A method for estimating mineral resource reserves by using a fractal algorithm

Abstract

The application discloses a kind of mineral resources reserves estimation method using fractal algorithm, step (S1): resource distribution map generation based on machine learning;Step (S2): feature matching mine site identification based on fractal geometry;Step (S3): reserve estimation based on regression analysis.The application is a new mineral resources reserves estimation calculation scheme used in the field of intelligent technology of mineral exploration.The scheme uses the "fractal" method in mathematics, combined with AI technology, to provide a complete solution for resource reserves calculation of target mine, especially weathered bauxite, weathered laterite nickel ore and other types of near-surface mine, to achieve the task goal of calculating the mineral resources reserves of target mining area without field exploration data or only using a small amount of artificial exploration data.

Description

Technical Field

[0001] This invention relates to the field of mineral exploration technology, specifically to a method for estimating mineral resource reserves using fractal algorithms. Background Technology

[0002] Mineral resource reserve calculation is a core technical step in geological exploration and mining development, and its results directly determine the economic feasibility of mining projects and the design of development plans. Currently, reserve calculation methods in the industry are undergoing three major paradigm shifts: from experience-based qualitative estimation to data-driven quantitative modeling; from static reserve snapshot assessment to dynamic full life cycle response; and from isolated analysis of single parameters to the integration of uncertainties from multi-source data.

[0003] In recent years, researchers both at home and abroad have been continuously tackling key challenges in three core areas: expanding parameter dimensions, deeply integrating data sources, and strengthening uncertainty quantification. This has resulted in three representative technical paths that together form the technological foundation of the field of storage computation: Multi-parameter collaborative inversion and cluster analysis: Breaking the dilemma of multiple solutions in single-parameter analysis This approach focuses on the complementarity of multi-source geophysical parameters and uses clustering algorithms to objectively classify mineral deposit traps. For example, patent CN 119357729 A couples magnetotelluric inversion and transient electromagnetic inversion data, and introduces the K-means clustering algorithm to automatically divide the probe area into mineral deposit areas, porosity areas, boundary areas, and surrounding rock areas, thus mitigating the ambiguity of a single resistivity parameter. Patent CN 114612258 A introduces Monte Carlo stochastic simulation into the reserve classification system, constructing a two-dimensional intersection map of "mineral reserves - occurrence probability" to replace the traditional empirical reduction factor method, thereby improving the objectivity of reserve level conversion.

[0004] Geochemical element-driven conditional modeling: Achieving adaptive structural analysis of orebody strike. This approach embeds prior knowledge of mineral deposit geology into spatial statistical models, enabling precise analysis of orebody morphology under complex geological structures. For example, patent CN 119864105 B identifies ore-forming target areas through elemental symbiosis analysis, modifies the Kriging method with segmented anisotropic interpolation, and introduces a dynamic block method and PID error feedback mechanism to achieve online self-correction of the reserve model; patent CN 110008596 B, based on the dominant wavelength theory of rock mechanics, establishes an arc length inversion model to correct the difference between the arc length of the ore layer and the surface projection, which is suitable for rapid exploration of folded and layered ore bodies.

[0005] Deep Coupling of Artificial Intelligence and Geostatistics: Constructing a Highly Robust Uncertainty Quantification Framework This approach leverages the strong fitting capabilities of AI models and the rigor of geostatistical methods to achieve high-precision reserve estimation for complex ore bodies. For example, patent CN 120632355 A constructs a five-step closed-loop process of "PCA feature extraction - 3D - CNN morphological learning - Bayesian model fusion - Bayesian hyperparameter optimization - MCMC uncertainty quantification," outputting a reserve confidence interval with rigorous mathematical significance. Patent CN 115345657 A constructs a holographic three-dimensional reserve model, introducing a "price-boundary grade calculation model" to dynamically solve for economically recoverable reserves, representing a cutting-edge paradigm of digital mining.

[0006] Although existing technologies have entered a stage of systematic innovation, they still have many shortcomings and defects: 1. Rigid model assumptions: The model assumes that the geological parameters are uniform or simply distributed, which fails to effectively characterize the non-Gaussianity, spatial correlation and multi-scale heterogeneity of the geological parameters, and is not adaptable to complex geological conditions. 2. Equivalent logic is one-way: It only realizes "downward equivalence of proven reserves", lacks a two-way uncertainty transmission and verification mechanism between reserves of different levels, has a high dependence on ideal geological conditions, and has poor robustness; 3. Strong dependence on exploration data: The model construction is highly dependent on high-density, high-quality exploration data. In areas with low exploration levels, interpolation smoothing effects and edge distortion are prone to occur, and the uncertainty is difficult to quantify. 4. Static boundary grade: A globally uniform boundary grade is adopted, without considering the dynamic changes in local boundary grade caused by factors such as grade variation within the ore body, mining sequence, and differences in beneficiation processes; 5. Manual Modeling Process: The construction and correction of 3D models rely heavily on manual intervention, resulting in low levels of automation and intelligence, lengthy processes, and difficulty in adapting to small and medium-sized mines or rapid evaluation projects. 6. Decision-making process is based on experience: The final reserve results still require comprehensive human judgment, and an automated, multi-round, and progressively approximating intelligent estimation system has not yet been formed.

[0007] Traditional reserve calculation methods are inherently limited by high costs and long cycles. They rely on specialized equipment and require large field teams for on-site verification, making it difficult to meet the market's demand for rapid, low-cost, and highly reliable mineral resource potential assessment in the context of globalized mineral resource allocation. Therefore, it is urgent to overcome the bottlenecks of existing methods and construct a more adaptable, efficient, and intelligent mineral resource reserve calculation system. Summary of the Invention

[0008] Addressing the bottleneck of "high cost and low timeliness" caused by reliance on on-site inspections in existing technologies, and systematically overcoming five key defects of existing technologies in terms of model adaptability, data dependence, and computational efficiency, the present invention aims to provide a mineral resource reserve prediction method using fractal algorithms, providing a fast, low-cost, and highly reliable intelligent assessment framework for mineral resource exploration.

[0009] To achieve the above objectives, the present invention provides a method for estimating mineral resource reserves using fractal algorithms, comprising the following steps: Step S1: Generation of resource distribution map based on machine learning A classification model is trained using satellite remote sensing hyperspectral data of known mines to perform pixel-level predictions of the target area and output a resource abundance distribution map of the target mineral. Step S2: Feature matching mine identification based on fractal geometry Extract feature parameters such as fractal dimension and multifractal spectrum from the resource distribution map obtained in step S1. Combine with auxiliary data such as geological structure and topography of the target area, calculate the fractal feature similarity between the target area and known mines. Select a batch of mines with similar mineralization environment from the list of mines with known reserves as the data source for subsequent reserve calculation. Step S3: Reserve Estimation Based on Regression Analysis Using the similar mines matched in step (S2) as samples, a regression model is established using the actual reserve data of the same target mineral in known mines and remote sensing / geological features to quantitatively estimate the total resource reserves of this type of mineral in the target mine.

[0010] Step S1 specifically includes: (1.1) Data Acquisition and Preprocessing Acquire satellite hyperspectral remote sensing images of the target area, including dozens of bands, covering visible light, near infrared, and shortwave infrared.

[0011] The images are radiometrically calibrated, atmospherically corrected, and geometrically finely corrected to eliminate environmental interference.

[0012] Mineral index and vegetation cover index were extracted as auxiliary features.

[0013] (1.2) Sample construction and model training Collect known mining areas of this type of mineral, mark the ore body range on remote sensing images, and extract the pixel spectrum within the range as positive samples.

[0014] Background pixels were randomly selected from the non-mineralized area as negative samples.

[0015] Construct a multidimensional feature vector, including original band reflectance, band ratio, principal component analysis dimensionality reduction components, and texture features.

[0016] Supervised learning was conducted using convolutional neural networks, support vector machines, and random forest algorithms to train a mineral probability prediction model.

[0017] The algorithm in step 1.2 can also use a lightweight CNN to automatically learn spectral-spatial features.

[0018] When the sample size is small, step 1.2 uses SVM or ensemble learning (XGBoost) supplemented by feature selection.

[0019] (1.3) Resource Abundance Prediction The trained model is applied to the whole image for pixel-by-pixel prediction, and the probability or abundance value of each pixel belonging to the target mineral type is output.

[0020] Spatial filtering is applied to the prediction results to remove salt-and-pepper noise, generating a continuous resource distribution probability map.

[0021] Step S2 specifically includes: (2.1) Delineation of the core area of ​​the target mining area The location of the target mining area is found on the resource distribution map obtained in step S1, the area with high probability of resource quantity is delineated, and the area for subsequent fractal calculation is finally selected based on the geological and geomorphological characteristics of the area.

[0022] The selected computational region is divided into one or more sub-regions. For mines with large areas and multiple different ore-controlling patterns, sub-regions are divided according to the different "ore-controlling patterns".

[0023] When the selected sub-region cannot be well matched with the sample mine group with known reserves after subsequent "similarity measurement and matching", the sub-region is re-divided and its similarity with the sample mine group is measured again. After multiple trials, the scheme with the highest similarity is selected to finally complete the matching.

[0024] (2.2) Fractal Feature Extraction The fractal dimension (D0) of the resource distribution prediction map obtained in step S1 is calculated using the profile line method and Fourier spectral analysis to characterize the complexity and self-similarity of the resource distribution.

[0025] Further calculations of the multifractal spectrum (f(α) curve) are performed to extract spectral width and maximum point parameters, providing a more detailed description of the non-uniformity of the distribution.

[0026] Simultaneously, geological auxiliary data of the target area and known mining areas are collected, and their fractal dimension (Di) and the angle αi between their directions and the fractal characteristics D0 of the resource distribution prediction map are calculated. For example: Lithological structures: fracture density D1, normal distance of fold axis orientation D2; Topography and landforms: surface texture features D3, elevation variation coefficient D4, slope D5, river network fractal dimension D6; Geophysical / geochemical anomalies: D7, D8… are calculated separately according to the numerical distribution of each monitoring item; The fractal parameters of the aforementioned auxiliary data are also calculated to form a multidimensional fractal eigenvector: D = (D0, D1, α1, D2, α2, D3, α3, D4, α4, D5, α5,…) (2.3) Similarity measurement and matching Using the fractal feature vector D of the known mine as a template, the corresponding feature vector of the unknown area is extracted by sliding window or by geological unit.

[0027] Euclidean distance, cosine similarity, and Mahalanobis distance are calculated as similarity criteria, and thresholds are set to filter out a list of "candidate blocks" with highly similar fractal features.

[0028] If a candidate block is located in the same metallogenic belt or has the same stratigraphic background as a known mine, it is determined to be a target mine of the same metallogenic environment. This rule is applied to filter the list of "candidate blocks", and the "candidate blocks" that meet the requirements and have the highest similarity are used as the data pool for subsequent reserve regression analysis.

[0029] Step S3 specifically includes: (3.1) Sample set construction From the known mining areas of the same type selected in step S2, several "candidate blocks" are chosen as training samples. These samples all come from mining areas with reliable reserve data.

[0030] Extract a feature vector X for each sample mine, including: Statistics of the resource distribution map obtained in step S1; The fractal features obtained from step S2; Fractal characteristics of auxiliary data such as geological and topographical data; The actual explored reserves of the corresponding mining area are used as label Y.

[0031] (3.2) Regression Modeling Construct a regression model to learn the mapping relationship between feature X and reserves Y.

[0032] Step 3.2 can also employ multiple linear regression or support vector regression.

[0033] In step 3.2, when the feature dimension is high, LASSO (Least Absolute Shrinkage and Selection Operator) is used to prevent overfitting.

[0034] Perform cross-validation on the model and select the optimal hyperparameters.

[0035] (3.3) Estimation of reserves in the target mining area Extract the corresponding feature vector X' of the target candidate mining area, input it into the trained regression model, and obtain the predicted reserve value.

[0036] Output the confidence interval of the prediction result; The geological rationality of the estimation results was verified by combining the regional metallogenic regularity.

[0037] The beneficial effects of this invention are: 1. For the first time, deep learning spectral analysis and fractal geostatistics are integrated, enabling regional resource potential assessment without any field drilling data.

[0038] 2. By matching fractal features, unknown areas similar to the mineralization environment of known mines can be automatically identified, solving the subjective problem of traditional methods that rely on expert experience to delineate target areas.

[0039] 3. A complete quantitative estimation process was constructed, encompassing remote sensing signals, resource distribution, similarity of mineralized environments, and reserves. This enables the calculation of mineral resource reserves in a target mining area without the need for on-site exploration data, or with only a small amount of manual exploration data. Attached Figure Description

[0040] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments; Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0042] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0043] Reference Figure 1 The specific implementation adopts the following technical solution: a method for estimating mineral resource reserves using fractal algorithms, comprising the following steps: Step S1: Generation of resource distribution map based on machine learning A classification model is trained using satellite remote sensing hyperspectral data of known mines (positive samples) to perform pixel-level predictions of the target area and output a resource abundance distribution map of the target mineral.

[0044] Step S2: Feature matching mine identification based on fractal geometry Extract feature parameters such as fractal dimension and multifractal spectrum from the resource distribution map obtained in step S1. Combine with auxiliary data such as geological structure and topography of the target area, calculate the fractal feature similarity between the target area and known mines. Select a batch of mines with similar mineralization environment from the list of mines with known reserves as the data source for subsequent reserve calculation.

[0045] Step S3: Reserve Estimation Based on Regression Analysis Using the similar mines matched in step (S2) as samples, a regression model is established using the actual reserve data of the same target mineral in known mines and remote sensing / geological features to quantitatively estimate the total resource reserves of this type of mineral in the target mine.

[0046] Step S1 specifically includes: (1.1) Data Acquisition and Preprocessing Acquire satellite hyperspectral remote sensing images of the target area (such as Sentinel2, Lansat8, ASTER, etc.), which contain dozens of bands and cover visible light, near infrared, shortwave infrared, etc.

[0047] The images are radiometrically calibrated, atmospherically corrected, and geometrically finely corrected to eliminate environmental interference.

[0048] Mineral indices (such as alteration mineral information, iron staining index, hydroxyl index, etc.) and vegetation cover index are extracted as auxiliary features.

[0049] (1.2) Sample construction and model training Collect known mining areas of this type of mineral (such as laterite nickel ore and sedimentary bauxite), mark the ore body range on remote sensing images, and extract the pixel spectrum within the range as positive samples.

[0050] Background pixels were randomly selected from the non-mineralized area as negative samples.

[0051] Construct multidimensional feature vectors, including original band reflectance, band ratio, principal component analysis (PCA) dimensionality reduction components, texture features, etc.

[0052] Supervised learning was conducted using algorithms such as Convolutional Neural Network (CNN), Support Vector Machine (SVM), and Random Forest (RF) to train a mineral probability prediction model.

[0053] Preferred approach: Use lightweight CNNs (such as 1D-CNN for spectral curves or 2D-CNN for multi-band image patches) to automatically learn spectral-spatial features.

[0054] Alternative approach: When the sample size is small, use SVM or ensemble learning (XGBoost) supplemented by feature selection.

[0055] (1.3) Resource Abundance Prediction The trained model is applied to the whole image for pixel-by-pixel prediction, and the probability or abundance value of each pixel belonging to the target mineral type is output.

[0056] Spatial filtering (such as median filtering and mode filtering) is applied to the prediction results to remove salt-and-pepper noise and generate a continuous resource distribution probability map.

[0057] Step S2 specifically includes: (2.1) Delineation of the core area of ​​the target mining area The location of the target mining area is found on the resource distribution map obtained in step S1, the area with high probability of resource quantity is delineated, and the area for subsequent fractal calculation is finally selected based on the geological and geomorphological characteristics of the area.

[0058] The selected computational region is divided into one or more sub-regions. This operation is crucial for resource calculations in mines with large areas and multiple different structural ore-controlling patterns. Subsequently, the fractal characteristics of these sub-regions are extracted, resource reserves are calculated separately, and then they are merged to form the resource reserves of the entire target mining area.

[0059] When the selected sub-region cannot be well matched with the sample mining area with known reserves after subsequent "similarity measurement and matching" (for example, the number of matched mining areas is less than 10), it is necessary to re-divide the sub-region and measure its similarity with the sample mining area again. After multiple trial calculations, the matching is finally completed.

[0060] (2.2) Fractal Feature Extraction The fractal dimension (D0) of the resource distribution prediction map obtained in step S1 is calculated using methods such as the profile line method and Fourier spectrum analysis (power spectral density analysis) to characterize the complexity and self-similarity of the resource distribution.

[0061] Further calculations of the multifractal spectrum (f(α) curve) are performed to extract parameters such as spectral width and maximum points, providing a more detailed description of the non-uniformity of the distribution.

[0062] Simultaneously, geological auxiliary data of the target area and known mining areas are collected, and their fractal dimension (Di) and the angle αi between their directions and the fractal characteristics D0 of the resource distribution prediction map are calculated. Lithological structure (fracture density D1, normal distance of fold axis orientation D2) Topography and landforms (surface texture features D3, elevation variation coefficient D4, slope D5, river network fractal dimension D6) Geophysical / geochemical anomalies (such as magnetic and geochemical data). Calculate the fractal dimension for each selected data category; if no data is available, it can be excluded from the calculation. The fractal parameters of the aforementioned auxiliary data are also calculated to form a multidimensional fractal eigenvector: D = (D0, D1, α1, D2, α2, D3, α3, D4, α4, D5, α5,…) (2.3) Similarity measurement and matching Using the fractal feature vector D of the known mine as a template, the corresponding feature vector of the unknown area is extracted by sliding window (or by geological unit).

[0063] Euclidean distance, cosine similarity, and Mahalanobis distance are calculated as similarity criteria, and thresholds are set to filter out a list of "candidate blocks" with highly similar fractal features.

[0064] If a candidate block is located in the same metallogenic belt or has the same stratigraphic background as a known mine, it is determined to be a target mine of the same metallogenic environment. This rule is applied to filter the list of "candidate blocks", and the "candidate blocks" that meet the requirements and have the highest similarity (e.g., the top 100) are used as the data pool for subsequent reserve regression analysis.

[0065] Step S3 specifically includes: (3.1) Sample set construction From the known mines of the same type selected in step S2, a number of "candidate blocks" (e.g., 100) are chosen as training samples. These samples all come from mining areas with reliable reserve data.

[0066] Extract a feature vector X for each sample mine, including: The statistics (mean, variance, kurtosis, etc.) of the resource distribution map obtained in step S1. The fractal features (fractal dimension, multifractal spectrum parameters) obtained in step S2. Fractal characteristics of geological, topographic and other auxiliary data The actual explored reserves of the corresponding mining area are used as label Y.

[0067] (3.2) Regression Modeling Construct a regression model to learn the mapping relationship between feature X and reserves Y.

[0068] Preferred options: Use multiple linear regression (high interpretability) or support vector regression (SVR) (strong non-linear fitting ability).

[0069] Alternative: When the feature dimensions are high, use LASSO (Least Absolute Shrinkage and Selection Operator) to prevent overfitting.

[0070] Perform cross-validation on the model and select the optimal hyperparameters.

[0071] (3.3) Estimation of reserves in the target mining area Extract the corresponding feature vector X' of the target candidate mining area, input it into the trained regression model, and obtain the predicted reserve value.

[0072] Output the confidence interval of the prediction results (e.g., based on the Bootstrap method or the model's built-in variance estimation).

[0073] The geological rationality of the estimation results was verified by combining the regional metallogenic regularity.

[0074] The overall process and data link of this specific implementation method are as follows: Input: Satellite remote sensing hyperspectral imagery, known mine locations and reserves, regional geological (rock strata map, fault distribution map, fold map, etc.) / geographical (elevation map, river and lake distribution map, etc.) data. Output: Mineral resource distribution map of the target area and predicted reserves of candidate mines. Specifically, it includes the following: Step S101: Data Input and Initialization After the program starts, it first reads the user-specified workspace directory, which should contain: Satellite remote sensing hyperspectral image file (.tif format); The known mine vector boundary file (.shp format) and the corresponding actual reserve data (.csv or .xlsx) are available. Regional geological maps, digital elevation models (DEMs), and other auxiliary data.

[0075] The program automatically checks the integrity of files; if any necessary files are missing, it will display an error message and terminate the program.

[0076] Step S102: Remote sensing image preprocessing The program calls the GDAL or ENVI algorithm library and executes them sequentially: Radiometric calibration: converting DN values ​​to reflectivity or radiance; Atmospheric correction: Use FLAASH or 6S models to eliminate the effects of water vapor and aerosols; Geometric fine correction: Using known control points as a reference, the image is registered to a unified projection coordinate system; Image cropping: The target area is cropped based on the vector boundary of the study area.

[0077] Step S103: Machine learning model training and resource distribution map generation The internal sub-processes of this step are as follows: S1031: Extract the spectral vectors of all pixels within the known mine boundary as positive samples; randomly select an equal number of pixels from the background non-mineralized area as negative samples.

[0078] S1032: Standardize / normalize the samples and reduce the dimensionality to 32~64 dimensions using PCA or an autoencoder.

[0079] S1033: Construct a 1D-CNN model (convolutional layer + pooling layer + fully connected layer) that takes a spectral curve as input and outputs the probability of mineral deposits. The program uses the Adam optimizer with cross-entropy as the loss function and iteratively trains until the accuracy on the validation set no longer improves.

[0080] S1034: Save the trained model as a .h5 file and predict the mineral probability distribution grid (0~1 continuous values) for the entire study area pixel by pixel.

[0081] S1035: Perform a 3×3 median filter on the raster to eliminate isolated noise and output the final resource distribution map (the map is automatically saved in GeoTIFF format).

[0082] Step S104: Fractal Feature Extraction The program reads the resource distribution map generated in step S103 and calculates its fractal dimension and multifractal spectrum: The profile line method is employed: by extracting profile lines in different directions, the one-dimensional fractal dimension is calculated for rapid diagnosis and local feature analysis. Simultaneously, by comparing the power spectral density analysis results with those from two-dimensional Fourier transform, global statistical properties and anisotropy analysis are provided.

[0083] Output the directional fractal characteristics of different geological units.

[0084] Step S105: Fractal processing of geological auxiliary data If the user provides geological / topographical data, the program will automatically perform the following operations: For raster data such as fracture density maps and elevation variation maps, the profile line method is also used to calculate their fractal dimension and the angle between them and the fractal parameters of the resource distribution map obtained in step S104. The calculation results are combined with the fractal parameters of the resource distribution map obtained in step S104 to form a multidimensional fractal feature vector of the region (in JSON or CSV format).

[0085] Step S106: Matching similar mining farms The program uses known mines as templates to calculate the fractal feature vectors of each known mine. The study area is divided into several candidate windows, and the fractal feature vectors of each window are extracted sequentially. Calculate the cosine similarity between each candidate window and each known mining template. If the maximum value is greater than a preset threshold (e.g., 0.85), then the window is determined to be a "candidate mining area", and the corresponding known mining template ID and similarity score are recorded.

[0086] Step S107: Regression Model Training The program collects all known mines that have been identified as "candidate mines" and have real reserve data, and constructs a training set.

[0087] For each training sample, the feature vector X is composed of three parts: Statistical characteristics of resource distribution maps (mean, variance, skewness); Fractal characteristics (fractal dimension); Geological auxiliary fractal characteristics (if any).

[0088] Label Y represents the actual explored reserves (tons / metal content) of the mine.

[0089] The program uses support vector regression (SVR) or random forest regression, determines the optimal hyperparameters through grid search and 5-fold cross-validation, trains the regression model, and saves it.

[0090] Step S108: Target Mineral Area Reserve Prediction For the unknown candidate mines without real reserve data selected in step S106, extract their corresponding feature vectors X'. Load the trained regression model and call the predict() method to output the predicted reserve value. Meanwhile, the Bootstrap method (resampled 1000 times) is used to estimate the 90% confidence interval of the predicted value.

[0091] Example 1: Taking a laterite nickel mine in Indonesia as an example, a 1D-CNN model was trained using Sentinel-2 MSI data (13 bands) to generate a probability map of nickel ore distribution. The box dimension D = 1.72~1.85, and the Euclidean distance between it and a known mine (D = 1.75) is <0.1, indicating it is a matching mining area. Finally, the SVR model was used to predict that the nickel metal reserves in this area are 2.7 million tons, with an error of <10% compared to the actual drilling results.

[0092] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for estimating mineral resource reserves using fractal algorithms, characterized in that, Includes the following steps: Step (S1): Generation of resource distribution map based on machine learning A classification model is trained using satellite remote sensing hyperspectral data of known mines to perform pixel-level predictions of the target area and output a resource abundance distribution map of the target mineral. Step (S2): Feature matching mine identification based on fractal geometry Extract fractal dimension and multifractal spectrum feature parameters from the resource distribution map obtained in step (S1). Combine the geological structure and topographic data of the target area to calculate the fractal feature similarity between the target area and known mines. Select a batch of mines with similar mineralization environments from the list of mines with known reserves as the data source for subsequent reserve calculation. Step (S3): Reserve Estimation Based on Regression Analysis Using the similar mines matched in step (S2) as samples, a regression model is established using the actual reserve data of the same target mineral in known mines and remote sensing / geological features to quantitatively estimate the total resource reserves of this type of mineral in the target mine.

2. The method for estimating mineral resource reserves using fractal algorithms according to claim 1, characterized in that, The aforementioned step (S1) specifically includes: (1.1) Data Acquisition and Preprocessing Acquire satellite hyperspectral remote sensing images of the target area, including dozens of bands, covering visible light, near infrared, and shortwave infrared; Radiometric calibration, atmospheric correction, and geometric fine correction are performed on the images to eliminate environmental interference; Mineral index and vegetation cover index were extracted as auxiliary features; (1.2) Sample construction and model training Collect known mining areas of this type of mineral, mark the ore body range on remote sensing images, and extract the pixel spectrum within the range as positive samples; Background pixels were randomly selected from non-mineralized areas as negative samples. Construct a multidimensional feature vector, including original band reflectance, band ratio, principal component analysis dimensionality reduction components, and texture features; Supervised learning was conducted using convolutional neural networks, support vector machines, and random forest algorithms to train a mineral probability prediction model. (1.3) Resource abundance prediction The trained model is applied to the whole image for pixel-by-pixel prediction, and the probability or abundance value of each pixel belonging to the target mineral type is output. Spatial filtering is applied to the prediction results to remove salt-and-pepper noise, generating a continuous resource distribution probability map.

3. The method for estimating mineral resource reserves using fractal algorithms according to claim 2, characterized in that, The algorithm in step (1.2) uses a lightweight CNN to automatically learn spectral-spatial features.

4. The method for estimating mineral resource reserves using fractal algorithms according to claim 2, characterized in that, When the sample size is small, step (1.2) uses SVM or ensemble learning (XGBoost) supplemented by feature selection.

5. The method for estimating mineral resource reserves using fractal algorithms according to claim 1, characterized in that, The aforementioned step (S2) specifically includes: (2.1) Delineation of the core area of ​​the target mining area Find the location of the target mining area on the resource distribution map obtained in step (S1), delineate the area where the high probability of resource quantity is located, and finally select the area for subsequent fractal calculation by combining the geological and geomorphological characteristics of the area; The selected calculation area is divided into one or more sub-regions; for mines with large mining areas and multiple different structural ore-controlling modes within the area, sub-regions are divided according to the different "ore-controlling modes". When the selected sub-region cannot be well matched with the sample mine group with known reserves after subsequent "similarity measurement and matching", the sub-region is re-divided and its similarity with the sample mine group is measured again. After multiple trials, the division scheme with the highest similarity is selected as the final scheme to complete the matching. (2.2) Fractal Feature Extraction The fractal dimension (D0) of the resource distribution prediction map obtained in step (S1) is calculated using the profile line method and Fourier spectrum analysis to characterize the complexity and self-similarity of the resource distribution. Further calculation of the multifractal spectrum (f(α) curve) is performed to extract spectral width and maximum point parameters, which can more precisely describe the non-uniformity of the distribution. Simultaneously, geological auxiliary data of the target area and known mining areas are collected, and their fractal dimension (Di) and the angle αi between their directions and the fractal characteristics D0 of the resource distribution prediction map are calculated. Lithological structures: fracture density D1, normal distance of fold axis orientation D2; Topography and landforms: surface texture features D3, elevation variation coefficient D4, slope D5, river network fractal dimension D6; Geophysical and geochemical anomalies: Calculate D7, D8, etc., according to the numerical distribution of each monitoring item; The fractal parameters of the aforementioned auxiliary data are also calculated to form a multidimensional fractal eigenvector: D = (D0, D1, α1, D2, α2, D3, α3, D4, α4, D5, α5,…); (2.3) Similarity measurement and matching Using the fractal feature vector D of the known mine as a template, the corresponding feature vector of the unknown area is extracted by sliding window or by geological unit; Euclidean distance, cosine similarity, and Mahalanobis distance were calculated as similarity criteria, and thresholds were set to filter out a list of "candidate blocks" with highly similar fractal features. If a candidate block is located in the same metallogenic belt or has the same stratigraphic background as a known mine, it is determined to be a target mine of the same metallogenic environment. This rule is applied to screen the list of "candidate blocks", and the "candidate blocks" that meet the requirements and have the highest similarity are used as the data pool for subsequent reserve regression analysis.

6. The method for estimating mineral resource reserves using fractal algorithms according to claim 1, characterized in that, The aforementioned step (S3) specifically includes: (3.1) Sample set construction From the known mines of the same type selected in step S2, several "candidate blocks" are selected as training samples; these samples all come from mining areas with reliable reserve data. Extract a feature vector X for each sample mine, including: The statistics of the resource distribution map obtained in step (S1); The fractal features obtained in step (S2); Fractal characteristics of geological and topographic auxiliary data; The actual explored reserves of the corresponding mining area are used as label Y; (3.2) Regression Modeling Construct a regression model to learn the mapping relationship between feature X and reserves Y; Perform cross-validation on the model and select the optimal hyperparameters; (3.3) Estimation of reserves in the target mining area Extract the corresponding feature vector X' of the target candidate mining area, input it into the trained regression model, and obtain the predicted reserve value; Output the confidence interval of the prediction result; The geological rationality of the estimation results was verified by combining the regional metallogenic regularity.

7. A method for estimating mineral resource reserves using fractal algorithms according to claim 6, characterized in that, The step (3.2) described above uses multiple linear regression or support vector regression.

8. A method for estimating mineral resource reserves using fractal algorithms according to claim 6, characterized in that, In step (3.2), when the feature dimension is high, LASSO is used to prevent overfitting.

Images