An intelligent prediction method for ore-forming target area based on multi-source geological relationship data

Abstract

The present application relates to the technical field of mineral resources exploration, and discloses a method for intelligent prediction of metallogenic target area based on multi-source geological relationship data. The present application comprises obtaining the metallogenic favorability probability value of a plurality of grid cells in a target area; performing reverse verification, using the reserved known ore deposit points to construct a standard mode library, and calculating the local confidence of each grid cell through structural similarity; performing secondary filtering, replacing the metallogenic favorability probability value of the grid cell with a local confidence lower than a preset threshold with the weighted median of the fusion weight in its neighborhood, wherein the fusion weight is determined by the spatial distance and the local confidence of the neighbor itself; dividing the core favorable unit and the peripheral favorable unit according to the filtered probability value, and delineating the metallogenic target area through spatial connectivity clustering. Through the cooperation of reverse verification and secondary filtering, the present application effectively suppresses the isolated noise in the prediction result, and improves the continuity and reliability of the target area delineation.

Description

Technical Field

[0001] This invention relates to the field of mineral resource exploration technology, and in particular to an intelligent prediction method for ore-forming target areas based on multi-source geological relationship data. Background Technology

[0002] With the decreasing availability of surface outcrops and easily identifiable shallow deposits, mineral resource exploration is extending towards deeper and more covered areas. Through comprehensive analysis of multi-source geological data, including geological, geochemical, geophysical, and remote sensing data, potential mineral resource-rich areas are identified, providing a scientific basis for exploration project deployment. As easily identifiable shallow mineral resources become increasingly depleted, mineral exploration work has shifted entirely to finding breakthroughs in deep, concealed ore bodies and areas with complex geological conditions. Traditional manual analysis methods that rely on the experience of geological experts have limitations such as strong subjectivity, low efficiency, weak multi-source data fusion capabilities, and insufficient accuracy in deep prediction, and can no longer meet the actual needs of current exploration work.

[0003] With the rapid development of information technology and artificial intelligence, technologies such as Geographic Information Systems (GIS), machine learning, deep learning, and knowledge graphs have been gradually applied to the field of mineral resource prediction. Intelligent mineral exploration has become the mainstream direction of industry development, and a large amount of research has been carried out in this direction, resulting in a number of related technical solutions: Patent application CN118194162A discloses a method for mineral exploration target area location based on multivariate data. This method processes geological, geochemical, and geophysical datasets to extract various predictive variables, including lithological variables, fault variables, geochemical clustering variables, and dimensionality reduction variables. These variables are then fused and input into multiple machine learning models such as random forest, support vector machine, and multilayer perceptron, and the best-performing model is selected to output a mineral distribution probability map. This method focuses on variable extraction from multi-source data and comparison of multiple models, but does not perform any post-processing verification or filtering optimization on the probability map output by the models.

[0004] Patent application CN119850863A discloses a method and system for predicting mineral exploration target areas based on three-dimensional geological modeling. This method constructs a three-dimensional geological model, combines deep learning to predict deep lithology, and after screening ore-forming factors, uses a graph convolutional network for mineral exploration prediction. However, this method directly uses the prediction results for three-dimensional target area delineation and lacks verification of the reliability of the prediction results and noise processing.

[0005] Patent application CN121706051A discloses an intelligent delineation method for gold prospecting target areas based on multi-source geological data. This method enhances mineralization-related feature information, constructs a mineralization correlation network to obtain mineralization correlation parameters, calculates mineralization probability values ​​point-by-point to generate a mineralization probability surface, and finally filters high-probability areas and combines them with typical deposit models to delineate the target area. However, this method directly performs threshold filtering after generating the probability surface, without considering the potential for isolated noise or overfitting issues arising from model predictions.

[0006] Patent application CN120144957A discloses a method for constructing a knowledge graph and intelligent mineral exploration prediction based on multi-source heterogeneous geological data. This method constructs a geological knowledge graph using natural language processing technology and utilizes a deep learning model for mineralization prediction. However, this method does not address the verification or filtering of the prediction results.

[0007] Patent CN120197969B discloses a mineral exploration prediction system based on multi-scale big data fusion. This system integrates multiple modules, including geological big data management, mining and analysis, multi-element extraction, multi-scale grid computing, intelligent mineral exploration prediction, intelligent evaluation, decision support, and scheduling and command, forming a comprehensive platform. While the mineral exploration prediction module employs knowledge graphs and various machine learning algorithms, the specific techniques for reverse verification or adaptive filtering of the prediction results are not disclosed.

[0008] Patent application CN116611697A discloses a method for delineating mineral exploration target areas based on multi-source heterogeneous information using machine learning. This method establishes a conceptual model of a metallogenic system, extracts multi-source spatial proxy mineralization indicator information from geosciences, integrates the data using neural networks, and then applies machine learning algorithms for hyperparameter optimization and target area delineation. This method uses known mineral deposits as training samples but does not use these known deposits for post-validation of the prediction results.

[0009] In summary, although existing intelligent mineral exploration target area prediction technologies have achieved a preliminary integration of multi-source geoscientific data and artificial intelligence technologies, the following technical problems exist in practical mineral exploration engineering applications: First, the model predictions lack systematic reverse verification based on known mineral deposits. Existing technologies generally use known mineral deposits only as positive samples for model training. Even when training and test sets are divided, internal accuracy assessment is only performed during model training. After the model outputs the mineralization probability map, it fails to utilize reserved, independent known mineral deposits that were not involved in model training to conduct systematic external reverse verification of the predictions. This deficiency makes it impossible to quantitatively assess the reliability of the predictions in actual exploration scenarios, and it also fails to effectively distinguish between local spurious high-value noise caused by model overfitting and genuine mineralization anomalies. Ultimately, the delineated prospecting target areas are prone to include a large number of invalid regions, significantly increasing the risk and cost of exploration engineering verification.

[0010] Second, existing technologies have limited methods for handling noise in prediction results. For isolated high-value points and mottled scattered noise commonly found in prediction probability maps, existing technologies mainly employ two methods: one is global smoothing filtering, such as median filtering and Gaussian filtering. While filtering out discrete noise, this method indiscriminately smooths the probability gradient of real mineralization anomaly areas, resulting in blurred boundaries of mineralization anomalies and weakened probability values ​​of core high-value areas, directly reducing the spatial accuracy of target area delineation. The other method is directly screening target areas using a fixed threshold. This method completely fails to eliminate isolated noise points within the threshold range, resulting in the delineated target areas containing a large number of false anomalies, failing to provide accurate and effective guidance for exploration project deployment.

[0011] Third, existing technologies lack confidence assessment methods based on local spatial structure patterns. According to classical metallogenic theory and long-term exploration practices, genuine metallogenic favorable areas generally exhibit high-probability clusters with a certain degree of spatial continuity. Furthermore, the metallogenic probability around known deposits typically follows a spatial distribution pattern of decreasing gradients outwards from the deposit center. In contrast, false high-value points caused by model overfitting and data noise are mostly isolated and discrete, with their surrounding probability spatial distribution patterns fundamentally different from the typical patterns of known deposits. However, existing technologies have not explored or utilized this core metallogenic law, nor have they constructed anomaly confidence assessment methods based on local spatial structure patterns. This results in an inability to quantitatively characterize the local reliability of prediction results, nor can it achieve accurate filtering of false noise and effective enhancement of genuine anomalies through pattern recognition. Ultimately, this leads to an inability to meet the accuracy and precision requirements of mineral exploration target area delineation under deep and complex geological conditions. Summary of the Invention

[0012] Given the existing technical problems such as the lack of a reliability verification mechanism for prediction results, inherent defects in noise processing of prediction results, and the lack of confidence assessment based on local spatial structure patterns, an intelligent prediction method for ore-forming target areas based on multi-source geological relationship data is proposed.

[0013] The purpose is to establish a standard pattern library by reserving known mineral deposit points, use structural similarity to back-verify the local probability distribution pattern of each grid cell, calculate the local confidence level, and then apply adaptive weighted median filtering to low confidence areas, thereby effectively suppressing isolated noise in model prediction and improving the spatial continuity and reliability of target area delineation.

[0014] The technical solution of this invention is an intelligent prediction method for ore-forming target areas based on multi-source geological relationship data, comprising the following steps: Obtain the mineralization probability values ​​of multiple grid cells in the target area, and calculate the mineralization probability value of each grid cell based on multi-source geological data.

[0015] Perform reverse verification: Obtain the reserved known mineral deposit points within the target area, extract the mineralization favorable probability values ​​of all grid cells within the first neighborhood window centered on each known mineral deposit point, arrange them in spatial order to form the first pattern vector, and collect the first pattern vectors of all known mineral deposit points into a standard pattern library.

[0016] For each grid cell, extract the mineralization probability values ​​of all grid cells within a second neighborhood window of the same size, centered on the current grid cell, and arrange them into a second pattern vector in the same spatial order. Compare the second pattern vector with each first pattern vector in the standard pattern library one by one, and take the maximum structural similarity obtained as the local confidence of the current grid cell.

[0017] The side lengths of the first and second neighborhood windows are set to odd numbers between 5 and 15. This is because if the side length is less than 5 (e.g., 3×3), the window contains only 9 grid cells, which is too little information and cannot effectively capture the spatial continuity and gradient change characteristics that the mineralization favorable area should have. It is easy to misjudge isolated noise points as valid anomalies.

[0018] If the side length is greater than 15 (e.g., 17×17), the window will contain more than 289 grid cells, which will introduce too much background information far from the center point, dilute the feature weight of the central region, and increase the computational cost quadratically. For common study area grid sizes, this will result in unacceptable computation time.

[0019] The structural similarity is used instead of a simple mean comparison because structural similarity examines three dimensions: brightness (overall probability level), contrast (the degree of contrast between probability values), and spatial structure (the order of probability values). This more comprehensively depicts the probability distribution around a known ore deposit, while mean comparison only focuses on the overall level and cannot distinguish different spatial patterns such as high center and low periphery or high periphery and center.

[0020] Perform secondary filtering: Compare the local confidence of each grid cell with a preset threshold. For grid cells with local confidence lower than the preset threshold, define a filtering window centered on the current grid cell. Calculate the distance weight of each neighboring grid cell within the filtering window based on the spatial distance between each neighboring grid cell and the central grid cell. Multiply the distance weight of each neighboring grid cell by its own local confidence to obtain the fusion weight of that neighboring grid cell. Sort the mineralization probability values ​​of all neighboring grid cells within the filtering window, along with the central grid cell's own fusion weight, in ascending order. Calculate the cumulative sum of the fusion weights. Take the probability value corresponding to the first time the cumulative sum reaches half of the total weight as the weighted median. Replace the original mineralization probability value of the central grid cell with this weighted median.

[0021] The specific method for calculating the weighted median is as follows: Let there be a total of [number] medians within the filter window. Each grid cell (including the central grid cell and its neighbors) The corresponding mineralization favorability probability value is The corresponding fusion weight is .

[0022] All grid cells are sorted by probability value Sort in ascending order to obtain the sequence The corresponding fusion weight is .

[0023] Calculate the total weight .

[0024] Starting from the minimum probability value, the weights are accumulated sequentially until the minimum index is found. Make Then the weighted median .

[0025] The preset threshold for local confidence is between 0.6 and 0.8. If the threshold is below 0.6, a large number of grid cells with medium to low confidence will not be filtered, resulting in a significant amount of residual noise.

[0026] If the threshold is higher than 0.8, only a very small number of grid cells that are highly similar to known mineral deposit patterns will be retained, which may lead to over-filtering and weaken the true mineralization anomalies.

[0027] The filtering window uses a 3×3 or 5×5 grid window. The 3×3 window has the least computational cost and is suitable for areas with high data density and dispersed noise distribution.

[0028] The 5x5 window provides a smoother effect and is suitable for areas with continuous noise distribution.

[0029] Larger windows (such as 7×7) can cause the filtered results to be too smooth, potentially erasing small but real mineralization anomalies.

[0030] The distance weighting adopts an exponential decay function (with the exponent being the negative of the square of the Euclidean distance). The influence of spatial distance on mineralization generally follows the rule that the closer the distance, the stronger the correlation. The exponential decay gives higher weights to nearby neighbors, while the weights of distant neighbors approach zero, which is consistent with the basic principle of spatial proximity in geology.

[0031] The fusion weights consider both distance and the local confidence of the neighbors themselves. Neighbors with higher confidence are closer to the known mineralization patterns and should contribute more to the filtering. The dual weighting strategy makes the filtering results more reliably approximate the real mineralization distribution pattern.

[0032] Grid cells with a mineralization probability value greater than the first preset value after secondary filtering are marked as core favorable cells. Grid cells with a mineralization probability value greater than the second preset value and less than or equal to the first preset value after secondary filtering are marked as peripheral favorable cells. Spatial connectivity clustering is performed on the core favorable cells to obtain continuous core target areas. Then, the peripheral favorable cells adjacent to each core target area are included in the corresponding target area range to obtain a complete mineralization target area.

[0033] The first preset value and the second preset value range from 0.6 to 0.8 and 0.3 to 0.5, respectively, and the first preset value is greater than the second preset value.

[0034] Areas with a mineralization probability value below 0.3 generally have no practical mineral exploration significance, areas above 0.8 are high-reliability mineralization anomaly areas, and areas in between are medium-potential areas.

[0035] By dividing the core favorable units into core and peripheral favorable units, we can ensure the priority identification of high-potential areas and avoid missing edge mineralization information due to overly strict thresholds.

[0036] Furthermore, the mineralization favorability probability values ​​of multiple grid cells in the target area are obtained, including: Extracting geological entities with spatial coordinates and attribute information from multi-source geological data, including stratigraphic units, fault structures, magmatic bodies, geochemical anomalies, geophysical anomalies, remote sensing alteration zones, and other spatial entities with clear geological significance; Based on the extracted geological entities, spatial topological relationships, genetic evolution relationships, and mineralization coupling relationships between geological entities are constructed respectively. Using geological entities as nodes and the three types of geological relationships constructed as weighted edges, a heterogeneous geological relationship map is generated.

[0037] The node features and edge features are fused using a multi-relation heterogeneous graph attention network to obtain the fused feature vector of each node. The target area is divided into multiple regular grid units, and the mineralization probability value of the corresponding grid unit is calculated based on the node fused feature vector contained in each grid unit.

[0038] Furthermore, when constructing spatial topological relationships, the spatial distance between any two geological entities is calculated, and the weight of the spatial topological relationship is calculated based on the spatial distance using an exponential decay function.

[0039] The formula for the exponential decay function is: in, Geological entities and Spatial topological relation weights, The shortest distance in space between two entities. This is the distance attenuation coefficient. It is a natural constant.

[0040] The influence of spatial distance on mineralization correlation generally follows the rule that the closer the distance, the stronger the correlation, and this correlation strength decreases nonlinearly with increasing distance. The exponential decay function fits this geological law well.

[0041] In contrast, linear decay overestimates the correlation between distant entities, while step functions (with a weight of 1 for distances less than a threshold and 0 otherwise) are too coarse to distinguish the differences in correlation between different distance ranges.

[0042] Furthermore, when constructing the causal evolution relationship, the formation age of each geological entity is obtained. The weight of the causal evolution relationship is calculated based on the difference in formation ages between two geological entities. The weight is calculated by taking the complement of the normalized age difference (i.e., 1 minus the normalized age difference). The calculation formula is as follows: in, Geological entities and The weight of the causal evolution relationship, , These are the formation ages of the two entities, The maximum difference in formation age of all geological entities within the study area; when the calculation result is less than 0, the weight is forcibly set to 0.

[0043] The closer the formation ages of two geological entities, the closer they are in time, and the greater the possibility of a causal relationship between them; therefore, their weight should be higher. Conversely, the greater the age difference, the smaller the weight should be.

[0044] The linear normalization method maintains the physical comparability of age differences, and the calculation results always fall between 0 and 1, which facilitates subsequent fusion calculations with other relation weights.

[0045] Furthermore, when constructing the mineralization coupling relationship, the frequency of co-occurrence of any two geological entities within a pre-defined buffer zone around a known ore deposit is statistically analyzed, and the weight of the mineralization coupling relationship is calculated based on the frequency of co-occurrence. The ratio of the frequency of co-occurrence to the total number of ore deposits is used as the weight, and the calculation formula is as follows: in, Geological entities and The weight of the mineralization coupling relationship, The frequency of common occurrence of the two entities within a preset buffer. This represents the total number of known mineral deposits within the study area that participated in model training.

[0046] The higher the frequency with which two geological entities co-occur around a known mineral deposit, the greater their coupling contribution to mineralization. This statistical relationship directly reflects the correlation between geological assemblage and mineralization.

[0047] Normalizing the frequency to the 0-1 range makes the weights of the three types of geological relationships comparable, which facilitates the unified processing of subsequent attention mechanisms.

[0048] Furthermore, when fusing node features and edge features in the heterogeneous geological relationship map, a multi-relationship heterogeneous graph attention network is adopted, and independent attention weight calculation branches are set for spatial topological relationships, genetic evolution relationships, and mineralization coupling relationships.

[0049] The three types of geological relationships have different physical meanings and data distribution characteristics: Spatial topology reflects the spatial proximity between entities, and its weight distribution is usually relatively continuous.

[0050] The causal evolution relationship reflects the degree of temporal proximity, and its weight distribution may exhibit bimodal or multimodal characteristics.

[0051] Mineralization coupling relationships reflect statistical co-occurrence frequencies, and their weight distribution is usually sparse.

[0052] If a shared attention parameter is set for the three types of relationships, the model cannot adapt to the distribution characteristics of different types of relationships, which will reduce the effect of feature fusion.

[0053] Independent branches allow each type of relationship to learn its own attention weight matrix and parameters, better capturing the differences in the contribution of various relationships to mineralization.

[0054] Furthermore, when calculating structural similarity, the mean brightness, contrast variance, and spatial structure correlation coefficient between the second mode vector and the first mode vector are compared respectively, and the three are multiplied together to obtain the structural similarity.

[0055] Structural similarity (SSIM) is calculated using the following formula: in, The second mode vector With the first pattern vector Structural similarity, , Let be the arithmetic mean of the two vectors. , Let be the variances of the two vectors, respectively. Let covariance be the variance of the two vectors. , To avoid calculating the constant term that causes instability.

[0056] The mean brightness reflects the overall level of probability values ​​within the window, the contrast variance reflects the degree of contrast between high and low probability values, and the spatial structure correlation coefficient reflects whether the spatial arrangement order of probability values ​​is consistent.

[0057] The three dimensions each characterize the probability distribution pattern from different perspectives, and any difference in any dimension will lead to a decrease in structural similarity.

[0058] It can effectively identify situations with the same mean but different spatial structures. For example, the probability distribution around a known mining point is high in the center and low around the edges, while an isolated noise point may be high in the center but also high around the edges, or high in the center but high on one side and low on the other side. Even if the mean of the two is the same, structural similarity can identify the noise point.

[0059] Furthermore, the first and second neighborhood windows are the same size, and the number of grid cells on the side of the window is odd.

[0060] Ensure that the window has a clearly defined central grid cell, which facilitates a unified spatial arrangement of the positional relationship between the central grid and other grids within the window.

[0061] If an even-numbered side length is used, and there is no unique central grid, the correspondence between elements in the pattern vector and their spatial positions will not be unique, affecting the accuracy and consistency of pattern matching.

[0062] Furthermore, the filtering window is a grid window of a preset size. When calculating the distance weight based on the spatial distance, the Euclidean distance between the center point of the neighboring grid cell and the center point of the central grid cell is calculated. The exponential function value is calculated using the negative of the square of this distance as the exponent, and the distance attenuation factor is obtained.

[0063] The distance weight calculation formula is as follows: in, For neighboring grid cells Relative to the central grid cell Distance weights, The Euclidean distance between the center points of the two grid cells. This is the Gaussian kernel bandwidth parameter.

[0064] The formula for calculating the fusion weight is: in, For neighboring grid cells The fusion weight, For neighboring grid cells The local confidence level.

[0065] The influence of spatial distance on mineralization correlation decays rapidly with increasing distance. The quadratic exponential decay gives higher weight to nearby neighbors, while the weight of medium and distant neighbors rapidly approaches zero, which is more in line with the actual law of the nearest neighbor effect in geology.

[0066] Furthermore, the reserved known mineral deposit points are randomly selected from all known mineral deposit points and are not involved in the process of constructing heterogeneous geological relationship maps, feature fusion, and calculating mineralization favorability probability values.

[0067] The proportion of reserved known mineral deposits is set at 10% to 30%.

[0068] If the reservation ratio is less than 10%, the sample size in the standard model library is too small to cover the diverse mineralization spatial patterns in the study area, resulting in insufficient reliability of the confidence assessment of reverse verification. If the reservation ratio is greater than 30%, the number of positive samples used for model training is reduced, which may affect the prediction accuracy of the mineralization favorability probability value itself.

[0069] Furthermore, the present invention also provides an intelligent prediction system for ore-forming target areas based on multi-source geological relationship data, used to execute the above method, including: The probability acquisition module is used to obtain the mineralization favorability probability values ​​of multiple grid cells in the target area. It uses the probability data output by the front-end prediction model as the input of this system, so that the system can run independently of the specific prediction method, which improves the system's compatibility and applicability. The probability acquisition module can adopt the form of a data interface to directly read the gridded probability data from external storage or upstream prediction models without repeated calculations.

[0070] The reverse verification module, connected to the probability acquisition module, is used to acquire reserved known mineral deposit points within the target area. It extracts the mineralization probability values ​​of all grid cells within a first neighborhood window centered on each known mineral deposit point and arranges them into a first pattern vector. The first pattern vectors of all known mineral deposit points are then compiled into a standard pattern library. The reverse verification module also extracts the mineralization probability values ​​of all grid cells within a second neighborhood window centered on each grid cell and arranges them into a second pattern vector. The second pattern vector is compared one by one with each first pattern vector in the standard pattern library, and the maximum structural similarity is taken as the local confidence of the current grid cell. The probability value output by the probability acquisition module is the only data input to the reverse verification module. There are no intermediate processing steps between the two; direct connection minimizes data transmission latency and improves processing efficiency.

[0071] The module is divided into two sub-units: a standard pattern library construction unit and a local confidence calculation unit. This division is because the standard pattern library only needs to be built once and can be reused by all grid units. Separating the two can avoid redundant calculations and reduce system overhead.

[0072] The secondary filtering module, connected to the reverse verification module, compares the local confidence score of each grid cell with a preset threshold. For grid cells with local confidence scores below the preset threshold, a filtering window is defined. The distance weight is calculated based on the spatial distance of each neighboring grid cell within the filtering window. The distance weight is multiplied by the local confidence score of the neighboring grid cell to obtain the fusion weight. The mineralization probability values ​​of all neighboring grid cells within the filtering window, along with the central grid cell's own mineralization probability values, are sorted in ascending order according to the fusion weights, and the weighted median is taken. The weighted median replaces the original mineralization probability value of the central grid cell. The local confidence score output by the reverse verification module is the core input of the secondary filtering module. The data flow between the two is unidirectional, and direct connection can ensure the real-time nature of the filtering decision.

[0073] The module comprises three sub-units: a threshold comparison unit, a weight calculation unit, and a median filtering unit. These sub-units are responsible for confidence screening, fusion weight calculation, and probability value replacement, respectively. This functional decomposition makes the module structure clear and facilitates individual debugging and optimization. The preset threshold value range in the threshold comparison unit is 0.6 to 0.8.

[0074] The target area delineation module, connected to the secondary filtering module, is used to mark grid cells with a mineralization probability value greater than a first preset value after secondary filtering as core favorable cells, and grid cells with a mineralization probability value greater than a second preset value but less than or equal to the first preset value after secondary filtering as peripheral favorable cells. Spatial connectivity clustering is performed on the core favorable cells to obtain continuous core target areas. Then, the peripheral favorable cells adjacent to each core target area are included in the corresponding target area range to obtain a complete mineralization target area. The mineralization favorable area has spatial zoning characteristics of a core high-potential area and a peripheral medium-potential area. The secondary classification more finely describes the internal structure of the target area and provides a priority basis for subsequent exploration deployment.

[0075] The clustering analysis unit employs the DBSCAN spatial density clustering algorithm. DBSCAN does not require pre-specifying the number of clusters, automatically identifies connected regions of arbitrary shapes, and is robust to noise points, making it suitable for application scenarios such as mineralized target areas with irregular spatial morphology and isolated noise. The target area merging unit incorporates peripheral advantageous units into the core target area. The rationale is that the mineralization halo of a mineralized system typically weakens gradually outward from the core ore body. Including spatially adjacent peripheral advantageous units into the target area can avoid missing edge mineralization information, making the delineated target area more complete.

[0076] The four modules are connected sequentially according to the data flow direction: probability acquisition module, reverse verification module, secondary filtering module, and target area delineation module.

[0077] Probability values ​​must be available before reverse verification can be performed, local confidence levels must be available before secondary filtering can be performed, and corrected probability values ​​must be available before the target area can be delineated. Serial connections conform to data dependencies, avoiding circular calls and resource conflicts between modules. The connections between modules adopt standard data interface definitions, including input data formats (grid coordinate arrays, probability value arrays, confidence arrays, etc.) and output data formats, which facilitates module replacement and system expansion.

[0078] The reverse verification step and the secondary filtering step of this invention work together: the local confidence level output by the reverse verification is used as the fusion weight multiplier of the neighboring units in the secondary filtering, so that the low confidence region is dominated by the high confidence neighbor for correction, while the high confidence region remains unchanged, thereby realizing adaptive filtering.

[0079] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention achieves a quantitative assessment of the reliability of prediction results through reverse verification. It uses the local probability distribution pattern in the neighborhood of a reserved known mineral deposit point as a standard and employs structural similarity to perform reverse verification on each grid cell. This effectively identifies abnormal regions in the model prediction results that are inconsistent with the known mineralization pattern. The structural similarity is comprehensively evaluated from three dimensions: mean brightness, variance of contrast, and correlation coefficient of spatial structure. This distinguishes cases where the mean is the same but the spatial structure is different, thus making up for the shortcomings of existing technologies that rely solely on the magnitude of probability values ​​or comparison of mean values.

[0080] 2. This invention achieves synergistic optimization of fidelity preservation and noise reduction through adaptive weighted median filtering. It employs adaptive weighted median filtering based on local confidence, filtering only grid cells with confidence levels below a preset threshold, while retaining the original values ​​in high-confidence regions. This avoids the weakening of true mineralization anomaly information by global smoothing. During the filtering process, the fusion weight of neighboring grid cells depends simultaneously on their spatial distance from the central grid cell and their own local confidence level. This ensures that low-confidence regions are primarily corrected by high-confidence neighbors, while high-confidence regions remain unchanged, forming a feedback loop that verifies and guides the filtering. This effectively maintains the boundary clarity and probability strength of true mineralization anomalies while filtering out isolated noise.

[0081] 3. Reverse verification and secondary filtering are used as post-processing steps in mineralization prediction, forming a complete prediction, verification, and correction closed loop with the front-end prediction model. Known mineral deposits are reserved and not used in the training of the front-end model, but only for reverse verification. This ensures the independence of the verification data and avoids overfitting and evaluation bias caused by using the same batch of data for both training and verification. Attached Figure Description

[0082] Figure 1 This is an overall flowchart of Embodiment 1 of the present invention; Figure 2 This is a detailed flowchart of the reverse verification steps of the present invention; Figure 3 This is a flowchart of the weighted median calculation process of this invention; Figure 4 This is a flowchart of the target area delineation process of the present invention; Figure 5 This is a diagram showing the connection structure of the system modules of the present invention. Detailed Implementation

[0083] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0084] The specific values ​​given in the following embodiments (including but not limited to neighborhood window side length, confidence threshold, filter window size, reservation ratio, etc.) are preferred examples set to facilitate understanding of the technical solution of the present invention, and are not intended to limit the present invention. Those skilled in the art can adjust these parameters within a reasonable range according to the grid size of the target area, the type of mineral deposit, the data quality, and the exploration target. As long as they do not depart from the technical solution defined in the claims of the present invention, they should be considered to fall within the protection scope of the present invention.

[0085] Basic environment and prerequisites for implementing this invention: Hardware and software operating environment: All embodiments of this invention utilize a general, open-source, and publicly available software and hardware environment, without any dedicated security equipment or code, and can be reproduced by those skilled in the art without creative effort. Hardware environment: Use general-purpose servers or workstations with multi-core CPUs, large-capacity RAM, and GPU acceleration to meet the basic computing power requirements for large-scale spatial geological data processing and deep learning model training.

[0086] Software environment: It is compatible with mainstream operating systems such as Windows and Linux, and comes with a general geographic information processing platform, Python numerical computing framework, and open-source dependency libraries related to spatial data processing, machine learning, and deep learning. It can fully realize the entire process of geological spatial data processing, graph neural network model training, spatial analysis, and mineralization prediction. The version can be adapted to the actual environment.

[0087] Simulation verification environment: Mineralization prediction simulation uses industry-standard software such as MORPAS and Geomodeller.

[0088] All geological data used in the embodiments of this invention are publicly available data that can be obtained free of charge through official channels by those skilled in the art. There is no confidential or proprietary data. Specific acquisition channels include: regional geological maps, mineral geological data: public regional geological survey results published by provincial geological survey institutes, and the publicly published series of "China Mineral Geology".

[0089] Geochemical data: Public dataset from the National Geochemical Benchmark Network of the China Geological Survey.

[0090] Geophysical data: National aeromagnetic and aeromagnetic datasets from the China Airborne Geophysical and Remote Sensing Center, Ministry of Natural Resources.

[0091] Remote sensing geological data: USGSEarthExplorer, Landsat series and Sentinel series remote sensing image data publicly available from the Geospatial Data Cloud.

[0092] To avoid ambiguity, the core terms and parameters in all embodiments of this invention are standardized: Grid cells: regular square grids with equal latitude and longitude projection, CGCS2000 national geodetic coordinate system, and each grid cell has a unique row and column number and center point coordinates; Structural similarity (SSIM): The calculation formula disclosed in this specification is uniformly adopted, with constant terms fixed as C1=(0.01×1)2=0.0001, C2=(0.03×1)2=0.0009, and the probability value of fitting the [0,1] interval is calculated.

[0093] Preserving known mineral deposit points: Strictly adhering to the three no's principle of not participating in the construction of heterogeneous geological relationship maps, not participating in feature fusion, and not participating in the training of mineralization favorability models, these points are only used for reverse verification to ensure the complete independence of verification data and avoid overfitting bias.

[0094] Evaluation index definition: Hit rate = (Number of reserved verification deposit points falling into the delineated target area / Total number of reserved verification deposit points) × 100%.

[0095] False alarm rate = (Number of grid cells without mineralization within the delineated target area / Total number of grid cells in the delineated target area) × 100%.

[0096] AUC value: Area under the ROC curve, with a value range of [0,1]. The closer the value is to 1, the better the model's predictive performance.

[0097] Example 1: This example is a basic application scenario of the regional survey stage of the present invention. It addresses the industry pain points of increasingly depleted shallow mineral deposits and the difficulty of mineral exploration in covered areas, and achieves accurate delineation of mineralized prospective areas at the regional scale.

[0098] This embodiment selects a 1:50,000 scale integrated exploration area in the Gansu section of the West Qinling metallogenic belt, with a geographical range of 104°30′ to 105°00′ east longitude and 34°10′ to 34°30′ north latitude, covering a total area of ​​approximately 2,600 km². 2 This region is an important gold-antimony polymetallic metallogenic belt in my country, with 36 gold and antimony deposits (occurrences) discovered, including 1 large gold deposit, 2 medium-sized antimony deposits, and 33 small deposits and occurrences. The core factors controlling the ore-forming are the Indosinian-Yanshanian fault structure, the contact zone between carbonate and clastic strata, and intermediate-acidic magmatic rocks, possessing typical orogenic geological conditions for gold-antimony mineralization.

[0099] All data in this embodiment are publicly available geological data, and the specific sources and preprocessing steps are as follows: Prerequisite parameter settings: Grid unit division: A 100m×100m regular grid is used to divide the study area into 300 rows×260 columns, totaling 78,000 effective grid units. Each unit corresponds to an actual ground area of ​​100m×100m.

[0100] Known mineral deposit point division: 36 known mineral deposit points are randomly divided into two groups, with 25 training samples (69.4% of the total, used for model training) and 11 reserved validation samples (30.6% of the total, not used for model training, only for reverse validation). The reserved ratio is within the preferred range of 10% to 30% of this invention, taking into account both the sample size of the standard model library and the accuracy of model training.

[0101] The preset values ​​for the core parameters are: a 7×7 grid for the neighborhood window, a 5×5 grid for the filter window, a confidence threshold of 0.7, a core probability threshold of 0.75, and a peripheral probability threshold of 0.4. All parameters are determined based on the metallogenic geological characteristics of the study area and can be adjusted by those skilled in the art within the scope disclosed in this invention.

[0102] like Figure 1 As shown, step S1: Obtaining the mineralization favorability probability value of the target area grid cell, the operation is as follows: Geological entity extraction: Geological entities with spatial coordinates and standardized attributes are extracted from the preprocessed multi-source data, including: 14 stratigraphic unit entities, 38 fault structure entities, 12 magmatic rock body entities, 18 geochemical anomaly entities, 22 geophysical anomaly entities, and 16 remote sensing alteration zone entities, totaling 120 effective geological entities. Each entity has a unique ID, spatial geometric information, and 16-dimensional standardized attribute features, serving as nodes in the heterogeneous geological relationship map.

[0103] Construction and weight calculation of three types of geological relationships: Spatial topological relationships: Calculating the shortest spatial distance between any two geological entities. (Two-dimensional plane Euclidean distance, unit: km), the weights are calculated using an exponential decay function, the formula is: in, Geological entities and Spatial topological relation weights, The shortest distance in space between two entities. This is the distance attenuation coefficient. natural constant In this embodiment, the distance attenuation coefficient =0.9, determined based on the effective influence range of the ore-controlling fault in the study area of ​​1-2km; when the spatial distance between two entities is greater than 10km, the weight is forcibly set to 0, and long-distance associations without geological significance are eliminated.

[0104] Genetic and evolutionary relationships: The formation age of each geological entity is obtained, and the weights are calculated using the age difference normalization formula, which is as follows: in, Geological entities and The weight of the causal evolution relationship, , These are the formation ages of the two entities, The maximum difference in formation age of all geological entities within the study area; when the calculation result is less than 0, the weight is forcibly set to 0.

[0105] In this embodiment, =800Ma represents the maximum difference in formation age between the oldest strata and the newest rock mass in the study area. When the calculated result is less than 0, the weight is forcibly set to 0 to ensure that the weight value range is [0,1].

[0106] Metallogenic Coupling Relationship: For the 25 ore deposit points in the training sample set, a buffer zone with a radius of 2 km was constructed for each point. The frequency of co-occurrence of any two geological entities within the buffer zone was counted. The weights are calculated using a frequency normalization formula, which is: in, Geological entities and The weight of the mineralization coupling relationship, The frequency of common occurrence of the two entities within a preset buffer. This represents the total number of known mineral deposits within the study area that participated in model training.

[0107] In this embodiment, 25 represents the total number of mineral deposits in the training sample set, and the weight of entities with no co-occurrence records is reset to 0.

[0108] Heterogeneous geological relationship map generation: Using 120 geological entities as nodes and the above three types of weighted geological relationships as edges, a heterogeneous geological relationship map is constructed. The initial features of the nodes are 16-dimensional standardized attribute features, and the edges contain the corresponding relation types and weight values.

[0109] Feature fusion in multi-relationship heterogeneous graph attention network: A multi-relation heterogeneous graph attention network (MR-HGAT) is adopted, with independent attention weight calculation branches set for spatial topological relationships, genetic evolution relationships, and mineralization coupling relationships. The specific network structure and training parameters are as follows: Each branch has two graph attention layers, with each layer having a 64-dimensional output. The activation function is LeakyReLU (negative slope 0.2), and the dropout rate is 0.3 to avoid overfitting.

[0110] The feature vectors output from the three branches are concatenated and then mapped to a 128-dimensional node fusion feature vector through a fully connected layer.

[0111] Training settings: The geological entities corresponding to the 25 ore deposit points in the training sample set are used as positive samples, and 25 geological entities without mineralization records are randomly selected as negative samples. The loss function is binary cross-entropy, and the optimizer is Adam (learning rate 0.001, weight decay 1e-4). The training epochs are 200, with 20 early stopping epochs. Training is stopped when the validation set loss does not decrease for 20 consecutive epochs to avoid overfitting.

[0112] After training, a 128-dimensional fused feature vector of 120 geological entity nodes is output.

[0113] Calculation of mineralization favorability of grid cells: For each 100m×100m grid cell, the number of geological entity nodes contained within its spatial range is counted, and the fused feature vectors of the corresponding nodes are aggregated by mean to obtain the 128-dimensional feature vector of the grid. For grid cells without geological entities, the fused feature vectors of the three closest geological entity nodes are aggregated by weighting the fusion feature vectors according to the inverse distance to obtain the grid feature vector.

[0114] Input all grid feature vectors into the trained MR-HGAT network classification head, and output the mineralization probability value of each grid cell (range [0,1]) through the Sigmoid activation function to generate a 100m resolution mineralization probability raster map, which serves as the input for subsequent reverse verification.

[0115] like Figure 2 As shown, step S2: Reverse verification and local confidence calculation: Standard pattern library construction: The first neighborhood window uses a 7×7 odd-numbered grid, corresponding to an actual ground area of ​​700m×700m, ensuring that the window has a unique central grid cell.

[0116] For the 11 reserved verification ore deposits, the grid cell where the ore deposit is located is located. With the grid as the center, the mineralization favorability probability values ​​of 49 grids in a 7×7 window are extracted and arranged in row priority spatial order (from left to right, from top to bottom) as a first pattern vector of length 49. Min-max normalization is performed on each vector.

[0117] The 11 normalized first pattern vectors are aggregated to build a standard pattern library, which only needs to be built once and can be reused.

[0118] Extraction of the second pattern vector to be tested: For each grid cell in the study area, with the current grid as the center, the probability values ​​in the 7×7 second neighborhood window that are exactly the same size as the first neighborhood window are extracted, arranged in the same row priority order as the second mode vector of length 49, and the same min-max normalization processing is performed.

[0119] For the edge grid of the study area, the portion of the neighborhood window that exceeds the boundary is filled by mirroring, that is, the grid data in the study area is symmetrically copied along the boundary to ensure that the pattern vector length is consistent and to avoid distortion of the boundary data.

[0120] Local confidence calculation: Using the structural similarity calculation formula, the second pattern vector for each grid is... , and each first pattern vector in the standard pattern library Calculate the SSIM value one by one; Structural similarity (SSIM) is calculated using the following formula: in, The second mode vector With the first pattern vector Structural similarity, , Let be the arithmetic mean of the two vectors. , Let be the variances of the two vectors, respectively. Let covariance be the variance of the two vectors. , To avoid calculating the constant term that causes instability.

[0121] Using formula The maximum SSIM value is taken as the local confidence of the current grid cell, with a value range of [0,1].

[0122] Traverse all 78,000 grid cells to generate a local confidence raster that matches the probability raster space, which serves as the input for subsequent secondary filtering.

[0123] like Figure 3 As shown, step S3: Adaptive quadratic filtering based on local confidence: Filtering threshold and window settings: The local confidence preset threshold is set to 0.7. The judgment rule is as follows: for high confidence grids with local confidence ≥ 0.7, the original mineralization favorable probability value is retained and no filtering is performed; for low confidence grids with local confidence < 0.7, adaptive weighted median filtering is performed; the filtering window uses a 5×5 odd-numbered grid, corresponding to a 500m×500m range on the ground.

[0124] Fusion weight calculation: For each low-confidence center grid, a 5×5 filtering window is defined centered on it, and the fusion weights of the 25 grid cells within the window are calculated. Distance weight calculation: Calculate the Euclidean distance between the center point of the neighboring grid and the center point of the central grid. (Unit: grid cell), the distance weight is calculated using a squared exponential decay function, the formula is: in, For neighboring grid cells Relative to the central grid cell Distance weights, The Euclidean distance between the center points of the two grid cells. This is the Gaussian kernel bandwidth parameter.

[0125] In this embodiment, the Gaussian kernel bandwidth =2, adapted to a 5×5 filter window.

[0126] Fusion weight calculation: Multiply the distance weight of each neighboring grid by its own local confidence to obtain the fusion weight. The formula is as follows: in, For neighboring grid cells The fusion weight, For neighboring grid cells The local confidence level.

[0127] In this embodiment, the distance weight of the central grid itself is 1, and the fusion weight is... .

[0128] The fusion weights of the 25 grids within the window are normalized so that the sum of the weights is 1.

[0129] Weighted median calculation and probability value substitution: The 25 grid cells within the filtering window are sorted in ascending order of their mineralization favorability probability values ​​to obtain the sorted probability sequence. and the corresponding sorted fusion weight sequence .

[0130] Starting from the minimum probability value, the weights are accumulated sequentially until the minimum index is found. Make Weighted median In this embodiment, 0.5 (half of the total weight) That is, the weighted median.

[0131] Replace the original mineralization probability value of the central grid cell with the weighted median.

[0132] The filtering process is completed by traversing all low-confidence grids, while the original values ​​are retained for high-confidence grids. Finally, a raster map of mineralization favorability probability after secondary filtering is generated.

[0133] like Figure 4 As shown, step S4: Intelligent delineation of the ore-forming target area: Two-level division of advantageous units: Set the first preset value (core threshold) to 0.75 and the second preset value (peripheral threshold) to 0.4; mark the grid with probability value > 0.75 as core advantageous units, mark the grid with probability value ≤ 0.75 as peripheral advantageous units, and mark the grid with probability value ≤ 0.4 as non-advantageous units, which are not included in the target area.

[0134] Spatial connectivity clustering: The DBSCAN spatial density clustering algorithm is used to cluster core advantageous units, with the parameter set as: neighborhood radius. = 1 grid cell (100m), minimum neighborhood sample number MinPts = 4; each continuous cluster is an independent core target area, and isolated core advantageous cells that do not meet the density conditions are judged as noise points and removed.

[0135] The DBSCAN clustering algorithm used in this invention is a well-known density-based spatial clustering algorithm, whose core parameter is the neighborhood radius. And the minimum number of points, MinPts.

[0136] In this embodiment, The value is the diagonal length of the grid cell, and MinPts is 1.

[0137] For a detailed implementation of this algorithm, please refer to the paper "ADensity-Based Algorithm for Discovering Clusters in Large Spatial Databases with Noise" published by Ester et al. in 1996. Complete target area generation: Construct an adjacency buffer with a grid cell width for each core target area, filter all favorable peripheral cells within the buffer, and include them in the corresponding core target area; complete target area topology checks and attribute statistics, and output parameters such as spatial vector boundary, area, core area proportion, and average mineralization favorability of each target area, thus completing the intelligent delineation of mineralization target areas.

[0138] Simulation verification and publicly available data results: The simulation experiment in this embodiment is fully reproducible. All code and data use open-source tools and public datasets. The simulation design and results are as follows: To verify the technical effect of the present invention, four sets of control experiments were set up. All experiments used the exact same input data, training set, and validation set division, only changing the core technical aspects, to form a comparison with the complete solution of the present invention. The control schemes all correspond to the core solutions of existing patents in the background art: The simulation experiment uses 10-fold cross-validation, randomly dividing the 25 ore deposit points in the training sample set into 10 parts. Each time, 9 parts are used for training and 1 part for validation, and the average value is taken after 10 repetitions to ensure the stability of the results. The 11 ore deposit points in the reserved validation set do not participate in the training and cross-validation process and are only used for the final hit rate verification.

[0139] Publicly available simulation data results: All experimental data are the average values ​​from 10-fold cross-validation. The results are shown in the table below. The data are reproducible and verifiable. The complete scheme of this invention achieves an AUC value of 0.942, which is 5.5 percentage points higher than control group 2 and 12.1 percentage points higher than control group 4. This proves that the heterogeneous geological relationship map and multi-relationship attention network of this invention can more effectively integrate the mineralization information of multi-source geological data and has better prediction performance than existing technologies.

[0140] The accuracy rate of the reserved mineral deposit points in this invention reaches 91%, and the accuracy rate of the core target area reaches 82%, which are higher than those of the control group 2. This proves that the reverse verification process of this invention can effectively identify the real mineralization mode, greatly improve the ability to identify effective mineral deposits, and solve the defect of existing technologies that lack independent reverse verification.

[0141] The false alarm rate of the target area of ​​the present invention is only 18%, which is 24 percentage points lower than that of the control group 2. The total target area is only 48% of that of the control group 2. This proves that the adaptive weighted median filtering of the present invention can eliminate false high-value noise, significantly reduce the range of invalid target areas, reduce the cost and risk of exploration engineering verification, and solve the inherent defects of noise processing in the prior art.

[0142] Control group 3 used global median filtering. Although the false alarm rate was lower than that of control group 2, the hit rate was significantly reduced. This proves that while global filtering removes noise, it indiscriminately weakens the real mineralization anomalies. In contrast, the adaptive filtering of this invention only processes low confidence regions, balancing the core requirements of noise filtering and fidelity preservation.

[0143] This embodiment ultimately delineated six gold-antimony polymetallic mineralization target areas within the study region, covering a total area of ​​108.6 km². 2 The core target area is 32.4 km². 2Of the 11 reserved independent verification deposit points, 10 fell within the delineated target area, achieving an overall hit rate of 91%. Among them, 9 fell within the core target area, achieving a core area hit rate of 82%. This effectively eliminated 106 isolated false high-value noise points in the original probability grid, while fully preserving the boundary and gradient characteristics of the real mineralization anomalies, thus fully realizing the expected technical effects of this invention.

[0144] Example 2: This example is a refined application scenario of deep mineral exploration according to the present invention. In response to the industry pain points of high difficulty in predicting deep concealed ore bodies, difficulty in fusion of three-dimensional data, and many false anomalies, the present invention is extended to a three-dimensional spatial scale to verify the applicability of the present invention in the prediction of deep concealed ore bodies.

[0145] This embodiment selects a 1:10,000 scale detailed survey area of ​​the Xiangshan copper-nickel ore field in the eastern Tianshan Mountains of Xinjiang, with a geographical range of 93°45′~93°52′ east longitude and 41°50′~41°55′ north latitude, covering a total area of ​​approximately 60 km². 2 This mining field is an important concentration area of ​​magmatic copper-nickel sulfide deposits in my country. Two large copper-nickel deposits have been discovered, with exploration depths reaching -1200m. The core prospecting target is concealed copper-nickel ore bodies deeper than -1500m. The core factors controlling the ore are basic to ultrabasic rock bodies, deep and large faults, and magmatic segregation mineralization. It has typical characteristics of deep concealed ore exploration and is a key and challenging scenario for current mineral exploration.

[0146] Data Sources and Preprocessing: All data in this embodiment are publicly available public geological results and publicly measured data from the industry. The specific sources and preprocessing are as follows: Pre-position 3D parameter settings: Three-dimensional mesh unit division: A three-dimensional regular mesh of 200m (X-axis) × 200m (Y-axis) × 50m (Z-axis) is adopted, covering the entire detailed survey area in planar range and from the ground surface to -2000m in depth, generating a total of 1,225,000 effective three-dimensional mesh units of 175 rows × 175 columns × 40 layers, each unit having unique three-dimensional spatial coordinates.

[0147] Known ore deposit point division: 28 known copper-nickel ore bodies / mineralization points (including 12 deep concealed ore bodies) in the mining area were randomly divided into two groups: a training sample set of 20 (accounting for 71.4%) and a reserved validation set of 8 (accounting for 28.6%, including 4 deep concealed ore bodies, which are not used for model training and are only used for reverse validation).

[0148] The preset values ​​for the three-dimensional core parameters are: 3×3×3 grid for the three-dimensional neighborhood window, 3×3×3 grid for the three-dimensional filter window, confidence threshold of 0.75, core probability threshold of 0.8, and peripheral probability threshold of 0.5, which are adapted to the three-dimensional spatial distribution characteristics of deep concealed ore bodies.

[0149] This embodiment, while fully retaining the core technical solution of the present invention, adapts to the three-dimensional deep exploration scenario and expands the technical aspects in three dimensions. All steps conform to the core inventive concept of the present invention. The specific operation is as follows: Obtaining the mineralization favorability probability value of a 3D mesh cell: 3D geological entity extraction: Geological entities with 3D spatial coordinates and attributes are extracted from the preprocessed 3D multi-source data, including: 8 stratigraphic unit entities, 16 fault structure entities, 22 basic-ultrabasic rock mass entities, 24 geochemical anomaly entities, 18 deep geophysical anomaly entities, and 12 alteration zone entities, totaling 100 effective 3D geological entities. Each entity has 18-dimensional standardized attribute features and serves as a node in the 3D heterogeneous geological relationship map.

[0150] 3D geological relationship construction and weight calculation: 3D spatial topological relationships: Calculate the shortest distance in 3D space between any two 3D geological entities. (Euclidean distance in three-dimensional space), the weights are calculated using an exponential decay function, the formula is: in, Geological entities and Spatial topological relation weights, The shortest distance in space between two entities. This is the distance attenuation coefficient. It is a natural constant.

[0151] In this embodiment, the distance attenuation coefficient =1.1, adapted to the close-range ore-controlling characteristics of magmatic copper-nickel ore bodies, with a three-dimensional spatial distance greater than 5km and a weight reset to 0.

[0152] Genetic evolution relationship: The formation age of each geological entity was obtained, and the weights were calculated using the same formula as in Example 1. The maximum difference in formation age in the study area was determined. =300Ma, adapted to the time window of magmatic mineralization events in the Tianshan orogenic belt.

[0153] Metallogenic coupling relationship: For the 20 ore bodies in the training sample set, a 3D buffer with a radius of 1.5 km was constructed. The co-occurrence frequency of two geological entities within the buffer was counted, and the formula is as follows: Three-dimensional ore control range adapted to copper-nickel ore bodies.

[0154] Construction and feature fusion of a 3D heterogeneous geological relationship map: A 3D heterogeneous geological relationship map is constructed using 100 3D geological entities as nodes and three types of weighted geological relationships as edges. A 3D-adapted multi-relationship heterogeneous graph attention network is adopted, with independent attention branches set for the three types of relationships. The network is adjusted to a single graph attention layer with a single branch output dimension of 32 dimensions to avoid overfitting with small samples. The training rounds are 150, with 15 early stopping rounds, and a 64-dimensional 3D fusion feature vector is output for each node.

[0155] 3D mesh mineralization favorability calculation: For each 3D mesh cell, the geological entity fusion feature vector within its spatial range is aggregated to obtain a 3D mesh feature vector, which is input into the trained network and outputs the mineralization favorability probability value of each 3D mesh cell, generating 3D mineralization favorability probability volume data, which serves as the input for subsequent reverse verification.

[0156] 3D reverse verification and local confidence calculation: Construction of the 3D standard pattern library: A 3×3×3 first neighborhood window is used, corresponding to a 3D spatial range of 600m×600m×150m, ensuring that the window has a unique central grid cell.

[0157] For the eight reserved verification ore bodies, the probability values ​​of 27 three-dimensional grids within a 3×3×3 window are extracted with the central grid of the ore body as the core. These values ​​are arranged in a three-dimensional spatial order of layer priority and row priority to form a first mode vector of length 27. After normalization, the vectors are collected into a three-dimensional standard mode library, which covers the mineralization modes of both shallow and deep concealed ore bodies.

[0158] 3D pattern vector extraction: For each 3D mesh cell, extract the probability values ​​within the same 3×3×3 3D second neighborhood window, arrange them in the same 3D spatial order as the second pattern vector, and normalize them.

[0159] The boundaries of the 3D mesh are filled with mirror images to ensure consistent vector lengths.

[0160] 3D Local Confidence Calculation: Using the SSIM calculation formula and confidence formula that are completely consistent with Example 1, the 3D standard pattern library is traversed, and the maximum SSIM value is taken as the local confidence of the current 3D mesh to generate 3D local confidence volume data.

[0161] Three-dimensional adaptive quadratic filtering: Filtering threshold and 3D window settings: The confidence threshold is set to 0.75. High-confidence 3D grids with a confidence level ≥ 0.75 retain their original probability values, while low-confidence grids with a confidence level < 0.75 undergo 3D adaptive weighted median filtering.

[0162] The three-dimensional filtering window uses a 3×3×3 grid, corresponding to a three-dimensional spatial range of 600m×600m×150m, to avoid large windows from erasing small-scale deep concealed ore body anomalies.

[0163] 3D fusion weight calculation: For each low-confidence center grid, calculate the 3D Euclidean distance of neighboring grids within the 3D filtering window, and use the squared exponential decay function to calculate the distance weight, as shown in the formula: in, For neighboring grid cells Relative to the central grid cell Distance weights, The Euclidean distance between the center points of the two grid cells. In this embodiment, the Gaussian kernel bandwidth parameter is used. =1 Adapts to a 3×3×3 window.

[0164] The formula for fusion weights is: This completes the weight normalization.

[0165] 3D weighted median calculation and probability replacement: Following the same weighted median calculation rules as in Example 1, the probability values ​​of the 27 grids within the 3D filtering window are arranged in ascending order, the cumulative weights are calculated, and the probability value corresponding to the first time the total weight reaches 50% is taken as the weighted median. The original probability value of the central grid is replaced to generate the filtered 3D mineralization favorable probability volume data.

[0166] Intelligent delineation of 3D mineralization target areas: Two-level division of advantageous units: Set the first preset value (core threshold) to 0.8 and the second preset value (peripheral threshold) to 0.5; 3D meshes with a probability value > 0.8 are marked as core advantageous units, and those with a probability value < 0.5 and ≤ 0.8 are marked as peripheral advantageous units.

[0167] Three-dimensional spatial clustering: The three-dimensional DBSCAN clustering algorithm is used to cluster the core advantageous units. The parameters are set as neighborhood radius ε=1 three-dimensional grid and minimum number of samples MinPts=3 to identify continuous three-dimensional core target areas and remove isolated noise points.

[0168] 3D target area generation: The surrounding favorable units adjacent to each 3D core target area are included in the target area range, and the spatial range, occurrence depth range, volume, average grade prediction value and other parameters of each 3D target area are output to complete the delineation of the 3D mineralization target area of ​​the deep concealed ore body.

[0169] 3D simulation verification and publicly available data results: 3D simulation experiment design: Three sets of control experiments were set up. All experiments used the same 3D input data, training set and validation set division to verify the technical effect of the present invention in 3D deep prediction. Simulation results and implementation effects: This embodiment ultimately delineates four deep copper-nickel mineralization target areas, with a depth range of -800m to -1800m and a total volume of 2.86km². 3 The core target area has a volume of 0.72 km². 3 All eight reserved verification ore bodies (including four deep concealed ore bodies) fell within the delineated target area, achieving an overall hit rate of 100%. All four deep concealed ore bodies fell within the core target area. The invention effectively eliminated 128 isolated false high-value noise points in the three-dimensional probabilistic volume, while completely preserving the abnormal boundaries of small-scale deep concealed ore bodies. This verifies the superior performance of the invention in three-dimensional deep concealed ore body prediction scenarios and solves the industry pain point of numerous false anomalies and low hit rate in deep mineral exploration prediction.

[0170] Example 3: This example is a system implementation scenario of the present invention. Based on Examples 1 and 2, this example provides an intelligent prediction system for ore-forming target areas based on multi-source geological relationship data, such as... Figure 5 As shown, it includes: The probability acquisition module corresponds to the method steps of obtaining the mineralization favorable probability values ​​of multiple grid cells in the target area.

[0171] This module uses a data interface to directly read gridded probability data from external storage or upstream prediction models (such as the MR-HGAT network in Example 1).

[0172] The module contains a data parsing unit and a format verification unit to ensure that the grid coordinates and probability value range of the input data meet the system requirements.

[0173] In this embodiment, the probability acquisition module reads the 100m resolution mineralization favorability probability raster generated in Example 1.

[0174] The reverse verification module is directly connected to the probability acquisition module, corresponding to the reverse verification step in the method steps. Internally, the module is divided into two sub-units: a standard pattern library construction unit and a local confidence calculation unit. The standard pattern library construction unit extracts probability values ​​within a first neighborhood window centered on each reserved known mineral deposit point, arranges them spatially into a first pattern vector, and normalizes them to construct a standard pattern library. This library only needs to be constructed once and is reused by all grid cells thereafter. The local confidence calculation unit extracts probability values ​​within a second neighborhood window of the same size for each grid cell, arranges them into a second pattern vector, and calculates the structural similarity (SSIM) between each second pattern vector and each first pattern vector in the standard pattern library. The maximum value is taken as the local confidence of that grid cell.

[0175] In this embodiment, a 7×7 neighborhood window is used, 11 verification deposit points are reserved, and the local confidence of each grid cell is calculated with a value range of [0,1]. The local confidence raster output by the reverse verification module is directly sent to the secondary filtering module.

[0176] The secondary filtering module is directly connected to the reverse verification module, corresponding to the secondary filtering step in the method. The module contains three sub-units: a threshold comparison unit, a weight calculation unit, and a median filtering unit. The threshold comparison unit compares the local confidence of each grid cell with a preset threshold (0.7 in this embodiment), marking grid cells with confidence below the threshold as to be filtered, while high-confidence grid cells skip the filtering process. The weight calculation unit, for each to-be-filtered grid cell, defines a filtering window centered on it (5×5 in this embodiment), calculates the distance weight of each neighboring grid cell within the window (using a squared exponential decay function, Gaussian kernel bandwidth β=2), multiplies the distance weight by the local confidence of the neighboring cells to obtain the fusion weight, and then performs normalization.

[0177] The median filtering unit sorts the probability values ​​of all grids within the window in ascending order of their fusion weights, calculates the cumulative weighted sum, and takes the probability value at which the cumulative sum first reaches half of the total weights as the weighted median, replacing the original probability value of the center grid. After filtering, the corrected probability raster is output.

[0178] The target delineation module is directly connected to the secondary filtering module, corresponding to the target delineation step in the method. The module internally includes a threshold division unit, a cluster analysis unit, and a target merging unit. The threshold division unit compares the probability value after secondary filtering with a first preset value (0.75) and a second preset value (0.4), marking them as core advantageous units and peripheral advantageous units, respectively.

[0179] The clustering analysis unit uses the DBSCAN algorithm to perform spatial connectivity clustering on the core advantageous units. The neighborhood radius is taken as the diagonal length of the grid (approximately 70.71m), and the minimum number of points (MinPts) is set to 4 to identify continuous core target areas. The target area merging unit constructs an adjacency buffer with a grid cell width for each core target area, incorporating all peripheral advantageous units within the buffer into the target area range, and outputting the complete ore-forming target area vector boundary.

[0180] This embodiment of the system automates the entire process of reverse verification, secondary filtering, and target delineation. Compared to executing each module individually, the system offers the following synergistic advantages: the local confidence score output by the reverse verification module is directly used as the multiplier of the fusion weights in the secondary filtering module, adaptively linking the filtering strength with local reliability; the corrected probability map output by the secondary filtering module is directly used as the input to the target delineation module, avoiding manual intervention in intermediate data and error accumulation. Standard data interfaces (grid coordinate arrays, probability value arrays, confidence score arrays, etc.) are used between modules, facilitating module replacement and system expansion.

[0181] Structural similarity (SSIM) is commonly used in image quality assessment to compare the similarity of two image windows in terms of brightness, contrast, and spatial structure. This invention applies it to the field of mineralization prediction, comparing the similarity between the probability distribution patterns of unknown grid cells and the standard patterns of known mineral deposit points. Genuine mineralization anomalies typically exhibit a spatial structure with a gradient decaying outwards from the mineralization point, while spurious noise lacks this structure. SSIM can quantify this structural difference, making it suitable as a measure of local confidence.

[0182] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for intelligent prediction of ore-forming target areas based on multi-source geological relationship data, characterized in that, Includes the following steps: Acquire multi-source geological data of the target area, and calculate the mineralization probability value of each grid cell based on the multi-source geological data; Perform reverse verification: Obtain the reserved known mineral deposit points within the target area, extract the mineralization favorability probability values ​​of all grid cells within the first neighborhood window centered on each known mineral deposit point, arrange them into a first pattern vector in spatial order, and collect the first pattern vectors of all known mineral deposit points into a standard pattern library; for each grid cell, extract the mineralization favorability probability values ​​of all grid cells within the second neighborhood window of the same size centered on the current grid cell, arrange them into a second pattern vector in the same spatial order, compare the second pattern vector with each first pattern vector in the standard pattern library one by one, and take the calculated maximum structural similarity as the local confidence of the current grid cell; Perform secondary filtering: Compare the local confidence of each grid cell with a preset threshold. For grid cells with local confidence lower than the preset threshold, define a filtering window centered on the current grid cell. Calculate the distance weight of each neighboring grid cell within the filtering window based on the spatial distance between each neighboring grid cell and the central grid cell. Multiply the distance weight of each neighboring grid cell by its own local confidence to obtain the fusion weight of that neighboring grid cell. Arrange the mineralization probability values ​​of all neighboring grid cells within the filtering window, along with the central grid cell's own fusion weight, in ascending order. Calculate the cumulative sum of the fusion weights. Take the probability value corresponding to the first time the cumulative sum reaches half of the total weight as the weighted median. Replace the original mineralization probability value of the central grid cell with this weighted median. Grid cells with a mineralization probability value greater than the first preset value after secondary filtering are marked as core favorable cells. Grid cells with a mineralization probability value greater than the second preset value and less than or equal to the first preset value after secondary filtering are marked as peripheral favorable cells. Spatial connectivity clustering is performed on the core favorable cells to obtain continuous core target areas. Then, the peripheral favorable cells adjacent to each core target area are included in the corresponding target area range to obtain a complete mineralization target area.

2. The intelligent prediction method for metallogenic target areas based on multi-source geological relationship data according to claim 1, characterized in that: Obtain the mineralization favorability probability values ​​of multiple grid cells in the target area, including: Extracting geological entities with spatial coordinates and attribute information from multi-source geological data; Based on the extracted geological entities, spatial topological relationships, genetic evolution relationships, and mineralization coupling relationships between the geological entities are constructed respectively. Using the geological entities as nodes and the three types of geological relationships as weighted edges, a heterogeneous geological relationship map is generated. The node features and edge features of the heterogeneous geological relationship map are fused to obtain the fused feature vector of each node; The target area is divided into multiple grid cells, and the mineralization probability value of each grid cell is calculated based on the node fusion feature vector contained in each grid cell. Among them, the reserved known mineral deposit points are those extracted from all known mineral deposit points and do not participate in the process of obtaining the mineralization favorability probability value.

3. The intelligent prediction method for metallogenic target areas based on multi-source geological relationship data according to claim 2, characterized in that: When constructing spatial topological relationships, the spatial distance between any two geological entities is calculated, and the weight of the spatial topological relationship is calculated based on the spatial distance using an exponential decay function.

4. The intelligent prediction method for metallogenic target areas based on multi-source geological relationship data according to claim 2, characterized in that: When constructing the genetic evolution relationship, the formation age of each geological entity is obtained, and the weight of the genetic evolution relationship is calculated based on the difference in formation ages of two geological entities.

5. The intelligent prediction method for metallogenic target areas based on multi-source geological relationship data according to claim 2, characterized in that: When constructing mineralization coupling relationships, the frequency of common occurrence of any two geological entities within a preset buffer zone around a known mineral deposit is counted, and the weight of the mineralization coupling relationship is calculated based on the frequency of common occurrence.

6. The intelligent prediction method for metallogenic target areas based on multi-source geological relationship data according to claim 2, characterized in that: When fusing node features and edge features in heterogeneous geological relationship maps, a multi-relationship heterogeneous graph attention network is adopted, and independent attention weight calculation branches are set for spatial topological relationships, genetic evolution relationships, and mineralization coupling relationships.

7. The intelligent prediction method for metallogenic target areas based on multi-source geological relationship data according to claim 1, characterized in that: When calculating structural similarity, the mean brightness, contrast variance, and spatial structure correlation coefficient between the second mode vector and the first mode vector are compared respectively, and the structural similarity is obtained by multiplying the three together.

8. The intelligent prediction method for metallogenic target areas based on multi-source geological relationship data according to claim 1, characterized in that: The first and second neighboring windows are the same size, and the number of grid cells on the side of the window is odd.

9. The intelligent prediction method for metallogenic target areas based on multi-source geological relationship data according to claim 1, characterized in that: The filtering window is a grid window of a preset size. When calculating the distance weight based on the spatial distance, the Euclidean distance between the center point of the neighboring grid cell and the center point of the central grid cell is calculated. The exponential function value is calculated using the negative of the square of this distance as the exponent, and the distance attenuation factor is obtained.

10. An intelligent prediction system for ore-forming target areas based on multi-source geological relationship data, characterized in that, The system is used to execute the intelligent prediction method for metallogenic target areas based on multi-source geological relationship data as described in any one of claims 1-9, the system comprising: The probability acquisition module acquires the mineralization favorability probability values ​​of multiple grid cells in the target area; The reverse verification module, connected to the probability acquisition module, acquires the reserved known mineral deposit points within the target area. It extracts the mineralization favorability probability values ​​of all grid cells within the first neighborhood window centered on each known mineral deposit point and arranges them into a first pattern vector. It then aggregates the first pattern vectors of all known mineral deposit points into a standard pattern library. Additionally, it extracts the mineralization favorability probability values ​​of all grid cells within the second neighborhood window centered on each grid cell and arranges them into a second pattern vector. Finally, it compares the second pattern vector with each first pattern vector in the standard pattern library and takes the maximum structural similarity as the local confidence of the current grid cell. The secondary filtering module, connected to the reverse verification module, compares the local confidence of each grid cell with a preset threshold. For grid cells with local confidence below the preset threshold, a filtering window is defined. The distance weight is calculated based on the spatial distance of each neighboring grid cell within the filtering window. The distance weight is multiplied by the local confidence of the neighboring grid cell to obtain the fusion weight. The mineralization probability values ​​of all neighboring grid cells within the filtering window, together with the mineralization probability values ​​of the central grid cell itself, are sorted in ascending order according to the fusion weight and the weighted median is taken. The weighted median replaces the original mineralization probability value of the central grid cell. The target area delineation module is connected to the secondary filtering module. It marks the grid cells whose mineralization probability value after secondary filtering is greater than the first preset value as core favorable cells and the grid cells whose mineralization probability value after secondary filtering is greater than the second preset value and less than or equal to the first preset value as peripheral favorable cells. It performs spatial connectivity clustering on the core favorable cells to obtain continuous core target areas. Then, it includes the peripheral favorable cells adjacent to each core target area into the corresponding target area range to obtain a complete mineralization target area.

Images