A spatial intelligence based mineral prospectivity method

Abstract

The application discloses a mineral prediction method based on spatial intelligence, relates to the technical field of spatial intelligence, and comprises the following steps: identifying the spatial relationship of a spatial entity element list, constructing a topological connected graph, extracting connected structure attributes, executing semantic relationship reasoning, and generating an object-level spatial intelligent scene graph; based on the object-level spatial intelligent scene graph, constructing an anisotropic cost field, calculating a minimum cost channel distance, mapping into a channel weight expression, and generating a channel weight layer; performing spatial alignment and multi-modal spatial correlation feature fusion on a multi-source mineralization evidence data set, the object-level spatial intelligent scene graph and the channel weight layer, executing mineral inference through a mineral prediction machine learning model, and generating a mineral prediction layer. The application realizes dynamic modeling of a mineralization path, and improves prediction accuracy and engineering practicability.

Description

Technical Field

[0001] This invention relates to the field of spatial intelligence technology, and in particular to a mineral prediction method based on spatial intelligence. Background Technology

[0002] Mineral resource prediction technology has evolved from early qualitative analysis relying on geologists' experience to a digital analysis system integrating multi-source spatial data and intelligent algorithms. With the widespread adoption of Geographic Information Systems (GIS) and breakthroughs in remote sensing technology, mineral resource prediction has undergone a paradigm shift from manually drawn geological maps to spatial database management, improving the efficiency of regional geological information integration and prediction accuracy. In recent years, the deep application of machine learning algorithms has further promoted the intelligentization of mineral resource prediction models, constructing data-driven prediction frameworks, optimizing resource exploration decision-making processes, and providing technical support for the sustainable development of global mineral resources. This marks an upgrade in mineral resource prediction from experience-based to data-intelligent-based.

[0003] However, mineral prediction technologies still have limitations in multi-source data fusion and spatial relationship modeling. First, multi-source mineralization evidence data often suffers from differences in coordinate systems, spatial resolution mismatches, and missing values, resulting in a lack of standardized preprocessing mechanisms for data integration, which limits the efficiency and reliability of extracting key mineralization clues. Second, the depth of spatial relationship processing is insufficient. Traditional technologies rely solely on geometric superposition, neglecting the topological connectivity and semantic dynamic interaction between spatial entities during the mineralization process. This makes it difficult to construct high-order spatial intelligent scenarios, thus restricting the interpretability and engineering practicality of predictions in complex tectonic areas. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a spatial intelligence-based mineral prediction method to address the problems of imperfect multi-source data fusion mechanisms and insufficient dynamism in spatial relationship processing.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides a spatial intelligence-based mineral prediction method, comprising: collecting a multi-source mineralization evidence dataset, extracting candidate spatial entity elements, performing centerline segmentation encoding, partitioning and boundary fitting, and calculating spatial entity element attributes to generate a list of spatial entity elements; identifying spatial relationships in the list of spatial entity elements, constructing a topological connectivity graph, extracting connectivity structure attributes, performing semantic relationship reasoning, and generating an object-level spatial intelligent scene graph; based on the object-level spatial intelligent scene graph, constructing an anisotropic cost field, calculating the minimum cost channel distance, mapping it to channel weight expression, and generating a channel weight layer; spatially aligning the multi-source mineralization evidence dataset, the object-level spatial intelligent scene graph, and the channel weight layer, fusing multimodal spatial association features, and performing mineral inference through a mineral prediction machine learning model to generate a mineral prediction layer.

[0008] As a preferred embodiment of the spatial intelligence-based mineral prediction method of the present invention, the specific steps for collecting multi-source mineralization evidence datasets are as follows:

[0009] Collect multi-source mineralization evidence data, perform consistency verification and coordinate system unification, and generate a coordinate system unification mineralization evidence package;

[0010] Spatial resolution resampling transformation and raster pixel scale unification processing are performed on the coordinate system mineralization evidence package to generate a unified resampling package;

[0011] The resampled unified package is subjected to grid alignment and grid starting point locking to obtain the grid aligned volume. Missing value processing and outlier removal are then performed to generate a multi-source mineralization evidence dataset.

[0012] As a preferred embodiment of the spatial intelligence-based mineral prediction method of the present invention, the specific steps of extracting candidate spatial entity elements and performing centerline segmentation encoding, partitioning, and boundary fitting are as follows:

[0013] Candidate linear features, candidate boundary features, and candidate anomaly features are extracted from multi-source mineralization evidence datasets, and cross-modal spatial response alignment and confidence propagation screening are performed to generate a candidate spatial entity feature layer.

[0014] Refine and repair the candidate linear features in the candidate spatial entity feature layer to obtain the connected skeleton, and perform segmentation and centerline segmentation encoding to generate a segmented encoding set;

[0015] Partitioning and boundary fitting are performed on candidate boundary features and candidate anomaly features to obtain fitted boundary curves. Spatial adsorption and position correction are then performed using a segmented encoding set to generate a boundary fitting set.

[0016] As a preferred embodiment of the spatial intelligence-based mineral prediction method of the present invention, the spatial entity element list is generated by calculating the spatial entity element attributes based on the boundary fitting set and the segmented coding set, and writing the source and quality tags.

[0017] As a preferred embodiment of the spatial intelligence-based mineral prediction method of the present invention, the specific steps for identifying the spatial relationships of the spatial entity element list and constructing a topological connectivity graph are as follows:

[0018] Based on the list of spatial entity elements, a candidate neighborhood retrieval range is constructed under a unified coordinate system, and spatial relationship identification between objects is performed to generate a spatial relationship record set;

[0019] The list of spatial entity elements and the set of spatial relationship records are registered as topological nodes and topological edges, respectively. The spatial relationships between duplicate objects are merged and prioritized to generate a topological connectivity graph.

[0020] As a preferred embodiment of the spatial intelligence-based mineral prediction method of the present invention, the specific steps for generating the object-level spatial intelligence scene map are as follows:

[0021] The connectivity structure attributes are extracted from the topological connectivity graph, and topological connectivity confidence propagation and semantic constraint adjudication are performed to generate a connectivity structure attribute table.

[0022] By combining the list of spatial entity elements, the set of spatial relationship records, the topological connectivity graph, and the connectivity structure attribute table, and performing semantic relationship reasoning and structured assembly, an object-level spatial intelligent scene graph is generated.

[0023] As a preferred embodiment of the spatial intelligence-based mineral prediction method of the present invention, the specific steps for constructing an anisotropic cost field based on an object-level spatial intelligence scene graph are as follows:

[0024] Based on object-level spatial intelligent scene map, spatial entity elements are filtered and the center line of the channel is rasterized and snapped to generate the starting point raster set.

[0025] By combining the starting grid set with semantic relations, weighted correction of the barrier boundary direction is performed to construct the initial anisotropic cost field, and direction-sensitive cost recalibration is performed to generate the anisotropic cost field.

[0026] As a preferred embodiment of the spatial intelligence-based mineral prediction method of the present invention, the specific steps for generating the channel weight layer are as follows:

[0027] Based on the anisotropic cost field, the minimum cost channel distance is calculated, and the intersection region is updated and the blocking and suppression constraints are applied to generate a channel distance map.

[0028] The minimum cost channel distance in the channel distance graph is mapped to a channel weight representation, and weight enhancement, weight suppression, upper limit pruning, and spatial smoothing are performed to generate a channel weight layer.

[0029] As a preferred embodiment of the spatial intelligence-based mineral prediction method of the present invention, the mineral prediction machine learning model includes an evidence raster coding layer, a scene graph structure coding layer, a feature modulation gating layer and a cross-modal spatial attention fusion layer.

[0030] The evidence raster coding layer performs multi-layer convolutional feature extraction on the multi-source mineralization evidence dataset and outputs evidence raster feature representation;

[0031] The scene graph structure encoding layer performs topological connectivity aggregation on the object-level spatial intelligent scene graph and outputs structural association features;

[0032] The feature modulation gating layer uses the channel weight layer as a constraint condition to perform feature weight gating and direction consistency modulation on the evidence grid feature representation and structural correlation features, and outputs channel constraint features;

[0033] The cross-modal spatial attention fusion layer performs spatial location-level attention allocation on channel-constrained features and outputs a mineral prediction numerical raster.

[0034] As a preferred embodiment of the spatial intelligence-based mineral prediction method of the present invention, the specific steps for generating the mineral prediction layer are as follows:

[0035] Spatial alignment and multimodal spatial correlation feature fusion are performed on the multi-source mineralization evidence dataset, object-level spatial intelligent scene map and channel weight layer to generate fused feature tensor;

[0036] The fused feature tensor is organized into inference batches and input into the mineral prediction machine learning model to perform mineral inference and obtain a mineral prediction numerical raster.

[0037] Spatial continuity correction and mineralization probability determination threshold segmentation are performed on the mineral prediction numerical raster to generate a mineral prediction layer.

[0038] The beneficial effects of this invention are as follows: by constructing an object-level spatial intelligent scene graph, the topological connectivity and semantic dynamic interaction between spatial entities are captured, and a high-order spatial intelligent scene is constructed to reflect the causal chain of mineralization, thereby improving the interpretability of predictions for complex tectonic areas; by constructing an anisotropic cost field to calculate the minimum cost channel distance, dynamic modeling of mineralization paths is realized, thereby improving prediction accuracy and engineering practicality. Attached Figure Description

[0039] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a flowchart of a spatial intelligence-based mineral prediction method.

[0041] Figure 2 A flowchart for collecting multi-source mineralization evidence datasets.

[0042] Figure 3 Scatter plot showing the relationship between the proportion of proxy for blocking and suppressing constraints and prediction accuracy (PR_AUC).

[0043] Figure 4 A heat map of the spatial distribution of numerical raster data for mineral resource prediction. Detailed Implementation

[0044] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0045] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0046] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0047] Reference Figures 1-4 This is one embodiment of the present invention, which provides a spatial intelligence-based mineral prediction method, including the following steps:

[0048] S1: Collect multi-source mineralization evidence datasets, extract candidate spatial entity features, perform centerline segmentation coding, partitioning and boundary fitting, calculate spatial entity feature attributes, and generate a list of spatial entity features;

[0049] S1.1: Collect multi-source mineralization evidence data, perform consistency verification and coordinate system one, and generate a coordinate system one mineralization evidence package;

[0050] Specifically, when collecting multi-source metallogenic evidence data (such as geophysical exploration data, mineral deposit and mineralization information data, etc.), the multi-source metallogenic evidence data from each data source is read and registered according to the data source identifier in order, and the field structure and coordinate reference identifier are registered. For missing fields, mismatched field types, duplicate records and coordinate reference identifier conflicts, consistency checks are performed and entries that do not meet the consistency check conditions are removed. For multi-source metallogenic evidence data that have passed the consistency check, the coordinate reference identifier is read, and coordinate transformation, axial direction correction and coordinate unit conversion (e.g., meters and kilometers) are performed according to the unified target coordinate reference. The boundary range of the converted data is verified and packaged into a coordinate system-metallogenic evidence package.

[0051] S1.2: Perform spatial resolution resampling transformation and raster pixel scale unification processing on the coordinate system mineralization evidence package to generate a resampling unified package;

[0052] Specifically, the metallogenic evidence package of the coordinate system is expanded line by line. The spatial resolution marker and raster cell scale marker corresponding to each multi-source metallogenic evidence data are read and registered in the transformation record. The unified target spatial resolution and unified target raster cell scale (e.g., 30 meters) are determined. For multi-source metallogenic evidence data with a spatial resolution higher than the unified target spatial resolution, the cell values ​​are merged according to the window range and backfilled to the unified raster position. For multi-source metallogenic evidence data with a spatial resolution lower than the unified target spatial resolution, the values ​​are filled in according to the unified raster position and the filling mark is written. For multi-source metallogenic evidence data whose raster cell scale does not meet the unified target raster cell scale, the horizontal scale conversion and vertical scale conversion are performed to complete the consistency of cell width and cell height and update the raster boundary range. For multi-source metallogenic evidence data that have completed spatial resolution resampling transformation and raster cell scale unification processing, the range coverage check and gap mark check are performed, and the transformation record and the processed multi-source metallogenic evidence data are packaged into a resampling unified package.

[0053] It should be noted that the window range is calculated by converting the unified target spatial resolution with the spatial resolution of the multi-source mineralization evidence data. It is used to limit the neighborhood range of raster cells that need to be merged or split during spatial resolution resampling conversion (for example, the window side length is calculated by converting the side length corresponding to the unified target spatial resolution).

[0054] S1.3: Perform grid alignment and grid starting point locking on the resampled unified package, obtain the grid aligned volume, and perform missing value processing and outlier removal to generate a multi-source mineralization evidence dataset;

[0055] Specifically, the resampling unified package is expanded line by line, and the unified target spatial resolution and unified target raster cell scale are read. Within the coverage area of ​​the resampling unified package, the coordinates of the grid starting point are selected. The grid starting point is locked and written as a fixed reference and remains unchanged. Grid alignment is based on the grid starting point coordinates and the unified target raster cell scale. The center coordinates of the raster cells in the resampling unified package are converted into row and column indices and mapped to the unified grid position. Boundary clipping is performed on raster cells that exceed the unified grid boundary. Missing marks are written to the gap positions within the unified grid boundary. The grid alignment outputs a grid alignment body. Missing value processing locates the gap position according to the missing marks in the grid alignment body. Filling is performed according to the valid values ​​of adjacent raster cells and the fill mark is registered. Outlier removal performs valid interval verification and mutation amplitude verification on the raster cell values ​​in the grid alignment body. Raster cell values ​​that do not meet the verification conditions are written as anomaly marks and set as missing marks. Missing value processing is performed again at the anomaly mark positions. The result is encapsulated into a multi-source mineralization evidence dataset.

[0056] S1.4: Extract candidate linear features, candidate boundary features, and candidate anomaly features from the multi-source metallogenic evidence dataset, and perform cross-modal spatial response alignment and confidence propagation filtering to generate a candidate spatial entity feature layer;

[0057] Furthermore, the multi-source mineralization evidence dataset is expanded grid-wise according to grid position, and the mineralization evidence values ​​corresponding to the grids of each data source are read and arranged side by side at the same grid position. For the high-response strips of the side-arranged mineralization evidence values ​​that extend continuously in the direction of adjacent grid positions, connectivity aggregation and centerline extraction are performed to form candidate linear features. For the side-arranged mineralization evidence values ​​in the high-low response transition zone, boundary tracking and curve solidification are performed to form candidate boundary features. For the side-arranged mineralization evidence values ​​in isolated areas with significant differences from the neighborhood, anomaly markers are written and aggregated to form candidate anomaly features. Cross-modal spatial response alignment is performed to overlay and verify the candidate linear features, candidate boundary features and candidate anomaly features according to the grid position correspondence and the offset distance is recorded. Confidence propagation screening starts from the overlapping segment of multiple data sources and gradually propagates and decays along the adjacent grid positions. Low confidence segments are eliminated by confidence threshold gating and high confidence segments are retained to generate a candidate spatial entity feature layer.

[0058] It should be noted that the confidence threshold is taken from the 90th percentile of the initial confidence distribution of the overlapping segment of multiple data sources, and the threshold is fixed on the validation samples to maximize the area under the precision-recall curve (example range: 0.6 to 0.9).

[0059] S1.5: Refine and repair the candidate linear features in the candidate spatial entity feature layer, obtain the connected skeleton, and perform segmentation and centerline segmentation encoding to generate a segmented encoding set;

[0060] Specifically, each candidate linear feature in the candidate spatial entity feature layer is expanded, and the strip region corresponding to the candidate linear feature is shrunk inward along the boundary while simultaneously removing burr segments and isolated short segments, so that the candidate linear features converge into a central line band at the grid position. The positions of the connected endpoints in the central line band are located and the endpoint directions are recorded. The consistency of the endpoint spacing and direction is checked, and the connection segments are written and merged into the central line band by connecting them point by point at the grid position. Connectivity is checked again on the central line band and adjacent connected segments are merged to obtain the connected skeleton. The connected skeleton is segmented according to the intersection node position, the direction change position, and the lower limit of the length constraint, and the start and end grid positions and intra-segment order of each central line segment are recorded. Each central line segment is combined according to the connected skeleton number and intra-segment order to generate a central line segment code, and the central line segment code, start and end grid positions, segment length, and source mark are summarized into a segment code set.

[0061] It should be noted that the grid position point-by-point connection method refers to gradually placing points between endpoint pairs according to adjacent grid positions and writing the connection segment, so that the gap between endpoints is continuously spliced ​​and closed by adjacent grid positions (e.g., four-neighbor step or eight-neighbor step).

[0062] S1.6: Perform partitioning and boundary fitting on candidate boundary features and candidate anomaly features to obtain fitted boundary curves, and combine them with segmented encoding sets for spatial adsorption and position correction to generate a boundary fitting set;

[0063] Specifically, candidate boundary elements and candidate anomaly elements in the candidate spatial entity element layer are expanded one by one. Candidate boundary elements are divided into partition boundary segments according to their connectivity and adjacency relationships and partition numbers are registered. Candidate anomaly elements are divided into partition anomaly regions according to the connectivity range of anomaly markers and partition numbers are registered. Within the partition boundary segments and partition anomaly regions, the boundary position is tracked grid by grid along the outer contour and the boundary direction is registered. After smoothing the folded and jagged segments in the boundary direction, boundary fitting is performed to map the boundary direction to the fitted boundary curve. Combined with the segmented coding set, the corresponding position of the centerline segment code is located according to the segmented coding set. The offset distance of the fitted boundary curve relative to the corresponding position of the centerline segment code is obtained, and the centerline segment code with the smallest offset distance is selected as the adsorption target. Spatial adsorption and position correction are performed on the fitted boundary curve segments whose offset distance does not exceed the adsorption threshold. When the fitted boundary curve crosses multiple centerline segment codes, it is cut according to the intersection position and spatial adsorption and position correction are performed separately. The partition number, fitted boundary curve and centerline segment code are summarized and encapsulated to generate a boundary fitting set.

[0064] It should be noted that the anomaly marker connectivity range refers to the aggregate range within which all grid locations marked with anomalies in the candidate anomaly features can reach each other in an adjacency manner.

[0065] The adsorption threshold is defined based on the distribution of offset distances at the corresponding positions of the fitted boundary curve relative to the center line segmented encoding and the maximum allowable offset distance for position correction (example range: 1 to 3 grid widths).

[0066] S1.7: Based on the boundary fitting set and segmented coding set, calculate the spatial entity feature attributes, write the source and quality tags, and generate a list of spatial entity features;

[0067] Specifically, the fitted boundary curves in the boundary fitting set are merged into corresponding segmented coding sets according to partition numbers and centerline segment codes, and spatial entity element entries are established and written with spatial entity element numbers. For each spatial entity element entry, the geometric contour is unfolded according to the outer contour of the fitted boundary curve and the segmented coding direction of the centerline. Based on the unified target raster pixel scale, the boundary span, centerline segment length, number of turning segments, and proportion of abnormal area coverage are statistically analyzed. A turning segment penalty coefficient is used to penalize the number of turning segments (for example, when the number of turning segments exceeds five, the penalty coefficient will reduce the total score of the spatial entity), and an abnormal coverage enhancement coefficient is used to enhance the proportion of abnormal area coverage (for example, when the abnormal area coverage exceeds 30%, the enhancement coefficient will increase). The proportion of spatial entities is calculated, and the statistical values ​​are converted into spatial entity element attributes under a unified spatial scale (reflecting the geometric characteristics and quality indicators of spatial entities in a numerical way, such as a boundary span of 500 meters, a center line segment length of 200 meters, a number of turning segments of 3, and an abnormal area coverage ratio of 25%). The geometric shape and quality level of different spatial entities can be quantified and compared through attribute values. The source markers recorded in the boundary fitting set and the source markers recorded in the segmented coding set are merged into a source marker. The confidence segment retention information, the number of breakpoint repair occurrences, and the integrity of spatial adsorption and position correction trajectories in the candidate spatial entity element layer are summarized into quality markers. After the fields are written, all spatial entity element entries are summarized to generate a spatial entity element list.

[0068] The formula for calculating the attributes of spatial entity features is:

[0069] ;

[0070] in, Represents the attributes of spatial entity elements. This represents a mapping compressed to 0 to 1. Indicates the length of the center line segment. Indicates a uniform target raster cell scale. Indicates the boundary span. Indicates the transition penalty coefficient. Indicates the number of transition segments. Indicates the anomalous coverage enhancement coefficient. This indicates the percentage of abnormal coverage areas.

[0071] It should be noted that the turning point penalty coefficient is defined based on the density of the number of turning points relative to the length of the centerline segment coding direction, and is used to apply a stronger penalty to spatial entity element entries with a high number of turning points in the spatial entity element attributes (example range: 0.3 to 2.0); the anomaly coverage enhancement coefficient is defined based on the proportion of anomaly area coverage within the coverage range of the fitted boundary curve, and is used to apply a stronger enhancement to spatial entity element entries with a higher proportion of anomaly area coverage in the spatial entity element attributes (example range: 0.2 to 3.0).

[0072] S2: Identify the spatial relationships of the list of spatial entity elements, construct a topological connectivity graph, extract connectivity structure attributes, perform semantic relationship reasoning, and generate an object-level spatial intelligent scene graph;

[0073] S2.1: Based on the list of spatial entity elements, construct the candidate neighborhood retrieval range under a unified coordinate system, perform spatial relationship identification between objects, and generate a spatial relationship record set;

[0074] Specifically, the spatial entity element number and attributes are read one by one. Under a unified coordinate system, the fitted boundary curve and centerline corresponding to each spatial entity element are segmented and coded into coordinate ranges. The coordinate ranges are then expanded and the expanded boundary ranges are fixed to obtain the candidate neighborhood retrieval range for each spatial entity element. The list of spatial entity elements is then paired according to the candidate neighborhood retrieval range. Neighborhood pairing is based on whether the candidate neighborhood retrieval ranges overlap or the distance between the candidate neighborhood retrieval ranges does not exceed the upper limit of the neighborhood distance (the upper limit of the neighborhood distance is taken in the distance unit of the unified target coordinate reference, and after the coordinate unit conversion is completed, it is written into the neighborhood distance upper limit field in the unified distance unit, example range: 1 km to 5 km). Within the candidate neighborhood retrieval range, combinations of spatial entity element numbers that overlap, intersect, are adjacent, and have the same direction are retrieved and written into the spatial relationship type tag and relationship strength tag to generate a spatial relationship record set.

[0075] It should be noted that the spatial relationship type tag is used to specify the spatial relationship category (e.g., spatial intersection, spatial adjacency) presented by the combination of spatial entity element numbers within the candidate neighborhood retrieval range.

[0076] The relation strength label is written as high strength when there is a large area of ​​spatial overlap or long distance spatial intersection within the candidate neighborhood retrieval range, the minimum spacing is close to zero, and the center line segment coding direction is consistent. It is written as medium strength when there is only local spatial overlap or short distance spatial intersection within the candidate neighborhood retrieval range, or the minimum spacing is in the middle of the upper limit of the neighborhood distance and the center line segment coding direction is partially consistent. It is written as low strength when there is only close boundary within the candidate neighborhood retrieval range, the minimum spacing is close to the upper limit of the neighborhood distance, and the center line segment coding direction is inconsistent.

[0077] S2.2: Register the list of spatial entity elements and the set of spatial relationship records as topological nodes and topological edges respectively, and perform spatial relationship merging and priority sorting between duplicate objects to generate a topological connectivity graph;

[0078] Furthermore, the list of spatial entity elements is expanded item by item, and the spatial entity element number, fitted boundary curve, centerline segment code, and spatial entity element attributes are read. The spatial entity element number is used as the topology node identifier and registered in the topology node record. At the same time, the fitted boundary curve range and the centerline segment code position are registered as the topology node spatial range. The spatial relationship record set is expanded item by item, and the combination of spatial entity element numbers, spatial relationship type marker, and relationship strength marker are read. The combination of spatial entity element numbers is registered as the topology edge connection endpoint, and the spatial relationship type marker and relationship strength marker are registered as topology edge attributes. When performing duplicate spatial relationship merging, the following is used: Duplicate records are located by combining spatial entity element numbers and spatial relationship type tags. Duplicate records are merged into a single topological edge record, retaining the topological edge record with higher relationship strength tags. The correspondence between records before and after merging is also recorded. Topological edge records are arranged according to the priority of spatial relationship type tags and the order of relationship strength tags, and a sorting number is written (e.g., spatial intersection takes precedence over spatial adjacency). Topological node records and topological edge records are encapsulated into a topological connectivity graph. The topological connectivity graph solidifies the spatial relationships between spatial entity elements into a sortable topological edge connection structure, enabling subsequent connectivity aggregation to pass and summarize structural information between connected nodes according to priority.

[0079] S2.3: Extract connectivity structure attributes from the topological connectivity graph, and perform topological connectivity confidence propagation and semantic constraint adjudication to generate a connectivity structure attribute table;

[0080] Furthermore, topological node records and topological edge records are read one by one from the topological connectivity graph. Connections are established between the topological node records according to the spatial entity element number combinations of the topological edge records. The number of connected topological edges, the index of adjacent topological nodes, and the distribution of spatial relationship type markers are recorded in the topological node records. Connectivity grouping numbers are written to reachable topological node records according to the connection comparisons, and the connectivity path level and sorting number are recorded for each topological edge record. Topological connectivity confidence propagation uses the relationship strength marker of the topological edge record as the starting confidence level, propagating it hop-by-hop between adjacent topological node records while performing attenuation and accumulation. The confidence marker after each propagation is written back to the topological node record. Semantic constraint adjudication performs mutual exclusion and duplicate checks on different spatial relationship type markers appearing in the same spatial entity element number combination. Topological edge records with higher relationship strength markers and earlier sorting numbers are retained. Adjudication markers are written to the unretained topological edge records, and the adjudication reasons are recorded. A connectivity structure attribute table is then generated.

[0081] It should be noted that the connectivity structure attribute represents the connectivity group number, the number of connected topological edges, the index of adjacent topological nodes, the spatial relationship type label distribution, the connectivity path level, and the confidence label information of propagation backwrite for each topological node and topological edge in the topological connectivity graph.

[0082] S2.4: Combine the list of spatial entity elements, the set of spatial relationship records, the topological connectivity graph and the connectivity structure attribute table, and perform semantic relationship reasoning and structured assembly to generate an object-level spatial intelligent scene graph;

[0083] Specifically, the list of spatial entity elements, the set of spatial relationship records, the topological connectivity graph, and the connectivity structure attribute table are read one by one. A comparison is established according to the spatial entity element number and the combination of spatial entity element numbers. The coordinate range of the fitted boundary curve, the segmented code of the center line, and the spatial entity element attributes corresponding to the spatial entity element number are written into the object node. The spatial relationship type label, relationship strength label, sorting number, connectivity group number, connectivity path level, and topological node confidence label corresponding to the spatial entity element number combination with an empty adjudication mark in the topological edge record are written into the relationship edge. Semantic relationship reasoning performs consistency checks on the spatial relationship type label and connectivity group number on the relationship edge, and performs a strong-weak consistency check on the relationship strength label and topological node confidence label. The relationship edge that passes the check is written into the reasoning relationship label. The structured assembly merges the object node, relationship edge, and connectivity path level according to the connectivity group number and solidifies the reference index to generate an object-level spatial intelligent scene graph.

[0084] S3: Based on the object-level spatial intelligent scene graph, construct an anisotropic cost field, calculate the minimum cost channel distance, map it to the channel weight expression, and generate a channel weight layer;

[0085] S3.1: Based on the object-level spatial intelligent scene map, filter spatial entity elements and perform rasterization of the channel centerline landing point snapping to generate the starting point raster set;

[0086] Specifically, each object node and relational edge is read sequentially. The inference relation markers in the relational edges are used as the filtering criteria. Relationships marked as valid are retained and their corresponding spatial entity element number combinations are registered. The spatial entity element number combinations are expanded to object nodes, and spatial entity element numbers that simultaneously meet the following criteria are filtered: the relation strength marker is high or medium, the topological node confidence marker is not low, and the connectivity path level is within the backbone level of the connectivity group number. The centerline segments of the selected spatial entity element numbers are then coded and summarized into channel centerline candidates. The channel centerline rasterization landing point snapping method concatenates the channel centerline candidates into continuous centerlines according to the connectivity path level and sorting sequence number. Based on the unified target raster cell scale, the continuous centerlines are mapped point by point to the grid position and written into the centerline landing point. When there is an offset between the centerline landing point and the grid center position, the centerline landing point is snapped to the nearest grid center position while maintaining the connectivity of adjacent centerline landing points. The endpoint landing points and intersection landing points are extracted from the centerline landing points, and duplicate landing points are merged. The results are then summarized to generate the starting point raster set.

[0087] S3.2: Combine the starting grid set with semantic relations, perform weighted correction of the barrier boundary direction, construct the initial anisotropic cost field, and perform direction-sensitive cost recalibration to generate the anisotropic cost field.

[0088] Specifically, based on the starting grid set, the grid position corresponding to each starting grid is used as the starting point for cost expansion. Based on the object-level spatial intelligent scene graph, the coordinate range of the fitting boundary curve pointing to the barrier boundary is expanded into the barrier boundary direction, and the centerline segmented encoding direction is expanded into the channel direction. The cost corresponding to the crossing direction of the barrier boundary direction is written as the rising state, the cost corresponding to the channel direction is written as the falling state, and the cost corresponding to the other directions is written as the intermediate state (e.g., the cost is 0.5 to 2.0). The initial anisotropic cost field construction writes the grid positions within the coverage area of ​​the starting grid set into the direction-related cost in the order of expansion of adjacent grid positions, and superimposes the barrier boundary direction weighted correction and the channel direction weighted correction during each expansion. The direction-related cost of adjacent grid positions is checked for bidirectional consistency and smoothed locally. The barrier boundary crossing direction is kept under rising constraints, and the channel direction is kept under falling constraints. The recalibrated direction-related cost is solidified to generate the anisotropic cost field.

[0089] S3.3: Based on the anisotropic cost field, calculate the minimum cost channel distance, update the intersection region and impose blocking and suppression constraints to generate a channel distance map;

[0090] Specifically, the starting grid set is used as the starting point for expansion. Within the coverage area of ​​the anisotropic cost field, progressive expansion of adjacent grid positions is performed on the grid locations. The progressive path of adjacent grid positions is used to connect the starting grid set to the target grid location. The progressive expansion moves continuously along the adjacent grid locations to form candidate ordered paths, and the step number and total number of steps are recorded for each adjacent movement in the candidate ordered paths. Each adjacent movement obtains the cost corresponding to the direction in the anisotropic cost field as the cost increment according to the movement direction. The progressive expansion continuously accumulates the cost increments according to the step number to form the cumulative cost corresponding to the target grid location, and the minimum cost channel distance is used. This represents the candidate value with the lowest cumulative cost within the progressive path range of adjacent grid positions, while retaining the movement direction marker for backtracking; the intersection region update forms an intersection region marker at the grid positions where multiple sources converge during the backtracking process of the movement direction marker, performs merging on the cumulative cost within the coverage area of ​​the intersection region marker, and performs continuity smoothing on the boundary of the intersection region marker; the barrier suppression constraint sets cross-suppression for the movement direction in the direction of the barrier boundary weighted correction that is in an elevated state, performs replacement on the movement direction marker corresponding to the cross-suppression direction and triggers re-expansion, and summarizes the minimum cost channel distance corresponding to each grid position to generate a channel distance map.

[0091] The formula for calculating the minimum cost channel distance is:

[0092] ;

[0093] in, This represents the minimum cost path distance. Represents the starting raster set, Represents the distance from the starting raster set to the grid position. The progressive path of adjacent grid positions, This represents an ordered path. Indicates from Move to adjacent grid position The cost increment, This represents the total number of steps taken along an ordered path within a grid location. Indicates the step number in the ordered path of the grid position. Indicates the grid position.

[0094] Figure 3The scatter plot of the relationship between the proportion of surrogate representation for the barrier inhibition constraint and the prediction accuracy (PR_AUC) is used to verify the impact of barrier inhibition constraints on prediction stability and discriminative ability from a statistical distribution perspective. During the solidification process of the mineral prediction machine learning model, the training and validation sets are batch-wise subjected to forward inference and their convergence is continuously monitored. The validation set is used to select the optimal weights. In the comparative experiment, the proportion of high-probability barrier areas is recorded for each run as a surrogate representation for the barrier inhibition constraint, and the corresponding prediction accuracy (PR_AUC) is recorded simultaneously (representing a comprehensive measure of the precision-recall curve formed using a pixel-level binary classification approach, used to characterize the positive class determination effect of mineralization). The scatter plots correspond to control group A, control group B, and the experimental group, respectively. A lower proportion of surrogate representation for the barrier inhibition constraint indicates a more effective suppression of the barrier boundary direction crossing, and the path modeling of the channel distance map is more consistent with the "barrier-channel" directional constraint. The degree of separation of the scatter plots in the vertical direction (PR_AUC) and horizontal direction (proportion) can intuitively reflect the difference in separability and robustness between the experimental group and the control group, thus supporting the improvement of engineering practicality with recordable objective indicators.

[0095] S3.4: Map the minimum cost channel distance in the channel distance graph to a channel weight representation, and perform weight enhancement, weight suppression, upper limit pruning and spatial smoothing to generate a channel weight layer;

[0096] Furthermore, the minimum cost channel distance at each grid location is converted into a channel weight expression according to the distance-to-weight mapping relationship (referring to establishing a monotonically decreasing normalized mapping based on the minimum cost channel distance and the upper limit of channel distance, and pruning the mapping result to a weight range of zero to one). The channel weight expression decreases as the minimum cost channel distance increases and increases as the minimum cost channel distance decreases (for example, the channel weight expression takes the example range of 0 to 1). Weight enhancement adjusts the increase of the channel weight expression within the coverage area of ​​the intersection region and the range adjacent to the channel direction and registers the enhancement mark. Weight suppression adjusts the decrease of the channel weight expression within the range where the blocking suppression constraint effect and the minimum cost channel distance in the channel distance map are close to the upper limit of channel distance and registers the suppression mark. Upper limit pruning truncates the grid locations where the channel weight expression exceeds the upper limit of channel weight and registers the pruning mark. Spatial smoothing performs continuity constraint smoothing on the channel weight expressions of adjacent grid locations and retains the boundary transition zone, and summarizes to generate a channel weight layer.

[0097] S4: Spatial alignment and multimodal spatial correlation feature fusion are performed on the multi-source mineralization evidence dataset, object-level spatial intelligent scene map and channel weight layer, and mineral inference is performed through the mineral prediction machine learning model to generate a mineral prediction layer;

[0098] S4.1: The mineral prediction machine learning model includes an evidence raster coding layer, a scene graph structure coding layer, a feature modulation gating layer, and a cross-modal spatial attention fusion layer;

[0099] Specifically, the mineral prediction machine learning model uses a fused feature tensor as a unified input and concatenates an evidence raster encoding layer, a scene graph structure encoding layer, a feature modulation gating layer, and a cross-modal space attention fusion layer in a fixed hierarchical order. The evidence raster encoding layer outputs evidence raster feature representations from the fused feature tensor and uses them as input to the feature modulation gating layer. The scene graph structure encoding layer outputs structural correlation features from the fused feature tensor and uses them as input to the feature modulation gating layer. The feature modulation gating layer receives the evidence raster feature representations, structural correlation features, and channel weight layer constraints from the fused feature tensor and outputs channel constraint features as input to the cross-modal space attention fusion layer. The cross-modal space attention fusion layer receives the channel constraint features and outputs a mineral prediction numerical raster.

[0100] It should be noted that the mineral prediction machine learning model is trained using a fused feature tensor as input. Supervision labels are formed using a grid of known mineralized points and a grid of control non-mineralized areas. Forward inference is performed in batches based on the training and validation sets. A binary cross-entropy loss is obtained based on the difference between the predicted and supervision labels, and this loss is weighted and summed to form the training loss. Parameters (batch size, number of iterations, and loss weights, etc.) are then updated in reverse. The validation set is continuously monitored for convergence, and the optimal weights are saved. The learning rate decreases according to the convergence status of the validation set, and training terminates when the validation set loss no longer decreases, thus completing the solidification of the mineral prediction machine learning model.

[0101] The training set refers to the set of samples extracted from the fused feature tensor and bound with supervisory labels for parameter updates. The validation set refers to the set of samples extracted from the fused feature tensor and bound with supervisory labels but not involved in parameter updates, used to monitor convergence and select the optimal weights (e.g., reserving 10% to 30% as the validation set according to spatial partitioning).

[0102] S4.2: The evidence raster coding layer performs multi-layer convolutional feature extraction on the multi-source mineralization evidence dataset and outputs evidence raster feature representation;

[0103] Specifically, the evidence raster coding layer unfolds the multi-source mineralization evidence dataset into a multi-channel raster arrangement according to grid positions, and extracts a fixed neighborhood range of raster blocks for each grid position as convolution input under a unified target raster cell scale. During the multi-layer convolutional feature extraction process, the evidence raster coding layer performs boundary padding on the raster blocks to maintain spatial alignment, performs convolution kernel sliding operation on the multi-channel raster arrangement within the neighborhood to generate the first layer feature map, performs nonlinear mapping and scale shaping on the first layer feature map, and continues to input the next layer convolution kernel sliding operation to generate a higher layer feature map. The evidence raster coding layer maintains the grid position correspondence in each layer and performs channel aggregation and spatial consistency on the multi-layer feature maps, so that each grid position obtains a corresponding high-dimensional feature vector. The evidence raster coding layer backfills the high-dimensional feature vector according to the grid position and encapsulates it to output the evidence raster feature representation.

[0104] S4.3: The scene graph structure encoding layer performs topological connectivity aggregation on the object-level spatial intelligent scene graph and outputs structural association features;

[0105] Specifically, the scene graph structure encoding layer receives the object-level spatial intelligent scene graph from the fusion feature tensor. It expands object nodes according to spatial entity element numbers and extracts the connected group number, connected path level, and topological node confidence label to form object node states. It expands relational edges according to spatial entity element numbers and extracts spatial relation type label, relation strength label, and inference relation label to form edge constraints. Within the connected group number range, it aggregates adjacent object node states along the relational edge connection endpoints. It weights the aggregated information according to the relation strength label and topological node confidence label and performs hierarchical consistency according to the connected path level. It then backfills the updated object node states. Finally, it assigns the object node states to corresponding grid positions according to the mapping relationship from the object-level spatial intelligent scene graph to the grid position and encapsulates and outputs structural association features.

[0106] S4.4: The feature modulation gating layer uses the channel weight layer as a constraint to perform feature weight gating and orientation consistency modulation on the evidence grid feature representation and structural correlation features, and outputs channel constraint features;

[0107] Specifically, in the feature modulation gating layer, the evidence raster feature representation and structural association feature are aligned according to the grid position. The channel weight expression of the channel weight layer at the same grid position is normalized to the gating coefficient range (e.g., 0 to 1). The gating coefficient is scaled channel by channel for the evidence raster feature representation and channel by channel for the structural association feature, while the missing marker position retains a placeholder value and does not participate in the scaling. The direction consistency modulation determines the channel direction by the direction of change of the channel weight expression of the channel weight layer at adjacent grid positions. The neighboring response of the evidence raster feature representation and the structural association feature along the channel direction is enhanced, and the neighboring response deviating from the channel direction is suppressed. The enhancement and suppression amplitudes are subject to upper limit constraints. The gating and modulated evidence raster feature representation and the gating and modulated structural association feature are fused in a fixed splicing order to output the channel constraint feature.

[0108] S4.5: The cross-modal spatial attention fusion layer performs spatial location-level attention allocation on channel-constrained features and outputs a numerical raster for mineral prediction.

[0109] Specifically, the cross-modal spatial attention fusion layer receives channel constraint features and expands them according to grid positions. For each grid position, three sets of linear projections are performed on the channel constraint features to obtain an attention query vector, an attention key vector, and an attention value vector. For each grid position, adjacent grid positions are enumerated within the attention neighborhood. The similarity between the attention query vector and the attention key vectors of adjacent grid positions is calculated and normalized to obtain spatial position-level attention weights. Spatial position-level attention weights for missing marker positions are masked and placeholder values ​​are maintained. The spatial position-level attention weights are weighted and converged on the attention value vector to obtain a fusion vector. Channel shaping and residual merging are performed on the fusion vector. The cross-modal spatial attention fusion layer backfills the fusion vector according to grid positions and outputs a mineral prediction numerical raster.

[0110] S4.6: Spatial alignment and multimodal spatial correlation feature fusion are performed on the multi-source mineralization evidence dataset, object-level spatial intelligent scene map and channel weight layer to generate fused feature tensor;

[0111] Specifically, the multi-source mineralization evidence dataset, object-level spatial intelligent scene map, and channel weight layer are established in a unified coordinate system according to the grid starting point locking relationship and the unified target raster cell scale to establish a grid position correspondence, so that the channel weight layer and the multi-source mineralization evidence dataset correspond one-to-one at the same grid position; the object-level spatial intelligent scene map is segmented and encoded and mapped to the covering grid position according to the fitted boundary curve coordinate range and centerline; at the covering grid position, the spatial entity element number, connected group number, connected path level, topological node confidence label, spatial relationship type label, relationship strength label, and inference relationship label are registered; at each grid position, the mineralization evidence values ​​of the multi-source mineralization evidence dataset, the channel weight expression of the channel weight layer, and the object node labels and relationship edge labels of the object-level spatial intelligent scene map are summarized in a fixed splicing order; for missing label positions, the missing labels are retained and placeholder values ​​are registered, and the resulting data are encapsulated to generate a fusion feature tensor.

[0112] S4.7: Organize the fused feature tensor into inference batches and input them into the mineral prediction machine learning model to perform mineral inference and obtain a mineral prediction numerical raster.

[0113] Specifically, the fused feature tensor is divided into inference sample blocks according to grid positions. The inference sample blocks cover continuous grid positions with a fixed spatial window and retain the corresponding missing markers and placeholder values. The inference sample blocks are grouped according to the inference batch size and the inference batch number is registered (e.g., the inference batch size is 32 to 256). The inference sample blocks corresponding to each inference batch number are encapsulated according to the input order of the mineral prediction machine learning model and input into the mineral prediction machine learning model. The mineral prediction machine learning model performs forward inference on the inference sample blocks and outputs the mineralization probability value of each inference sample block. The mineralization probability value is backfilled to the corresponding grid position according to the spatial window position of the inference sample block in the fused feature tensor. For overlapping window positions, weighted fusion is used (the weight is determined by the relative distance of each grid position in the overlapping window to the center of the window, with the center position having a larger weight and the edge position having a smaller weight, and multiple outputs of overlapping positions are merged into a single output according to the weight) and the placeholder values ​​of missing marker positions are kept out of the fusion. The mineral prediction value raster is obtained by summarizing.

[0114] S4.8: Perform spatial continuity correction and mineralization probability determination threshold segmentation on the mineral prediction numerical raster to generate a mineral prediction layer;

[0115] Furthermore, the mineral prediction numerical raster is expanded according to grid position, and missing marker positions are retained and not involved in continuity processing. Neighborhood consistency correction is performed on the mineralization probability values ​​of adjacent grid positions, reverting isolated high-value points to the neighborhood principal values ​​and raising isolated low-value points to the neighborhood principal values, while maintaining smooth changes in the continuous gradient direction to avoid broken stripes. At the same time, local smoothing is performed on the boundary transition zone to reduce jagged boundaries. The mineralization probability judgment threshold segmentation compares the mineralization probability values ​​that have completed spatial continuity correction with the mineralization probability judgment threshold for each grid position. Grid positions that meet the mineralization probability judgment threshold are written as mineralized area markers, and grid positions that do not meet the mineralization probability judgment threshold are written as non-mineralized area markers. Connectivity aggregation is performed on the mineralized area markers and adjacent fragments are merged. Isolated connected regions with too small areas are removed and the boundaries of connected regions are smoothed to generate a mineral prediction layer.

[0116] It should be noted that the mineralization probability determination threshold is defined based on the distribution of mineralization probability values ​​in the mineral prediction numerical raster and the coverage range of mineralization probability values ​​of known mineralization point grids (example range: 0.6 to 0.9).

[0117] Figure 4 The spatial distribution heatmap of the mineral prediction numerical raster is used to display the inference output distribution of the mineral prediction machine learning model at a unified grid location. The horizontal axis represents the unified grid location as a column index, the vertical axis as a row index, and the color bars represent the mineralization probability intensity. After fusing feature tensors and inputting them into the mineral prediction machine learning model in batches, a mineral prediction numerical raster is obtained. Spatial continuity correction and mineralization probability threshold segmentation are then performed on the raster to generate a mineral prediction layer. The heatmap visually reflects the strengthening effect of spatial continuity correction on the coherence of high-probability areas. These high-probability areas exhibit a continuous distribution in bands or clusters, facilitating interpretation in relation to the causal chain of mineralization expressed by the object-level spatial intelligent scene map. This supports the interpretability and engineering practicality of predictions for complex tectonic zones.

[0118] It should be noted that spatial continuity correction uses a fixed neighborhood window to check the neighborhood consistency of each grid location. High-value points that differ greatly from the principal value of the neighborhood and have no similar high-value support in the neighborhood are reverted to the principal value of the neighborhood. Low-value points that differ greatly from the principal value of the neighborhood and have no similar low-value support in the neighborhood are raised to the principal value of the neighborhood. Neighborhood smoothing is performed on step changes at the boundary.

[0119] In summary, this invention achieves the capture of topological connectivity and semantic dynamic interaction between spatial entities by constructing an object-level spatial intelligent scene graph, constructing a high-order spatial intelligent scene to reflect the causal chain of mineralization, and improving the interpretability of predictions for complex tectonic regions; and realizes dynamic modeling of mineralization paths by constructing an anisotropic cost field to calculate the minimum cost channel distance, thereby improving prediction accuracy and engineering practicality.

[0120] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A mineral resource prediction method based on spatial intelligence, characterized in that: include, Collect multi-source mineralization evidence datasets, extract candidate spatial entity features, perform centerline segmentation coding, partitioning and boundary fitting, calculate spatial entity feature attributes, and generate a list of spatial entity features; Identify the spatial relationships of the list of spatial entity elements, construct a topological connectivity graph, extract connectivity structure attributes, perform semantic relationship reasoning, and generate an object-level spatial intelligent scene graph; Based on the object-level spatial intelligent scene graph, an anisotropic cost field is constructed, and the minimum cost channel distance is calculated, mapped to the channel weight expression, and a channel weight layer is generated. The multi-source mineralization evidence dataset, object-level spatial intelligent scene graph, and channel weight layer are spatially aligned and multimodal spatial correlation features are fused. Mineral inference is then performed through a mineral prediction machine learning model, which includes an evidence raster coding layer, a scene graph structure coding layer, a feature modulation gating layer, and a cross-modal spatial attention fusion layer. The evidence raster coding layer performs multi-layer convolutional feature extraction on the multi-source mineralization evidence dataset and outputs evidence raster feature representation; The scene graph structure encoding layer performs topological connectivity aggregation on the object-level spatial intelligent scene graph and outputs structural association features; The feature modulation gating layer uses the channel weight layer as a constraint condition to perform feature weight gating and direction consistency modulation on the evidence grid feature representation and structural correlation features, and outputs channel constraint features; The cross-modal spatial attention fusion layer performs spatial location-level attention allocation on the channel-constrained features and outputs a mineral prediction numerical raster to generate a mineral prediction layer.

2. The mineral prediction method based on spatial intelligence as described in claim 1, characterized in that: The specific steps for collecting the multi-source mineralization evidence dataset are as follows. Collect multi-source mineralization evidence data, perform consistency verification and coordinate system unification, and generate a coordinate system unification mineralization evidence package; Spatial resolution resampling transformation and raster pixel scale unification processing are performed on the coordinate system mineralization evidence package to generate a unified resampling package; The resampled unified package is subjected to grid alignment and grid starting point locking to obtain the grid aligned volume. Missing value processing and outlier removal are then performed to generate a multi-source mineralization evidence dataset.

3. The mineral prediction method based on spatial intelligence as described in claim 2, characterized in that: The specific steps for extracting candidate spatial entity features, performing centerline segmentation encoding, partitioning, and boundary fitting are as follows: Candidate linear features, candidate boundary features, and candidate anomaly features are extracted from multi-source mineralization evidence datasets, and cross-modal spatial response alignment and confidence propagation screening are performed to generate a candidate spatial entity feature layer. Refine and repair the candidate linear features in the candidate spatial entity feature layer to obtain the connected skeleton, and perform segmentation and centerline segmentation encoding to generate a segmented encoding set; Partitioning and boundary fitting are performed on candidate boundary features and candidate anomaly features to obtain fitted boundary curves. Spatial adsorption and position correction are then performed using a segmented encoding set to generate a boundary fitting set.

4. The mineral prediction method based on spatial intelligence as described in claim 3, characterized in that: The list of spatial entity elements is generated by calculating the attributes of spatial entity elements based on the boundary fitting set and the segmented coding set, and writing the source and quality tags.

5. The mineral prediction method based on spatial intelligence as described in claim 4, characterized in that: The specific steps for identifying the spatial relationships of the spatial entity element list and constructing a topological connectivity graph are as follows. Based on the list of spatial entity elements, a candidate neighborhood retrieval range is constructed under a unified coordinate system, and spatial relationship identification between objects is performed to generate a spatial relationship record set; The list of spatial entity elements and the set of spatial relationship records are registered as topological nodes and topological edges, respectively. The spatial relationships between duplicate objects are merged and prioritized to generate a topological connectivity graph.

6. The mineral prediction method based on spatial intelligence as described in claim 5, characterized in that: The specific steps for generating the object-level spatial intelligent scene graph are as follows: The connectivity structure attributes are extracted from the topological connectivity graph, and topological connectivity confidence propagation and semantic constraint adjudication are performed to generate a connectivity structure attribute table. By combining the list of spatial entity elements, the set of spatial relationship records, the topological connectivity graph, and the connectivity structure attribute table, and performing semantic relationship reasoning and structured assembly, an object-level spatial intelligent scene graph is generated.

7. The mineral prediction method based on spatial intelligence as described in claim 6, characterized in that: The specific steps for constructing an anisotropic cost field based on the object-level spatial intelligent scene graph are as follows. Based on object-level spatial intelligent scene map, spatial entity elements are filtered and the center line of the channel is rasterized and snapped to generate the starting point raster set. By combining the starting grid set with semantic relations, weighted correction of the barrier boundary direction is performed to construct the initial anisotropic cost field, and direction-sensitive cost recalibration is performed to generate the anisotropic cost field.

8. The mineral prediction method based on spatial intelligence as described in claim 7, characterized in that: The specific steps for generating the channel weight layer are as follows. Based on the anisotropic cost field, the minimum cost channel distance is calculated, and the intersection region is updated and the blocking and suppression constraints are applied to generate a channel distance map. The minimum cost channel distance in the channel distance graph is mapped to a channel weight representation, and weight enhancement, weight suppression, upper limit pruning, and spatial smoothing are performed to generate a channel weight layer.

9. The mineral prediction method based on spatial intelligence as described in claim 1, characterized in that: The specific steps for generating the mineral prediction layer are as follows. Spatial alignment and multimodal spatial correlation feature fusion are performed on the multi-source mineralization evidence dataset, object-level spatial intelligent scene map and channel weight layer to generate fused feature tensor; The fused feature tensor is organized into inference batches and input into the mineral prediction machine learning model to perform mineral inference and obtain a mineral prediction numerical raster. Spatial continuity correction and mineralization probability determination threshold segmentation are performed on the mineral prediction numerical raster to generate a mineral prediction layer.

Images