A method and device for intelligent determination of land use change patches

Abstract

This invention discloses an intelligent method and apparatus for determining land use change patches, relating to the fields of natural resource management and artificial intelligence. Based on a user-input query, information about the patch to be determined, and a preset retrieval strategy, it retrieves target determination rules from a target policy and regulation rule base. This target policy and regulation rule base is constructed by extracting structured rules and building a multi-dimensional index from original policy and regulation documents used for determining land use change patches. Based on the query, the information about the patch to be determined, the target determination rules, and a preset determination prompt word template, target determination prompt words are generated. These prompt words are then input into a large language model to obtain the target determination result for the information about the patch to be determined. This invention improves the accuracy and recall of rule retrieval by constructing a target policy and regulation rule base instead of original document retrieval, avoids the problem of incomplete evidence caused by scattered determination rules, and enhances the accuracy of the determination.

Description

Technical Field

[0001] This invention belongs to the fields of natural resource management and artificial intelligence technology, and particularly relates to an intelligent determination method and device for land use change patches. Background Technology

[0002] With the rapid development of artificial intelligence technology, large language models are being applied more and more widely in professional fields. Retrieval-Augmented Generation (RAG) technology has become the mainstream technical approach for knowledge-intensive tasks. By combining external knowledge bases with large language models, it effectively solves the problems of lagging model knowledge updates and insufficient professional domain knowledge.

[0003] In the field of natural resource management, determining the illegality of land use changes is a crucial operational task. This work requires the comprehensive application of multiple laws, regulations, and policy documents, including the "Land Administration Law of the People's Republic of China" and the "Procedures for Investigating and Handling Illegal Acts Related to Natural Resources," to assess the compliance of land use change plots. Due to the numerous legal provisions and complex judgment logic involved, traditional manual judgment methods are inefficient.

[0004] Currently, existing methods typically employ Search Augmentation Generation (RAG) technology. However, due to the significant semantic gap between user questions and legal rules, directly using user questions for retrieval often fails to effectively match relevant judgment rules, resulting in poor retrieval results. Furthermore, since determining the illegality of map features requires the integration of multiple policy documents and legal provisions, the relevant judgment criteria are scattered across different documents. Traditional document segmentation methods may disrupt the integrity and logical coherence of the rules, leading to fragmented search results and an inability to obtain complete judgment criteria. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, the present invention provides an intelligent determination method and apparatus for land use change patches.

[0006] The technical problem to be solved by this invention is achieved through the following technical solution: In a first aspect, the present invention provides an intelligent determination method for land use change patches, comprising: Obtain the user's query question and the information of the image patches to be judged; Based on the query question, the information of the land use change plots to be determined, and the preset search strategy, target determination rules are obtained from a pre-built target policy and regulation rule library. The target policy and regulation rule library is constructed by extracting structured rules and using multi-dimensional indexes from the original policy and regulation documents used for determining land use change plots. The target determination rules are used to determine whether the information of the land use change plots to be determined conforms to the policies or regulations in the original policy and regulation documents. Based on the query question, the information of the map features to be judged, the target judgment rules, and the preset judgment prompt word template, generate target judgment prompt words; Input the target determination prompts into the large language model to obtain the target determination results for the image patch information to be determined.

[0007] Secondly, the present invention provides an intelligent determination device for land use change patches, comprising: The acquisition module is used to acquire the user's query question and the information of the patch to be judged; The judgment rule retrieval module is used to retrieve target judgment rules from a pre-built target policy and regulation rule library based on the query question, the information of the land use change plot, and the preset retrieval strategy. The target policy and regulation rule library is constructed by extracting structured rules and using multi-dimensional indexes from the original policy and regulation documents used for land use change plot judgment. The target judgment rules are used to determine whether the information of the land use change plot conforms to the policies or regulations in the original policy and regulation documents. The prompt word generation module is used to generate target determination prompt words based on the query question, the information of the patch to be determined, the target determination rules, and the preset determination prompt word template; The judgment result output module is used to input the target judgment prompt words into the large language model to obtain the target judgment result for the information of the patch to be judged.

[0008] This invention provides an intelligent method and apparatus for determining land use change patches. It replaces traditional document retrieval by constructing a dedicated target policy and regulation rule base. The policy and regulation content is extracted using structured rules, resulting in clear semantics and well-defined boundaries. This solves the retrieval failure problem caused by the semantic mismatch between user questions and rule content in traditional RAG technology, significantly improving the accuracy and recall of rule retrieval. Through unified management of the target policy and regulation rule base, determination rules scattered across multiple policy and regulation documents are centrally stored, resolving the incomplete basis problem caused by scattered determination rules and ensuring that all necessary rules can be obtained during patch determination. By querying the question, information on the patch to be determined, target determination rules, and preset determination prompt word templates, target determination prompt words are generated, enabling the large language model to better understand and apply the rules for reasoning and determination. This solves the problem of insufficient context utilization in traditional RAG solutions and improves the accuracy of determination.

[0009] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating an intelligent determination method for land use change patches provided in an embodiment of the present invention. Figure 2This is a schematic diagram of the processing procedure of an intelligent determination method for land use change patches provided in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the construction process of the target policy and regulation rule base provided in an embodiment of the present invention; Figure 4 This is a structural block diagram of an intelligent determination device for land use change patches provided in an embodiment of the present invention. Detailed Implementation

[0011] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0012] This invention provides an intelligent method for determining land use change patches. See also... Figure 1 and Figure 2 The method includes the following steps: S10. Obtain the query question input by the user and the information of the plot to be judged.

[0013] For example, users can input query questions using natural language, such as "Is the current land parcel illegal?" or "What type of illegality does this land parcel belong to?" The land parcel information to be judged consists of the attribute information of the land parcel to be judged, which may include land type, location information, area, change time, change approval status, etc.

[0014] S20. Based on the query question, the information of the map features to be determined, and the preset retrieval strategy, obtain the target determination rules from the pre-built target policy and regulation rule library.

[0015] The target policy and regulation rule base is constructed by extracting structured rules and building multi-dimensional indexes from the original policy and regulation documents used for land use change patch determination; the target determination rules are used to determine whether the information of the patch to be determined conforms to the policies or regulations in the original policy and regulation documents.

[0016] Optionally, the process of constructing the target policy and regulatory rule base includes: A1. Preprocess the original policy and regulatory documents to obtain preprocessed policy and regulatory documents.

[0017] For example, refer to Figure 3The original policy and regulatory documents can be in formats including, but not limited to, PDF files, Word documents, and text files. The structural characteristics of these documents are: each law, regulation, or policy is typically described as an independent clause with a hierarchical numbering structure, such as "Chapter 1, Article 1, Paragraph 1, Item 1 of the Natural Resources Law". First, the document format is preprocessed: PDF files are parsed to extract text content, Word documents are formatted, and text files undergo encoding standardization to obtain the initial preprocessed text. Next, document structure recognition is performed to identify the hierarchical relationship between chapters, sections, articles, paragraphs, and items in the initial preprocessed text. Finally, text encoding is unified by converting all initial preprocessed text to UTF-8 encoding format to obtain the preprocessed policy and regulatory document.

[0018] Specifically, let the original set of policy and regulatory documents be... ,in, This represents the maximum number of documents indexed, or the total number of documents. For any original policy or regulatory document... , The entire preprocessing process is defined as three consecutive transformation operations: , where format conversion functions Used to uniformly convert PDF, Word, and other formats into plain text, i.e., initial preprocessing of text, structure recognition function. Used to identify and label hierarchical structures, encoding unified functions. Used to output UTF-8 encoded text, i.e., preprocessed policy and regulatory documents. The set of preprocessed policy and regulatory documents is represented as... , Indicates the first A preprocessing policy and regulatory document.

[0019] A2. Input the preset rule-extracted prompt words and preprocessed policy and regulatory documents into the large language model for structured rule extraction to obtain multiple judgment rule information.

[0020] For example, preprocessed policy and regulatory documents are submitted to a large language model (e.g., Qwen3-235B-A22B-instruct-2507). Here, the preprocessed policy and regulatory documents need to be processed in batches. Long documents are divided into batches according to a predetermined length, and the content of each batch does not exceed the context window limit of the large language model. Preset rules are used to extract prompt words to guide the large language model in fully extracting all policy or regulatory rule information, resulting in multiple decision rule information.

[0021] Specifically, for the first Preprocessing policy and regulatory documents Rule extraction function of large language model Used to extract prompt words according to preset rules Extract the set of judgment rules information ,in To process policy and regulatory documents The maximum amount of judgment rule information extracted from it. To process policy and regulatory documents The extracted first Information on each decision rule. Information on each decision rule. Represented as structured tuples The fields correspond to the rule's unique identifier, source document, chapter information, clause number, rule content, set of applicable conditions, set of associated rules, effective date, and version number, respectively. Ultimately, the set of judgment rule information extracted from all pre-processed policy and regulatory documents is... ,in To determine the total number of rule information.

[0022] The structured format of the decision rule information may include the following fields: ① Unique rule identifier (rule_id): A unique number automatically generated by the system.

[0023] ② Source file: The name of the file from which the rule originates.

[0024] ③ Chapter Information: The chapter number to which the rule belongs.

[0025] ④ Article number: The article number of the rule.

[0026] ⑤ Rule content: The complete text content of the rule.

[0027] ⑥ Conditions: A list of prerequisites for the rule to apply.

[0028] ⑦ Related rules: A list of other rule identifiers that have a referencing relationship with this rule.

[0029] ⑧ Effective Date (effective_date): The effective date of the rule.

[0030] ⑨ Version number: The version number of the rule.

[0031] Here are some examples of prompt words extracted using preset rules: Role: You are a professional legal and compliance document analysis expert. Your core competency lies in accurately identifying, understanding, and extracting policies, regulations, and business rules from texts in a structured manner. You possess exceptional rigor, able to distinguish between "rule clauses" and "background descriptions," ensuring the completeness and accuracy of the extracted content.

[0032] Goal: Your task is to read the given preprocessed policy and regulatory document fragments and extract all the policy or regulatory rules information from them. Before outputting the final structured results, you must first conduct detailed, step-by-step analysis and reasoning.

[0033] Workflow: Please strictly follow these steps in your thinking and execution: 1. Preliminary scanning and localization: Read through the entire [preprocessed policy and regulatory document fragment] and identify all paragraphs or sentences that may contain rules, clauses, definitions, requirements, or procedures.

[0034] The initial assessment determines whether each identified element constitutes an independent and complete rule.

[0035] 2. Analyze and extract key points one by one: For each potential rule identified in the first step, conduct in-depth analysis: What is the core content? Extract its most essential constraints or guiding statements.

[0036] What are the applicable conditions? Identify the prerequisites that trigger or limit the rule (such as "when...", "if...", "applies to...").

[0037] Is there a reference? Check if other clauses or provisions are explicitly referenced (such as "refer to Article X" or "complies with provision Y").

[0038] What is metadata? It determines the chapter, clause number, effective date, and version number (if included in the fragment).

[0039] 3. Structure and Formatting: Organize the information extracted from the second step strictly according to the JSON format specified below.

[0040] Generate a temporary unique identifier (rule_id) for each rule.

[0041] Ensure all fields are filled. For information not found in the fragment, fill with an empty string or "unmarked".

[0042] 4. Final review: Check that the generated JSON conforms to the format requirements, ensuring that no rules are omitted and that the content is accurate.

[0043] The output confirms that it is only the final JSON object and does not contain any text related to the analysis process.

[0044] Output Format: The final output must be a standard JSON object, without any Markdown tags (such as json ...), and without any analysis process, introduction, or conclusion.

[0045] The JSON object contains a root field called rules, whose value is a list, where each element represents a rule to be extracted.

[0046] Field description: rule_id: Generates a temporary unique identifier in the format TEMP_ + source file name hash prefix + _ + clause number (or the serial number if there is no clause number).

[0047] source: Enter the source file name provided by the user.

[0048] chapter: The chapter number or title to which the rule belongs (leave an empty string if not mentioned).

[0049] article: The clause number of the rule (such as "Article 5", "3.1").

[0050] content: The complete original text of the rule, without any deletions.

[0051] conditions: A list of strings that lists all applicable conditions.

[0052] related_rules: A list of strings that lists the rule numbers that are being referenced.

[0053] effective_date: Date string or "unmarked".

[0054] version: Version string or "unspecified".

[0055] Example output structure: { "rules": [ { "rule_id": "TEMP_PolicyA_Art5", "source": "Enterprise Data Security Management Regulations.pdf", "chapter": "Chapter Three", "article": "Article 15", "content": "All core business data must be encrypted during transmission." "conditions": ["Core business data", "During transmission"], "related_rules": ["Article 10"], "effective_date": "2024-01-01", "version": "V2.0" } ] } The user input is as follows: [Source file name] {{source_file_name}} [Excerpt from preprocessing policy and regulatory documents] {{content}} A3. Based on multiple judgment rule information, vectorized index, keyword index and knowledge graph index are established respectively to obtain the target policy and regulation rule base.

[0056] For example, a multi-dimensional index is established for each extracted judgment rule information and stored in the target policy and regulation rule base. The types of indexes established include: 1) Vectorized index: The judgment rule information is converted into a high-dimensional vector representation through an embedding model, and the high-dimensional vector representation is stored in a vector index library, supporting rule retrieval based on semantic similarity; 2) Keyword index: An inverted index is established for the judgment rule information and stored in a full-text search engine, supporting keyword exact matching and fuzzy matching retrieval; 3) Knowledge graph index: A graph of relationships between judgment rule information is constructed, establishing a triple relationship of "rule-applicable conditions-legal consequences," stored in a graph database, supporting graph-based reasoning queries.

[0057] Specifically, the metadata of the decision rule information is stored in a relational database and associated with various indexes. For the set of decision rule information... Vectorized indexes are embedded in the model. The first Convert the judgment rule information into dimensional regular vector ,in Ultimately, this forms the vector index library. ,in, express The Middle Information on the judgment rules, The rule represents a unique identifier; the keyword index is constructed using the inverted index building function. By establishing a mapping relationship between keywords and decision rules, the constructed full-text search engine is represented as follows: Knowledge graph index is defined as , where the set of nodes It includes nodes for determining rules, applicable conditions, and legal consequences. This represents the node containing the decision rule information. Indicates the applicable condition node, Nodes representing legal consequences, edge sets This indicates the relationships between decision rule information, the relationships between decision rule information and applicable conditions, and the relationships between decision rule information and legal consequences. Ultimately, the complete index structure of the target policy and regulatory rule base is as follows: ,in A relational database for storing metadata about decision rules.

[0058] Optionally, the method in this embodiment further includes: obtaining newly released original policy and regulatory documents and extracting structured rules to obtain judgment rule information to be updated; determining the update type of the judgment rule information to be updated through a large language model or a preset update type determination strategy; wherein the update type of the judgment rule information to be updated includes addition, revision, supplementation and repeal; and updating the judgment rule information to be updated to the target policy and regulatory rule library according to the update type of the judgment rule information to be updated.

[0059] For example, the system can receive newly published original policy and regulatory documents uploaded by users and extract structured rules using a large language model to obtain the judgment rule information to be updated; alternatively, users can directly input the judgment rule information to be updated, and this embodiment does not impose any restrictions on this. Then, a strategy can be determined using a large language model or a preset update type to compare the judgment rule information to be updated with the judgment rule information in the target policy and regulatory rule base to determine the update type of the judgment rule information to be updated. Specifically, the large language model can analyze prompt words based on the input preset update type and compare the judgment rule information to be updated with the judgment rule information in the target policy and regulatory rule base to output the update type of the judgment rule information to be updated.

[0060] Optionally, a preset update type determination strategy is used to determine the update type of the judgment rule information to be updated, including: matching the judgment rule information to be updated with the number of each judgment rule information in the target policy and regulation rule base; if the matching result is empty, the update type of the judgment rule information to be updated is determined to be addition; if the matching is successful, the successfully matched judgment rule information is determined to be matched judgment rule information, and the semantic similarity between the judgment rule information to be updated and the matched judgment rule information is calculated to obtain the matching similarity; if the matching similarity is less than a preset similarity threshold, the update type of the judgment rule information to be updated is determined to be revision; if the matching similarity is greater than or equal to the preset similarity threshold, the update type of the judgment rule information to be updated is determined to be supplement; if the latest version status of the matched judgment rule information corresponding to the judgment rule information to be updated is invalid, the update type of the judgment rule information to be updated is determined to be obsolescence.

[0061] For example, suppose the set of judgment rule information already existing in the target policy and regulation rule base is represented as follows: , express Maximum index in, superscript The existing set of decision rule information to be updated can be represented as follows: , express Maximum index in, superscript Indicates whether it is pending update or has been updated. Rule comparison function. For the Information on judgment rules to be updated To make a judgment, The update type for obtaining the judgment rule information to be updated can be represented as: The determination is based on a matching function using the file number (source document + chapter information + clause number). Search for rules with the same terms; if the matching result is empty, that is... If the match is successful, the update type will be Add (ADD); ,in Then the determination rule information for successful matching is the matching determination rule information. It also calculates the semantic similarity between the decision rule information to be updated and the matching decision rule information. ,like The update type is then a revision; if The update type is then supplementary. A preset similarity threshold is set. If the latest version status (STATUS) of the matching rule information corresponding to the rule information to be updated is invalid (INACTIVE), then... The update type is then REPEAL.

[0062] In addition, this embodiment can also utilize a large language model to match the matching rules based on preset conflict detection prompts. Perform a conflict check. If a conflict is found, prompt the user to resolve the conflict.

[0063] The following are examples of preset conflict detection prompts: Role: You are an expert in analyzing conflicts in land use policies, regulations, and rules. You are proficient in the logic of map patch determination rules and can accurately compare two rules to identify conflicts and clarify the basis for the analysis.

[0064] Goal: Compare the rule to be updated (RJNEW) with the matching rule (RmOLD), check for conflicts, and return the check results briefly in natural language without redundant content.

[0065] Workflow: 1. Rule parsing: Extract the core elements of two rules (rule content, applicable conditions, related rules, effective date, and version number).

[0066] 2. Core Comparison: Focus on checking four types of conflicts: content conflict (contradictory judgment criteria in the same scenario), applicable condition conflict (overlapping conditions and inconsistent logic), related rule conflict (related rules or core content of the other party conflict), and timeliness conflict (overlapping effective dates and inconsistent content).

[0067] 3. Conflict determination: Determine whether a conflict exists, and mark the type of conflict and the specific points of conflict; if there is no conflict, explain the reasons.

[0068] 4. Output Results: Return results concisely in natural language, including: conflict check results (conflict present / no conflict), conflict type (omit if none), brief conflict points (omit if none), judgment criteria, and processing prompts. The language should be concise and without redundancy.

[0069] The core elements of rule information (for reference) for both RJNEW and RmOLD include: rule content, applicable conditions, associated rules, effective date, and version number.

[0070] Output Format (must be a concise natural language description): Briefly describe the conflict check results. If a conflict exists, mention the conflict type, the core conflict point, and the basis for judgment, along with additional handling suggestions. If no conflict exists, explain the reasons and provide handling suggestions. Keep it concise and avoid redundancy.

[0071] Notes: 1. Judgment is based solely on input rules, without adding additional assumptions or omitting potential conflicts; 2. If the applicable conditions do not overlap, even if the content is different, it is considered conflict-free; 3. New and old versions must be judged in conjunction with the effective date. If the old version has been abolished, there is no conflict; 4. The expression is precise and concise, meeting the requirements for brief returns in natural language.

[0072] Optionally, based on the update type of the judgment rule information to be updated, the judgment rule information to be updated is updated to the target policy and regulation rule base, including: if the update type of the judgment rule information to be updated is addition, a vectorized index, a keyword index, and a knowledge graph index are created for the judgment rule information to be updated, and added to the target policy and regulation rule base; if the update type of the judgment rule information to be updated is revision or supplementation, the rule content contained in the vectorized index, keyword index, and knowledge graph index in the target policy and regulation rule base is updated based on the judgment rule information to be updated; if the update type of the judgment rule information to be updated is repeal, the status of the vectorized index, keyword index, and knowledge graph index corresponding to the judgment rule information to be updated in the target policy and regulation rule base is marked as invalid.

[0073] For example, when updating the judgment rule information to be updated to the target policy and regulation rule base, a vectorized index, keyword index, and knowledge graph index are created for the judgment rule information to be updated with the update type of "new"; the corresponding index content is updated for the judgment rule information to be updated with the update type of "revision" or "supplement"; and the judgment rule information to be updated with the update type of "obsolete" is marked as invalid in the index, retained in historical records but not participating in regular searches. In addition, incremental update or full update mode can be selected according to the number of judgment rule information to be updated.

[0074] Specifically, the set of judgment rule information to be updated Divided into three subsets according to update type ,in, This indicates that the update type is newly added judgment rule information to be updated. This indicates update-type or supplementary judgment rule information to be updated. This indicates update rules for rules to be updated if the update type is obsolete. The incremental index update mode is implemented through functions, specifically including: vectorized index update representation as... ,in, , , They represent respectively The corresponding rule unique identifier, rule vector, and rule content. This represents the existing vector index library. Represents the embedded model; keyword index update is represented as , This indicates an existing keyword index. This represents the inverted index update function, which encapsulates the logic for adding, deleting, and modifying keywords in the index; the knowledge graph index update is represented as... , This indicates the updated node. Represents the updated edge, where: , ,in, This indicates existing nodes. Indicates the node to be deleted. This indicates a newly added node. Indicates existing edges. Indicates the edge to be deleted. This represents a newly added edge. This represents the difference set.

[0075] The index update mode can be selected based on the change ratio: when Incremental update mode is used when the condition is met; otherwise, full update mode is used. The preset reconstruction threshold is (usually set to 0.3). This indicates the total amount of judgment rule information contained in the existing target policy, regulation and rule base.

[0076] Optionally, the method in this embodiment further includes: determining the version change information and current version status of each judgment rule information according to the update type of each judgment rule information in the target policy and regulation rule base; the version change information includes the creation time, modification time, and description of the modification content; the current version status is effective, repealed, or pending effectiveness.

[0077] For example, this embodiment can perform version management on each judgment rule in the target policy and regulation rule base. For example, 1) Version history chain maintenance: Each judgment rule maintains a complete version history; 2) Change information recording: Records version change information of the judgment rule, including creation time, modification time, and description of modification content; 3) Version status management: Marks the current version status of the judgment rule as "effective", "repealed", or "pending effectiveness"; 4) Version backtracking support: Supports querying the historical version information of the judgment rule. The version information (including version change information and current version status) structure may include: ① Rule identifier (rule_id); ② Current version number; ③ Rule status; ④ Effective date (effective_date); ⑤ Expiry date (expiry_date, if applicable); ⑥ Change log (change_log): Records the version number, date, and change description for each version change.

[0078] Specifically, regarding the judgment rule information Its version sequence is represented as a time-ordered historical chain. ,in For version number, For timestamps, For explanation of the changes, This represents the total number of historical versions. The current version status is determined by the conversion function. Based on the change type of each judgment rule information Confirmed: When At that time, the current version status is active. At that time, the current version status is INACTIVE, and the effective date will be [date missing]. At this time, the current version status is PENDING. Indicates the current time. Version numbers are incremented. Update, in which The major or minor version number is incremented based on the change type. For example, if the initial version of each decision rule is 1.0, and the change type is REVISE, then the minor version number is incremented, such as 1.1; if the change type is SUPPLEMENT, then the major version number is incremented, such as 2.0.

[0079] Optionally, the method in this embodiment further includes: receiving a rule deletion instruction input by a user; parsing the rule deletion instruction using a large language model, and retrieving and displaying at least one candidate judgment rule information from the target policy and regulation rule base; receiving a selection instruction from the user for at least one candidate judgment rule information, and determining the judgment rule information to be deleted; deleting the judgment rule information to be deleted, and deleting the vectorized index, keyword index, and knowledge graph index corresponding to the judgment rule information to be deleted in the target policy and regulation rule base.

[0080] For example, a rule deletion instruction can be a user describing the characteristics of the decision rule information to be deleted in natural language. Alternatively, the user can also enter a unique identifier or number of the decision rule information to directly search for and delete it.

[0081] Specifically, let's say the user inputs a rule deletion command. The large language model parses the rule deletion instructions and obtains candidate decision rule information through similarity matching. , can be represented as ,in, To locate the similarity threshold, express The corresponding rule vector. Additionally, the deletion operation function. Supports two modes: setting during logical deletion. and ,in, This indicates the information of the rules for determining whether to delete. The status has been changed to deleted. This indicates the information of the rules for determining whether to delete. The deletion time record is the current time; physical deletion is performed. During deletion, a function can be used to detect association rules. Detect the reference relationships of the rules for determining deletion, where... express Any of the judgment rule information, express The set of judgment rule information associated with or referenced, if Then prompt the user that there is a problem. The judgment rule information references the judgment rule information to be deleted.

[0082] After deleting the decision rule information to be deleted, the vectorized index, keyword index, and knowledge graph index corresponding to the decision rule information to be deleted in the target policy and regulation rule base are deleted. Specifically, let the set of decision rule information to be deleted be... Index cleanup and updates are performed via functions. Implementation, where vectorized index cleanup and update can be represented as Keyword index cleaning and updating can be represented as Knowledge graph index cleaning and updating can be represented as ,in, , They represent The rule's unique identifier and rule vector. This indicates the node corresponding to the deletion determination rule information. express , Indicates the starting node of the edge. This indicates the end node of the edge. The deletion log is also updated. Record the current time Operator and deletion mode .

[0083] In one optional embodiment, the preset retrieval strategy in step S20 may include: S201. Determine whether the number of judgment rule information in the target policy and regulation rule library whose current version status is effective does not exceed a preset threshold.

[0084] For example, the preset threshold is determined based on the capacity of the large language model context.

[0085] S202. If so, then all judgment rule information that is currently in effect in the current version shall be determined as the target judgment rule.

[0086] For example, if the number of decision rules in the target policy and regulation rule base that are currently in effect is small, they can be fully incorporated into the context of the large language model. In this case, all decision rules that are currently in effect can be read and used as the basis for subsequent decisions. The advantage of this approach is that the decision basis is complete and no applicable rules are missed.

[0087] S203. If not, then based on the query question and the information of the map feature to be determined, retrieve the rules from the pre-built target policy and regulation rule library to obtain the target determination rules.

[0088] For example, if the number of judgment rules that are currently in effect in the target policy and regulation rule base is large, then it is necessary to search for rules in the pre-built target policy and regulation rule base to obtain the target judgment rules.

[0089] Optionally, step S203 may specifically include: S2031. Analyze the query question and the information of the map features to be judged to obtain key information; S2032. Convert the key information into a key information vector, and retrieve the first judgment rule information based on the similarity between the key information vector and the rule vector in the vectorized index; wherein, the vectorized index represents the mapping relationship between each judgment rule information and the corresponding rule vector; S2033. Based on key information and keyword index, retrieve the second judgment rule information; wherein, the keyword index represents the mapping relationship between each judgment rule information and the corresponding keyword; S2034. Based on the first decision rule information, the second decision rule information, and the knowledge graph index, the third decision rule information is retrieved; wherein, the nodes of the knowledge graph index represent the decision rule information, and the edges of the knowledge graph index represent the association between the decision rule information. S2035. Sort and deduplicate the first, second, and third determination rule information to obtain the target determination rule.

[0090] For example, the process first parses the query question and the information of the map features to be judged, extracting key information such as land type, region, time, and change type. Then, it retrieves first judgment rule information from the vectorized index based on vector similarity; second judgment rule information from the keyword index based on keywords; and third judgment rule information from the knowledge graph index based on graph relationships. Finally, it merges the first, second, and third judgment rule information, sorts them, and removes duplicates. The advantage of this approach is that it can accurately focus on relevant rules, improving judgment efficiency.

[0091] In addition, search enhancement strategies can be adopted during the search process, such as expanding the query by using synonyms and related words: multi-hop search can be achieved by expanding the association rules based on the preliminary search results.

[0092] Specifically, let the query question entered by the user be... The information of the patch to be determined is The rule retrieval function executes according to different strategies: 1) Full rule mode directly retrieves all effective judgment rule information in the target policy and regulation rule base to obtain the target judgment rule, i.e. .

[0093] 2) Retrieval rule pattern: First, analyze the query question and the information of the map features to be judged to extract key information. , This represents the key information extraction function. Indicates land category, Indicates a region. Indicates time, The change type is indicated, and then multi-dimensional retrieval and fusion are performed to obtain the target determination rule. , This represents a multi-dimensional rule fusion and ranking operation, where vectorized index retrieval obtains the Top-K results to get the first decision rule information, denoted as... Here, Indicates taking the first indivual Information on the rules for determining the largest value. for The latest version status, keyword index retrieval yields the second judgment rule information, represented as , This represents a keyword retrieval function. Knowledge graph indexing extends retrieval to obtain information on third-party decision rules. , Indicates association rules, Indicates and Related decision-making rule information. The final result is merged and sorted to obtain the target decision-making rule, represented as follows: Among them, the fusion score , This represents the raw semantic similarity value of the search results obtained during vector retrieval. This represents the score obtained during keyword retrieval, a normalized result calculated based on the number of matched keywords and the weight of each keyword. This represents the score indicating the degree of correlation with the target determination rule information, and the weights satisfy the following: .

[0094] S30. Based on the query question, the information of the map features to be determined, the target determination rules, and the preset determination prompt word template, generate target determination prompt words.

[0095] For example, the target judgment prompts are obtained by concatenating the query question, the information of the patch to be judged, the target judgment rules, and the preset judgment prompt word template. The target judgment prompt word assembly structure includes: (a) Preset judgment prompt word template: Define the roles and task requirements of the large language model; (ii) Rule base content: List all acquired target determination rules in sequence; (III) User Questions and Map Plot Information: This includes user query questions and map plot information to be determined (land type code, location information, area, change time, approval status, etc.); (iv) Output requirements: Specify the format and content requirements for the output of the large language model.

[0096] Target determination prompt word construction function Assemble the components into a complete target determination prompt. Its structure is formalized into ordered tuples. ,in For preset judgment prompt word templates, For the sequence of target determination rules, For query questions, Let be a tuple containing the information of the patch to be determined, where A unique code for a map patch. This refers to the land use type of the map patch. This is the location information of the map patch. Let be the area of ​​the patch. Land use change type For land use change time, The approval status of changes to map features. For output requirements, The maximum sequence number. Target determination prompts must satisfy the context length constraint of the large language model. ,in Limits the length of the window.

[0097] S40. Input the target determination prompt words into the large language model to obtain the target determination result for the information of the patch to be determined.

[0098] For example, the target determination prompt is submitted to the large language model for inference to obtain the target determination result, and then returned to the user in a structured format.

[0099] The structure of the target determination result includes: ① Judgment: Illegal / Legal / Pending verification; ② Type of violation (violation_type): such as "illegally occupying cultivated land" or "unauthorized change of land use"; ③ Applied rules: List the specific rules on which the judgment is based, including the rule identifier, rule content, and rules applicable instructions; ④ Reasoning process: Explain in detail the reasoning logic behind the judgment; ⑤ Confidence level: High / Medium / Low; ⑥ Suggestions: Provide suggestions for subsequent handling.

[0100] Specifically, suppose the structured output of the target determination result is defined as This includes the judgment conclusion. Types of violations Applicable rules Reasoning process Confidence level and recommendations for handling ,in, It is illegal. For legality, This needs further verification. The decision-making logic is based on a conditional matching function. To achieve, among which, This is the set of conditions used for the rule. If a target determination rule exists that satisfies this condition... If all target determination rules are not met, then it is judged as ILLEGAL. If it is, it is judged as LEGAL; otherwise, it is PENDING.

[0101] The present invention provides an intelligent determination method for land use change patches, which has the following advantages compared with the prior art: 1) Improved accuracy of rule retrieval: By constructing a dedicated target policy and regulation rule base to replace the original document retrieval, the content of the regulations and policies has been refined and structured through a large language model, resulting in clear semantics and well-defined boundaries. This solves the problem of retrieval failure caused by the semantic mismatch between user questions and rule content in traditional RAG technology, and significantly improves the accuracy and recall rate of rule retrieval.

[0102] 2) Ensures the completeness of the judgment basis: Through the unified management of the target policy and regulation rule library, the judgment rules scattered in multiple policy and regulation documents are centrally stored, and the full rule mode is supported to obtain the complete judgment basis. This solves the problem of incomplete basis caused by the dispersion of judgment rules and ensures that all the necessary rules can be obtained when judging the map patch.

[0103] 3) Implemented dynamic rule updates: Innovatively introduced a rule version management mechanism, supporting automatic comparison and update of rules, manual update, and historical version tracing, solving the problem of difficulty in synchronizing the rule base after policy and regulation updates, and ensuring the timeliness and accuracy of the judgment basis.

[0104] 4) Improved rule extraction efficiency: By leveraging the document understanding capabilities of large language models, structured rule information is automatically extracted from policy and regulatory documents. Compared with manual compilation, the extraction efficiency is significantly improved, the completeness is better, and the human resource costs of building and maintaining the rule base are reduced.

[0105] 5) Improved the quality of target determination prompt word construction: By organizing rules as independent units and using specially designed prompt word templates, the large language model can better understand and apply rules for reasoning and determination, solving the problem of insufficient context utilization in traditional RAG solutions and improving the accuracy of determination.

[0106] 6) Enhanced interpretability of the judgment: The judgment results include the specific rule clauses cited, the rules applicable instructions and the complete reasoning process, which has good interpretability, facilitates the review and confirmation by business personnel and the subsequent handling of appeals, and improves the credibility and traceability of the judgment results.

[0107] Secondly, embodiments of the present invention also provide an intelligent determination device for land use change patches, such as... Figure 4 As shown, the device may include: The acquisition module 401 is used to acquire the query question input by the user and the information of the patch to be judged; The judgment rule retrieval module 402 is used to retrieve target judgment rules from a pre-built target policy and regulation rule library based on the query question, the information of the land use change ... The prompt word generation module 403 is used to generate target determination prompt words based on the query question, the information of the patch to be determined, the target determination rules, and the preset determination prompt word template; The judgment result output module 404 is used to input the target judgment prompt words into the large language model to obtain the target judgment result for the information of the patch to be judged.

[0108] For details regarding the device, please refer to the steps of the intelligent determination method for land use change patches provided in the first aspect; these will not be elaborated upon here.

[0109] The present invention provides an intelligent determination device for land use change patches, which has the following advantages compared with the prior art: 1) Improved accuracy of rule retrieval: By constructing a dedicated target policy and regulation rule base to replace the original document retrieval, the content of the regulations and policies has been refined and structured through a large language model, resulting in clear semantics and well-defined boundaries. This solves the problem of retrieval failure caused by the semantic mismatch between user questions and rule content in traditional RAG technology, and significantly improves the accuracy and recall rate of rule retrieval.

[0110] 2) Ensures the completeness of the judgment basis: Through the unified management of the target policy and regulation rule library, the judgment rules scattered in multiple policy and regulation documents are centrally stored, and the full rule mode is supported to obtain the complete judgment basis. This solves the problem of incomplete basis caused by the dispersion of judgment rules and ensures that all the necessary rules can be obtained when judging the map patch.

[0111] 3) Implemented dynamic rule updates: Innovatively introduced a rule version management mechanism, supporting automatic comparison and update of rules, manual update, and historical version tracing, solving the problem of difficulty in synchronizing the rule base after policy and regulation updates, and ensuring the timeliness and accuracy of the judgment basis.

[0112] 4) Improved rule extraction efficiency: By leveraging the document understanding capabilities of large language models, structured rule information is automatically extracted from policy and regulatory documents. Compared with manual compilation, the extraction efficiency is significantly improved, the completeness is better, and the human resource costs of building and maintaining the rule base are reduced.

[0113] 5) Improved the quality of target determination prompt word construction: By organizing rules as independent units and using specially designed prompt word templates, the large language model can better understand and apply rules for reasoning and determination, solving the problem of insufficient context utilization in traditional RAG solutions and improving the accuracy of determination.

[0114] 6) Enhanced interpretability of the judgment: The judgment results include the specific rule clauses cited, the rules applicable instructions and the complete reasoning process, which has good interpretability, facilitates the review and confirmation by business personnel and the subsequent handling of appeals, and improves the credibility and traceability of the judgment results.

[0115] It should be noted that the terms "first," "second," etc., are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention.

[0116] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0117] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings and the disclosure in carrying out the claimed invention. In the description of the invention, the word "comprising" does not exclude other components or steps, "a" or "an" does not exclude a plurality, and "a plurality" means two or more, unless otherwise explicitly specified. Furthermore, while different embodiments may describe certain measures, this does not mean that these measures cannot be combined to produce good results.

[0118] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. An intelligent method for determining land use change patches, characterized in that, include: Obtain the user's query question and the information of the image patches to be judged; Based on the query question, the information of the land use change plot to be determined, and the preset retrieval strategy, target determination rules are obtained from a pre-built target policy and regulation rule base; the target policy and regulation rule base is constructed by extracting structured rules and multi-dimensional indexes from the original policy and regulation documents used for determining land use change plots. The target determination rule is used to determine whether the information of the patch to be determined conforms to the policies or regulations in the original policy and regulation document; Based on the query question, the information of the patch to be determined, the target determination rules, and the preset determination prompt word template, generate target determination prompt words; The target determination prompt is input into the large language model to obtain the target determination result for the image patch information to be determined.

2. The intelligent determination method for land use change patches according to claim 1, characterized in that, The process of constructing the target policy and regulation rule base includes: The original policy and regulatory documents are preprocessed to obtain preprocessed policy and regulatory documents; The preset rule-extracted prompt words and the preprocessed policy and regulation documents are input into the large language model for structured rule extraction, resulting in multiple judgment rule information; Based on the aforementioned multiple judgment rule information, a vectorized index, a keyword index, and a knowledge graph index are established respectively to obtain the target policy and regulation rule base.

3. The intelligent determination method for land use change patches according to claim 2, characterized in that, The step of obtaining target determination rules from a pre-built target policy and regulation rule base based on the query question, the information of the patch to be determined, and a preset retrieval strategy includes: Determine whether the number of judgment rule information in the target policy and regulation rule base whose current version status is effective does not exceed a preset threshold; If so, then all judgment rule information that is currently in effect in the current version will be determined as the target judgment rule; If not, then based on the query question and the information of the patch to be determined, a rule retrieval is performed from the pre-built target policy and regulation rule library to obtain the target determination rule.

4. The intelligent determination method for land use change patches according to claim 3, characterized in that, The step of retrieving target determination rules from a pre-built target policy and regulation rule base based on the query question and the information of the map feature to be determined includes: The query question and the information of the patch to be determined are analyzed to obtain key information; The key information is converted into a key information vector, and the first judgment rule information is retrieved based on the similarity between the key information vector and the rule vector in the vectorized index; wherein, the vectorized index represents the mapping relationship between each judgment rule information and the corresponding rule vector; Based on the key information and the keyword index, second determination rule information is retrieved; wherein, the keyword index represents the mapping relationship between each determination rule and the corresponding keyword; Based on the first determination rule information, the second determination rule information, and the knowledge graph index, the third determination rule information is retrieved; wherein, the nodes of the knowledge graph index represent the determination rule information, and the edges of the knowledge graph index represent the association relationships between the determination rule information. The first determination rule information, the second determination rule information, and the third determination rule information are sorted and deduplicated to obtain the target determination rule.

5. The intelligent determination method for land use change patches according to claim 2, characterized in that, The method further includes: Obtain newly released original policy and regulatory documents and extract structured rules to obtain the judgment rule information to be updated; The update type of the judgment rule information to be updated is determined by using a large language model or a preset update type determination strategy; wherein, the update type of the judgment rule information to be updated includes addition, revision, supplementation and repeal. According to the update type of the judgment rule information to be updated, the judgment rule information to be updated is updated to the target policy and regulation rule library.

6. The intelligent determination method for land use change patches according to claim 5, characterized in that, The update type of the judgment rule information to be updated is determined by a preset update type determination strategy, including: The determination rule information to be updated is matched with the determination rule information in the target policy and regulation rule library by number; If the matching result is empty, then the update type of the judgment rule information to be updated is determined to be new; If a match is successful, the matching rule information is determined as the matching rule information, and the semantic similarity between the rule information to be updated and the matching rule information is calculated to obtain the matching similarity. If the matching similarity is less than a preset similarity threshold, the update type of the rule information to be updated is determined to be revision. If the matching similarity is greater than or equal to the preset similarity threshold, the update type of the rule information to be updated is determined to be supplement. If the latest version status of the matching rule information corresponding to the rule information to be updated is invalid, the update type of the rule information to be updated is determined to be obsolescence.

7. The intelligent determination method for land use change patches according to claim 5, characterized in that, The step of updating the judgment rule information to the target policy and regulation rule base according to the update type of the judgment rule information to be updated includes: If the update type of the judgment rule information to be updated is "add", then a vectorized index, a keyword index, and a knowledge graph index are established for the judgment rule information to be updated, and added to the target policy and regulation rule base; If the update type of the judgment rule information to be updated is revision or supplementation, then the rule content contained in the vectorized index, keyword index and knowledge graph index in the target policy and regulation rule base is updated based on the judgment rule information to be updated. If the update type of the judgment rule information to be updated is obsolete, then the vectorized index, keyword index, and knowledge graph index corresponding to the judgment rule information to be updated in the target policy and regulation rule base will be marked as invalid.

8. The intelligent determination method for land use change patches according to claim 5, characterized in that, The method further includes: Based on the update type of each judgment rule in the target policy and regulation rule base, determine the version change information and current version status of each judgment rule; the version change information includes the creation time, modification time, and description of the modification content; the current version status is effective, repealed, or pending effectiveness.

9. The intelligent determination method for land use change patches according to claim 2, characterized in that, The method further includes: Receive rule deletion commands input by the user; The rule deletion instruction is parsed using a large language model, and at least one candidate judgment rule is retrieved from the target policy and regulation rule base and displayed. Receive a user's selection instruction for the at least one candidate judgment rule information and determine the judgment rule information to be deleted; Delete the decision rule information to be deleted, and delete the vectorized index, keyword index and knowledge graph index corresponding to the decision rule information to be deleted in the target policy and regulation rule base.

10. An intelligent determination device for land use change patches, characterized in that, include: The acquisition module is used to acquire the user's query question and the information of the patch to be judged; The judgment rule retrieval module is used to retrieve target judgment rules from a pre-built target policy and regulation rule library based on the query question, the information of the land use change plot to be judged, and a preset retrieval strategy. The target policy and regulation rule library is constructed by extracting structured rules and using multi-dimensional indexes from the original policy and regulation documents used for judging land use change plots. The target judgment rules are used to determine whether the information of the land use change plot to be judged conforms to the policies or regulations in the original policy and regulation documents. The prompt word generation module is used to generate target judgment prompt words based on the query question, the information of the patch to be judged, the target judgment rule, and the preset judgment prompt word template; The judgment result output module is used to input the target judgment prompt words into the large language model to obtain the target judgment result for the image patch information to be judged.

Images