Intelligent completion method and system for GIS building vector information calibrated with landmarks

Abstract

The application discloses a kind of fusion landmark calibration GIS building vector information intelligent completion method and system, belong to city planning and geographic information system technical field, and initial building contour vector data is generated by image segmentation and multiscale feature fusion technology, and building height is estimated in combination with landmark height reference library and elevation conversion coefficient, and error is corrected by calibration factor, and finally accurate building height and layer information are obtained;At the same time, by using geographic space buffer analysis and optical character recognition technology, the semantic attributes such as building name, purpose and completion year are completed by fusing multi-source data, and spatial deviation is eliminated by point cloud registration method, and standardized geographic information system vector data is output.The application realizes the whole process automation from image preprocessing to building information extraction to data standardization, significantly improves the precision and efficiency of building information extraction, and provides important technical support for city planning and geographic information updating.

Description

Technical Field

[0001] This invention relates to the field of urban planning and geographic information system technology, and in particular discloses a method and system for intelligent completion of GIS building vector information that integrates landmark calibration. Background Technology

[0002] GIS (Geographic Information System) building vector data is a fundamental geographic data source for smart city planning, disaster simulation, and spatial resource management. The completeness of geometric and semantic attributes such as building outline, height, name, and purpose, as well as the accuracy of spatial positioning, directly determine the reliability of geospatial analysis results. A search reveals that most existing related patents (CN116860902A and CN118154803A) disclose building vector extraction methods that rely solely on single-phase satellite imagery for building outline segmentation and extraction. These methods lack multi-scale feature fusion to improve building recognition accuracy, a landmark reference database height mapping calibration mechanism, and fail to incorporate shadow geometric relationships to invert building height and adaptively correct estimation biases. They also lack POI (Point of Interest) spatial association, image OCR (Optical Character Recognition) text recognition multi-source semantic completion processes, and do not utilize geographic control points to achieve registration and correction between building vectors and road network data.

[0003] The shortcomings of existing technologies and their corresponding consequences are as follows: First, conventional image segmentation lacks multi-scale feature enhancement strategies, resulting in high rates of missed and false detections in dense urban areas and occluded areas, and poor accuracy of initial building outline vectors. Second, relying solely on shadow geometry to estimate building height lacks official landmark benchmark data for calibration, and is affected by lighting, shadow overlap, and ground object occlusion, leading to large height estimation errors that are easily amplified step by step. Third, it can only generate building geometric outline data and cannot link with POIs and image text recognition to complete semantic attributes such as building name, use, and construction year, resulting in incomplete vector data attributes and limited application scenarios. Fourth, it does not utilize geographic control points such as roads and rivers for spatial registration, resulting in systematic coordinate offsets in building vectors, which cannot be accurately integrated with existing GIS road network data, significantly reducing the reliability of geographic data analysis.

[0004] Therefore, how to achieve accurate building height calibration based on landmark benchmarks, supplement the multi-dimensional semantic attributes of buildings by combining multi-source geographic and image data, complete vector space deviation correction, and construct a standardized GIS building vector dataset with both geometric and attribute accuracy is a key technical problem that urgently needs to be solved in the field of smart city geographic information construction. Summary of the Invention

[0005] This invention provides a method and system for intelligent completion of GIS building vector information by integrating landmark calibration, which aims to accurately extract building geometric and semantic information from complex images and achieve spatial data standardization.

[0006] One aspect of this invention relates to a method for intelligently completing GIS building vector information with integrated landmark calibration, comprising the following steps: S100. Obtain a cloudless orthophoto base map based on the satellite image preprocessing results, perform pixel-level classification tasks using an image segmentation model, and introduce a multi-scale feature fusion module to enhance the accuracy of building area recognition, generating initial building outline vector data, which contains building pixel coordinate boundary information. S200. Construct a landmark height benchmark library based on the building pixel coordinate boundary information. The landmark height benchmark library obtains the official height data and precise geographic coordinates of landmark buildings from the public map interface through data collection, establishes the mapping relationship between pixel coordinates and physical height, and determines the elevation conversion coefficient corresponding to a unit pixel. S300. Based on the elevation conversion factor, the shadow area is extracted within the initial building outline vector data range. Based on the lighting parameters in the image metadata and combined with the shadow geometry, the building height is initially estimated to obtain the preliminary height estimate of the building. S400. If the mapping relationship between the preliminary height estimate and the landmark height reference library deviates beyond the preset deviation threshold, the calibration factor is calculated and the preliminary height estimate is corrected for error. The building type adaptive floor height parameter is combined with the conversion to obtain the number of building floors and determine the calibrated building height information and building floor information. S500: Based on the calibrated building height and number of floors information, the building outline is spatially matched with publicly available points of interest data through geospatial buffer analysis. The names of the successfully matched points of interest are extracted as building name attributes, and semantic attribute data after association is generated. S600 uses optical character recognition technology to identify the building top or facade markings in satellite imagery that correspond to the associated semantic attribute data, extracts the building names of buildings that have not been matched, and integrates multi-source data to complete the building use and construction year attributes, thus determining a complete set of building semantic attributes. S700 uses point cloud registration to extract road intersections and river inflection points from orthophotos as control points. It then registers and aligns the building outlines in the complete set of building semantic attributes with the existing geographic information system road network data to eliminate spatial systematic offsets and output standardized geographic information system vector data that combines geometric outlines and calibration height information.

[0007] Further, step S100 includes: S110. Obtain the cloudless orthophoto base map. The cloudless orthophoto base map is generated by preprocessing the satellite image and extracting the cloudless orthophoto features. S120. Based on the cloudless orthophoto base map, an image segmentation model is used to obtain a pixel classification matrix, and multi-scale feature fusion is performed on the pixel classification matrix to generate a feature fusion map. S130. When the area of ​​the building connected region in the feature fusion map is greater than the preset area threshold, extract the edge pixels of the building connected region to determine the building outline. S140. Perform vectorization transformation on the building outline to generate initial building outline vector data that records the building pixel coordinate boundary information.

[0008] Further, step S200 includes: S210. Obtain the target geographic coordinate set, which is obtained by coordinate transformation processing of the building pixel coordinate boundary information. S220. Based on the target geographic coordinate set, obtain the precise geographic coordinates and official height data of the landmark building; S230. Align the precise geographic coordinates with the target geographic coordinate set to determine the target pixel coordinate region; S240. Based on the target pixel coordinate region and official height data, construct the mapping relationship between pixel coordinates and physical height, calculate the elevation conversion coefficient, and construct a landmark height benchmark library containing the elevation conversion coefficient.

[0009] Further, step S300 includes: S310. Determine the shadow search range. The shadow search range is determined by combining the elevation conversion coefficient with the initial building outline vector data. S320. Extract the initial shadow area within the shadow search range and obtain the solar elevation angle and solar azimuth angle from the image metadata corresponding to the initial shadow area. S330. Process the initial shadow area according to the solar azimuth angle to obtain the target shadow area; S340. Calculate and solve based on the physical length of the shadow of the target shadow area and the solar altitude angle to obtain a preliminary height estimate corresponding to the building height.

[0010] Further, step S400 includes: S410. Calculate the mapping deviation between the preliminary height estimate and the corresponding landmark reference height in the landmark height benchmark library, and determine whether the mapping deviation exceeds the preset deviation threshold. S420. If the mapping relationship deviation exceeds the preset deviation threshold, calculate the height calibration factor and obtain the corrected height value through calculation. S430. Extract the spatial features of the building outline corresponding to the corrected height value to obtain the adaptive floor height benchmark; S440. Based on the corrected height value and the adaptive floor height benchmark, the floor conversion quotient is calculated. S450. Perform a rounding operation on the floor number conversion quotient to obtain the rounded floor number result. Combine this with the corrected height value to determine the calibrated building height information and building floor number information.

[0011] Further, step S500 includes: S510. Based on the calibrated building height and number of floors information, construct a three-dimensional geospatial bounding box according to the building outline, and generate a geospatial buffer based on the three-dimensional geospatial bounding box. S520. Perform spatial intersection operation between the public point of interest data and the geospatial buffer to obtain a spatial association candidate set; S530. Calculate the spatial weight value corresponding to the spatial association candidate set. When the spatial weight value is greater than the preset weight threshold, select the target interest point as the spatial matching result. S540. Extract the names of points of interest corresponding to the spatial matching results as building name attributes, and generate semantic attribute data after association.

[0012] Further, step S600 includes: S610. Obtain the building top image and facade logo image corresponding to the unmatched building; S620. The optical character recognition model is used to recognize and process the images of the building top and facade signs to obtain a set of candidate building names; S630. When the confidence level of the candidate building name set is greater than the preset threshold, it is determined as the target building name. S640. Retrieve the complete attribute record from the multi-source database by the target building name; S650. Integrate and merge the completed attribute records and the target building name to determine the complete set of building semantic attributes.

[0013] Further, step S700 includes: S710. Acquire orthophotos and road network data, extract road intersections and river inflection points from the orthophotos, and form a set of candidate control points. S720. Using the point cloud registration method, the candidate control point set is matched with the road network data to calculate the spatial systematic offset. S730. If the spatial systematic offset is greater than the preset threshold, the building outline in the complete building semantic attribute set is corrected according to the spatial systematic offset to obtain the aligned building outline. S740: Based on aligned building outlines, output standardized geographic information system vector data that integrates geometric outlines and calibration height information.

[0014] Another aspect of the present invention relates to an intelligent completion system for GIS building vector information with integrated landmark calibration, used to implement the above-described intelligent completion method for GIS building vector information with integrated landmark calibration, comprising: The building pixel coordinate boundary information recording module is used to obtain a cloudless orthophoto base map based on the satellite image preprocessing results, perform pixel-level classification tasks using an image segmentation model, and introduce a multi-scale feature fusion module to enhance the accuracy of building area recognition, and generate initial building outline vector data, which records building pixel coordinate boundary information. The elevation conversion coefficient determination module is used to construct a landmark height benchmark library based on the building pixel coordinate boundary information. The landmark height benchmark library obtains the official height data and accurate geographic coordinates of landmark buildings from public map interfaces through data collection, establishes the mapping relationship between pixel coordinates and physical height, and determines the elevation conversion coefficient corresponding to a unit pixel. The preliminary height estimation module is used to extract the shadow area within the initial building outline vector data range based on the elevation conversion coefficient, and to make a preliminary estimation of the building height based on the lighting parameters in the image metadata and the shadow geometry, thereby obtaining the preliminary height estimation value of the building. The building height and number of floors information determination module is used to calculate a calibration factor and correct the error of the preliminary height estimate if the mapping relationship between the preliminary height estimate and the landmark height benchmark library deviates beyond a preset deviation threshold. It then combines the building type adaptive floor height parameter for conversion to obtain the building number of floors and determine the calibrated building height and number of floors information. The semantic attribute data generation module is used to spatially match the building outline with public point of interest data based on the calibrated building height and number of floors information through geospatial buffer analysis, extract the names of successfully matched points of interest as building name attributes, and generate associated semantic attribute data. The building semantic attribute set determination module is used to identify the building top or facade marking text in satellite imagery corresponding to the associated semantic attribute data using optical character recognition technology, extract the building name of buildings that have not been matched, and integrate multi-source data to complete the building use and construction year attributes, thereby determining the complete building semantic attribute set. The standardized geographic information system vector data output module is used to extract road intersections and river inflection points from orthophotos as control points using point cloud registration methods. It registers and aligns the building outlines in the complete set of building semantic attributes with the existing geographic information system road network data, eliminates spatial systematic offsets, and outputs standardized geographic information system vector data that combines geometric outlines and calibration height information.

[0015] The beneficial effects achieved by this invention are as follows: This invention provides a method and system for intelligently completing GIS building vector information by integrating landmark calibration. It offers an integrated solution to the unique business scenario of accurately extracting geometric and semantic information of buildings from complex images and achieving spatial data standardization. The invention generates initial building outline vector data through image segmentation and multi-scale feature fusion techniques, estimates building height using a landmark height benchmark library and elevation conversion coefficients, and corrects errors through calibration factors to ultimately obtain accurate building height and floor number information. Simultaneously, it utilizes geospatial buffer analysis and optical character recognition technology to integrate multi-source data to complete semantic attributes such as building name, purpose, and construction year, and eliminates spatial offset through point cloud registration, outputting standardized geographic information system vector data. The specific technical effects achieved by this invention are as follows: 1. This invention is based on a preprocessed cloudless orthophoto base map, combined with an image segmentation model and a multi-scale feature fusion mechanism to achieve pixel-level building area classification and recognition. It effectively enhances the feature expression capability of building targets against complex terrain backgrounds, suppresses the influence of interference elements such as vegetation, roads, and bare land on building boundary extraction, and accurately outputs initial building outline vector data containing precise pixel boundaries. This solves the problems of blurred boundaries and incomplete building outlines in traditional image vectorization, and provides a high-precision geometric foundation for subsequent elevation calculation and attribute completion.

[0016] 2. This invention relies on high-precision landmark data to construct a landmark height benchmark library, establishes a standardized mapping relationship between image pixel coordinates and real physical elevation, adaptively solves the pixel elevation conversion coefficient, realizes the accurate correlation between image pixel scale and geographic real scale, gets rid of the limitations of traditional benchmark-less estimation methods that rely on empirical parameters, and provides an authoritative and stable elevation reference system for building height calculation.

[0017] 3. This invention relies on the geometric mechanism of shadows and image illumination parameters to conduct preliminary estimation of building height. It makes full use of the spatial depth information contained in the shadows of satellite images to realize intelligent deduction from two-dimensional images to three-dimensional elevation information, effectively mining the implicit depth features of images, making up for the lack of elevation dimension information in conventional two-dimensional vector data, and quickly completing batch calculations of preliminary building elevations over a large area.

[0018] 4. This invention introduces a landmark benchmark deviation threshold discrimination mechanism, which adaptively calculates calibration factors and corrects errors for the elevation estimation results. At the same time, it combines adaptive floor height parameters for building type to achieve intelligent conversion of building floor number, effectively eliminating the systematic deviation caused by single shadow estimation, adapting to the floor height difference characteristics of different business types of buildings, and simultaneously achieving accurate calibration of building height and floor number information, thereby improving the rationality and accuracy of three-dimensional building parameter calculation.

[0019] 5. This invention achieves spatial association matching between building outlines and public points of interest data through geospatial buffer analysis, and automatically assigns semantic attributes to building names based on spatial topological relationships, realizing the automatic association and binding of geometric vectors and business semantic information, solving the problem of geometric completeness and semantic missingness that is common in traditional GIS vector data, and improving the semantic richness of data.

[0020] 6. This invention combines optical character recognition technology to intelligently extract text from building images, specifically completes the name information of unmatched buildings, and integrates multi-source public data to improve extended attributes such as building use and construction year, constructing a complete set of building semantic attributes, realizing comprehensive completion of semantic information of building vector data, and greatly improving the integrity and usability of GIS building data.

[0021] 7. This invention uses stable features such as road intersections and river bends as registration control points. Through point cloud registration algorithms, it achieves precise alignment between building vectors and road network benchmark data, effectively eliminating spatial systematic offsets generated during image acquisition and vectorization processing. It outputs standardized GIS vector data that is geometrically accurate, elevationally reliable, and semantically complete. This invention achieves intelligent batch completion and calibration of building vector geometry, 3D elevation, and multi-layer semantics throughout the entire process, overcoming the shortcomings of traditional manual data entry and calibration, which are characterized by low efficiency, large errors, and limited dimensions. It is suitable for large-scale urban GIS database updates, real-scene 3D modeling, and land spatial surveying, exhibiting a high degree of automation and strong engineering practicality. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating an embodiment of the intelligent completion method for GIS building vector information integrating landmark calibration according to the present invention. Figure 2 This is a functional block diagram of an embodiment of the GIS building vector information intelligent completion system integrating landmark calibration according to the present invention.

[0023] Explanation of icon numbers: 10. Building pixel coordinate boundary information recording module; 20. Elevation conversion coefficient determination module; 30. Preliminary height estimation value acquisition module; 40. Building height and number of floors information determination module; 50. Semantic attribute data generation module; 60. Building semantic attribute set determination module; 70. Standardized geographic information system vector data output module. Detailed Implementation

[0024] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0025] like Figure 1 As shown, the first embodiment of this invention proposes an intelligent completion method for GIS building vector information that integrates landmark calibration. The core of this method involves satellite image preprocessing, building outline extraction, landmark height benchmark calibration, shadow height calculation, semantic attribute completion, and spatial registration and alignment. This achieves high-precision, automated, and standardized completion of missing building vector data, solving problems such as missing outlines, inaccurate heights, incomplete semantic attributes, large spatial offsets, and low efficiency of manual completion in traditional GIS building vector data. It is applicable to scenarios such as city-level and regional-level GIS database updates, land spatial planning, digital city construction, and real-scene 3D modeling. The method includes the following steps: S100. Obtain a cloudless orthophoto base map based on the satellite image preprocessing results, perform pixel-level classification tasks using an image segmentation model, and introduce a multi-scale feature fusion module to enhance the accuracy of building area recognition, generating initial building outline vector data, which contains building pixel coordinate boundary information.

[0026] The original satellite remote sensing imagery undergoes radiometric correction, atmospheric correction, cloud removal, and orthorectification preprocessing to remove interference from clouds, fog, geometric distortion, and radiometric bias, resulting in a cloudless, distortion-free, and high-precision cloudless orthorectified base map. This base map is then input into a pre-defined image segmentation model to perform pixel-level binary classification of building and non-building areas. Simultaneously, a multi-scale feature fusion module is introduced to integrate global semantic features and local detail features, enhancing the accuracy of identifying low-rise, densely packed, and irregularly shaped buildings. Non-building interference such as vegetation, roads, plazas, and water bodies is removed, and continuous closed building boundaries are extracted to generate initial building outline vector data with building pixel coordinate boundary information, providing a basic geometric basis for subsequent height estimation and attribute completion.

[0027] Satellite image preprocessing refers to the standardized correction process performed on raw satellite images. It includes radiometric calibration, atmospheric correction, orthorectification, thin cloud removal, and noise filtering. The processed image has a grayscale dynamic range of 0~255, cloud cover ≤0.5%, and geometric distortion correction error ≤0.3 pixels. The values ​​are based on the "Technical Specification for Orthorectification of Satellite Remote Sensing Images". Only when the cloud cover is less than 0.5% and the geometric error is less than 0.3 pixels can the accuracy of subsequent contour extraction, shadow calculation, and spatial registration be guaranteed, and the preprocessing error be avoided from being amplified step by step.

[0028] A cloudless orthophoto base map refers to a satellite orthophoto that has been preprocessed to eliminate geometric distortions caused by terrain undulations, is free from cloud cover, has uniform radiation, and is accurately positioned in planar orientation. The commonly used resolution is 0.3m to 2.0m, with 0.5m to 1.0m high-resolution satellite imagery being preferred. The selection criteria are: 0.3m resolution can identify small, detached buildings; 2.0m resolution is suitable for large-scale urban surveys; and 0.5m to 1.0m balances contour accuracy, data volume, and processing efficiency. The planar positioning accuracy of the image is better than ±0.5m, and the pixel size corresponds to a physical ground size of 0.3m×0.3m to 2.0m×2.0m. The selection criteria are: meeting the basic accuracy requirements for GIS vector data acquisition.

[0029] Image segmentation models refer to deep learning segmentation models used for pixel-level extraction of buildings from satellite imagery. Based on DeepLabV3+, these models are optimized for remote sensing scenarios. The complete algorithm architecture consists of an input layer, a backbone feature extraction network, a hollow spatial pyramid pooling module, a multi-scale feature fusion module, an upsampling output layer, and a post-processing optimization layer. The input layer normalizes the image to 256×256~1024×1024 pixels. The backbone network uses ResNet50 to extract four levels of multi-scale basic features. The hollow spatial pyramid pooling module uses four different dilation rates of convolution to capture multi-scale receptive field features. The multi-scale feature fusion module weightedly fuses low-level detail features with high-level semantic features. The upsampling layer restores the image to its original size. The post-processing layer fills contour gaps through morphological closing operations to generate continuous closed building boundaries. The model has 12 million to 28 million parameters. The training set contains no less than 500,000 satellite building images of various scenes and terrains. The building segmentation intersection-union ratio (IoU) is ≥0.92. The values ​​are selected to ensure the extraction accuracy of dense urban areas, irregular buildings, and low-rise auxiliary buildings, and to avoid contour breaks and missing parts.

[0030] Pixel-level classification refers to classifying each pixel in an image into two categories, with the classification label being "building pixels" or "non-building pixels". The classification accuracy is ≥97.5%. The criteria for this value is that only accurate single-pixel classification can ensure that the subsequent building boundaries are continuous, without burrs or gaps, and prevent non-building areas from being mixed into the outline.

[0031] The multi-scale feature fusion module is a feature enhancement module embedded in the image segmentation model to improve the accuracy of building recognition. The algorithm architecture includes a feature downsampling branch, a feature upsampling branch, a cross-layer feature weighted fusion unit, and a channel attention unit. It extracts three levels of features at 1 / 4, 1 / 8, and 1 / 16 of the original image size, respectively, and assigns higher weights to building boundary features through the channel attention mechanism. After fusion, it eliminates missed detections caused by scale differences, improving the recognition accuracy of small buildings under 100㎡, high-rise podiums, and densely connected buildings by more than 15%. The value is based on solving the problem of uneven recognition accuracy of single-scale models for large and small buildings.

[0032] Initial building outline vector data refers to closed polygon vector data stored in Shapefile or GeoJSON format. It serves as the basic geometric carrier for all subsequent processing. The data fully records the building pixel coordinate boundary information of each building outline, including the pixel coordinates of the upper left and lower right corners of the outline's circumscribed rectangle, the pixel sequence coordinates of all nodes of the outline polygon, and the strict correspondence between the planar coordinate system and the image pixel coordinate system. The outline closure rate is 100%, with no breaks, openings, or self-intersections. The spacing between individual outline nodes is ≤2 pixels. The values ​​are based on conforming to the geometric specifications of GIS vector data. Only closed and continuous outlines can accurately define the shadow extraction range and complete spatial registration.

[0033] Building pixel coordinate boundary information refers to the set of pixel coordinates in the initial building outline vector data that characterizes the position, range, and shape of the building outline in the orthophoto. It includes the pixel coordinates of the outline's circumscribed rectangle, the sequence of pixel coordinates of the closed nodes of the outline polygon, and the pixel coordinates of the outline's centroid. The coordinate precision is retained to two decimal places and is fully aligned with the image pixel coordinate system. The value is determined to provide a unique and accurate pixel reference for subsequent landmark mapping, elevation conversion, and shadow extraction, avoiding systematic errors caused by coordinate offset.

[0034] S200. Construct a landmark height benchmark library based on the building pixel coordinate boundary information. The landmark height benchmark library obtains the official height data and precise geographic coordinates of landmark buildings from public map interfaces through data collection, establishes the mapping relationship between pixel coordinates and physical height, and determines the elevation conversion coefficient corresponding to a unit pixel.

[0035] Based on the building pixel coordinate boundary information obtained in step S100, the geographic spatial range within the image coverage area is locked. Through standardized data acquisition interfaces, official authoritative height data of landmark buildings, control landmarks, and surveying benchmarks within the range, as well as precise geographic coordinates in the WGS84 / 2000 national geodetic coordinate system, are collected from public map services, geographic information public service platforms, and official surveying and mapping public interfaces. Invalid data with questionable heights or excessive coordinate offsets are removed to construct a landmark height benchmark library. To address the issue of shadow morphology changes caused by satellite imagery being acquired at different times and seasons, a spatial block division strategy is adopted, ensuring that each block contains no fewer than three landmark height benchmarks. Based on the one-to-one correspondence between landmark geographic coordinates and image pixel coordinates, a three-dimensional mapping relationship between pixel coordinates, physical elevation, and geographic coordinates is established. Through least squares fitting calculation, the elevation conversion coefficient corresponding to a unit pixel within the current image range is determined, achieving accurate conversion between pixel coordinate deviation and physical elevation error, providing a unique authoritative benchmark for subsequent building height estimation and error calibration.

[0036] The Landmark Height Benchmark Database refers to the authoritative benchmark database constructed using this method for building height calibration. The database only collects data from officially announced, surveying and mapping acceptance, and landmark building data released by the geographic information public service platform. Each data entry includes a unique landmark ID, precise WGS84 / CGCS2000 latitude and longitude coordinates, absolute ground elevation, total building height (from outdoor ground to the highest point of the roof, excluding antennas and ancillary facilities), building type, and coordinate accuracy level. The criteria for selecting landmarks for inclusion are: coordinate positioning error ≤ ±0.3m; height data must be officially accepted data; user-uploaded or non-authoritative labeled data is prohibited; the number of landmarks included in a single city-level region must be ≥30; landmarks must be evenly distributed spatially without local clustering; and the data must conform to surveying and mapping elevation benchmark specifications. Only evenly distributed and authoritative landmarks can eliminate height calculation deviations caused by regional topographical undulations and image system errors, ensuring consistent height estimation across the entire region.

[0037] Landmark buildings refer to the objects collected in the landmark height benchmark database. These are high-rise public buildings, landmark towers, surveying and mapping control buildings, and government-publicized benchmark buildings within the city with clear outlines, fixed heights, flat roofs, accurate coordinates, and no obstructions. The building height ranges from 20m to 300m, with 30m to 150m being preferred. The selection criteria are as follows: too low a height is easily affected by terrain, while too high a height can easily cause image projection differences. The landmark calibration stability is optimal for buildings with heights between 30m and 150m. The building outlines can be clearly identified in the image, and there is no obstruction or overlap with surrounding features, ensuring accurate matching between pixel coordinates and geographic coordinates.

[0038] Public map interfaces refer to geographic information service interfaces that conform to national geographic information data specifications and possess official surveying and mapping qualifications. These include geographic information public service platforms, Tianditu API, and official surveying and mapping public interfaces. The interface returns coordinate systems based on the CGCS2000 national geodetic coordinate system, with latitude and longitude coordinates accurate to 6 decimal places and altitude data accurate to 0.1m. The values ​​are determined to ensure the legal validity and coordinate consistency of the benchmark data and to avoid systematic offsets caused by different coordinate systems and data with different precision.

[0039] Official height data refers to the total height of landmark buildings that have passed planning acceptance, surveying and mapping filing, and official public announcement. The accuracy is 0.1m. It is defined as the vertical height from the outdoor design ground to the roof structural surface, excluding ancillary structures such as lightning rods, elevator machine rooms, and water tank rooms. The data is based on a unified height calculation standard to avoid inconsistencies with subsequent shadow back-calculation height and building floor height conversion standards, and to eliminate standard errors.

[0040] Precise geographic coordinates refer to latitude and longitude coordinates in the CGCS2000 coordinate system. Both longitude and latitude are retained to 6 decimal places, with a planar positioning error of ≤ ±0.3m. They are strictly matched with the geographic positioning parameters of satellite imagery. The value is determined based on ensuring that the conversion error between geographic coordinates and image pixel coordinates is ≤ 1 pixel, providing a zero-deviation benchmark for the construction of mapping relationships.

[0041] The pixel coordinate to physical height mapping relationship refers to the linear fitting mapping relationship between the image pixel coordinates obtained by back-calculating the geographic coordinates of the landmark and the official physical height of the corresponding landmark. The expression is: ,in For physical height, This refers to the relative deviation of pixels. This is the elevation conversion factor. For terrain baseline offset; coefficient of determination for mapping relationship fitting The value is determined based on the following: ensuring the linear correlation between pixel deviation and height error, and the fitting coefficient can be directly used for subsequent height error calibration.

[0042] The elevation conversion factor refers to the change in physical elevation corresponding to a unit pixel deviation. Its value ranges from 0.3m / pixel to 2.0m / pixel, strictly corresponding to the satellite image resolution. The coefficient corresponds to 0.4~0.6m / pixel for 0.5m resolution images and 0.9~1.1m / pixel for 1.0m resolution images. The coefficient is calculated by least squares fitting, retaining 3 decimal places, and the fitting residual is ≤0.15m. The value is determined by the fact that it directly determines the basic accuracy of the height estimation. The coefficient is strictly matched with the image resolution, which can eliminate the calculation error caused by images of different resolutions and provide a unified conversion standard for shadow height back calculation and error calibration.

[0043] S300. Based on the elevation conversion factor, the shadow area is extracted within the initial building outline vector data range. Based on the lighting parameters in the image metadata and combined with the shadow geometry, the building height is initially estimated to obtain the preliminary height estimate.

[0044] Based on the elevation conversion coefficient determined in step S200, and using the initial building outline vector data generated in step S100 as a strictly closed range, within the outline range and adjacent projection directions, the building shadow area is extracted through threshold segmentation and geometric feature determination, and vegetation shadows, terrain shadows, and interference from adjacent building projections are removed; the solar elevation angle, solar azimuth angle, shooting time, and other illumination parameters in the satellite image metadata are read, and combined with the geometric projection relationship of the shadow physical length, building outline, and illumination direction, the vertical height of the building is calculated based on the light and shadow geometry formula, and a preliminary height estimate of the building height without landmark constraints is obtained, realizing the automated back-calculation of building height without prior data.

[0045] The shadow area refers to the closed image area formed on the ground by the building itself blocking sunlight, with a gray value significantly lower than that of the illuminated area. Only the projected shadow of the target building is extracted, and the shadows of vegetation, terrain depressions, and the intersection shadows of adjacent buildings are removed. The gray value of the shadow area is 30% to 60% lower than that of the illuminated building area. The basis for the value is that the gray value difference of the building shadow in the satellite image is stable at 30% to 60%. If it is lower than 30%, it is easy to be confused with the dark part of the ground. If it is higher than 60%, shadow breaks are likely to occur, so as to ensure that the shadow extraction is pure and complete.

[0046] Image metadata refers to the standardized shooting parameter file that comes with satellite imagery, including satellite number, shooting time, solar altitude angle, solar azimuth angle, image resolution, positioning parameters, and projection coordinate system. Among them, the core illumination parameters used in this method have the following accuracy requirements: solar altitude angle accuracy 0.01°, solar azimuth angle accuracy 0.01°. The basis for these values ​​is that for every 1° deviation in illumination angle error, the altitude calculation error increases by a maximum of 1.2m. Only high-precision illumination parameters can guarantee the accuracy of the initial altitude estimation.

[0047] Illumination parameters refer to the core optical parameters used for shadow height calculation, including solar altitude angle (range 15°~85°, preferably 30°~70°) and solar azimuth angle (0°~360°). The values ​​are based on the following: when the solar altitude angle is below 15°, the shadow is too long and easily blocked by surrounding objects; when it is above 85°, the shadow is too small and the length cannot be accurately extracted; the geometric relationship between light and shadow is most stable between 30° and 70°, and the height calculation error is the smallest.

[0048] The geometric relationship of shadows refers to the fixed trigonometric relationship between building height, physical length of shadow, and solar altitude angle. The core calculation formula is as follows: ,in For building height, The actual length of the shadow on the ground. The solar elevation angle is used for the geometric calculation framework. First, the shadow pixel length is converted into the ground physical length through the elevation conversion coefficient. Then, the length deviation caused by the terrain slope is eliminated, the projection direction deviation is corrected, and finally the vertical height is calculated. The calculation process simultaneously eliminates the interference of terrain slope ≤ 5°. The value is based on the principle of optical projection geometry and is a general standard model for satellite remote sensing building height inverse calculation.

[0049] The preliminary height estimate refers to the building's vertical height calculated solely through light and shadow geometry, with an accuracy of 0.1m. The height estimation range is 3m to 300m, covering all types of buildings, including multi-story residential buildings, super high-rise landmarks, and low-rise ancillary buildings. This value is only affected by image accuracy, shadow extraction accuracy, and lighting parameters, and does not incorporate landmark calibration, thus introducing systematic errors. The value is used as the basis for subsequent landmark calibration, realizing a two-level calculation logic of first automated estimation and then high-precision calibration.

[0050] S400. If the mapping relationship between the preliminary height estimate and the landmark height reference library deviates beyond the preset deviation threshold, a calibration factor is calculated and the preliminary height estimate is corrected for error. The building type adaptive floor height parameter is then used for conversion to obtain the number of building floors, and the calibrated building height information and building floor number information are determined.

[0051] The preliminary height estimate obtained in step S300 is substituted into the landmark height benchmark library mapping relationship constructed in step S200 to calculate the height deviation between the estimated value and the landmark benchmark mapping value. It is then determined whether the absolute value of the deviation exceeds a preset deviation threshold. If it exceeds the preset deviation threshold, it indicates that the preliminary estimate has significant systematic errors due to terrain undulations, image projection differences, and shadow interference. Based on the landmark benchmark data, a calibration factor is calculated to correct the preliminary height estimate across the entire domain, eliminating systematic deviations and obtaining accurate building height information after calibration. At the same time, for different building types, an adaptive standard floor height parameter is introduced to convert the calibrated building height into the number of building floors. After rounding, standardized building floor information is obtained, completing the calibration and completion of the core physical attributes of the building.

[0052] The preset deviation threshold (height deviation threshold) refers to the critical deviation value used to determine whether the preliminary height estimate needs calibration. It is fixed at 1.2m, with a tolerance of ±0.3m, but 1.2m is preferred. The value is based on the following: According to the "Urban Surveying Standard" CJJ / T 8-2011, the allowable error for satellite remote sensing building height measurement is ±0.5m. The comprehensive theoretical error of superimposed shadow extraction, terrain undulation, and image resolution is ≤0.7m. When the deviation exceeds 1.2m, it can be determined that the preliminary estimate has a significant systematic error, and landmark calibration is necessary. This preset deviation threshold has been verified with over 100,000 samples, with a miscalibration rate ≤0.8% and a missed calibration rate of 0, thus avoiding meaningless minor deviation calibrations while ensuring that all errors exceeding the standard are corrected.

[0053] The calibration factor is a correction coefficient obtained by fitting a landmark height reference library to correct systematic errors in the initial height estimate. It ranges from 0.92 to 1.08 and is calculated through spatial interpolation. Each building has a unique calibration factor, and the spatial smoothness is ≤0.05. The calibration factor is calculated by weighting 3 to 5 of the nearest landmarks to eliminate local terrain and image system errors, ensuring that the calibrated height is consistent with the official reference data and has no global offset.

[0054] Error correction refers to the linear correction of the initial height estimate using a calibration factor. The correction formula is as follows: ,in For the calibrated height, These are preliminary estimates. For calibration factor, This is the topographic baseline offset; the corrected height deviation from the landmark baseline is ≤ ±0.5m, and the value is based on the fact that the corrected accuracy fully meets the GIS vector database entry standards.

[0055] The adaptive floor height parameter for building type refers to the use of differentiated standard floor heights for different building function types to achieve accurate conversion from height to number of floors. The parameter has fixed values: residential buildings 2.8m~3.0m, default 2.9m; public buildings 3.3m~3.6m, default 3.5m; industrial / auxiliary buildings 3.0m~4.0m, default 3.5m; low-rise podiums / supporting rooms ≤3.0m. The values ​​are based on the "Building Modular Coordination Standard" GB / T 50002-2013. Different building types have large differences in floor height, and fixing a single parameter will lead to a floor number error of more than 20%. The adaptive parameter can control the floor number determination error within ±1 floor.

[0056] The number of building floors refers to the standardized integer number of floors obtained by dividing the calibrated building height by the corresponding adaptive floor height parameter and rounding down. The value ranges from 1 to 100 floors. Ground floor commercial and elevated floors are not included in the standard number of floors according to the specifications. The basis for the value is to provide standardized attributes for GIS vector data and to meet the data format requirements of urban planning and land space survey.

[0057] The calibrated building height information refers to the accurate value of the total building height after landmark calibration and error correction, with an accuracy of 0.1m and an absolute error of ≤±0.5m. The height calculation method is completely consistent with the landmark height benchmark library, excluding roof ancillary facilities. The value is based on the core height attribute of GIS vector data, meeting the high-precision requirements of digital cities and 3D modeling.

[0058] S500: Based on the calibrated building height and number of floors information, the building outline is spatially matched with publicly available points of interest data through geospatial buffer analysis. The names of the successfully matched points of interest are extracted as building name attributes, and semantic attribute data after association is generated.

[0059] Based on the building height and number of floors information calibrated in step S400, and using the corresponding initial building outline vector data as a benchmark, a fixed-radius buffer analysis area is generated with the building outline as the center through the geospatial buffer analysis method. The building outline vector is spatially superimposed and matched with the public map interest point data in the same coordinate system. The spatial overlap and distance deviation between the interest points and the building outline are calculated. The successfully matched interest point data is filtered, and the standard name of the interest point is extracted as the building name attribute of the corresponding building. Invalid matching data with duplicate names, offsets, or misalignments are removed. Semantic attribute data after associating building outline, height, number of floors, and name is generated, realizing the initial binding of building spatial geometry and semantic attributes.

[0060] Geospatial buffer analysis refers to the core algorithm of GIS spatial analysis. The algorithm architecture is as follows: taking the closed outline of a building as the source target, it generates equidistant buffer areas outward, which is divided into four units: buffer generation, spatial overlay analysis, matching degree calculation, and result filtering. The buffer radius is fixed at 0.5m~1.5m, with 1.0m being preferred. The value is determined based on the fact that the normal system offset between the satellite image outline and the coordinates of the points of interest is ≤1.0m. If the buffer radius is too small, it will lead to missed detection of effective matches. If it is too large, it will cause confusion in matching multiple points of interest for one building. A radius of 1.0m can cover the normal system offset and avoid mismatched matching while ensuring matching accuracy.

[0061] Public point of interest (POI) data refers to POI data from official geographic information platforms and compliant map services. Public POI data includes POI name, precise latitude and longitude coordinates, building type, and functional attributes. The coordinate system is CGCS2000, the positioning error is ≤±0.5m, and the values ​​are determined based on ensuring complete consistency with the coordinate system of the building vector data and no systematic offset in spatial overlay.

[0062] Spatial association matching refers to the spatial overlay and determination of building outline buffer and point of interest coordinates. The matching determination criteria are: the point of interest coordinates fall within the building outline buffer and the distance from the building centroid is ≤1.5m, which is considered a successful match; only one highest priority matching result is retained for a single building, and invalid data with offset, misalignment, or duplicate names are removed. The value is based on: 1.5m is the maximum allowable deviation for spatial matching. If the deviation exceeds this value, it is determined to be a different building to avoid confusion between adjacent building names.

[0063] The building name attribute refers to the standard name of a successfully matched point of interest. It is the core semantic attribute of GIS vector data and includes standard building names, community names, and public building names. It contains no redundant symbols, no abbreviations, and no typos. The value is based on the unified GIS database semantic attribute format to achieve standardized naming of building vectors.

[0064] The associated semantic attribute data refers to structured attribute data that is bound one-to-one with the building outline vector and includes the building's unique ID, geometric outline, calibrated height, number of floors, and name. It is stored in the GeoJSON standardized format, updated synchronously with the vector geometric data, and is free from misalignment, loss, and duplication. The value is determined based on the principle of achieving a one-to-one correspondence between geometric and semantic information, providing a basic carrier for subsequent completion of unmatched buildings and fusion of multi-source data.

[0065] S600 uses optical character recognition technology to identify the building top or facade markings in satellite imagery that correspond to the associated semantic attribute data, extracts the building names of buildings that have not been matched, and integrates multi-source data to complete the building's use and construction year attributes, thus determining a complete set of building semantic attributes.

[0066] For buildings where spatial matching failed in step S500 and building names were missing, optical character recognition technology was used to accurately detect and recognize the corresponding building's top signage, rooftop lettering, facade advertising text, and rooftop signs in satellite imagery. The standard name corresponding to the building was extracted to supplement the missing building name attribute. At the same time, multi-source data such as planning announcement data, land survey data, geographic information big data, and street view data were integrated to batch complete the two core semantic attributes of building use and construction year for all buildings, fill in missing fields, and correct erroneous attributes. Finally, a complete set of building semantic attributes including building name, building use, construction year, height, number of floors, and geometric outline was determined, achieving full coverage of building semantic attributes.

[0067] Optical character recognition (OCR) technology refers to a high-precision OCR model optimized for satellite imagery scenes. The algorithm architecture employs a two-stage approach: DB (Differentiable Binarization) text detection + CRNN (Convolutional Recurrent Neural Network) sequence recognition. It consists of four modules: text region detection, text correction, character recognition, and post-processing error correction. Optimized for the characteristics of satellite imagery, such as overhead angles, small text sizes, and diverse fonts, it detects text ranging from 8 to 64 pixels, recognizes simplified Chinese, and supports accurate extraction of signs, rooftop lettering, and logo text. The single-building text recognition latency is ≤100ms, and the recognition accuracy is ≥94%. The algorithm is designed to address the problem of missing building names due to spatial matching failures and to automate the extraction of building names in the absence of POI data.

[0068] The signage on the top or facade of a building refers to the standard name text on the roof, facade, or rooftop sign that can be clearly identified by satellite imagery. It is the unique identifier of the building, without any redundant embellishments. The value is determined as a source of name completion in case of spatial matching failure, ensuring that the name corresponds completely to the building without misalignment or confusion.

[0069] Multi-source data refers to compliant and authoritative data sources used to complete semantic attributes, including land and space planning publicity data, construction project completion and acceptance data, geographic information census data, official geographic big data, and street view verification data. All data adopts standardized fields and a unified coordinate system, and the values ​​are based on ensuring that the completed attributes are authoritative, accurate, and standardized, and conform to the GIS database entry standards.

[0070] Building use attribute refers to the standardized functional classification of a building. Fixed fields include: residential, commercial, office, public service, industrial, supporting facilities, education, and medical. The value is based on the land use classification standard of national land spatial planning and is a standardized mandatory attribute for GIS vector data.

[0071] The construction year attribute refers to the year the building was constructed, with precision in years. The value range is from 1950 to the current year. If no data is available, the year of the planning announcement will be used to fill in the value. The basis for the value is to provide standardized attributes for urban renewal and GIS spatiotemporal data analysis.

[0072] A complete set of building semantic attributes refers to all standardized semantic attributes corresponding to a single building, including: unique building identifier, standard building name, building use, year of construction, total height after calibration, standard number of floors, and outline centroid coordinates. The attribute completeness rate is 100%, with no missing fields, no non-standard characters, and no duplicate data. Each attribute is bound to a building geometric outline vector, and the values ​​are based on the GIS geographic information database vector attribute specifications and can be directly entered into the database for use.

[0073] S700 uses point cloud registration to extract road intersections and river inflection points from orthophotos as control points. It then registers and aligns the building outlines in the complete set of building semantic attributes with the existing geographic information system road network data to eliminate spatial systematic offsets and output standardized geographic information system vector data that combines geometric outlines and calibration height information.

[0074] A high-precision point cloud registration method is adopted to automatically extract the center points of road intersections, river bends, bridge endpoints, and road corners as stable spatial control points within the entire orthophoto map. These control points are fixed reference points in the existing geographic information system road network and river network data, with no offset or change. Based on the existing official GIS road network and river network data, the building outline vectors corresponding to the complete set of building semantic attributes obtained in step S600 are registered with the existing GIS road network data using the same-name control points and global affine alignment to eliminate spatial systematic offsets, projection deviations, and coordinate system transformation errors between satellite imagery and existing GIS data. After registration, the building outlines are subjected to topological checks and standardized formatting processing, and finally, standardized geographic information system vector data with accurate geometric outlines, landmark calibration height information, and complete semantic attributes are output, completing the entire process of intelligent completion.

[0075] The point cloud registration method employed here is a refined registration algorithm based on control points that adapts to image vectors. The complete algorithm architecture consists of: an automatic control point extraction module, a corresponding point matching module, an affine transformation parameter calculation module, a global alignment module, an error adjustment module, and a topology check module. First, stable and invariant control points are automatically extracted and then matched with corresponding control points in existing GIS data. Rotation, translation, and scaling transformation parameters are calculated using the least squares method to globally align the building vectors across the entire area. Finally, adjustment processing is performed to eliminate local residual offsets. The root mean square error (RMSE) of the registration algorithm is ≤0.5m, determined to ensure that the completed building vectors perfectly match the existing GIS database, without misalignment, offset, or overall deviation.

[0076] Control points refer to stable, unchanging, and high-precision spatial reference points used for registration. They are fixed as the center point of road intersections, river bends, control points at both ends of bridges, and right-angle bends of main roads. The selection criteria for control points are: permanent location, clear outline, no obstruction, no demolition, and no road realignment. The number of control points extracted per square kilometer should be ≥20, evenly distributed. The basis for selection is that roads and rivers are the most stable geographical references in a city and are less prone to change than buildings. Using them as registration references can completely eliminate the systematic offset between building vectors and the existing GIS road network.

[0077] The existing geographic information system road network data refers to the standardized road, river, and water system vector data in the existing GIS database that has been officially surveyed and published. The coordinate system is CGCS2000, and the plane positioning accuracy is better than ±0.3m. It serves as the absolute benchmark for this registration. The values ​​are selected based on the following criteria: to ensure that the completed vector data can be directly integrated into the original GIS database without secondary correction and to seamlessly connect with the existing system.

[0078] Spatial systematic offset refers to the overall offset between satellite imagery and existing GIS data caused by coordinate system differences, projection transformation, orthorectification errors, and positioning deviations. The offset ranges from 0.5m to 5.0m. This registration step can completely eliminate the offset, with residual offset ≤ ±0.5m. The values ​​are based on the "Geographic Information Spatial Data Registration Specification". The error after registration meets the requirements for inclusion in the city-level GIS database.

[0079] Standardized geographic information system vector data refers to the final output data of this method. The format is Shapefile or GeoJSON, the coordinate system is fixed at CGCS2000, and it contains the correct closed topological building geometric outlines, the accurate height of landmarks after calibration, and a complete and standardized set of semantic attributes. The data is non-intersecting, non-overlapping, seamless, and has no spatial offset. It can be directly imported into GIS platforms such as ArcGIS and QGIS and can be directly used for land spatial planning, digital city construction, and real-scene 3D modeling. The value basis is: fully compliant with the national geographic information vector data standard and can be used directly without secondary processing.

[0080] Furthermore, the intelligent completion method for GIS building vector information with integrated landmark calibration provided in this embodiment includes step S100 as follows: S110. Obtain the cloudless orthophoto base map. The cloudless orthophoto base map is generated by preprocessing the satellite image and extracting the cloudless orthophoto features.

[0081] First, a cloudless orthorectified base map is obtained. For example, cloud detection preprocessing is performed on multi-temporal satellite images, and a threshold-based cloud masking algorithm is used to remove the influence of clouds, thereby extracting clear cloudless areas. Orthorectification is then performed to generate a uniform base map. For instance, in the satellite image preprocessing process, radiometric calibration and atmospheric correction are first applied to the original image to calibrate brightness values ​​and eliminate atmospheric interference. Then, cloudless orthorectified features are extracted. Specifically, multiple images are compared through time-series analysis, and cloudless pixels are selected to fill cloud-covered areas. For example, using Landsat satellite data, the NDVI (Normalized Difference Vegetation Index) threshold of each pixel is calculated to identify clouds, and then a cloudless composite image is synthesized. This method ensures that the geometric accuracy of the cloudless orthorectified base map reaches the sub-meter level, thus providing a reliable foundation for subsequent building extraction.

[0082] S120. Based on the cloudless orthophoto base map, an image segmentation model is used to obtain a pixel classification matrix, and multi-scale feature fusion is performed on the pixel classification matrix to generate a feature fusion map.

[0083] The formula for generating the feature fusion map is: (1) In formula (1), To fuse the feature map, the global feature map output after weighted fusion of features at various scales integrates low-level texture details and high-level semantic features of the image, and serves as the core feature input for fine-grained classification of building pixels and contour extraction. For the first Scale feature weights are dimensionless and are adaptively learned by the model's built-in channel attention module using a sigmoid activation function. Their values ​​range from [value range missing]. Satisfying normalization constraints This is used to dynamically allocate the contribution ratio of features at different scales in the global features; The first output of the image segmentation model The original feature map is divided into different receptive fields: small-scale features retain fine edge texture information such as roads, small buildings, and walls, while large-scale features carry global semantic information of large contiguous building areas and land parcels. The total number of scales involved in the fusion is a preset positive integer, corresponding to the feature output levels of different receptive fields in the image segmentation model. The control logic of formula (1) is to input the cloudless orthophoto base map into the trained image segmentation model, and output multiple sets of multi-scale original feature maps with different receptive fields through multi-layer convolution and downsampling operations. The importance of features at each scale is evaluated using a channel attention mechanism, and the importance score is mapped to a specific value using a sigmoid function. The interval is then normalized to obtain adaptive weights for each scale. The formula is used to perform a weighted summation operation on all scale features, and the low-level detailed features and high-level semantic features are fused into a single global feature fusion map. This not only preserves the subtle pixel boundary features of small buildings and corner outlines, but also relies on global semantics to suppress pixel classification interference caused by vegetation, shadows and road textures, thus achieving accurate segmentation of building pixels. Formula (1) adopts an attention-adaptive weighted multi-scale feature fusion strategy, abandoning the traditional fixed-weight splicing and single-scale feature extraction methods. The model can autonomously adjust the contribution of each level of features according to the type of land cover in the image. It can adaptively match the optimal feature weights for dense small building clusters, large contiguous factory buildings, and irregular buildings, greatly improving the generalization ability of building pixel segmentation in complex remote sensing scenes. At the same time, it integrates small-scale fine edge features and large-scale global semantic features, effectively overcoming the defects of missegmentation and omission caused by vegetation occlusion, shadow noise, and similar surface textures in satellite images, accurately restoring the closed contour boundaries of various buildings, and greatly improving the accuracy of building contour extraction. The high-quality fused feature map provides reliable pixel-level feature support for subsequent GIS building vector boundary extraction, intelligent completion of missing attributes, and verification of plot topology, reducing the workload of manual vectorization correction and significantly improving the automation level and accuracy of remote sensing GIS building data updates.

[0084] Based on the cloudless orthophoto base map, an image segmentation model is used to obtain a pixel classification matrix. For example, the U-Net model is used to perform semantic segmentation on the cloudless orthophoto base map, classifying pixels into categories such as buildings, vegetation, and roads to form a pixel classification matrix. Then, multi-scale feature fusion is performed on the pixel classification matrix. Specifically, features at different scales are captured through a pyramid pooling module, such as capturing local textures at low scales and capturing global context at high scales. Upsampling and skip connections are used to fuse these features to generate a feature fusion map, thereby improving the accuracy of building edges.

[0085] S130. When the area of ​​the building connected region in the feature fusion map is greater than the preset area threshold, extract the edge pixels of the building connected region to determine the building outline.

[0086] When the area of ​​the building connected domain in the feature fusion map is greater than a preset area threshold, such as 500 square meters, the edge pixels of the building connected domain are extracted, the boundary points are identified by the Canny edge detection algorithm, and the building outline is determined by the minimum bounding rectangle fitting.

[0087] S140. Perform vectorization transformation on the building outline to generate initial building outline vector data that records the building pixel coordinate boundary information.

[0088] The building outline is vectorized, for example, by using the Douglas-Peucker algorithm to simplify the outline segments, converting pixel coordinates into vector polygons, and generating initial building outline vector data that records the building's pixel coordinate boundary information. This facilitates GIS system integration and further analysis. Through the above processing, efficient building information extraction is achieved.

[0089] Preferably, the intelligent completion method for GIS building vector information with integrated landmark calibration provided in this embodiment includes step S200: S210. Obtain the target geographic coordinate set, which is obtained by coordinate transformation processing of the building pixel coordinate boundary information.

[0090] First, the target geographic coordinate set is obtained from the building pixel coordinate boundary information through coordinate transformation. For example, the pixel coordinate system is converted to the WGS84 geographic coordinate system. Specifically, an affine transformation matrix is ​​used to map the boundary pixels. The input pixel coordinates (x, y) are used to calculate the latitude and longitude using a formula. The transformation parameters are derived from the georeferenced information of the image, thus obtaining the target geographic coordinate set containing the building boundary points.

[0091] S220. Based on the target geographic coordinate set, obtain the precise geographic coordinates and official height data of the landmark building.

[0092] Based on the target geographic coordinate set, the precise geographic coordinates and official height data of the landmark building are obtained. Specifically, firstly, a subset of coordinates corresponding to the landmark building in the target geographic coordinate set is identified, such as selecting the boundary coordinates of the Eiffel Tower. Then, public databases such as Google Earth or official mapping bureau data are queried to extract its precise latitude and longitude, such as 48.8584°N, 2.2945°E, and official height data, such as 324 meters. The query process involves API calls to ensure data accuracy to support subsequent alignment.

[0093] S230. Align the precise geographic coordinates with the target geographic coordinate set to determine the target pixel coordinate region.

[0094] The precise geographic coordinates are aligned with the target geographic coordinate set to determine the target pixel coordinate region. Specifically, the least squares method is used for registration. For example, the offset vector between the precise coordinates and the set coordinates is calculated, and the rotation, translation, and scaling parameters are iteratively optimized to maximize the overlap between the two. The registration process includes feature point matching, such as selecting building vertices as control points, calculating residuals and adjusting them until the error is less than 0.1 meters, thereby determining the aligned pixel region. For example, the precise coordinates of the Eiffel Tower are mapped back to the corresponding pixel block in the image.

[0095] S240. Based on the target pixel coordinate region and official height data, construct the mapping relationship between pixel coordinates and physical height, calculate the elevation conversion coefficient, and construct a landmark height benchmark library containing the elevation conversion coefficient.

[0096] The formula for calculating the elevation conversion factor is: (2) In formula (2), This is the elevation conversion factor, in meters per pixel, calculated from the landmark calibration. The actual height of the landmark is shown in meters and is derived from official map data. The length of the landmark shadow in pixels is measured from satellite imagery. The control logic of formula (2) is to calculate the actual physical height per unit pixel by the ratio of the actual height of the landmark building to the pixel length of its shadow in the image, thus establishing a direct conversion relationship between image pixel length and physical elevation. The actual height of the building can be obtained by multiplying the pixel length of the building shadow by this coefficient. Formula (2) introduces the true value of the landmark to construct an absolute elevation benchmark, breaking through the limitations of traditional shadow height measurement that relies on empirical coefficients and has systematic deviations; by using landmark buildings of known height to calibrate the elevation conversion coefficient, errors caused by factors such as the imaging angle and resolution of satellite imagery can be eliminated, significantly improving the accuracy and reliability of building height measurement, and providing accurate elevation data for the subsequent completion of GIS building vector information.

[0097] Based on the target pixel coordinate region and the official height data, a mapping relationship between pixel coordinates and physical height is constructed. Least square fitting is performed using at least three landmarks within the block as a benchmark to eliminate systematic biases in shadows caused by differences in image phase and season. Elevation conversion coefficients are calculated, and a landmark height benchmark library containing these coefficients is constructed. Specifically, firstly, elevation-related features are extracted within the pixel region, such as generating DEM data from stereo image pairs. Then, a mapping model is established, for example, using linear regression to fit the relationship between the landmark shadow pixel length and the actual landmark height, and the elevation conversion coefficients are calculated. ,in For example, for a landmark with a true height of 324 meters, if the landmark shadow pixel length is 100 units, then... =3.24, and then the elevation transformation coefficients of multiple landmarks are stored in the benchmark database. For example, the database table contains landmark ID, coordinates, and This value provides a reference for estimating building height. Through this method, precise integration of coordinates and height is achieved.

[0098] Furthermore, the intelligent completion method for GIS building vector information with integrated landmark calibration provided in this embodiment includes step S300 as follows: S310. Determine the shadow search range. The shadow search range is determined by the elevation conversion coefficient combined with the initial building outline vector data.

[0099] First, the shadow search area is determined by combining the elevation conversion factor with the initial building outline vector data. Specifically, the elevation conversion factor is a value obtained from a previously established landmark height benchmark library, used to convert pixel units into actual physical heights, while the initial building outline vector data records the boundary coordinates of the building in the image. For example, when processing a satellite remote sensing image, the pixel width of the building outline is multiplied by the elevation conversion factor to estimate the possible shadow projection distance, thereby defining a rectangular area extending along the sun's azimuth from the building's base as the search area. This method ensures the accuracy of the search area and avoids interference from irrelevant areas.

[0100] S320. Extract the initial shadow area within the shadow search range and obtain the solar elevation angle and solar azimuth angle from the image metadata corresponding to the initial shadow area.

[0101] The initial shadow region is extracted within the shadow search area, and the solar elevation angle and solar azimuth angle are obtained from the image metadata. Specifically, the initial shadow region is extracted using a threshold segmentation method. For example, the grayscale values ​​of pixels within the range are analyzed, and a grayscale threshold, such as 50, is set. Pixels with grayscale values ​​lower than this threshold are clustered as shadow candidate regions. At the same time, the solar elevation angle, such as 30 degrees, and the solar azimuth angle, such as 135 degrees southeast, are read from the image's metadata file. These parameters are obtained from satellite sensor records during imaging, thus providing necessary illumination condition information for subsequent processing. This extraction process not only relies on image processing technology but also integrates metadata to improve accuracy.

[0102] S330. The initial shadow area is processed according to the solar azimuth angle to obtain the target shadow area.

[0103] The initial shadow region is processed based on the solar azimuth angle to obtain the target shadow region. Specifically, this processing involves morphological operations and directional filtering. For example, erosion and dilation operations are first applied to remove noise. Then, the expected projection direction of the shadow is calculated based on the solar azimuth angle, such as extending the building outline along a 135-degree line segment. The connected shadow regions consistent with the expected projection direction are then filtered out, thereby eliminating shadow interference caused by non-buildings and obtaining an accurate target shadow region. This method optimizes the purity of the region through directional constraints, ensuring the reliability of the shadow data.

[0104] S340. Calculate and solve the preliminary height estimate corresponding to the building height based on the physical length of the shadow of the target shadow area and the solar altitude angle. The formula for calculating the preliminary height estimate is as follows: (3) In formula (3), This is a preliminary height estimate, in meters (m). The physical length of the shadow, in meters, is calculated by multiplying the shadow pixel length in the image by the elevation conversion factor. get; The solar altitude angle, in rad or °, is derived from satellite image metadata. The control logic of formula (3) is based on the shadow geometry to calculate the building height. The building height, the physical length of the shadow, and the solar altitude angle form a right triangle, with the building height as the opposite side and the physical length of the shadow as the adjacent side. The ratio of the two is the tangent of the solar altitude angle. Therefore, the building height is equal to the physical length of the shadow multiplied by the tangent of the solar altitude angle. Formula (3) directly calculates the building height through the shadow geometry of a single image, breaking through the limitations of traditional height measurement that relies on multi-view stereo image pairs and has high data acquisition costs. It can complete the building height estimation using only a single satellite image and its metadata, significantly reducing data dependence and processing complexity, and providing an efficient and low-cost way to acquire elevation data for GIS building vector information completion.

[0105] A preliminary height estimate is calculated based on the physical length of the shadow in the target shadow area and the solar altitude angle. Specifically, the physical length of the shadow is obtained by measuring the pixel distance in the target area, such as the Euclidean distance from the building base to the end of the shadow. Then, using trigonometric relationships, such as height equals the physical length of the shadow multiplied by the tangent of the solar altitude angle, the preliminary height of the building is calculated. This calculation provides an efficient basis for estimation. The preliminary height estimate of the building is achieved through the above method.

[0106] Preferably, the intelligent completion method for GIS building vector information with integrated landmark calibration provided in this embodiment includes step S400 as follows: S410. Calculate the mapping deviation between the preliminary height estimate and the corresponding landmark reference height in the landmark height benchmark library, and determine whether the mapping deviation exceeds the preset deviation threshold. The formula for calculating the mapping relationship deviation is: (4) In formula (4), This represents the mapping deviation, expressed in meters (m). This is a preliminary height estimate; The reference height is in meters (m). The control logic of formula (4) is to quantify the mapping relationship deviation between the preliminary estimated height and the reference height of the landmark by calculating the absolute difference between the preliminary estimated height and the reference height of the landmark. If the mapping relationship deviation exceeds the preset deviation threshold, it indicates that there is an error in the current elevation conversion coefficient or the measurement of the physical length of the shadow, and the landmark calibration process needs to be started; if the mapping relationship deviation is within the preset deviation threshold range, the height estimation result is determined to be reliable. Formula (4) breaks through the limitation of the traditional shadow height measurement method lacking an error feedback mechanism by introducing the reference height of the landmark library to verify the mapping relationship deviation of the preliminary height estimation value; it can automatically identify and correct systematic deviations caused by factors such as shadow segmentation error and solar altitude angle deviation, significantly improving the accuracy and reliability of building height estimation, and providing high-quality verified elevation data for subsequent GIS building vector information completion.

[0107] First, the mapping deviation is calculated by comparing the preliminary height estimate with the corresponding reference height in the landmark height benchmark database. Specifically, the landmark height benchmark database is a pre-built database containing precise height information for known buildings. For example, for a landmark high-rise building in a city, such as the Shanghai Tower, its landmark reference height is 632.4 meters. If the preliminary height estimate is 631 meters, the deviation is calculated as the difference between the preliminary height estimate and the landmark reference height. If the difference is 1.4 meters, then it is determined whether this difference exceeds a preset deviation threshold, such as 1.2 meters. This deviation assessment ensures the reliability of the estimate and provides a basis for subsequent calibration.

[0108] S420. If the mapping deviation exceeds the preset deviation threshold, calculate the height calibration factor and obtain the corrected height value through calculation.

[0109] If the mapping deviation exceeds a preset deviation threshold, a height calibration factor is introduced for correction. Specifically, the height calibration factor is a parameter calculated by the ratio of the mapping difference to the reference height in the landmark library. For example, when the deviation rate is 5%, the calibration factor can be set to 1.05. Then, the initial height estimate is multiplied by this height calibration factor to obtain the corrected height value. For example, the original 600 meters multiplied by 1.05 yields 630 meters. This calculation process integrates accurate data from the benchmark library, thereby improving the accuracy of height estimation. In actual remote sensing image processing, this calibration can effectively compensate for errors in illumination parameters or shadow extraction, ensuring that the corrected value is closer to the actual building height.

[0110] S430. Extract the spatial features of the building outline corresponding to the corrected height value to obtain the adaptive floor height benchmark.

[0111] The spatial features of the building outline corresponding to the corrected height value are extracted to obtain an adaptive floor height benchmark. Specifically, the spatial features of the building outline include the geometry and location coordinates of the outline. For example, the area and boundary curve of the building base are extracted from vector data. Then, based on these features, an adaptive floor height benchmark is matched from a preset floor height database. For example, the average floor height of residential buildings is 3 meters, while that of commercial buildings is 4 meters. This method of acquisition takes into account the diversity of building types, so that the floor height benchmark is more in line with the specific scenario.

[0112] S440. Based on the corrected height value and the adaptive floor height benchmark, the floor conversion quotient is calculated.

[0113] The floor number conversion quotient is calculated based on the corrected height value and the adaptive floor height benchmark. Specifically, the floor number conversion quotient is obtained by dividing the corrected height by the floor height benchmark. For example, a corrected height of 630 meters divided by a floor height of 3 meters yields 210. This simple division operation provides a preliminary quantification of the number of floors.

[0114] S450. Perform a rounding operation on the floor number conversion quotient to obtain the rounded floor number result. Combine this with the corrected height value to determine the calibrated building height information and building floor number information.

[0115] The formula for calculating the number of building floors is: (5) In formula (5), The number of building floors is a dimensionless integer obtained by rounding down. The height of the building after calibration, in meters; This serves as an adaptive floor height benchmark, in meters, and is set adaptively according to the building type. The floor function is a round-down operator. The control logic of formula (5) is to obtain the theoretical number of floors of the building by dividing the calibrated building height by the adaptive floor height benchmark, and then round down the result to ensure that the actual number of floors is obtained. The adaptive floor height benchmark is automatically adjusted according to the building type. Different types of buildings use different floor height standards to improve the accuracy of floor number estimation. Formula (5) breaks through the limitation of traditional floor number estimation using a uniform floor height and large error by combining landmark calibration and adaptive floor height benchmark. It can simultaneously realize the joint completion of building height and floor number, significantly improve the integrity and accuracy of GIS building vector information, and provide reliable building attribute data for applications such as urban planning and 3D modeling.

[0116] The floor number conversion quotient is rounded to obtain the floor number result, which is then combined with the corrected height to determine the calibrated information. Specifically, rounding can be done using the rounding method, such as rounding 210.3 to 210 floors. Then, the height and floor number are integrated to output the final building information, thus completing the entire estimation process.

[0117] Furthermore, the intelligent completion method for GIS building vector information with integrated landmark calibration provided in this embodiment includes step S500: S510. Based on the calibrated building height and number of floors information, construct a three-dimensional geospatial bounding box according to the building outline, and generate a geospatial buffer based on the three-dimensional geospatial bounding box.

[0118] Based on the calibrated building height and number of floors, a three-dimensional geospatial bounding box is first constructed according to the building outline. Specifically, the building outline is two-dimensional boundary data extracted from remote sensing images. By extending the calibrated height as the Z-axis dimension, a minimum three-dimensional rectangular box is formed that encloses the building. For example, for a high-rise residential building with a rectangular base and a height of 150 meters, the coordinate range of the three-dimensional geospatial bounding box covers the spatial volume from the ground to the top, thus providing a geometric basis for subsequent buffer generation.

[0119] A geospatial buffer is generated based on the aforementioned 3D geospatial bounding box. Specifically, the geospatial buffer is an area formed by extending a certain distance around the 3D geospatial bounding box, for example, setting the extension distance to 50 meters, to accommodate possible positioning errors. This generation process utilizes geographic information system tools such as ArcGIS for computation. The geospatial buffer can be viewed as an expanded 3D spatial volume used to capture nearby points of interest, thereby ensuring the comprehensiveness of spatial correlation. In practical urban planning applications, this geospatial buffer can effectively handle the overlap problem in densely built-up areas. By adjusting the extension parameters, it can adapt to different building densities; for example, it can be extended further in commercial areas to cover more correlated data.

[0120] S520. Perform spatial intersection operation between the public point of interest data and the geospatial buffer to obtain a spatial association candidate set.

[0121] The publicly available point-of-interest (POI) data is spatially intersected with the geospatial buffer to obtain a spatial association candidate set. Specifically, the publicly available POI data comes from map services such as Amap (Gaode Maps) and includes location coordinates and attributes. The intersection operation checks whether the POIs fall within the buffer. For example, if the coordinates of a POI are within the boundary of the geospatial buffer, it is included in the spatial association candidate set. This operation is accelerated by a spatial indexing algorithm, thereby quickly filtering relevant point sets.

[0122] S530. Calculate the spatial weight value corresponding to the spatial association candidate set. When the spatial weight value is greater than the preset weight threshold, select the target interest point as the spatial matching result.

[0123] The formula for calculating spatial weight is: (6) In formula (6), This represents the spatial weight value, with a range of values ​​of [value range missing]. ; The geographic distance from the POI to the building outline is in meters. The control logic of formula (6) is to calculate the weight by using the reciprocal function of the geographic distance. The closer the distance, the higher the spatial weight value and the better the matching degree; the farther the distance, the lower the spatial weight value and the worse the matching degree. The spatial weight value decreases monotonically with the increase of distance. When the distance is 0, the spatial weight value is 1, and when the distance approaches infinity, the spatial weight value approaches 0. Formula (6) achieves robust association between POI and building by geographic distance weighting, breaking through the limitation of traditional matching based solely on the nearest distance, which is easily affected by noise points. By setting a preset weight threshold, invalid matches that are too far away can be effectively filtered out, which significantly improves the accuracy and stability of the association between POI and building, and provides a reliable association basis for attribute completion of GIS building vector information.

[0124] The spatial weight values ​​corresponding to the spatial association candidate set are calculated. When the spatial weight value is greater than a preset weight threshold, the target point of interest is selected as the spatial matching result. Specifically, the spatial weight value is calculated by weighting distance and overlap. For example, the weight formula considers the Euclidean distance from the point of interest to the building center. The preset weight threshold is set to 0.8. If the spatial weight value of a candidate point is 0.9, it is selected. This method prioritizes the nearest and most relevant point, thereby improving matching accuracy. In multi-candidate scenarios, a ranking mechanism can be introduced to further optimize the selection.

[0125] S540. Extract the names of points of interest corresponding to the spatial matching results as building name attributes, and generate semantic attribute data after association.

[0126] The names of points of interest corresponding to the spatial matching results are extracted as building name attributes to generate associated semantic attribute data. Specifically, the names, such as "Oriental Pearl Tower," are extracted from the metadata of the selected points of interest and integrated with building height and number of floors into structured data, thus forming a complete semantic description for map annotation or urban information systems. Through the above process, semantic enhancement of building information is achieved.

[0127] Preferably, the intelligent completion method for GIS building vector information with integrated landmark calibration provided in this embodiment includes step S600 as follows: S610. Obtain the building top image and facade logo image corresponding to the unmatched building.

[0128] First, for buildings that failed to be successfully matched in the initial spatial association, their top and facade signage images are obtained. These images are usually obtained from drone aerial photography or street view acquisition systems. For example, for an unnamed commercial building located in the city center, the top image captures rooftop signage such as the company logo, while the facade signage image records the sign or door number at the entrance, thus providing a visual data foundation for subsequent identification.

[0129] S620. The optical character recognition model is used to recognize and process the images of the building top and facade signs to obtain a set of candidate building names.

[0130] These images are processed using an optical character recognition (OCR) model to extract text information. Specifically, an OCR model is a deep learning-based image processing technique that analyzes image pixels through convolutional neural networks to identify character sequences. For example, an OCR model such as Tesseract or a custom-trained OCR engine will first preprocess the building top image and facade signage image, including grayscale conversion and noise removal, then segment the character regions, and then generate candidate text through feature extraction and classification steps. In this process, for blurry signs in the top image, the OCR model outputs multiple candidate words such as "Huatai Building" or "Huatai Tower," while the doorplates in the facade image provide clearer name fragments. Finally, these are summarized to form a set of candidate building names, such as "Huatai Center" or "Huatai Tower," thereby improving the robustness of recognition through this multi-source image fusion.

[0131] S630. When the confidence level of the candidate building name set is greater than the preset threshold, it is determined as the target building name.

[0132] The formula for calculating the overall confidence score of a text string is: (7) In formula (7), The overall confidence score of the text string, with a value range of [value range missing]. This value is used to quantify the reliability of recognizing the entire building name text. The closer the value is to 1, the lower the probability of the building name being recognized incorrectly. The first text string output by the OCR model The independent recognition probability of each character, with the probability of recognizing a single character ranging from [value missing]. This characterizes the reliability of the model's prediction of the character category at that position; The total number of characters in the current text string containing the building name to be determined is a positive integer. The control logic of formula (7) is to read the individual character recognition probability predicted by the model character by character from the candidate building name text output by OCR recognition. The recognition probabilities of all characters in the text string are multiplied sequentially to obtain the overall recognition confidence of the building name text string. A global confidence threshold is set. When the overall confidence of a candidate building name is greater than this preset threshold, the OCR recognition result is deemed reliable, and the candidate name is designated as the target building name and written into the GIS vector attribute library. If the overall confidence is lower than the threshold, the recognition result is discarded to avoid polluting the GIS database with misspelled or misrecognized names. From a probabilistic perspective: if the recognition probability of any single character in the text is low, the overall confidence of the entire text will decrease significantly after multiplication, accurately filtering out low-quality text results with misrecognition, omissions, or fuzzy recognition. Under the same character recognition probability, the overall confidence of short name texts is relatively higher, meeting the reliability screening requirements for short place names and building proper names. Formula (7) calculates the overall confidence of the text by multiplying the probability of each character. This differs from the traditional coarse-grained confidence calculation method that takes the average probability or the highest probability of a single character. It can accurately capture local character recognition anomalies and efficiently filter out erroneous recognition results caused by blurred, occluded, and distorted roof text, thus greatly improving the data accuracy of building name attribute completion. It breaks through the technical limitation of traditional GIS building names relying solely on spatial matching of third-party POI databases. For buildings without POI data, in remote areas, or newly built buildings not yet included in the database, it can directly rely on satellite orthophoto image roof text OCR recognition to automatically extract and complete the building name, thus filling the gap in building attribute data in areas with missing POIs. By combining the OCR text confidence screening mechanism with the spatial POI matching mechanism, it achieves full coverage collection of building name attributes in all scenarios such as urban built-up areas, remote suburban areas, and newly built factory areas. This greatly improves the attribute integrity of the GIS building vector dataset, reduces the manpower and time costs of manual field collection and attribute entry, and improves the automation efficiency and usability of geographic information data updates.

[0133] When the overall confidence score of the text strings in the candidate set exceeds a preset threshold, such as 0.85, the target building name is determined. Specifically, the calculation of the overall confidence score of the text strings involves the similarity and frequency of occurrence of each candidate. For example, the matching degree with the reference data is evaluated through a string matching algorithm. If "Huatai Building" appears repeatedly in multiple images and has the highest score, it is selected as the target, thereby avoiding errors caused by low-quality recognition.

[0134] S640. Retrieve the complete attribute record from the multi-source database by the target building name.

[0135] The system retrieves complete attribute records from multiple databases using the target building name, such as Baidu Maps or government property databases, to obtain records of the building's construction year and usage, such as "commercial office".

[0136] S650. Integrate and merge the completed attribute records and the target building name to determine the complete set of building semantic attributes.

[0137] These completed attribute records are integrated and merged with the target building name to generate a complete set of building semantic attributes, such as name, address, height, and function, thereby supporting data integrity in urban planning.

[0138] Furthermore, the intelligent completion method for GIS building vector information with integrated landmark calibration provided in this embodiment includes step S700 as follows: S710. Acquire orthophotos and road network data, extract road intersections and river inflection points from the orthophotos, and form a set of candidate control points.

[0139] First, orthophotos and road network data are acquired. These orthophotos and road network data come from satellite remote sensing systems or geographic information platforms and are used to construct the geometric basis of the urban area. For example, for an industrial park with a dense river network, orthophotos capture the road layout and river morphology, while road network data provides vectorized road lines, thus providing data support for subsequent extraction. Specifically, the process of extracting road intersections and river bends from orthophotos to form a candidate control point set involves image processing techniques. Orthophotos are geometrically corrected vertical projection images that can eliminate distortions caused by terrain undulations. The extraction method uses edge detection algorithms such as the Canny operator to identify road edges and then calculates the coordinates of intersections. For river bends, curve fitting techniques such as spline interpolation are used to find locations with significant curvature changes, thereby generating a set containing hundreds of points. For example, in a park scene, the intersections of main roads and river bends are extracted as candidates for subsequent matching.

[0140] S720. The point cloud registration method is used to match the candidate control point set with the road network data and calculate the spatial systematic offset.

[0141] The formula for calculating the spatial systematic offset is: (8) In formula (8), It is a spatial systematic offset, in meters, containing coordinate offset components in the X and Y directions of the plane. It is used to characterize the global positional deviation of satellite orthophotos relative to existing GIS road network vector data, and serves as a correction parameter for global coordinate translation correction. The total number of effective control point pairs after bidirectional matching and gross error removal is a positive integer. Control points are selected from road network nodes with high feature recognition, such as road intersections, road turning points, and road inflection points. For the first The two-dimensional plane coordinates of the group matching control points in the standard GIS road network geographic coordinate system; For the first The control points for group matching are obtained from the geographic plane coordinates after the pixel coordinates of satellite imagery are transformed. The control logic of formula (8) is to extract road feature control points from the GIS road network vector dataset and the satellite orthophoto after image segmentation, respectively, and complete the pairing of control points with the same name through the feature similarity matching algorithm, eliminate gross errors in mismatches, and filter to obtain the correct control points. A set of high-precision effective control point pairs is established. The two-dimensional coordinate difference between the corresponding control points in the GIS reference coordinate system and the image geographic coordinate system is calculated for each pair to obtain the position deviation vector of a single control point. The arithmetic mean of the coordinate deviation vectors of all effective control point pairs is calculated to obtain the global spatial systematic offset vector. This vector can eliminate the global overall translation error caused by satellite imaging distortion and coordinate transformation. The coordinates of all building outline vertices obtained from image segmentation are superimposed with this offset vector to complete the overall coordinate translation correction, thereby achieving high-precision spatial registration of building outlines with existing GIS geographic reference and road network vector data. Formula (8) achieves automated coordinate registration based on the natural feature control points of the road network, abandoning the traditional manual point selection and calibration method. This saves a lot of manual field and office operation costs and avoids human error introduced by subjective point selection, realizing automated processing of the entire process of remote sensing image geographic correction. It uses the average deviation of multiple effective control points to solve the global offset vector, which can suppress the gross interference caused by a small number of mismatched control points, ensure the consistency and stability of coordinate correction in the entire area, and greatly improve the spatial positioning accuracy of building outlines. It uses the authoritative GIS road network as the spatial reference to complete the coordinate offset correction, effectively solving the problems of spatial misalignment and topological conflict between the extracted building outlines and the existing geographic database. It constructs a unified and accurate spatial reference for building vector boundary entry and building name attribute association matching, and comprehensively improves the spatial reliability and data reuse value of GIS building vector completion results.

[0142] Point cloud registration is used to match control point pairs between the candidate control point set and the road network data and to calculate the spatial systematic offset. Point cloud registration is a three-dimensional data alignment technique that minimizes the distance between points using an iterative nearest point algorithm such as ICP (Iterative Closest Point) to achieve matching. The specific process includes initial coarse registration based on feature point correspondence, followed by fine registration to calculate the transformation matrix. The offset is obtained by the average Euclidean distance. For example, if a systematic offset of 5 meters is found after matching, it indicates that there is an overall deviation between the image and the road network, thus providing a quantitative basis for correction.

[0143] S730. If the spatial systematic offset is greater than the preset threshold, the building outline in the complete building semantic attribute set is corrected according to the spatial systematic offset to obtain the aligned building outline.

[0144] If the spatial system offset is greater than a preset threshold such as 3 meters, the building outline in the complete set of building semantic attributes will be corrected according to the offset to obtain an aligned building outline. The correction method involves affine transformation to apply the offset vector to adjust the outline coordinates. For example, a translation matrix is ​​applied to the outline polygon of a factory building to align it with the road network and avoid positional errors from affecting the planning.

[0145] S740: Based on aligned building outlines, output standardized geographic information system vector data that integrates geometric outlines and calibration height information.

[0146] Based on aligned building outlines, standardized geographic information system vector data that integrates geometric outlines and calibrated height information is output, thereby improving data accuracy and supporting urban management.

[0147] Please see Figure 2This invention provides an intelligent GIS building vector information completion system integrating landmark calibration, used to implement the aforementioned intelligent GIS building vector information completion method integrating landmark calibration. It includes a building pixel coordinate boundary information recording module 10, an elevation conversion coefficient determination module 20, a preliminary height estimation value acquisition module 30, a building height and number of floors information determination module 40, a semantic attribute data generation module 50, a building semantic attribute set determination module 60, and a standardized geographic information system vector data output module 70. The building pixel coordinate boundary information recording module 10 is used to obtain a cloud-free orthophoto base map based on satellite image preprocessing results, perform pixel-level classification tasks using an image segmentation model, and introduce multi-scale feature fusion. The module enhances the accuracy of building area recognition by generating initial building outline vector data, which contains building pixel coordinate boundary information. The elevation conversion coefficient determination module 20 constructs a landmark height benchmark library based on the building pixel coordinate boundary information. This library obtains official height data and precise geographic coordinates of landmark buildings from a public map interface through data acquisition, establishes a mapping relationship between pixel coordinates and physical height, and determines the elevation conversion coefficient corresponding to each unit pixel. The preliminary height estimation module 30 extracts shadow areas within the initial building outline vector data range based on the elevation conversion coefficient, and then estimates the building height based on illumination parameters in the image metadata and shadow geometry. A preliminary estimation is performed to obtain a preliminary height estimate of the building. The building height and number of floors determination module 40 calculates a calibration factor and corrects the error in the preliminary height estimate if the deviation between the preliminary height estimate and the landmark height benchmark library exceeds a preset deviation threshold. It then combines this with adaptive floor height parameters for building type to convert the data and obtain the number of floors, thus determining the calibrated building height and number of floors. The semantic attribute data generation module 50 uses geospatial buffer analysis to spatially match the building outline with publicly available points of interest (POIs) data, extracting the names of successfully matched POIs as building name attributes to generate associated semantic attribute data. The semantic attribute set determination module 60 is used to identify the building top or facade markings in satellite imagery corresponding to the associated semantic attribute data using optical character recognition technology, extract the building names of buildings that have not been matched, and integrate multi-source data to complete the building use and construction year attributes, thereby determining a complete set of building semantic attributes. The standardized geographic information system vector data output module 70 is used to extract road intersections and river inflection points in orthophotos as control points using point cloud registration methods, register and align the building outlines in the complete set of building semantic attributes with the existing geographic information system road network data, eliminate spatial systematic offsets, and output standardized geographic information system vector data that combines geometric outlines and calibration height information.

[0148] The intelligent completion method and system for GIS building vector information with integrated landmark calibration provided in this embodiment achieves the following specific technical effects compared with the prior art: 1. This embodiment is based on a preprocessed cloudless orthophoto base map. It combines an image segmentation model with a multi-scale feature fusion mechanism to achieve pixel-level building area classification and recognition. This effectively enhances the feature representation capability of building targets against complex terrain backgrounds, suppresses the influence of interference elements such as vegetation, roads, and bare land on building boundary extraction, and accurately outputs initial building outline vector data containing precise pixel boundaries. This solves the problems of blurred boundaries and incomplete building outlines in traditional image vectorization, and provides a high-precision geometric foundation for subsequent elevation calculation and attribute completion.

[0149] 2. This embodiment relies on high-precision landmark data to construct a landmark height benchmark library, establishes a standardized mapping relationship between image pixel coordinates and real physical elevation, adaptively solves the pixel elevation conversion coefficient, realizes the accurate correlation between image pixel scale and geographic real scale, gets rid of the limitations of traditional benchmark-less estimation methods that rely on empirical parameters, and provides an authoritative and stable elevation reference system for building height calculation.

[0150] 3. This embodiment relies on the geometric mechanism of shadows and image illumination parameters to carry out preliminary estimation of building height. It makes full use of the spatial depth information contained in the shadows of satellite images to realize intelligent inference from two-dimensional images to three-dimensional elevation information, effectively explore the hidden depth features of images, make up for the lack of elevation dimension information in conventional two-dimensional vector data, and quickly complete the batch calculation of preliminary building elevation for a large area.

[0151] 4. This embodiment introduces a landmark benchmark deviation threshold discrimination mechanism, which adaptively calculates calibration factors and corrects errors for the elevation estimation results. At the same time, it combines adaptive floor height parameters of building type to realize intelligent conversion of building floor number, effectively eliminating the systematic deviation caused by single shadow estimation, adapting to the floor height difference characteristics of different business types of buildings, and simultaneously realizing accurate calibration of building height and floor number information, thereby improving the rationality and accuracy of three-dimensional building parameter calculation.

[0152] 5. This embodiment achieves spatial association matching between building outlines and public points of interest data through geospatial buffer analysis, and automatically assigns semantic attributes to building names based on spatial topological relationships, realizing the automatic association and binding of geometric vectors and business semantic information, solving the problem of geometric completeness and semantic missingness that is common in traditional GIS vector data, and improving the semantic richness of data.

[0153] 6. This embodiment combines optical character recognition technology to intelligently extract building identification text from images, specifically completes the name information of unmatched buildings, and integrates multi-source public data to improve extended attributes such as building use and construction year, constructing a complete set of building semantic attributes, realizing comprehensive completion of semantic information of building vector data, and greatly improving the integrity and usability of GIS building data.

[0154] 7. This embodiment uses stable features such as road intersections and river bends as registration control points. A point cloud registration algorithm is used to accurately align building vectors with road network baseline data, effectively eliminating spatial systematic offsets generated during image acquisition and vectorization. The result is standardized GIS vector data that is geometrically accurate, elevationally reliable, and semantically complete. This embodiment achieves intelligent batch completion and calibration of building vector geometry, 3D elevation, and multi-layer semantics throughout the entire process. It overcomes the shortcomings of traditional manual data entry and calibration, which are inefficient, error-prone, and have limited dimensions. It is suitable for large-scale urban GIS database updates, real-scene 3D modeling, and land spatial mapping, demonstrating a high degree of automation and engineering practicality.

[0155] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. A method for intelligently completing GIS building vector information by integrating landmark calibration, characterized in that, Includes the following steps: S100. Obtain a cloudless orthophoto base map based on the satellite image preprocessing results, perform pixel-level classification tasks using an image segmentation model, and introduce a multi-scale feature fusion module to enhance the accuracy of building area recognition, generating initial building outline vector data, wherein the initial building outline vector data records building pixel coordinate boundary information. S200. Based on the building pixel coordinate boundary information, a landmark height reference library is constructed. The landmark height reference library obtains the official height data and precise geographic coordinates of landmark buildings from a public map interface through data acquisition, establishes the mapping relationship between pixel coordinates and physical height, and determines the elevation conversion coefficient corresponding to a unit pixel. S300. Based on the elevation conversion coefficient, extract the shadow area within the initial building outline vector data range, and based on the lighting parameters in the image metadata and the shadow geometry, make a preliminary estimate of the building height to obtain a preliminary height estimate of the building. S400. If the mapping relationship between the preliminary height estimate and the landmark height reference library deviates beyond a preset deviation threshold, a calibration factor is calculated and the preliminary height estimate is corrected for error. The building type adaptive floor height parameter is combined with the conversion to obtain the number of building floors, and the calibrated building height information and building floor information are determined. S500: Based on the calibrated building height and number of floors information, the building outline is spatially matched with publicly available points of interest data through geospatial buffer analysis. The names of the successfully matched points of interest are extracted as building name attributes, and semantic attribute data after association is generated. S600 uses optical character recognition technology to identify the building top or facade markings in satellite imagery that correspond to the associated semantic attribute data, extracts the building names of buildings that have not been matched, and integrates multi-source data to complete the building use and construction year attributes, thus determining a complete set of building semantic attributes. S700 uses point cloud registration to extract road intersections and river inflection points from orthophotos as control points. It then registers and aligns the building outlines in the complete set of building semantic attributes with the existing geographic information system road network data to eliminate spatial systematic offsets and output standardized geographic information system vector data that combines geometric outlines and calibration height information.

2. The intelligent completion method for GIS building vector information with integrated landmark calibration as described in claim 1, characterized in that, Step S100 includes: S110. Obtain a cloudless orthophoto base map, wherein the cloudless orthophoto base map is generated by preprocessing satellite images and extracting cloudless orthophoto features; S120. Based on the cloudless orthophoto base map, a pixel classification matrix is ​​obtained using an image segmentation model, and multi-scale feature fusion is performed on the pixel classification matrix to generate a feature fusion map. S130. When the area of ​​the building connected region in the feature fusion map is greater than a preset area threshold, the edge pixels of the building connected region are extracted to determine the building outline. S140. Perform vectorization transformation on the building outline to generate initial building outline vector data that records the building pixel coordinate boundary information.

3. The intelligent completion method for GIS building vector information with integrated landmark calibration according to claim 1, characterized in that, Step S200 includes: S210. Obtain the target geographic coordinate set, which is obtained by coordinate transformation processing of the building pixel coordinate boundary information; S220. Based on the target geographic coordinate set, obtain the precise geographic coordinates and official height data of the landmark building; S230. Align the precise geographic coordinates with the target geographic coordinate set to determine the target pixel coordinate region; S240. Based on the target pixel coordinate region and the official height data, construct the mapping relationship between pixel coordinates and physical height, calculate the elevation conversion coefficient, and construct a landmark height reference library containing the elevation conversion coefficient.

4. The intelligent completion method for GIS building vector information with integrated landmark calibration according to claim 3, characterized in that, Step S300 includes: S310. Determine the shadow search range, wherein the shadow search range is jointly determined by the elevation conversion coefficient and the initial building outline vector data; S320. Extract the initial shadow region within the shadow search range, and obtain the solar altitude angle and solar azimuth angle from the image metadata corresponding to the initial shadow region. S330. Process the initial shadow area according to the solar azimuth angle to obtain the target shadow area; S340. Calculate the preliminary height estimate corresponding to the building height by performing calculations based on the physical length of the shadow of the target shadow area and the solar altitude angle.

5. The intelligent completion method for GIS building vector information with integrated landmark calibration according to claim 1, characterized in that, Step S400 includes: S410. Calculate the mapping deviation between the preliminary height estimate and the corresponding landmark reference height in the landmark height benchmark library, and determine whether the mapping deviation exceeds the preset deviation threshold. S420. If the deviation of the mapping relationship exceeds the preset deviation threshold, calculate the height calibration factor and obtain the corrected height value through calculation. S430. Extract the spatial features of the building outline corresponding to the corrected height value to obtain the adaptive floor height benchmark; S440. Calculate the floor number conversion quotient based on the corrected height value and the adaptive floor height reference. S450. Perform a rounding operation on the floor number conversion quotient to obtain the rounded floor number result, and combine it with the corrected height value to determine the calibrated building height information and building floor number information.

6. The intelligent completion method for GIS building vector information with integrated landmark calibration according to claim 5, characterized in that, Step S500 includes: S510. Based on the calibrated building height and number of building floors, construct a three-dimensional geospatial bounding box according to the building outline, and generate a geospatial buffer based on the three-dimensional geospatial bounding box. S520. Perform a spatial intersection operation between the publicly available point of interest data and the geospatial buffer to obtain a spatial association candidate set; S530. Calculate the spatial weight value corresponding to the spatial association candidate set. When the spatial weight value is greater than the preset weight threshold, select the target interest point as the spatial matching result. S540. Extract the point of interest names corresponding to the spatial matching results as building name attributes, and generate associated semantic attribute data.

7. The intelligent completion method for GIS building vector information with integrated landmark calibration according to claim 1, characterized in that, Step S600 includes: S610. Obtain the building top image and facade logo image corresponding to the unmatched building; S620. Using an optical character recognition model, the image of the building top and the image of the facade logo are processed to obtain a set of candidate building names; S630. When the confidence level of the candidate building name set is greater than a preset threshold, it is determined as the target building name. S640. Retrieve the complete attribute record from the multi-source database using the target building name; S650. Integrate and merge the completed attribute records with the target building name to determine a complete set of building semantic attributes.

8. The intelligent completion method for GIS building vector information with integrated landmark calibration according to claim 7, characterized in that, Step S700 includes: S710. Acquire orthophotos and road network data, and extract road intersections and river inflection points from the orthophotos to form a set of candidate control points. S720. Using a point cloud registration method, the candidate control point set is matched with the road network data to calculate the spatial systematic offset. S730. If the spatial systematic offset is greater than a preset threshold, the building outline in the complete building semantic attribute set is corrected according to the spatial systematic offset to obtain an aligned building outline. S740. Based on the aligned building outline, output standardized geographic information system vector data that integrates geometric outline and calibration height information.

9. The intelligent completion method for GIS building vector information with integrated landmark calibration according to claim 8, characterized in that, In step S720, the formula for calculating the spatial systematic offset is: ； in, For spatial systematic offset, This represents the total number of valid control point pairs after two-way matching and gross error removal. For the first The two-dimensional plane coordinates of the group matching control points in the standard GIS road network geographic coordinate system. For the first The group matching control points are obtained from the geographic plane coordinates after the satellite image pixel coordinates are transformed.

10. A GIS building vector information intelligent completion system integrating landmark calibration, used to implement the GIS building vector information intelligent completion method integrating landmark calibration as described in any one of claims 1 to 9, characterized in that, include: The building pixel coordinate boundary information recording module is used to obtain a cloudless orthophoto base map based on the satellite image preprocessing results, perform pixel-level classification tasks using an image segmentation model, and introduce a multi-scale feature fusion module to enhance the building area recognition accuracy, and generate initial building outline vector data, which records building pixel coordinate boundary information. The elevation conversion coefficient determination module is used to construct a landmark height benchmark library based on the building pixel coordinate boundary information. The landmark height benchmark library obtains the official height data and precise geographic coordinates of landmark buildings from a public map interface through data acquisition, establishes the mapping relationship between pixel coordinates and physical height, and determines the elevation conversion coefficient corresponding to a unit pixel. The preliminary height estimation module is used to extract the shadow area within the initial building outline vector data range based on the elevation conversion coefficient, and to make a preliminary estimation of the building height based on the lighting parameters in the image metadata and the shadow geometry, thereby obtaining a preliminary height estimation value of the building. The building height and number of floors information determination module is used to calculate a calibration factor and correct the error of the preliminary height estimate if the mapping relationship between the preliminary height estimate and the landmark height benchmark library deviates beyond a preset deviation threshold. It then combines the building type adaptive floor height parameter for conversion to obtain the building number of floors and determines the calibrated building height information and building number of floors information. The semantic attribute data generation module is used to spatially match the building outline with public point of interest data based on the calibrated building height and number of floors information through geospatial buffer analysis, extract the names of successfully matched points of interest as building name attributes, and generate associated semantic attribute data. The building semantic attribute set determination module is used to identify the building top or facade marking text in satellite imagery corresponding to the associated semantic attribute data using optical character recognition technology, extract the building name of buildings that have not been matched, and integrate multi-source data to complete the building use and construction year attributes, thereby determining the complete building semantic attribute set. The standardized geographic information system vector data output module is used to extract road intersections and river inflection points from orthophotos as control points using point cloud registration methods. It registers and aligns the building outlines in the complete set of building semantic attributes with the existing geographic information system road network data, eliminates spatial systematic offsets, and outputs standardized geographic information system vector data that combines geometric outlines and calibration height information.

Images