Intelligent evaluation method for ore-forming target area

Abstract

The present application relates to the technical field of mineral resource analysis, and more particularly to a method for intelligent evaluation of ore-forming target areas, which generates an ore-forming probability map based on a deep learning model, segments the ore-forming probability prediction map, and generates a plurality of candidate target area units; calculates the geological logic score of each candidate target area unit and the matching degree with known ore deposit points in the study area; and finally calculates a comprehensive score based on the objective weighting of typical ore deposits and sorts the candidate target area units to classify them. The present application solves the problems of uneven scale and distorted feature extraction in intelligent ore-forming prediction target areas, reduces the complexity of the model, improves the accuracy of ore-forming target area sorting and geological interpretability, and is suitable for intelligent ore-forming prediction and batch target area screening and verification driven by multi-source data.

Description

Technical Field

[0001] This invention belongs to the field of mineral resource analysis technology, and in particular relates to an intelligent evaluation method for ore-forming target areas. Background Technology

[0002] With the deepening of deep mineral exploration, mineralization prediction methods based on machine learning (such as random forests and support vector machines) and deep learning (such as convolutional neural networks and residual neural networks) have become a research hotspot in the field of mineral exploration. These methods extract nonlinear mineralization patterns from multi-source geoscientific data, including geological, geophysical, geochemical, and remote sensing data, to generate mineralization probability prediction maps covering the study area. This provides crucial data support for mineral exploration and effectively promotes the transformation of mineralization prediction technology from traditional qualitative evaluation to quantitative intelligent evaluation.

[0003] However, current research on mineralization prediction largely focuses on the construction of front-end mineralization prediction models and the improvement of algorithm accuracy. The crucial link connecting theoretical achievements in mineralization prediction with practical field exploration—post-processing of mineralization probability prediction maps, automated target area delineation, and quantitative evaluation and optimization—lacks systematic research and reliable technical solutions. This has become a core bottleneck restricting the engineering application of intelligent mineralization prediction technology. In actual exploration work, the mineralization probability prediction maps output by deep learning models often contain large, connected high-probability regions and discrete high-probability noise patches. How to scientifically delineate mineralization target areas with reasonable spatial scales and clear engineering verification value from a continuous and complex probability field, and how to objectively and quantitatively rank and optimize the mineralization potential of these target areas, are key technical challenges that urgently need to be overcome in the current mineral exploration field. Currently, target area screening and evaluation still heavily rely on the manual experience and judgment of exploration personnel. This not only suffers from strong subjectivity, high degree of blindness, and low efficiency, but also makes it difficult to balance the conformity with the geological laws of mineralization and the prediction accuracy of data-driven models. Problems such as uneven target area scale, distorted feature extraction, and poor geological interpretability of evaluation results are prone to occur, which seriously restricts the efficiency and accuracy of mineral exploration and discovery.

[0004] In summary, the key technical problem that urgently needs to be solved in the process of transforming current intelligent mineralization prediction technology into engineering applications is: how to scientifically and rationally scale and automatically segment the continuous mineralization probability field, accurately extract the deep mineralization characteristics of irregular target areas under the premise of effectively eliminating background noise interference, and then construct a comprehensive evaluation system that combines geological logic and high data mining accuracy, so as to achieve objective and quantitative ranking and optimization of mineralization target areas and break through the bottleneck between the theoretical results of mineralization prediction and engineering exploration practice. Summary of the Invention

[0005] In view of this, the present invention aims to provide an intelligent evaluation method for ore-forming target areas. It achieves adaptive scale regularization of ore-forming target areas through a probabilistic topological watershed algorithm, and combines an object-based weighted evidence model with deep feature cross-similarity matching technology to accurately extract core ore-forming features of the target area. This solves the problems of uneven target area scale, distorted feature extraction, and insufficient geological interpretability of evaluation results in existing technologies. The present invention achieves a technological breakthrough in post-processing of ore-forming prediction by constructing a two-dimensional mutual trust evaluation system that integrates explicit geological logic and implicit data mining, moving from coarse data sorting to refined intelligent selection. This provides a high-precision, highly interpretable, and engineering-applicable intelligent decision-making tool for mineral resource exploration.

[0006] To achieve the above objectives, the technical solution created by this invention is implemented as follows: A method for intelligent evaluation of ore-forming target areas includes: S1: Obtain multi-source geoscientific data raster layers of the study area and input the multi-source geoscientific data raster layers into a deep learning model to obtain a mineralization probability prediction map; S2: Perform probabilistic topological watershed segmentation on the mineralization probability prediction map obtained in step S1, dividing the high-probability connected domain into several candidate target area units. S3: Perform a geological logic score on each candidate target area unit obtained in step S2 to reflect the matching degree of geological metallogenic regularity; based on the deep learning model in step S1, determine the similarity between each candidate target area unit and the known mineral deposit points in the study area in step S1; S4: The geological logic score and similarity weighted fusion of each candidate target area unit obtained in step S3 are used to obtain the comprehensive score of each candidate target area unit; based on the known mineral deposit distribution in the study area in step S1, the weights of the geological logic score and similarity are optimized according to the comprehensive score to obtain the optimal weights and the corresponding optimal comprehensive score of each candidate target area unit. S5: Sort the corresponding optimal comprehensive scores of all candidate target units obtained in step S4, and classify the candidate target units according to the set classification threshold.

[0007] Furthermore, in step S1, the deep learning model adopts the ResNet-50 model, and the input multi-source geoscientific data raster layers include geological layers, geochemical layers, geophysical layers, and remote sensing layers of the study area.

[0008] Furthermore, step S2 includes: S21: Using known mineral deposit points within the study area as positive samples, randomly select multiple spatially independent background points within the study area as negative samples, generate multiple candidate thresholds to distinguish between positive and negative samples, and calculate the receiver operation feature curve corresponding to each candidate threshold. S22: Based on the receiver operation characteristic curve obtained in step S21, calculate the Youden index corresponding to each candidate threshold, and determine the candidate threshold with the largest Youden index as the optimal binarization threshold; use the optimal binarization threshold to binarize the mineralization probability prediction map to obtain the mineralization prospect mask. S23: Using the mineralization probability map as the three-dimensional terrain surface and the probability value of the mineralization probability prediction map as the elevation value, a three-dimensional mineralization probability field is formed; the gradient amplitude of the three-dimensional mineralization probability field is calculated, and the local maximum points in the three-dimensional mineralization probability field are determined. S24: Based on the local maxima points obtained in step S23, the mineralization foreground mask obtained in step S22 is used to perform image segmentation on the mineralization probability prediction map to obtain several independent candidate target units.

[0009] Furthermore, step S3, which involves performing a geological logic score on the candidate target area units, includes: S311: Using the spatial vector boundary of each candidate target area unit as a mask, obtain the distribution of continuous attribute values ​​covered by the multi-source geoscience data raster layer of each candidate target area unit in step S1; S312: Extract statistics from the continuous attribute value distribution obtained in step S311 that can represent the overall high-value anomaly characteristics of the candidate target area unit; S313: Based on the statistics obtained in step S312, the continuous attribute values ​​are discretized into several segments using the natural breakpoint method, and the studentization contrast of each segment is calculated. S313: For each multi-source geoscience data raster layer, based on the distribution of effective pixel values ​​of the layer within the study area, the continuous attribute values ​​are discretized into several segments, and the studentized contrast of each segment is calculated; then, based on the segment to which the object-level statistics of each candidate target unit on the corresponding layer belong, the studentized contrast value of the candidate target unit on the corresponding layer is determined. S314: Accumulate the studentized contrast values ​​of each candidate target area unit obtained in step S313 on each multi-source geoscience data raster layer to obtain the geological logic score of each candidate target area unit.

[0010] Furthermore, the process of extracting statistics from the attribute value distribution in step S312 includes: for each candidate target area unit and each multi-source geoscientific data raster layer, using the spatial vector boundary of the candidate target area unit as a mask, extracting all effective pixel attribute values ​​of the candidate target area unit within the coverage area of ​​the corresponding multi-source geoscientific data raster layer; sorting and statistically analyzing the effective pixel attribute values ​​or constructing an empirical cumulative distribution, and using the 90th quantile corresponding to the cumulative frequency reaching 90% as the object-level statistics of the candidate target area unit on the corresponding layer.

[0011] Furthermore, the process of calculating the studentized contrast of each segment in step S313 includes: using known mineral deposit points within the study area as positive samples, randomly selecting multiple spatially independent background points within the study area as negative samples, and calculating the positive weights of evidential features appearing in the segment using the following formula: ; Among them, W + B represents positive weights, B represents evidence features appearing in the current segment, and D represents positive samples. Let P(B|D) represent a negative sample, and let P(B|D) represent the probability that the evidence feature appears in a positive sample. This indicates the probability of an evidential feature appearing in a negative sample. An evidential feature is the event that the attribute value in a multi-source geoscientific data raster layer falls into a certain segment. This can be further understood as the event that the attribute value in a geological layer, geochemical layer, geophysical layer, or remote sensing layer falls into a certain segment. The negative weight of a segment where no evidentiary features appear is calculated using the following formula: ; Among them, W - Indicates negative weight. This indicates that no evidentiary features were found in the current section. This indicates the probability that the evidence feature does not appear in the positive sample. This indicates the probability that the evidence feature does not appear in the negative sample. The student-specific contrast is calculated using the following formula: ; Among them, C s σ represents the contrast in studentization. 2 Indicates variance.

[0012] Furthermore, the process of calculating similarity in step S3 includes: S321: According to the set rule grid window, the multi-source geoscience data raster layer is windowed so that the known mineral deposit points are within the range of the grid window; S322: After the result of the division in step S321 is processed by the deep learning model in step S1, the first semantic feature vector containing the known mineral deposit points is obtained. S323: Keeping the regular grid window size, step size, and channel order set in step S321 unchanged, determine the regular grid points that fall inside the current candidate target area unit according to the geometric boundary of the current candidate target area unit; then, using the regular grid points as the window center, crop the corresponding multi-source geoscience data windows from the multi-source geoscience data raster layer; input the cropped multi-source geoscience data windows into the deep learning model trained in step S1 to obtain the second semantic feature vector; S324: Stack the first semantic feature vectors obtained in step S322 by row to form a deposit feature matrix; stack the second semantic feature vectors obtained in step S323 by row to form a target feature matrix for the current candidate target area unit; S325: Calculate the cosine similarity between the ore deposit feature matrix and the target area feature matrix obtained in step S324 to obtain the cross-similarity matrix. Then, aggregate the cross-similarity matrix to obtain the similarity.

[0013] Furthermore, in step S325, the local maximum matching mean rule is used to aggregate the cross-similarity matrix to obtain the similarity.

[0014] Furthermore, the process of determining the weights based on the distribution of known mineral deposits in the study area in step S4 includes: using known mineral deposits in the study area as positive samples, randomly selecting multiple spatially independent background points in the background of the study area as negative samples, and calculating the mean and variance of the comprehensive scores of the two types of samples respectively; constructing an objective function based on the mean and variance of the two types of samples according to the Fisher discrimination criterion; selecting the weight that maximizes the objective function as the optimal weight, and using the optimal weight to weight and fuse the geological logic score and similarity to obtain the optimal comprehensive score.

[0015] Furthermore, the grading process in step S5 includes: using the optimal weights obtained in step S4, calculating the optimal comprehensive score of the known ore deposits in the study area to obtain the optimal comprehensive score distribution; calculating the receiver operating characteristic curve of the optimal comprehensive score distribution and using the Youden index to determine the grading threshold; and combining the statistical distribution characteristics of the natural discontinuities of the optimal comprehensive score distribution with the grading threshold to complete the target grading.

[0016] Compared with the prior art, the present invention can achieve the following beneficial effects: This invention presents an intelligent evaluation method for mineralized target areas. It achieves adaptive scale regularization of candidate target areas through a probabilistic topological watershed algorithm, solving the problems of uneven target area scale and poor comparability in traditional methods. By constructing a two-dimensional evaluation system based on geological logic matching and ore deposit feature fingerprint similarity matching, it ensures the geological interpretability of the evaluation results based on a weighted evidence model of the object. Furthermore, by reusing the complete feature extraction module of a deep learning model through the cross-similarity matrix method, it achieves accurate quantitative matching of deep mineralization patterns, solving the problem of feature extraction distortion. Finally, it uses the Fisher discriminant criterion to achieve objective weight assignment, completing the quantitative ranking and optimization of target areas. This invention represents a technological breakthrough in post-processing of mineralization prediction, moving from "extensive data ranking" to "refined intelligent optimization." It effectively bridges the bottleneck between theoretical achievements in intelligent mineralization prediction and practical field engineering exploration, providing high-precision, geologically interpretable, and engineering-practical technical support for deep mineral exploration. Attached Figure Description

[0017] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments and descriptions of the invention are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 A schematic flowchart of the intelligent evaluation method for ore-forming target areas described in the embodiments of the present invention; Figure 2 This is a schematic diagram illustrating the effect of the segmentation process described in the embodiment of the present invention. Figure 3 A schematic diagram of the overall process of steps S3 and S4 as described in the embodiment of the present invention; Figure 4 A schematic diagram of the generated candidate target classification and optimization results described in the embodiments of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not constitute a limitation thereof. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0019] In the description of this invention, it should be understood that the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0020] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0021] like Figure 1 As shown in the embodiments of the present invention, the intelligent evaluation method for ore-forming target areas includes: S1: Obtain multi-source geoscientific data raster layers of the study area and input the multi-source geoscientific data raster layers into the deep learning model to obtain a mineralization probability prediction map.

[0022] In some embodiments, the deep learning model adopts the ResNet-50 model. In this embodiment of the invention, the Shizigou copper-polymetallic ore cluster in Qinghai Province is taken as the study area. This area has a complex geological structure and complete geological, geophysical, geochemical and remote sensing multi-source geoscientific data. Seventeen typical copper-polymetallic ore deposits have been discovered in the study area, which has a mature foundation for the study of regional metallogenic regularities and can provide reliable positive and negative sample support for the implementation of this invention.

[0023] In some embodiments, the input multi-source geological data raster layer includes a geological layer, a geochemical layer, a geophysical layer, and a remote sensing layer of the study area. The geological layer in this embodiment includes a fault buffer map of the study area (distance from the fault 0-500m, 500-1000m, 1000-2000m), a raster map of the distribution of favorable metallogenic strata and lithologies in the study area, and a stratigraphic lithology entropy map. Further, the raster map of the distribution of favorable metallogenic strata and lithologies is obtained by assigning values ​​to strata or lithological units spatially associated with mineralization based on regional metallogenic geological data and the spatial overlay relationship between known ore deposits and stratigraphic lithological units. This raster map is used to characterize the indicative role of stratigraphic lithological conditions on mineralization. The stratigraphic lithology entropy map characterizes the complexity of lithological type combinations within a preset neighborhood window. It is obtained by statistically analyzing the area proportion of each lithological category within the neighborhood window and calculating the information entropy, and is used to reflect the indicative role of lithological contact zones or multi-lithological combination areas on mineralization. Specifically, the geochemical layer in this embodiment of the invention employs the inverse distance weighting method to interpolate 1:50,000 stream sediment measurement data of the study area, generating a geochemical anomaly map including major ore-forming elements (Cu, Au, Ag, Pb, Zn) and indicator elements (As, Sb, Bi). The geophysical layer in this embodiment includes a 1:50,000 Bouguer gravity anomaly map and an aeromagnetic total field anomaly map of the study area, as well as the vertical first derivative and horizontal total gradient anomaly maps of the aeromagnetic total field anomaly map. Further, the aeromagnetic total field anomaly map refers to a magnetic anomaly raster map obtained by subtracting the geomagnetic normal field after diurnal variation correction, normal field correction, and leveling of the observed total magnetic field strength obtained through airborne magnetic surveys. The vertical first derivative anomaly map of the aeromagnetic total field anomaly map is used to enhance the anomalous response of shallow or local magnetic bodies, and the horizontal total gradient anomaly map of the aeromagnetic total field anomaly map is used to highlight the lateral variation boundaries of magnetic anomalies. These vertical and horizontal anomaly maps provide geophysical evidence for identifying fault structures, rock mass boundaries, and mineralization-related magnetic anomalies. The remote sensing layers in this embodiment of the invention include iron staining alteration anomaly maps and hydroxyl alteration anomaly maps extracted from Landsat-8 satellite remote sensing data of the study area using principal component analysis (PCA) and band ratio method.

[0024] In this embodiment of the invention, the acquired geological, geochemical, geophysical, and remote sensing layers are resampled to a unified spatial resolution of 30m×30m, and the unified spatial coordinate system is WGS84 UTM projection. The values ​​of each layer are normalized in the 0-1 range to ensure that the spatial range, resolution, and coordinate system of all input data are completely consistent, providing a standardized data foundation for subsequent mineralization prediction, target area segmentation, and quantitative evaluation.

[0025] S2: Perform probabilistic topological watershed segmentation on the mineralization probability prediction map obtained in step S1, dividing the high-probability connected region into several candidate target area units.

[0026] In some embodiments, to address the problem of high-probability region connectivity forming large patches ("pancake-like") or fragmented noise ("pepper-like") caused by the traditional single threshold method, step S2 employs a probabilistic topological watershed algorithm to segment the mineralization probability map, generating candidate target area units of comparable scale; step S2 includes: S21: Using known mineral deposits within the study area as positive samples, randomly select multiple spatially independent background points within the study area as negative samples, generate multiple candidate thresholds to distinguish between positive and negative samples, and calculate the receiver operation feature curve corresponding to each candidate threshold.

[0027] The Receiver Operating Characteristic (ROC) curve is a common term in the field of computer-based mineralization prediction, used to evaluate the ability of different mineralization probability thresholds to distinguish between known mineral deposits and background points. Specifically, several candidate mineralization probability thresholds are traversed, and areas in the mineralization probability prediction map with probability values ​​greater than or equal to the current candidate threshold are identified as mineralization prospective areas. The proportion of positive samples falling into mineralization prospective areas is calculated as the true positive rate, and the proportion of negative samples identified as mineralization prospective areas is calculated as the false positive rate. The ROC curve is constructed with the false positive rate on the x-axis and the true positive rate on the y-axis. This curve is used in the subsequent step S22 to determine a binarization threshold that can balance the ability to identify known mineral deposits and the ability to exclude background areas. In this embodiment of the invention, 17 known typical mineral deposits within the study area are used as positive samples, and 200 spatially independent background points far from the mineralized area within the study area are randomly selected as negative samples. S22: Based on the receiver operation feature curve obtained in step S21, calculate the Youden index corresponding to each candidate threshold, and determine the candidate threshold with the largest Youden index as the optimal binarization threshold; use the optimal binarization threshold to binarize the mineralization probability prediction map to obtain a mineralization foreground mask. The mineralization foreground mask is used to limit the effective range of subsequent segmentation and eliminate interference from low-probability background areas. In this embodiment of the invention, the optimal binarization threshold obtained is 0.68.

[0028] S23: Using the mineralization probability map as the three-dimensional topographic surface and the probability values ​​of the mineralization probability prediction map as the elevation values, a three-dimensional mineralization probability field is formed; the gradient amplitude of the three-dimensional mineralization probability field is calculated to determine the local maxima points in the three-dimensional mineralization probability field. Regions with larger gradients in the three-dimensional mineralization probability field correspond to "steep slopes" of probability changes, which are often the boundaries of different mineralization centers.

[0029] S24: Based on the local maxima points obtained in step S23, the mineralization foreground mask obtained in step S22 is used to segment the mineralization probability prediction map, resulting in several independent candidate target area units. In this embodiment of the invention, a watershed segmentation algorithm is used to segment the mineralization probability prediction map. Specifically, local maxima points are marked as seed points for catchment basins. Under the constraint of the mineralization foreground mask, a marker-controlled watershed segmentation algorithm is run on the mineralization probability prediction map. This algorithm simulates water flow spreading from the local maxima seed points outwards. When water flows from different catchment basins meet, a watershed line is constructed, thereby naturally segmenting the connected high-probability areas in the mineralization probability prediction map into several independent candidate target area units with comparable scale spatial distribution characteristics. The result obtained after image segmentation in step S24 is as follows: Figure 2 As shown, Figure 2 (a) is a schematic diagram of the effect before segmentation. Figure 2 (b) is a schematic diagram showing the effect after the segmentation process in step S24. From Figure 2 As can be seen, before the segmentation process, high-probability regions are interconnected, and the boundaries between different potential mineralization centers are unclear. After step S24, the continuous high-probability regions are divided into several spatially independent candidate target area units by local maxima seed points and probability gradient boundaries. Therefore, the present invention provides a basis for candidate target areas to have clearer spatial boundaries and better scale comparability, thus providing a basis for subsequent geological logic scoring, similarity calculation, and ranking optimization based on target area objects.

[0030] In this embodiment of the invention, morphological opening operations are also performed on several independent candidate target units to remove tiny noise patches with an area of ​​less than 0.1 km², smooth the boundaries of the candidate target units, and finally generate 26 spatially independent candidate target units with comparable scale distribution characteristics.

[0031] S3: Perform a geological logic score on each candidate target area unit obtained in step S2 to reflect the matching degree of geological mineralization regularity; based on the deep learning model in step S1, determine the similarity between each candidate target area unit and the known mineral deposit points in the study area in step S1.

[0032] This invention employs an object-based weighted evidence model to calculate geological logic scores. The core of this method is to transform the evaluation from pixel-level to target area object-level evaluation, quantifying the degree of matching between the target area and known mineralization patterns. Specifically, in some embodiments, step S3, which involves geological logic scoring of candidate target area units, includes: S311: Using the spatial vector boundary of each candidate target area unit as a mask, obtain the distribution of continuous attribute values ​​covered by the multi-source geoscience data raster layer of each candidate target area unit in step S1. 。In this embodiment of the invention, the specific method involves statistically analyzing the distribution of continuous attribute values ​​of all covered pixels within the geological layer, geochemical layer, geophysical layer, and remote sensing layer for each candidate target area unit.

[0033] S312: Extract statistical quantities that represent the overall high-value anomaly characteristics of candidate target area units from the continuous attribute value distribution obtained in step S311, thereby realizing the conversion from pixel-level data to object-level data. In this embodiment of the invention, statistical quantities that represent the overall high-value anomaly characteristics of candidate target area units are extracted from the attribute value distributions corresponding to geological layers, geochemical layers, geophysical layers, and remote sensing layers, respectively.

[0034] In some embodiments, the process of extracting statistics from the attribute value distribution in step S312 includes: for each candidate target area unit and each multi-source geoscientific data raster layer, using the spatial vector boundary of the candidate target area unit as a mask, extracting all effective pixel attribute values ​​of the candidate target area unit within the coverage area of ​​the corresponding multi-source geoscientific data raster layer; sorting and statistically analyzing the effective pixel attribute values ​​or constructing an empirical cumulative distribution, and using the 90th quantile corresponding to a cumulative frequency of 90% as the object-level statistics of the candidate target area unit on the corresponding layer. Compared with the mean and median, the 90th quantile can reduce the influence of low-value noise in the non-mineralized background within the target area on the statistical results, and retain high-value information reflecting mineralization anomalies, thereby realizing the conversion from pixel-level data to object-level data. In this embodiment of the invention, effective pixel attribute values ​​of each candidate target area unit within the coverage area of ​​the corresponding layer are extracted for geological layers, geochemical layers, geophysical layers, and remote sensing layers, and the 90th quantile is calculated as the object-level statistics of the corresponding layer; it can be understood that each candidate target area unit corresponds to a statistic on each type of layer.

[0035] S313: For each multi-source geoscientific data raster layer, based on the distribution of effective pixel values ​​of the layer within the study area, the continuous attribute values ​​are discretized into several segments, and the studentized contrast of each segment is calculated. Then, based on the segment to which the object-level statistics of each candidate target area unit on the corresponding layer belong, obtained in step S312, the studentized contrast value of that candidate target area unit on the corresponding layer is determined. The studentized contrast is used to reflect the degree of significant correlation between the multi-source geoscientific data raster layer and the spatial distribution of mineralization. In this embodiment of the invention, for geological, geochemical, geophysical, and remote sensing layers respectively, based on the distribution of effective pixel values ​​of the corresponding layers within the study area, several discrete segments are determined using the natural breakpoint method or the natural breakpoint approximation method based on quantiles; based on the distribution relationship between known ore deposit points and effective background pixels in the study area within and outside each discrete segment, the studentized contrast corresponding to each discrete segment in each layer is calculated; then, the 90th quantile statistic of each candidate target area unit obtained in step S312 on the layer is assigned to the corresponding discrete segment, and the studentized contrast corresponding to the discrete segment is used as the studentized contrast value of the candidate target area unit on the layer.

[0036] In some embodiments, the process of calculating the studentization contrast of each segment in step S313 includes: Using known mineral deposits within the study area as positive samples, and randomly selecting multiple spatially independent background points within the study area as negative samples, the positive weights of evidential features appearing in the segment are calculated using the following formula: ; Among them, W + The positive weights represent the degree of support for mineralization when the evidence features appear. B indicates that the evidence features appear in the current segment, and D represents positive samples. Let P(B|D) represent a negative sample, and let P(B|D) represent the probability that an evidentiary feature appears in a positive sample. This indicates the probability of the evidence feature appearing in the negative sample; the evidence feature is the event that the attribute value in the multi-source geoscientific data raster layer falls into a certain segment, which can be further understood as the event that the attribute value in the geological layer, geochemical layer, geophysical layer or remote sensing layer falls into a certain segment. The negative weight of a segment where no evidentiary features appear is calculated using the following formula: ; Among them, W - This indicates a negative weight, used to characterize the degree of influence on mineralization when evidential features are absent. This indicates that no evidentiary features were found in the current section. This represents the probability that no evidential features appear in the positive samples. This indicates the probability that the evidence feature does not appear in the negative sample. The student-specific contrast is calculated using the following formula: ; Among them, C s σ represents the student-specific contrast, used to evaluate the significance of the correlation between the evidence segment and the spatial distribution of mineralization. 2 The variance is represented by the difference between positive and negative weights, which reflects the overall indicative strength of the evidence segment for mineralization.

[0037] S314: Accumulate the studentized contrast values ​​of each candidate target area unit obtained in step S313 on each multi-source geoscience data raster layer to obtain the geological logic score of each candidate target area unit.

[0038] In this embodiment of the invention, the corresponding studentized contrast values ​​of each candidate target area in the geological layer, geochemical layer, geophysical layer and remote sensing layer are summed to obtain the geological logic score of each candidate target area.

[0039] In this embodiment of the invention, the obtained geological logic score is further normalized to the range of [0, 100]. The closer the score is to 100, the higher the similarity between the target area and the regional metallogenic geological patterns.

[0040] This invention reuses the deep learning model in step S1 and calculates the similarity score between the candidate target area unit and the known ore deposit points within the study area in step S1 using the cross-similarity matrix method. This method quantifies the similarity of the deep mineralization patterns of the candidate target area and known typical ore deposits by constructing an ore deposit sample fingerprint database. Specifically, in some embodiments, the process of calculating the similarity in step S3 includes: S321: According to the set rule grid window, the multi-source geoscientific data raster layer is windowed so that the known mineral deposit points are within the range of the grid window.

[0041] In this embodiment of the invention, the parameters of the regular grid window (such as size, step size, and multi-source geoscientific data channel order) remain fixed, ensuring that each window contains multi-source geoscientific data at the corresponding spatial location. In this embodiment, the pixel size of the regular grid window is 32×32, corresponding to a 900m×900m area on the ground. For any known mineral deposit point, a multi-source geoscientific data window containing the current known mineral deposit point is extracted, with the current known mineral deposit point's location or its neighboring regular grid points as the window center.

[0042] S322: After the result of the division in step S321 is processed by the deep learning model in step S1, the first semantic feature vector containing known mineral deposit points is obtained.

[0043] Further explanation is that the first semantic feature vector refers to the deep mineralization semantic representation obtained by the deep learning model trained in step S1 after extracting features from a multi-source geoscientific data window containing known mineral deposits. Specifically, the multi-source geoscientific data window is a multi-channel data window composed of multiple geoscientific raster layers. After being input into the deep learning model, the model does not use the final classification result, but extracts the deep feature output before the classification layer as the semantic feature vector corresponding to the window. This semantic feature vector is used to characterize the comprehensive spatial combination features of multi-source geoscientific information such as geochemistry, geophysics, remote sensing, and geological structure near the known mineral deposits.

[0044] In this embodiment of the invention, the dimension of the first semantic feature vector is 512. When the study area contains 17 known typical mineral deposit points, and only one window containing the mineral deposit point is extracted for each mineral deposit point, 17 first semantic feature vectors can be obtained. When a buffer sampling area is further set with each mineral deposit point as the center, more first semantic feature vectors can be obtained. Specifically, in this embodiment of the invention, the buffer range of the mineral deposit fingerprint database is 31×31 grids, that is, with the grid where the known mineral deposit point is located or its neighboring regular grids as the center, 31×31 grid positions are selected around it, and the corresponding multi-source geoscientific data windows are extracted respectively. The first semantic feature vectors extracted from these windows together constitute the mineral deposit sample fingerprint database.

[0045] S323: Keeping the regular grid window size, step size, and channel order set in step S321 unchanged, determine the regular grid points that fall inside the current candidate target area unit according to the geometric boundary of the current candidate target area unit; then, using the regular grid points as the window center, crop the corresponding multi-source geoscience data windows from the multi-source geoscience data raster layer; input the cropped multi-source geoscience data windows into the deep learning model trained in step S1 to obtain the second semantic feature vector.

[0046] Further explanation is that the second semantic feature vector refers to the deep mineralization semantic representation obtained by the deep learning model trained in the same step S1 after extracting features from the multi-source geoscientific data windows within the candidate target area unit. The second semantic feature vector and the first semantic feature vector use the same input window size, the same multi-source geoscientific data channel order, the same data preprocessing method, and the same deep learning feature extraction model. Therefore, they are located in the same feature space and can be used to calculate cosine similarity.

[0047] In this embodiment of the invention, the current candidate target area unit may contain multiple regular grid points or pixel locations, thus allowing for the cropping of multiple multi-source geoscientific data windows and correspondingly generating multiple second semantic feature vectors. Each second semantic feature vector also has a dimension of 512. These multiple second semantic feature vectors collectively characterize the deep mineralization semantic features of different local locations within the current candidate target area unit.

[0048] S324: Stack the first semantic feature vectors obtained in step S322 by row to form a deposit feature matrix; stack the second semantic feature vectors obtained in step S323 by row to form a target feature matrix of the current candidate target area unit.

[0049] In this embodiment of the invention, the size of the ore deposit feature matrix is ​​M×512, where M represents the number of first semantic feature vectors in the ore deposit sample fingerprint database, and the size of the target area feature matrix is ​​N×512, where N represents the number of second semantic feature vectors extracted within the current candidate target area unit. In this embodiment, both the ore deposit feature matrix and the target area feature matrix are further L2 normalized to eliminate the influence of feature amplitude differences on subsequent similarity calculations.

[0050] S325: Calculate the cosine similarity between the ore deposit feature matrix and the target area feature matrix obtained in step S324 to obtain the cross-similarity matrix. Then, aggregate the cross-similarity matrix to obtain the similarity of the current candidate target area units.

[0051] In this embodiment of the invention, the normalized ore deposit feature matrix H deposit With the normalized target feature matrix H target Multiplying the transposes of the matrices yields the cross-similarity matrix S: S=H deposit ·H target T ; The cross-similarity matrix S has a size of M×N, and each element in the matrix represents the cosine similarity between a first semantic feature vector and a second semantic feature vector.

[0052] In some embodiments, the cross-similarity matrix is ​​aggregated using a local maximum matching mean rule to obtain the similarity. Specifically, in this embodiment, for each row of the cross-similarity matrix S, the maximum value is selected along the target window dimension, representing the highest similarity that the first semantic feature vector can match among all the second semantic feature vectors of the current candidate target unit; thus, a locally optimal matching vector of length M is obtained. Then, the arithmetic mean of each component in this locally optimal matching vector is calculated to obtain the comprehensive similarity of the current candidate target unit relative to a known typical mineral deposit sample fingerprint database.

[0053] In this embodiment of the invention, the comprehensive similarity is further normalized to the [0, 100] interval through linear transformation to obtain the final similarity score. The closer the similarity score is to 100, the more similar the deep mineralization semantic features of the current candidate target area unit are to the deep mineralization semantic features of known typical deposits.

[0054] S326: Repeat steps S323 to S325 to traverse all candidate target units and obtain the similarity corresponding to each candidate target unit.

[0055] In this embodiment of the invention, after training the ResNet-50 model, all model parameters are frozen, and the model is switched to evaluation mode to maintain the stability and consistency of feature extraction. Subsequently, the final classification layer of the model is replaced with an identity mapping, while retaining the complete feature extraction module from the initial convolutional layer, batch normalization layer to the residual block layer, adaptive average pooling layer, and flattening operation. This module serves as a general mineralization feature extractor, providing a foundation for feature similarity score calculation. This feature extractor can map input windows of arbitrary size to fixed-dimensional deep semantic feature vectors, realizing the transformation from multi-source geoscientific data to mineralization pattern representations. In this embodiment of the invention, this feature extractor is used to obtain a first semantic feature vector and a second semantic feature vector.

[0056] S4: The geological logic score and similarity weighted fusion of each candidate target area unit obtained in step S3 are used to obtain the comprehensive score of each candidate target area unit; based on the known mineral deposit distribution in the study area in step S1, the weights of the geological logic score and similarity are optimized according to the comprehensive score to obtain the optimal weights and the corresponding optimal comprehensive score for each candidate target area unit. The overall process of steps S3 and S4 is as follows: Figure 3 As shown.

[0057] To achieve objective assignment of evaluation weights and avoid subjective bias from manual weighting based on experience, this invention employs the Fisher discriminant criterion to optimize the weight coefficients of geological logic scores and similarity. Specifically, in some embodiments, the process of determining the weights based on the known distribution of mineral deposits in the study area in step S4 includes: Using known mineral deposits within the study area as positive samples, and randomly selecting multiple spatially independent background points within the study area as negative samples, the mean and variance of the combined scores for the two classes of samples are calculated respectively: Based on the mean and variance of the two classes of samples, an objective function is constructed according to Fisher's discrimination criterion to maximize the inter-class separation of positive and negative samples and minimize the intra-class aggregation. The optimal weights are selected based on the weights that maximize the objective function. The geological logic score and similarity score are then weighted and fused using these optimal weights to obtain the optimal comprehensive score.

[0058] In this embodiment of the invention, the comprehensive score is obtained by the following formula: Score = α × S logic +(1-α)×S sim ; Where Score represents the overall score, S logic S represents the geological logic score. sim The similarity is represented by α, and the weight is represented by α. Using the target area units containing 17 known typical mineral deposits in the study area as positive samples and 100 randomly selected background area units as negative samples, the mean and variance of the comprehensive scores of the two types of samples were calculated using the above formula. The objective function is: ; Where J(α) represents the objective function, μ pos and μ neg Let σ represent the mean of the positive sample and the mean of the negative sample, respectively. pos 2 and σ neg 2 These represent the variances of the positive and negative samples, respectively. The weight α is set in the interval (0,1), and a grid search is performed with a step size of 0.01 to traverse all possible values ​​of weight α. The optimal weight obtained is 0.58, which indicates that the target area evaluation of this study area needs to take into account both the conformity of the metallogenic geological regularity (weight 58%) and the similarity of the data-driven deep metallogenic model (weight 42%).

[0059] S5: Sort the corresponding optimal comprehensive scores of all candidate target units obtained in step S4, and classify the candidate target units according to the set classification threshold.

[0060] In some embodiments, the grading process includes: Using the optimal weights obtained in step S4, calculate the optimal comprehensive score of the known mineral deposit points in the study area to obtain the optimal comprehensive score distribution; Calculate the receiver operating characteristic curve (ROC) of the optimal comprehensive score distribution, and use the Youden index to determine the grading threshold; By combining the statistical distribution characteristics of the natural breakpoints of the optimal comprehensive score distribution with the classification threshold, target classification is completed.

[0061] In this embodiment of the invention, step S5 specifically includes: The best comprehensive scores of the 27 candidate target units are sorted in descending order. Based on the comprehensive score distribution of the target areas where 17 known typical mineral deposits in the study area are located, the grading threshold was determined to be 76.2 again by using the ROC curve Youden index. Combining the statistical distribution characteristics of natural breakpoints in the optimal comprehensive score distribution with the classification threshold, target classification is completed, specifically including: First-level metallogenic prospective areas: the best comprehensive score is ≥76.2. In this embodiment of the invention, a total of 9 candidate target area units were delineated. These target areas have high similarity in geological and metallogenic regularity, and the deep metallogenic characteristics are highly similar to those of known deposits. They are priority engineering verification target areas. Secondary mineralization prospective areas: 60.0 ≤ optimal comprehensive score < 76.2. In this embodiment of the invention, a total of 9 candidate target area units were delineated, which are key target areas of interest. For the third-level mineralization prospective area: 40.0≤optimal comprehensive score<60.0, the embodiment of the present invention delineates a total of 9 candidate target area units, which are potential reserve target areas.

[0062] The final candidate target differentiation and optimization results generated in this embodiment of the invention are as follows: Figure 4 As shown. From Figure 4 It can be seen that, after comprehensive scoring and ranking, the high-level candidate target areas of this invention are mainly distributed in areas with high mineralization probability, strong geological evidence combinations, and high similarity between deep mineralization characteristics and known typical deposits. Furthermore, their spatial locations show good agreement with known deposits within the study area. This result indicates that the method provided by this invention can effectively identify favorable mineralization areas consistent with known mineralization facts. Relatively low-level candidate target areas typically exhibit weaker mineralization probability, geological logic score, or deep feature similarity, thus resulting in lower overall priority. Compared to manual delineation based solely on mineralization probability prediction maps, the method provided by this invention not only reflects the mineralization favorability of candidate target areas but also demonstrates their matching relationship with known deposit mineralization models, thereby converting multiple candidate areas in a continuous probability field into a prioritized engineering verification sequence. The good correspondence between high-level candidate target areas and known deposits further verifies the effectiveness and geological rationality of the method in target area selection.

[0063] In this embodiment of the invention, the grading results are exported as a grading target area spatial vector boundary file (Shapefile format), and a matching geological attribute table is generated for each target area. The attribute table contains core information such as target area number, area, average mineralization probability, geological logic score, feature similarity score, comprehensive mineralization potential score, and combination of major mineralization anomalies.

[0064] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0065] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for intelligent evaluation of ore-forming target areas, characterized in that, include: S1: Obtain multi-source geoscientific data raster layers of the study area and input the multi-source geoscientific data raster layers into a deep learning model to obtain a mineralization probability prediction map; S2: Perform probabilistic topological watershed segmentation on the mineralization probability prediction map obtained in step S1, dividing the high-probability connected domain into several candidate target area units. S3: Perform a geological logic score on each candidate target area unit obtained in step S2 to reflect the matching degree of geological metallogenic regularity; based on the deep learning model in step S1, determine the similarity between each candidate target area unit and the known mineral deposit points in the study area in step S1; S4: The geological logic score and similarity weighted fusion of each candidate target area unit obtained in step S3 are used to obtain the comprehensive score of each candidate target area unit; based on the known mineral deposit distribution in the study area in step S1, the weights of the geological logic score and similarity are optimized according to the comprehensive score to obtain the optimal weights and the corresponding optimal comprehensive score of each candidate target area unit. S5: Sort the corresponding optimal comprehensive scores of all candidate target units obtained in step S4, and classify the candidate target units according to the set classification threshold.

2. The intelligent evaluation method for ore-forming target areas according to claim 1, characterized in that, In step S1, the deep learning model used is the ResNet-50 model, and the input multi-source geoscience data raster layers include geological layers, geochemical layers, geophysical layers, and remote sensing layers of the study area.

3. The intelligent evaluation method for ore-forming target areas according to claim 1, characterized in that, Step S2 includes: S21: Using known mineral deposit points within the study area as positive samples, randomly select multiple spatially independent background points within the study area as negative samples, generate multiple candidate thresholds to distinguish between positive and negative samples, and calculate the receiver operation feature curve corresponding to each candidate threshold. S22: Based on the receiver operation characteristic curve obtained in step S21, calculate the Youden index corresponding to each candidate threshold, and determine the candidate threshold with the largest Youden index as the optimal binarization threshold; use the optimal binarization threshold to binarize the mineralization probability prediction map to obtain the mineralization prospect mask. S23: Using the mineralization probability map as the three-dimensional terrain surface and the probability value of the mineralization probability prediction map as the elevation value, a three-dimensional mineralization probability field is formed; the gradient amplitude of the three-dimensional mineralization probability field is calculated, and the local maximum points in the three-dimensional mineralization probability field are determined. S24: Based on the local maxima points obtained in step S23, the mineralization foreground mask obtained in step S22 is used to perform image segmentation on the mineralization probability prediction map to obtain several independent candidate target units.

4. The intelligent evaluation method for ore-forming target areas according to claim 1, characterized in that, Step S3, which involves performing geological logic scoring on candidate target units, includes: S311: Using the spatial vector boundary of each candidate target area unit as a mask, obtain the distribution of continuous attribute values ​​covered by the multi-source geoscience data raster layer of each candidate target area unit in step S1; S312: Extract statistics from the continuous attribute value distribution obtained in step S311 that can represent the overall high-value anomaly characteristics of the candidate target area unit; S313: For each multi-source geoscience data raster layer, based on the distribution of effective pixel values ​​of the layer within the study area, the continuous attribute values ​​are discretized into several segments, and the studentized contrast of each segment is calculated; then, based on the segment to which the object-level statistics of each candidate target unit on the corresponding layer belong, the studentized contrast value of the candidate target unit on the corresponding layer is determined. S314: Accumulate the studentized contrast values ​​of each candidate target area unit obtained in step S313 on each multi-source geoscience data raster layer to obtain the geological logic score of each candidate target area unit.

5. The intelligent evaluation method for ore-forming target areas according to claim 4, characterized in that, The process of extracting statistics from the attribute value distribution in step S312 includes: for each candidate target area unit and each multi-source geoscientific data raster layer, using the spatial vector boundary of the candidate target area unit as a mask, extracting all effective pixel attribute values ​​of the candidate target area unit within the coverage area of ​​the corresponding multi-source geoscientific data raster layer; sorting and statistically analyzing the effective pixel attribute values ​​or constructing an empirical cumulative distribution, and using the 90th quantile corresponding to the cumulative frequency reaching 90% as the object-level statistics of the candidate target area unit on the corresponding layer.

6. The intelligent evaluation method for ore-forming target areas according to claim 4, characterized in that, The process of calculating the student-specific contrast of each segment in step S313 includes: Using known mineral deposits within the study area as positive samples, and randomly selecting multiple spatially independent background points within the study area as negative samples, the positive weights of evidential features appearing in the segment are calculated using the following formula: ； Among them, W + B represents positive weights, B represents evidence features appearing in the current segment, and D represents positive samples. Let P(B|D) represent a negative sample, and let P(B|D) represent the probability that the evidence feature appears in a positive sample. This indicates the probability of an evidential feature appearing in a negative sample. An evidential feature is the event that the attribute value in a multi-source geoscientific data raster layer falls into a certain segment. This can be further understood as the event that the attribute value in a geological layer, geochemical layer, geophysical layer, or remote sensing layer falls into a certain segment. The negative weight of a segment where no evidentiary features appear is calculated using the following formula: ； Among them, W - Indicates negative weight. This indicates that no evidentiary features were found in the current section. This indicates the probability that the evidence feature does not appear in the positive sample. This indicates the probability that the evidence feature does not appear in the negative sample. The student-specific contrast is calculated using the following formula: ； Among them, C s σ represents the contrast in studentization. 2 Indicates variance.

7. The intelligent evaluation method for ore-forming target areas according to claim 1, characterized in that, The process of calculating similarity in step S3 includes: S321: According to the set rule grid window, the multi-source geoscience data raster layer is windowed so that the known mineral deposit points are within the range of the grid window; S322: After the result of the division in step S321 is processed by the deep learning model in step S1, the first semantic feature vector containing the known mineral deposit points is obtained. S323: Keeping the regular grid window size, step size, and channel order set in step S321 unchanged, determine the regular grid points that fall inside the current candidate target area unit according to the geometric boundary of the current candidate target area unit; then, using the regular grid points as the window center, crop the corresponding multi-source geoscience data windows from the multi-source geoscience data raster layer; input the cropped multi-source geoscience data windows into the deep learning model trained in step S1 to obtain the second semantic feature vector; S324: Stack the first semantic feature vectors obtained in step S322 by row to form a deposit feature matrix; stack the second semantic feature vectors obtained in step S323 by row to form a target feature matrix for the current candidate target area unit; S325: Calculate the cosine similarity between the ore deposit feature matrix and the target area feature matrix obtained in step S324 to obtain the cross-similarity matrix. Then, aggregate the cross-similarity matrix to obtain the similarity.

8. The intelligent evaluation method for ore-forming target areas according to claim 7, characterized in that, In step S325, the local maximum matching mean rule is used to aggregate the cross-similarity matrix to obtain the similarity.

9. The intelligent evaluation method for ore-forming target areas according to claim 1, characterized in that, Step S4, which involves determining the weights based on the known distribution of mineral deposits in the study area, includes: Using known mineral deposits within the study area as positive samples, and randomly selecting multiple spatially independent background points within the study area as negative samples, the mean and variance of the combined scores for the two classes of samples are calculated respectively: Based on the mean and variance of the two types of samples, an objective function is constructed according to Fisher's discrimination criterion; The optimal weights are selected based on the weights that maximize the objective function. The geological logic score and similarity score are then weighted and fused using these optimal weights to obtain the optimal comprehensive score.

10. The intelligent evaluation method for ore-forming target areas according to claim 1, characterized in that, The grading process in step S5 includes: Using the optimal weights obtained in step S4, calculate the optimal comprehensive score of the known mineral deposit points in the study area to obtain the optimal comprehensive score distribution; Calculate the receiver operating characteristic curve of the optimal comprehensive score distribution, and use the Youden index to determine the grading threshold; By combining the statistical distribution characteristics of the natural breakpoints of the optimal comprehensive score distribution with the classification threshold, target classification is completed.

Images