A Geophysical-Geologically Constrained Method and System for Weak Information Extraction and Skarn-Type Iron Deposit Ore Formation Prediction
By employing a geophysical-geological joint constraint method, low-rank sparse decomposition and multi-source feature fusion were performed on gravity and magnetic data. This solved the problem of extracting weak signals and predicting mineralization of concealed iron deposits under low and gentle gravity and magnetic anomalies, achieving prediction results with high signal-to-noise ratio and geophysical consistency.
Patent Information
- Application Number
- CN202611230468.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-13
- Publication Date
- 2026-11-17
Smart Images

Figure CN123115643A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of mineral resource prediction technology, and more specifically to a method and system for weak information extraction and skarn-type iron ore mineralization prediction based on geophysical-geological joint constraints. Background Technology
[0002] In covered areas and deep within old mines, the exploration of concealed iron ore deposits faces three core challenges: weakness, dispersion, and concealment. Low-level, gentle gravity and magnetic anomalies refer to gravity and magnetic anomaly signals with low amplitude, gentle gradient, and diffuse spatial distribution. Their fundamental cause lies in the large burial depth, medium scale, or gentle dip of the concealed iron ore body, resulting in a much weaker gravity and magnetic response observed at the surface compared to shallower or exposed ore bodies. These anomalies are easily obscured by interference signals from the regional background field and shallow geological bodies, making the identification and extraction of weak mineralization information a long-standing technical bottleneck in exploration geophysics. Specifically, iron ore mineralization prediction under low-level, gentle gravity and magnetic anomaly conditions faces three levels of technical difficulties: First, weak signals have low detectability. When the amplitude of mineral-induced gravity and magnetic anomalies is close to or below the noise level, traditional anomaly detection methods based on amplitude thresholds have difficulty distinguishing between effective signals and random noise, resulting in a high false negative rate. Second, the diversity of anomaly sources is confusing. Low-grade gravity and magnetic anomalies can originate not only from concealed iron ore bodies, but also from non-ore-bearing magnetic or high-density rock bodies, regional tectonic undulations, or human disturbances. The lack of effective anomaly source identification methods makes it difficult to control the false alarm rate. Third, the coupling and superposition of the background field and the anomalous field: In the sedimentary cover area, the undulation of the basement beneath the thick non-magnetic cap layer will generate a large-scale, gentle regional background field, which is highly coupled with the local weak anomalies of the deep ore body in space and frequency, making it difficult for traditional frequency domain filtering methods to achieve effective separation.
[0003] Currently, the technical approaches for extracting weak geophysical information can be mainly divided into the following three categories: 1. Frequency Domain Filtering Methods: Frequency domain filtering methods, represented by high-pass filtering, band-pass filtering, matched filtering, and Wiener filtering, are currently the most widely used techniques for potential field separation. Their basic principle is to utilize the difference in energy distribution between the regional field and the local field in the wavenumber domain, and to achieve separation by designing filters with specific frequency responses. However, these methods have two inherent drawbacks: first, the selection of the cutoff frequency or filtering parameters is highly dependent on subjective experience, and different parameter settings can lead to drastically different separation results; second, frequency domain filtering is based on the global nature of Fourier transform and cannot locally adapt in the spatial domain, easily producing "ringing effects" or introducing truncation artifacts. Especially for non-stationary, slow, weak, and anomalous signals, this method struggles to achieve a precise balance between suppressing noise and preserving the signal. 2. Adaptive methods based on time-frequency analysis: Time-frequency analysis methods, such as empirical mode decomposition, variational mode decomposition, wavelet transform, and curvelet transform, decompose the signal into multiple components with different frequency characteristics, achieving more flexible signal separation than fixed filtering. The advantage of these methods is that they do not require a preset cutoff frequency and have varying degrees of multi-scale analysis capabilities. However, they share common problems: the physical meaning of the decomposed components is unclear, making it impossible to directly distinguish between "mineral-induced anomaly components" and "anomaly components caused by non-mineral geological bodies"; the component selection or reconstruction rules often rely on statistical characteristics or energy criteria, lacking constraints in a geophysical and geological sense. In practical applications, personnel often need to judge which component "may represent mineral-induced anomalies" based on subjective experience, which reduces the objectivity of the method and increases the risk of misjudgment. 3. Low-Rank Sparse Decomposition Methods: In recent years, low-rank sparse decomposition methods, represented by Robust Principal Component Analysis (RPCA) and GoDec (GoDecomposition), have shown new potential in the field of geophysical data processing. These methods decompose a data matrix into a low-rank matrix representing the background field or regional information, and a sparse matrix representing local anomalies or noise. Compared with frequency domain filtering and time-frequency analysis methods, the advantages of low-rank sparse decomposition are: the decomposition results have a clear physical correspondence—the low-rank components vary gently in space, corresponding to the regional background field; the sparse components contain only a small number of non-zero elements, corresponding to the fields generated by local anomalies. This model closely matches the actual composition of geophysical potential field data, thus possessing a natural interpretive advantage.
[0004] However, existing low-rank sparse decomposition methods still have the following fundamental shortcomings in extracting weak information from slow magnetic anomalies: 1. Blindness in Hyperparameter Setting: Both RPCA and GoDec require presetting key hyperparameters—low rank and sparsity cardinality—which determine the complexity of the background field and the number of non-zero elements allowed to be retained in the sparse components. In existing methods, these settings rely entirely on empirical trial and error or cross-validation, without any connection to prior geological or geophysical knowledge. In areas with complex geological conditions, blind parameter setting easily leads to two failure scenarios: if the low rank is preset too high and the sparsity cardinality is preset too low, some mineral-induced weak signals will be incorrectly included in the background field and filtered out; if the low rank is preset too low and the sparsity cardinality is preset too high, some background field information will be mixed into the sparse components, forming false anomalies that interfere with prediction. 2. Spatial Uniformity of Sparse Constraints: The standard GoDec algorithm applies globally uniform and isotropic sparsity constraints to sparse components—that is, it only restricts the total number of non-zero elements in the component, without distinguishing the spatial location and distribution of these non-zero elements. This "one-size-fits-all" constraint completely ignores the spatial differences in geological mineralization patterns: in favorable mineralization locations such as fault intersections and rock mass contact zones, even weak anomalies with low amplitude and scattered shapes have significant prospecting value; while in areas with large-area sedimentary cover and no tectonic activity, anomalies of the same or even higher amplitude are likely to be shallow magnetic inhomogeneities or human interference. The standard GoDec cannot perform this kind of spatial differentiation processing based on geological background.
[0005] 3. Lack of Geophysical Consistency Verification for Decomposition Results: Existing low-rank decomposition methods only optimize for reconstruction error, such as minimizing it. They completely fail to verify whether the decomposed sums are physically plausible. For example, for gravity and magnetic data, there exists a theoretical relationship described by Poisson's theorem, where the ratio of the vertical derivative of the magnetic anomaly to the horizontal gradient of the gravity anomaly is a constant, known as the Poisson ratio. If the Poisson ratios of the decomposed sparse components of the magnetic and gravity anomalies deviate significantly from the statistical range of the known study area, it indicates that the decomposition results are physically unreliable. Existing methods lack any mechanism to guarantee such physical consistency.
[0006] In the mineralization prediction stage, existing technologies mainly include two types of deep learning solutions: 1. Purely data-driven deep learning prediction methods: These methods are represented by convolutional neural networks (CNNs) and their variants (such as ResNet, U-Net, etc.), which regard mineralization prediction as an end-to-end image segmentation or classification problem from multivariate geoscience data to mineralization probability. Its training process relies entirely on labeled known mineral deposits (i.e., positive samples) and randomly selected non-mineral deposits (i.e., negative samples). The network automatically learns the nonlinear mapping relationship between data and labels through backpropagation. The main drawbacks of the pure data-driven method are: First, it is highly sensitive to the number of positive samples—in geological prospecting, the number of known mineral deposits is usually extremely limited, typically only tens to hundreds of boreholes with mineralized points, which is difficult to meet the training requirements of deep networks; Second, the learned features lack interpretability—the features extracted by the network often have "black box" attributes, making it difficult to establish a correspondence with known metallogenic geological elements, such as ore-controlling structures and ore-bearing rock series, making it difficult to verify the reliability of the prediction results; Third, it is prone to geologically unreasonable predictions—such as outputting a high probability of mineralization in thick non-magnetic sedimentary areas, or predicting that the target area has a fragmented "salt and pepper" spatial distribution, which contradicts the spatial continuity that mineralization should have.
[0007] 2. Knowledge-Data Coupled Deep Learning Prediction Method: This method proposes integrating geological knowledge into a deep learning prediction model. Its core techniques include: performing buffer analysis on ore-controlling elements, statistically analyzing the power-law relationship between ore deposit distribution density and distance within each buffer zone, and constructing a knowledge-constrained input layer and a knowledge-constrained loss function based on this.
[0008] However, this type of "knowledge-data coupling" method has two shortcomings: 1. Insufficient Knowledge Dimension: Existing methods rely solely on geostatistical knowledge, lacking constraints from geophysical laws. The "knowledge" utilized is limited to the spatial statistical laws governing ore-controlling factors and deposit distribution, such as the power-law decay relationship that "the closer to a fault, the higher the probability of mineralization." These statistical laws are empirical and qualitative, failing to incorporate quantitative laws of geophysics. For instance, Poisson's theorem for gravity and magnetic anomalies requires a specific quantitative relationship between gravity and magnetic anomalies originating from the same source. This physical law can serve as a strong constraint for determining whether anomalies are "co-originating" or "ore-induced," but existing knowledge coupling methods completely neglect the utilization of such geophysical knowledge.
[0009] 2. Insufficient Data Dimensionality: Existing methods primarily process geochemical or geological layers, lacking in-depth processing of raw geophysical signals. Current methods often input pre-processed rasterized layers, such as geochemical anomaly maps and structural density maps, without directly processing raw geophysical signals. Under conditions of low-grade magnetic anomalies, mineralization information is buried in strong background fields and noise. Without specialized weak signal enhancement and extraction before the data enters the model, deep learning models struggle to effectively learn the mapping relationship between weak mineralization information and mineralization. Existing methods lack a knowledge-guided signal preprocessing module for weak information extraction.
[0010] Based on the above analysis, existing technologies have the following four systemic defects in predicting iron ore formation with low-magnetic anomalies: Existing low-rank decomposition methods for weak information extraction are characterized by "mathematical models but lack of knowledge guidance." Methods like RPCA or GoDec lack prior geological and geophysical knowledge in their hyperparameter settings and constraints, making it difficult to reliably separate mineralization anomalies under low signal-to-noise ratio conditions. Existing knowledge-data coupling prediction methods are characterized by "statistical knowledge but lack of physical constraints." The knowledge is limited to geological statistical laws and does not incorporate geophysical laws. For example, Poisson's theorem and potential field extension frequency band constraints cannot guarantee the physical rationality of the prediction results. Existing prediction processes are characterized by "separation of front-end and back-end." Weak information extraction and deep learning prediction are two independent, unidirectionally connected modules; once the front-end processing parameters are set, they are not adjusted. Even if the back-end model detects that some weak signals have been mistakenly filtered out, it cannot provide feedback to correct the front-end extraction parameters, resulting in irreversible loss of mineralization information. Existing methods lack specific design for iron ore mineralization mechanisms. Neither general low-rank decomposition methods nor general mineralization prediction deep learning models are specifically designed for the specific mineralization geological conditions and specific geophysical response characteristics of iron ore, especially skarn-type deposits, resulting in insufficient domain adaptability. Therefore, how to provide a method and system for weak information extraction and skarn-type iron ore mineralization prediction based on geophysical-geological joint constraints is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0011] In view of this, the present invention provides a method and system for weak information extraction and skarn-type iron ore mineralization prediction based on geophysical-geological joint constraints. Geophysical laws and the geological understanding of skarn-type iron ore mineralization are simultaneously encoded into the constraints within the algorithm, which are integrated throughout the entire process of front-end gravity and magnetic weak information extraction, mid-end multi-source feature fusion, and back-end model training. Furthermore, a closed-loop collaborative iterative optimization mechanism is established between the front-end and back-end to solve the problems of low-rank decomposition without knowledge guidance, lack of physical constraints in prediction, and front-end and back-end separation in existing technologies.
[0012] To achieve the above objectives, the present invention provides the following technical solution: A method for weak information extraction and skarn-type iron ore mineralization prediction based on geophysical-geological joint constraints includes the following steps: S1. Obtain multivariate geoscientific data of the study area, construct a geological complexity matrix to characterize the mineralization favorability, a structural strike tensor field to characterize the direction of ore-controlling structures, and a geological impossibility mask to characterize areas where mineralization is impossible. S2. Using the geological complexity matrix and the structural strike tensor field as constraints, low-rank sparse decomposition is performed on the gravity and magnetic data in multivariate geoscience data to extract mineral-induced weak information components. S3. Using the mineralized weak information component, geological complexity matrix and structural strike tensor field as constraint orbits and multivariate geoscience data as feature orbits, a dual-gated attention mechanism is used to fuse the constraint orbits and feature orbits to obtain multi-source fused features. S4. The prediction network is trained based on a composite loss function that includes geophysical consistency constraints and geological structure constraints. The multi-source fusion features are input into the trained prediction network, and the mineralization prediction probability map is output. S5. Extract high-confidence target areas based on the mineralization prediction probability map, update the geological complexity matrix based on the high-confidence target areas, and repeat S2 to S4 until the preset convergence condition is met to obtain the final mineralization prediction result.
[0013] Optionally, the geological complexity matrix is constructed in S1 as follows: Using the boundary between the intermediate-acidic rock mass and the carbonate rock contact zone as the center, calculate the shortest Euclidean distance from each point in space to the contact zone. Construct the contact zone tightness factor: ; In the formula, For position Contact zone tightness factor, This is the distance attenuation parameter; Based on fracture line features, multi-scale buffer analysis and density calculation are performed to construct the fracture density factor: ; In the formula, For position fracture density factor, For the first k Under each scale The length of the break line or the number of intersections per unit area centered on the center. These are the weighting coefficients for each scale; Lithological favorable factors are assigned based on the lithological category of the rock mass. ; Contact belt tightness factor fracture density factor and lithological favorable factor The geological complexity matrix is obtained by weighted fusion and normalization.
[0014] Optionally, the construction of the orientation tensor field in S1 is specifically as follows: Merge the fracture structural lines and contact zone lines to generate a ore-controlling structural framework line layer; extend the strike vectors on the framework lines to the entire region through spatial interpolation and perform direction consistency processing to obtain the normalized strike vectors at each spatial location; rotate the strike vectors by 90 degrees to obtain the direction vectors perpendicular to the strike of the ore-controlling structures, thus forming the structural strike tensor field. The specific steps for constructing a geological impossibility mask in S1 are as follows: ; In the formula, For position Impossibility mask.
[0015] Optionally, S2 is as follows: Gravity and magnetic data are decomposed into low-rank background field components and overall anomalous field components using low-rank sparse decomposition: In the formula, Represents gravity and magnetic data. Indicates the low-rank background field component. Represents the overall anomaly field components. To represent random noise components, when separating the overall anomaly field, the geological complexity matrix is used to spatially adaptively adjust the threshold of sparse constraints in order to reduce the threshold in favorable mineralization locations. The overall anomalous field component is further decomposed into mineral-induced weakness information components. Non-mineral interference anomaly components : , To represent random noise components, an anisotropic regularization term based on the constructed orientation tensor field is introduced into the decomposition objective function to constrain the mineral-induced weak information components. The spatial morphology is continuously distributed along the direction of the ore-controlling structure; An adaptive parameter optimization strategy combining geophysical and geological methods is used to automatically determine the low-rank sparse decomposition. and sparse cardinality Mineral-induced anomaly base number Non-mineral anomaly base The optimization strategy constructs a joint objective function that includes a data fitting error term, a model complexity penalty term, a geophysical consistency check term, and a geological rationality assessment term. Then, it uses Bayesian optimization to search for the parameter combination that minimizes the joint objective function within the parameter space.
[0016] Optionally, the threshold of the sparse constraint can be spatially adaptively adjusted using the geological complexity matrix, specifically as follows: ; In the formula, For the first t Threshold for sparse constraints in +1 rounds For the standard soft threshold operator, Based on the global threshold, This is a geological complexity adjustment coefficient. Geological complexity matrix The threshold scaling factor that constitutes spatial adaptation. Let be the low-rank background field component of round t.
[0017] Optionally, an anisotropic regularization term based on constructing the orientation tensor field can be introduced into the decomposition objective function. Specifically: ; In the formula, Weak information components for minerals In position gradient vector, For position The unit direction vector, This represents the vector dot product operation. Take the absolute value or L1 norm; The objective of the second, constrained decomposition optimization is: ; In the formula, To achieve a consistent regularization coefficient and control the strength of geological constraints, For the weak information component matrix of the ore The cardinality, specifically, refers to the number of non-zero elements in the matrix. This is a preset upper limit for the base number of the weak information component caused by minerals. Non-mineral interference anomaly component matrix The cardinality represents the number of non-zero elements in the matrix. This is the preset upper limit of the base number for non-mineral interference anomaly components.
[0018] Optionally, S3 employs a dual-gated attention mechanism to fuse the constraint trajectory and the feature trajectory, specifically as follows: The geological complexity matrix in the constraint orbit and constructing the tensor field As input, a geological attention map is generated after passing through two convolutional layers and a sigmoid activation function. In the formula, For the sigmoid function, For activation function, For convolution operations, This represents channel splicing, where multi-source fusion features are obtained through cross-attention operations. : ; In the formula, For cross-attention operations, For element-wise multiplication, As a geological attention enhancer, Geochemical characteristics and remote sensing features The fusion and splicing As a geophysical credibility enhancement factor, and This is a learnable scaling factor.
[0019] Optional, the composite loss function in S4 Spatial adaptive weighted classification loss Geophysical consistency loss , geological structural continuity loss and geological impossibility loss The weighted summation consists of: ; In the formula, , , These are the weighting coefficients for the corresponding items; Spatial Adaptive Weighted Classification Loss Focal Loss is used as the basic classification loss, and the geological complexity matrix is utilized. Constructing spatially adaptive weights: ; In the formula, The total number of pixels, For the model to the first The predicted probability of each pixel. The focusing parameters for Focal Loss. For spatially adaptive sample weights, by modulation: , Based on the weights of positive samples, As a geological complexity adjustment factor; Geophysical consistency loss Defined as: ; In the formula, This is a set of high-probability prediction regions. For indicator functions, The number of pixels in the high-probability prediction area. This is an estimate of Poisson's ratio. The reasonable range of Poisson's ratio for skarn-type iron ore in the known study area; Geological structural continuity loss Using the construction of a direction tensor field Direction vector in Perform the calculation: ; In the formula, Calculate the projection of the probability gradient onto the direction perpendicular to the strike of the ore-controlling structure; Geological Impossibility Loss Using geological impossibility mask Perform the calculation: .
[0020] Optional, S5 specifically includes: S51, Mineralization Prediction Probability Chart The predicted probability is higher than the preset threshold. Connected regions were extracted into a set of high-confidence target regions. ; S52, for each high-confidence target region Calculate the overall credibility score These include anomaly intensity consistency score, geological favorableness consistency score, and geophysical consistency score. S53, Filter out the overall credibility Target areas exceeding a preset confidence threshold are designated as reliable target areas, and the geological complexity matrix is then analyzed. Update: ; In the formula, As a credibility threshold, To update the step size, target area Spatial indicator function, make sure Always stay within range; S54, with the updated geological complexity matrix replace Repeat steps S2 to S4 to obtain a new mineralization prediction probability map. ; S55. Calculate the spatial overlap (IoU) of high-probability regions in the mineralization prediction probability map obtained from two adjacent iterations. If the IoU is greater than the preset convergence threshold or the number of iterations reaches the preset maximum value, stop the iteration and output the mineralization prediction probability map of the current round as the final mineralization prediction result.
[0021] A geophysical-geologically constrained weak information extraction and skarn-type iron ore mineralization prediction system, applying the aforementioned geophysical-geologically constrained weak information extraction and skarn-type iron ore mineralization prediction method, includes: The data acquisition module is used to acquire multi-dimensional geoscientific data of the study area; The structured characterization module, connected to the data acquisition module, is used to construct the geological complexity matrix, the structural strike tensor field, and the geological impossibility mask. The weak information extraction module, connected to the structured characterization module, uses the geological complexity matrix and the tectonic strike tensor field as constraints to perform low-rank sparse decomposition on gravity and magnetic data in multivariate geoscience data and extract mineral-induced weak information components. The multi-source feature fusion module is connected to the weak information extraction module. It uses mineral-induced weak information components and geological complexity matrix as constraint orbits and multivariate geoscience data as feature orbits. It uses a dual-gated attention mechanism to fuse the data and obtain multi-source fused features. The model training and prediction module is connected to the multi-source feature fusion module. It is trained based on a composite loss function that includes geophysical consistency constraints and geological structure constraints. The multi-source fused features are input into the prediction network, and the output is a mineralization prediction probability map. The iterative optimization module, connected to the model training and prediction module and the weak information extraction module, extracts high-confidence target areas based on the mineralization prediction probability map and updates the geological complexity matrix. This triggers the weak information extraction module, the multi-source feature fusion module, and the model training and prediction module to work repeatedly until the convergence condition is met, and then outputs the final mineralization prediction result.
[0022] As can be seen from the above technical solution, compared with the prior art, the present invention provides a method and system for weak information extraction and skarn-type iron ore mineralization prediction based on geophysical-geological joint constraints, which has the following beneficial effects: 1. Significantly Improved Signal-to-Noise Ratio for Weak Information Extraction: This invention effectively preserves weak mineralized signals in favorable mineralization areas such as contact zones and strongly suppresses non-mineralized interference in sedimentary cover areas through spatially adaptive thresholding guided by geological complexity and anisotropic regularization constraints. The expected signal-to-noise ratio improvement is 3-5 dB compared to the GoDec standard. 2. Physical and geological self-consistency of prediction results: Through geophysical consistency loss and geological structure continuity loss, the high-probability target area output by the model satisfies the gravity and magnetic Poisson relationship and is continuously distributed along the ore-controlling structure direction, avoiding the "salt and pepper noise" and physically unreasonable predictions common in pure data-driven methods. 3. Stable generalization under small sample conditions: Multi-level geological knowledge constraints effectively reduce the model's dependence on a large number of labeled samples, and can still maintain stable prediction ability even when the proportion of positive samples is less than 5%. 4. Automated closed-loop iterative optimization: Bayesian optimization automatically determines decomposition parameters, and the closed-loop feedback mechanism simulates the cognitive process of geologists, greatly improving the automation level of the entire process from data processing to target area output. Attached Figure Description
[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0024] Figure 1This is a flowchart of the method for weak information extraction and skarn-type iron ore mineralization prediction according to the present invention; Figure 2 The flowchart of the improved three-level decomposition process for weak information extraction in GoDec is shown below. Figure 3 This is a schematic diagram of the dual-track input and dual-gated attention fusion structure of the present invention; Figure 4 This is a schematic diagram of the closed-loop collaborative iterative optimization mechanism of the present invention; Figure 5 This is a schematic diagram comparing the effects of embodiments of the present invention with existing methods. Detailed Implementation
[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0026] This invention discloses a method for weak information extraction and skarn-type iron ore mineralization prediction based on geophysical-geological joint constraints, such as... Figure 1 As shown, it includes the following steps: S1. Obtain multivariate geoscientific data of the study area, construct a geological complexity matrix to characterize the mineralization favorability, a structural strike tensor field to characterize the direction of ore-controlling structures, and a geological impossibility mask to characterize areas where mineralization is impossible. S2. Using the geological complexity matrix and the structural strike tensor field as constraints, low-rank sparse decomposition is performed on the gravity and magnetic data in multivariate geoscience data to extract mineral-induced weak information components. S3. Using the mineralized weak information component, geological complexity matrix and structural strike tensor field as constraint orbits and multivariate geoscience data as feature orbits, a dual-gated attention mechanism is used to fuse the constraint orbits and feature orbits to obtain multi-source fused features. S4. The prediction network is trained based on a composite loss function that includes geophysical consistency constraints and geological structure constraints. The multi-source fusion features are input into the trained prediction network, and the mineralization prediction probability map is output. S5. Extract high-confidence target areas based on the mineralization prediction probability map, update the geological complexity matrix based on the high-confidence target areas, and repeat S2 to S4 until the preset convergence condition is met to obtain the final mineralization prediction result.
[0027] In this embodiment of the invention, the multivariate geoscientific data acquired in S1 specifically includes: 1. Geophysical Data: Acquire aeromagnetic ΔT data (scale no less than 1:50,000, grid spacing no greater than 100m × 100m) and Bouguer gravity data (scale no less than 1:200,000, grid spacing no greater than 200m × 200m) for the study area, and uniformly resample to the same grid spacing (e.g., 50m × 50m) to establish a unified spatial reference. Perform pole-switching on the aeromagnetic data to eliminate the influence of magnetization direction on anomaly morphology; calculate transformation parameters such as the vertical first derivative, horizontal gradient modulus, and analytical extensions (upward extensions of 100m, 200m, and 500m) for the gravity and aeromagnetic data respectively. These parameters together constitute the geophysical data set; 2. Geological Data: Collect geological maps of the study area at a scale of 1:50,000 or larger, and digitally extract the following ore-controlling elements: fault lines, including strike, dip, length, and other attributes; boundaries of intermediate-acidic rock masses, i.e., diorite, granodiorite, quartz diorite, etc.; distribution range of carbonate rock strata, such as limestone, dolomite, marble, etc.; and contact zone boundaries between rock masses and carbonate rocks. 3. Geochemical data: Collect 1:50,000 geochemical data of the study area and extract indicator element anomaly maps related to skarn-type iron deposits, including but not limited to Fe element anomalies, Co element anomalies, V element anomalies, Ti element anomalies, and Cu-Pb-Zn combination anomalies indicating hydrothermal activity; 4. Remote sensing data: Collect multispectral remote sensing data of the study area, extract iron staining alteration anomaly information and hydroxyl alteration anomaly information, as well as remote sensing linear structure interpretation results.
[0028] Furthermore, the geological complexity matrix is constructed in S1 as follows: Skarn-type iron ore deposits exhibit a clear ore-controlling pattern: the ore bodies are strictly located at or near the contact zone between intermediate-acidic rocks and carbonate rocks, and are subject to the dual control of the contact zone morphology and fault structure. Geological complexity matrix. It is a matrix with the exact same size as the gravity and magnetic data grid. m OK n (column), each element This represents the geological favorableness of a spatial location from the perspective of the mineralization regularity of skarn-type iron ore deposits. The larger the value, the more favorable the geological conditions for mineralization at that location.
[0029] Using the boundary between the intermediate-acidic rock mass and the carbonate rock contact zone as the center, calculate the shortest Euclidean distance from each point in space to the contact zone. Construct the contact zone tightness factor: ; In the formula, For position Contact zone tightness factor, This is a distance attenuation parameter; studies show that most skarn-type iron ore bodies are distributed within a range of 0-500m outward from the contact zone, and the closer to the contact zone, the higher the probability of mineralization. In this embodiment, the typical value is set based on the statistical regularity of known deposits. The physical meaning of this factor is: it takes a value of 1 on the contact strip line and decreases Gaussianly as it moves away from the contact strip.
[0030] Based on fracture line features, multi-scale buffer analysis and density calculation are performed to construct the fracture density factor: ; In the formula, For position fracture density factor, For the first k Under each scale This refers to the length of the break line or the number of intersections within a unit area centered on the center. For example, the buffer zone radii are taken as 100m, 200m, and 500m, respectively. The weighting coefficients for each scale are used. Favorable mineralization sites for skarn-type iron ore deposits are often located at the intersection of fault structures or at the composite sites of faults and contact zones. This is because faults provide channels for the migration of ore-bearing hydrothermal fluids. The multi-scale design aims to capture the control effect of different levels of faults on mineralization, that is, regional deep and large faults control the mineralization zone, while secondary faults control the emplacement space of specific ore bodies.
[0031] Lithological favorable factors are assigned based on the lithological category of the rock mass. ; Contact belt tightness factor fracture density factor and lithological favorable factor The geological complexity matrix is obtained by weighted fusion and normalization.
[0032] In this embodiment, intermediate-acidic rock masses of different compositions exhibit varying mineralization specificity for skarn-type iron ore deposits. Based on statistics of known deposits in the study area, different basic favorable values are assigned to the interior of the rock mass according to lithology (e.g., diorite = 0.8, granodiorite = 0.6, quartz monzonite = 0.5, other lithologies = 0.1), while the sedimentary rock area outside the rock mass is assigned a value of 0. This factor is only used in the construction of... It provides supplementary information about the internal regions of the rock mass.
[0033] The above three factors are weighted and fused to generate a geological complexity matrix: ; in, Based on the fundamental principle that "the contact zone is the main controlling factor" in skarn-type iron ore deposits, the default recommended weight is [weight missing]. , , .
[0034] The fusion result, after normalization to the [0,1] interval: .
[0035] Furthermore, the construction of the directional tensor field in S1 is specifically as follows: Merge fault lines and contact zone lines to generate a ore-controlling structural framework layer; extend the strike vectors on the framework lines to the entire area through spatial interpolation and perform direction consistency processing to obtain a normalized strike vector for each spatial location; rotate the strike vectors by 90 degrees, i.e. , representing the direction perpendicular to the strike of the ore-controlling structure. Signal abrupt changes in this direction will be penalized, resulting in a direction vector perpendicular to the strike of the ore-controlling structure, which constitutes the structural strike tensor field; The morphology and orientation of skarn-type iron ore bodies are controlled by the strike of ore-controlling structures, such as faults and contact zones. (Tectonic strike tensor field) It is a spatially continuous two-dimensional vector field, where the vector at each point is... This indicates the strike direction of the ore-controlling structure at that point, i.e., the normalized unit vector.
[0036] In this embodiment, spatial interpolation specifically involves finding the nearest skeleton line segment to any point in space and assigning the direction vector of that line segment to that point; for the skeleton line segment itself, its direction vector is determined by the direction of the line connecting the two endpoints of the line segment.
[0037] Due to the direction and In a constructive sense, the vector field is treated with sign consistency to ensure that the directional vectors of adjacent regions do not suddenly reverse.
[0038] The specific steps for constructing a geological impossibility mask in S1 are as follows: ; In the formula, For position Impossibility mask.
[0039] The geological impossibility mask marks areas that are "impossible" or "extremely unlikely" to mineralize from the perspective of skarn-type iron ore mineralization theory. For skarn-type iron ore, this mainly includes: large areas covered by Quaternary sediments, where the possibility of a contact zone between the intrusive body and carbonate rocks has been confirmed; thick sedimentary areas within basins, too far from known intrusive bodies or carbonate rock basements; and the core of fracture zones in regional major faults, where ore-forming materials have already escaped. This mask will serve as a penalty term in the back-end loss function, directly suppressing the model from outputting a high probability of mineralization in these areas.
[0040] Furthermore, such as Figure 2As shown, S2 is specifically: Gravity and magnetic data are decomposed into low-rank background field components and overall anomalous field components using low-rank sparse decomposition: In the formula, Represents gravity and magnetic data. This represents the low-rank background field component, indicating a gently varying gravity and magnetic field resulting from the regional geological background. , Represents the overall anomaly field components. To represent random noise components, when separating the overall anomaly field, the geological complexity matrix is used to spatially adaptively adjust the threshold of sparse constraints in order to reduce the threshold in favorable mineralization locations. The overall anomalous field component is further decomposed into mineral-induced weakness information components. Non-mineral interference anomaly components : , To represent random noise components, an anisotropic regularization term based on the constructed orientation tensor field is introduced into the decomposition objective function to constrain the mineral-induced weak information components. The spatial morphology is continuously distributed along the direction of the ore-controlling structure; The localized weak anomaly representing the formation of concealed skarn-type iron ore bodies is characterized by its sparseness and continuity along the ore-controlling structural direction. This represents local anomalies caused by non-mineral factors such as superficial magnetic inhomogeneity and human interference.
[0041] An adaptive parameter optimization strategy combining geophysical and geological methods is used to automatically determine the low-rank sparse decomposition. and sparse cardinality Mineral-induced anomaly base number Non-mineral anomaly base The optimization strategy constructs a joint objective function that includes a data fitting error term, a model complexity penalty term, a geophysical consistency check term, and a geological rationality assessment term. Then, it uses Bayesian optimization to search for the parameter combination that minimizes the joint objective function within the parameter space.
[0042] Furthermore, the threshold for sparse constraints is spatially adaptively adjusted using the geological complexity matrix, specifically as follows: ; In the formula, For the first t Threshold for sparse constraints in +1 rounds For the standard soft threshold operator, Based on the global threshold, This is a geological complexity adjustment coefficient. Geological complexity matrix The threshold scaling factor that constitutes spatial adaptation. Let be the low-rank background field component of round t.
[0043] The sparse components of standard GoDec are updated as follows: The physical meaning of this operation is: to transform the residual matrix... The absolute value is less than Elements with absolute values greater than zero are considered noise. Elements that are considered abnormal signals are treated as such. However, this "one-size-fits-all" threshold setting completely ignores the differences in spatial geological background.
[0044] After the improvement, in Areas with high mineralization values, such as contact zones and fault intersections, are favorable locations for mineralization. Approaching The effective threshold is significantly reduced, and more weak signals are preserved, even if their amplitude is much lower than the conventional threshold; Areas with low values, i.e., sedimentary cover areas and other unfavorable mineralization sites. As the threshold approaches 1, it remains unchanged, and noise is effectively suppressed. This mechanism essentially achieves a differentiated processing strategy of "lowering the detection threshold and increasing the detection probability in geologically favorable locations, and maintaining a high threshold and controlling the false alarm rate in geologically unfavorable locations."
[0045] Furthermore, an anisotropic regularization term based on constructing a directional tensor field is introduced into the decomposition objective function. Specifically: ; In the formula, Weak information components for minerals In position The gradient vector is calculated using the Sobel operator or central difference. For position The unit direction vector, This represents the vector dot product operation. Take the absolute value or L1 norm; The objective of the second, constrained decomposition optimization is: ; In the formula, To achieve a consistent regularization coefficient and control the strength of geological constraints, For the weak information component matrix of the ore The cardinality, specifically the number of non-zero elements in the matrix, represents the sparsity of the spatial distribution of the extracted mineral-induced anomaly signal. The upper limit of the preset cardinality for the mineral-induced weak information component, i.e., the limit allowed by the algorithm. The maximum number of non-zero elements to retain in the ore body. This value can be set based on prior geological knowledge of the known size of the ore body or the proportion of the target exploration area. Non-mineral interference anomaly component matrix The cardinality represents the number of non-zero elements in the matrix. This is a preset upper limit for the base number of non-mineral interference anomaly components, i.e., the number allowed by the algorithm. The maximum number of non-zero elements to retain.
[0046] Standard GoDec only constrains The sparsity, i.e., the upper limit of the number of non-zero elements. There are no restrictions on its spatial distribution. This isotropic sparse constraint does not conform to geological reality. In fact, mineralized anomalies should be continuous along the contact zone and fault strike direction and should not be interrupted, while they can abruptly change in the direction perpendicular to the structural strike because the ore body has a boundary.
[0047] when exist gradient direction at and When the vertical structural strike direction is consistent, the absolute value of the dot product is the largest, indicating an abrupt change across the strike of the ore-controlling structure at that location. This is unfavorable in the metallogenic regularity of skarn-type iron ore deposits, as the ore body will not abruptly cut off at the center of the structural line unless displaced by a later fault. The regularization term imposes the maximum penalty on such signal abrupt changes, suppressing their inclusion. .when gradient direction and When the direction is consistent along the structural strike, the dot product approaches zero. That is, the change in the signal along the structural strike is a normal extension or pinch-out of the ore body and is not penalized.
[0048] Effect: encourage The anomalies are continuously distributed along the contact zone and the direction of the fault. We can extract weak magnetic anomalies that are "beaded" or "banded" and conform to the morphological characteristics of skarn-type iron ore bodies. At the same time, we can effectively suppress non-mineral interference anomalies that are spatially disordered and have no relation to the tectonic framework.
[0049] In this embodiment of the invention, the key hyperparameter in the first two layers of decomposition is the low-rank parameter. and sparse cardinality (First layer), mineral-induced anomaly baseline Non-mineral anomaly base (The second layer) is no longer set by human experience, but is automatically determined by minimizing the geophysical-geological joint objective function; With first-level parameters Taking optimization as an example, construct a joint objective function: ; The meanings and functions of each item are as follows: 1. BIC Item – The Balance Between Model Complexity and Data Fit ; The first term represents the data fitting error, and the second term represents the model complexity penalty. The BIC criterion strikes a balance between fitting accuracy and model complexity to prevent overfitting. , Excessive noise can lead to the inclusion of noise in the effective signal, resulting in underfitting. , Too small a value can lead to the loss of mining information.
[0050] 2. Item—Geophysical Consistency Check
[0051] This test verifies whether the decomposition results satisfy the fundamental laws of geophysics. For the combined gravity and magnetic treatment, the consistency test using the Poisson theorem for gravity and magnetic fields is employed. When the statistical range of the magnetic susceptibility to density ratio (i.e., Poisson's ratio) of the study area is known, the vertical derivative of the magnetic anomaly... A linear relationship should be satisfied between the magnetic anomaly and the gravity anomaly at a specific level. Calculate the Poisson's ratio estimates of the decomposed sparse components of the magnetic anomaly and gravity anomaly at their corresponding spatial locations, and statistically analyze those falling within a known reasonable range. The proportion outside of that, as Measurement: ; in, For the point-by-point estimation of Poisson's ratio, This is the effective anomaly amplitude threshold. The physical basis for this test is: if the separated magnetic anomaly sparse components and gravitational anomaly sparse components do indeed originate from the same field source, i.e., a skarn-type iron ore body, then they should satisfy Poisson's theorem; if... An excessively large value indicates that the decomposition result contains a large number of non-homogeneous interference signals, resulting in poor decomposition quality.
[0052] When only magnetic data is available and no gravity data is available, a potential field extension band consistency check can be performed: for the decomposed sparse components... If downward continuation is performed and the continuation results show divergence or violent oscillation within the known ore body depth range, then it indicates... It contains high-frequency noise that does not conform to known geological models.
[0053] 3. Item - Geological Rationality Assessment ; in, The Pearson correlation coefficient is used. The absolute value of the sparse component represents the anomaly intensity. The idea behind this evaluation is: if the extracted anomaly... If it is indeed a mineral-induced anomaly, then its spatial distribution should be related to geological favorable conditions. A positive correlation—strong anomalies appear in contact zones and fault-developed areas, while the anomalies are weak in sediment-covered areas. A low correlation coefficient indicates... It contains a large number of anomalies unrelated to geological patterns, which may originate from non-mineral disturbances.
[0054] Weighting coefficient , The settings are as follows: and The weighting coefficients for physical and geological constraints range from [0.1, 0.5]. The rationale for this setting is: when reliable gravity and magnetic data for the study area are available and the Poisson's ratio is known, The possible values are relatively large, such as 0.3-0.5; when only magnetic data is available, It can take a small value or be set to 0. The value is based on The reliability is determined.
[0055] Parameter search space for: ; The setting of the search scope is also guided by geological priors: lower limit Based on the complexity estimation of the regional background field of the study area, smaller values are taken for stable blocks with simple structures, and larger values are taken for orogenic belts with complex structures. upper limit The parameters are set based on prior estimates of the known number and size of ore bodies. In practical implementation, grid search or Bayesian optimization can be used. Find the optimal combination of parameters internally.
[0056] Based on the above three-layer design, the complete process of the improved GoDec algorithm with geophysical-geological joint constraints is as follows: Algorithm: Improved GoDec Weak Information Extraction Based on Geophysical-Geological Joint Constraints Input: Gravity and magnetic data matrix Geological complexity matrix Constructing a tensor field ; Output: Mineral-induced weak information components Background field components ; 1. Initialization: Setting the parameter search space Weighting coefficient , , , ; 2. Parameter optimization: In Inner search minimizes the joint objective function of ; 3. First-level decomposition, weighted background field separation: initialization , ; Repeat the following steps until convergence: truncation of SVD before preservation One singular value; Spatial adaptive soft threshold; Check convergence conditions: ; 4. First layer output: Obtain the background field and overall anomaly field ; 5. Determining the parameters for the second layer: Similar to step 2, in Optimal search ; 6. Second-level decomposition, namely, separation of anisotropic mineral-induced information: initialization , ; Repeat the following steps until convergence: renew Solving the problem The optimization problem of regularization terms; renew : ; Check the convergence conditions; 7. Return and ; 2.3.4 Parallel processing of gravity data; Gravity data and magnetic data share the same potential field mathematical properties and are equally applicable to the aforementioned low-rank sparse decomposition framework. For the Bouguer gravity anomaly data matrix... The improved GoDec process, identical to that used for magnetic data analysis, was employed to obtain the gravitational background field in parallel. and gravity minerals weaken information The advantage of parallel processing of gravity and magnetic paths is that subsequent... Physical consistency testing can be performed by cross-validating the corresponding positions of the two results. If the magnetic sparse component and the gravitational sparse component are highly correlated in space and the Poisson's ratio is reasonable, then they are from the same source, and the confidence level of the skarn-type iron ore body will be greatly improved.
[0057] Furthermore, such as Figure 3As shown, S3 employs a dual-gated attention mechanism to fuse the constraint trajectory and the feature trajectory, specifically as follows: The geological complexity matrix in the constraint orbit and constructing the tensor field As input, a geological attention map is generated after passing through two convolutional layers and a sigmoid activation function. In the formula, For the sigmoid function, For activation function, For convolution operations, This represents channel splicing, where multi-source fusion features are obtained through cross-attention operations. : ; In the formula, For cross-attention operations, geophysical augmented features are used as queries, and geochemical and remote sensing augmented features are used as keys and values. Attention weights are calculated and weighted fusion is performed to achieve adaptive interaction between the two types of features. For element-wise multiplication, This is a geological attention enhancement factor that amplifies physical features in geologically favorable locations and maintains its original value in geologically unfavorable locations. Geochemical characteristics and remote sensing features The fusion and splicing As a geophysical reliability enhancement factor, it amplifies the contributions of geochemical and remote sensing features in regions with significant geophysical gradients, while suppressing their contributions in regions with gentle gradients. and As a learnable scaling factor, in this embodiment, the initial value is set to 0.5, and it is automatically adjusted through training.
[0058] In this embodiment of the invention, the constraint orbit contains geophysical features and geological constraint information, which is used to provide spatial attention guidance for multi-source feature fusion, and specifically includes the following channels: S2 outputs information on weakened magnetic mineralization, S2 outputs information on weakened gravity mineralization, S1 constructs the geological complexity matrix, and S1 constructs the structural strike tensor field.
[0059] The feature track contains all the original and derived features of the four types of data, used to provide multidimensional mineralization indication information. Specifically, it includes the following channels: Geophysical characteristics: polarization magnetic anomaly, vertical first derivative of magnetic anomaly, Bouguer gravity anomaly, vertical first derivative of gravity, and background field reflecting the regional tectonic pattern output by S2. Geological features: It also serves as an input channel in the characteristic track, including the contact zone distance field and the fracture density field; Geochemical characteristics: Fe, Co, V, Ti elemental anomalies; Remote sensing features: iron staining alteration anomaly map, hydroxyl alteration anomaly map; The geophysical-geological dual-gated attention fusion mechanism includes a physical feature encoding branch, a geological attention map generation branch, and a dual-gated fusion operation. The physical feature encoding branch uses shallow convolutional networks to perform preliminary feature extraction on different geophysical data channels, obtaining geophysical feature maps. This shallow structure is designed based on the physical nature of geophysical data—the existence of definite potential field transformation relationships between different derived parameters of gravity and magnetic anomalies; excessively deep feature extraction may disrupt this physical correlation. In contact zones and fault-developed areas, the geological gated indicator network should pay special attention to weak anomalies in geophysical data. The physical gated indicator network should only seriously consider surface geochemical and remote sensing anomalies where a physical interface truly exists at depth; otherwise, these surface signals may simply be weathering or anthropogenic interference.
[0060] Since geochemical and remote sensing data primarily reflect surface or near-surface information, geochemical sampling depths are typically <1m, and remote sensing, while reflecting surface conditions, has limited ability to indicate deep, concealed ore bodies. This embodiment introduces the gradient amplitude of geophysical data as a confidence-gating signal: ; in, For polar magnetic anomalies or Bouguer gravity anomalies, this signal exhibits high values near the contact zone at the boundary of the mineralized anomaly, and low values over a large background field area. After normalization to [0,1], This indicates whether the geophysical anomaly gradient at that location is significant. Locations with significant gradients are more likely to have deep physical interfaces.
[0061] The fused feature map is input into the subsequent encoder-decoder prediction network. The encoder backbone network can adopt any deep convolutional network architecture with multi-scale feature extraction capabilities, including but not limited to ResNet-50, ResNet-101, or the DeepLabV3+ encoder with dilated convolutions. The decoder gradually restores the high-level semantic features of the encoder to the original input size through upsampling and skip connections, outputting a mineralization probability map of the same size as the input. , Indicates spatial location The probability of the presence of a skarn-type iron ore body.
[0062] Furthermore, the composite loss function in S4 Spatial adaptive weighted classification loss Geophysical consistency loss , geological structural continuity loss and geological impossibility loss The weighted summation consists of: ; In the formula, , , These are the weighting coefficients for the corresponding items; Spatial Adaptive Weighted Classification Loss Focal Loss is used as the basic classification loss, and the geological complexity matrix is utilized. Constructing spatially adaptive weights: ; In the formula, The total number of pixels, For the model to the first The predicted probability of a pixel if that pixel is a positive sample label. To predict the probability of "mineral existence"; if it is a negative sample label, To predict the probability of "no mineral deposits"; This is the focusing parameter of Focal Loss, which reduces the loss contribution of easily classified samples, allowing the model to focus on difficult-to-classify samples. For spatially adaptive sample weights, by modulation: , This is the base positive sample weight, used to address the imbalance between the total number of positive and negative samples. In this embodiment, it is set to 0.8. This is a geological complexity adjustment factor, which is set to 1.0 in this embodiment; The spatial adaptive weighted classification loss implemented in this method To address the problem of scarce and spatially uneven distribution of positive samples in skarn-type iron ore deposits, it is important to note that positive samples (known ore bodies) in skarn-type iron ore deposits typically account for only a very small proportion of the total area of the study area, usually less than 5%, and are concentrated near the contact zone. Standard binary cross-entropy loss treats all samples equally, which leads to the model being dominated by a large number of easily classified negative samples, such as sedimentary overburden areas, while neglecting the learning of difficult-to-classify samples, such as weak anomaly areas near the contact zone. In areas with high values, such as near the contact zone, the sample weight is amplified—even if the probability of the sample being correctly classified is already high, the model is still forced to give it more attention because these areas are "high-value areas" for mineral exploration, and the cost of underreporting is extremely high. In areas with low values, such as sediment-covered areas, the sample weight remains at the baseline value—even if these areas are falsely reported, the verification cost is low.
[0063] Geophysical consistency loss Defined as: ; In the formula, This is a set of high-probability prediction regions. For indicator functions, The number of pixels in the high-probability prediction area. This is an estimate of Poisson's ratio. The Poisson's ratio range for skarn-type iron ore in the known study area is given. If the model outputs a high mineralization probability in a certain area, but the gravity and magnetic anomalies in that area do not satisfy the co-origin Poisson relationship, it indicates that the prediction is likely a false alarm anomaly, or the magnetic anomaly originates from a shallow magnetic body, or the gravity anomaly originates from a basement uplift rather than a concealed iron ore body. The loss term penalizes such predictions, forcing the model to learn to output physically consistent prediction results. In this embodiment, when magnetic and gravity data are available for the study area, Poisson's theorem is used to verify the physical plausibility of the prediction results, and the predicted probability is... In high-probability regions, the ratio of the vertical derivative of the local magnetic anomaly to the horizontal gradient of the gravity anomaly is calculated point by point to obtain the Poisson's ratio estimate. When only magnetic data is available, the local magnetic anomaly signal in the predicted high-probability region is extended downwards to the known ore body depth. If the extension results show divergence, violent oscillations, or abnormally increased amplitude, it indicates that the signal contains too much high-frequency noise and cannot be a response from a deep ore body. The loss is then changed to: ; in, For the results of downward extension, The maximum allowable amplitude increase ratio, To avoid dividing by zero for small constants.
[0064] Geological structural continuity loss Using the construction of a direction tensor field Direction vector in Perform the calculation: ; In the formula, Calculate the projection of the probability gradient onto the direction perpendicular to the strike of the ore-controlling structure; Geological structural continuity loss The constrained model outputs a mineralization probability map that spatially conforms to the geological pattern of skarn-type iron ore deposits distributed along the ore-controlling structural direction. If the model exhibits drastic probability changes perpendicular to the structural strike, such as a sudden truncation of a high-probability predicted area at the center of the structural line, the projected value will be large, resulting in a severe penalty. By minimizing this loss, the model is encouraged to output continuous high-probability predictions along the strike of contact zones and faults, rather than isolated, fragmented points.
[0065] The geological basis for skarn-type iron ore deposits is that skarn-type iron ore bodies are continuously distributed along the contact zone in a layered, quasi-layered, or lenticular manner, and it is extremely rare for them to suddenly pinch out at the center of the contact zone without any tectonic displacement. This loss term encodes this morphological regularity of the ore body as a differentiable constraint condition.
[0066] Geological Impossibility Loss Using geological impossibility mask Perform the calculation: .
[0067] Geological Impossibility Mask In Quaternary-covered areas and within basins, the value is close to 1; otherwise, it is 0. The physical meaning of this loss is simple and direct: if in If the model outputs a high probability in a region, the loss term contributes significantly; if the model outputs a low probability in these regions, the loss term automatically approaches zero, with no penalty.
[0068] In this embodiment, the weighting coefficients of each loss term are determined using the following strategy: Physical consistency loss weight When reliable gravity and magnetic data are available for the study area, take 0.05-0.10; when only magnetic data are available, take 0.02-0.05, because the potential field extension constraint is not as precise as the Poisson theorem constraint. Structural continuity loss weights The value should be between 0.03 and 0.08. This value should not be too large, otherwise the predicted probability map may become too smooth, losing the true details of the ore body boundaries. Impossibility Loss Weight : Take a value between 0.02 and 0.05. This value represents a soft prohibition, meaning that the model is not forced to output zero probability in these regions, but a significant penalty is imposed on high probability outputs.
[0069] All weight coefficients can be kept fixed during training, or they can be automatically adjusted using methods such as uncertainty weighting or gradient normalization.
[0070] Furthermore, such as Figure 4 As shown, S5 specifically refers to: S51, Mineralization Prediction Probability Chart The predicted probability is higher than the preset threshold. Connected regions were extracted into a set of high-confidence target regions. In this embodiment, The value is 0.75; S52, for each high-confidence target region Calculate the overall credibility score These include anomaly intensity consistency score, geological favorableness consistency score, and geophysical consistency score. In this embodiment, the overall credibility score is used. Specifically: 1. Consistency score for anomaly intensity ; This score measures the similarity between the amplitude of mineralization weakness information in the target area and known ore body anomalies; the closer the value is to 1, the better the amplitude characteristics match.
[0071] 2. Geological Favorability Consistency Score ; This score measures the average geological favorableness of the target area. The closer the value is to 1, the closer the spatial relationship between the target area and the ore-controlling elements of skarn-type iron ore.
[0072] 3. Geophysical Consistency Score
[0073] When gravity and magnetic data are available, calculate the proportion of gravity and magnetic anomalies within the target area that satisfy the Poisson relation: ; Overall credibility ; in, The default weights are 1 / 3 each.
[0074] S53, Filter out the overall credibility Target areas exceeding a preset confidence threshold are designated as reliable target areas, and the geological complexity matrix is then analyzed. Update: ; In the formula, The confidence threshold is set at 0.6. Only target regions with a confidence level higher than this threshold are included in the update process. In this embodiment, the confidence threshold is 0.6. To update the step size and control the adjustment range in each iteration, target area Spatial indicator function, make sure Always stay within range; High-confidence target areas mean that the system is confident that there is a very high probability of an iron ore body at that location. Therefore, the geological favorableness of these locations should be enhanced to guide the next round of GoDec decomposition to further reduce the detection threshold in these areas and mine weaker mineralized signals.
[0075] S54, with the updated geological complexity matrix replace Repeat steps S2 to S4 to obtain a new mineralization prediction probability map. ; S55. Calculate the spatial overlap (IoU) of high-probability regions in the mineralization prediction probability map obtained from two adjacent iterations. If the IoU is greater than the preset convergence threshold or the number of iterations reaches the preset maximum value, stop the iteration and output the mineralization prediction probability map of the current round as the final mineralization prediction result.
[0076] In this embodiment of the invention, the calculation of the target region spatial overlap (IoU) between two adjacent iterations is specifically as follows: ; In the formula, For the first wheel The set of high-probability prediction regions. When (like When the target area spatial distribution has stabilized and the iteration has converged, the final mineralization prediction probability map is output. Also set the maximum number of iterations (e.g., ...). ), to prevent excessive iteration.
[0077] The comparison results between the method of this embodiment and existing methods are as follows: Figure 5 As shown.
[0078] In this embodiment, the low-rank decomposition algorithm can be replaced by RPCA, IALM-RPCA, Semi-NMF, etc., provided that geological constraints are also introduced; the prediction network architecture can be replaced by DenseNet, Swin Transformer, U-Net++, DeepLabV3+, etc.; the Bayesian optimization in the parameter optimization strategy can be replaced by grid search, random search; in the knowledge representation method It can be generated using fuzzy logic or supervised learning methods.
[0079] and Figure 1 Corresponding to the method described above, this embodiment of the invention also discloses a geophysical-geologically constrained weak information extraction and skarn-type iron ore mineralization prediction system. The system, applying the aforementioned geophysical-geologically constrained weak information extraction and skarn-type iron ore mineralization prediction method, includes: The data acquisition module is used to acquire multi-dimensional geoscientific data of the study area; The structured characterization module, connected to the data acquisition module, is used to construct the geological complexity matrix, the structural strike tensor field, and the geological impossibility mask. The weak information extraction module, connected to the structured characterization module, uses the geological complexity matrix and the tectonic strike tensor field as constraints to perform low-rank sparse decomposition on gravity and magnetic data in multivariate geoscience data and extract mineral-induced weak information components. The multi-source feature fusion module is connected to the weak information extraction module. It uses mineral-induced weak information components and geological complexity matrix as constraint orbits and multivariate geoscience data as feature orbits. It uses a dual-gated attention mechanism to fuse the data and obtain multi-source fused features. The model training and prediction module is connected to the multi-source feature fusion module. It is trained based on a composite loss function that includes geophysical consistency constraints and geological structure constraints. The multi-source fused features are input into the prediction network, and the output is a mineralization prediction probability map. The iterative optimization module, connected to the model training and prediction module and the weak information extraction module, extracts high-confidence target areas based on the mineralization prediction probability map and updates the geological complexity matrix. This triggers the weak information extraction module, the multi-source feature fusion module, and the model training and prediction module to work repeatedly until the convergence condition is met, and then outputs the final mineralization prediction result.
[0080] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.
[0081] Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for weak information extraction and skarn-type iron ore mineralization prediction based on geophysical-geological joint constraints, characterized in that, Includes the following steps: S1. Obtain multivariate geoscientific data of the study area, construct a geological complexity matrix to characterize the mineralization favorability, a structural strike tensor field to characterize the direction of ore-controlling structures, and a geological impossibility mask to characterize areas where mineralization is impossible. S2. Using the geological complexity matrix and the structural strike tensor field as constraints, low-rank sparse decomposition is performed on the gravity and magnetic data in multivariate geoscience data to extract mineral-induced weak information components. S3. Using the mineralized weak information component, geological complexity matrix and structural strike tensor field as constraint orbits and multivariate geoscience data as feature orbits, a dual-gated attention mechanism is used to fuse the constraint orbits and feature orbits to obtain multi-source fused features. S4. The prediction network is trained based on a composite loss function that includes geophysical consistency constraints and geological structure constraints. The multi-source fusion features are input into the trained prediction network, and the mineralization prediction probability map is output. S5. Extract high-confidence target areas based on the mineralization prediction probability map, update the geological complexity matrix based on the high-confidence target areas, and repeat S2 to S4 until the preset convergence condition is met to obtain the final mineralization prediction result.
2. The method for weak information extraction and skarn-type iron ore mineralization prediction based on geophysical-geological joint constraints according to claim 1, characterized in that, The geological complexity matrix is constructed in S1 as follows: Using the boundary between the intermediate-acidic rock mass and the carbonate rock contact zone as the center, calculate the shortest Euclidean distance from each point in space to the contact zone. Construct the contact zone tightness factor: ; In the formula, For position Contact zone tightness factor, This is the distance attenuation parameter; Based on fracture line features, multi-scale buffer analysis and density calculation are performed to construct the fracture density factor: ; In the formula, For position fracture density factor, For the first k Under each scale The length of the break line or the number of intersections per unit area centered on the center. These are the weighting coefficients for each scale; Lithological favorable factors are assigned based on the lithological category of the rock mass. ; Contact belt tightness factor fracture density factor and lithological favorable factor The geological complexity matrix is obtained by weighted fusion and normalization.
3. The method for weak information extraction and skarn-type iron ore mineralization prediction based on geophysical-geological joint constraints according to claim 1, characterized in that, The construction of the orientation tensor field in S1 is specifically as follows: Merge the fracture structural lines and contact zone lines to generate a ore-controlling structural framework line layer; extend the strike vectors on the framework lines to the entire region through spatial interpolation and perform direction consistency processing to obtain the normalized strike vectors at each spatial location; rotate the strike vectors by 90 degrees to obtain the direction vectors perpendicular to the strike of the ore-controlling structures, thus forming the structural strike tensor field. The specific steps for constructing a geological impossibility mask in S1 are as follows: ; In the formula, For position Impossibility mask.
4. The method for weak information extraction and skarn-type iron ore mineralization prediction based on geophysical-geological joint constraints according to claim 1, characterized in that, S2 specifically refers to: Gravity and magnetic data are decomposed into low-rank background field components and overall anomalous field components using low-rank sparse decomposition: In the formula, Represents gravity and magnetic data. Indicates the low-rank background field component. Represents the overall anomaly field components. To represent random noise components, when separating the overall anomaly field, the geological complexity matrix is used to spatially adaptively adjust the threshold of sparse constraints in order to reduce the threshold in favorable mineralization locations. The overall anomalous field component is further decomposed into mineral-induced weakness information components. Non-mineral interference anomaly components : , To represent random noise components, an anisotropic regularization term based on the constructed orientation tensor field is introduced into the decomposition objective function to constrain the mineral-induced weak information components. The spatial morphology is continuously distributed along the direction of the ore-controlling structure; An adaptive parameter optimization strategy combining geophysical and geological methods is used to automatically determine the low-rank sparse decomposition. and sparse cardinality Mineral-induced anomaly base number Non-mineral anomaly base The optimization strategy constructs a joint objective function that includes a data fitting error term, a model complexity penalty term, a geophysical consistency check term, and a geological rationality assessment term. Then, it uses Bayesian optimization to search for the parameter combination that minimizes the joint objective function within the parameter space.
5. The method for weak information extraction and skarn-type iron ore mineralization prediction based on geophysical-geological joint constraints according to claim 4, characterized in that, The specific method for spatially adaptively adjusting the threshold of sparse constraints using the geological complexity matrix is as follows: ; In the formula, For the first t Threshold for sparse constraints in +1 rounds For the standard soft threshold operator, Based on the global threshold, This is a geological complexity adjustment coefficient. Geological complexity matrix The threshold scaling factor that constitutes spatial adaptation. Let be the low-rank background field component of round t.
6. The method for weak information extraction and skarn-type iron ore mineralization prediction based on geophysical-geological joint constraints according to claim 4, characterized in that, An anisotropic regularization term based on constructing a directional tensor field is introduced into the decomposition objective function. Specifically: ; In the formula, Weak information components for minerals In position gradient vector, For position The unit direction vector, This represents the vector dot product operation. Take the absolute value or L1 norm; The objective of the second, constrained decomposition optimization is: ; In the formula, To achieve a consistent regularization coefficient and control the strength of geological constraints, For the weak information component matrix of the ore The cardinality, specifically, refers to the number of non-zero elements in the matrix. This is a preset upper limit for the base number of the weak information component caused by minerals. Non-mineral interference anomaly component matrix The cardinality represents the number of non-zero elements in the matrix. This is the preset upper limit of the base number for non-mineral interference anomaly components.
7. The method for weak information extraction and skarn-type iron ore mineralization prediction based on geophysical-geological joint constraints according to claim 1, characterized in that, S3 employs a dual-gated attention mechanism to fuse constraint tracks and feature tracks, specifically as follows: The geological complexity matrix in the constraint orbit and constructing the tensor field As input, a geological attention map is generated after passing through two convolutional layers and a sigmoid activation function. In the formula, For the sigmoid function, For activation function, For convolution operations, This represents channel splicing, where multi-source fusion features are obtained through cross-attention operations. : ; In the formula, For cross-attention operations, For element-wise multiplication, As a geological attention enhancer, Geochemical characteristics and remote sensing features The fusion and splicing As a geophysical credibility enhancement factor, and This is a learnable scaling factor.
8. The method for weak information extraction and skarn-type iron ore mineralization prediction based on geophysical-geological joint constraints according to claim 1, characterized in that, Composite loss function in S4 Spatial adaptive weighted classification loss Geophysical consistency loss , geological structural continuity loss and geological impossibility loss The weighted summation consists of: ; In the formula, , , These are the weighting coefficients for the corresponding items; Spatial Adaptive Weighted Classification Loss Focal Loss is used as the basic classification loss, and the geological complexity matrix is utilized. Constructing spatially adaptive weights: ; In the formula, The total number of pixels, For the model to the first The predicted probability of each pixel. The focusing parameters for Focal Loss. For spatially adaptive sample weights, by modulation: , Based on the weights of positive samples, As a geological complexity adjustment factor; Geophysical consistency loss Defined as: ; In the formula, This is a set of high-probability prediction regions. For indicator functions, The number of pixels in the high-probability prediction area. This is an estimate of Poisson's ratio. The reasonable range of Poisson's ratio for skarn-type iron ore in the known study area; Geological structural continuity loss Using the construction of a direction tensor field Direction vector in Perform the calculation: ; In the formula, Calculate the projection of the probability gradient onto the direction perpendicular to the strike of the ore-controlling structure; Geological Impossibility Loss Using geological impossibility mask Perform the calculation: 。 9. The method for weak information extraction and skarn-type iron ore mineralization prediction based on geophysical-geological joint constraints according to claim 1, characterized in that, S5 specifically refers to: S51, Mineralization Prediction Probability Chart The predicted probability is higher than the preset threshold. Connected regions were extracted into a set of high-confidence target regions. ; S52, for each high-confidence target region Calculate the overall credibility score These include anomaly intensity consistency score, geological favorableness consistency score, and geophysical consistency score. S53. Filter out the overall credibility Target areas exceeding a preset confidence threshold are designated as reliable target areas, and the geological complexity matrix is then analyzed. Update: ; In the formula, As a credibility threshold, To update the step size, target area Spatial indicator function, make sure Always stay within range; S54, with the updated geological complexity matrix replace Repeat steps S2 to S4 to obtain a new mineralization prediction probability map. ; S55. Calculate the spatial overlap (IoU) of high-probability regions in the mineralization prediction probability map obtained from two adjacent iterations. If the IoU is greater than the preset convergence threshold or the number of iterations reaches the preset maximum value, stop the iteration and output the mineralization prediction probability map of the current round as the final mineralization prediction result.
10. A geophysical-geologically constrained weak information extraction and skarn-type iron ore mineralization prediction system, characterized in that, The method for weak information extraction and skarn-type iron ore mineralization prediction based on geophysical-geological joint constraints according to any one of claims 1-9 includes: The data acquisition module is used to acquire multi-dimensional geoscientific data of the study area; The structured characterization module, connected to the data acquisition module, is used to construct the geological complexity matrix, the structural strike tensor field, and the geological impossibility mask. The weak information extraction module, connected to the structured characterization module, uses the geological complexity matrix and the tectonic strike tensor field as constraints to perform low-rank sparse decomposition on gravity and magnetic data in multivariate geoscience data and extract mineral-induced weak information components. The multi-source feature fusion module is connected to the weak information extraction module. It uses mineral-induced weak information components and geological complexity matrix as constraint orbits and multivariate geoscience data as feature orbits. It uses a dual-gated attention mechanism to fuse the data and obtain multi-source fused features. The model training and prediction module is connected to the multi-source feature fusion module. It is trained based on a composite loss function that includes geophysical consistency constraints and geological structure constraints. The multi-source fused features are input into the prediction network, and the output is a mineralization prediction probability map. The iterative optimization module, connected to the model training and prediction module and the weak information extraction module, extracts high-confidence target areas based on the mineralization prediction probability map and updates the geological complexity matrix. This triggers the weak information extraction module, the multi-source feature fusion module, and the model training and prediction module to work repeatedly until the convergence condition is met, and then outputs the final mineralization prediction result.
