of each layer to generate a comprehensive map by multi-channel superposition;

[0039] Based on the comprehensive map, a deep learning model is used to delineate the target area.

[0040] Mineral distribution layers, lithology and structure layers, and alteration layers are all commonly used layers in geological analysis, and those skilled in the art can draw them according to actual needs and based on experience.

[0041] In a preferred embodiment, the mineral distribution layer, lithology structure layer and alteration layer are all obtained based on remote sensing images.

[0042] Remote sensing images, especially those combined with multispectral and hyperspectral imaging technologies, can provide detailed analysis of the geological characteristics of mining areas at a macro scale.

[0043] Among them, the mineral distribution layer is a layer that identifies the types and distribution of surface minerals. The mineral distribution layer can intuitively display the distribution pattern of mineral combinations in the mining area. Especially in the identification of mineralized zones, the mineral distribution layer can often provide key clues for subsequent geological exploration work.

[0044] In a preferred embodiment, based on remote sensing image analysis, those skilled in the art can combine field survey data to draw a mineral distribution map.

[0045] The lithologic structure layer is used to identify the structural features within the mining area, which can help predict the trend and depth distribution of the ore body and infer possible mineralization control factors.

[0046] The alteration layer is a layer for identifying alteration characteristics. Through analysis of remote sensing images, it can quickly and widely identify alteration zones formed by typical altered minerals, such as sericite, chlorite, limonite, etc., to assist in the exploration of mineral resources.

[0047] In a preferred embodiment, the remote sensing image is preprocessed before being used.

[0048] The pre-processing comprises:

[0049] Image correction: geometric correction and radiation correction are performed on remote sensing images to improve the accuracy of the images;

[0050] Image fusion: fuse images of different resolutions to obtain more comprehensive surface information;

[0051] Noise removal: filtering removes noise from remote sensing images to improve data quality;

[0052] Feature extraction: extracting geological features related to mineral resources through image processing technology. The image processing technology can be any existing image processing method and is not limited in the present invention.

[0053] In a preferred embodiment, a hydroxy-dyed layer is also applied.

[0054] The hydroxyl dyeing layer is a layer for marking iron dyeing and hydroxyl groups.

[0055] Abnormal areas of iron-stained and hydroxyl minerals are usually closely related to hydrothermal alteration zones, which are important indicators of mineralization. Iron-stained minerals (such as limonite and hematite) and hydroxyl minerals (such as sericite and chlorite) are usually formed in alteration environments under oxidizing conditions, especially under the action of hydrothermal fluids during the mineralization process. Therefore, hydroxyl-stained maps can help identify the location and intensity of mineralized alteration zones, indicating areas where deep ore bodies may exist.

[0056] Preferably, the hydroxyl-dyed layer is obtained by analyzing remote sensing images.

[0057] According to the present invention, the mineral distribution layer, lithology layer, alteration layer and classification layer can all represent the mineralized target area at a certain angle. However, since each layer represents different contents, how to combine these layers to obtain more accurate positioning is the difficulty of the present invention.

[0058] In a preferred embodiment, the layer also includes a remote sensing image.

[0059] In a preferred embodiment, the layer also includes geological images or geological related images to further improve accuracy.

[0060] In a preferred embodiment, the preliminary weight is determined based on metallogenic theory and is set as:

[0061] Mineral distribution layer weight: 20%-25%;

[0062] Lithology and structure layer weight: 30%-35%;

[0063] Alteration layer weight: 25%-30%;

[0064] The weight of the hydroxy dye layer is 15%-20%.

[0065] The above weights are based on metallogenic theory and combined with metallogenic control factors, and are obtained through empirical summary.

[0066] The mineral distribution layer can identify the distribution of surface minerals, especially secondary minerals or minerals related to alteration (such as oxides and hydroxyl minerals). It can indicate possible alteration zones and mineralization zones and help determine secondary signs of prospecting. However, the mineral distribution layer mainly provides superficial information, so its role in deep ore body prediction is relatively limited and its weight is slightly lower.

[0067] The lithology and structure layer can reveal the distribution of parent rocks and ore-controlling structures related to mineralization. Since structure and lithology are the basis for controlling mineralization, their weight should be the highest. Through the lithology and structure layer, the potential favorable ore body distribution area can be delineated;

[0068] Hydrothermal alteration zones are usually closely related to ore-forming fluid activities, especially minerals related to nickel-copper sulfide mineralization (such as chlorite, sericite, etc.). Alteration layers can reflect the scope and intensity of hydrothermal activities and help determine the alteration zones related to mineralization. Therefore, the weight of alteration layers should be high when predicting potential mineralization zones.

[0069] Iron staining and hydroxyl mineral anomalies are often related to superficial oxidation zones and alteration. Although iron staining and hydroxyl mineral anomalies are not always direct signs of mineralization, they can indicate potential alteration zones and mineralized areas. Since hydroxyl-stained layers are usually more suitable for the identification of surface oxidation and shallow alteration zones, their weight can be appropriate, but lower than lithology and alteration maps.

[0070] In a preferred embodiment, the preliminary weights are further fine-tuned, specifically:

[0071] When searching for deep concealed ore bodies, increase the weight of lithology and structure layers and alteration layers, because deep mineralization is more dependent on the indications of structure and alteration.

[0072] When a layer corresponds to a higher resolution remote sensing image, the weight of the layer is increased.

[0073] The random forest method is an ensemble learning algorithm based on decision trees. It performs classification or regression tasks by combining multiple decision trees. It is a commonly used optimization method. In the present invention, its specific process is not described in detail.

[0074] In the present invention, multi-source data are trained by random forest method, and weights are adjusted according to the characteristics of the actual mineralization area to improve the accuracy of model prediction.

[0075] According to the present invention, the comprehensive map covers multi-source features, including geological structures (such as faults, folds, etc.), lithology distribution, and mineral spectral curves (reflecting the spatial distribution of minerals). The comprehensive input of these multi-source features provides an important basis for mineralization prediction for the deep learning model.

[0076] Among them, geological structure lines: geological structure lines extracted from remote sensing images represent important geological structures such as faults and folds. These structures are usually closely related to the formation and distribution of ore bodies and are one of the important input features of the model;

[0077] Lithology distribution: The distribution of different lithologies plays a decisive role in the existence of ore bodies. By combining remote sensing images with geological data, the model can identify various lithology areas and provide stratigraphic environment information for mineralization prediction;

[0078] Mineral spectral curve: The mineral spectral curve reflects the unique spectral characteristics of the mineral. The mineral distribution map can be obtained through remote sensing images. These distribution maps provide mineral composition information for the model and are an important basis for delineating prospecting target areas.

[0079] In a preferred embodiment, the comprehensive map is also marked with known mineral point data, and the known mineral point locations are provided to the model as label data for supervised learning. The data of these known mineral points can help the model identify potential mineralized areas during the training process.

[0080] In a preferred embodiment, in addition to using known mining area data for annotation, the label data is also obtained in the following manner: for copper-nickel ore, the mine site is identified by identifying iron staining and hydroxyl anomalies, and the identification of iron staining and hydroxyl anomalies includes spectral band selection, image fusion and feature enhancement, and data comparison and analysis.

[0081] Preferably, mineral points can also be identified through structural control, including the extraction of fault structures, joint analysis of DEM and images, and association of fault-intensive areas with mineralization.

[0082] The spectral band selection, through the infrared bands (such as mid-wave infrared and near-infrared) in remote sensing images, identifies typical reflectance spectra associated with iron-stained minerals (such as limonite and hematite) and hydroxyl minerals (such as sericite and chlorite). The spectral characteristics of these minerals usually include strong absorption in the near-infrared band and characteristic reflection peaks in the mid-infrared band, which can help effectively distinguish different alteration minerals.

[0083] The image fusion and feature enhancement utilizes the fusion of multispectral images and hyperspectral images to combine the surface structure in the high-resolution image with the spectral information in the multispectral image to enhance the features of hydroxyl and iron-stained minerals, and uses principal component analysis (PCA) and other techniques to extract the main mineralization anomaly information, so as to more prominently and clearly observe the mineral distribution and alteration features of the target area;

[0084] The data comparison analysis extracts the hydroxyl and iron-stained mineral reflectance features in the image and compares them with the spectral features of known mineralized areas to verify the mineralization potential in the new area. For example, the spectral angle matching method (SAM) is used to compare the spectra of different areas to confirm which areas have similar mineralization environment characteristics.

[0085] The extraction of the fault structure uses linear features such as faults and folds identified in remote sensing images, combined with digital elevation models (DEMs) to conduct detailed three-dimensional terrain analysis, using multi-angle illumination direction synthesis (shaded relief) and slope analysis technology to reveal potential geological structures;

[0086] The DEM and image joint analysis further determines the location and trend of the fault zone through the joint application of DEM data and remote sensing images, and superimposes the low-lying or high-convex areas identified in the DEM with the linear features in the remote sensing images to locate possible mineralization structural areas;

[0087] The fault-intensive area is associated with mineralization. Spatial analysis tools such as buffer analysis are used to evaluate the positional relationship between the fault-intensive area and known mineral deposits to determine whether there is mineralization control. Special attention is paid to the mineralization and alteration phenomena near the fault zone, thereby delineating possible mineralized areas.

[0088] For gold mines, the mineral point identification is carried out by combining linear fault structures with altered mineral zones, including multi-scale analysis of fault structures and classification and superposition of altered zones. Preferably, the mineral point identification is also carried out by the spectral characteristics of altered minerals, including refined spectral matching of hyperspectral images and spectral interpretation based on machine learning.

[0089] Among them, the multi-scale analysis of the fault structure focuses on identifying the positional relationship between the fault structure and the altered mineral belt through the mid-wave infrared (SWIR) and near infrared bands (NIR). Gold deposits are often related to fault structures. Multi-scale analysis of fault structures is performed at different scales (macro, meso, and micro), especially the continuity and intersection of linear features. Therefore, these locations usually have higher prospecting potential.

[0090] The classification and superposition of the alteration zones mentioned above, spatially superimpose the alteration zone information of limonite and hematite with the fault structure, and identify possible alteration and structural composite areas. Apply the weighted overlay analysis in GIS to give higher weights to the intersection points of different alteration types and structures, thereby improving the accuracy of comprehensive prospecting potential assessment.

[0091] The refined spectral matching of the hyperspectral image uses the hyperspectral image to finely match the spectral characteristics of chlorite and quartz veins. The spectral characteristic index is used to evaluate the anomalies in the hyperspectral image, and the spectral characteristics are compared with the known ore deposits in combination with the actual mineralization model to delineate the mineralization target area.

[0092] The machine learning-based spectral interpretation introduces support vector machine (SVM) or random forest algorithm (RandomForest) to automatically classify altered minerals through labeled training data. The application of machine learning algorithms can more precisely identify mineralized alteration areas in hyperspectral data and improve the accuracy and reliability of identifying mineralization zones.

[0093] For uranium mines, mine site identification is carried out through the combination of hydrothermal alteration and radiation intensity anomalies, including remote sensing monitoring of radiation intensity and detailed analysis of hydrothermal alteration zones.

[0094] Preferably, the identification of the mineral points is also performed by identifying the oxidation zone, including extracting the spectral features of the oxidation zone and combining the oxidation zone with geological exploration data.

[0095] Among them, the remote sensing monitoring of radiation intensity uses remote sensing images to detect the radiation intensity of radon, focusing on analyzing the spatial distribution characteristics of surface radiation intensity. Abnormally high values ​​of radiation intensity may indicate uranium mineralization activity below the surface. The radiation anomaly is superimposed and analyzed with the hydrothermal alteration zone, and the area where the high-value radiation area overlaps with the alteration zone is circled.

[0096] The oxidation zones and geological exploration data, hydrothermal alteration zones can be identified by abnormal spectral reflections in remote sensing images, such as the strong absorption bands of sericite and limonite. By analyzing the continuous distribution characteristics of these hydrothermal alteration zones, possible mineralization channel maps are drawn as a reference for subsequent detailed exploration.

[0097] The spectral feature extraction of the oxidation zone identifies the oxidized sedimentary areas containing limonite and hematite through the long-wave infrared band (LWIR) in remote sensing images. Uranium deposits are usually closely related to oxidation zones. Through hyperspectral analysis, the characteristic spectra of these oxidized minerals can be extracted to ensure the accuracy of mineral identification.

[0098] The oxidation zones are combined with geological exploration data and ground exploration data, such as soil geochemical analysis results, to evaluate the mineralization potential of the oxidation zones. Chemical anomalies (such as high content of oxides) in the oxidation zones are cross-validated with anomaly areas identified by remote sensing to determine the prospecting potential of the oxidation zones, and the spatial distribution of these oxidation zones is estimated using geostatistical methods (such as Kriging interpolation) to determine their relationship with possible mineralization.

[0099] In a preferred embodiment, the deep learning model adopts the self-developed MINER_CNNET model, which is a deep learning model based on convolutional neural network (CNN) and is specially designed for remote sensing geological prospecting. The model combines geological structure lines, lithology information, mineral spectral curves and known mineral point information in the comprehensive map, and automatically identifies important geological features through deep learning, thereby realizing the delineation of prospecting target areas. This method can identify subtle geological features that are difficult to detect by traditional remote sensing interpretation, such as weak reflection signals of mineralized zones, complex lithology transition zones or hidden structural features.

[0100] According to the present invention, the input of the MINER_CNNET model includes a comprehensive map, and preferably, also includes a remote sensing image.

[0101] Preferably, the MINER_CNNET model includes an autoencoder and a CNN network.

[0102] The autoencoder is used to extract features from remote sensing images, achieve feature dimension reduction and denoising, and obtain remote sensing low-latitude features as output of the autoencoder;

[0103] Furthermore, the autoencoder includes multiple convolutional layers, through which data and dimensions are gradually reduced to extract and retain key features in remote sensing images. Through the autoencoder, not only can the noise in the data be effectively removed, but also the computational complexity of the model can be reduced, making the subsequent deep learning process more efficient. In addition, the extracted low-dimensional features are easier to understand and visualize, allowing geological experts to more intuitively understand which features play a key role in prospecting prediction. This feature dimensionality reduction can not only increase the training speed of the model, but also enhance the generalization ability of the model in different geological environments, avoiding overfitting caused by data redundancy.

[0104] Further preferably, the input of the MINER_CNNET model also includes one or more of hyperspectral images, digital elevation models (DEMs) and geological layers, which are respectively subjected to feature extraction and feature dimensionality reduction through autoencoders, and the dimensionality reduction results are fused with remote sensing low-latitude features as autoencoder outputs.

[0105] According to a preferred embodiment of the present invention, the autoencoder output is superimposed on the comprehensive map, and the superimposed result is used as the input of the CNN network, and the CNN network outputs the potential mineralization area.

[0106] Superimposing the autoencoder output with the comprehensive map effectively combines the low-dimensional features with the mineral spectrum, geological structure and other data in the comprehensive map, so that the CNN network can focus on areas related to mineralization from the beginning of training. This method can not only reduce the training time of the model, but also improve the recognition accuracy.

[0107] In a preferred embodiment, an attention mechanism is introduced into the CNN network, and the feature map output by the autoencoder network is used as a reference to guide the CNN network to better focus on local features related to mineralization, avoid redundant calculations on irrelevant areas, and thus improve the prospecting effect of the model.

[0108] The CNN network and self-attention mechanism are commonly used structures in neural networks. In the present invention, their specific structures are not described in detail.

[0109] In a preferred embodiment, when the mineral resources to be evaluated are copper-nickel deposits, gold deposits and uranium deposits, the weight of the convolution kernel is set to be concentrated on the part that can capture the linear characteristics of the alteration zone and the fault, so that the model can more easily extract key geological information.

[0110] In a preferred embodiment, the convolution kernel size and stride are specifically designed according to the mineral resources to be evaluated. For example, for gold deposits, the convolution kernel design focuses on extracting fault structural lines and alteration zones so as to better capture these mineralization indicator features in the early convolution layer.

[0111] In a preferred embodiment, in the CNN network, the size of the convolution kernel and the pooling layer is dynamically adjusted. For example, for small but geologically significant details (such as alteration boundaries), a smaller convolution kernel can be used to ensure the complete extraction of spatial details. When dealing with large-scale geological structures (such as large fault zones), the convolution kernel size is increased to capture larger geographical area features.

[0112] In a preferred embodiment, in the CNN network, the pooling layer adopts a variable stride and a variable pooling window strategy. For fine features, such as alteration zones, a smaller pooling window and a smaller stride are used to maximize the retention of feature details. For broader geological features, a larger pooling window is used to simplify the features and reduce computational complexity.

[0113] In a preferred embodiment, the Bayesian optimization method is used to dynamically adjust the parameters in the CNN network, and the weight ratios of different features are dynamically adjusted by evaluating the impact of different features on the model output. For example, for copper-nickel deposits, the weights of alteration layers and structural layers are increased; for uranium deposits, the weights of radiation intensity anomalies and oxidation zones are increased. This weight adjustment can be continuously optimized in an iterative manner to improve the performance of the model in different deposit types.

[0114] In a preferred embodiment, during the training of the CNN network, gradient back propagation is used to calculate the contribution of different input features to the loss function, thereby adjusting the weight of each input feature. This approach ensures that in each step of model training, the input features most relevant to mineral prospecting are focused on.

[0115] In a preferred embodiment, when the geological conditions in the area to be evaluated are complex, the number of convolution layers of the CNN network is increased so that the model can learn more complex high-order features. For example, for areas with complex mineralization backgrounds, the number of convolution layers can reach more than 10 layers to ensure that the diversity and depth of geological features are fully captured. For simpler geological backgrounds (such as a small number of fault structures and stable rock masses), the network depth is reduced to avoid overfitting of features.

[0116] In a preferred embodiment, when the geological features of the region to be evaluated are relatively simple (such as near a known mineralization belt), the model's ability to capture overall features is enhanced by increasing the width of the CNN network, especially by adding more convolution kernels in the lower-level convolution layers to ensure that the model has a better capture of the global mineralization trend. At the same time, the design of a multi-branch structure (such as the jump connection in ResNet) can avoid information loss caused by network widening.

[0117] According to the present invention, the final output of MINER_CNNET is a classification map, in which each pixel is divided into different categories to identify its possible mineralization potential area. This prospecting target area map can provide geologists with an intuitive prospecting reference.

[0118] MINER_CNNET extracts the spatial features of input data through the convolution layer, reduces the data dimension through the pooling layer, and processes the features comprehensively through the fully connected layer, and finally outputs a mineralization target map to identify areas with potential mineralization. The model integrates multi-source information and can effectively capture deep geological features, thereby improving the accuracy of prospecting.

[0119] By integrating different geological information in remote sensing images, MINER_CNNET can not only automatically learn geological features from complex data, but also improve the accuracy of the model by using known mineral point data. This enables the model to provide intuitive guidance for geological prospecting, significantly improve prospecting efficiency, and reduce risks and costs. Example

[0120] Example 1

[0121] The mineral resource potential of a certain area is evaluated. The area is nickel-copper sulfide deposit with complex mineralization background. The mineral resource potential is evaluated through the following steps:

[0122] Draw mineral distribution layers, lithology and structure layers, alteration layers and hydroxyl dyeing layers;

[0123] Use the random forest algorithm to adjust the weights of each layer;

[0124] Each layer is superimposed to obtain a comprehensive picture;

[0125] Based on the comprehensive map, the self-developed MINER_CNNET deep learning model is used to delineate the prospecting target area.

[0126] Among them, the alteration zone obtained from the alteration layer is located near the fault structure and extends to the known mineralized area. Through in-depth research on these alteration zones, it is speculated that they may be the channels of deep mineralization hydrothermal fluids. A series of hidden ore bodies have been discovered through drilling verification, indicating that the application of alteration markers in ore prospecting can effectively indicate deep mineralized zones;

[0127] Based on the lithology and structure map, we identified the obvious spectral difference between the rock mass and the surrounding rocks, and inferred that the region may have similar mineralization conditions as the known ore bodies. Subsequent geological drilling confirmed this speculation and discovered new mineralized zones, indicating that lithology and structure maps play an important role in delineating ore-controlling rock masses and predicting mineralized zones.

[0128] Several possible oxidation zone locations were identified based on the hydroxyl-stained layer. Subsequent field verification and geochemical sampling work showed that these abnormal areas corresponded to the outcropping areas of shallow mineralized zones. Further geological surveys confirmed the existence of high-grade nickel-copper ore bodies beneath these areas.

[0129] Among them, the self-developed MINER_CNNET deep learning model, the comprehensive graph is obtained through the following steps:

[0130] Set the initial weights of each layer;

[0131] The random forest method is used to adjust the weights of each layer;

[0132] The layers are superimposed with the adjusted weights to obtain a comprehensive picture.

[0133] The initial weights are set to:

[0134] Mineral distribution layer weight: 10%;

[0135] Lithology and structure layer weight: 35%;

[0136] Alteration layer weight: 30%;

[0137] The weight of the hydroxy dye layer is 15%.

[0138] The deep learning model adopts the MINER_CNNET model. The input of the MINER_CNNET model includes a comprehensive map and a remote sensing image. The autoencoder is used to extract features of the remote sensing image, realize feature dimensionality reduction and denoising, and obtain remote sensing low-dimensional features as the output of the autoencoder. The autoencoder includes multiple convolutional layers, which gradually reduce data and dimensions through multiple convolutional layers to extract and retain key features in the remote sensing image.

[0139] The autoencoder output is superimposed with the comprehensive map, and the superimposed result is used as the input of the CNN network, which outputs the potential mineralization area.

[0140] Through the above methods, several areas with potential mineral resources were successfully identified, such as Figure 2 As shown in the figure, it is consistent with the results of on-site exploration, and a previously unrecorded large fault zone has been successfully identified. The strike of the fault zone coincides with the strike of the known ore body, and it shows obvious tensile properties in geology. It is speculated that it may be one of the main channels of ore-forming fluids. Subsequent geological surveys and drilling verifications have found multiple hidden ore bodies near the fault zone, thus expanding the prospecting scope of the mining area. Through the analysis of the comprehensive map, some weak but potential mineralization-significant abnormal areas were revealed. According to the guidance of these characteristic maps, detailed geophysical exploration and drilling were carried out, and new mineralized zones were successfully discovered, which significantly improved the efficiency and accuracy of prospecting.

[0141] In the description of the present invention, it should be noted that the terms "upper", "lower", "inner", "outer", "front", "rear", etc. indicate positions or positional relationships based on the working state of the present invention, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", "third", and "fourth" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0142] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected" and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0143] The present invention has been described above in conjunction with preferred embodiments, but these embodiments are only exemplary and serve only as an illustration. On this basis, the present invention may be subjected to a variety of substitutions and improvements, all of which fall within the scope of protection of the present invention.

Claims

1. A method for evaluating mineral resource potential based on remote sensing and deep learning, characterized in that: The following steps are involved: Draw mineral distribution layers, lithology and structure layers, and alteration layers; Determine the preliminary weights of each layer, and optimize and adjust the preliminary weights using a random forest algorithm; Treat each layer as a channel and fuse the data of each layer to generate a comprehensive map by multi-channel superposition; Based on the comprehensive map, a deep learning model is used to delineate the target area; The deep learning model is a MINER_CNNET model, which is based on a convolutional neural network; The input of the MINER_CNNET model also includes remote sensing images; The MINER_CNNET model includes an autoencoder and a convolutional neural network. The autoencoder is used to extract features from remote sensing images, obtain remote sensing low-latitude features as the output of the autoencoder, superimpose the autoencoder output with the comprehensive map, and use the superimposed result as the input of the CNN network, which outputs potential mineralization areas.

2. The method for evaluating mineral resource potential based on remote sensing and deep learning according to claim 1, characterized in that: The mineral distribution layer, lithology structure layer and alteration layer are all obtained based on remote sensing images.

3. The method for evaluating mineral resource potential based on remote sensing and deep learning according to claim 1, characterized in that: Before using the remote sensing images, they are also preprocessed.

4. The method for evaluating mineral resource potential based on remote sensing and deep learning according to claim 3 is characterized in that: The pre-processing comprises: Image correction: geometric correction and radiation correction of remote sensing images; Image fusion: fusion of images with different resolutions; Noise removal: filtering to remove noise from remote sensing images; Feature extraction: Extract geological features related to mineral resources through image processing technology.

5. The method for evaluating mineral resource potential based on remote sensing and deep learning according to claim 1, characterized in that: Before determining the preliminary weights of each layer, it also includes drawing the hydroxyl dye layer.

6. The method for evaluating mineral resource potential based on remote sensing and deep learning according to claim 1, characterized in that: During the superposition process, the weights of different layers are determined according to the mineralization theory.

7. The method for evaluating mineral resource potential based on remote sensing and deep learning according to claim 1, characterized in that: The comprehensive graph is obtained by following the steps below: Set the initial weights of each layer according to the metallogenic theory; The random forest method is used to adjust the weights of each layer; The layers are superimposed with the adjusted weights to obtain a comprehensive picture.

Images