A method and system for ore prospecting based on intelligent fusion of multi-source data and target area delineation

Abstract

This invention belongs to the field of geological exploration technology, specifically disclosing a mineral exploration method and system based on intelligent fusion of multi-source data and target area delineation. The method involves collecting multi-source mineral exploration information and preprocessing it under a unified spatiotemporal reference to obtain a multi-source fusion basic dataset; extracting features and performing multimodal representation on the aforementioned dataset to construct a fused feature vector field; obtaining candidate mineralized anomaly areas based on spatial pattern recognition and initial anomaly selection using the aforementioned vector field; extracting the target area boundary point set of the aforementioned anomaly areas using a graph-based geometric reconstruction method to generate an initial target area range; constructing a multi-objective optimization model to iteratively optimize the aforementioned range to obtain the optimal target area delineation result, which is then output to guide deep and peripheral mineral exploration work. The system includes modules for data acquisition, constructing a fused feature vector field, identification and initial selection, generating an initial target area range, and target area delineation. This invention features intelligent efficiency, high target area identification accuracy, and reasonable boundary delineation.

Description

Technical Field

[0001] This invention belongs to the field of geological exploration technology, specifically relating to a mineral exploration method and system based on intelligent fusion of multi-source data and target area delineation, which is intelligent, efficient, has high target area identification accuracy, and reasonable boundary delineation. Background Technology

[0002] In the process of mineral resource exploration, commonly used geophysical, geochemical and remote sensing methods each have their advantages. However, a single data source is often limited by its physical and chemical mechanisms or measurement conditions, making it difficult to fully reflect the comprehensive information of the geological body.

[0003] Multi-source mineral exploration data fusion is a key technology that integrates multi-type, multi-scale, and multi-modal mineral exploration data from geology, geophysics, geochemistry, remote sensing, drilling, and other sources. By eliminating contradictions and redundancies between data and uncovering implicit correlations, it ultimately improves the accuracy, efficiency, and reliability of mineral exploration. Its core objective is to construct an integrated model of "multi-source information - metallogenic regularity - target area prediction," thereby solving problems such as limited information and multiple solutions inherent in traditional single-source data sources. Therefore, multi-source mineral exploration data fusion technology is increasingly widely used in metallogenic prediction.

[0004] However, existing multi-source mineral exploration data fusion methods often remain at the stage of simple superposition or weighted addition, failing to fully utilize the inherent correlation and spatial structural features between data, resulting in the following problems: 1. The multi-source information is not deeply integrated, leading to severe loss of feature information, resulting in blurred boundaries of the obtained anomalies, and ultimately insufficient accuracy in target area identification; 2. The lack of effective geometric reconstruction and optimization mechanisms leads to the target area being easily too large or too small, affecting not only exploration efficiency and accuracy, but also unreasonable boundary delineation; 3. There is often a lack of automated analysis and target area delineation algorithms for high-dimensional, multimodal data, thus relying heavily on manual interpretation, resulting in highly subjective and inefficient results with insufficient intelligence.

[0005] In existing technologies, to address the aforementioned problems in multi-source mineral exploration data fusion, there are methods such as GOCAD and Leapfrog that construct three-dimensional geological models based on borehole, geophysical, and remote sensing data to intuitively display the spatial distribution of strata, structures, and ore bodies. These methods offer strong spatial visualization, support three-dimensional overlay analysis of multi-source data (such as profile cutting and volume calculation), and are convenient for dynamic updates (the model can be corrected in real time with the addition of boreholes). They also assist in reserve estimation with higher accuracy than two-dimensional methods. However, modeling complex structures (such as folds and thrust zones) is difficult and prone to geometric inconsistencies; it also relies heavily on data density (models in sparse borehole areas have low reliability); and the depth of multi-source data fusion is limited (mostly visualization overlays, lacking intelligent correlation analysis). In addition, there are multi-source information integration systems (GIS-based platforms) that use GIS as a spatial framework to integrate geological maps, geophysical maps, geochemical maps, remote sensing maps, and other multi-source maps, supporting functions such as spatial querying, statistical analysis, and buffer analysis. Its data management capabilities are strong, supporting unified storage of multiple formats (Shapefile, GeoTIFF, borehole database); it can flexibly integrate third-party tools (such as evidence rights law, Kriging plugin) to adapt to the needs of different exploration stages; and it is low-cost, making it suitable for small and medium-sized exploration units. However, data quality (such as coordinate accuracy and error range) directly affects the results, requiring strict quality control; moreover, its deep fusion capabilities are limited (mostly map overlay, lacking intelligent algorithm-driven processing); and its real-time performance is insufficient, making it difficult to support rapid decision-making in the field (such as weak mobile integration capabilities). In addition, there is AI and big data fusion technology (machine learning + parallel computing) that utilizes high-performance computing (HPC) to process massive multi-source data (such as TB-level satellite imagery, millions of geochemical samples), combined with deep learning (such as CNN, Transformer) to automatically identify mineralization patterns. Its processing efficiency far exceeds traditional methods (such as GPU-accelerated deep learning can complete hyperspectral classification in minutes); and it can discover hidden patterns that are difficult for the human eye to recognize (such as the nonlinear coupling between geophysical anomalies and remote sensing alteration); it also supports real-time dynamic fusion (such as real-time remote sensing data from UAVs being integrated into model updates). However, it has extremely high requirements for computing resources (GPU clusters, storage), making it difficult for small units to deploy; moreover, model training requires a large amount of labeled data, and rare minerals still face a "data shortage"; and the insufficient embedding of geological logic (such as not considering the control of plate tectonics on mineralization) may lead to pseudo-correlation that is "data-driven but does not conform to geological laws".

[0006] Therefore, there is an urgent need for a new method that can efficiently integrate multi-source data, intelligently identify mineralization anomalies, and accurately delineate target area boundaries to improve the prediction accuracy and reliability of deep and concealed mineral resources. Summary of the Invention

[0007] To address the problems mentioned in the background section, this invention provides a mineral exploration method based on intelligent fusion of multi-source data and target area delineation, which is intelligent, efficient, has high target area identification accuracy, and reasonable boundary delineation. It also provides a mineral exploration system based on intelligent fusion of multi-source data and target area delineation.

[0008] The mineral exploration method based on intelligent fusion of multi-source data and target area delineation of this invention is implemented as follows: It includes data acquisition, construction of a fused feature vector field, identification and initial selection, generation of initial target area range, and target area delineation steps. The specific steps are as follows: A. Data Acquisition: Collect multi-source mineral exploration information, including geophysical data, geochemical data, and remote sensing image data, and preprocess the collected data under a unified spatiotemporal reference to obtain a multi-source fusion basic dataset. B. Constructing a fused feature vector field: Perform feature extraction and multimodal representation on the multi-source fusion basic dataset to construct a fused feature vector field; C. Identification and preliminary selection: Spatial pattern recognition and anomaly selection are performed based on the fused feature vector field to obtain candidate mineralization anomaly areas; D. Generate the initial target area range: The target area boundary point set of the candidate mineralization anomaly area is extracted using a graph-based geometric reconstruction method, and the initial target area range is generated. E. Target Area Delineation: Construct a multi-objective optimization model to iteratively optimize the initial target area range, obtain the optimal target area delineation result, and output it to guide deep and peripheral mineral exploration work.

[0009] Furthermore, in step A, the specific process of preprocessing the collected data under a unified spatiotemporal reference to obtain the multi-source fusion basic dataset is as follows: A10. Noise suppression, outlier removal, and coordinate system adjustment are performed on the collected multi-source mineral exploration information data respectively; A20. Spatial resampling is performed on the collected raster and vector data of different resolutions to achieve uniform grid accuracy; A30. Based on the feature scale matching strategy, the preprocessed geophysical, geochemical and remote sensing data are mapped to the same spatial reference system to form a multi-source fusion basic dataset.

[0010] Furthermore, the specific process of step B is as follows: B10. Use deep feature encoding methods to extract high-dimensional features of various types of data in the multi-source fusion basic dataset; B20. Use a multimodal feature fusion network to jointly map the aforementioned high-dimensional features to obtain a fused feature tensor; B30. Perform spatial interpolation and normalization on the fused feature tensor to form a fused feature vector field covering the target region.

[0011] Furthermore, the specific process of step C, which involves spatial pattern recognition and initial anomaly selection based on the fused feature vector field to obtain candidate mineralization anomaly regions, is as follows: C10. Perform eigenvalue decomposition on the fused feature vector field and extract the main feature components; C20. Based on the main feature components, construct a feature adjacency graph based on spatial similarity calculation, and use spectral clustering method to identify potential abnormal clusters; C30. Project the identified potential anomaly clusters back into three-dimensional space to restore the spatial range of the corresponding anomaly body, i.e., the candidate mineralization anomaly region.

[0012] Furthermore, in step D, a graph-based geometric reconstruction method is used to extract the target area boundary point set of the candidate mineralization anomaly region, specifically as follows: D10. Extract the boundary feature points of the candidate mineralization anomaly zone and calculate the spatial gradient of each feature point; D20. Filter boundary feature points using adaptive thresholds to construct a boundary point set; D30. Perform 3D topological reconstruction on the boundary point set to generate a meshed boundary model; D40. Construct a target area map structure based on a gridded boundary model, and use a community partitioning algorithm to identify relatively independent target area units.

[0013] Furthermore, the specific process of step D40 is as follows: D41. Initialize each grid cell in the meshed boundary model as an independent community; D42. Calculate the modularity gain after merging adjacent communities, and iteratively merge them according to the maximum gain principle; D43. After the iterative merging is completed, extract the set of center points for each community and generate the fitted surface of the target area through principal component analysis. D44. Extract the ridge lines and key nodes on the aforementioned fitted surface as the target area boundary point set.

[0014] Furthermore, in step D, generating the initial target area range involves first performing surface interpolation on the target area boundary point set to obtain an initial envelope surface; then, the initial envelope surface is segmented and discretized to form the initial target area range.

[0015] Furthermore, in step E, a multi-objective optimization model is constructed to iteratively optimize the initial target area range, obtain the optimal target area delineation result, and output it to guide deep and peripheral mineral exploration work. The specific process is as follows: E10. Construct a multi-objective optimization model with target area coverage, boundary compactness, and mineralization anomaly intensity as the core optimization objectives. Iteratively optimize the initial target area range to obtain the Pareto optimal solution set. Based on the needs of the exploration stage, select the optimal target area range from the Pareto optimal solution set as the optimal target area delineation result. E20. The optimal target area range is selected based on the distribution and evaluation index of the Pareto optimal solution set, and is used as the optimal target area delineation result. E30: Output the optimal target area delineation results to guide deep and peripheral mineral exploration work.

[0016] The mineral exploration system based on multi-source data intelligent fusion and target area delineation of this invention is implemented as follows: it includes a data acquisition module, a module for constructing a fused feature vector field, a module for identification and preliminary selection, a module for generating the initial target area range, and a target area delineation module. The data acquisition module is connected to the fusion feature vector field construction module. It is used to collect multi-source mineral exploration information, including geophysical data, geochemical data and remote sensing image data, and to preprocess the collected data under a unified spatiotemporal reference to obtain a multi-source fusion basic dataset. A fusion feature vector field module is constructed and connected to the recognition and preliminary selection module. It is used to extract features and represent multimodal data from a multi-source fusion basic dataset and construct a fusion feature vector field. The identification and initial selection module is connected to the initial target area range generation module. It is used to perform spatial pattern recognition and anomaly initial selection based on the fused feature vector field to obtain candidate mineralized anomaly areas. The module for generating the initial target area range is connected to the target area delineation module. It is used to extract the target area boundary point set of candidate mineralization anomaly areas using a graph-based geometric reconstruction method and generate the initial target area range. The target area delineation module is used to construct a multi-objective optimization model, iteratively optimize the initial target area range, obtain the optimal target area delineation result, and output it to guide deep and peripheral mineral exploration work.

[0017] Furthermore, the module for generating the initial target area range is used to extract boundary feature points of candidate mineralization anomaly areas and calculate the spatial gradient of each feature point; then, boundary feature points are filtered through adaptive thresholding to construct a boundary point set; next, the boundary point set is subjected to three-dimensional topological reconstruction to generate a meshed boundary model; subsequently, each mesh cell in the meshed boundary model is initialized as an independent community; the modularity gain after merging adjacent communities is calculated, and the communities are iteratively merged according to the maximum gain principle; after the iterative merging is completed, the center point set of each community is extracted, and a fitting surface of the target area is generated through principal component analysis; the ridges and key nodes on the aforementioned fitting surface are extracted as the target area boundary point set; finally, the target area boundary point set is subjected to surface interpolation to obtain the initial envelope surface; and the initial envelope surface is segmented and discretized to form the initial target area range.

[0018] The present invention has the following beneficial effects: 1. Existing technologies often rely on simple data overlay or weighting, failing to uncover the inherent correlations within the data and resulting in significant loss of feature information. This invention addresses this deficiency by employing noise suppression, outlier removal, coordinate system unification, and spatial resampling techniques to map multi-source data from geophysics, geochemistry, and remote sensing to a unified spatiotemporal reference and grid precision. This eliminates data inconsistencies and redundancy, laying the foundation for deep fusion and preventing feature misalignment due to differences in data format and resolution, thereby ensuring data consistency. Furthermore, by employing deep feature encoding methods to extract high-dimensional features from various data sources and combining them with a multimodal feature fusion network for joint mapping, the dispersed single-source features are transformed into structured fusion feature tensors. Then, through spatial interpolation and normalization, a fusion feature vector field covering the target area is formed. This approach not only fully utilizes the complementary information between different data (such as geophysical anomalies reflecting deep structures, geochemical anomalies indicating element enrichment, and remote sensing alteration revealing the distribution of alteration zones), effectively preserving the physical meaning and spatial correlation of the data, but also uncovers the nonlinear coupling relationships between data (such as the correlation between geophysical anomalies and remote sensing alteration). This avoids feature fragmentation under traditional overlay methods, significantly reduces information loss, and ultimately achieves a breakthrough in the bottleneck of shallow data fusion and deep correlation of multi-source information, improving the completeness of feature extraction.

[0019] 2. Existing technologies lack effective methods for constructing and optimizing target area boundaries, often resulting in targets that are too large or too small, or have blurred boundaries. This invention innovates a geometric reconstruction and optimization mechanism. It extracts boundary feature points from candidate mineralization anomalies, calculates spatial gradients, and uses adaptive threshold filtering to construct a high-quality set of boundary points. Then, it generates a meshed boundary model through 3D topological reconstruction and introduces a community partitioning algorithm (based on modularity gain iterative merging of mesh cells). Combined with principal component analysis, it generates a fitted surface, extracts ridges and key nodes, ultimately forming an accurate set of target area boundary points. This fully utilizes spatial structural features, avoids boundary distortion caused by traditional 2D modeling or simple 3D overlay, effectively solves the problem of blurred anomaly boundaries, and ensures the rationality of the delineated boundaries. Moreover, this invention employs a non-dominated sorting evolutionary optimization algorithm to iteratively optimize the initial target area range, generating a Pareto optimal solution set. Then, it combines evaluation indicators to select the optimal target area. Compared with the existing "one-time delineation" mode, this invention can balance "target area integrity" and "accuracy," avoiding the blind expansion or contraction of the target area range under single data-driven conditions, as well as the range deviation caused by human experience. This ensures that the target area neither misses potential mineralized areas nor contains too many irrelevant areas. The final output target area boundary is more in line with the actual mineralized body morphology, significantly improving exploration efficiency and accuracy.

[0020] 3. Existing technologies often rely heavily on manual interpretation, resulting in low efficiency, high subjectivity, and insufficient intelligence. This invention extracts key feature components based on the intrinsic decomposition of fused feature vector fields, constructs a feature adjacency graph using spatial similarity calculation, and automatically identifies potential anomaly clusters using spectral clustering algorithms, projecting them back into 3D space to obtain candidate mineralization anomaly areas. This eliminates the need for manual intervention and avoids result biases caused by differences in interpreter experience. Furthermore, its processing efficiency is far superior to traditional manual methods, making it particularly suitable for high-dimensional, multimodal, and massive data scenarios. Moreover, from initial target area generation (graph structure reconstruction) to optimal result output (multi-objective optimization), the entire process is automated, requiring no manual parameter adjustment or boundary correction. Compared to existing GIS systems that require manual tool usage and AI technologies that require manual data annotation, this invention significantly reduces labor costs while ensuring the objectivity and consistency of results, providing efficient decision support for rapid field exploration.

[0021] 4. The mineral exploration system of this invention is divided into five major modules: data acquisition, fusion feature vector field construction, identification and preliminary selection, initial target area generation, and target area delineation. Each module is functionally independent yet closely integrated, which not only reduces operational complexity (e.g., the data acquisition module supports multi-format data access without additional format conversion) but also facilitates later functional operation and maintenance (e.g., when adding drilling data, only the preprocessing algorithm of the data acquisition module needs to be expanded). Addressing the problems of low reliability in sparse borehole areas and difficulty in modeling complex structures in existing technologies (such as 3D geological modeling), this invention deeply integrates multi-source data (remote sensing, geophysical exploration, and geochemical exploration). Even in areas with insufficient borehole data, it can identify hidden mineralization anomalies through multi-source feature correlation. Simultaneously, the 3D target area delineation and optimization mechanism can accurately locate the spatial distribution of deep ore bodies, solving the problem that traditional 2D methods are insufficient for deep exploration and providing technical support for deep mineral resource prediction. Furthermore, compared to the high-cost model of AI and big data integration technology that relies on GPU clusters, the algorithm design of this invention takes into account both efficiency and resource consumption (such as spectral clustering and non-dominated ranking algorithms that do not require high-performance computing support), enabling small and medium-sized exploration units to deploy and use it without additional hardware investment.

[0022] In summary, this invention, through technological innovations such as deep data fusion, precise boundary reconstruction, intelligent full-process operation, and modular system, not only solves the core difficulties of existing multi-source mineral exploration data fusion, but also provides a "high-precision, high-efficiency, and high-reliability" technical solution for mineral resource exploration (especially deep and concealed minerals), and provides key technical support for the accurate prediction of deep and concealed mineral resources. Attached Figure Description

[0023] Figure 1 This is a flowchart of the mineral exploration method based on intelligent fusion of multi-source data and target area delineation according to the present invention. Figure 2This is a flowchart of the multi-objective optimization model in the mineral exploration method of the present invention; Figure 3 This is a schematic diagram of the mineral exploration system based on intelligent fusion of multi-source data and target area delineation according to the present invention. Detailed Implementation

[0024] The present invention will be further described below with reference to embodiments, but this is not intended to limit the present invention in any way. Any changes or improvements made based on the teachings of the present invention shall fall within the protection scope of the present invention.

[0025] like Figure 1 and 2 As shown, the mineral exploration method based on intelligent fusion of multi-source data and target area delineation of this invention includes data acquisition, construction of a fused feature vector field, identification and preliminary selection, generation of initial target area range, and target area delineation steps. The specific steps are as follows: A. Data Acquisition: Collect multi-source mineral exploration information, including geophysical data, geochemical data, and remote sensing image data, and preprocess the collected data under a unified spatiotemporal reference to obtain a multi-source fusion basic dataset. B. Constructing a fused feature vector field: Perform feature extraction and multimodal representation on the multi-source fusion basic dataset to construct a fused feature vector field; C. Identification and preliminary selection: Spatial pattern recognition and anomaly selection are performed based on the fused feature vector field to obtain candidate mineralization anomaly areas; D. Generate the initial target area range: The target area boundary point set of the candidate mineralization anomaly area is extracted using a graph-based geometric reconstruction method, and the initial target area range is generated. E. Target Area Delineation: Construct a multi-objective optimization model to iteratively optimize the initial target area range, obtain the optimal target area delineation result, and output it to guide deep and peripheral mineral exploration work.

[0026] In step A, the specific process of preprocessing the collected data under a unified spatiotemporal reference to obtain the multi-source fusion basic dataset is as follows: A10. Noise suppression, outlier removal, and coordinate system adjustment are performed on the collected multi-source mineral exploration information data respectively; A20. Spatial resampling is performed on the collected raster and vector data of different resolutions to achieve uniform grid accuracy; A30. Based on the feature scale matching strategy, the preprocessed geophysical, geochemical and remote sensing data are mapped to the same spatial reference system to form a multi-source fusion basic dataset.

[0027] It should be noted that in step A10, noise suppression can employ a multi-strategy combined noise reduction technique. For example, in the artificial source electromagnetic method, multi-frequency signals are transmitted through high-order pseudo-random signal encoding, and pure noise data is collected in conjunction with noise reference channel observation technology. Then, the noise is removed from the observation data using a least squares algorithm. For gravity data, an improved nonlocal mean filter (MNLM) can be used to calculate pixel similarity through Euclidean distance, and integral image technology can be combined to reduce computational complexity, effectively removing random noise while preserving geological structural details.

[0028] Outlier removal can employ multivariate statistical methods to identify outliers. For example, a "multivariate outlier map" can be constructed, using Mahalanobis distance to measure the deviation of the sample from the data center, and combining spatial distribution characteristics to distinguish between natural background values ​​and mineralization outliers. Meanwhile, commonly used laboratory methods such as the G-test (based on the normality assumption) or the IQR method (based on quantiles) can be used to quickly screen for univariate outliers. For multivariate data, principal component analysis (PCA) or the isolated forest algorithm should be used for outlier detection.

[0029] One aspect of coordinate systems is unifying data from different sources into the CGCS2000 geodetic coordinate system. For example, for Xi'an 80 coordinate system data, a seven-parameter transformation model (translation, rotation, scaling) is needed to achieve datum transformation; for raster data, the gdalwarp tool in the GDAL library can be used for projection transformation to ensure spatial alignment with UTM or Gauss-Kruger projection; for geophysical data containing elevation information (such as magnetic surveys), the elevation needs to be converted to ellipsoidal height using a geoid model (such as EGM2008) to achieve three-dimensional spatial unification with remote sensing DEM data.

[0030] It should be noted that a multi-resolution fusion strategy can be adopted in step A20. For example, for high-resolution remote sensing images (such as Sentinel-2 with a 10m resolution) and low-resolution geochemical data (such as a 1km grid), the low-resolution data can be upsampled to the high-resolution grid using cubic convolution interpolation, or the high-resolution data can be downsampled to the low-resolution grid using a majority voting method. The specific method depends on the data type: for categorical data (such as lithological classification), nearest neighbor interpolation is used to avoid interpolation generating new categories; while for continuous data (such as elemental concentrations), bilinear interpolation or cubic convolution interpolation is used to balance smoothing effects and detail preservation; for vector data such as geological structural lines, the area overlay method can be used to convert them into raster format, and the raster cell value can be set to structural line density or buffer distance to achieve grid alignment with other raster data.

[0031] It should be noted that the feature-scale matching strategy in step A30 uses the minimum bounding rectangle of the target mineral exploration area as a benchmark and uses ArcGIS's ExtractbyMask tool to trim each data source to ensure that all data cover the same geographical area. For multi-temporal remote sensing data (such as Landsat images from different years), it is necessary to select the temporal phase closest to the acquisition time of geophysical and geochemical data, or eliminate seasonal noise through time series analysis (such as Savitzky-Golay filtering).

[0032] Preprocessed geophysical, geochemical, and remote sensing data are mapped to the same spatial reference frame to form a multi-source fusion dataset. This involves Z-score standardization or Min-Max normalization of geophysical (e.g., resistivity), geochemical (e.g., elemental content), and remote sensing (e.g., NDVI index) data to eliminate dimensional differences and make different modal data comparable on a spatial grid. For example, the ppm unit of geochemical data and the DN value of remote sensing images are uniformly mapped to the [0,1] interval. Principal component analysis (PCA) or independent component analysis (ICA) are used to project the high-dimensional features of multimodal data into a low-dimensional space, extracting principal components that reflect common information such as geological structure and mineralization alteration. For example, in multi-source satellite data fusion, a feature vector is constructed using an improved Normalized Difference Water Index (MNDWI) and bipolar backscattering coefficient to effectively distinguish between water and non-water bodies. Semantic annotation is then applied to the data in conjunction with geological mineralization models. For example, high magnetic anomaly areas in geophysical data are spatially overlaid with fault lines interpreted by remote sensing, and labeled as "tectonic-magmatic activity zones," providing prior knowledge for subsequent pattern recognition. Convolutional neural networks (CNNs) are used to automatically learn deep feature associations between multi-source data. For instance, in haze source tracing research, a cross-attention mechanism is used to realize feature interaction between satellite AOD and ground-based PM2.5 components, identifying the coupling relationship between regional transport paths and local emissions. Ultimately, a multi-source fusion basic dataset is formed.

[0033] The specific process of step B is as follows: B10. Use deep feature encoding methods to extract high-dimensional features of various types of data in the multi-source fusion basic dataset; B20. Use a multimodal feature fusion network to jointly map the aforementioned high-dimensional features to obtain a fused feature tensor; B30. Perform spatial interpolation and normalization on the fused feature tensor to form a fused feature vector field covering the target region.

[0034] It should be noted that in step B10, the deep feature encoding method refers to a feature extraction method based on deep learning techniques (such as Convolutional Neural Networks (CNNs), Transformers, etc.), rather than traditional manually designed features (such as simple statistics or texture operators). Its function is to automatically mine hidden, mineralization-related high-dimensional features (such as nonlinear correlation features and complex spatial pattern features) from different types of raw data (such as gravity anomalies in geophysical data, elemental distribution in geochemical data, and spectral / texture information in remote sensing images) using deep learning models. For example, for remote sensing images, traditional methods might extract average brightness, while deep encoding might extract spectral-texture combinations similar to those around known mineral deposits, which is closer to the needs of mineral exploration.

[0035] High-dimensional features refer to features with higher dimensions (usually hundreds to thousands of dimensions), containing richer information. The original information of single-class data (such as the content of a certain element) is low-dimensional and difficult to reflect the complex process of mineralization (such as the superposition effect of multi-element symbiosis and tectonic-fluid activity); while high-dimensional features can integrate multi-level and multi-related information in the data (such as the combination of high value of element A + low value of element B + specific tectonic direction).

[0036] It should be noted that in step B20, "multimodal" refers to different types of data such as geophysical, geochemical, and remote sensing data (they have different modalities; for example, geophysical exploration uses physical field signals, geochemical exploration uses chemical element signals, and remote sensing uses image signals). The fusion network is a specially designed deep learning network (such as cross-modal attention networks, feature stitching-mapping networks, etc.) used to address the heterogeneity problem of different modal data (such as significant differences in data format, dimensions, and physical meaning). Through joint mapping of the network, high-dimensional features from different modalities are integrated into a unified feature space, preserving the correlation between features of each modality (such as the spatial correspondence between high gravity anomalies in a certain region and elemental anomalies in that region), ultimately outputting a fused feature tensor.

[0037] The fusion feature tensor is a higher-dimensional data structure than vectors (1-dimensional) and matrices (2-dimensional) (usually 3-dimensional or higher, such as including spatial location, feature dimension, and modal correlation). It is an intermediate product after the fusion of multimodal high-dimensional features, containing key information from each individual class of data and integrating the correlation information between different data. It is a comprehensive digital expression of the mineralization-related features of the target area.

[0038] It should be noted that in step B30, due to the potentially different original sampling densities of the multi-source data (e.g., high resolution in remote sensing images and sparse sampling points in geophysical data), the fused feature tensor may have spatially blank areas. Interpolation (e.g., Kriging interpolation, inverse distance weighting) ensures that the feature tensor is continuously distributed within the target region, covering every spatial point. Large differences in the dimensions of features from different modalities (e.g., one feature value ranges from 0 to 100, while another ranges from 0 to 10000) can lead to excessive focus on features with large values ​​in subsequent analyses. Normalization (e.g., mapping to the 0-1 interval) can eliminate the influence of dimensions, allowing for a more balanced weighting of each feature in subsequent calculations. After the above processing, each spatial point within the target region (e.g., a point corresponding to latitude and longitude coordinates) corresponds to a feature vector (containing the fused multimodal high-dimensional features), and the vectors of all spatial points together constitute a vector field.

[0039] Vector field function: It is the basis for subsequent spatial pattern recognition and anomaly selection. By analyzing the differences in feature vectors of different spatial points, it is possible to identify anomaly pattern regions related to mineralization (such as regions whose feature vectors are similar to those of known mineral deposits).

[0040] The specific process of step C, which involves spatial pattern recognition and initial anomaly selection based on the fused feature vector field to obtain candidate mineralization anomaly regions, is as follows: C10. Perform eigenvalue decomposition on the fused feature vector field and extract the main feature components; C20. Based on the main feature components, construct a feature adjacency graph based on spatial similarity calculation, and use spectral clustering method to identify potential abnormal clusters; C30. Project the identified potential anomaly clusters back into three-dimensional space to restore the spatial range of the corresponding anomaly body, i.e., the candidate mineralization anomaly region.

[0041] It should be noted that in step C10, the fused feature vector field is formed by fusing high-dimensional features from multiple sources (geophysical, geochemical, and remote sensing). This can be understood as each spatial point covering the target area carrying a dataset of feature vectors containing multiple types of geological information (similar to how each point in three-dimensional space has a set of numbers describing its geological attributes). The essence of intrinsic decomposition is a mathematical process for reducing the dimensionality of the high-dimensional feature vector field and extracting key information (similar to the approach of principal component analysis (PCA)). Through decomposition, redundant and secondary information is removed from the features, retaining the main feature components that reflect the core characteristics of mineralization anomalies (such as key signals related to mineralization, such as geophysical field anomalies, elemental enrichment features, and remote sensing alteration information). The aim is to simplify the data while preserving core patterns.

[0042] It should be noted that in step C20, spatial similarity calculation measures whether the main feature components of different spatial points are similar (for example, the closer the mineralization-related features of two points are, the higher the similarity). Common calculation methods include cosine similarity and Euclidean distance. A feature adjacency graph treats each spatial point as a node. If the spatial similarity between two nodes exceeds a threshold (i.e., their features are similar), they are connected by edges to form a graph structure describing the spatial feature relationships.

[0043] Spectral clustering is a graph-based clustering algorithm that analyzes the spectral properties (such as eigenvalues ​​and eigenvectors) of an adjacency graph to group nodes with similar features and spatial proximity into clusters (i.e., contiguous regions). Potentially anomalous clusters here refer to contiguous regions whose features differ significantly from their surroundings and may exhibit mineralization (because mineralization leads to local geological anomalies, these clusters are candidate areas for mineralization).

[0044] It should be noted that in step C30, since the preliminary feature processing may be carried out in a high-dimensional feature space (not the actual geographic space), the abnormal clusters obtained by clustering need to be mapped to the actual three-dimensional geographic coordinates (longitude, latitude, and depth) so that they become regions with clear spatial locations and ranges.

[0045] The regions obtained through the above steps, which have clear boundaries and abnormal features in three-dimensional space, serve as the preliminary basis for subsequent target area delineation (further verification and optimization are required), and are thus considered as candidate mineralization anomaly regions.

[0046] In step D, a graph-based geometric reconstruction method is used to extract the target area boundary point set of the candidate mineralization anomaly region, specifically as follows: D10. Extract the boundary feature points of the candidate mineralization anomaly zone and calculate the spatial gradient of each feature point; D20. Filter boundary feature points using adaptive thresholds to construct a boundary point set; D30. Perform 3D topological reconstruction on the boundary point set to generate a meshed boundary model; D40. Construct a target area map structure based on a gridded boundary model, and use a community partitioning algorithm to identify relatively independent target area units.

[0047] It should be noted that in step D10, boundary feature points refer to key points on the edge of candidate mineralization anomaly zones (potentially mineralized areas that have been preliminarily identified). These points exhibit characteristics (such as elemental content, physical field intensity, and remote sensing alteration index) that significantly differ from the surrounding non-anomaly zones, serving as critical points distinguishing anomaly zones from background zones. For example, if the copper content in a region abruptly drops from 500 ppm in the anomaly zone to 50 ppm in the background zone, this sudden change is a typical boundary feature point. Spatial gradient describes the rate and direction of change of a feature in space (similar to the concept of slope). For boundary feature points, calculating their spatial gradient quantifies the severity of the feature abrupt change—the larger the gradient value, the more significant the feature change at that point, and the more likely it is to be a genuine mineralization boundary (rather than a random change caused by noise). For example, the magnetic anomaly gradient at a mineralization boundary is typically much larger than the gradual change in the background zone.

[0048] It should be noted that in step D20, the adaptive threshold differs from a fixed threshold (such as manually setting a gradient value > 10). The adaptive threshold is automatically adjusted based on the distribution characteristics of the data itself. For example, by analyzing the gradient value distribution (such as mean and standard deviation) of all boundary feature points, a reasonable critical value is dynamically determined to ensure that only points with sufficiently large gradients and significant feature mutations are retained. Its purpose is to eliminate pseudo-boundary points caused by noise and measurement errors (such as accidental fluctuations with small gradients), retaining the core points that truly reflect the mineralization boundary, ultimately forming a boundary point set (a series of discrete but crucial boundary points).

[0049] It should be noted that in step D30, 3D topological reconstruction connects discrete sets of boundary points to construct a spatially related 3D structure. Here, topology refers to the connections between points, lines, and surfaces (e.g., which points are adjacent to form an edge and which edges enclose a surface). For example, boundary points at different depths on the Earth's surface and underground are spatially associated to form a 3D outline encompassing the anomaly zone. The meshed boundary model transforms the reconstructed 3D boundary into a regular or irregular mesh (similar to a 3D fishing net). Each mesh cell contains local features of the boundary (e.g., location, gradient direction), transforming the originally abstract set of boundary points into a quantifiable and computable 3D model, providing a structured carrier for subsequent analysis.

[0050] It should be noted that in step D40, the target area graph structure transforms the gridded boundary model into a graph data structure. Each grid cell in the gridded boundary model is considered a node in the graph, and the spatial adjacency relationships between grid cells (such as two grids being physically adjacent or sharing a boundary) are considered edges. The weight of each edge represents the feature similarity between grids (such as the consistency of mineralization intensity). This node-edge graph structure allows for a quantitative description of the spatial relationships between grid cells.

[0051] Community partitioning algorithms are analytical methods in graph theory used to identify tightly connected subgroups in a graph (similar to finding close-knit circles of friends in a social network). The core idea is to divide closely related nodes (grid cells) in the graph into communities (subgroups), ensuring that nodes within the same community are tightly connected, while nodes between different communities are sparsely connected. In mineral exploration scenarios, mineralization anomalies may not be a continuous whole, but rather multiple independent ore bodies (such as multiple veins controlled by different fault zones). This algorithm can divide tightly connected grid cells in the graph into communities, with each community corresponding to a relatively independent target area unit (i.e., a potential independent ore body range). In this invention, the aim is to merge spatially closely related grid cells into relatively independent target area units (i.e., a complete potential mineralization region), avoiding misclassification of spatially unrelated areas as the same target area.

[0052] The specific process of step D40 is as follows: D41. Initialize each grid cell in the meshed boundary model as an independent community; D42. Calculate the modularity gain after merging adjacent communities, and iteratively merge them according to the maximum gain principle; D43. After the iterative merging is completed, extract the set of center points for each community and generate the fitted surface of the target area through principal component analysis. D44. Extract the ridge lines and key nodes on the aforementioned fitted surface as the target area boundary point set.

[0053] In step D, generating the initial target area range involves first performing surface interpolation on the target area boundary point set to obtain the initial envelope surface; then, the initial envelope surface is segmented and discretized to form the initial target area range.

[0054] It should be noted that in step D41, the gridded boundary model refers to the model obtained by performing three-dimensional topological reconstruction on the boundary point set. This transforms the boundary point set of the candidate mineralization anomaly zone into regular or irregular grid cells (similar to a grid puzzle). Each grid cell represents a local region on the boundary, and the grids form an overall structure through spatial relationships (such as adjacency or connection). Initially, each grid cell is considered a separate community (the smallest unit), meaning that the relationships between grids are not considered for the time being; each local grid is treated as an independent analysis object.

[0055] It should be noted that in step D42, modularity is an indicator of the quality of community partitioning: a higher value indicates a closer connection between nodes (grids) within the same community and a sparser connection between different communities (a more reasonable partitioning). Modularity gain refers to the increase in overall modularity after merging two adjacent communities (grid unit groups). If the modularity increases after merging, the merger is more reasonable; otherwise, it is unreasonable. In each iteration, the modularity gain after merging all adjacent communities is calculated, and the group of adjacent communities with the largest gain is selected for merging (i.e., the overall partitioning is more reasonable after merging these two communities); this process is repeated until merging any adjacent community can no longer increase the modularity. In this way, the final community (target unit) is ensured to have the closest spatial connection within its area.

[0056] It should be noted that in step D43, after the iterative merging is completed, each community (target area unit) consists of multiple grid units. The center point set refers to the set of geometric centers of these grid units (such as the centroid of each grid), which is used to represent the spatial distribution characteristics of the community.

[0057] Principal component analysis (PCA) is a dimensionality reduction and feature extraction method. By analyzing the spatial distribution patterns of a set of centroids, it extracts the main distribution trends (such as overall orientation and undulation characteristics) and fits a smooth surface. This surface can approximately represent the spatial morphology of the target area unit (such as the overall outline of the mineralized region).

[0058] It should be noted that in step D44, the ridgeline refers to a relatively prominent and continuous line on the fitted surface (similar to the ridge of a mountain range), which is the skeleton of the surface shape; the key node refers to the turning point, endpoint, or curvature change point of the ridgeline (such as the steepest or gentlest position of the surface). These points can precisely control the boundary shape of the target area unit, and therefore serve as the key control point set of the target area, providing a basis for the subsequent generation of the initial target area range. The target area unit is a prerequisite for the extraction of the final target area boundary point set: only by completing the independent division of the target area unit can we avoid merging multiple unrelated mineralization anomalies (such as two unconnected veins) into the same boundary, ensuring that each independent ore body can extract its own exclusive and accurate boundary point set.

[0059] In step D, generating the initial target area range involves first performing surface interpolation on the target area boundary point set to obtain the initial envelope surface; then, the initial envelope surface is segmented and discretized to form the initial target area range.

[0060] It should be noted that surface interpolation refers to using mathematical algorithms (such as Kriging interpolation, cubic spline interpolation, etc.) to connect a discrete set of key control points (isolated points distributed in space) in a target area into a continuous, smooth surface. The goal is to make the surface conform to the spatial distribution trend of all control points, accurately reflect the overall morphology of the target area (such as the orientation and depth changes of mineralized areas), fill the gaps between control points, and form a complete spatial outline. The initial envelope surface, where "envelope" means to cover, is a continuous surface obtained through interpolation that can completely cover all key control points, providing a preliminary outline of the target area's spatial extent. For example, if the key control points are distributed on the surface of an approximately ellipsoidal mineralized body, the initial envelope surface is an ellipsoidal continuous surface that covers these points, initially defining the spatial boundary of the target area.

[0061] Piecewise discretization refers to the process where a continuous initial envelope surface cannot be directly used as computable and optimizable boundary data, and must be decomposed into a series of discrete segments. In three-dimensional space, this is typically done by dividing the surface into multiple small geometric units (such as triangular patches or quadrilateral meshes); in two-dimensional planar projection, it may be decomposed into polylines or polygonal line segments. This step transforms the continuous surface into discrete boundary units that can be digitally processed. The initial target boundary set refers to the sum of all discrete boundary segments (such as the edges of a 3D mesh or line segments of a 2D polyline) after piecewise discretization. It is the original version of the target area, containing the approximate outline of the target area, but may contain redundant or inaccurate parts, providing basic data for subsequent multi-objective optimization.

[0062] In step E, the multi-objective optimization model comprehensively considers target area coverage, boundary compactness, and mineralization anomaly intensity.

[0063] It should be noted that target area coverage refers to the proportion of the optimized target area that covers the candidate mineralization anomaly areas. High coverage means the target area can contain as much potential mineralization information as possible, reducing the risk of missed detections; however, excessive coverage may lead to an overly large target area, introducing irrelevant regions. Boundary compactness describes the regularity of the target area boundary, usually quantified by indicators such as the ratio of target area area to perimeter. High compactness indicates a more concentrated target area shape (e.g., close to a circle or a regular polygon), reducing the dispersion of the exploration range and lowering fieldwork costs; conversely, loose boundaries increase ineffective exploration areas. Mineralization anomaly intensity refers to the strength of mineralization-related anomaly features within the target area (e.g., elemental enrichment, geophysical field anomaly amplitude, remote sensing alteration index, etc.). High intensity indicates a high mineralization potential in the area, making it a core area that should be prioritized during optimization; if the target area contains too many low-intensity anomalies, it will reduce mineral exploration efficiency.

[0064] In step E, a multi-objective optimization model is constructed to iteratively optimize the initial target area range, obtain the optimal target area delineation result, and output it to guide deep and peripheral mineral exploration work. The specific process is as follows: E10. Construct a multi-objective optimization model with target area coverage, boundary compactness, and mineralization anomaly intensity as the core optimization objectives. Iteratively optimize the initial target area range to obtain the Pareto optimal solution set. Based on the needs of the exploration stage, select the optimal target area range from the Pareto optimal solution set as the optimal target area delineation result. E20. The optimal target area range is selected based on the distribution and evaluation index of the Pareto optimal solution set, and is used as the optimal target area delineation result. E30: Output the optimal target area delineation results to guide deep and peripheral mineral exploration work.

[0065] It should be noted that in step E10, the evolutionary optimization algorithm based on non-dominated sorting is an intelligent algorithm for multi-objective optimization: evolutionary optimization simulates the biological evolution process (such as selection, crossover, and mutation in genetic algorithms), generating better target boundary schemes through iteration; non-dominated sorting is the core mechanism: in multi-objective (coverage, compactness, strength) optimization, if scheme A is not inferior to scheme B in all indicators, and is superior in at least one indicator, then A dominates B. The algorithm retains non-dominated solutions (high-quality schemes not dominated by other schemes) through sorting, ensuring that the iteration direction always advances towards better solutions.

[0066] In multi-objective optimization, there is no perfect solution where all metrics are optimal (e.g., increasing coverage may decrease compactness). The Pareto optimal solution set is a set of solutions that cannot be further improved: for any given solution, improving one metric (e.g., coverage) inevitably leads to a decrease in another metric (e.g., compactness). These solutions constitute the optimal frontier, reflecting the balance between different objectives.

[0067] It should be noted that in step E20, the optimal target area is determined by weighting the distribution and evaluation indicators of the Pareto optimal solution set. The Pareto optimal solution set contains multiple candidate schemes, which need to be screened in combination with actual needs: the distribution of the solution set is analyzed (such as the numerical distribution of each scheme on the three indicators), and the scheme that takes into account the core needs is identified (such as when the exploration budget is limited, the scheme with high compactness and meeting the intensity standard is given priority); combined with knowledge from fields such as geological mineralization models, the target area range most suitable for guiding actual mineral exploration work is finally determined.

[0068] like Figure 3 As shown, the mineral exploration system based on intelligent fusion of multi-source data and target area delineation of the present invention is implemented as follows: it includes a data acquisition module, a module for constructing a fused feature vector field, a module for identification and preliminary selection, a module for generating the initial target area range, and a target area delineation module. The data acquisition module is connected to the fusion feature vector field construction module. It is used to collect multi-source mineral exploration information, including geophysical data, geochemical data and remote sensing image data, and to preprocess the collected data under a unified spatiotemporal reference to obtain a multi-source fusion basic dataset. A fusion feature vector field module is constructed and connected to the recognition and preliminary selection module. It is used to extract features and represent multimodal data from a multi-source fusion basic dataset and construct a fusion feature vector field. The identification and initial selection module is connected to the initial target area range generation module. It is used to perform spatial pattern recognition and anomaly initial selection based on the fused feature vector field to obtain candidate mineralized anomaly areas. The module for generating the initial target area range is connected to the target area delineation module. It is used to extract the target area boundary point set of candidate mineralization anomaly areas using a graph-based geometric reconstruction method and generate the initial target area range. The target area delineation module is used to construct a multi-objective optimization model, iteratively optimize the initial target area range, obtain the optimal target area delineation result, and output it to guide deep and peripheral mineral exploration work.

[0069] The module for generating the initial target area range is used to extract boundary feature points of candidate mineralization anomaly areas and calculate the spatial gradient of each feature point. Then, boundary feature points are filtered through adaptive thresholding to construct a boundary point set. Next, the boundary point set is reconstructed in three dimensions to generate a gridded boundary model. Subsequently, each grid cell in the gridded boundary model is initialized as an independent community. The modularity gain after merging adjacent communities is calculated and iteratively merged according to the maximum gain principle. After the iterative merging is completed, the center point set of each community is extracted, and a fitting surface of the target area is generated through principal component analysis. The ridges and key nodes on the aforementioned fitting surface are extracted as the target area boundary point set. Finally, the target area boundary point set is interpolated to obtain the initial envelope surface. The initial envelope surface is then segmented and discretized to form the initial target area range.

[0070] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A mineral exploration method based on intelligent fusion of multi-source data and target area delineation, characterized in that: The process includes data acquisition, construction of a fused feature vector field, identification and initial selection, generation of the initial target area, and target delineation. The specific steps are as follows: A. Data Acquisition: Collect multi-source mineral exploration information, including geophysical data, geochemical data, and remote sensing image data, and preprocess the collected data under a unified spatiotemporal reference to obtain a multi-source fusion basic dataset. B. Constructing a fused feature vector field: Perform feature extraction and multimodal representation on the multi-source fusion basic dataset to construct a fused feature vector field; C. Identification and preliminary selection: Spatial pattern recognition and anomaly selection are performed based on the fused feature vector field to obtain candidate mineralization anomaly areas; D. Generate the initial target area range: The target area boundary point set of the candidate mineralization anomaly area is extracted using a graph-based geometric reconstruction method, and the initial target area range is generated. E. Target Area Delineation: Construct a multi-objective optimization model to iteratively optimize the initial target area range, obtain the optimal target area delineation result, and output it to guide deep and peripheral mineral exploration work.

2. The mineral exploration method based on intelligent fusion of multi-source data and target area delineation according to claim 1, characterized in that: In step A, the specific process of preprocessing the collected data under a unified spatiotemporal reference to obtain the multi-source fusion basic dataset is as follows: A10. Noise suppression, outlier removal, and coordinate system adjustment are performed on the collected multi-source mineral exploration information data respectively; A20. Spatial resampling is performed on the collected raster and vector data of different resolutions to achieve uniform grid accuracy; A30. Based on the feature scale matching strategy, the preprocessed geophysical, geochemical and remote sensing data are mapped to the same spatial reference system to form a multi-source fusion basic dataset.

3. The mineral exploration method based on intelligent fusion of multi-source data and target area delineation according to claim 1, characterized in that: The specific process of step B is as follows: B10. Use deep feature encoding methods to extract high-dimensional features of various types of data in the multi-source fusion basic dataset; B20. Use a multimodal feature fusion network to jointly map the aforementioned high-dimensional features to obtain a fused feature tensor; B30. Perform spatial interpolation and normalization on the fused feature tensor to form a fused feature vector field covering the target region.

4. The mineral exploration method based on intelligent fusion of multi-source data and target area delineation according to claim 1, characterized in that: The specific process of step C, which involves spatial pattern recognition and initial anomaly selection based on the fused feature vector field to obtain candidate mineralization anomaly regions, is as follows: C10. Perform eigenvalue decomposition on the fused feature vector field and extract the main feature components; C20. Based on the main feature components, construct a feature adjacency graph based on spatial similarity calculation, and use spectral clustering method to identify potential abnormal clusters; C30. Project the identified potential anomaly clusters back into three-dimensional space to restore the spatial range of the corresponding anomaly body, i.e., the candidate mineralization anomaly region.

5. The mineral exploration method based on intelligent fusion of multi-source data and target area delineation according to claim 1, characterized in that: In step D, a graph-based geometric reconstruction method is used to extract the target area boundary point set of the candidate mineralization anomaly region, specifically as follows: D10. Extract the boundary feature points of the candidate mineralization anomaly zone and calculate the spatial gradient of each feature point; D20. Filter boundary feature points using adaptive thresholds to construct a boundary point set; D30. Perform 3D topological reconstruction on the boundary point set to generate a meshed boundary model; D40. Construct a target area map structure based on a gridded boundary model, and use a community partitioning algorithm to identify relatively independent target area units.

6. The mineral exploration method based on intelligent fusion of multi-source data and target area delineation according to claim 5, characterized in that: The specific process of step D40 is as follows: D41. Initialize each grid cell in the meshed boundary model as an independent community; D42. Calculate the modularity gain after merging adjacent communities, and iteratively merge them according to the maximum gain principle; D43. After the iterative merging is completed, extract the set of center points for each community and generate the fitted surface of the target area through principal component analysis. D44. Extract the ridge lines and key nodes on the aforementioned fitted surface as the target area boundary point set.

7. The mineral exploration method based on intelligent fusion of multi-source data and target area delineation according to claim 6, characterized in that: In step D, generating the initial target area range involves first performing surface interpolation on the target area boundary point set to obtain the initial envelope surface; then, the initial envelope surface is segmented and discretized to form the initial target area range.

8. The mineral exploration method based on intelligent fusion of multi-source data and target area delineation according to any one of claims 1 to 7, characterized in that: In step E, a multi-objective optimization model is constructed to iteratively optimize the initial target area range, obtain the optimal target area delineation result, and output it to guide deep and peripheral mineral exploration work. The specific process is as follows: E10. Construct a multi-objective optimization model with target area coverage, boundary compactness, and mineralization anomaly intensity as the core optimization objectives. Iteratively optimize the initial target area range to obtain the Pareto optimal solution set. Based on the needs of the exploration stage, select the optimal target area range from the Pareto optimal solution set as the optimal target area delineation result. E20. The optimal target area range is selected based on the distribution and evaluation index of the Pareto optimal solution set, and is used as the optimal target area delineation result. E30: Output the optimal target area delineation results to guide deep and peripheral mineral exploration work.

9. A mineral exploration system based on intelligent fusion of multi-source data and target area delineation, characterized in that: It includes a data acquisition module, a fusion feature vector field construction module, an identification and initial selection module, an initial target area generation module, and a target area delineation module. The data acquisition module is connected to the fusion feature vector field construction module. It is used to collect multi-source mineral exploration information, including geophysical data, geochemical data and remote sensing image data, and to preprocess the collected data under a unified spatiotemporal reference to obtain a multi-source fusion basic dataset. A fusion feature vector field module is constructed and connected to the recognition and preliminary selection module. It is used to extract features and represent multimodal data from a multi-source fusion basic dataset and construct a fusion feature vector field. The identification and initial selection module is connected to the initial target area range generation module. It is used to perform spatial pattern recognition and anomaly initial selection based on the fused feature vector field to obtain candidate mineralized anomaly areas. The module for generating the initial target area range is connected to the target area delineation module. It is used to extract the target area boundary point set of candidate mineralization anomaly areas using a graph-based geometric reconstruction method and generate the initial target area range. The target area delineation module is used to construct a multi-objective optimization model, iteratively optimize the initial target area range, obtain the optimal target area delineation result, and output it to guide deep and peripheral mineral exploration work.

10. The mineral exploration system based on intelligent fusion of multi-source data and target area delineation according to claim 9, characterized in that: The module for generating the initial target area range is used to extract the boundary feature points of the candidate mineralization anomaly area and calculate the spatial gradient of each feature point; then, the boundary feature points are filtered by an adaptive threshold to construct a boundary point set; next, the boundary point set is reconstructed in three dimensions to generate a meshed boundary model; then, each mesh cell in the meshed boundary model is initialized as an independent community. Calculate the modularity gain after merging adjacent communities and iteratively merge them according to the maximum gain principle; after the iterative merging is completed, extract the set of center points of each community and generate the fitting surface of the target area through principal component analysis; extract the ridges and key nodes on the aforementioned fitting surface as the set of boundary points of the target area; finally, perform surface interpolation on the set of boundary points of the target area to obtain the initial envelope surface; and then discretize the initial envelope surface into segments to form the initial target area range.

Images