Multi-modal large model fine-tuning corpus production method for planning and natural resource field

Abstract

The invention provides a multi-modal large model fine-tuning corpus production method for the field of planning and natural resources, and aims at solving the problems that existing corpus lacks professional semantics, manual annotation is low in efficiency, and no standardized process exists. According to the method, an exclusive VQA corpus is generated through four core modules including construction of an industry exclusive cognitive task and VQA template library, image-text pairing generation, template-driven VQA sample generation and automatic quality control optimization in combination with an industry planning standard and a cognitive hierarchy, namely perception-reasoning-association-application. The method comprises the following steps: firstly, automatically extracting visual elements and policy texts of a planning graph, and matching a professional template to generate candidate Qamp; the method comprises the following steps: A, finally outputting a standardized corpus package through semantic detection and expert re-check optimization; according to the method, professional semantic accurate alignment is realized, the labor cost is greatly reduced, the corpus quality is reliable, the method can be expanded to similar fields, and efficient support is provided for field multi-mode large model fine adjustment and capability evaluation.

Description

Technical Field

[0001] This invention relates to the field of multimodal corpus construction technology, and in particular to a method for producing a multimodal large-scale model fine-tuning corpus for the planning and natural resources field. Background Technology

[0002] Current multimodal corpus construction mainly focuses on general visual language tasks, such as COCO Caption, VisualGenome, and ChartQA. These datasets share common characteristics: image content mainly consists of natural scenes, everyday objects, and general charts; text annotation emphasizes object recognition and general description, lacking planned semantics and spatial logic; and generation methods are mainly manual or crowdsourced annotation, making it difficult to cover specialized symbol systems.

[0003] In the fields of planning and natural resources, some auxiliary annotation and map reading systems have also emerged, such as tools for map patch recognition, automatic mapping and indicator extraction in geographic information mapping systems (GIS). However, these methods are mainly oriented towards vector data and spatial computing, have not established image-text VQA mapping relationships, and lack a multi-level understanding of policy semantics, spatial structure logic and planning intent, and cannot support end-to-end large model fine-tuning or question-and-answer generation tasks.

[0004] In summary, while Multimodal Large Language Models (MLLMs) currently perform well in general-purpose scenarios, they still have the following challenges in specialized fields such as urban planning and natural resource management: 1) Insufficient professional semantic understanding ability: General models cannot accurately identify the symbol system, spatial elements and policy intentions in planning maps; 2) Lack of systematic construction methods for industry corpora: planning documents and image data are from multiple sources and are heterogeneous, lacking standardized image-text pairing rules and knowledge system mapping; 3) Manual annotation is inefficient and costly: Existing data construction relies on expert manual annotation and lacks automated and semi-automated auxiliary mechanisms; 4) Lack of data generation process for VQA (Visual Question Answering) tasks: especially for complex visual text content such as planning documents, spatial layout diagrams, and design schemes, there are no reusable corpus production standards.

[0005] Therefore, there is an urgent need for a technical system that can automatically transform knowledge systems in fields such as planning images, policy texts, and exam questions into VQA training samples, so as to achieve standardized, multi-stage, and reusable corpus production. Summary of the Invention

[0006] The purpose of this invention is to propose a multimodal corpus production and structured generation process for the planning and natural resources industry, which supports the fine-tuning and evaluation of large-scale models in the field, laying the foundation for subsequent automatic understanding, design assistance, and policy generation in urban planning.

[0007] To achieve the above objectives, this invention proposes a method for producing a multimodal large-scale model fine-tuning corpus in the fields of planning and natural resources, comprising the following steps: S1. Construct industry-specific cognitive tasks and VQA template library: Based on industry standards, professional examination systems and planning text and image tasks in the planning and natural resources field, define a four-level progressive cognitive task system covering perception, reasoning, association and application. At the same time, construct a VQA question template library corresponding to each cognitive level, including planning map classification and planning map element recognition, spatial relationship, professional reasoning and policy association. S2. Image-text pairing generation: Obtain multi-format raw data in the fields of planning and natural resources, filter, parse and extract elements from the raw data, establish a corresponding mapping relationship between image data and text data including policy text, planning indicators and context descriptions, and obtain standardized image-text pairing data; S3. Template-driven VQA sample generation: The multimodal large model is called to generate candidate question-answer pairs based on the VQA question template library and image-text pairing data. The candidate question-answer pairs are screened according to the planning industry professional semantics. The accuracy of the question-answer expression is corrected by the industry technical requirements in the template library. The initial VQA training samples containing image-question-answer are output. S4. Automated Quality Control Optimization: The initial VQA training samples are subjected to automatic semantic consistency detection and expert review of logical rationality. Finally, they are classified and labeled according to question type, difficulty and four-level cognitive task system to obtain high-quality standardized corpus. S5. Scalable Corpus Output: Output the standardized corpus in a preset format. The standardized corpus includes image ID, question text, answer text, task type, and annotation source information, which is used for fine-tuning of multimodal large models in the planning and natural resources field.

[0008] Furthermore, by replacing industry standards and professional terms in the VQA problem template library, we can adapt and fine-tune multimodal large models in the fields of architectural design and ecological restoration.

[0009] Furthermore, in step S1, the four cognitive levels specifically include: Perception layer: This includes element recognition and image description. The element recognition is used to evaluate the model's ability to recognize the layout structure, text annotations, basic geographic features and drawing elements in the planning map, and to establish semantic alignment between image content and natural language. The image description is used to extract details based on the image itself and generate an objective description. Reasoning levels include: planning map classification, spatial relationship reasoning, and professional reasoning. The planning map classification is based on the "five-level, three-category" planning system, dividing planning maps into master plans, detailed plans, and special plans. Provincial planning maps are further divided into basic analysis maps and planning result maps, while municipal planning maps are divided into survey maps, control maps, and schematic maps. Spatial relationship reasoning includes topological spatial relationship reasoning, sequential spatial relationship reasoning, and metric spatial relationship reasoning. Professional reasoning includes four dimensions: spatial layout, functional organization, transportation system, and ecological environment. Association hierarchy: used to evaluate the model's ability to associate planning maps with underlying policies, regulations, and planning indicators, and to establish cross-domain associations between spatial elements and corresponding policy frameworks; Application level: This includes scheme evaluation and decision making. Scheme evaluation is used to assess the merits of planning schemes based on visual input and spatial context. Decision making is used to make value- and principle-based choices in constrained design situations.

[0010] Furthermore, in step S1, the VQA question template library specifically includes: Planning map classification templates include: provincial basic analysis maps (12 categories such as location analysis maps and topographic maps), provincial planning result maps (20 categories such as territorial spatial development and protection pattern maps and three control line maps), municipal survey maps (5 categories such as municipal territorial spatial land and sea use status maps), municipal control maps (20 categories such as municipal territorial spatial control line planning maps), and municipal schematic maps (5 categories such as municipal main functional zoning maps). Spatial Relationship Templates: These include templates for determining topological spatial relationships, querying sequential spatial relationships, and comparing metric spatial relationships. They also provide standardized problem descriptions for determining adjacent / containment / intersection relationships between geographic entities, querying directional distribution, and comparing distances. Professional reasoning templates include spatial layout analysis templates, functional organization assessment templates, transportation system adaptation templates, and ecological environment adaptation templates, which correspond to the problem templates for analyzing the characteristics and causes of urban layout, assessing the rationality of functional area planning, analyzing the support of transportation facility layout for economic and social development, and analyzing the guarantee of ecological protection planning for sustainable development. Policy-related templates: These include templates for linking planning indicators, templates for adapting to regulatory requirements, a query function for linking planning indicators in the corresponding planning map, and templates for determining compatibility with land use regulations and ecological protection regulations.

[0011] Furthermore, step S2 specifically includes: extracting and filtering valid planning maps through a custom script, extracting text information from the original data using OCR recognition technology, and extracting professional visual elements from the planning maps through semantic segmentation or layer parsing technology, thereby establishing a mapping relationship between image data and corresponding policy texts, planning indicators, and contextual descriptions; The elements of the planning map include administrative boundaries, government seats, annotations, scale, legend, contour lines, and water system; The visual elements include roads, green spaces, plots of land, and transportation facilities; The multi-format raw data includes planning PDF files, JPG image files, and DOC text files; The original data comes from official planning documents issued by municipal governments, planning agencies, and academic institutions.

[0012] Furthermore, in step S3, the multimodal large model includes at least one of the Qwen2.5-VL-72B-Instruct model and the GPT-4V model. The screening process specifically involves eliminating candidate question-answer pairs that do not conform to the professional terminology standards of the planning industry or deviate from the technical requirements of planning. The correction process includes supplementing planning industry-specific expressions and adjusting the consistency between the question-answer logic and the professional logic of planning.

[0013] Furthermore, in step S4, the semantic consistency detection model is implemented through a pre-trained planning industry semantic matching model; the review content of the logical rationality expert review includes whether the question-answer pair conforms to the planning industry technical specifications, whether it accurately reflects the core information of the planning map, and whether it conforms to the requirements of the policy text.

[0014] Furthermore, the spatial layout in the professional reasoning includes centralized layout and decentralized layout; the centralized layout includes grid-like and ring-radial layouts; the decentralized layout includes cluster-like, strip-like, star-like, ring-like, satellite-like, multi-center, and cluster city layouts. The functional organization in the professional reasoning includes industrial land, residential areas, warehousing areas, and public facilities land. Among them, the layout of industrial land must meet the requirements of pollution isolation from residential areas and the requirements of optimized transportation links, while the layout of warehousing areas must meet the requirements of hazardous materials isolation and convenient logistics transportation. The transportation system in the professional reasoning includes the site selection and layout assessment of railways, highways, ports, and airports, which must meet the adaptability requirements of transportation facilities and urban space in the planning industry. The ecological and environmental reasoning in the professional reasoning includes the assessment of the protection and utilization of the natural environment and ecosystems, which must comply with the requirements of ecological protection red lines and nature reserve planning.

[0015] Furthermore, the planning indicators in the associated hierarchy include indicators for cultivated land area, permanent basic farmland protection area, ecological protection red line area, and urban development boundary expansion multiple. The policy framework includes territorial spatial planning outline documents, industry technical regulations, and related legal documents.

[0016] This invention also proposes a corpus production system for fine-tuning multimodal large models in the fields of planning and natural resources, comprising: Data input module: Receives raw data in multiple formats from the planning domain; Task and Template Building Module: Defines four-level progressive cognitive tasks and builds an industry-specific VQA question template library; Image-text matching module: Filters and optimizes raw data, extracts text and professional visual elements, and establishes accurate cross-modal mapping; Sample generation module: calls the multimodal large model to generate and filter candidate question-answer pairs, and outputs initial VQA samples; Quality control module: This module optimizes samples through automatic detection and professional review to complete classification and labeling. Scalable output module: Outputs standardized corpora in a structured format, supporting cross-domain template replacement and adaptation.

[0017] Furthermore, the image-text matching module includes a custom script unit, an OCR recognition unit, a semantic segmentation unit, and a mapping establishment unit.

[0018] Furthermore, the quality control module includes a semantic consistency detection unit and a professional review unit.

[0019] Compared with the prior art, the advantages of the present invention are: 1. This invention deeply embeds urban and rural planning professional standards, the "five-level, three-category" planning system, and the cognitive logic of urban and rural planner examinations into the corpus production process. It designs exclusive VQA templates based on planning map types (including professional questions such as map classification, element identification, and policy correlation), and constructs a standardized question bank for core elements in the planning field (administrative boundaries, land use types, transportation systems, ecological red lines, etc.). This achieves professional semantic alignment, enabling the model's professional reasoning to cover the four core dimensions of planning: spatial layout, functional organization, transportation systems, and environmental ecology, rather than simple reasoning for general scenarios. It realizes cross-domain integration of general multimodal corpus generation and planning industry professional knowledge, establishing a professional "image-text" VQA mapping relationship. This solves the problems of missing professional semantics and insufficient spatial logic in existing general corpora, generating corpora that accurately match the fine-tuning needs of large-scale models in the field.

[0020] 2. Compared to existing corpora that primarily focus on basic tasks and lack the advanced cognitive tasks required in the planning field, this invention cannot support the deep training and evaluation of large models in professional scenarios. This invention constructs a four-tiered cognitive task system encompassing "perception, reasoning, association, and application," covering multiple levels of cognitive tasks such as element recognition, image description (perception), planning map classification, spatial relationship / professional reasoning (reasoning), policy and planning map association (association), and scheme evaluation and decision-making (application). This transforms the professional cognitive logic of the planning industry into a quantifiable and reusable corpus task system, enabling the generated corpus to be used not only for model fine-tuning but also for professional capability evaluation, filling the gap in existing technologies for producing corpora for advanced cognitive tasks in professional fields.

[0021] 3. This invention utilizes existing large models to automatically generate candidate Q&A. Through this template-driven and multimodal large model-assisted generation method, automated corpus generation is achieved, significantly reducing reliance on expert manual annotation. Compared with traditional manual methods, this invention significantly improves efficiency and reduces costs.

[0022] 4. This invention establishes a high-quality control mechanism combining automatic semantic detection with expert review. It automatically verifies the logical consistency between questions and answers, corrects professional deviations with the help of an expert team, and outputs a standardized corpus format, ensuring the accuracy and consistency of the generated corpus. By designing standardized question templates (such as provincial planning map classification questions, spatial relationship reasoning questions, and professional policy association questions) for different types, elements, and spatial relationships of provincial and municipal planning maps, it outputs a unified format corpus package containing core fields such as "image ID, question text, answer, task type, and annotation source," which can be directly connected to the fine-tuning interface of multimodal large models. This solves the problems in existing technologies where the lack of systematic quality control over professional corpora leads to semantic contradictions, professional errors, and messy formats in the generated samples, making them unsuitable for direct use in large model fine-tuning.

[0023] 5. The method and system of this invention are highly scalable. Their core processes, including task decomposition, image-text matching, template generation, and quality control, adopt a modular design, allowing direct adaptation to multimodal corpus production in related fields such as architectural design, ecological restoration, and land space management. Simply replacing the corresponding professional standard templates (such as architectural design codes and ecological restoration technical regulations) allows for rapid generation of multimodal corpora in that field. This solves the problems of existing professional corpus production technologies being specific to a single scenario and lacking reusability. Attached Figure Description

[0024] Figure 1 This is a flowchart of a method for producing a multimodal large model fine-tuning corpus in the field of planning and natural resources, according to an embodiment of the present invention. Figure 2The map shown is from page 27 of the "Jiangsu Provincial Territorial Spatial Planning (2021-2035)" used in the method of this embodiment of the invention. Figure 3 This is a partial view of page 18 of the "Jiangsu Provincial Territorial Spatial Planning (2021-2035)" as described in the embodiments of the present invention. Figure 4 This is an example diagram of the standardized corpus package output in the method of this embodiment of the invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be further described below.

[0026] Example 1 This embodiment 1 proposes a system and method for producing a multimodal large-scale model fine-tuning corpus in the fields of planning and natural resources, including the following steps: S1. Construct industry-specific cognitive tasks and VQA template library: Based on industry standards, professional examination systems and planning text and image tasks in the planning and natural resources field, define a four-level progressive cognitive task system covering perception, reasoning, association and application. At the same time, construct a VQA question template library corresponding to each cognitive level, including planning map classification and planning map element recognition, spatial relationship, professional reasoning and policy association. In this embodiment, the four hierarchical dimensions of perception, reasoning, association, and application respectively reflect different levels of image understanding, as follows: (1) Perception level, which includes two subclasses: element recognition and image description; Element recognition reflects the model's ability to identify layout structure, text annotations, basic geographic features, and drawing elements in planning maps, and is used to establish semantic alignment between image content and natural language. Image description refers to extracting as much detail as possible from an image and generating a description based solely on the image itself, rather than identifying a specific problem. The quality of the description reflects whether the model has truly "seen" the image.

[0027] (2) Reasoning levels, which include three categories: classification, spatial relationship reasoning, and professional reasoning: The classification reflects the model's ability to identify different types of planning maps. It focuses only on statutory spatial planning and does not include conceptual planning, urban design, regional strategic planning, or thematic studies. Specifically, based on China's "five-level, three-category" planning system, maps are divided into master plans, detailed plans (including regulatory plans and site plans), and special plans. For maps without a clearly marked type (such as those in exam questions), experts will label the type based on the scale and planning content. Spatial relationship reasoning reflects a model’s ability to understand spatial relationships between geographic elements, including topological, ordinal, and metric relationships. Professional reasoning encompasses four aspects: spatial layout, functional organization, transportation system, and ecological environment, reflecting the core dimensions of professional planning interpretation.

[0028] (3) The association level reflects the model's ability to associate planning maps with the policies and contextual documents behind them. It examines policies, regulations, and planning indicators. Unlike tasks that only focus on visual recognition, this association task requires the model to connect spatial elements with the corresponding policy framework (such as administrative levels, regulatory classifications, and land use standards), so that the model can establish cross-domain associations between visual and textual modalities, aligning spatial content with normative planning knowledge.

[0029] (4) Application level, which reflects the model’s ability to compare, evaluate and optimize planning schemes. Unlike identification or fact recall, the application level requires the model to weigh ecological protection, land use efficiency and development goals, thereby demonstrating its ability to generate strategic, selective and professionally based judgments.

[0030] This application layer includes two types of tasks: scheme evaluation and decision making. Scheme evaluation refers to testing whether the model can evaluate the merits of planning schemes based on visual input and spatial context, based on factors such as land use distribution, spatial continuity, accessibility, and compatibility. Decision making refers to evaluating whether the model can make value- and principle-based choices in a constrained design context.

[0031] Based on the aforementioned four-level progressive cognitive task system, a VQA question template library corresponding to each cognitive level is constructed, specifically including: V1. Planning Map Classification Template The classification of provincial territorial spatial planning maps is shown in Table 1. The "Technical Regulations for the Compilation of Provincial Territorial Spatial Planning" divides provincial territorial spatial planning maps into basic analysis maps and planning result maps. Basic analysis maps cover location analysis maps, topographic and geomorphological maps, administrative division maps, current status maps of land and sea use, mineral resource distribution maps, current status maps of nature reserves, current status maps of urban systems, current status maps of comprehensive transportation, current status maps of historical and cultural protection, current status maps of water conservancy infrastructure, and other current status maps such as geological, hydrological, disaster, and marine environmental quality maps, as well as resource and environmental carrying capacity and territorial spatial development suitability evaluation maps, including single-factor evaluation and comprehensive carrying capacity evaluation maps. The planning outcome maps include: a map of the spatial development and protection pattern of the national territory; a map of the red line for the protection of arable land and permanent basic farmland; a map of the red line for ecological protection; a map of the urban development boundary; a map of the three control lines; a map of the distribution of national and provincial main functional zones; a map of the optimized pattern of major agricultural production areas; a map of the optimized pattern of key ecological functional zones; a map of the optimized pattern of urbanized areas; a map of the planning area for key prevention and control of major disasters; a map of the functional layout of marine space; a map of the overall protection spatial system of cultural and natural heritage; a map of the planning system of nature reserves; a map of biodiversity conservation; a map of the planning for water resource security and water source conservation; a map of the planning for key infrastructure; a map of the planning for coastal zone protection and utilization; a map of the planning for ecological restoration and comprehensive land management; a map of the planning for energy resource security; and a map of the strategic pattern of integrated land and sea management.

[0032] Table 1 Classification of Provincial Territorial Spatial Planning Maps

[0033] According to the "Standard for Mapping Municipal Territorial Spatial Master Plans," municipal territorial spatial master plan maps are classified into survey maps, control maps, and schematic maps. Survey maps comprise five types: a map showing the current status of land and sea use within the city limits, a map showing the current status of land and sea use in the central urban area, a map showing the distribution of nature reserves within the city limits, a map showing the distribution of historical and cultural heritage sites within the city limits, and a map showing the distribution of natural disaster risks within the city limits. The control-type maps are more abundant, totaling 20 maps (categories), covering the city's territorial spatial control line planning map, city's ecosystem protection planning map, city's agricultural (pastoral) spatial planning map, city's historical and cultural protection planning map, city's comprehensive transportation planning map, city's infrastructure planning map, city's territorial spatial planning zoning map, city's ecological restoration and comprehensive management planning map, city's mineral resources planning map, central urban area land use planning map, central urban area territorial spatial planning zoning map, central urban area development intensity zoning planning map, central urban area control line planning map, central urban area green space system and open space planning map, central urban area public service facility system planning map, central urban area historical and cultural protection planning map, central urban area road traffic planning map, central urban area municipal infrastructure planning map, central urban area comprehensive disaster prevention and mitigation planning map, and central urban area underground space planning map. There are 5 schematic maps (categories), including the city's main functional zoning map, the city's overall territorial spatial pattern planning map, the city's urban system planning map, the city's urban and rural living circle and public service facility planning map, and the central urban area urban renewal planning map. These maps together form the visual foundation of territorial spatial planning, providing an important basis for the formulation, implementation and management of the plan.

[0034] For example, if the claimed PDF is at the provincial level, the standard question would be: The "Technical Regulations for the Compilation of Provincial Territorial Spatial Planning" divides provincial territorial spatial planning maps into basic analysis maps and planning outcome maps. Basic analysis maps encompass location analysis maps, topographic and geomorphological maps, administrative division maps, current land and sea use maps, mineral resource distribution maps, current nature reserve maps, current urban system maps, current comprehensive transportation maps, current historical and cultural preservation maps, current water conservancy infrastructure maps, and other current status maps such as geological, hydrological, disaster, and marine environmental quality maps, as well as resource and environmental carrying capacity and territorial spatial development suitability evaluation maps, including single-factor evaluation and comprehensive carrying capacity evaluation maps. The planning outcome maps include: a map of the spatial development and protection pattern of the national territory; a map of the red line for the protection of arable land and permanent basic farmland; a map of the ecological protection red line; a map of the urban development boundary; a map of the three control lines; a map of the distribution of national and provincial main functional zones; a map of the optimized pattern of major agricultural production areas; a map of the optimized pattern of key ecological functional zones; a map of the optimized pattern of urbanized areas; a planning map of key areas for the prevention and control of major disasters; a map of the functional layout of marine space; a map of the overall protection spatial system of cultural and natural heritage; a planning map of the natural protected area system; a planning map of biodiversity conservation; a planning map of water resource security and water source conservation; a planning map of key infrastructure; a planning map of coastal zone protection and utilization; a planning map of ecological restoration and comprehensive land management; a planning map of energy resource security; and a strategic pattern map of integrated land and sea development. Which type of map does this particular map belong to? Planning map element identification: The element classification template for the overall land use plan map is shown in Table 2. The "Specifications for Mapping the Overall Land Use Plan of a Municipality" divides the elements of the overall land use plan map of a municipality into basic geographical elements, annotations, map sheet configuration, and drawing elements.

[0035] Table 2 Elements and Main Contents of the Overall Territorial Spatial Planning Map

[0036] Basic geographic elements include administrative boundaries, government seats, elevation features, contour lines and isobaths, and other land features. Regarding administrative boundaries, within the mapping area, they should be represented down to the district (county) or township level; outside the mapping area, they should be represented down to the province, city, or district (county) level. Border cities must indicate their national borders. The representation requirements for government seats are the same as for administrative boundaries. Elevation features include important mountain ranges, peaks, and passes, and their names and elevation values ​​should be labeled. In areas where elevation and underwater topography have a significant impact on the national land space, contour lines and isobaths can be added respectively. Other land features can be represented by water systems, coastlines, etc., depending on the regional situation, and should refer to relevant topographic map specifications.

[0037] Notes: All maps should include necessary annotations, mainly covering the names of city (prefecture), county (district), and township (town) government seats; railway stations, airports, ports, highways and railways and their accessible places; major water conservancy facilities, rivers, lakes, reservoirs, canals, sea areas; national parks, nature reserves, natural parks, etc.; and other important geographical features. The types of text used for annotations within the same drawing file should be limited to four. Chinese characters should preferably use Song typeface, with alternatives including Heiti, Kaiti, Fangsong, or Lishu; English and numbers should preferably use Times New Roman, with alternatives including Arial Black. Annotations of the same type within different drawing files should maintain consistent font and size. Annotation text for base map elements should primarily be gray or white, and clearly distinguishable in color and size from annotations for other elements. The representation of marine elements should comply with relevant regulations.

[0038] Map Sheet Configuration: The map sheet configuration of a municipal-level territorial spatial master plan map includes the map title, map frame, north arrow and wind rose diagram, scale, legend, signature, and date of creation. The map title should be located above the map frame and include the plan name and subject name. Chinese characters should be in bold, and English and numbers should be in Times New Roman. The map frame consists of an outer frame and an inner frame, drawn with thick and thin solid lines respectively. The north arrow and wind rose diagram are usually located in the upper right or upper left corner of the map sheet. Areas with wind direction data use a 16-direction or 8-direction wind rose diagram; other areas use a north arrow. A linear scale can be used, and its total length should preferably be 1 / 10 of the map frame width. The legend includes graphics (lines, color blocks, or symbols) and text, usually located below the map frame. The map should include the official name of the planning unit and the date of creation, located in the lower left or lower right corner outside the map frame.

[0039] Map elements include base map elements, mandatory elements for expressing the main content, and optional elements. Base map elements generally include administrative boundaries, natural geography, transportation, land use, and zoning. Regarding administrative boundary elements, city-level base maps must show district (county) level and above administrative boundaries and government seats, and the administrative boundaries of the mapping area should be shaded. Coastal cities should also include the coastline and municipal sea areas. Central urban area base maps must show township (town) level and above administrative boundaries and government seats, and be appropriately processed. Natural geographical elements include mountains and water systems. Regarding transportation elements, except for specific types of land use planning maps, other existing base maps must show existing airports, railways and stations, intercity rail, ports and wharves, highways, and urban backbone road networks, and can be selectively classified or expressed using different symbols superimposed on the same land use. Regarding land use and zoning elements, except for specific types of planning drawings, other existing status maps should express the existing construction land, including urban and rural construction land, regional infrastructure land, and other construction land. Planning maps should express urban development zones, and cities with the necessary conditions may add the expression of village construction zones.

[0040] If any mandatory elements are specified in this standard within the mapping area, they should be expressed as required. Optional elements may be selected and expressed according to the actual situation, or other elements may be added.

[0041] V2. Spatial Relationship Template: In this embodiment, spatial relationship refers to the interaction between geographic spatial entities. The main spatial relationships include: ① Topological spatial relationships: used to describe the relationships between entities such as adjacency, connection, inclusion, and intersection.

[0042] Topological relationships on a map refer to the property that graphic relationships remain unchanged despite deformations (scaling, rotation, and stretching) that maintain continuity. The shapes and sizes of various graphics on a map change with deformation, but the adjacency, association, containment, and connection relationships between graphic elements remain constant. Besides logically defining nodes, radians, and polygons to describe the topological relationships of graphic elements, topological relationships also exist between different types of spatial entities.

[0043] There are five possible relationships between points, lines, and planes: separation, adjacency, overlap, containment or coverage, and intersection. These relationships are detailed below: (i) Point-to-point relationship: There are only two relationships between point entities: disjoint and coincident.

[0044] (ii) Point-line relationship: There are only three relationships between point entities and line entities: adjacent, disjoint, and contained (sometimes also called intersecting).

[0045] (iii) Point-surface relationship: There are three types of relationships between point entities and surface entities: adjacent (sometimes called intersecting), disjoint, and contained.

[0046] (iv) Line-to-line relationships: There are five types of relationships between line entities: adjacent, intersecting, separated, contained, and overlapping.

[0047] (v) Line-surface relationship: There are four types of relationships between line entities and surface entities: adjacent, intersecting, disjoint, and contained.

[0048] (vi) Face-to-face relationship: There are 5 types of relationships between face entities: adjacent, intersecting, disjoint, contained, and overlapping.

[0049] ② Sequential spatial relationship: used to describe the order in which entities are arranged in geographic space, such as the directional relationships between entities, such as front and back, up and down, left and right, and east, west, south and north.

[0050] Sequential spatial relationships are based on the distribution of spatial entities in geographic space and are described using directional terms such as up / down, left / right, front / back, east / west / north / south. Similar to the formal description of topological spatial relationships, the sequential relationships between different types of spatial entities can be examined through various combinations such as point-to-point, point-to-line, point-to-area, line-to-line, line-to-area, and area-to-area. Since sequential spatial relationships require calculations of the orientations between spatial entities to arrive at corresponding directional descriptions, and such calculations are very complex, there is currently no good solution for constructing sequential spatial relationships between entities. Furthermore, sequential spatial relationships change with the projection and geometric transformation of spatial data. Therefore, sequential spatial relationships are not currently described or expressed in GIS.

[0051] ③Measure spatial relationships: used to describe the relationships such as distance between spatial entities.

[0052] Measuring spatial relationships primarily refers to the distance relationships between spatial entities. Different combinations of point-to-point, point-to-line, point-to-plane, line-to-line, line-to-plane, and plane-to-plane relationships can also be established in topological spatial relationships to examine the metric relationships between different types of spatial entities.

[0053] The corresponding example questions are shown in Table 3: Table 3. Examples of Spatial Relationships in the Overall Territorial Spatial Planning Map

[0054] V3. Professional Reasoning Template This includes templates for spatial layout analysis, functional organization assessment, transportation system adaptation, and ecological environment adaptation, corresponding to templates for analyzing urban layout morphological characteristics and causes, assessing the rationality of functional area planning, analyzing the support of transportation facility layout for economic and social development, and analyzing the guarantee of ecological protection planning for sustainable development. Details are as follows: A. Spatial Layout Analysis Template The overall urban layout is a comprehensive reflection of a city's social, economic, environmental, and engineering technology and architectural spatial combination. The concentrated and decentralized development of urban spatial structure has always been two important forces. All existing ideal urban forms can be traced back to these two basic development models. Many types of research have been conducted on urban layout forms. Synthesizing different research findings, based on the city's land use patterns and road network structure, they can be broadly categorized into two main types: concentrated and decentralized.

[0055] Centralized urban layout refers to the concentrated arrangement of major urban land uses in contiguous areas. Its advantages include ease of establishing comprehensive living service facilities, compact and economical land use, and efficient connections between various urban functions and residents. Generally, this type of urban development is encouraged in small and medium-sized cities. However, the layout of such cities needs to consider the relationship between short-term and long-term goals, allowing for flexibility in planning and leaving room for future development. This avoids a situation where, although the layout is compact in the short term, functional mixing and interference may occur in the long term.

[0056] Meanwhile, this centralized urban layout can be further divided into grid-like, ring-radial, and other types: (1) Grid-like Grid-like cities are the most common and traditional spatial layout pattern, consisting of a network of perpendicular roads. The city's regular shape easily accommodates the arrangement of various buildings, but if poorly managed, it can easily lead to a monotonous layout. This urban form generally forms easily in plains areas without external constraints and is unsuitable for areas with complex terrain. This form allows for urban expansion in all directions and is more suitable for the development of vehicular traffic. Due to the uniformity of the road network and similar accessibility across areas, it is not easy to form a significant, concentrated central area. Major examples include Los Angeles and Milton Keynes. Washington, D.C., has added radial roads to its grid-like road network, which can be seen as an improvement on this form.

[0057] (2) Circular radial The radial-ring road network is a common urban form in large and medium-sized cities. It features good urban accessibility, a strong tendency towards compact, centripetal development, and often a high-density, vibrant, and expressive city center. This type of city easily utilizes radial roads to organize the city's axial system and landscape. However, the biggest problem is the potential for congestion and over-concentration in the city center, along with poor land use regularity, which is detrimental to building layout. This form is generally unsuitable for small cities. Major examples include Beijing and Paris.

[0058] The most prominent feature of a decentralized urban layout is that the urban space is distributed in a non-clustered manner, including various forms such as clusters, belts, stars, rings, satellites, multi-center cities, and group cities.

[0059] (1) Clustered A cluster-like urban form refers to a city divided into several discontinuous urban land parcels, each separated by farmland, mountains, wide rivers, large forests, etc. The planning layout of this type of city can be flexibly formulated according to land conditions, making it easier to handle the short-term and long-term relationships of urban development, facilitating access to nature, and ensuring that each land use is properly allocated. The key is to strike a balance between concentration and dispersion, achieving both a reasonable division of labor and strengthened connections, while also ensuring a certain scale within each cluster, allowing departments with similar functions and natures to be relatively clustered. Convenient transportation links between clusters are essential. (2) Strip-shaped (linear) Linear cities are mostly formed due to topographical constraints, confining them to a narrow geographical space. They develop longitudinally along both sides of a main transportation axis, exhibiting strong directional characteristics in both their planar landscape and traffic flow. This spatial organization has certain advantages, but its size should be limited; it shouldn't be too long, otherwise transportation costs would be excessive. Therefore, it's essential to develop transportation lines parallel to the main axis. Key examples include Shenzhen and Lanzhou.

[0060] (3) Star-shaped (finger-shaped) Star-shaped cities typically emerge from a core urban area and expand outwards along multiple transportation corridors, reserving significant amounts of non-construction land between these corridors. This form can be viewed as a development pattern formed by superimposing multiple linear cities onto a ring-radial urban framework. The establishment of a radial, high-capacity public transportation system significantly influences this pattern, and strengthening control over non-construction land along the development corridors is crucial for ensuring its success. A prime example is Copenhagen.

[0061] (4) Circular Ring-shaped cities typically develop in a circular pattern around core elements such as lakes, mountains, and farmland. Structurally, they can be seen as the result of a linear city developing end-to-end under specific circumstances. Compared to linear cities, the closed ring shape facilitates connections between functional areas. Since the central part of the ring is predominantly natural space, it can create beautiful landscapes and favorable ecological conditions for the city. However, unless there are specific natural constraints or strict control measures, the pressure for urban land to expand towards the center of the ring is extremely high. Typical examples include Singapore, Taizhou in Zhejiang Province, and the Randstad region of the Netherlands.

[0062] (5) Satellite-like Satellite-shaped cities are generally formed around a large or megacity, with several smaller cities developing around it. Typically, the central city has a strong dominant position, while the smaller peripheral cities are relatively independent but maintain close ties with the central city in terms of production, work, culture, and daily life. This form is essentially a spatial model proposed by Howard's Garden City theory and Unwin's satellite city theory. This model is conducive to a balanced distribution of population and productivity within the large city and its surrounding hinterland.

[0063] (6) Multicenter and cluster cities This spatial form is the result of the continuous sprawl and development of cities in multiple directions. Multiple different districts or clusters develop independently under certain conditions, gradually forming diverse focal points, centers, and axes. Typical cities with this spatial form include Detroit and Los Angeles. In some densely populated urban areas, a more pronounced cluster development characteristic emerges, such as the Kyoto-Osaka-Kobe region of Japan. Centered on Osaka, this crescent-shaped area along a 50km radius northeast of Osaka Bay includes Kyoto, Kobe, and the historical capital Nara, forming the Osaka metropolitan area with a population of 17 million. With the completion of major projects such as the Kansai International Airport, the Kansai Cultural and Academic Research City, and the Osaka Bay inter-regional development, a multi-center network-type metropolitan area structure with highly concentrated population, industry, and culture has formed along the interconnected axes of these cities. The goal is to build a central city for international exchange, stimulating urban vitality and creating a favorable urban environment.

[0064] B. Functional Organization Assessment Template It refers to the process of rationally laying out and arranging different functional areas of a city in urban planning. It aims to optimize the spatial structure of the city, improve the city's operational efficiency, and enhance the quality of life for its residents.

[0065] Specifically, the land uses include the following: (1) Industrial land, the layout requirements are as follows: The industrial zone and residential area are conveniently connected, and employees have convenient transportation to and from get off work.

[0066] To avoid interference and pollution from industrial areas to residential areas, measures should be taken such as positioning the industrial zone upstream and downstream, establishing green protective barriers, and designating wastewater discharge points downstream of the river and at one end of the residential area.

[0067] In industrial zones, industries with large labor forces or a high proportion of female labor should be located close to residential areas. Electronics, sewing, and handicrafts should be scattered in separate residential areas, while machinery and textiles should be located in separate areas on the outskirts of the city.

[0068] The specific layout of industrial and residential areas should facilitate employees' walking commute and consider the development of public transportation routes to balance traffic load. However, industrial and residential areas should be neither too close nor too far apart to avoid one-way transportation and prevent the industrial area from encircling the city. Emphasis should be placed on transportation links within industrial areas (energy consumption: for goods with an annual transport volume of over 100,000 tons directly from the railway, dedicated railway lines should be laid or industrial marshalling yards should be established. Attention should be paid to the direction of entry lines to avoid perpendicular or direct intersections with main roads entering the industrial area. Railway freight should be close to the industrial area and distributed in several locations as needed to reduce transshipment and alleviate traffic pressure on urban roads).

[0069] Factories along the river or near the river mainly include shipyards, paper mills, timber mills, fertilizer plants, and printing and dyeing plants (pay attention to the rational use of the shoreline). Factories that mainly rely on road transportation can be located further away to avoid occupying the shoreline.

[0070] Factories along external transportation routes: These are usually located on the outskirts of the city. The intersections of roads leading to and from the factory entrances and outside the factory should be organized in a reasonable way to avoid excessive interference with external transportation.

[0071] Industrial isolation: Independent locations far from cities; chemical and metallurgical plants should be kept at a distance of 500-800 meters or more from urban areas. Factories with collaboration: Centralized location nearby reduces transshipment during production, lowers production costs, reduces pressure on urban traffic, and forms a supply chain.

[0072] Factories in the old city area: those scattered in several locations need to be relocated and consolidated, or conditions should be created for their relocation and reconstruction. Basin and canyon areas: where calm winds are frequent, it is not suitable for placing polluting industries.

[0073] (2) Residential areas have the following layout requirements: A good location is desirable for residence; avoid unfavorable conditions such as floods, earthquakes, landslides, swamps, and windy areas.

[0074] It takes up less farmland and is located close to workplaces.

[0075] When the city is small, the layout should be concentrated; when the city is large, the layout should be dispersed, leaving room for flexibility.

[0076] (3) The layout requirements for the storage area are as follows: Small cities can be located separately on the city's periphery. Large and medium-sized cities should combine centralized and decentralized approaches, avoiding excessive concentration. Dangerous goods warehouses are located in secluded areas in the suburbs to avoid transporting them through the city. The cold storage facility has an odor and contains sewage; it is located along a river in the suburbs. Vegetable warehouse: At the entrance of the main road leading from the city edge to the suburbs; Fuel and flammable material storage: located in a separate area in the suburbs, downwind or crosswind of the city during the windy season. Oil depot: away from residential areas, substations, important transportation hubs, airports, large reservoirs and hydropower stations, important bridges, medium-sized enterprises, mining areas, military facilities, preferably in low-lying areas with protective forest belts. The land use ratio should not be too large or too small; The logistics center is located on the city's outer ring road and on highways leading to other cities.

[0077] (4) Land for public facilities Commercial centers are located near the distribution center of the population they serve. Note that large cities may have more than one center, and possibly a secondary center as well. The road at the location should be classified as a major traffic artery; The company will arrange high-tech industrial zones near science and technology and higher education areas to facilitate functional integration; Public facilities that directly serve the public (libraries, science and technology museums) are located near residential areas and science and technology and higher education areas; Comply with land use quota regulations.

[0078] (5) Artistic land use Make good use of the unique natural landscapes of each city, such as highlands, hills, rivers, lakes, and water bodies, and use them as the visual focus and activity center of the overall layout to create the characteristics of various cities, such as plains, mountains, and water towns.

[0079] By utilizing mountains, rivers, lakes, scenic spots, historical sites, parks, green spaces, and historical and cultural areas, the city's aesthetics are fully reflected, forming the overall framework of the urban landscape.

[0080] The layout of roads and bridges can be well integrated with the mountains, water surfaces, and forests, creating an artistic urban landscape. Emphasis should be placed on the urban landscape positioning of the city center and main roads, creating distinctive artistic features for these areas.

[0081] An orderly and rhythmic urban axis reflects the nature and characteristics of the city.

[0082] C. Transportation System Adaptation Template The transportation system is a collective term for the infrastructure and management systems used in urban planning to ensure the efficient flow of people and goods. It includes various modes of transportation such as roads, railways, waterways, and aviation, as well as related transportation facilities and management measures.

[0083] Railways should pass through the outskirts of cities and should not fragment the city too much.

[0084] Passenger stations: Large cities may have multiple passenger stations, located on the outskirts of the central area, 2-3 km from the city center. Medium and small cities have them located on the outskirts of the urban area, using a through-passage system. Passenger stations must be connected to the city's main roads, coordinating with urban public transportation, long-distance buses, and commercial services, emphasizing intra-city transfers, and providing direct access to the city center, bus stations, docks, and subway lines that can be directly accessed by the passenger station.

[0085] Freight stations: Small and medium-sized cities have only one comprehensive freight station, while large cities have stations located on the outskirts of the city, depending on their function. Comprehensive freight stations, primarily for arrivals and departures, are located near cargo sources or in conjunction with cargo distribution centers; those not serving the city are located in the suburbs. Dangerous goods stations are located in the suburbs with designated safety isolation zones.

[0086] Marshalling yards: These allow for future expansion and are typically located at the junction of main railway lines, avoiding being surrounded by dedicated lines, freight yards, industrial areas, or warehouses. They ensure convenient access for main traffic flows and minimize corner traffic. They include passing and overtaking stations with a spacing of approximately 8-12 km.

[0087] Ports: Site selection should avoid floating timber yards, bridges, sluice gates, and water source protection areas. Deep water should be used for deep-water purposes, and shallow water for shallow-water purposes, with sufficient shoreline for residential use. A main port access road should be available. The water depth should be 10 meters to accommodate ships of 10,000 tons. Oilfield operations should be located downstream of cities, port areas, anchorages, and important bridges. Shipyards should have designated waterways and land areas.

[0088] Airports: Located on either side of the city along the prevailing wind direction, with a minimum tangent distance of 5-7 km and a transit distance of 15 km, and a distance of 10-30 km from the city center. Connected by a dedicated highway within 30 minutes. Maintaining an appropriate distance from marshalling yards. Airports should be moderately centralized rather than dispersed.

[0089] Highways: Expressways are designed for speeds of 100-120 km / h (60 km / h in mountainous areas). Mega-cities should have expressway ring roads connecting various expressways and linking them to the urban expressway network, but these should not pass through the city center. Medium and small cities should use grade-separated interchanges away from the city center, with dedicated main access roads connecting to the city. Expressways should connect to urban expressways, while general-grade highways should connect to urban arterial roads with normal-speed traffic. Mega-cities should ideally pass tangentially through densely trafficked areas, rather than penetrating deep into those areas. Highways and urban roads should form separate systems, without interference, and should not directly contact the city, but connect at designated entrances. Highways cannot be used as urban arterial roads. The highway grade and cross-sectional form must be appropriate and reasonable. Large cities should have multiple highway passenger stations located on the edge of the city center, connected to highways by urban arterial roads.

[0090] Highway passenger transport stations: One highway passenger transport station shall be set up in medium and small cities. Passenger transport hubs must have convenient connections with urban passenger transport routes and grade-separated intersections may be adopted.

[0091] Highway freight station: on the edge of the central area for daily necessities.

[0092] D. Ecological Environment Adaptation Template Environmental ecology refers to the protection and utilization of the natural environment and ecosystems in urban planning, aiming to achieve sustainable urban development and improve the quality of life for residents.

[0093] The example questions corresponding to the above professional reasoning templates are shown in Table 3: Table 3. Example Questions on Professional Reasoning Based on the Overall Territorial Spatial Planning Map

[0094] V4. Policy-related template It includes templates for associating planning indicators, templates for adapting to regulatory requirements, a query function for the association of planning indicators in the corresponding planning map, and templates for determining the compatibility with land use regulations and ecological protection regulations.

[0095] In this embodiment, taking Jiangsu Province as an example, if the fourth image (such as...) is found... Figure 2 The document shown on page 27 of the "Jiangsu Provincial Territorial Spatial Planning (2021-2035)" is a collection of the first three maps, titled "Three Control Lines Map." Based on page 27 of the "Jiangsu Provincial Territorial Spatial Planning (2021-2035)," relevant content can be found by searching the context of the maps, such as the specific content on page 18 (e.g., ...). Figure 3 As shown in the figure, the spatial foundation for Jiangsu's high-quality development has been strengthened. The "Jiangsu Provincial Territorial Spatial Planning (2021-2035)" proposes the creation of a national ecological civilization pilot zone, prioritizing ecological, agricultural, and security-related protective spaces based on bottom-line constraints and safety resilience. Following the priority order of arable land and permanent basic farmland, ecological protection red lines, and urban development boundaries, the "three zones and three lines" are comprehensively delineated: by 2035, Jiangsu Province will maintain no less than 59.77 million mu of arable land, of which no less than 53.44 million mu will be protected as permanent basic farmland; the ecological protection red line will be no less than 18,200 square kilometers, of which the marine ecological protection red line will be no less than 9,500 square kilometers; and the expansion multiple of the urban development boundary will be controlled within 1.3 times the scale of urban construction land in 2020. Strengthening the overall protection and efficient utilization of natural resources, and coordinating the allocation of emergency space for disaster prevention, mitigation, and relief, as well as major public emergencies, provides a fundamental guarantee for ensuring food security, ecological security, and energy security. "This passage is related to the theme and legend of the drawing. Therefore, relevant questions (QA) were written based on this passage."

[0096] S2. Image-text pairing generation: Acquire multi-format raw data in the fields of planning and natural resources, filter, parse and extract elements from the raw data, establish a corresponding mapping relationship between image data and text data including policy texts, planning indicators and context descriptions, and obtain standardized image-text pairing data.

[0097] In this embodiment, a custom script is written to extract planning maps from official planning documents issued by municipal governments, planning agencies, and academic institutions. Since many documents contain miscellaneous visual elements such as icons, background images, and real-world photographs, manual screening is required to retain clean planning maps. When necessary, the clarity of low-resolution images and the legibility of embedded text can be improved by sharpening or tracing back to the source file. OCR recognition technology is used to extract textual information from the original data, and semantic segmentation or layer analysis technology is used to extract professional visual elements (such as roads, green spaces, and plots of land) from the planning maps, thereby automatically establishing a mapping relationship between image data and corresponding policy texts, planning indicators, and contextual descriptions. S3. Template-driven VQA sample generation: Multimodal large models (such as Qwen-VL, GPT-4V, etc.) are called to generate candidate question-answer pairs based on the VQA question template library and image-text pairing data. Then, the candidate question-answer pairs are filtered according to the planned industry professional semantics (industry templates). The accuracy of the question-answer expression is corrected by the industry technical requirements in the template library, thereby outputting the initial VQA training samples containing image-question-answer.

[0098] S4. Automated Quality Control Optimization: The questions and answers in the initial VQA training samples above are automatically detected for semantic consistency and reviewed by experts for logical rationality. Finally, they are classified and labeled according to question type, difficulty and four-level cognitive task system to obtain a high-quality standardized corpus package.

[0099] S5. Scalable Corpus Output: Outputs standardized corpus in a preset format. This standardized corpus includes image IDs, question text, answer text, task type, and annotation source information. (Example follows) Figure 4 As shown, it is used for fine-tuning of multimodal large models in the fields of planning and natural resources.

[0100] Additionally, the planning and natural resources industry standards and professional terms in the VQA problem template library of this embodiment can be replaced to adapt to the fine-tuning of multimodal large models in the fields of architectural design and ecological restoration.

[0101] Example 2 This embodiment 2 proposes a multimodal large-scale model fine-tuning corpus production system for the planning and natural resources field, used to implement the above-mentioned fine-tuning corpus production method. The system includes: Data input module: Receives raw data in multiple formats from the planning domain; Task and Template Building Module: Defines four-level progressive cognitive tasks and builds an industry-specific VQA question template library; Image-text matching module: Filters and optimizes raw data, extracts text and professional visual elements, and establishes accurate cross-modal mapping; Sample generation module: calls the multimodal large model to generate and filter candidate question-answer pairs, and outputs initial VQA samples; Quality control module: This module optimizes samples through automatic detection and professional review to complete classification and labeling. Scalable output module: Outputs standardized corpora in a structured format, supporting cross-domain template replacement and adaptation.

[0102] The above are merely preferred embodiments of the present invention and do not constitute any limitation on the present invention. Any equivalent substitutions or modifications made by those skilled in the art to the technical solutions and content disclosed in the present invention without departing from the scope of the present invention shall be deemed to have remained within the protection scope of the present invention.

Claims

1. A method for producing a multimodal large-scale model fine-tuning corpus for the fields of planning and natural resources, characterized in that, Includes the following steps: S1. Construct industry-specific cognitive tasks and VQA template library: Based on industry standards, professional examination systems and planning text and image tasks in the planning and natural resources field, define a four-level progressive cognitive task system covering perception, reasoning, association and application. At the same time, construct a VQA question template library corresponding to each cognitive level, including planning map classification and planning map element recognition, spatial relationship, professional reasoning and policy association. S2. Image-text pairing generation: Obtain multi-format raw data in the fields of planning and natural resources, filter, parse and extract elements from the raw data, establish a corresponding mapping relationship between image data and text data including policy text, planning indicators and context descriptions, and obtain standardized image-text pairing data; S3. Template-driven VQA sample generation: The multimodal large model is called to generate candidate question-answer pairs based on the VQA question template library and image-text pairing data. The candidate question-answer pairs are screened according to the planning industry professional semantics. The accuracy of the question-answer expression is corrected by the industry technical requirements in the template library. The initial VQA training samples containing image-question-answer are output. S4. Automated Quality Control Optimization: The initial VQA training samples are subjected to automatic semantic consistency detection and expert review of logical rationality. Finally, they are classified and labeled according to question type, difficulty and four-level cognitive task system to obtain high-quality standardized corpus. S5. Scalable Corpus Output: Output the standardized corpus in a preset format. The standardized corpus includes image ID, question text, answer text, task type, and annotation source information, which is used for fine-tuning of multimodal large models in the planning and natural resources field.

2. The method for producing a multimodal large-scale model fine-tuning corpus in the field of planning and natural resources according to claim 1, characterized in that, By replacing industry standards and professional terms in the VQA issue template library, we can adapt and fine-tune multimodal large models in the fields of architectural design and ecological restoration.

3. The method for producing a multimodal large-scale model fine-tuning corpus in the field of planning and natural resources according to claim 1, characterized in that, In step S1, the four cognitive levels specifically include: Perception layer: This includes element recognition and image description. The element recognition is used to evaluate the model's ability to recognize the layout structure, text annotations, basic geographic features and drawing elements in the planning map, and to establish semantic alignment between image content and natural language. The image description is used to extract details based on the image itself and generate an objective description. Reasoning levels include planning map classification, spatial relationship reasoning, and professional reasoning. The planning map classification is based on the "five-level, three-category" planning system, dividing the planning map into master plan, detailed plan, and special plan. The spatial relationship reasoning includes topological spatial relationship reasoning, sequential spatial relationship reasoning, and metric spatial relationship reasoning. The professional reasoning includes four dimensions: spatial layout, functional organization, transportation system, and ecological environment. Association hierarchy: used to evaluate the model's ability to associate planning maps with underlying policies, regulations, and planning indicators, and to establish cross-domain associations between spatial elements and corresponding policy frameworks; Application level: This includes scheme evaluation and decision making. Scheme evaluation is used to assess the merits of planning schemes based on visual input and spatial context. Decision making is used to make value- and principle-based choices in constrained design situations.

4. The method for producing a multimodal large-scale model fine-tuning corpus in the field of planning and natural resources according to claim 1, characterized in that, In step S1, the VQA question template library specifically includes: Planning map classification templates: including classification and determination templates for provincial basic analysis maps, provincial planning result maps, municipal survey maps, municipal control maps, and municipal schematic maps; Spatial Relationship Templates: These include templates for determining topological spatial relationships, querying sequential spatial relationships, and comparing metric spatial relationships. They also provide standardized problem descriptions for determining adjacent / containment / intersection relationships between geographic entities, querying directional distribution, and comparing distances. Professional reasoning templates include spatial layout analysis templates, functional organization assessment templates, transportation system adaptation templates, and ecological environment adaptation templates, which correspond to the problem templates for analyzing the characteristics and causes of urban layout, assessing the rationality of functional area planning, analyzing the support of transportation facility layout for economic and social development, and analyzing the guarantee of ecological protection planning for sustainable development. Policy-related templates: These include templates for linking planning indicators, templates for adapting to regulatory requirements, a query function for linking planning indicators in the corresponding planning map, and templates for determining compatibility with land use regulations and ecological protection regulations.

5. The method for producing a multimodal large-scale model fine-tuning corpus for planning and natural resources as described in claim 1, characterized in that, Step S2 specifically includes: extracting and filtering valid planning maps through a custom script, extracting text information from the original data using OCR recognition technology, and extracting professional visual elements from the planning maps through semantic segmentation or layer parsing technology, thereby establishing a mapping relationship between image data and corresponding policy texts, planning indicators, and contextual descriptions; The elements of the planning map include administrative boundaries, government seats, annotations, scale, legend, contour lines, and water system; The visual elements include roads, green spaces, plots of land, and transportation facilities; The multi-format raw data includes planning PDF files, JPG image files, and DOC text files; The original data comes from official planning documents issued by municipal governments, planning agencies, and academic institutions.

6. The method for producing a multimodal large-scale model fine-tuning corpus in the field of planning and natural resources according to claim 1, characterized in that, In step S3, the screening process specifically involves eliminating candidate question-and-answer pairs that do not conform to the professional terminology standards of the planning industry or deviate from the technical requirements of planning; the correction process includes supplementing planning industry-specific expressions and adjusting the consistency between the question-and-answer logic and the professional logic of planning.

7. The method for producing a multimodal large-scale model fine-tuning corpus for planning and natural resources as described in claim 1, characterized in that, In step S4, the semantic consistency detection model is implemented through a pre-trained semantic matching model for the planning industry; the review content of the logical rationality expert review includes whether the question-answer pair conforms to the planning industry technical specifications, whether it accurately reflects the core information of the planning map, and whether it conforms to the requirements of the policy text.

8. The method for producing a multimodal large-scale model fine-tuning corpus for planning and natural resources as described in claim 3, characterized in that, The spatial layout in the professional reasoning includes centralized layout and decentralized layout; the centralized layout includes grid-like and ring-radial layouts; the decentralized layout includes cluster-like, strip-like, star-like, ring-like, satellite-like, multi-center, and cluster city layouts; The functional organization in the professional reasoning includes industrial land, residential areas, warehousing areas, and public facilities land. Among them, the layout of industrial land must meet the requirements of pollution isolation from residential areas and the requirements of optimized transportation links, while the layout of warehousing areas must meet the requirements of hazardous materials isolation and convenient logistics transportation. The transportation system in the professional reasoning includes the site selection and layout assessment of railways, highways, ports, and airports, which must meet the adaptability requirements of transportation facilities and urban space in the planning industry. The ecological and environmental reasoning in the professional reasoning includes the assessment of the protection and utilization of the natural environment and ecosystems, which must comply with the requirements of ecological protection red lines and nature reserve planning.

9. The method for producing a multimodal large-scale model fine-tuning corpus for planning and natural resources as described in claim 3, characterized in that, The planning indicators in the associated hierarchy include indicators for cultivated land area, permanent basic farmland protection area, ecological protection red line area, and urban development boundary expansion multiple. The policy framework includes territorial spatial planning outline documents, industry technical regulations, and related legal documents.

10. A system for producing a multimodal large-scale model fine-tuning corpus for the fields of planning and natural resources, characterized in that, include: Data input module: Receives raw data in multiple formats from the planning domain; Task and Template Building Module: Defines four-level progressive cognitive tasks and builds an industry-specific VQA question template library; Image-text matching module: Filters and optimizes raw data, extracts text and professional visual elements, and establishes accurate cross-modal mapping; Sample generation module: calls the multimodal large model to generate and filter candidate question-answer pairs, and outputs initial VQA samples; Quality control module: This module optimizes samples through automatic detection and professional review to complete classification and labeling. Scalable output module: Outputs standardized corpora in a structured format, supporting cross-domain template replacement and adaptation.

Images