Natural resource archive retrieval and utilization method combining large model and knowledge graph technology

Abstract

The invention provides a natural resource archive retrieval and utilization method combining a large model and a knowledge graph technology, and relates to the technical field of data processing, the method comprises the following steps: identifying business entities in the field of natural resources and semantic association thereof, and constructing an RDF triple with attribute constraint; performing geometric feature preserving type coordinate simplification on the plot entity vector boundary coordinate point sequence through a Douglas-Peucker compression algorithm to generate a key coordinate point set after noise reduction; on the basis of the key coordinate point set, vector geometric superposition operation is carried out on polygon boundaries of the two plot entities, the area proportion of overlapped regions is calculated according to the intersection-to-union ratio, and when the intersection-to-union ratio exceeds a preset threshold value, entity space alignment is carried out, so that a triple set of space alignment is obtained; and fusing the triple set with the aligned space, and constructing the natural resource domain knowledge graph. According to the invention, redundant information interference is effectively reduced.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a natural resource archive retrieval and utilization method that combines a large model with knowledge graph technology. Background Art

[0002] Constructing domain knowledge graphs enables the associative organization of archival information. For example, some existing technologies extract triples (entity-relationship-entity) from archives to construct natural resource knowledge graphs encompassing spatiotemporal and business relationships, providing a semantically relevant foundation for archival retrieval. Simultaneously, the advantages of large language models in text understanding and information extraction have been introduced into this field. Pre-trained models enhance the accuracy of professional terminology recognition and entity attribute extraction, assisting in the construction and completion of knowledge graphs.

[0003] However, existing technologies still have some limitations in combining large models with knowledge graphs to achieve efficient utilization of natural resource archives: For example, natural resource archives are complex in format (e.g., scanned documents, PDFs, CAD drawings, etc.). Existing technologies have limited automatic parsing capabilities for document directory hierarchies and nested structures, making it difficult to accurately decompose them into sub-documents according to the "Document Object Model topology" and "directory hierarchy rules." This results in coarse-grained semantic analysis. Traditional knowledge extraction methods, some of which rely on rule-based matching or shallow machine learning models, lack a sufficient understanding of the specialized semantics implicit in archives (e.g., the logic of land ownership changes and the relevance of planning policies). While some technologies incorporate large models, they lack dedicated processing mechanisms tailored to the specific characteristics of the natural resource domain (e.g., boundary point coordinates and land use codes), resulting in insufficient completeness and accuracy in the extracted structured metadata. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a natural resource archive retrieval and utilization method that combines large models with knowledge graph technology, which effectively reduces redundant information interference, makes deep semantic understanding more targeted, and greatly improves the degree of automation and accuracy of archive preprocessing.

[0005] In order to solve the above technical problems, the technical solutions of the present invention are as follows: A natural resource archive retrieval and utilization method combining a large model and knowledge graph technology, the method comprising: Step 1: Obtain a target natural resource archive set, wherein the archive set includes at least one natural resource archive to be processed; Step 2: Based on the target natural resource archive set obtained in step 1, the document object model topology structure is parsed and the corresponding sub-document set is generated according to the directory hierarchy rules; Step 3: Using the pre-set natural resources domain model, perform deep semantic understanding and structured information extraction on the sub-document set generated in step 2, and output structured metadata describing the core elements of the archive; Step 4: Based on the structured metadata output from step 3, identify business entities in the natural resources field and their semantic associations, and construct RDF triples with attribute constraints; the identified business entities include land parcel entities with vector boundary coordinate point sequences; Step 5: Perform geometric feature-preserving coordinate simplification on the vector boundary coordinate point sequence of the land entity extracted in step 4 using the Douglas-Peucker compression algorithm to generate a set of key coordinate points after noise reduction. Step 6: Based on the denoised key coordinate point set from step 5, perform vector geometric overlay operations on the polygonal boundaries of the two land parcel entities. Calculate the area ratio of the overlapping region based on the intersection-and-union ratio. When the intersection-and-union ratio exceeds a preset threshold, perform spatial alignment of the entities to obtain a set of spatially aligned triplets. Step 7: Integrate the spatially aligned triplet set completed in step 6 to construct a knowledge graph in the field of natural resources.

[0006] A natural resource archive retrieval and utilization system that combines large models with knowledge graph technology, including: An acquisition module, configured to acquire a target natural resource archive set, wherein the archive set includes at least one natural resource archive to be processed; A parsing module is used to parse the document object model topology structure based on the target natural resource archive set, and to generate corresponding sub-document sets according to the directory hierarchy rules; The execution module is used to perform deep semantic understanding and structured information extraction on the sub-document collection through the preset natural resources domain model, and output structured metadata describing the core elements of the archive; A construction module is used to identify business entities and their semantic associations in the natural resources field based on structured metadata, and construct RDF triples with attribute constraints; wherein the identified business entities include land parcel entities with vector boundary coordinate point sequences; The denoising module is used to perform geometric feature-preserving coordinate simplification on the vector boundary coordinate point sequence of the land entity through the Douglas-Peucker compression algorithm to generate a set of key coordinate points after denoising; The alignment module is used to perform vector geometric overlay operations on the polygonal boundaries of two land entities based on the denoised key coordinate point set. The overlapping area ratio is calculated based on the intersection-and-union ratio. When the intersection-and-union ratio exceeds a preset threshold, the entity space is aligned to obtain a set of spatially aligned triplets. The fusion module is used to fuse the spatially aligned triple sets and construct a knowledge graph in the field of natural resources.

[0007] The above solution of the present invention includes at least the following beneficial effects: By parsing the topological structure of the document object model and disassembling sub-documents according to directory hierarchy rules, we have broken through the limitations of complex formats and rough structural analysis in traditional archive processing and achieved refined splitting of archives. This collection of sub-documents layered according to business logic effectively reduces redundant information interference, makes deep semantic understanding more targeted, and greatly improves the automation and accuracy of archive preprocessing.

[0008] A preset large model of the natural resources field is used to perform deep semantic understanding. Compared with traditional rule matching or general models, it can accurately identify field-specific terms (such as "boundary point coordinates" and "land use nature codes") and implicit business logic (such as ownership change associations). The output structured metadata is more in line with the core elements of the archives. This process solves the problem of insufficient understanding of professional semantics by traditional methods and improves the integrity of metadata.

[0009] Constructing RDF triples with attribute constraints not only records the semantic associations between entities but also enhances the accuracy of these relationships through attribute constraints (such as the area and coordinate range of a plot). In particular, incorporating the "sequence of vector boundary coordinate points" of a plot entity into the entity definition enables the fusion of spatial and semantic features.

[0010] The Douglas-Peucker algorithm is used to perform geometric feature-preserving simplification on the vector boundary coordinate point sequence. While removing noise points (such as redundant coordinates caused by measurement errors), the core geometric outline of the plot is fully preserved. This process reduces the amount of coordinate data and avoids boundary matching deviations caused by coordinate noise.

[0011] Entity space alignment is achieved based on vector geometric overlay operations based on the intersection-union ratio. Compared with traditional alignment methods that rely on text similarity, it solves the identification problems of "same place with different names" and "same name in different places" based on the essential characteristics of space. The quantitative alignment logic when the intersection-union ratio exceeds the preset threshold improves the cross-archival matching accuracy of land entities and realizes knowledge association in the spatial dimension. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is a flow chart of a natural resource archive retrieval and utilization method that combines a large model with knowledge graph technology, provided by an embodiment of the present invention.

[0013] Figure 2 This is a schematic diagram of a natural resource archive retrieval and utilization system that combines a large model with knowledge graph technology, provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0014] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0015] like Figure 1 As shown, an embodiment of the present invention proposes a natural resource archive retrieval and utilization method that combines a large model with knowledge graph technology, and the method includes the following steps: Step 1: Obtain a target natural resource archive set, wherein the archive set includes at least one natural resource archive to be processed; Step 2: Based on the target natural resource archive set obtained in step 1, the document object model topology structure is parsed and the corresponding sub-document set is generated according to the directory hierarchy rules; Step 3: Using the pre-set natural resources domain model, perform deep semantic understanding and structured information extraction on the sub-document set generated in step 2, and output structured metadata describing the core elements of the archive; Step 4: Based on the structured metadata output from step 3, identify business entities in the natural resources field and their semantic associations, and construct RDF triples with attribute constraints; the identified business entities include land parcel entities with vector boundary coordinate point sequences; Step 5: Perform geometric feature-preserving coordinate simplification on the vector boundary coordinate point sequence of the land entity extracted in step 4 using the Douglas-Peucker compression algorithm to generate a set of key coordinate points after noise reduction. Step 6: Based on the denoised key coordinate point set from step 5, perform vector geometric overlay operations on the polygonal boundaries of the two land parcel entities. Calculate the area ratio of the overlapping region based on the intersection-and-union ratio. When the intersection-and-union ratio exceeds a preset threshold, perform spatial alignment of the entities to obtain a set of spatially aligned triplets. Step 7: Integrate the spatially aligned triplet set completed in step 6 to construct a knowledge graph in the field of natural resources.

[0016] In an embodiment of the present invention, by integrating a collection containing at least one archive to be processed, omission of archival information is avoided; it is compatible with data sets containing single or multiple natural resource archives, adapting to archive processing needs of different scales, and by parsing the Document Object Model (DOM) topological structure, breaking through the limitations of the "whole file" extensive processing in traditional archive processing, and can accurately identify the hierarchical relationship within the document (such as chapters, clauses, attachments, etc.); sub-document collections are generated according to directory hierarchy rules, so that each sub-document focuses on a specific business topic (such as "approval process" and "proof of ownership"), reducing redundant information interference, providing more accurate granularity for deep semantic understanding, and improving the efficiency and accuracy of subsequent analysis; using a preset large model of the natural resources field, compared with general models or rule matching methods, it can accurately identify domain-specific terms (such as "boundary points" and "land use pre-examination") and implicit business logic (such as "the relationship between approval process and policy basis"); the output structured metadata directly corresponds to the core elements of the archive (such as plot coordinates, approval date, and owner), solving the problem of information fragmentation extracted by traditional methods.

[0017] Based on structured metadata, business entities (such as plots, departments, policies) and their semantic associations are identified, and RDF triples with attribute constraints are constructed (such as "Plot A-Approval Unit-Natural Resources Bureau (Approval Date: 2023-01-01)"). Compared with triples without attribute constraints, they more accurately reflect business realities. For the first time, "vector boundary coordinate point sequence" is used as the core attribute of the plot entity, breaking through the limitation of traditional entities relying solely on text descriptions and realizing the fusion expression of spatial features and semantic features. While maintaining the geometric outline of the plot (such as boundary shape and inflection point position), measurement errors or redundant coordinate points are removed, reducing the data volume (usually compressible by 40%-60%) and reducing the complexity of subsequent calculations. Through "geometric feature-preserving" denoising, the interference of noisy coordinates on boundary matching is avoided, ensuring that the key coordinate point set can truly reflect the spatial form of the plot and provide high-quality data for precise spatial alignment.

[0018] Through vector geometric overlay operations and intersection-and-union calculations, the relevance of plots is judged based on their essential spatial characteristics, solving the misjudgment problem of "same place, different name" or "same name, different place" caused by traditional reliance on text similarity (such as name matching); using the intersection-and-union threshold (such as 80%) as the basis for alignment, quantitative decision-making of spatial alignment is achieved, and the accuracy of cross-archival plot entity matching is increased to over 90%, effectively breaking down spatial information silos; integrating the spatially aligned triples enables the knowledge graph to contain both semantic associations between entities (such as "plot-policy basis" and "plot-ownership change") and accurate spatial relationships (such as "adjacent plots" and "inclusion relationship"), breaking through the limitation of traditional knowledge graphs that can only express semantic associations. This knowledge graph can support spatial location-based retrieval (such as "querying all industrial land archives in the XX area"), time dimension tracing (such as "historical approval records of plot A"), and cross-entity association analysis (such as "the impact of policy changes on plot use").

[0019] In a preferred embodiment of the present invention, step 2 includes: Step 21: Based on the natural resource archive file obtained in step 1, parse and generate a Document Object Model (DOM) tree structure; recursively calculate the depth value of each node using the DOM tree as the processing object to generate a DOM tree with depth annotations, wherein the root node depth is initialized to 0, and the child node depth value = the parent node depth value + 1; Step 22, using the depth-annotated DOM tree output from step 21 as input: Filter the node set with a depth value ≥ 2, and execute the following for each filtered node: Extract the text blocks of all leaf nodes in the node subtree, arrange the text blocks in the original character offset order to form an ordered sequence, and calculate the distance between adjacent text blocks in the sequence, where the distance value = the starting offset of the subsequent text block - the ending offset of the previous text block; Record the maximum spacing value of the current node. When the maximum spacing value is less than the preset threshold, merge all text blocks in the sequence to generate sub-documents, and output the sub-document set and the corresponding DOM node path and maximum spacing value. Step 23, taking the sub-document set outputted in step 22 as the processing object, assigning a globally unique identifier to each sub-document, associating the DOM node path and the maximum spacing value outputted in step 22, and binding the source file number to generate structured traceability metadata.

[0020] In the embodiment of the present invention, the specific implementation process of the above step 21 is as follows: First, for each natural resource archive file obtained in step 1 (e.g., PDF, Word, scanned document to text, etc.), its internal structure is parsed using a document parsing tool (e.g., PDFBox for PDF, POI for Word, etc.) to generate a corresponding Document Object Model (DOM) tree structure. The nodes of this DOM tree correspond to different content units in the archive. For example, the root node represents the entire archive file, the first-level child nodes can correspond to the archive's chapter titles, the second-level child nodes can correspond to the paragraphs within the chapters, the third-level child nodes can correspond to the sentences or phrases within the paragraphs, and the leaf nodes correspond to the smallest text fragments.

[0021] Next, the generated DOM tree is used as the processing object, and the depth value of each node is calculated recursively: the depth of the root node of the DOM tree is initialized to 0; starting from the root node, all its direct child nodes are traversed in sequence, and a depth value is assigned to each child node. The depth value of a child node is equal to the depth value of its parent node plus 1; then these child nodes are used as parent nodes, and their child nodes are recursively processed. The depth is also calculated according to the rule of "child node depth value = parent node depth value + 1" until all nodes in the DOM tree are traversed; finally, a DOM tree is generated in which each node is clearly marked with a depth, for example, the root node (the entire document) has a depth of 0, the "Chapter 1" node has a depth of 1, the "Section 1.1" node has a depth of 2, the "1.1.1 Paragraph" node has a depth of 3, and so on.

[0022] The specific implementation process of the above step 22 is as follows: Using the depth-annotated DOM tree output from step 21 as input, first filter out all nodes with a depth value greater than or equal to 2 to form a set of nodes to be processed (these nodes usually correspond to sub-content with a certain level in the archive, such as sub-paragraphs under a chapter, details under a clause, etc.).

[0023] For each filter node in the collection, do the following: Extract all leaf nodes in the subtree corresponding to the current node (that is, nodes without child nodes in the DOM tree, usually corresponding to specific text fragments, such as sentences, phrases, data items, etc.), and obtain the text block content of each leaf node; arrange these text blocks according to their character offset order in the original archive file to form an ordered sequence of text blocks. The "original character offset order" here refers to the order in which the text blocks actually appear in the archive file. For example, the text blocks that appear first in the archive are arranged at the front of the sequence, and those that appear later are arranged at the back.

[0024] Calculate the spacing between two adjacent text blocks in the ordered sequence. That is, for any two consecutive text blocks in the sequence (previous text block and subsequent text block), the spacing value is equal to the starting character offset of the subsequent text block in the original file minus the ending character offset of the previous text block in the original file. For example, if the previous text block ends at the 100th character position in the original file and the subsequent text block starts at the 105th character position, then the spacing value between the two is 5.

[0025] Traverse the spacing values ​​of all adjacent text blocks and record the maximum spacing value corresponding to the current node.

[0026] Compare the maximum spacing value of the current node with a preset threshold (the threshold is set according to the file layout characteristics, for example, it can be set to 10 characters for formal official documents, and 15 characters for handwritten scan-to-text files, etc.): If the maximum spacing value is less than the preset threshold, it means that the text blocks in the node subtree are highly coherent as a whole and belong to the same subject content. Therefore, all text blocks in the sequence are merged into a complete text content to generate a sub-document; at the same time, the DOM node path corresponding to the sub-document (that is, the hierarchical path from the root node to the current node, such as "Archive File>Chapter 2>Section 2.3>Paragraph 3") and the maximum spacing value of the current node are recorded as the attribute information of the sub-document, and finally a sub-document set consisting of all qualified sub-documents, as well as the DOM node path and maximum spacing value corresponding to each sub-document are output.

[0027] The specific implementation process of the above step 23 is as follows: The sub-document set output in step 22 is taken as the processing object, and a globally unique identifier is assigned to each sub-document (the identifier can be generated by combining timestamps, random strings, node feature codes, etc. to ensure that it is not repeated in the entire archive processing system); then, the globally unique identifier of each sub-document is associated with the corresponding DOM node path (such as "Archive File>Chapter 1>Section 1.2>Article 2") and the maximum spacing value output in step 22 to form a one-to-one mapping relationship; at the same time, the number of the original natural resource archive from which the sub-document comes (that is, the unique identifier of the archive in the target natural resource archive set in step 1, such as "Land and Resources Archive 2023-001") is obtained, and the source archive number is bound to the above-mentioned globally unique identifier, DOM node path, and maximum spacing value to form structured data containing "sub-document unique identifier, source archive number, DOM node path, and maximum spacing value", that is, the structured traceability metadata of the sub-document; through this process, each sub-document can be clearly traced back to its position in the original archive, the text coherence characteristics when it was generated, and the source archive information.

[0028] The present invention recursively calculates the node depth and accurately quantifies the hierarchical structure of the archive content (such as the nested relationship between chapters, paragraphs, and clauses), providing a clear hierarchical basis for subsequent content splitting according to business logic, avoiding the content fragmentation problem caused by traditional splitting according to fixed length; the DOM tree with depth annotation can intuitively reflect the internal structure of the archive, laying the foundation for screening valuable middle and deep content (such as nodes with depth ≥ 2), and improving the pertinence of sub-document splitting; focusing on nodes with depth ≥ 2 ensures that the split sub-documents focus on specific business details (such as "approval conditions" and "boundary point description"), avoiding the interference of redundant information contained in root nodes or shallow nodes; by calculating the distance between adjacent text blocks and judging the coherence based on the maximum distance value, it can achieve It automatically identifies "semantically complete units" and merges text blocks when the spacing is less than a threshold, ensuring the thematic consistency of sub-document content and solving the semantic break problem caused by traditional format splitting (such as by page or line); it records the DOM node path and the maximum spacing value, and provides structured "content source" and "coherence feature" labels for sub-documents; the globally unique identifier ensures that each sub-document can be accurately located, avoiding duplication or confusion of archive content, and associates the DOM node path, the maximum spacing value and the source archive number to build a complete traceability chain of "sub-document-original location-content feature-source archive"; the structured traceability metadata provides a standardized index for cross-archive content association (such as the description of the same plot in different archives).

[0029] In a preferred embodiment of the present invention, step 3 includes: Step 31: Using the structured provenance metadata generated in step 23 as input, the associated sub-documents are input into a pre-set natural resources domain macro-model. Through the domain adaptation decoding layer of the macro-model, three core entities, namely, plot number, ownership entity, and spatial use category, are extracted to identify the ownership relationship between "ownership entity and plot" and the functional relationship between "plot and spatial use." Step 32: Based on the entities and relationships output in step 31, extract the three attributes of area value, coordinate system type, and issuance date for each land parcel entity; convert the issuance date into a timestamp format, and together with the coordinate system type, form a space-time dual constraint; Step 33: Standardize and encapsulate the structured features outputted in step 32, and output a structured metadata set containing entities, relationships, and constraints. The structured metadata set includes: Entity dimension: Generates a triple of {entity type, entity identifier, attribute set}; Relationship dimension: Generate {subject entity identifier, relationship type, object entity identifier} triples; Constraint dimension: Bind the spatial-temporal dual constraints to the corresponding land entity identifiers.

[0030] In the embodiment of the present invention, the specific implementation process of the above step 31 is as follows: First, the sub-document contents in the structured traceability metadata generated in step 23 are input one by one into the preset natural resources field model. The large model has been adapted and trained for the natural resources field, and its domain adaptation decoding layer contains specific entity recognition and relationship extraction modules.

[0031] For each subdocument in the input, the large model performs the following operations: Through domain dictionary matching combined with contextual semantic analysis, the numbering information that uniquely identifies the plot in the document is identified; using named entity recognition (NER) technology, combined with domain knowledge graphs (such as "Natural Resources Center", "XX Company", "Zhang San" and other organizational or natural person names), the subject that owns the rights to the plot is identified; based on the preset spatial use classification system (such as "residential land", "commercial service facility land", "industrial land", etc.), the use function of the plot is determined through text classification algorithms.

[0032] By analyzing the verb phrases in the text that describe the relationship between the ownership subject and the plot (such as "own", "hold", "registered under the name of..."), and combining it with grammatical dependency analysis, the ownership relationship of "ownership subject-plot" is determined; by identifying the verb structure that describes the use of the plot (such as "used for construction", "planned for") or prepositional phrases (such as "used as..."), the functional relationship of "plot-spatial use" is established; finally, the large model outputs three core entities including plot number, ownership subject, and spatial use category, as well as structured data of two types of relationships: "ownership subject-plot" and "plot-spatial use".

[0033] The process of building a large model for the natural resources domain can be divided into six core steps: domain data construction, basic model selection, domain adaptation training, multimodal fusion, spatiotemporal constraint integration, and entity relationship enhancement. The specific implementation is as follows: Multi-source data collection and integration of three core data types: Unstructured documents (e.g., PDFs and Word documents) such as natural resource approval documents, ownership certificates, and laws and regulations were collected to build a text library containing over 100,000 files. Geographic information data, such as plot vector boundary coordinates and administrative division maps, was collected to form a spatial feature library with a coordinate system. Timestamp information, such as the issuance date and approval time, was extracted to construct a time series annotation dataset. Over 10,000 samples were manually annotated, defining three entity labels: plot number, ownership entity, and spatial use, as well as two relationship labels: ownership and use association. Spatial descriptions in the text (e.g., "East to XX Road") were aligned with vector coordinates to establish a "text-to-space" mapping. A natural resource knowledge graph containing over 500,000 triples was constructed, covering semantic associations such as plot-policy-department relationships, and used as an external knowledge base. For example, the "Land Use Preliminary Review" policy was linked to the "Natural Resources Bureau" approval process node.

[0034] Universal model base: BERT-wwm-ext is used as the basic language model, which has an F1 value of up to 92% in Chinese entity recognition tasks and supports long text processing (up to 512 tokens). If remote sensing imagery needs to be processed, the CLIP model is integrated to build a multimodal encoder to achieve "text-image" joint representation. For example, the land parcel description text and satellite image features are aligned through comparative learning. For vector coordinates, PointNet++ is used to extract geometric features, and the coordinate sequence is converted into a 128-dimensional vector, which is then embedded and spliced ​​into the input model with the text. The issuance date is converted into a timestamp (such as "2023-01-01" → 1672531200) and injected into the model through position encoding. The ROIAlign technology is used to extract the key inflection points of the land parcel coordinates, generate a geometric feature vector, and fuse it with the text semantic features.

[0035] A natural resources dictionary containing more than 20,000 terms (such as "boundary point" and "land use pre-examination") was constructed, and pre-training on domain texts was continued through the MLM (masked language model) task, which increased the model's semantic understanding accuracy of professional vocabulary by 30%; Denoising Autoencoder was used to reduce the noise of spatial coordinates, remove redundant coordinate points caused by measurement errors, and retain geometric contour features; entity recognition (NER) and relationship extraction (RE) were jointly trained, and the CasRel framework was used to handle overlapping relationships, such as simultaneously identifying the "ownership subject-land plot" and "land plot-spatial use" relationships.

[0036] Predict the intersection over union (IoU) of two plots' coordinates to supervise the model's learning of spatial correlation; determine the temporal order of issuance dates to improve logical understanding of the time dimension; during the training phase, use triples in the knowledge graph (such as "Plot A - Approval Unit - Natural Resources Bureau") as soft constraints, and use PromptTuning to guide the model to generate relationships that conform to business logic; use Graph Convolutional Network (GCN) to encode the structural information of the knowledge graph into node embeddings and fuse them with text embeddings.

[0037] The FuzzyWuzzy library is used to implement fuzzy matching to solve the problem of "same place, different names" (such as "No. 123, XX Road" and "No. 123-1, XX Road"); the BiLSTM-CRF layer is used to capture entity boundaries, for example, to identify "XX City Natural Resources and Planning Bureau" as a complete organizational entity; and Stanford CoreNLP is used to extract the subject, predicate, and object structure of a sentence, for example, to parse the "ownership entity-land plot" relationship from "The land plot is held by XX Company."

[0038] Geometric features of plot coordinates (such as area and shape) are used as auxiliary features for relationship extraction. For example, the "plot-contains-subplot" relationship requires spatial inclusion judgment. A spatiotemporal gating unit (ST-Gate) is added to the model output layer to dynamically adjust the confidence level of entity relationships based on timestamp and coordinate features. For example, if the coordinates of two plots overlap but the issuance dates differ by 10 years, the probability of the "same plot" relationship is automatically reduced.

[0039] The F1 value (target ≥90%) and boundary accuracy (coordinate point matching error ≤5 meters) are used; relationship extraction: Hits@10 is used to evaluate the accuracy of triple prediction, and the accuracy of the "owner-plot" relationship is required to be ≥92%; in the spatial alignment task, when the IoU threshold is set to 0.8, the cross-archival plot matching accuracy must be ≥90%; adversarial sample generation is introduced, such as inserting interference terms in the text (such as changing "industrial land" to "industrial land") to improve model robustness; temperature scaling is used to calibrate the model output probability to reduce misjudgment caused by "same place, different names".

[0040] Regular updates to the domain dictionary and knowledge graph are used, and continuous pretraining is used to prevent the model from forgetting new knowledge. For example, when new terms related to "ecological red lines" are added, the F1 value drop is controlled within 2%; model distillation is used to compress model parameters to 1 / 4 of the original volume, increasing the inference speed by 3 times while maintaining an F1 value of ≥88%; during deployment, NVIDIA TensorRT is combined for inference acceleration, and the processing time for a single archive is ≤2 seconds; services are provided through RESTful API, supporting batch processing (1,000 archives / hour) and returning structured metadata (such as plot number, ownership subject, and coordinate set); and integration with GIS platforms (such as ArcGIS) enables visual analysis of the spatial relationships of plots (such as highlighting adjacent plots). Through the above construction process, the model achieved an F1 value of 91.2% for entity recognition, 90.5% for relationship extraction, and 92.3% for spatial alignment in natural resource archive processing tasks, significantly outperforming traditional rule matching methods (accuracy increased by more than 40%), and effectively supporting intelligent retrieval and association analysis of natural resource archives.

[0041] The specific implementation process of step 32 above is as follows: Based on the entity and relationship data output in step 31, for each identified land parcel entity, further extract and process its attributes: Area value extraction: Locate the numerical description associated with the plot number in the subdocument and extract the area value by combining quantifier matching (such as "square meters," "hectares," and "mu") with numerical regular expressions. If unit conversion is required (such as "10,000 square meters (equivalent to 15 mu)"), convert it to a standard unit (such as square meters).

[0042] Coordinate system type identification: The reference system used to describe the plot coordinates is identified by text matching against a preset list of coordinate system names (such as "2000 National Geodetic Coordinate System", "Region A 54 Coordinate System", and "Region B 80 Coordinate System"). If the document does not explicitly mention it, the possible coordinate system type is inferred through the coordinate numerical features (such as latitude and longitude ranges and projection zone numbers).

[0043] Issuance date processing: Identify date expressions related to land rights in documents (such as "registration date", "issuance date", and "approval date"), and extract specific dates using date regular expressions (such as "YYYY-MM-DD" and "YYYY year MM month DD day"). Convert the extracted date string to a standard timestamp format (such as Unix timestamp) to facilitate subsequent comparison and calculation of time dimensions.

[0044] Construction of space-time dual constraints: The standardized coordinate system type (such as "CGCS2000") and the timestamp (such as 1640995200) are combined to form the "space-time dual constraint" attribute of each plot entity, which is used to limit the valid time range of the plot in a specific coordinate system.

[0045] The specific implementation process of step 33 above is as follows: The structured features output in step 32 are standardized and encapsulated to form a structured metadata set in a unified format: Entity dimension encapsulation: Generate a triple of {entity type, entity identifier, attribute set} for each identified entity (land parcel, ownership subject, spatial use category).

[0046] Entity type: clearly marked as "land parcel", "ownership entity" or "spatial use category".

[0047] Entity identification: Assign a unique identifier (such as a UUID) to each entity to ensure traceability of entities across documents.

[0048] Property collection: Land parcel entity: contains attributes such as area value, coordinate system type, timestamp (issuance date), etc.

[0049] Ownership subject: includes attributes such as name and type (organization / individual).

[0050] Space use category: includes attributes such as category code and Chinese name.

[0051] Relational dimension encapsulation: Generate a {subject entity identifier, relationship type, object entity identifier} triple for each identified relationship.

[0052] Principal Entity Identifier: The entity identifier that points to the initiator of the relationship.

[0053] Relationship type: marked as "attributive relationship" or "functional relationship".

[0054] Object entity identifier: The entity identifier that points to the recipient of the relationship.

[0055] Constraint dimension binding: The "spatial-temporal dual constraint" attribute generated in step 32 is bound to the corresponding plot entity identifier to form a {plot entity identifier, spatial constraint (coordinate system type), temporal constraint (timestamp)} mapping relationship. The final output structured metadata set integrates information from the three dimensions of entity, relationship, and constraint.

[0056] The present invention uses a preset large model of the natural resources field and an adaptive decoding layer to accurately extract professional entities such as plot numbers, ownership entities, and spatial use categories. Compared with general models, the recognition accuracy of field terms such as "real estate certificate number" and "commercial service facility land" is improved by more than 30%; it automatically identifies two key relationships, "ownership entity-plot" and "plot-spatial use", and associates discrete entities into semantic networks, providing a relationship basis for subsequent knowledge graph construction, supporting complex searches such as "querying all commercial land under a certain enterprise"; it uniformly extracts three attributes, namely area value, coordinate system type, and issuance date, and converts the date into a timestamp format, which solves the problem of inconsistent date formats in different files (such as "January 1, 2023" and "2023-01-01"). It improves data comparability; combines the coordinate system type with the timestamp to form a "space-time dual constraint", giving the plot entity a validity limit in the space-time dimension (such as the validity period of a plot in the 2000 coordinate system), solving the problem of coordinate system consistency in historical archives and supporting cross-time spatial analysis of plots; through the standardized encapsulation of the three dimensions of entities, relationships, and constraints, a triple set in a unified format is formed, which provides a standardized interface for subsequent knowledge graph construction and reduces data integration costs; binds the space-time constraints to the plot entity identifier, so that the attributes and relationships of each plot have the ability to verify the space-time validity, avoiding the defect of "static" entity relationships in traditional knowledge graphs, and supporting space-time sensitive queries such as "querying industrial land registered before 2020 using the Beijing 54 coordinate system".

[0057] In summary, step 3 realizes the transformation of natural resource archives from "text information" to "structured knowledge" through the domain big model. In particular, the innovative "space-time dual constraint" mechanism solves the multi-source and inconsistency problems of the space-time dimensions in archival data, and significantly improves the knowledge graph's ability to express complex natural resource information and query flexibility.

[0058] In a preferred embodiment of the present invention, step 4 includes: Step 41: Using the structured metadata set output from step 3 as input, parse the entity identifier, entity type, and attribute set in the entity dimension to identify the core business entity types in the natural resources field. The land parcel entity type is identified as a key entity containing a sequence of vector boundary coordinate points. The ownership entity type is also identified. Step 42: Taking the business entity type identified in step 41 as the processing object, for each land parcel entity, extract a vector boundary coordinate point sequence from the structured metadata set as the core spatial attribute of the land parcel entity, and add the coordinate point sequence to the attribute set of the land parcel entity to form an updated land parcel entity attribute set; Step 43: Based on the updated land parcel entity attribute set in step 42 and the relationship dimension and constraint dimension output in step 3, a set of RDF triples with attribute constraints is constructed, where: Each attribute value in the entity attribute set is converted into a triple with the entity identifier as the subject, the attribute name as the predicate, and the attribute value as the object; Each relation in the relational dimension is converted into a triple with the subject entity identifier as the subject, the relation type as the predicate, and the object entity identifier as the object; The spatial-temporal dual constraints in the constraint dimension are bound to the corresponding plot entity identifier, generating a triple with the plot entity identifier as the subject, the constraint type as the predicate, and the constraint value as the object.

[0059] In the embodiment of the present invention, the specific implementation process of the above step 41 is as follows: First, using the structured metadata set output in step 3 as input, we parse entity-level information. For each entity, we extract its unique entity identifier, predefined entity type (e.g., "land parcel," "ownership entity," "spatial use category"), and a set of attributes containing specific characteristics.

[0060] During the parsing process, the core business entity types are identified by matching the preset natural resources domain ontology: 1. Land parcel entity: When the entity type field is clearly marked as "land parcel" and its attribute set contains the "vector boundary coordinate point sequence" attribute (such as the coordinate data extracted from the domain model in step 3), the entity is determined to be a land parcel entity type.

[0061] 2. Ownership entity: When the entity type field is marked as "ownership entity" and the attribute set contains features such as "name" and "type (organization / individual)", it is determined to be the ownership entity type.

[0062] In addition, other possible business entity types (such as spatial use categories) will be identified, but the land parcel entity and the ownership entity entity will be the focus of subsequent processing. The final output is the identified business entity type and its corresponding entity identifier set.

[0063] The specific implementation process of the above step 42 is as follows: For the land parcel entity type identified in step 41 , its core spatial attributes are further extracted from the structured metadata set.

[0064] For each parcel entity: In the structured metadata set output in step 3, locate the attribute set corresponding to the plot entity; extract the "vector boundary coordinate point sequence" attribute value from it, which is usually an ordered list of multiple coordinate points (such as longitude and latitude pairs) that describes the polygonal boundary of the plot; add the extracted vector boundary coordinate point sequence to the attribute set of the plot entity, replacing or supplementing the original possibly incomplete spatial information, to form an updated attribute set of the plot entity.

[0065] For example, if the original attribute set of a parcel entity contains information such as area and coordinate system type, this step will supplement the complete "vector boundary coordinate point sequence" to make its attribute set more comprehensive in describing the spatial characteristics of the parcel. Ultimately, the updated attribute set of all parcel entities is output.

[0066] The specific implementation process of the above step 43 is as follows: Based on the updated set of land parcel entity attributes in step 42 and the relationship dimension and constraint dimension information output in step 3, a set of RDF triples with attribute constraints is constructed.

[0067] Attribute triple generation: For each entity (including plot entities, ownership entity, etc.), traverse each attribute value in its attribute set and generate a triple with the entity's unique identifier as the subject, the attribute name as the predicate, and the attribute value as the object; for example, the plot entity ID is "plot_001", and its attribute "area value" is 10,000 square meters, then the generated triple is: (plot_001, area value, 10,000 square meters).

[0068] Relation triple generation: For each relationship in the relationship dimension in step 3, extract its subject entity identifier, relationship type, and object entity identifier; generate a triple with the subject entity identifier as the subject, the relationship type as the predicate, and the object entity identifier as the object; for example, if there is a relationship "Enterprise A - owns - Plot_001", generate the triple: (Enterprise A, owns, Plot_001).

[0069] Constrained triple generation: For each spatial-temporal dual constraint in the constraint dimension of step 3, bind it to the corresponding plot entity identifier; generate a triple with the plot entity identifier as the subject, the constraint type (such as "coordinate system type" and "timestamp") as the predicate, and the constraint value as the object; for example, the coordinate system type of plot_001 is "2000 National Geodetic Coordinate System" and the timestamp is 1640995200, then two triples are generated: (plot_001, coordinate system type, 2000 National Geodetic Coordinate System), (plot_001, timestamp, 1640995200).

[0070] Finally, all generated attribute triples, relationship triples and constraint triples are integrated to form an RDF triple set with attribute constraints.

[0071] The present invention accurately identifies two core business entities in the field of natural resources: plots and ownership entities, by parsing entity dimension information; uses "sequences of vector boundary coordinate points" as the key identifier of plot entities, strengthening the core position of spatial features in natural resource management; incorporates vector boundary coordinate point sequences into the set of plot entity attributes, so that the plot description is transformed from "textual" to "spatial", realizing the fusion expression of semantic information and geometric information; complete vector boundary data provides the necessary conditions for spatial overlay operations in subsequent steps (such as calculating the overlapping area of ​​plots), breaking through the limitation of traditional knowledge graphs that can only process textual relations; entity attributes, relationships and constraints are uniformly converted into RDF triple format to form a standardized knowledge expression framework, providing a unified interface for the construction of knowledge graphs; by binding the dual constraints of space and time, the triples not only express "what" but also limit "when it is valid" and "under what spatial reference it is valid", enhancing the semantic expression ability of knowledge graphs for complex natural resource information.

[0072] In a preferred embodiment of the present invention, step 5 includes: Step 51: Using the RDF triple set constructed in step 4 as input, process the vector boundary coordinate point sequence corresponding to each land parcel entity; initialize the first and last points of the coordinate point sequence as key points to form an initial key point set; Step 52: Based on the initial key point set determined in step 51, recursively partition the line segment formed by the first and last points, that is, traverse all non-key coordinate points between the first and last points of the line segment, calculate the vertical distance of each point to the line segment, and identify the point with the largest vertical distance; Step 53: When the maximum vertical distance value identified in step 52 exceeds the preset distance threshold, the corresponding point is marked as a new key point and added to the key point set; then, the original line segment is split into two sub-segments, front and back, with the new key point as the boundary, and the maximum vertical distance calculation and key point determination process described in step 52 is recursively performed on each sub-segment; In step 54 , after the recursive process is terminated, the key points generated by all levels of judgment are merged and arranged in the order of the original coordinate sequence to generate a denoised key coordinate point set as a simplified boundary representation of the land entity.

[0073] In the embodiment of the present invention, the specific implementation process of the above step 51 is as follows: From the RDF triples generated in step 4, filter out all plot entities that contain vector boundary coordinate point sequences. The coordinate point sequences of each plot entity are arranged in the order recorded in the original archive (e.g., clockwise / counterclockwise coordinate points from the starting point to the end point of the plot boundary). For each plot entity's coordinate point sequence, directly mark the first point (first point) and the last point (last point) in the sequence as initial key points. These two points are the basic endpoints of the plot boundary, ensuring that the starting and ending positions of the boundary outline are not lost. The first point and the last point form the initial key point set, which only contains these two coordinate points and serves as the starting basis for subsequent recursive partitioning.

[0074] The specific implementation process of the above step 52 is as follows: Using the first and last points in the initial key point set determined in step 51 as endpoints, a virtual line segment is formed (i.e., a straight line connecting the two points). All non-key coordinate points except the first and last points in the original coordinate point sequence corresponding to the line segment (these points are the original recorded points in the middle of the line segment) are traversed. For each non-key coordinate point, the perpendicular distance to the virtual line segment is calculated (i.e., the shortest straight-line distance between the point and the line segment, reflecting the degree of deviation of the point from the line segment). Among all the calculated perpendicular distances, the one with the largest value is found, and the corresponding coordinate point is recorded. This point is the point in the current line segment that deviates most significantly from the basic line segment.

[0075] The specific implementation process of the above step 53 is as follows: The maximum vertical distance value identified in step 52 is compared with a preset distance threshold (the threshold is set according to the accuracy requirements of the plot, such as 0.5 meters for urban construction land and 2 meters for rural collective land); if the maximum vertical distance value exceeds the threshold, it means that the point has a greater impact on the morphology of the plot boundary contour and needs to be retained, so the point is marked as a new key point and added to the key point set; using this new key point as the dividing point, the original first point-end point line segment is divided into two new sub-segments: one is the line segment from the first point to the new key point, and the other is the line segment from the new key point to the end point; for the two divided sub-segments, the operation of step 52 is repeated respectively, and the non-key coordinate points within the range of each sub-segment are traversed, the vertical distance of each point to the corresponding sub-segment is calculated, and the point with the maximum vertical distance is found.

[0076] The specific implementation process of the above step 54 is as follows: The recursive process of steps 52 and 53 is continuously repeated. Each time a sub-segment is segmented, it is checked whether there are points that deviate beyond the threshold. If there are, the segmentation is continued and key points are added until the vertical distances of the remaining non-key coordinate points in all sub-segments to the corresponding segments are less than the preset threshold, and the recursive process is terminated. All coordinate points marked as key points in the recursive process are collected (including the initial first point, the last point, and the key points newly added in each level of segmentation). These key points are arranged in the order of their sequence in the original coordinate point sequence (keeping the direction consistent with the original boundary, such as the clockwise order unchanged) to form the final key coordinate point set. This key coordinate point set removes redundant points with small deviations in the original sequence (such as small fluctuation points caused by measurement errors), and simplifies the coordinate data while keeping the overall geometric shape of the land boundary (such as corners and protrusions) unchanged, so as to serve as a simplified boundary representation of the land entity.

[0077] The present invention uses the first and last points as the initial benchmark, and selects points exceeding the threshold as key points by recursively calculating the vertical distance. The main turning points and protruding parts of the plot boundary can be retained during the simplification process, ensuring that the simplified boundary can still accurately reflect the basic shape characteristics of the plot; non-key coordinate points close to the line segment are eliminated, which greatly reduces the number of coordinate points describing the plot boundary, reduces the pressure of data storage and processing, and makes subsequent analysis, transmission and other operations of the plot entity more efficient; filters out subtle fluctuation points caused by measurement errors or redundant records, making the plot boundary representation more concise and smooth, and avoiding these irrelevant points from interfering with plot shape judgment and related applications (such as plot area calculation, boundary comparison, etc.); finally, the key points are arranged in the order of the original coordinate sequence, ensuring the logical consistency of the simplified boundary coordinate point set with the original sequence, without changing the overall direction and relative position relationship of the plot boundary, and ensuring the reliability of the simplification result.

[0078] In a preferred embodiment of the present invention, step 6 includes: Step 61: Using the denoised key coordinate point set of each land parcel entity output in step 5 as input, two land parcel entities to be subjected to spatial relationship analysis are selected to form an entity pair; Step 62: for the entity pairs selected in step 61, reconstruct the corresponding first polygon boundary and second polygon boundary using their respective key coordinate point sets; Step 63, performing a vector geometric superposition operation on the first polygon boundary and the second polygon boundary reconstructed in step 62, calculating the overlapping area polygon formed by the intersection of the two polygons, and respectively calculating the area value of the overlapping area polygon, the area value of the first polygon boundary, and the area value of the second polygon boundary, including: step 631, according to the key coordinate points of the first polygon boundary reconstructed in step 62, traversing all the key coordinate points of the first polygon boundary, identifying the point set located on the outermost boundary, connecting the point set on the outermost boundary to form a minimum convex polygon outline surrounding all internal points, and generating a first candidate convex hull; step 632, according to the key coordinate point set of the second polygon boundary reconstructed in step 62, traversing all the key coordinate points of the second polygon boundary, identifying the point set located on the outermost boundary, connecting the point set on the outermost boundary to form a minimum convex polygon outline surrounding all internal points. The minimum convex polygon outline is used to generate a second candidate convex hull; step 633, calculating the area value of the first candidate convex hull, calculating the original area value of the key coordinate point set of the first polygon boundary, calculating the absolute difference between the area of ​​the first candidate convex hull and the original area, and calculating the area error rate. When the area error rate is less than the preset convex hull simplification tolerance threshold, the first candidate convex hull is determined to be valid. Otherwise, the key coordinate point set of the first polygon boundary is used as the first candidate convex hull, and the same validity verification process is performed on the second candidate convex hull, and the final valid first convex hull boundary and the second convex hull boundary are output; step 634, calculating the overlapping area polygon formed by the intersection of the final valid first convex hull boundary and the second convex hull boundary; step 635, respectively calculating the area value of the overlapping area polygon obtained in step 634, the area value of the first convex hull boundary output in step 633, and the area value of the second convex hull boundary; Step 64: Calculate an intersection-over-union ratio based on the area value calculated in step 63, where the intersection-over-union ratio = overlapping area / (area of ​​the first polygon + area of ​​the second polygon - area of ​​the overlapping area); Step 65, comparing the IoU value calculated in step 64 with a preset IoU threshold; Step 66: When the IoU value exceeds a preset IoU threshold, it is determined that the two land parcel entities in the entity pair have a spatial overlap relationship, and an entity spatial alignment operation is performed, i.e., a semantic triple representing the spatial alignment relationship is generated, where the subject entity is the first land parcel entity identifier, the predicate is the preset spatial alignment relationship type, and the object entity is the second land parcel entity identifier. In step 67 , all semantic triples generated in step 66 that meet the intersection-over-union ratio threshold condition and complete the spatial alignment operation are collected to form a spatially aligned triple set.

[0079] In the embodiment of the present invention, the above step 61, entity pair selection, may include: Extract the coordinate point set. From all the plot entities output in step 5, extract the key coordinate point set of each plot entity after noise reduction. Each coordinate point set is an ordered list containing the simplified plot boundary coordinate points (such as [(x1, y1), (x2, y2), ...]). Based on business needs (such as plot merging, ownership verification, spatial relationship analysis), determine the scope of the plots that require spatial relationship analysis. For example: To verify changes in land parcels within the same area over different periods, select pairs of entities with the same parcel number from adjacent years. To identify parcels with overlapping ownership, select parcels with different ownership within the same area. From the extracted parcel entities, select two parcels to form an entity pair. For example, select "Parcel A" and "Parcel B" as the entity pair for analysis and record their key coordinate point set.

[0080] The above step 62, polygon boundary reconstruction, may include: Sorting coordinate points: Arrange the key coordinate point sets of each plot in the entity pair selected in step 61 in the original sequence order (e.g., clockwise). Ensure that the coordinate points can correctly form a closed polygon after being connected.

[0081] Closed polygon construction: For the first plot (e.g., "Plot A"), connect the sorted coordinate points in sequence, closing the last point with the first point to form the first polygon boundary. For the second plot (e.g., "Plot B"), perform the same operation to form the second polygon boundary. Check whether the reconstructed polygon meets the following requirements: Contains at least three coordinate points (forming a basic polygon), and all edges do not intersect (a simple polygon). If there are any intersecting edges, they need to be corrected (such as adjusting the order of the coordinate points or deleting the points that cause the intersection).

[0082] The above step 63, vector geometric overlay operation and area calculation, includes: Step 631, the first candidate convex hull is generated: Identify the outermost point, traverse all key coordinate points of the first polygon boundary, for each point P: Check whether all other points are on the same side of a line passing through point P. If such a line exists, then P is the outermost point.

[0083] Convex hull construction: Collect all the outermost points, arrange them in the order of the original coordinate point set, and connect these points in sequence to form a closed convex polygon, which is the first candidate convex hull.

[0084] The above step 632, generating the second candidate convex hull, includes: Repeat the operation of step 631 to process the key coordinate point set of the second polygon boundary to generate a second candidate convex hull.

[0085] Step 633: Convex hull validity verification Area calculation: Calculate the area of ​​the first candidate convex hull (calculated by the convex hull boundary coordinate points), and calculate the area of ​​the original coordinate point set of the first polygon boundary.

[0086] Error rate calculation: Calculate the absolute difference in area: |convex hull area - original area|.

[0087] Calculate the area error rate: (absolute difference / original area) × 100%.

[0088] Validity judgment: If the error rate is less than a preset threshold (such as 5%), the first candidate convex hull is valid. If the error rate is greater than or equal to the threshold, the candidate convex hull is discarded and the first polygon boundary is directly used as the first convex hull boundary. The same verification process is performed on the second candidate convex hull to determine the final valid second convex hull boundary.

[0089] The above step 634, determining the overlapping area polygons, may include: Line segment intersection detection: Traverse each line segment (edge) of the first convex hull boundary and check whether it intersects with all line segments of the second convex hull boundary. If the two line segments intersect, calculate the coordinates of the intersection point, collect all intersection points, and the boundary points that are simultaneously inside the two convex hulls; arrange the selected points in clockwise or counterclockwise order (the direction can be determined by the cross product), and connect these points to form a closed polygon, that is, the overlapping area polygon.

[0090] The area value calculation in step 635 includes: Overlapping area: Calculate the area of ​​the overlapping polygons based on their coordinate points.

[0091] Area of ​​the first convex hull: If the first candidate convex hull is valid, its area is used; otherwise, the area of ​​the first polygon boundary is used.

[0092] Second convex hull area: Similarly, the area value to be used is determined based on the validity of the second candidate convex hull.

[0093] Step 64, calculating the intersection-over-union ratio, may include: Calculation of union area: The union area = the area of ​​the first convex hull + the area of ​​the second convex hull - the area of ​​the overlapping area (avoiding repeated counting of the overlapping part).

[0094] Intersection-over-union calculation: Intersection-over-union ratio = overlapping area / union area × 100%.

[0095] Step 65, comparing the IoU threshold value may include: The intersection-over-union ratio calculated in step 64 is compared with a preset threshold (eg, 80%) to determine whether the spatial alignment condition is met.

[0096] Step 66, entity space alignment and triple generation may include: Alignment determination: If the intersection-union ratio is greater than the threshold, it is determined that the two land entities have a spatial overlap relationship and need to be aligned.

[0097] Triplet generation: Main entity: the unique identifier of the first plot (such as the plot number).

[0098] Predicate: The preset spatial alignment relationship type (such as "spatial overlap", "same plot", "contains", etc.).

[0099] Object entity: the unique identifier of the second plot of land.

[0100] Attribute additions: Metadata such as the intersection-union ratio and the area of ​​overlapping regions are added to the triples to enhance the semantic expression of the relationship.

[0101] Step 67, forming a set of spatially aligned triples may include: Results collection: Summarize the semantic triples generated by all entity pairs that meet the threshold conditions in step 66.

[0102] Deduplication processing checks whether there are repeated triplets (such as "A-spatial overlap-B" and "B-spatial overlap-A"), and retains one of them or merges them.

[0103] Set construction, the final processed triples are stored in a unified format to form a spatially aligned triple set for subsequent knowledge graph construction.

[0104] The present invention can accurately determine the spatial overlap between land entities and improve the reliability of spatial relationship analysis by reconstructing polygons based on key coordinate points, calculating convex hulls and intersection-over-union ratios, etc.; it uses the denoised key coordinate point set for processing to reduce the amount of data and improve computational efficiency; at the same time, through mechanisms such as convex hull validity verification, it ensures the accuracy of the analysis results; it presents spatial overlapping relationships in the form of semantic triples, which facilitates the subsequent storage, query and further analysis and utilization of spatial relationships; through spatial alignment operations, it associates the boundary information of land entities that may originally have differences, thereby improving the usability of land data in scenarios such as spatial analysis and integration.

[0105] In a preferred embodiment of the present invention, step 7 includes: Step 71: using the spatially aligned triple set output in step 6 as a first input and the RDF triple set with attribute constraints constructed in step 4 as a second input; Step 72: Perform triple fusion processing on the first input and the second input provided in step 71, that is, identify triples describing the same entity identifier in the two input sets and merge their attributes and relationship information; for the spatially aligned relationship triples generated in step 6, incorporate them as new relationships into the fused triple set; Step 73: Using the triplet set fused in step 72 as the basis for graph construction, each unique entity identifier in the triplet is mapped to a node in the graph database, and each entity attribute is mapped to an attribute key-value pair of the corresponding node; Step 74: Based on the nodes created in step 73, the semantic relationships defined in the triples and the spatial alignment relationships added in step 72 are mapped into directed edges in the graph database, i.e., the node corresponding to the subject entity identifier is the edge start point, the relationship type is the edge label, and the node corresponding to the object entity identifier is the edge end point. Step 75: For the entity node of the land parcel, the key coordinate point set after noise reduction in step 5 is attached to the corresponding node as a spatial geometric attribute; Step 76, after completing the node, edge and attribute mapping of steps 73 to 75, instantiate and generate a knowledge graph in the field of natural resources in the graph database. The graph contains nodes such as plot entities and ownership subject entities, and a semantic network structure composed of directed edges such as ownership relationships, functional relationships and spatial alignment relationships.

[0106] In the embodiment of the present invention, the above step 71, entity pair selection, may include: Extract the set of spatially aligned triplets. From the output of step 6, obtain all spatially aligned triplets. Each triplet has the structure {subject entity identifier, spatial relationship type, object entity identifier}, for example, "Plot A - Spatial Overlap - Plot B." Some triplets also include attribute information such as the intersection-over-union ratio (e.g., 0.85) and the overlap area (e.g., 500 square meters).

[0107] Extract the set of RDF triples with attribute constraints. From the result of step 4, obtain all RDF triples with attribute constraints. These triples include: Entity attribute information, such as "Plot A - Area - 1000 square meters" and "Plot A - Usage - Industrial land"; Semantic relationship information, such as "Company X-Ownership-Plot A" (ownership relationship), "Plot A-Function-Industrial Production" (functional relationship), and some relationships come with constraints (such as "Approval Date: 2023-01-01").

[0108] The triple fusion process in step 72 may include: Entity ID matching: Iterate over each entity identifier in the spatially aligned triples (e.g., "Plot A," "Company X") and search for entities with the same identifier in the RDF triples with property constraints. For example, use the unique number of "Plot A" (e.g., "G2023001") as the matching criterion to ensure that entities across the collection are accurately associated.

[0109] Merge attributes with relationships: For the same matching entity, all its attributes are merged. For example: In the RDF triple, "Plot A" has the attributes "Area: 1000 square meters" and "Purpose: Industrial Land". In the spatial alignment triple, "Plot A" is associated with the attribute "Intersection-over-Union Ratio: 0.85". After merging, the attribute set of "Plot A" is {Area: 1000 square meters, Purpose: Industrial Land, Intersection-over-Union Ratio: 0.85}.

[0110] The original semantic relationship in the RDF triples (such as "Company X-Owned-Plot A") is retained, and the new relationship in the spatial alignment triples (such as "Plot A-Spatial Overlap-Plot B") is added to the fused set to form a complete relationship network.

[0111] Conflict resolution: If the same attribute of the same entity is inconsistent in the two sets (for example, "Plot A - Purpose - Industrial Land" in the RDF triple, and "Purpose - Commercial Land" in the document associated with the spatial alignment triple), the attribute of the RDF triple in step 4 shall prevail (because it is extracted from the core elements of the original archive and is more authoritative), and the conflict information shall be recorded for subsequent manual verification.

[0112] The node mapping in step 73 may include: Traverse the fused triple set, extract all unique entity identifiers, and remove duplicates; for example, filter out 4 unique entities from "Plot A", "Company X", "Plot B", and "Policy Y"; classify unique identifiers by entity type, such as "Plot Entity" (Plot A, Plot B), "Rights Entity" (Company X), "Policy Entity" (Policy Y), etc., and define a unified node label for each type (such as "Plot", "Rights Entity", and "Policy").

[0113] Node creation: In a graph database, a node is created for each unique entity identifier. For example, a node labeled "Land Plot A" is created for "Land Plot A," and a node labeled "Ownership Entity" is created for "Company X."

[0114] Property key-value pair mapping: Convert the attribute information of each entity into a node's attribute key-value pair. For example: The node attributes of "Plot A" are {Number: Plot A, Area: 1,000 square meters, Purpose: Industrial land, Coordinate system: WGS84}; the node attributes of "Company X" are {Name: XX Co., Ltd., Unified Social Credit Code: 91XXXXXXXXXXX, Registered Address: XX Road, XX City}.

[0115] The above step 74, edge mapping may include: Relation extraction and classification: From the fused triples, all semantic relations are extracted, including: The original relationship in step 4 (such as "attribution relationship" and "functional relationship"); The newly added spatial alignment relationships in step 72 (such as "spatial overlap", "contains", and "adjacent").

[0116] Define unique edge labels for each relationship type, such as "belongs", "function", "spatial overlap".

[0117] Directed edge creation: For each triple {subject entity identifier, relationship type, object entity identifier}, create a directed edge in the graph database: The starting point of an edge is the node corresponding to the subject entity identifier (e.g., the "Company X" node); the label of the edge is the relationship type (e.g., "ownership"); and the endpoint of the edge is the node corresponding to the object entity identifier (e.g., the "Plot A" node). For example, the triple "Company X - Ownership - Plot A" corresponds to the edge "Company X node (starting point) - Ownership (label) - Plot A node (end point)."

[0118] Edge attribute addition: If a relationship has attributes (such as the intersection-over-union ratio in a spatial alignment relationship or the date in an approval relationship), these attributes are used as key-value pairs for the edge. For example, the attributes for the edge "Plot A - Spatial Overlap - Plot B" are {Intersection-over-union ratio: 0.85, Overlap area: 500 square meters}.

[0119] The above step 75, adding spatial geometric attributes, may include: Coordinate point set matching: From the output of step 5, extract the key coordinate point set for each land entity after noise reduction. For example, the key coordinate point set for "Land A" is [(x1, y1), (x2, y2), (x3, y3)] (arranged in clockwise order).

[0120] Plot node attribute update: In the graph database, find all nodes labeled "Plot" and match the corresponding key coordinate point set using the entity identifier (such as "Plot A"). This matched point set is used as a spatial geometry attribute and appended to the properties of the corresponding plot node. For example, the "Plot A" node has a new attribute {Key coordinate point set: [(x1, y1), (x2, y2), (x3, y3)], Geometry type: Polygon}.

[0121] Spatial index association: Associate a spatial index (such as an R-tree index based on coordinate range) with the "key coordinate point set" attribute of the plot node to ensure that subsequent spatial queries (such as "query all plots in area XX" and "find plots adjacent to plot A") can be executed quickly.

[0122] Step 76 above, instantiating the knowledge graph, may include: Graph structure integration: In the graph database, the nodes created in step 73, the directed edges created in step 74, and the spatial geometric attributes added in step 75 are integrated to form a complete graph structure. For example: The nodes include "Plot A" and "Plot B" (plot entities), "Company X" (ownership subject), and "Policy Y" (policy entity); the directed edges include "Company X-ownership-Plot A" (ownership relationship), "Plot A-Function-Industrial Production" (functional relationship), and "Plot A-Spatial Overlap-Plot B" (spatial alignment relationship).

[0123] Integrity Verification: Check that all nodes have at least one associated edge (to avoid isolated nodes). For example, if "Plot C" has no associated edges, trace back to its original triples from steps 4 and 6 and fill in the missing relationships (e.g., "Company Y - Ownership - Plot C"). Check that all plot nodes have key coordinate point sets attached. If missing, rematch the results from step 5 and fill in the missing relationships. Create indexes for frequently queried fields (e.g., plot number, ownership entity name, and spatial relationship type) to improve query speed. Partition the graph by business scenario (e.g., administrative division into "XX District Sub-Graph" and "XX County Sub-Graph") for easier management and scalability. After completing all the above mapping and verification in the graph database, a formal natural resources domain knowledge graph is generated. This graph is a semantic network consisting of nodes (entities), directed edges (relationships), and attributes. It can directly support complex queries (e.g., "Find all plots in XX City that overlap with industrial land and their ownership entities") and association analysis (e.g., "The impact path of policy Y on the change of use of plot A").

[0124] like Figure 2 As shown, a natural resource archive retrieval and utilization system combining large models and knowledge graph technology includes: An acquisition module, configured to acquire a target natural resource archive set, wherein the archive set includes at least one natural resource archive to be processed; A parsing module is used to parse the document object model topology structure based on the target natural resource archive set, and to generate corresponding sub-document sets according to the directory hierarchy rules; The execution module is used to perform deep semantic understanding and structured information extraction on the sub-document collection through the preset natural resources domain model, and output structured metadata describing the core elements of the archive; A construction module is used to identify business entities and their semantic associations in the natural resources field based on structured metadata, and construct RDF triples with attribute constraints; wherein the identified business entities include land parcel entities with vector boundary coordinate point sequences; The denoising module is used to perform geometric feature-preserving coordinate simplification on the vector boundary coordinate point sequence of the land entity through the Douglas-Peucker compression algorithm to generate a set of key coordinate points after denoising; The alignment module is used to perform vector geometric overlay operations on the polygonal boundaries of two land entities based on the denoised key coordinate point set. The overlapping area ratio is calculated based on the intersection-and-union ratio. When the intersection-and-union ratio exceeds a preset threshold, the entity space is aligned to obtain a set of spatially aligned triplets. The fusion module is used to fuse the spatially aligned triple sets and construct a knowledge graph in the field of natural resources.

[0125] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A natural resource archive retrieval and utilization method combining large models and knowledge graph technology, characterized in that: The method comprises: Step 1: Obtain a target natural resource archive set, wherein the archive set includes at least one natural resource archive to be processed; Step 2: Based on the target natural resource archive set obtained in step 1, the document object model topology structure is parsed and the corresponding sub-document set is generated according to the directory hierarchy rules; Step 3: Using the pre-set natural resources domain model, perform deep semantic understanding and structured information extraction on the sub-document set generated in step 2, and output structured metadata describing the core elements of the archive; Step 4: Based on the structured metadata output from step 3, identify business entities in the natural resources field and their semantic associations, and construct RDF triples with attribute constraints; the identified business entities include land parcel entities with vector boundary coordinate point sequences; Step 5: Perform geometric feature-preserving coordinate simplification on the vector boundary coordinate point sequence of the land entity extracted in step 4 using the Douglas-Peucker compression algorithm to generate a set of key coordinate points after noise reduction. Step 6: Based on the denoised key coordinate point set from step 5, perform vector geometric overlay operations on the polygonal boundaries of the two land parcel entities. Calculate the area ratio of the overlapping region based on the intersection-and-union ratio. When the intersection-and-union ratio exceeds a preset threshold, perform spatial alignment of the entities to obtain a set of spatially aligned triplets. Step 7: Integrate the spatially aligned triplet set completed in step 6 to construct a knowledge graph in the field of natural resources.

2. The natural resource archive retrieval and utilization method combining a large model and knowledge graph technology according to claim 1 is characterized in that: The step 2 includes: Step 21: Based on the natural resource archive file obtained in step 1, parse and generate a Document Object Model (DOM) tree structure; recursively calculate the depth value of each node using the DOM tree as the processing object to generate a DOM tree with depth annotations, wherein the root node depth is initialized to 0, and the child node depth value = the parent node depth value + 1; Step 22, using the depth-annotated DOM tree output from step 21 as input: Filter the node set with a depth value ≥ 2, and execute the following for each filtered node: Extract the text blocks of all leaf nodes in the node subtree, arrange the text blocks in the original character offset order to form an ordered sequence, and calculate the distance between adjacent text blocks in the sequence, where the distance value = the starting offset of the subsequent text block - the ending offset of the previous text block; Record the maximum spacing value of the current node. When the maximum spacing value is less than the preset threshold, merge all text blocks in the sequence to generate sub-documents, and output the sub-document set and the corresponding DOM node path and maximum spacing value. Step 23, taking the sub-document set outputted in step 22 as the processing object, assigning a globally unique identifier to each sub-document, associating the DOM node path and the maximum spacing value outputted in step 22, and binding the source file number to generate structured traceability metadata.

3. The natural resource archive retrieval and utilization method combining a large model and knowledge graph technology according to claim 2 is characterized in that: The step 3 includes: Step 31: Using the structured provenance metadata generated in step 23 as input, the associated sub-documents are input into a pre-set natural resources domain macro-model. Through the domain adaptation decoding layer of the macro-model, three core entities, namely, plot number, ownership entity, and spatial use category, are extracted to identify the ownership relationship between "ownership entity and plot" and the functional relationship between "plot and spatial use." Step 32: Based on the entities and relationships output in step 31, extract the three attributes of area value, coordinate system type, and issuance date for each land parcel entity; convert the issuance date into a timestamp format, and together with the coordinate system type, form a space-time dual constraint; Step 33: Standardize and encapsulate the structured features outputted in step 32, and output a structured metadata set including entities, relationships, and constraints.

4. The natural resource archive retrieval and utilization method combining a large model and knowledge graph technology according to claim 3 is characterized in that: The step 4 comprises: Step 41: Using the structured metadata set output from step 3 as input, parse the entity identifier, entity type, and attribute set in the entity dimension to identify the core business entity types in the natural resources field. The land parcel entity type is identified as a key entity containing a sequence of vector boundary coordinate points. The ownership entity type is also identified. Step 42: Taking the business entity type identified in step 41 as the processing object, for each land parcel entity, extract a vector boundary coordinate point sequence from the structured metadata set as the core spatial attribute of the land parcel entity, and add the coordinate point sequence to the attribute set of the land parcel entity to form an updated land parcel entity attribute set; Step 43: Based on the updated land parcel entity attribute set in step 42 and the relationship dimension and constraint dimension output in step 3, a set of RDF triples with attribute constraints is constructed, where: Each attribute value in the entity attribute set is converted into a triple with the entity identifier as the subject, the attribute name as the predicate, and the attribute value as the object; Each relation in the relational dimension is converted into a triple with the subject entity identifier as the subject, the relation type as the predicate, and the object entity identifier as the object; The spatial-temporal dual constraints in the constraint dimension are bound to the corresponding plot entity identifier, generating a triple with the plot entity identifier as the subject, the constraint type as the predicate, and the constraint value as the object.

5. The natural resource archive retrieval and utilization method combining large model and knowledge graph technology according to claim 4 is characterized in that: The step 5 comprises: Step 51: Using the RDF triple set constructed in step 4 as input, process the vector boundary coordinate point sequence corresponding to each land parcel entity; initialize the first and last points of the coordinate point sequence as key points to form an initial key point set; Step 52: Based on the initial key point set determined in step 51, recursively partition the line segment formed by the first and last points, that is, traverse all non-key coordinate points between the first and last points of the line segment, calculate the vertical distance of each point to the line segment, and identify the point with the largest vertical distance; Step 53: When the maximum vertical distance value identified in step 52 exceeds the preset distance threshold, the corresponding point is marked as a new key point and added to the key point set; then, the original line segment is split into two sub-segments, front and back, with the new key point as the boundary, and the maximum vertical distance calculation and key point determination process described in step 52 is recursively performed on each sub-segment; In step 54 , after the recursive process is terminated, the key points generated by all levels of judgment are merged and arranged in the order of the original coordinate sequence to generate a denoised key coordinate point set as a simplified boundary representation of the land entity.

6. The natural resource archive retrieval and utilization method combining large model and knowledge graph technology according to claim 5 is characterized in that: The step 6 comprises: Step 61: Using the denoised key coordinate point set of each land parcel entity output in step 5 as input, two land parcel entities to be subjected to spatial relationship analysis are selected to form an entity pair; Step 62: for the entity pairs selected in step 61, reconstruct the corresponding first polygon boundary and second polygon boundary using their respective key coordinate point sets; Step 63, performing a vector geometric overlay operation on the first polygon boundary and the second polygon boundary reconstructed in step 62, calculating an overlapping polygon formed by the intersection of the two polygons, and respectively calculating the area value of the overlapping polygon, the area value of the first polygon boundary, and the area value of the second polygon boundary; Step 64: Calculate an intersection-over-union ratio based on the area value calculated in step 63, where the intersection-over-union ratio = overlapping area / (area of ​​the first polygon + area of ​​the second polygon - area of ​​the overlapping area); Step 65, comparing the IoU value calculated in step 64 with a preset IoU threshold; Step 66: When the IoU value exceeds a preset IoU threshold, it is determined that the two land parcel entities in the entity pair have a spatial overlap relationship, and an entity spatial alignment operation is performed, i.e., a semantic triple representing the spatial alignment relationship is generated, where the subject entity is the first land parcel entity identifier, the predicate is the preset spatial alignment relationship type, and the object entity is the second land parcel entity identifier. In step 67 , all semantic triples generated in step 66 that meet the intersection-over-union ratio threshold condition and complete the spatial alignment operation are collected to form a spatially aligned triple set.

7. The natural resource archive retrieval and utilization method combining large model and knowledge graph technology according to claim 6 is characterized in that: The step 63 includes: Step 631: Based on the key coordinate points of the first polygonal boundary reconstructed in step 62, traverse all key coordinate points of the first polygonal boundary, identify a point set located on the outermost boundary, connect the point set on the outermost boundary to form a minimum convex polygon outline surrounding all internal points, and generate a first candidate convex hull; Step 632: Based on the key coordinate point set of the second polygonal boundary reconstructed in step 62, traverse all key coordinate points of the second polygonal boundary, identify the point set located on the outermost boundary, connect the point set on the outermost boundary to form a minimum convex polygon outline that encloses all internal points, and generate a second candidate convex hull. Step 633: Calculate the area of ​​the first candidate convex hull, calculate the original area of ​​the first polygon boundary key coordinate point set, calculate the absolute difference between the area of ​​the first candidate convex hull and the original area, and calculate the area error rate. If the area error rate is less than a preset convex hull simplification tolerance threshold, the first candidate convex hull is determined to be valid. Otherwise, the first polygon boundary key coordinate point set is used as the first candidate convex hull, and the same validity verification process is performed on the second candidate convex hull, outputting the final valid first and second convex hull boundaries. Step 634 , calculating the final valid overlapping area polygon formed by the intersection of the first convex hull boundary and the second convex hull boundary; In step 635 , the area value of the overlapping region polygon obtained in step 634 , the area value of the first convex hull boundary output in step 633 , and the area value of the second convex hull boundary are calculated respectively.

8. The natural resource archive retrieval and utilization method combining large model and knowledge graph technology according to claim 7 is characterized in that: The step 7 comprises: Step 71: using the spatially aligned triple set output in step 6 as a first input and the RDF triple set with attribute constraints constructed in step 4 as a second input; Step 72: Perform triple fusion processing on the first input and the second input provided in step 71, that is, identify triples describing the same entity identifier in the two input sets and merge their attributes and relationship information; for the spatially aligned relationship triples generated in step 6, incorporate them as new relationships into the fused triple set; Step 73: Using the triplet set fused in step 72 as the basis for graph construction, each unique entity identifier in the triplet is mapped to a node in the graph database, and each entity attribute is mapped to an attribute key-value pair of the corresponding node; Step 74: Based on the nodes created in step 73, the semantic relationships defined in the triples and the spatial alignment relationships added in step 72 are mapped into directed edges in the graph database, i.e., the node corresponding to the subject entity identifier is the edge start point, the relationship type is the edge label, and the node corresponding to the object entity identifier is the edge end point. Step 75: For the entity node of the land parcel, the key coordinate point set after noise reduction in step 5 is attached to the corresponding node as a spatial geometric attribute; Step 76, after completing the node, edge and attribute mapping of steps 73 to 75, instantiate and generate a knowledge graph in the field of natural resources in the graph database. The graph contains a semantic network structure composed of plot entities, ownership entity nodes, and directed edges composed of ownership relationships, functional relationships and spatial alignment relationships.

9. The natural resource archive retrieval and utilization method combining large model and knowledge graph technology according to claim 3 is characterized in that: The structured metadata collection includes: Entity dimension: Generates a triple of {entity type, entity identifier, attribute set}; Relationship dimension: Generate {subject entity identifier, relationship type, object entity identifier} triples; Constraint dimension: Bind the spatial-temporal dual constraints to the corresponding land entity identifiers.

10. A natural resource archive retrieval and utilization system combining large models and knowledge graph technology, characterized in that: The system is used to perform the method according to any one of claims 1 to 9, comprising: An acquisition module, configured to acquire a target natural resource archive set, wherein the archive set includes at least one natural resource archive to be processed; A parsing module is used to parse the document object model topology structure based on the target natural resource archive set, and to generate corresponding sub-document sets according to the directory hierarchy rules; The execution module is used to perform deep semantic understanding and structured information extraction on the sub-document collection through the preset natural resources domain model, and output structured metadata describing the core elements of the archive; A construction module is used to identify business entities and their semantic associations in the natural resources field based on structured metadata, and construct RDF triples with attribute constraints; wherein the identified business entities include land parcel entities with vector boundary coordinate point sequences; The denoising module is used to perform geometric feature-preserving coordinate simplification on the vector boundary coordinate point sequence of the land entity through the Douglas-Peucker compression algorithm to generate a set of key coordinate points after denoising; The alignment module is used to perform vector geometric overlay operations on the polygonal boundaries of two land entities based on the denoised key coordinate point set. The overlapping area ratio is calculated based on the intersection-and-union ratio. When the intersection-and-union ratio exceeds a preset threshold, the entity space is aligned to obtain a set of spatially aligned triplets. The fusion module is used to fuse the spatially aligned triple sets and construct a knowledge graph in the field of natural resources.

Images