Intelligent topographic map labeling and generating method based on knowledge graph

Abstract

The invention relates to the technical field of map generation, in particular to a topographic map intelligent labeling and generating method based on a knowledge graph, and the method comprises the steps: obtaining data in real time; constructing a knowledge graph; generating a to-be-labeled element; carrying out preliminary annotation integration; carrying out low-confidence correction; dynamically updating the annotation; generating a primary topographic map; adjusting the historical duration; and generating a final plate. According to the method, a knowledge graph structure is introduced, spatial geometry, texture and spectral features of image data are fused with semantic features in text terrain description information, a feature vector library and hierarchical attribute relationship is constructed, and a labeling confidence calculation and dynamic updating mechanism is introduced in a labeling stage; and the spatial distribution and feature difference degree of the low-confidence region is corrected, so that the problems of inaccurate terrain change capture and limited labeling precision caused by single data source processing, static knowledge representation and insufficient utilization of multi-temporal data are effectively solved.

Description

Technical Field

[0001] This invention relates to the field of map generation technology, and in particular to a method for intelligent annotation and generation of topographic maps based on knowledge graphs. Background Technology

[0002] With the increasing demand for high-precision topographic information in areas prone to geological disasters, traditional methods that rely on manual annotation or a single data source to generate topographic maps are clearly insufficient in terms of efficiency and accuracy. Especially in the case of large-scale, multi-temporal, and multi-data source scenarios, it is difficult to accurately capture topographic changes and complex relationships between elements, which poses a significant challenge to geological disaster prediction, prevention and control, and emergency management.

[0003] Chinese Patent Application Publication No. CN115495584A discloses a map knowledge representation method based on a knowledge graph. The method includes: acquiring map data and extracting map knowledge from it; wherein the map knowledge includes conceptual knowledge and rule-based knowledge, the conceptual knowledge being common knowledge in the field of cartography and including map element knowledge, the map element knowledge including entities, entity relationships, and entity features; the rule-based knowledge being various models, rules, and principles formed during the cartographic process; generating a map knowledge network graph based on the map element knowledge and a knowledge graph; and performing knowledge reasoning based on the rule-based knowledge to complete the generated map knowledge network graph, thereby generating new entities and / or new entity relationships, thus obtaining a completed map knowledge network graph.

[0004] Therefore, the knowledge graph-based map knowledge representation method has the following problems: the method only focuses on the extraction and representation of map knowledge and does not fully combine multi-source image data and text description information, resulting in limited accuracy of terrain feature recognition; the method has limited processing capability for multi-temporal data, easily ignores terrain change features, and is difficult to adapt to rapid changes in terrain over time or caused by disaster events. Summary of the Invention

[0005] To address this, the present invention provides a knowledge graph-based intelligent topographic map annotation and generation method. This method integrates multi-source image data with textual topographic description information, constructs a topographic element knowledge graph, and combines it with a dynamic update mechanism to achieve high-precision identification and spatial relationship expression of topographic elements. This overcomes the problems in existing technologies where inaccurate topographic change capture and limited annotation accuracy are caused by single data source processing, static knowledge representation, and insufficient utilization of multi-temporal data.

[0006] To achieve the above objectives, this invention provides a knowledge graph-based intelligent annotation and generation method for topographic maps, comprising: Real-time acquisition of image data and textual terrain description information for a preset historical period within areas prone to geological disasters; A terrain feature knowledge graph is constructed based on the terrain feature features in all the image data and all the text terrain description information, and a feature vector library is established for the terrain features according to the preset hierarchical relationship and preset attribute relationship. A set of elements to be labeled is generated based on the spatial geometric features of the image data and the semantic features of the text terrain description information. Based on the set of elements to be labeled, the terrain element knowledge graph, and the feature vector library, the element category, spatial location, and interrelationship of each terrain element are determined and integrated to obtain preliminary labeling results; Based on the preliminary annotation results, several low-confidence annotation regions are identified, and local annotation corrections are performed based on the spatial distribution characteristics and feature differences of the low-confidence annotation regions to obtain corrected annotation results. The degree of deviation is determined based on the image feature vectors and semantic feature vectors in the corrected annotation results within the next preset update cycle, and local re-annotation is performed based on the degree of deviation to obtain dynamic update results; A preliminary topographic map is generated based on the dynamic update results, and the preliminary confidence level of each topographic element in the preliminary topographic map is calculated. The preset historical duration is adjusted based on the distribution characteristics of terrain areas where the initial confidence level is lower than the preset confidence threshold, or the final terrain map is determined, and the terrain element knowledge graph is updated based on the spatial and semantic attributes of each terrain element in the final terrain map.

[0007] Furthermore, the process of acquiring real-time imagery data and textual topographic descriptions of areas prone to geological disasters for a predetermined historical period includes: The high-incidence geological disaster area was acquired by using drones, and the high-incidence geological disaster area was subjected to multi-temporal high-resolution images for a preset historical period of time. The images were then denoised, enhanced, and stitched together to obtain image sub-data. Multispectral images of the geological disaster-prone areas over a predetermined historical period are obtained using satellite remote sensing. The multispectral images are then calibrated for geographic coordinates and dehazed to obtain image sub-data. The slope, vegetation type, water system changes, and distribution of artificial facilities in the geological disaster high-incidence areas are collected by ground sensors over a preset historical period. The data are then deduplicated and pre-processed by word segmentation to obtain the original descriptive information. The image segment data and the original description information are synchronized in time and aligned in space to obtain the image data and the text terrain description information.

[0008] Furthermore, the process of constructing a terrain feature knowledge graph based on the terrain feature features in all the image data and all the text terrain description information includes: Spatial geometric features, texture features, and spectral features are extracted from all the image data to obtain feature extraction results. Based on the feature extraction results, each of the terrain elements is classified and labeled to obtain an image feature vector set. Extract the element categories, interrelationships, and dynamic changes from all the text terrain description information to obtain the description extraction results, and standardize and encode the description extraction results to obtain a set of semantic feature vectors; Based on the image feature vector set and semantic feature vector set, each terrain element is hierarchically divided and attribute-encoded to form a hierarchical index and an attribute index. Based on the hierarchical index, the attribute index, the image feature vector set, and the semantic feature vector set, terrain element nodes, node attributes, and relationships between nodes are generated, and the nodes and relationships are stored in a preset initial knowledge graph to construct the terrain element knowledge graph.

[0009] Furthermore, the process of generating terrain feature nodes, node attributes, and relationships between nodes based on the hierarchical index, the attribute index, the image feature vector set, and the semantic feature vector set includes: The hierarchical position and parent-child relationship of each terrain feature are determined according to the hierarchical index to generate the terrain feature node; The attribute type and attribute value of each terrain feature are determined based on the attribute index, the image feature vector set, and the semantic feature vector set to generate the node attributes; Based on the spatial association features, semantic association features, and hierarchical relationships among the various terrain elements, the association type and direction between nodes are determined to generate the relationships between the nodes.

[0010] Furthermore, the process of establishing a feature vector library for terrain features according to preset hierarchical and attribute relationships includes: Based on the terrain feature nodes and node attributes, the spatial geometric features, texture features, spectral features, and semantic features of each node are extracted to obtain the extracted feature vector. The extracted feature vectors are standardized and encoded, and then classified and grouped according to the preset hierarchical relationship and the preset attribute relationship to establish the feature vector library.

[0011] Furthermore, the process of generating a set of elements to be labeled based on the spatial geometric features of the image data and the semantic features of the textual terrain description information includes: Identify potential terrain boundaries, slope variation areas, texture patterns, and spectral anomaly areas in the spatial geometric features of the image data to generate a set of image candidate features; Extract the element category, element location description, and dynamic change information from the semantic features of the text terrain description information to generate a set of semantic candidate elements; The image candidate feature set and the semantic candidate feature set are fused together, and spatial matching and semantic consistency verification are performed to generate the feature set to be labeled.

[0012] Further, the process of determining and integrating the feature categories, spatial locations, and interrelationships of each terrain feature based on the set of features to be labeled, the terrain feature knowledge graph, and the feature vector library to obtain preliminary labeling results includes: Feature matching and attribute mapping are performed based on the set of features to be labeled and the terrain feature knowledge graph to obtain feature similarity and context consistency. Based on the feature similarity, the contextual consistency, and the feature vector library, the feature category, spatial location, and interrelationship of each terrain feature are determined and integrated to obtain preliminary annotation results.

[0013] Further, based on the preliminary annotation results, several low-confidence annotation regions are determined, and local annotation corrections are performed based on the spatial distribution characteristics and feature differences of these low-confidence annotation regions to obtain corrected annotation results. This process includes: Based on the feature similarity and contextual consistency between each terrain element and the corresponding node in the terrain element knowledge graph in the preliminary annotation results, as well as the matching degree of similar terrain elements in the feature vector library, the annotation confidence of the terrain element is calculated. Identify terrain regions whose label confidence is less than a preset label confidence threshold to determine a number of low-confidence labeled regions; Statistical analysis is performed on the image feature vector set and the semantic feature vector set within the low confidence region to calculate the feature difference degree; The feature category, spatial location, or association of the unlabeled features in the low-confidence region is adjusted according to the spatial distribution characteristics and the feature difference to obtain the corrected labeling result.

[0014] Furthermore, the process of determining the degree of deviation based on the image feature vectors and semantic feature vectors in the corrected annotation results within the next preset update cycle, and performing local re-annotation based on the degree of deviation to obtain dynamic update results includes: Obtain the newly added image data and text terrain description information within the next preset update cycle, and extract the corresponding image feature vector and semantic feature vector; The image feature vector and the semantic feature vector are compared with the corresponding feature vectors in the previous cycle's corrected annotation results. The vector difference, semantic drift rate, and spatial offset are calculated to comprehensively determine the degree of deviation. When the deviation exceeds a preset deviation threshold, the corresponding terrain area is marked as an area to be re-labeled. The image feature vectors and semantic feature vectors in the area to be re-annotated are re-matched and attribute-mapped to calculate new feature similarity and annotation confidence. When the recalculated label confidence score is less than the preset label confidence threshold, the corresponding area to be relabeled is locally relabeled. The terrain feature knowledge graph is updated based on the feature categories, spatial locations, and interrelationships of the terrain region after local relabeling to obtain the dynamic update result.

[0015] Furthermore, the process of adjusting the preset historical duration based on the distribution characteristics of terrain areas where the initial confidence level is lower than the preset confidence threshold includes: Obtain the Euclidean distance between the two-dimensional coordinates and the preset reference coordinates of the terrain region whose initial confidence level is lower than the preset confidence threshold, so as to obtain several distribution distances; Calculate the standard deviation of all the distribution distances to obtain the distribution concentration; When the distribution concentration is greater than a preset concentration threshold, the terrain area is determined to be concentrated, and the preset historical duration is increased according to the relative deviation between the distribution concentration and the preset concentration threshold.

[0016] Compared with existing technologies, the advantages of this invention lie in its ability to automatically extract terrain features and their relationships from multi-source heterogeneous terrain information by introducing a knowledge graph structure and fusing the spatial geometric, texture, and spectral features of image data with the semantic features of textual terrain description information. This process enables high-precision matching and identification of terrain features at both the spatial scale and semantic level through the construction of a feature vector library and hierarchical attribute relationships. The introduction of a labeling confidence calculation and dynamic update mechanism during the labeling stage ensures that the spatial location, morphological features, and semantic descriptions of terrain features maintain adaptive consistency across multi-temporal data. Simultaneously, by correcting the spatial distribution and feature differences in low-confidence areas, dynamic compensation and continuous optimization of local labeling errors are achieved. Finally, this method establishes a quantitative correspondence between image feature changes, semantic drift rate, and labeling confidence, enabling terrain labeling results to evolve continuously in the temporal dimension and maintain overall coordination in the spatial dimension. This significantly improves the accuracy, robustness, and timeliness of automatic terrain map labeling, effectively solving the problems of inaccurate terrain change capture and limited labeling accuracy caused by single data source processing, static knowledge representation, and insufficient utilization of multi-temporal data.

[0017] Furthermore, through the collaborative acquisition of multi-source data from UAVs, satellite remote sensing, and ground sensors, a comprehensive spatiotemporal information fusion of high-risk geological disaster areas was achieved. High-resolution multi-temporal UAV imagery provided detailed changes in surface morphology, satellite multispectral imagery reflected differences in surface reflectivity and material distribution, and ground sensors supplemented information on surface attributes such as slope, vegetation type, water system changes, and the distribution of artificial facilities. By preprocessing the above data through denoising, enhancement, stitching, calibration, and semanticization, and by synchronizing in the temporal dimension and aligning coordinates in the spatial dimension, data from different sources and scales were correlated under a unified spatiotemporal benchmark. This not only improved the spatiotemporal consistency and geometric accuracy of the data but also enabled a quantitative mapping between image change trends and changes in surface parameters. This provides accurate, continuous, and verifiable basic data support for subsequent terrain feature identification, feature extraction, and dynamic evolution analysis.

[0018] Furthermore, by fusing the spatial geometric, textural, and spectral features of image data with the element categories, interrelationships, and dynamic changes described in the textual terrain description information, a terrain element knowledge graph with hierarchical and attribute indexes is formed, realizing a multi-dimensional mapping and association from visual spatial information to semantic information. Through the joint encoding of image feature vectors and semantic feature vectors, the intrinsic connections between terrain elements in terms of spatial distribution, morphological changes, and semantic dependencies can be revealed, enabling a systematic expression of terrain structure, geomorphic evolution, and their interactions. This enhances the interpretability and reasoning ability of terrain data, providing accurate data support and a knowledge foundation for subsequent dynamic terrain identification and evolutionary trend analysis.

[0019] Furthermore, by generating terrain element nodes, node attributes, and inter-node relationships based on hierarchical indexes, attribute indexes, and image and semantic feature vector sets, the spatial structure, attribute features, and semantic information of the terrain can be systematically mapped into a knowledge graph. This allows each terrain element to establish a hierarchical and relational network according to the actual terrain's height, slope, category, and interdependencies, thereby achieving precise matching of spatial location, feature type, and functional attributes. This method ensures that the numerical values ​​and types of node attributes correspond to the actual geometric shape, texture pattern, spectral characteristics, and semantic description of the terrain. Simultaneously, the spatial and semantic relationships between nodes reflect the proximity, inclusion relationships, and functional dependencies between terrain elements, providing accurate references for subsequent annotation, correction, and dynamic updates. This ensures that the generated knowledge graph can scientifically and coherently reflect the terrain structure and its changing patterns.

[0020] Furthermore, by establishing a feature vector library based on preset hierarchical and attribute relationships, the spatial geometric, textural, spectral, and semantic features of terrain elements can be systematically integrated and encoded, enabling various elements to form a distinguishable distribution structure in a multidimensional feature space. The feature weights, spatial distributions, and attribute differences between elements at different levels are quantified and mapped into feature vectors, thereby ensuring that parameters such as elevation changes, slope, vegetation cover, water system distribution, and artificial facilities can mutually verify and coordinate during the annotation and matching process. This reflects the hierarchical structure of the terrain and improves the accuracy of element identification and classification, providing stable and comparable basic data for subsequent matching of elements to be annotated and knowledge graph updates.

[0021] Furthermore, by fusing image geometric features with text semantic features, the collaborative recognition and matching of spatial location, morphological features and semantic attributes are achieved. This enables precise correspondence between terrain boundaries, slope changes, texture patterns and spectral anomalies and feature categories and dynamic change information, thereby improving the integrity and accuracy of the set of features to be labeled. At the same time, it ensures that the terrain labeling results generated under different times and data sources can maintain consistency and reliability, and supports dynamic updates and continuous monitoring.

[0022] Furthermore, precise matching of elements to be labeled is achieved through feature similarity and contextual consistency, enabling effective fusion of image data and textual terrain description information. This ensures the accuracy of terrain element category determination, the rationality of spatial location, and the scientific validity of correlations. By comprehensively utilizing spatial geometric features, texture patterns, spectral anomalies, and semantic features, the interactions between different parameters are quantified and unified, reflecting the inherent laws of terrain objects in spatial distribution, proximity relationships, and type attributes. This ensures that the initial labeling results not only conform to the actual situation of terrain changes but also provide a reliable data foundation for subsequent low-confidence correction and dynamic updates.

[0023] Furthermore, through confidence level calculation and feature difference analysis, terrain areas with high uncertainty are effectively identified, and targeted local corrections are made, significantly improving the accuracy and consistency of the annotation results. The comprehensive application of spatial geometric features, texture patterns, spectral information, and semantic features ensures that the adjustment of low-confidence areas not only considers a single parameter but also coordinates the mutual influence of category determination, spatial location, and feature relationships. This achieves a high degree of consistency in the spatial distribution, type classification, and interrelationships of terrain feature annotations, thereby enhancing the matching degree between the knowledge graph and image data and providing a reliable foundation for subsequent dynamic updates and topographic map generation.

[0024] Furthermore, by comparing and comprehensively analyzing image features and semantic features within different update cycles, the variation deviation of topographic regions can be accurately quantified, and local re-annotation can be selectively performed according to the degree of deviation, thereby ensuring the continuity and accuracy of the annotation results. Parameters such as vector dissimilarity, semantic drift rate, and spatial offset are interrelated with annotation confidence, reflecting the true changes in regional topography and their spatial distribution characteristics. This enables adaptive control of annotation accuracy and topographic information integrity during dynamic updates, improving the timeliness and reliability of topographic maps.

[0025] Furthermore, by analyzing the spatial distribution characteristics of low-confidence areas in the initial topographic map, the historical data collection time is appropriately extended for areas with high distribution concentration. This allows the system to utilize more temporal imagery and textual information for feature accumulation and verification, thereby improving the labeling confidence of these areas. Specifically, the difference between the distribution concentration and a preset threshold is multiplied by an adjustment coefficient and weighted by the historical time interval. This ensures that image features, texture features, spectral features, and semantic features are fully covered and compared within concentrated low-confidence areas, thus reasonably optimizing the labeling results. This makes the determination of spatial location, feature category, and correlation more stable and accurate, while avoiding labeling bias caused by a single time point or a small amount of data. Attached Figure Description

[0026] Figure 1 This is a flowchart of the knowledge graph-based intelligent topographic map annotation and generation method in this embodiment; Figure 2 This embodiment defines the logic diagram for determining the low-confidence labeled region. Figure 3 This is a logic diagram for determining the region to be re-annotated in this embodiment; Figure 4 The logic diagram for increasing the preset historical duration in this embodiment is shown. Detailed Implementation

[0027] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0028] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0029] Please see Figure 1 The diagram shown is a flowchart of the knowledge graph-based intelligent topographic map annotation and generation method of this embodiment. This embodiment provides a knowledge graph-based intelligent topographic map annotation and generation method, including: Real-time acquisition of image data and textual topographic descriptions of geological disaster-prone areas for a predetermined historical period; construction of a topographic feature knowledge graph based on topographic feature features in all the image data and textual topographic descriptions, and establishment of a feature vector library for topographic features according to predetermined hierarchical and attribute relationships; generation of a set of elements to be labeled based on the spatial geometric features of the image data and the semantic features of the textual topographic descriptions; determination and integration of the feature category, spatial location, and interrelationships of each topographic feature based on the set of elements to be labeled, the topographic feature knowledge graph, and the feature vector library to obtain preliminary labeling results; and determination of several low-confidence labels based on the preliminary labeling results. The system first identifies regions and performs local annotation correction based on the spatial distribution characteristics and feature differences of low-confidence annotated regions to obtain corrected annotation results. Then, it determines the degree of deviation based on the image feature vectors and semantic feature vectors within the corrected annotation results in the next preset update cycle, and performs local re-annotation based on the degree of deviation to obtain dynamic update results. A preliminary topographic map is generated based on the dynamic update results, and the preliminary confidence level of each topographic element in the preliminary topographic map is calculated. The preset historical duration is adjusted based on the distribution characteristics of topographic regions whose preliminary confidence level is lower than a preset confidence threshold, or a final topographic map is determined. Finally, the topographic element knowledge graph is updated based on the spatial and semantic attributes of each topographic element in the final topographic map.

[0030] The preset history duration is a parameter used to limit the time range of image data and textual terrain description information. It depends on the rate of terrain change in the geological disaster area and the frequency of multi-temporal image acquisition. It is usually set between 6 months and 3 years. In this embodiment, it is set to 1 year, which can balance the time span and the continuity of change characteristics while ensuring the representativeness of the data. The preset update cycle is a time parameter used to control the frequency of dynamic updates and relabeling. It depends on the update frequency of monitoring data and the rate of terrain change. It is usually set between 7 days and 90 days. In this embodiment, it is set to 30 days, which can capture terrain changes in a timely manner and maintain the real-time accuracy of the labeling results.

[0031] In this embodiment, the preset hierarchical relationship refers to a hierarchical system established based on the hierarchical and inclusive relationships of topographic elements in the natural geographical structure, used to reflect the structural hierarchy and spatial organization among topographic elements. For example, "mountain system" is set as the upper-level element, which can include sub-elements such as "mountain range" and "peak"; "river" is set as the upper-level element, which can include hierarchical units such as "tributary", "river section", and "river bend". Through the setting of this hierarchical relationship, the system can construct a topographic element network structure with spatial inclusiveness and logical continuity in the knowledge graph, realizing layer-by-layer identification and labeling from the macro-whole to the micro-local. The preset attribute relationship refers to a multi-dimensional feature mapping established based on the attribute associations of topographic elements in natural characteristics and interactions, used to describe the semantic and physical connections between topographic elements. For example, "slope", "altitude", "vegetation coverage", and "soil type" are numerical or categorical attributes that form attribute mappings with elements such as "mountain", "valley", and "plain"; "water flow direction" forms a functional association with "watershed catchment area"; and "lithology" forms a causal association with "surface morphology". By setting this attribute relationship, similarity calculation and semantic reasoning of different terrain features in the attribute dimension can be realized in the feature vector library, thereby supporting the subsequent feature matching, confidence assessment and dynamic relabeling process.

[0032] In this embodiment, the process of generating an initial topographic map based on the dynamic update results includes integrating and drawing the feature categories, spatial locations, and interrelationships of various topographic regions to form a complete two-dimensional or three-dimensional topographic map. Simultaneously, the initial confidence level of each topographic feature is calculated using the feature similarity, semantic consistency, and historical annotation confidence levels of the topographic features in the dynamic update results. This confidence level is obtained by weighted comprehensive calculation of the similarity between the image feature vector and the feature vector of the knowledge graph node, the consistency of the semantic feature vector, and the spatial location matching degree. Each indicator is normalized according to preset weights and a weighted average is calculated to obtain a confidence value falling within the range of 0 to 1. This value quantifies the reliability of the topographic feature annotation at that point, providing a basis for subsequent low-confidence area identification and historical duration adjustment.

[0033] In this embodiment, the preset weights of each indicator in the initial confidence calculation are as follows: The weight corresponding to image feature similarity depends on the image resolution, the richness of spectral information, and the effect of image noise processing. High-resolution, low-noise image data will have a higher weight, usually set between 0.4 and 0.6, and in this embodiment, it is set to 0.5; The weight corresponding to semantic feature consistency depends on the detail of the text description, the accuracy of standardized coding, and the degree of matching between the description and the image. The more complete and accurate the text information, the higher its weight can be, usually set between 0.3 and 0.5, and in this embodiment, it is set to 0.35; The weight corresponding to spatial location matching depends on the accuracy of geographic coordinate calibration, the accuracy of sensor positioning, and the spatial distribution complexity of terrain features. The more accurate the spatial positioning and the more complex the distribution, the higher the weight can be, usually set between 0.1 and 0.3, and in this embodiment, it is set to 0.15. The sum of the three weights is 1, which is used to calculate the initial confidence score, thereby comprehensively reflecting the labeling reliability of terrain features in different dimensions.

[0034] By introducing a knowledge graph structure, the spatial geometric, textural, and spectral features of image data are fused with the semantic features of textual terrain descriptions, enabling the automatic extraction of terrain elements and their relationships from multi-source heterogeneous terrain information. By constructing a feature vector library and hierarchical attribute relationships, high-precision matching and identification of terrain elements can be achieved at both spatial and semantic levels. In the annotation stage, an annotation confidence calculation and dynamic update mechanism are introduced to ensure adaptive consistency of the spatial location, morphological features, and semantic descriptions of terrain elements across multiple temporal data. Simultaneously, by correcting the spatial distribution and feature differences in low-confidence areas, dynamic compensation and continuous optimization of local annotation errors are achieved. Finally, this method establishes a quantitative correspondence between image feature changes, semantic drift rate, and annotation confidence, enabling terrain annotation results to evolve continuously in the temporal dimension and maintain overall coordination in the spatial dimension. This significantly improves the accuracy, robustness, and timeliness of automatic terrain map annotation, effectively solving the problems of inaccurate terrain change capture and limited annotation accuracy caused by single data source processing, static knowledge representation, and insufficient utilization of multi-temporal data.

[0035] Specifically, the process of acquiring real-time image data and textual terrain description information for a preset historical period in high-risk geological disaster areas includes: acquiring multi-temporal high-resolution images of the high-risk geological disaster areas over the preset historical period using drones, and performing denoising, enhancement, and stitching processing on the multi-temporal high-resolution images to obtain image sub-data; acquiring multispectral images of the high-risk geological disaster areas over the preset historical period using satellite remote sensing, and performing geographic coordinate calibration and dehazing processing on the multispectral images to obtain image sub-data; collecting information on slope, vegetation type, water system changes, and distribution of artificial facilities in the high-risk geological disaster areas over the preset historical period using ground sensors, and performing deduplication and word segmentation preprocessing to obtain original description information; and synchronizing and spatially aligning the image sub-data and the original description information to obtain the image data and the textual terrain description information.

[0036] Through collaborative data acquisition from multiple sources including UAVs, satellite remote sensing, and ground sensors, a comprehensive spatiotemporal information fusion of high-risk geological disaster areas was achieved. High-resolution multi-temporal UAV imagery provided detailed variations in landform morphology, satellite multispectral imagery reflected differences in surface reflectivity and material distribution, and ground sensors supplemented information on surface attributes such as slope, vegetation type, water system changes, and the distribution of man-made structures. By preprocessing the data through denoising, enhancement, stitching, calibration, and semanticization, and by synchronizing the data temporally and aligning it spatially, data from different sources and scales were correlated under a unified spatiotemporal reference. This not only improved the spatiotemporal consistency and geometric accuracy of the data but also enabled a quantitative mapping between image change trends and changes in surface parameters. This provides accurate, continuous, and verifiable foundational data for subsequent terrain feature identification, feature extraction, and dynamic evolution analysis.

[0037] Specifically, the process of constructing a terrain feature knowledge graph based on terrain feature features from all the image data and all the text terrain description information includes: extracting spatial geometric features, texture features, and spectral features from all the image data to obtain feature extraction results, and classifying and labeling each terrain feature according to the feature extraction results to obtain an image feature vector set; extracting feature categories, interrelationships, and dynamic change descriptions from all the text terrain description information to obtain description extraction results, and standardizing and encoding the description extraction results to obtain a semantic feature vector set; performing hierarchical division and attribute encoding on each terrain feature according to the image feature vector set and the semantic feature vector set to form a hierarchical index and an attribute index; generating terrain feature nodes, node attributes, and relationships between nodes according to the hierarchical index, the attribute index, the image feature vector set, and the semantic feature vector set, and storing the nodes and relationships in a preset initial knowledge graph to construct the terrain feature knowledge graph.

[0038] In this embodiment, the preset initial knowledge graph refers to a topographic feature knowledge framework constructed based on historical image data and textual topographic description information of existing high-risk geological disaster areas. It includes topographic feature nodes, node attributes, and relationships between nodes. Nodes represent different types of topographic features, such as ridges, slopes, rivers, or artificial structures. Node attributes include spatial geometric features, texture features, spectral features, and semantic features. Relationships between nodes reflect the spatial association, hierarchical relationship, and semantic dependency of topographic features. After standardized coding and hierarchical division, this initial knowledge graph can provide a unified reference system for matching subsequent sets of elements to be labeled, generating preliminary labeling results, and dynamically updating, achieving effective fusion and systematic management of image features and semantic information.

[0039] By fusing the spatial geometric, textural, and spectral features of image data with the element categories, interrelationships, and dynamic changes described in textual terrain descriptions, a terrain element knowledge graph with hierarchical and attribute indexes is formed, achieving a multi-dimensional mapping and association from visual spatial information to semantic information. Through joint encoding of image feature vectors and semantic feature vectors, the intrinsic connections between terrain elements in terms of spatial distribution, morphological changes, and semantic dependencies can be revealed. This allows for a systematic expression of terrain structure, geomorphic evolution, and their interactions, thereby enhancing the interpretability and reasoning ability of terrain data and providing accurate data support and a knowledge foundation for subsequent dynamic terrain identification and evolutionary trend analysis.

[0040] Specifically, the process of generating terrain feature nodes, node attributes, and inter-node relationships based on the hierarchical index, the attribute index, the image feature vector set, and the semantic feature vector set includes: determining the hierarchical position and parent-child relationship of each terrain feature based on the hierarchical index to generate the terrain feature nodes; determining the attribute type and attribute value of each terrain feature based on the attribute index, the image feature vector set, and the semantic feature vector set to generate the node attributes; and determining the association type and direction between nodes based on the spatial association features, semantic association features, and hierarchical relationships between the terrain features to generate the inter-node relationships.

[0041] By generating terrain element nodes, node attributes, and inter-node relationships based on hierarchical indexing, attribute indexing, and image and semantic feature vector sets, this method systematically maps the spatial structure, attribute features, and semantic information of terrain into a knowledge graph. This allows each terrain element to establish a hierarchical and interconnected network according to its actual height, slope, category, and interdependencies, achieving precise matching of spatial location, feature type, and functional attributes. This method ensures that the numerical values ​​and types of node attributes correspond to the actual geometric shape, texture pattern, spectral characteristics, and semantic description of the terrain. Simultaneously, the spatial and semantic relationships between nodes reflect the proximity, inclusion relationships, and functional dependencies between terrain elements, providing accurate references for subsequent annotation, correction, and dynamic updates. This guarantees that the generated knowledge graph can scientifically and coherently reflect the terrain structure and its changing patterns.

[0042] Specifically, the process of establishing a feature vector library for terrain elements according to preset hierarchical and attribute relationships includes: extracting spatial geometric features, texture features, spectral features, and semantic features of each node based on the terrain element nodes and node attributes of the terrain elements to obtain extracted feature vectors; standardizing and encoding the extracted feature vectors, and classifying and grouping them according to the preset hierarchical and attribute relationships to establish the feature vector library.

[0043] In this embodiment, by extracting multi-dimensional features from terrain element nodes and their attributes, including spatial geometric features, texture features, spectral features, and semantic features, feature vectors that comprehensively reflect the attributes and spatial structure of terrain elements can be obtained. Subsequently, these feature vectors are standardized and encoded, and terrain elements are divided into upper and lower levels according to a preset hierarchical relationship. Elements are also grouped by category and attribute according to a preset attribute relationship, thereby constructing a structured and searchable feature vector library. This enables efficient and accurate correspondence and integration of terrain elements of different types and levels in subsequent annotation, matching, and knowledge graph updates, while ensuring the uniformity and consistency of spatial location, attribute features, and semantic information.

[0044] By establishing a feature vector library based on preset hierarchical and attribute relationships, the spatial geometric, textural, spectral, and semantic features of terrain elements can be systematically integrated and encoded, enabling various elements to form a distinguishable distribution structure in a multidimensional feature space. The feature weights, spatial distributions, and attribute differences between elements at different levels are quantified and mapped into feature vectors, thereby ensuring that parameters such as elevation changes, slope, vegetation cover, water system distribution, and artificial facilities can mutually verify and coordinate during the annotation and matching process. This reflects the hierarchical structure of the terrain and improves the accuracy of element identification and classification, providing stable and comparable basic data for subsequent matching of elements to be annotated and knowledge graph updates.

[0045] Specifically, the process of generating a set of elements to be labeled based on the spatial geometric features of the image data and the semantic features of the text terrain description information includes: identifying potential terrain boundaries, slope change areas, texture patterns, and spectral anomaly areas in the spatial geometric features of the image data to generate a set of image candidate elements; extracting element categories, element location descriptions, and dynamic change information from the semantic features of the text terrain description information to generate a set of semantic candidate elements; and fusing the set of image candidate elements with the set of semantic candidate elements, and performing spatial matching and semantic consistency verification to generate the set of elements to be labeled.

[0046] By integrating image geometric features with text semantic features, the collaborative recognition and matching of spatial location, morphological features and semantic attributes are achieved. This enables precise correspondence between terrain boundaries, slope changes, texture patterns and spectral anomalies with feature categories and dynamic change information, thereby improving the integrity and accuracy of the set of features to be labeled. At the same time, it ensures that the terrain labeling results generated under different times and data sources can maintain consistency and reliability, and supports dynamic updates and continuous monitoring.

[0047] Specifically, the process of determining and integrating the feature categories, spatial locations, and interrelationships of each terrain feature based on the set of features to be labeled, the terrain feature knowledge graph, and the feature vector library to obtain preliminary labeling results includes: performing feature matching and attribute mapping based on the set of features to be labeled and the terrain feature knowledge graph to obtain feature similarity and contextual consistency; and determining and integrating the feature categories, spatial locations, and interrelationships of each terrain feature based on the feature similarity, contextual consistency, and the feature vector library to obtain preliminary labeling results.

[0048] In this embodiment, during the initial annotation process based on the set of elements to be annotated, the terrain feature knowledge graph, and the feature vector library, each element in the set of elements to be annotated is first matched with the corresponding node in the knowledge graph. Feature similarity and contextual consistency are calculated through attribute mapping to quantify the correspondence between each element and elements in the existing knowledge graph. Subsequently, by combining spatial geometric features, texture features, spectral features, and semantic features from the feature vector library, the element category, spatial location, and interrelationships of each element to be annotated are comprehensively determined and integrated to form a complete preliminary annotation result, reflecting the type distribution, spatial arrangement, and interaction relationships of terrain features.

[0049] By accurately matching the elements to be labeled using feature similarity and contextual consistency, effective fusion of image data and textual terrain description information is achieved, thereby ensuring the accuracy of terrain element category determination, the rationality of spatial location, and the scientific validity of correlation. Through the comprehensive utilization of spatial geometric features, texture patterns, spectral anomalies, and semantic features, the interactions between different parameters are quantified and unified, reflecting the inherent laws of terrain objects in spatial distribution, proximity relationships, and type attributes. This ensures that the initial labeling results not only conform to the actual situation of terrain changes but also provide a reliable data foundation for subsequent low-confidence correction and dynamic updates.

[0050] Please see Figure 2 As shown, this is the logic diagram for determining low-confidence labeled areas in this embodiment. In this embodiment, the process of determining several low-confidence labeled areas based on the preliminary labeling results and performing local labeling correction based on the spatial distribution characteristics and feature differences of the low-confidence labeled areas to obtain the corrected labeling results includes: calculating the labeling confidence of the terrain features based on the feature similarity, contextual consistency, and matching degree of similar terrain features in the feature vector library between each terrain feature in the preliminary labeling results and the corresponding nodes in the terrain feature knowledge graph; identifying terrain areas where the labeling confidence is less than a preset labeling confidence threshold to determine several low-confidence labeled areas; performing statistical analysis on the image feature vector set and the semantic feature vector set in the low-confidence areas to calculate the feature difference; and adjusting the feature category, spatial location, or association relationship of the elements to be labeled in the low-confidence areas based on the spatial distribution characteristics and the feature difference to obtain the corrected labeling results.

[0051] The preset label confidence threshold is a reference value used to judge the reliability of terrain feature labels. It depends on the image resolution, terrain complexity and knowledge graph matching accuracy, and is usually set between 0.6 and 0.9. In this embodiment, it is set to 0.75, which can effectively distinguish between high confidence and low confidence areas and achieve targeted local correction.

[0052] In this embodiment, statistical analysis is performed on the image feature vector set and semantic feature vector set in the low confidence region, including calculating the mean and variance of each feature dimension, statistically analyzing the Euclidean distance and cosine similarity between vectors, and combining the feature distribution difference index and spatial weighted average value to obtain the feature difference degree of the region, which is used to determine whether to perform local annotation correction on the low confidence region.

[0053] In this embodiment, during the process of identifying low-confidence labeled areas based on the preliminary annotation results, the annotation confidence of each terrain feature is first calculated using the feature similarity, contextual consistency, and matching degree of similar terrain features in the feature vector library between the terrain features and the corresponding nodes in the knowledge graph. Areas with annotation confidence below a preset threshold are designated as low-confidence labeled areas. Within these areas, statistical analysis is performed on the image feature vector set and semantic feature vector set to calculate the feature difference. Combined with the spatial distribution characteristics of the low-confidence areas, the feature category, spatial location, and interrelationships of the elements to be labeled are locally adjusted, thereby generating corrected annotation results and optimizing and improving the preliminary annotation results.

[0054] By employing confidence level calculations and feature difference analysis, high-uncertainty terrain areas are effectively identified, and targeted local corrections are made, significantly improving the accuracy and consistency of the annotation results. The comprehensive application of spatial geometric features, texture patterns, spectral information, and semantic features ensures that adjustments to low-confidence areas not only consider a single parameter but also coordinate the interrelationships between category determination, spatial location, and feature relationships. This achieves a high degree of consistency in the spatial distribution, type classification, and interrelationships of terrain feature annotations, thereby enhancing the matching degree between the knowledge graph and image data and providing a reliable foundation for subsequent dynamic updates and topographic map generation.

[0055] Please see Figure 3As shown, this is the logic diagram for determining the area to be re-annotated in this embodiment. In this embodiment, the process of determining the degree of deviation based on the image feature vector and semantic feature vector in the corrected annotation result within the next preset update cycle, and performing local re-annotation based on the degree of deviation to obtain the dynamic update result includes: acquiring the newly added image data and text terrain description information within the next preset update cycle, and extracting the corresponding image feature vector and semantic feature vector; comparing the image feature vector and semantic feature vector with the corresponding feature vector in the corrected annotation result of the previous cycle, and calculating the vector difference and semantic feature vector. The drift rate and spatial offset are used to comprehensively determine the degree of deviation. When the degree of deviation is greater than a preset deviation threshold, the corresponding terrain region is marked as a region to be re-labeled. The image feature vector and semantic feature vector in the region to be re-labeled are re-matched and attribute-mapped to calculate a new feature similarity and recalculate the labeling confidence. When the recalculated labeling confidence is less than a preset labeling confidence threshold, the corresponding region to be re-labeled is locally re-labeled. The terrain feature knowledge graph is updated according to the feature category, spatial location, and interrelationship of the terrain region after local re-labeling to obtain the dynamic update result.

[0056] The preset deviation threshold is a standard used to determine whether a terrain area needs to be re-labeled. It depends on the sensitivity of image feature vectors, semantic feature vectors, and spatial location changes, and is usually set between 0.05 and 0.20. In this embodiment, it is set to 0.10, which can ensure the accuracy of labeling while avoiding unnecessary re-labeling operations due to minor deviations.

[0057] In this embodiment, the vector difference is obtained by calculating the difference between the Euclidean distance or cosine similarity between the newly added image feature vector and the corresponding image feature vector in the previous cycle of correction annotation results; the semantic drift rate is calculated by comparing the ratio of changes in the category and attribute of each element in the newly added semantic feature vector and the semantic feature vector in the previous cycle; the spatial offset is determined by measuring the displacement of the terrain element in two-dimensional or three-dimensional spatial coordinates; the above three parameters are weighted and fused according to preset weights to obtain the comprehensive deviation degree, which is used to determine whether to enter the area to be re-annotated.

[0058] In this embodiment, the preset vector difference weight corresponding to the vector difference degree is used to measure the importance of the impact of image feature changes on the overall deviation. It depends on the image resolution, terrain complexity, and historical annotation stability, and is usually set between 0.3 and 0.6. In this embodiment, it is set to 0.5, which can ensure that sensitive terrain changes are captured in a timely manner while avoiding unnecessary re-annotation caused by minor noise. The preset semantic weight corresponding to the semantic drift rate is used to measure the impact of semantic information changes on the overall deviation. It depends on the completeness of the text description and the feature update frequency, and is usually set between 0.2 and 0.5. In this embodiment, it is set to 0.3, which can balance the response speed of semantic updates and annotation stability. The preset spatial weight corresponding to the spatial offset is used to measure the impact of terrain feature position changes on the overall deviation. It depends on the terrain feature scale and spatial sampling accuracy, and is usually set between 0.2 and 0.4. In this embodiment, it is set to 0.2, which can ensure that minor geographical location drifts will not excessively trigger re-annotation. The sum of the three weights is 1, which is used to calculate the comprehensive deviation degree by weighting, thereby scientifically determining the triggering conditions for dynamic updates and local re-annotation.

[0059] In this embodiment, based on the correction annotation results in the next preset update cycle, newly added image data and textual terrain description information are first acquired, and the corresponding image feature vectors and semantic feature vectors are extracted. Then, these feature vectors are compared with the correction annotation results of the previous cycle, and the vector difference, semantic drift rate, and spatial offset are calculated to comprehensively determine the degree of deviation of each terrain region. When the degree of deviation is greater than a preset threshold, the region is marked as a region to be re-annotated, and feature matching and attribute mapping are re-performed in it to calculate new feature similarity and annotation confidence. When the recalculated annotation confidence is lower than a preset threshold, the region is locally re-annotated, and the terrain feature knowledge graph is updated according to the re-annotated feature categories, spatial locations, and interrelationships to obtain dynamically updated terrain data results.

[0060] By comparing and comprehensively analyzing image features and semantic features within different update cycles, the variation deviation of topographic regions can be accurately quantified. Local re-annotation can then be selectively performed based on the degree of deviation, ensuring the continuity and accuracy of the annotation results. Parameters such as vector dissimilarity, semantic drift rate, and spatial offset are interrelated with annotation confidence, reflecting the true changes in regional topography and their spatial distribution characteristics. This enables adaptive control of annotation accuracy and topographic information integrity during dynamic updates, improving the timeliness and reliability of topographic maps.

[0061] Please see Figure 4As shown, this is the logic diagram for increasing the preset historical duration in this embodiment. In this embodiment, the process of adjusting the preset historical duration based on the distribution characteristics of the terrain areas whose initial confidence level is lower than the preset confidence threshold includes: obtaining the Euclidean distance between the two-dimensional coordinates and the preset reference coordinates of the terrain areas whose initial confidence level is lower than the preset confidence threshold to obtain several distribution distances; calculating the standard deviation of all the distribution distances to obtain the distribution concentration; when the distribution concentration is greater than the preset concentration threshold, determining that the terrain areas are concentrated, and increasing the preset historical duration based on the relative deviation between the distribution concentration and the preset concentration threshold, where T'=T×(1+k×︱U-U0︱ / U0), T' is the increased preset historical duration, T is the original preset historical duration, k is the preset duration adjustment coefficient, U is the distribution concentration, and U0 is the preset concentration threshold.

[0062] The preset concentration threshold is a reference value used to determine whether the spatial distribution of low-confidence terrain areas shows a concentrated trend. It depends on the average spatial distribution density of the terrain area and the fluctuation range of historical annotation confidence. It is usually set between 0.2 and 0.5. In this embodiment, it is set to 0.35, which can effectively identify highly concentrated areas that need to have their historical duration extended. The preset duration adjustment coefficient is an adjustment factor used to control the impact of distribution concentration on the historical duration increment. It depends on the complexity of terrain features and the multi-temporal coverage requirements of image data. It is usually set between 0.1 and 0.5. In this embodiment, it is set to 0.2, which can ensure annotation accuracy while avoiding excessive extension of historical duration and data redundancy.

[0063] By analyzing the spatial distribution characteristics of low-confidence areas in the initial topographic map, the historical data collection time is appropriately extended for areas with high distribution concentration. This allows the system to utilize more temporal imagery and textual information for feature accumulation and verification, thereby improving the labeling confidence of these areas. Specifically, the difference between the distribution concentration and a preset threshold is multiplied by an adjustment coefficient and weighted by the historical time interval. This ensures that image features, texture features, spectral features, and semantic features are fully covered and compared within concentrated low-confidence areas, thus reasonably optimizing the labeling results. This makes the determination of spatial location, feature category, and correlation more stable and accurate, while avoiding labeling bias caused by a single time point or a small amount of data.

[0064] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A knowledge graph-based topographic map intelligent labeling and generation method, characterized in that, The method comprises the following steps: real-time acquisition of image data and text terrain description information in a preset historical time length in a high-geological-disaster area; constructing a terrain element knowledge graph according to the terrain element features in all the image data and all the text terrain description information, and establishing a feature vector library according to the preset hierarchical relationship and the preset attribute relationship of the terrain elements; generating a set of elements to be labeled according to the spatial geometric features of the image data and the semantic features of the text terrain description information; determining the element category, spatial position, and mutual correlation of each terrain element according to the set of elements to be labeled, the terrain element knowledge graph, and the feature vector library, and integrating them to obtain a preliminary labeling result; determining a number of low-confidence labeling areas according to the preliminary labeling result, and performing local labeling correction according to the spatial distribution characteristics and feature difference degree of the low-confidence labeling areas to obtain a corrected labeling result; determining the degree of deviation according to the image feature vectors and semantic feature vectors in the corrected labeling result in the next preset update period, and performing local re-labeling according to the degree of deviation to obtain a dynamic update result; generating a preliminary terrain map according to the dynamic update result, and calculating the preliminary confidence of each terrain element in the preliminary terrain map; adjusting the preset historical time length according to the distribution characteristics of the terrain area with a preliminary confidence lower than the preset confidence threshold, or determining a final terrain map, and updating the terrain element knowledge graph according to the spatial attributes and semantic attributes of each terrain element in the final terrain map.

2. The knowledge graph-based topographic map intelligent labeling and generating method according to claim 1, characterized in that, The process of real-time acquisition of image data and text terrain description information in a preset historical time length in a high-geological-disaster area comprises: acquiring multi-temporal high-resolution images of the high-geological-disaster area in the preset historical time length by a drone, and performing denoising, enhancement, and splicing processing on the multi-temporal high-resolution images to obtain image data; acquiring multi-spectral images of the high-geological-disaster area in the past preset historical time length by satellite remote sensing, and performing geographic coordinate calibration and cloud and fog removal processing on the multi-spectral images to obtain image data; collecting the slope, vegetation type, water system change, and artificial facility distribution of the high-geological-disaster area in the past preset historical time length by ground sensors, and performing de-duplication and word segmentation preprocessing to obtain original description information; synchronizing the image data and the original description information in time and aligning them in space to obtain the image data and the text terrain description information. 3.The knowledge graph-based topographic map intelligent labeling and generating method according to claim 2, characterized in that, The process of constructing a terrain element knowledge graph according to the terrain element features in all the image data and all the text terrain description information comprises: extracting spatial geometric features, texture features, and spectral features from all the image data to obtain feature extraction results, and classifying and identifying each terrain element according to the feature extraction results to obtain an image feature vector set; extracting element categories, mutual relationships, and dynamic change descriptions from all the text terrain description information to obtain description extraction results, and standardizing and encoding the description extraction results to obtain a semantic feature vector set; According to the image feature vector set and the semantic feature vector set, each of the terrain features is hierarchically divided and attribute coded to form a hierarchical index and an attribute index; According to the hierarchical index, the attribute index, the image feature vector set and the semantic feature vector set, terrain feature nodes, node attributes and inter-node relationships are generated, and the nodes and relationships are stored in a preset initial knowledge graph to construct the terrain feature knowledge graph.

4. The knowledge graph-based topographic map intelligent labeling and generating method according to claim 3, characterized in that, The process of generating terrain feature nodes, node attributes and inter-node relationships according to the hierarchical index, the attribute index, the image feature vector set and the semantic feature vector set includes: According to the hierarchical index, the hierarchical positions and parent-child relationships of each of the terrain features are determined to generate the terrain feature nodes; According to the attribute index, the image feature vector set and the semantic feature vector set, the attribute types and attribute values of each of the terrain features are determined to generate the node attributes; According to the spatial correlation features, semantic correlation features and hierarchical relationships between each of the terrain features, the correlation types and directions between nodes are determined to generate the inter-node relationships.

5. The knowledge graph-based topographic map intelligent labeling and generating method according to claim 4, characterized in that, The process of establishing a feature vector library according to the preset hierarchical relationships and preset attribute relationships of terrain features includes: According to the terrain feature nodes and the node attributes of the terrain features, spatial geometric features, texture features, spectral features and semantic features of each node are extracted to obtain extracted feature vectors; The extracted feature vectors are standardized coded, and classified and grouped according to the preset hierarchical relationships and the preset attribute relationships to establish the feature vector library.

6. The knowledge graph-based topographic map intelligent labeling and generating method according to claim 5, characterized in that, The process of generating a set of elements to be labeled according to the spatial geometric features of the image data and the semantic features of the text terrain description information includes: Potential terrain boundaries, slope change areas, texture patterns and spectral anomaly areas in the spatial geometric features of the image data are identified to generate a set of image candidate elements; Element categories, element position descriptions and dynamic change information in the semantic features of the text terrain description information are extracted to generate a set of semantic candidate elements; The set of image candidate elements and the set of semantic candidate elements are fused, and spatial matching and semantic consistency verification are performed to generate the set of elements to be labeled.

7. The knowledge graph-based topographic map intelligent labeling and generating method according to claim 6, characterized in that, The process of determining and integrating the element categories, spatial positions and mutual correlation relationships of each of the terrain features according to the set of elements to be labeled, the terrain feature knowledge graph and the feature vector library to obtain a preliminary labeling result includes: According to the set of elements to be labeled and the terrain feature knowledge graph, element matching and attribute mapping are performed to obtain feature similarity and context consistency; According to the feature similarity, the context consistency and the feature vector library, the element categories, the spatial positions and the mutual correlation relationships of each of the terrain features are determined and integrated to obtain a preliminary labeling result. 8.The knowledge graph-based topographic map intelligent labeling and generating method according to claim 7, characterized in that, The process of determining a number of low-confidence labeling regions according to the preliminary labeling result and locally correcting the labeling according to the spatial distribution characteristics and feature difference degree of the low-confidence labeling regions to obtain a corrected labeling result comprises: According to the feature similarity, context consistency and matching degree of similar topographic features in the feature vector library between each topographic feature in the preliminary labeling result and the corresponding node in the topographic feature knowledge graph, the labeling confidence of the topographic feature is calculated; The topographic regions with a labeling confidence less than a preset labeling confidence threshold are identified to determine a number of low-confidence labeling regions; The image feature vector set and the semantic feature vector set in the low-confidence region are statistically analyzed to calculate the feature difference degree; According to the spatial distribution characteristics and the feature difference degree, the element category, spatial position or correlation of the topographic elements in the low-confidence region are adjusted to obtain the corrected labeling result. 9.The knowledge graph based topographic map intelligent labeling and generating method according to claim 8, characterized in that, The process of determining the deviation degree according to the image feature vector and the semantic feature vector in the corrected labeling result in the next preset update period and locally re-labeling according to the deviation degree to obtain a dynamic update result comprises: The image data and the text topographic description information newly added in the next preset update period are obtained, and the corresponding image feature vector and semantic feature vector are extracted; The image feature vector and the semantic feature vector are compared with the corresponding feature vectors in the corrected labeling result in the last period, respectively, to calculate the vector difference degree, the semantic drift rate and the spatial offset, so as to comprehensively determine the deviation degree; When the deviation degree is greater than a preset deviation degree threshold, the corresponding topographic region is marked as a region to be re-labeled; The image feature vector and the semantic feature vector in the region to be re-labeled are re-matched and re-mapped, the new feature similarity is calculated, and the labeling confidence is recalculated; When the recalculated labeling confidence is less than a preset labeling confidence threshold, the corresponding region to be re-labeled is locally re-labeled; According to the element category, spatial position and mutual correlation of the topographic region after local re-labeling, the topographic feature knowledge graph is updated to obtain the dynamic update result.

10. The knowledge graph-based topographic map intelligent labeling and generating method according to claim 9, characterized in that, The process of adjusting the preset historical time length according to the distribution characteristics of the topographic region with a preliminary confidence lower than a preset confidence threshold comprises: The Euclidean distance between the two-dimensional coordinates of the topographic region with a preliminary confidence lower than a preset confidence threshold and the preset reference coordinates is obtained to obtain a number of distribution distances; The standard deviation of all the distribution distances is calculated to obtain a distribution concentration degree; When the distribution concentration degree is greater than a preset concentration degree threshold, it is determined that the topographic region is concentratedly distributed, and the preset historical time length is increased according to the relative deviation between the distribution concentration degree and the preset concentration degree threshold.

Images