Intelligent checking system and method for cultivated land quality evaluation map based on multi-modal model

Abstract

The application discloses a cultivated land quality evaluation map intelligent checking system and method based on a multi-modal model. The system comprises: a front-end interaction module for receiving a cultivated land quality evaluation map and extracting metadata; an image preprocessing module for standardization processing; a multi-modal model service module for automatically matching and calling a single or multiple visual understanding models as an adaptive model according to the map type and complexity of the map through a model scheduler, and converting structured checking rules in a checking rule library into visual task instructions executable by the adaptive model for map element identification; a dynamic knowledge base module for carrying out normative checking based on the identification result in combination with the rules in the dynamic knowledge base to obtain a checking result; and a checking result classification module for grade division, display of defect position screenshots and provision of targeted repair based on the checking result. The application can improve the standardization level and work efficiency of cultivated land quality evaluation map checking.

Description

Technical Field

[0001] This invention relates to the field of image inspection and processing technology, and in particular to an intelligent inspection system and method for farmland quality evaluation maps based on a multimodal model. Background Technology

[0002] As a crucial component of the evaluation results, the quality of farmland quality assessment maps directly impacts the scientific validity and credibility of the assessment outcomes. According to the "Farmland Quality Grading" (GB / T 33469-2016) and related technical regulations, farmland quality assessment maps must meet multiple technical requirements, including completeness of results, accuracy of mathematical foundations, integrity of map elements, and standardization of layout. Currently, the following technical solutions are primarily used for the review of farmland quality assessment maps:

[0003] 1. Manual visual inspection mode

[0004] In existing technologies, map review mainly relies on manual visual inspection by professional technicians. Reviewers need to check the completeness and standardization of elements such as coordinate system, scale, legend, north arrow, farmland boundaries, and quality grade zoning on the map one by one. This method is highly intuitive, but it has technical drawbacks such as low review efficiency, strong subjectivity, and the possibility of missed inspections due to visual fatigue.

[0005] 2. Automated tools based on image processing

[0006] Some existing technologies employ automated image processing tools for map verification, identifying map elements through techniques such as edge detection, template matching, and feature extraction. While these tools can achieve a certain degree of automation, they have significant shortcomings in semantic understanding of complex map elements, spatial relationship analysis, and professional standard compliance judgment.

[0007] 3. Traditional computer vision technology

[0008] Some technical solutions attempt to incorporate traditional computer vision techniques, such as Histogram of Oriented Gradients (HOG), Scale Invariant Feature Transform (SIFT), and image classification, for map analysis. While these techniques possess basic image recognition capabilities, their recognition accuracy and adaptability are limited when faced with the complex semantics, multi-scale features, and domain-specific elements of professional maps.

[0009] 4. Application of general visual understanding models

[0010] With the development of deep learning, some existing technologies attempt to directly call general visual understanding models (such as CNN, ViT, etc.) for map understanding. However, general models lack professional knowledge in the field of farmland quality assessment, have difficulty accurately understanding the specific requirements of technical specifications such as "Farmland Quality Grade" for maps, and have weak semantic parsing capabilities for professional legends, annotations, and spatial layouts.

[0011] A comprehensive analysis of the above technical solutions reveals the following main shortcomings of existing technologies in the intelligent verification of farmland quality evaluation maps:

[0012] 1. Insufficient depth of visual understanding: Existing technologies mostly remain at the pixel level or shallow feature processing, making it difficult to achieve deep semantic understanding of the content of the map and to accurately identify the professional meaning of the map elements and their relationships.

[0013] 2. Difficulty in identifying professional elements: Farmland quality evaluation maps contain a large number of professional elements (such as farmland boundaries, quality grade zones, evaluation index markings, etc.). Existing technologies cannot accurately identify these elements, let alone determine whether they meet technical specifications.

[0014] 3. Limited multi-scale feature processing capabilities: Map verification needs to consider both the overall layout and local details. Existing methods are ineffective in handling the collaborative analysis of the macroscopic structure and microscopic annotations of maps, and are prone to neglecting one aspect for the other.

[0015] 4. Lack of spatial relationship analysis capability: The spatial positional relationships between map elements (such as the correspondence between hierarchical zoning and legend, and the relative position of labels and polygons) are important aspects of the verification. Existing technology lacks effective spatial relationship reasoning capabilities, making it difficult to assess the rationality of the layout.

[0016] 5. Poor adaptability to professional standards: The general visual understanding model is not optimized for the field of farmland quality evaluation, and the accuracy of judging compliance with technical standards is low; while rule-based tools require a large number of rules to be defined manually, which is difficult to adapt to the needs of changing map types and standard updates, resulting in high maintenance costs.

[0017] 6. Insufficient interpretability: Existing automated tools often output "black box" conclusions, lacking visual evidence and explanations of reasoning processes, making it difficult for auditors to understand the basis for inspection and making review and traceability difficult.

[0018] 7. A significant conflict exists between processing efficiency and accuracy: manual review is highly accurate but inefficient, making it difficult to scale; automated tools, while highly efficient, lack sufficient accuracy. Current technologies cannot achieve both high efficiency and high accuracy simultaneously, making it difficult to balance the two.

[0019] In summary, in the field of intelligent verification of farmland quality assessment maps, there is an urgent need for an intelligent verification method and system that can achieve a unified understanding of macro-structure and micro-annotation, possess intelligent visual analysis capabilities for professional maps, accurately judge the rationality of element layout, and accurately understand the requirements of technical specifications such as "Farmland Quality Grades," in order to fill the existing technological gaps and meet the actual needs of farmland protection work. Summary of the Invention

[0020] To address the aforementioned technical problems, this invention proposes an intelligent verification system and method for farmland quality evaluation maps based on a multimodal model. The system and method achieve intelligent processing of the entire farmland quality evaluation map process, from uploading, preprocessing, multimodal analysis to rule verification, result grading, and feedback, significantly improving the standardization level and work efficiency of farmland quality evaluation map verification.

[0021] To achieve the above objectives, the technical solution of the present invention is as follows:

[0022] An intelligent verification system for farmland quality assessment maps based on a multimodal model includes:

[0023] The front-end interaction module is used to receive farmland quality evaluation maps uploaded by users and extract the metadata of the farmland quality evaluation maps.

[0024] The image preprocessing module is used to standardize the uploaded farmland quality evaluation maps;

[0025] The multimodal model service module is used to automatically match and call one or more corresponding visual understanding models as adaptation models based on the map type and complexity of the farmland quality evaluation map through the model scheduler, and convert the structured inspection rules in the inspection rule base into visual task instructions that the adaptation model can execute to identify map elements of the preprocessed farmland quality evaluation map and obtain the recognition results.

[0026] The dynamic knowledge base module is used to construct a verification rule base based on the "Cultivated Land Quality Grade" (GB / T 33469-2016) and corresponding technical regulations; it is used to conduct multi-dimensional standardization verification of the map based on the identification results and in combination with the rules in the dynamic knowledge base to obtain the verification results.

[0027] The inspection result classification module is used to classify the inspection results into grades, display screenshots of defect locations, and provide targeted repair suggestions.

[0028] Preferably, the front-end interaction module includes:

[0029] The image preview component is used to receive farmland quality evaluation images and provide interactive operations, including zooming, panning, rotating and contrast adjustment.

[0030] The metadata automatic extraction unit is used to extract and parse the metadata of the farmland quality evaluation map and present it in real time through the display panel. The metadata includes file format, resolution, size, color mode, projection parameters, scale information, and geographic coordinates.

[0031] Batch processing queue is used for parallel uploading of multiple images and tracking of inspection progress.

[0032] Preferably, the image preprocessing module includes:

[0033] The resolution standardization unit is used to uniformly adjust farmland quality evaluation maps of different resolutions to a preset standard size;

[0034] The color correction and enhancement unit is used for unified color space conversion and image enhancement.

[0035] The format conversion and encoding unit is used to provide mutual conversion between different image formats and convert image data into the Base64 encoding format required by the visual understanding model;

[0036] The geographic information parsing unit is used to format and store the metadata of the extracted farmland quality evaluation map.

[0037] Preferably, the multimodal model service module includes:

[0038] The model scheduler is used to dynamically select one or more visual understanding models as the adaptation model based on the type and complexity of the graph.

[0039] A multimodal cue word builder is used to convert structured check rules in the check rule base into visual task instructions that can be executed by the adapted model;

[0040] The result parser is used to parse the unstructured results output by the adaptation model into structured JSON format recognition results, which include map feature categories, location coordinates, and confidence scores.

[0041] A confidence evaluator is used to quantitatively assess the reliability of the identification results.

[0042] Preferably, the check rule base includes:

[0043] The completeness rule for results is used to check whether the necessary element layers are included.

[0044] Mathematical foundational rules are used to verify the accuracy of geographic coordinates, projection parameters, and scales;

[0045] Map feature rules are used to check the integrity of map features, which include geographic features, label features, and professional features.

[0046] The finishing and finishing rules are used to verify the standardization of drawing layout, font style, annotation and color matching;

[0047] Data consistency rules are used to check the consistency between farmland quality evaluation maps and text reports and tabular data.

[0048] Preferably, the dynamic knowledge base module is also used to periodically or irregularly add, modify, abolish, or optimize the inspection rule base based on changes in the business operating environment, technical specifications, and historical data feedback.

[0049] Preferably, the inspection result classification module includes:

[0050] The results grading unit is used to classify issues into grades based on severity and confidence scores, including compliance, defects, and risks.

[0051] The problem localization unit is used to accurately locate problem areas at the pixel level on the original farmland quality evaluation map using heat maps and bounding boxes.

[0052] The evidence screenshot generation unit is used to automatically capture local images of the problem area and present them as visual evidence along with the inspection conclusions.

[0053] The report generation unit is used to generate a comprehensive inspection report that includes a list of issues, grading results, location coordinates, visual evidence, repair suggestions, and confidence scores. It supports online viewing and export.

[0054] Based on the above, this invention also discloses an intelligent verification method for farmland quality evaluation maps based on a multimodal model, comprising the following steps:

[0055] Receive farmland quality assessment maps uploaded by users and extract the metadata of the farmland quality assessment maps;

[0056] Standardize the farmland quality evaluation maps;

[0057] Several visual understanding models are integrated. Based on the map type and complexity of the farmland quality evaluation map, the model scheduler automatically matches and calls the corresponding single or multiple visual understanding models as the adaptation models. The structured check rules in the check rule base are converted into visual task instructions that the adaptation models can execute to identify map elements in the preprocessed farmland quality evaluation map and obtain the recognition results.

[0058] Based on the "Grading of Cultivated Land Quality" (GB / T 33469-2016) and corresponding technical specifications, a verification rule base is constructed. On the basis of the identification results, combined with the rules in the dynamic knowledge base, multi-dimensional standardization verification of cultivated land quality evaluation maps is carried out to obtain verification results.

[0059] Based on the inspection results, the system classifies the defects, displays screenshots of the defect locations, and provides targeted repair suggestions.

[0060] Preferably, the method further includes the following steps:

[0061] Based on changes in the business operating environment, technical specifications, and historical data feedback, the inspection rule base is regularly or irregularly updated with additions, modifications, repeals, or optimizations.

[0062] Preferably, the method further includes the following steps:

[0063] Save the recognition results and verification results of the visual understanding model's reasoning;

[0064] Based on the identification results and verification results, an inspection history archive is established, and based on the inspection history archive, result comparison, trend analysis, version management and audit tracking functions are provided.

[0065] Based on the above technical solution, the beneficial effects of the present invention are:

[0066] 1) Deep visual understanding capability: By introducing an advanced multimodal large model, this invention achieves a leap from pixel-level processing to semantic-level understanding, enabling deep understanding of map content, accurate identification and analysis of the professional meaning of various map elements, surpassing traditional rule-based simple inspection methods.

[0067] 2) Deep adaptation to professional fields: It is specifically optimized for the characteristics of maps in the field of farmland quality evaluation, and integrates professional standard knowledge such as "Farmland Quality Grade". It can effectively handle professional content such as farmland boundaries, quality grade zoning, and evaluation indicator labeling, which significantly improves the accuracy of verification and field adaptability.

[0068] 3) Multi-scale collaborative analysis capability: It supports multi-level visual understanding from the overall layout of the map to the annotation of local details. It can collaboratively analyze the rationality of the macro structure and the standardization of micro elements within the same framework, overcoming the technical defects of traditional methods that focus on one aspect while neglecting another.

[0069] 4) Spatial relationship reasoning ability: Possesses the ability to analyze the spatial positional relationship between map elements, and can accurately judge whether the layout of elements meets the standard requirements, such as the correspondence between legend and map patch, the relative position of label and geographic elements, etc., thus realizing the effective assessment of the rationality of the layout.

[0070] 5) Strong interpretability and traceability: Each inspection conclusion of this invention is accompanied by visual evidence such as pixel-level location markings and screenshots of the problem area, and the reasoning process is presented in a structured manner, making the inspection results transparent and understandable, greatly improving the credibility and reproducibility of the review conclusions, and making it easier for review experts to quickly confirm the problems.

[0071] 6) Excellent scalability and adaptability: The system adopts a modular design, decoupling the dynamic knowledge base from the multimodal model service. When technical standards are updated or new map types are added, only the rule configuration in the knowledge base needs to be updated or the adapted visual understanding model needs to be replaced, without refactoring the core system code, effectively ensuring the long-term adaptability and low maintenance cost of the technical solution. Attached Figure Description

[0072] Figure 1 This is a schematic diagram of the structure of an intelligent inspection system for farmland quality evaluation maps based on a multimodal model in one embodiment;

[0073] Figure 2 This is a schematic diagram of the intelligent verification method for farmland quality evaluation maps based on a multimodal model in one embodiment. Detailed Implementation

[0074] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0075] like Figure 1 As shown, this embodiment provides an intelligent verification system for farmland quality evaluation maps based on a multimodal model, including:

[0076] 1. Front-end interaction module: This module receives farmland quality assessment maps uploaded by users, supporting uploads and previews of various image formats including GeoTIFF, JPG, PNG, and PDF. It provides map thumbnail display, batch upload of multiple maps, and drag-and-drop upload functionality. It also integrates an automatic map metadata extraction function, capable of recognizing map resolution, color mode, file format, and geographic coordinates, including:

[0077] The image preview component is used to receive farmland quality evaluation images and provide interactive operations, including zooming, panning, rotating and contrast adjustment.

[0078] The metadata automatic extraction unit is used to extract and parse the metadata of the farmland quality evaluation map and present it in real time through the display panel. The metadata includes file format, resolution, size, color mode, projection parameters, scale information, and geographic coordinates.

[0079] Batch processing queue is used for parallel uploading of multiple images and tracking of inspection progress.

[0080] 2. Image preprocessing module, used to standardize uploaded farmland quality assessment maps, eliminate differences in image sources, and provide high-quality input data for subsequent model analysis, including:

[0081] The resolution standardization unit is used to uniformly adjust farmland quality evaluation maps of different resolutions to a preset standard size (such as 2048×2048 pixels) to ensure consistent input size.

[0082] The color correction and enhancement unit is used for unified color space conversion and image enhancement.

[0083] The format conversion and encoding unit is used to provide mutual conversion between different image formats and convert image data into the Base64 encoding format required by the visual understanding model;

[0084] The geographic information parsing unit is used to format and store the metadata of the extracted farmland quality evaluation map, providing a basis for mathematical verification.

[0085] 3. The multimodal model service module integrates multiple advanced visual understanding models. Based on the map type and complexity of the farmland quality assessment map, it automatically matches and calls one or more corresponding visual understanding models as adaptation models through a model scheduler. It also converts the structured check rules in the check rule base into visual task instructions executable by the adaptation models to perform map element recognition, text content extraction, and spatial relationship analysis on the preprocessed farmland quality assessment map, obtaining the recognition results. This includes:

[0086] The model scheduler is used to dynamically select one or more visual understanding models as the adaptation model based on the type and complexity of the graph.

[0087] A multimodal cue word builder is used to convert structured check rules in the check rule base into visual task instructions that can be executed by the adapted model;

[0088] The result parser is used to parse the unstructured results output by the adaptation model into structured JSON format verification results. These verification results include map feature categories, location coordinates, and confidence scores. The map features include:

[0089] 1) Geographical elements: basic geographical spatial elements such as farmland boundaries, water systems, roads, and settlements;

[0090] 2) Labeling elements: map title, legend, scale, north arrow, grid lines, connecting tables, and other auxiliary map elements;

[0091] 3) Professional elements: professional elements such as the boundaries and color blocks of cultivated land quality grade zones, sampling points of evaluation indicators, soil type zones, and annotations.

[0092] The confidence level evaluator is used to quantitatively assess the reliability of inspection results and provide a reference for the reliability of inspection conclusions.

[0093] 4. Dynamic Knowledge Base Module: This module is used to construct an inspection rule base based on the "Cultivated Land Quality Grades" (GB / T 33469-2016) and corresponding technical regulations. This inspection rule base adopts a structured design, supports dynamic updates and version management, and includes:

[0094] The rules for completeness of results are used to check whether essential element layers are included. These essential element layers refer to the thematic layers that must be included in the farmland quality evaluation map results according to the requirements of "Farmland Quality Grades" (GB / T 33469-2016) and corresponding technical regulations. They are the core components of the completeness of the evaluation results and usually include farmland quality grade distribution maps, evaluation index sampling point maps, etc.

[0095] Mathematical foundational rules are used to verify the accuracy of geographic coordinates, projection parameters, and scales;

[0096] Map element rules are used to check the integrity of map elements such as north arrow, legend, scale bar, and map title;

[0097] The finishing and finishing rules are used to verify the standardization of drawing layout, font style, annotation and color matching;

[0098] Data consistency rules are used to check the consistency between farmland quality evaluation maps and text reports and tabular data.

[0099] 5. Inspection result classification module, used to classify the inspection results into grades, display screenshots of defect locations, and provide targeted repair suggestions, including:

[0100] The results grading unit is used to classify issues into levels based on the severity and confidence level of the problem. The levels include three categories: compliance, defect, and risk, which helps users focus on addressing key issues.

[0101] The problem localization unit is used to accurately locate problem areas at the pixel level on the original farmland quality evaluation map using heat maps and bounding boxes.

[0102] The evidence screenshot generation unit is used to automatically capture local images of the problem area and present them as visual evidence along with the inspection conclusions.

[0103] The report generation unit is used to generate a comprehensive inspection report that includes a list of issues, grading results, location coordinates, visual evidence, repair suggestions, and confidence scores. It supports online viewing and export.

[0104] In the above embodiments, each module of the intelligent verification system for farmland quality evaluation maps based on multimodal models can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0105] like Figure 2 As shown, in one embodiment, an intelligent verification method for farmland quality evaluation maps based on a multimodal model is also provided for the standardization review of county-level farmland quality grade evaluation maps, including the following steps:

[0106] Step 1: Receive the farmland quality evaluation map uploaded by the user and extract the metadata of the farmland quality evaluation map.

[0107] In this embodiment, the user uploads a 2023 farmland quality evaluation map of a county (GeoTIFF format, 300dpi resolution, approximately 8000×6000 pixels) through the system's web interface. The front-end interaction module automatically recognizes the image format and calls the metadata automatic extraction unit to read and parse key metadata such as geographic coordinate information (CGCS2000 coordinate system, central meridian 117°), projection parameters (Gauss-Kruger projection), and scale (1:50000) from the GeoTIFF file, and presents it in real time through the display panel. The system also extracts basic information such as map resolution, color mode (RGB), and file size.

[0108] Step 2: Standardize the farmland quality evaluation map.

[0109] In this embodiment, the image preprocessing module performs standardization processing on the uploaded farmland quality evaluation map: First, the resolution standardization unit compresses the large-format map proportionally to the standard resolution (2048×2048 pixels) required by the model input, while maintaining the original aspect ratio; Second, the color correction and enhancement unit performs a unified color space conversion and automatically corrects any color cast that may exist in the scanned image, using an adaptive histogram equalization algorithm to improve the image contrast and clarity; Finally, the format conversion and encoding unit converts the image into the Base64 encoding format required by the multimodal model, while the metadata extracted by the geographic information parsing unit is formatted and stored for subsequent mathematical basis verification.

[0110] Step 3: Integrate several visual understanding models. Based on the map type and complexity of the farmland quality evaluation map, the model scheduler automatically matches and calls one or more corresponding visual understanding models as adaptation models. The structured verification rules in the verification rule base are converted into visual task instructions that the adaptation models can execute to identify map elements in the preprocessed farmland quality evaluation map and obtain the recognition results.

[0111] In this embodiment, the multimodal model service module receives the preprocessed map data and generates a visual understanding instruction through the multimodal cue word builder: "Please identify the following elements in this farmland quality evaluation map: 1. Auxiliary elements such as map title, legend, scale, and north arrow; 2. Basic geospatial elements such as farmland boundaries, water systems, roads, and settlements; 3. Professional elements such as farmland quality grade zoning boundaries and color blocks, evaluation index sampling points, soil type zoning, and annotations. Please output the recognition results in JSON format, including element category, location coordinates (boundary box), and recognition confidence." The model scheduler selects Gemini Vision Pro to execute the task based on the map complexity. The results parser parses the natural language results returned by the model into structured data, identifying: the map title is located in the center of the top of the map (confidence 0.98), containing "Distribution Map of Cultivated Land Quality Grades in XX County in 2023"; the legend is located in the lower right corner of the map (confidence 0.95), containing color blocks labeled with land quality grades 1-10; the north arrow is located in the upper left corner of the map (confidence 0.92); and a total of 7 grade color block areas for cultivated land quality grade are identified, with precise location coordinates.

[0112] Step 4: Construct a verification rule base according to the "Grading of Cultivated Land Quality" (GB / T 33469-2016) and corresponding technical regulations; based on the identification results, combine the rules in the dynamic knowledge base to conduct multi-dimensional standardization verification of cultivated land quality evaluation maps and obtain verification results.

[0113] In this embodiment, the recognition results are verified in multiple dimensions based on the rules for completeness of results, mathematical foundation, map element, finishing specifications, and data consistency in the dynamic knowledge base, as follows:

[0114] Completeness verification of results: Upon checking the essential element layers, it was found that the "sampling point distribution map" was not included in the main map as an appendix. However, according to the specifications, county-level evaluation results should include a sampling point distribution map, which was determined to be a defect in the completeness of the appendix.

[0115] Mathematical basis verification: Comparing the metadata extracted in step one with the rule base requirements (county-level maps should use the CGCS2000 coordinate system, 1:50,000 scale), it was found that the coordinate system meets the requirements, but the metadata lacks a detailed description of the projection parameters, indicating that the mathematical basis description is incomplete.

[0116] Map element verification: The inspection found that the north arrow was in the correct position but its style did not conform to the specifications (the wind rose style is required, but it is actually a single arrow), which was determined to be a defect in the standardization; the legend lacked an explanation of the "construction land" plot, which was determined to be an incomplete legend.

[0117] Finishing standard verification: Analysis of the color blocks in the grade zones revealed that the color blocks in the fifth grade (which should be light orange according to standard) appeared reddish on the map, which differed from the standard color spectrum, indicating that the use of color was not standardized.

[0118] Data consistency verification: This is used to check the consistency between farmland quality evaluation maps and text reports / table data, including: the consistency between the area proportion of each grade identified on the map and the statistical data in the report, the consistency between the grade distribution pattern and the text description, the consistency between the attribute labeling of sampling points and the data table, the consistency between the map metadata and the report description, and the spatial correspondence consistency of the same element among multiple maps.

[0119] Step 5: Based on the inspection results, classify the defects into levels, display screenshots of the defect locations, and provide targeted repair suggestions.

[0120] The inspection results classification and display module summarizes all findings and generates a comprehensive inspection report. The report includes: an overall score (85 points out of 100), inspection status (requires review), and a list of issues. Each issue entry includes: an issue description, a classification label (defect / risk), pixel-level location annotations on the original map (presented as a heatmap or bounding box), screenshot evidence of the issue area, a confidence score (0-1), and specific remediation suggestions (e.g., "It is recommended to change the compass style to a standard wind rose diagram"). The report is output in an interactive HTML format, supports map zooming to view issue details, and can be exported as a PDF archive.

[0121] Comparative experiments have verified that this embodiment demonstrates significant advantages in the following aspects: In terms of processing speed, the entire process of checking a single high-resolution farmland quality evaluation map takes approximately 45 seconds, while manual visual inspection takes an average of 25-30 minutes; in terms of problem detection rate, the system detects minor issues such as inconsistent compass styles, missing legend content, and color discrepancies, among which the color discrepancy issue was confirmed in subsequent manual review to be a color configuration error during the mapping process; in terms of accuracy, the system's element recognition accuracy rate reaches 94%, which is highly consistent with the results of expert manual interpretation.

[0122] In one embodiment of the intelligent verification method for farmland quality evaluation maps based on a multimodal model, the following steps are also included:

[0123] Based on changes in the business operating environment, technical specifications, and historical data feedback, the inspection rule base is regularly or irregularly updated with additions, modifications, repeals, or optimizations.

[0124] In one embodiment of the intelligent verification method for farmland quality evaluation maps based on a multimodal model, the following steps are also included:

[0125] Save the recognition results and verification results of the visual understanding model's reasoning;

[0126] Based on the identification results and verification results, an inspection history archive is established, and based on the inspection history archive, result comparison, trend analysis, version management and audit tracking functions are provided.

[0127] It should be understood that although the steps in the flowchart above are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart above may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0128] In one embodiment, a method for intelligent verification of farmland quality evaluation maps based on a multimodal model is also provided for batch verification and comparative analysis of multiple farmland quality evaluation results at the municipal level. Specifically:

[0129] This embodiment focuses on batch verification of a set of farmland quality evaluation maps (12 maps in total, including thematic maps such as grade distribution maps, sampling point distribution maps, and soil organic matter distribution maps) covering six districts and counties in a prefecture-level city. The emphasis is on consistency verification between maps and cross-map comparative analysis, placing higher demands on batch processing efficiency and correlation analysis capabilities.

[0130] Parallel processing mechanism: The batch processing queue schedules 12 inspection tasks to run simultaneously, the image preprocessing module adopts multi-threaded parallel processing, and the model scheduler dynamically allocates computing resources according to the complexity of each image, with the overall processing time controlled within 8 minutes.

[0131] Inter-map correlation analysis: Establish an inter-map correlation analysis mechanism to perform spatial location registration and attribute comparison for the same geographical feature (such as the same administrative boundary or the same sampling point location) in different maps. For example, overlay analysis can be performed on the township boundaries in the hierarchical distribution map and the sampling point locations in the sampling point distribution map.

[0132] Cross-map consistency verification: Design a special verification task to compare the farmland grade distribution of a certain township in the grade distribution map with the grade data reflected by the sampling points of the same township in the sampling point distribution map, and identify potential problems of insufficient sampling representativeness.

[0133] The system successfully detected the following issues: The "First-Class" land quality distribution area shown on the grade distribution map of a certain district has no corresponding sampling points covered in the sampling point distribution map, indicating a risk of insufficient evaluation data support; the descriptions of farmland quality grades on both sides of the boundary between two adjacent districts show abrupt changes in grade, lacking a buffer transition and inconsistent with geographical continuity; the correlation between the soil organic matter distribution map and the corresponding area's color block grade on the grade distribution map is weak, suggesting a possible abnormal setting of evaluation indicator weights. The generated comparative analysis report presents the consistency issues between the various maps in matrix form, providing a comprehensive basis for the overall quality control of municipal-level results.

[0134] In one embodiment, an intelligent verification method for farmland quality evaluation maps based on a multimodal model is also provided for specialized verification of professional elements in farmland quality grade zoning.

[0135] This embodiment focuses on the core professional element in farmland quality evaluation maps—quality grade zoning. It emphasizes the accuracy of grade boundary delineation, the standardization of grade labeling, the rationality of grade transitions, and the consistency between grade area statistics and map representation. The system employs enhanced visual analysis technology, combined with a domain knowledge base, to achieve a deep understanding of professional elements. This domain knowledge base refers to a collection of professional knowledge in the field of farmland quality evaluation, covering professional knowledge in areas such as farmland quality grade classification standards, evaluation indicator systems, soil geography laws, land use classification rules, evaluation unit division principles, and cartographic standards. It provides professional semantic support for the visual understanding model, assisting in achieving a deep understanding and intelligent judgment of professional elements. Specifically, it may include: farmland quality grade classification standards and evaluation indicator systems, soil type distribution patterns and texture gradation patterns, land use classification systems and land use boundary rules, evaluation unit division principles (such as the principle of plot integrity), common sense about geospatial relationships (such as the continuity of grade transitions), and the meaning and expression standards of professional map symbols. Specialized checks of professional elements include:

[0136] Grade boundary identification and verification: Employing a combination of edge detection and semantic segmentation techniques, the boundary lines of each grade zone are accurately extracted. The identified boundary lines are then registered and compared with the digital base map to determine whether the boundary delineation conforms to natural boundaries such as land use boundaries, administrative boundaries, and linear features.

[0137] Grade labeling standardization assessment: Identify grade numerical labels (such as "First-class" and "Level II") on the map, and verify whether the label position is within the zone, whether the label font style conforms to the standard, and whether the correspondence between the label and the legend color block is correct.

[0138] Grade transition analysis: Analyze the attribute changes at the boundaries of adjacent grade zones to identify any unreasonable situations such as grade jumps (e.g., a direct transition from Grade 1 to Grade 4 without transitions to Grade 2 or 3).

[0139] Area consistency verification: Identify the map features of each level from the map, estimate the area percentage, and compare it with the level area data in the report text to identify any possible data inconsistencies.

[0140] The system accurately identified the following issues: The grading line at the boundary of a certain township crossed a complete plot of cultivated land, failing to follow the plot boundary division and violating the principle of evaluation unit integrity; a "second-class" label was incorrectly placed within a third-class color block area; the cultivated land grade along both sides of the river abruptly dropped from first-class to fifth-class, lacking an intermediate grade buffer transition, which is inconsistent with the gradual change in soil texture; the estimated area percentage of a certain grade on the map (23.5%) differed significantly from the corresponding data in the report text (27.8%), indicating potential problems with the data source or statistical methods. The system provided improvement suggestions with professional justification for each issue, such as "It is recommended to re-examine the grading of cultivated land on both sides of the river and supplement intermediate grade sampling data," providing strong assurance for the professional technical quality of the maps.

[0141] The above are merely preferred embodiments of the present application and are not intended to limit the embodiments of the present application. For those skilled in the art, the embodiments of the present application can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of the present application should be included within the protection scope of the embodiments of the present application.

Claims

1. An intelligent verification system for farmland quality evaluation maps based on a multimodal model, characterized in that, include: The front-end interaction module is used to receive farmland quality evaluation maps uploaded by users and extract the metadata of the farmland quality evaluation maps. The image preprocessing module is used to standardize the uploaded farmland quality evaluation maps; The multimodal model service module is used to automatically match and call one or more corresponding visual understanding models as adaptation models based on the map type and complexity of the farmland quality evaluation map through the model scheduler, and convert the structured inspection rules in the inspection rule base into visual task instructions that the adaptation model can execute to identify map elements of the preprocessed farmland quality evaluation map and obtain the recognition results. The dynamic knowledge base module is used to construct a verification rule base based on the "Cultivated Land Quality Grade" (GB / T 33469-2016) and corresponding technical regulations; Based on the recognition results, and combined with the rules in the dynamic knowledge base, the map is subjected to multi-dimensional normative verification to obtain the verification results; The inspection result classification module is used to classify the inspection results into grades, display screenshots of defect locations, and provide targeted repair suggestions.

2. The intelligent verification system for farmland quality evaluation maps based on a multimodal model according to claim 1, characterized in that, The front-end interaction module includes: The image preview component is used to receive farmland quality evaluation images and provide interactive operations, including zooming, panning, rotating and contrast adjustment. The metadata automatic extraction unit is used to extract and parse the metadata of the farmland quality evaluation map and present it in real time through the display panel. The metadata includes file format, resolution, size, color mode, projection parameters, scale information, and geographic coordinates. Batch processing queue is used for parallel uploading of multiple images and tracking of inspection progress.

3. The intelligent verification system for farmland quality evaluation maps based on a multimodal model according to claim 1, characterized in that, The image preprocessing module includes: The resolution standardization unit is used to uniformly adjust farmland quality evaluation maps of different resolutions to a preset standard size; The color correction and enhancement unit is used for unified color space conversion and image enhancement. The format conversion and encoding unit is used to provide mutual conversion between different image formats and convert image data into the Base64 encoding format required by the visual understanding model; The geographic information parsing unit is used to format and store the metadata of the extracted farmland quality evaluation map.

4. The intelligent verification system for farmland quality evaluation maps based on a multimodal model according to claim 1, characterized in that, The multimodal model service module includes: The model scheduler is used to dynamically select one or more visual understanding models as the adaptation model based on the type and complexity of the graph. A multimodal cue word builder is used to convert structured check rules in the check rule base into visual task instructions that can be executed by the adapted model; The result parser is used to parse the unstructured results output by the adaptation model into structured JSON format recognition results, which include map feature categories, location coordinates, and confidence scores. A confidence evaluator is used to quantitatively assess the reliability of the identification results.

5. The intelligent verification system for farmland quality evaluation maps based on a multimodal model according to claim 1, characterized in that, The inspection rule base includes: The completeness rule for results is used to check whether the necessary element layers are included. Mathematical foundational rules are used to verify the accuracy of geographic coordinates, projection parameters, and scales; Map feature rules are used to check the integrity of map features, which include geographic features, label features, and professional features. The finishing and finishing rules are used to verify the standardization of drawing layout, font style, annotation and color matching; Data consistency rules are used to check the consistency between farmland quality evaluation maps and text reports and tabular data.

6. The intelligent verification system for farmland quality evaluation maps based on a multimodal model according to claim 1 or 5, characterized in that, The dynamic knowledge base module is also used to periodically or irregularly add, modify, abolish, or optimize the inspection rule base based on changes in the business operating environment, technical specifications, and historical data feedback.

7. The intelligent verification system for farmland quality evaluation maps based on a multimodal model according to claim 1, characterized in that, The inspection result classification module includes: The results grading unit is used to classify issues into grades based on severity and confidence scores, including compliance, defects, and risks. The problem localization unit is used to accurately locate problem areas at the pixel level on the original farmland quality evaluation map using heat maps and bounding boxes. The evidence screenshot generation unit is used to automatically capture local images of the problem area and present them as visual evidence along with the inspection conclusions. The report generation unit is used to generate a comprehensive inspection report that includes a list of issues, grading results, location coordinates, visual evidence, repair suggestions, and confidence scores. It supports online viewing and export.

8. A method for intelligent verification of farmland quality evaluation maps based on multimodal models, characterized in that, Includes the following steps: Receive farmland quality assessment maps uploaded by users and extract the metadata of the farmland quality assessment maps; Standardize the farmland quality evaluation maps; Several visual understanding models are integrated. Based on the map type and complexity of the farmland quality evaluation map, the model scheduler automatically matches and calls the corresponding single or multiple visual understanding models as the adaptation models. The structured check rules in the check rule base are converted into visual task instructions that the adaptation models can execute to identify map elements in the preprocessed farmland quality evaluation map and obtain the recognition results. Based on the "Grading of Cultivated Land Quality" (GB / T 33469-2016) and corresponding technical specifications, a verification rule base is constructed. On the basis of the identification results, combined with the rules in the dynamic knowledge base, multi-dimensional standardization verification of cultivated land quality evaluation maps is carried out to obtain verification results. Based on the inspection results, the system classifies the defects, displays screenshots of the defect locations, and provides targeted repair suggestions.

9. The intelligent verification method for farmland quality evaluation maps based on a multimodal model according to claim 8, characterized in that, It also includes the following steps: Based on changes in the business operating environment, technical specifications, and historical data feedback, the inspection rule base is regularly or irregularly updated with additions, modifications, repeals, or optimizations.

10. The intelligent verification method for farmland quality evaluation maps based on a multimodal model according to claim 8, characterized in that, It also includes the following steps: Save the recognition results and verification results of the visual understanding model's reasoning; Based on the identification results and verification results, an inspection history archive is established, and based on the inspection history archive, result comparison, trend analysis, version management and audit tracking functions are provided.

Images