A method and system for automatically generating a three-dimensional white model of a building based on multi-source satellites

Abstract

This invention discloses an automatic method and system for generating 3D white models of buildings based on multi-source satellites. The method includes: acquiring dual-line array stereo image data and laser altimetry data of a target area to form a stereo observation dataset; fusing the processed dual-line array stereo image data and laser altimetry data; automatically identifying and finely extracting the roof outline of each building within the target area and outputting it as vector polygon data; extracting the vector outline of each building and performing spatial analysis on the digital surface model; and using the difference between its corresponding top elevation and bottom elevation as the building height to form a 3D white model dataset of buildings in the target area. This method achieves automated, high-precision, and robust extraction of building outlines and heights, automatically generating structurally complete and semantically rich city-level 3D white models, and enabling rapid, accurate, and large-scale acquisition of large-scale urban 3D information.

Description

Technical Field

[0001] This invention relates to the field of building modeling technology, specifically to a method and system for automatically generating 3D white models of buildings based on multi-source satellites. Background Technology

[0002] Currently, mainstream urban 3D modeling technologies mainly rely on aerial photogrammetry and airborne LiDAR. Aerial photogrammetry reconstructs 3D information by capturing overlapping images from different angles, while airborne LiDAR directly acquires high-precision 3D point clouds by actively emitting laser pulses. With the development of my country's domestic stereo mapping satellite constellations such as Gaofen-7 and Ziyuan-3, their dual-line array stereo cameras and laser altimeters have made it possible to acquire global, all-weather stereo observation data from space. Existing technologies typically use this type of satellite data to produce large-scale terrain-level digital surface model products. However, in densely built-up and structurally complex urban areas, existing satellite-based automated processing technologies have significant shortcomings:

[0003] (1) The automatic extraction accuracy and completeness of building outlines are not high, and they are easily affected by shadows, vegetation and complex roof structures;

[0004] (2) The automated inversion accuracy of building height is insufficient, making it difficult to stably and accurately separate the top and bottom elevations of buildings from satellite stereo images;

[0005] (3) The automation and intelligence of the overall processing flow are limited, making it difficult to effectively integrate multi-source satellite data to leverage their respective advantages. As a result, the final generated 3D white model of the building is unable to meet the urgent needs of emerging applications such as smart low-altitude management for the rapid construction and updating of city-level 3D data bases in terms of accuracy, efficiency and reliability. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for automatically generating three-dimensional white models of buildings based on multi-source satellites, so as to solve the technical problems of high data acquisition cost, low efficiency, many limitations, insufficient accuracy of automated satellite data processing, and insufficient utilization of multi-source satellite data fusion in the prior art.

[0007] To solve the above-mentioned technical problems, the present invention specifically provides the following technical solution:

[0008] The first aspect of this invention provides a method for automatically generating three-dimensional white models of buildings based on multi-source satellites, comprising the following steps:

[0009] Acquire dual-line array stereo image data and laser altimetry data of the target area, and perform radiometric correction, geometric correction and joint adjustment on the acquired raw data to form a high-precision stereo observation dataset with unified geographic reference;

[0010] The fused dual-line array stereo image data and the laser altimetry data are used to generate a digital surface model covering the target area through stereo matching and laser point cloud interpolation. The laser altimetry data is used to correct and enhance the elevation accuracy of the initial DSM generated from the stereo image.

[0011] Using the digital surface model and related multispectral images as input, a deep learning semantic segmentation model is used to automatically identify and refine the roof outline of each building in the target area, and output as vector polygon data. The deep learning semantic segmentation model is optimized for the spectral, texture and elevation features of the building tops in the satellite images.

[0012] Extract the vector outline of each building, perform spatial analysis on the digital surface model, calculate the statistical characteristic value of the elevation value of the digital surface model within the outline of the building, determine the elevation value representing the main body of the roof as the top elevation of the building, and determine the base elevation of the building in contact with the ground by analyzing the elevation of the area adjacent to the edge of the building outline, which is used as the bottom elevation of the building.

[0013] Using the vector outline of each building as the base, and the difference between its corresponding top elevation and bottom elevation as the building height, vertical stretching is performed to generate a three-dimensional block model with a real planar position and height, thus forming a three-dimensional white model dataset of buildings in the target area.

[0014] As a preferred embodiment of the present invention, dual-line array stereo image data and laser altimetry data of the target area are acquired, and radiometric correction, geometric correction, and joint adjustment processing are performed on the acquired raw data to form a stereo observation dataset with a unified geographic reference, including:

[0015] The constellation of stereo mapping satellites is scheduled to conduct quasi-synchronous observations of the target area, acquire a pair of stereo images consisting of a dual-line array of forward-looking, backward-looking, and front-looking cameras, as well as discrete laser point cloud data acquired by the onboard laser altimeter;

[0016] Based on the satellite's precise orbit and attitude data, the dual-line array stereo image pairs are initially geo-registered with the laser point cloud to ensure that they cover the same geographical area.

[0017] Radiometric correction is performed on the dual-line array stereo image pair, and the dual-line array stereo image and laser point cloud are integrated for geometric correction.

[0018] Using ground control points and the laser point cloud as high-precision control benchmarks, the interior and exterior orientation elements of the image are jointly optimized using the regional network adjustment method to obtain a stereo observation dataset.

[0019] As a preferred embodiment of the present invention, a regional network adjustment method is used to jointly optimize the interior and exterior orientation elements of the image, including:

[0020] A joint adjustment model based on bundle adjustment is constructed, in which each image point in the dual-linear array stereo image and each laser foot point in the laser point cloud are regarded as observations originating from the spatial object point.

[0021] In the joint adjustment model, the three-dimensional coordinates of the laser point cloud are used as strong constraints, and sparsely distributed ground control points are used as absolute control references to establish a unified error equation, the expression of which is:

[0022]

[0023] in, Represents the residual vector of observations. This represents the design matrix corresponding to the image point observations, where the unknowns are the refined values ​​of the camera parameters and the error corrections for the satellite attitude and orbit system. This represents the correction vector for the camera and orbit attitude parameters. The design matrix represents the coordinates of the corresponding object points. This represents the correction vector indicating the coordinates of the connection points and the laser point cloud data. This represents a vector of constant terms representing the difference between the observed values ​​and the corresponding calculated values.

[0024] Based on the aforementioned error equation and weight matrix, an iterative solution is performed using the weighted least squares criterion, the expression of which is:

[0025]

[0026] in, This represents the weight matrix determined based on the prior accuracy of the observed values. The setting principle is: assign higher weights to the three-dimensional coordinate observation values ​​of laser point clouds and assign the highest fixed weights to the coordinate observation values ​​of ground control points;

[0027] The camera parameter refinement values, satellite attitude and orbit system error corrections, and optimized three-dimensional coordinates of all connection points of the stereo image are iteratively calculated to geometrically align the multi-source observation data with high precision.

[0028] In a preferred embodiment of the present invention, the fused dual-line array stereo image data and the laser altimetry data are used to generate a digital surface model covering the target area through stereo matching and laser point cloud interpolation. The laser altimetry data is used to correct and enhance the elevation accuracy of the initial DSM generated from the stereo image, including:

[0029] For the dual-linear array stereo image pair, a semi-global stereo matching algorithm is used to calculate the disparity pixel by pixel to generate an initial digital surface model. ;

[0030] The laser point cloud in the stereo observation dataset is filtered, and the filtered laser point cloud is gridded to generate a laser elevation control surface in the form of a regular grid. ;

[0031] The initial digital surface model With the laser elevation control surface Accurate spatial registration is performed, and the elevation difference between two DSMs at the same plane location is calculated to form an elevation residual field.

[0032] Based on the elevation residual field, a trend surface fitting method is used to establish the initial... Elevation system error model within the target area ;

[0033] Using the elevation system error model For the initial Perform global correction on the corrected digital surface model With laser elevation control surface The data is then fused together, and interpolation and smoothing are performed on the entire area to generate a digital surface model.

[0034] As a preferred embodiment of the present invention, the corrected digital surface model is compared with the laser elevation control surface. The fusion is performed, and the fusion strategy is as follows:

[0035] In the area covered by the laser point cloud, the laser elevation control surface is used. The elevation value is the primary factor in the correction of the digital surface model. The elevation value of the corresponding location in the middle The value is adaptively approximated to preserve the high precision of the laser and smooth the local noise of the optical DSM;

[0036] In areas not covered by the laser point cloud, the corrected digital surface model is used directly. Fill in the elevation values.

[0037] As a preferred embodiment of the present invention, the digital surface model and related multispectral imagery are used as input, and a deep learning semantic segmentation model is used to automatically identify and refine the roof outline of each building within the target area, and output as vector polygon data. The deep learning semantic segmentation model is optimized for the spectral, texture, and elevation features of the building rooftops in the satellite imagery, including:

[0038] Create a multi-channel input feature image for model training and inference. The multi-channel input feature image includes red, green, blue and near-infrared bands from multispectral images, providing spectral and vegetation index information, as well as normalized elevation and slope channels derived from the digital surface model.

[0039] The multi-channel input feature image is input into the deep learning semantic segmentation model to learn and construct the mapping relationship from elevation features to buildings, save the optimal model weights, and obtain the initial binary segmentation mask for buildings.

[0040] The binary segmentation mask is post-processed, and the building boundaries are optimized using an edge thinning algorithm to obtain the optimized binary segmentation mask;

[0041] The optimized binary mask is vectorized into polygons to generate an independent vector outline for each building, which is then assigned a unique identifier. The output is structured vector data.

[0042] As a preferred embodiment of the present invention, the vector contour of each building is extracted, spatial analysis is performed on the digital surface model, and statistical characteristic values ​​of the elevation values ​​of the digital surface model within the building contour range are calculated, including:

[0043] Each extracted building vector profile is spatially analyzed on the digital surface model, and the extraction of top and bottom elevations is divided into two parallel sub-processes.

[0044] The calculated top elevation of each building With bottom elevation It is stored as an attribute field in the corresponding building vector outline data.

[0045] In a preferred embodiment of the present invention, the elevation value characterizing the main body of the roof is determined as the top elevation of the building, and the base elevation where the building contacts the ground is determined by analyzing the elevation of the area adjacent to the building's outline edge, which is then used as the bottom elevation of the building. This includes:

[0046] The building top elevation is automatically extracted, specifically as follows:

[0047] For the current building vector outline, extract the elevation values ​​of all pixels inside the digital surface model to form the roof elevation sample set of the building;

[0048] Calculate the statistical distribution of the roof elevation sample set, and based on the statistical distribution, identify and remove obvious abnormal low-value points in the sample. The abnormal low-value points correspond to depressions caused by skylights, vents or vegetation on the roof.

[0049] For the filtered set of valid roof elevation samples, calculate its statistical characteristic value, and determine this characteristic value as the roof elevation of the building. ;

[0050] Automatic extraction of the building's base elevation, specifically:

[0051] Create an annular buffer zone with a width of d along the boundary of the current building vector outline. The width d is adaptively set according to the building size to ensure that the buffer zone falls into the ground area around the building.

[0052] Within the annular buffer, the elevation values ​​of all pixels are extracted from the digital surface model to form a candidate ground elevation sample set.

[0053] Calculate the statistical characteristics of the candidate ground elevation sample set, and determine all points in the candidate sample set that are lower than the statistical characteristic value plus a threshold Th as reliable ground points;

[0054] Calculate the average elevation of all reliable ground points and determine it as the base elevation where the building contacts the ground. .

[0055] As a preferred embodiment of the present invention, the vector contour of each building is used as the base surface, and the difference between its corresponding top elevation and bottom elevation is used as the building height. Vertical stretching is then performed to generate a three-dimensional block model with a true planar position and height, forming a three-dimensional white model dataset of buildings in the target area, including:

[0056] Perform geometric checks and repairs on each building vector profile with attached top elevation H_top and bottom elevation H_base attributes to ensure that it is a closed, non-self-intersecting planar polygon.

[0057] The vector contour polygon is used as the base surface, and its geometric shape is defined by its planar coordinates. The bottom elevation H_base is used as the reference elevation surface where the base surface is located.

[0058] Based on top elevation With bottom elevation The difference is used as the stretching height, and the stretching is performed in the vertical direction to generate a three-dimensional prism with the bottom surface as the base and the top surface parallel to the bottom surface, which serves as the basic white model of the building.

[0059] For each generated 3D prism, the projected coordinates of its base polygon and its base elevation are used as the basis for the calculation. Precisely placed in a three-dimensional coordinate system;

[0060] The unique identifier, outline area, and calculated height value of the building are used as attributes and stored in the three-dimensional block model.

[0061] The three-dimensional block models of all buildings within the target area are spatially integrated, and geometric conflicts or overlaps are eliminated to generate a continuous three-dimensional white model dataset of buildings covering the entire target area.

[0062] An automatic 3D white model generation system for buildings based on multi-source satellites, used to implement the automatic 3D white model generation method for buildings based on multi-source satellites as described in any one of claims 1-9, the system comprising:

[0063] The data acquisition and preprocessing module performs radiometric correction, geometric correction and joint adjustment on multi-source satellite data to form a high-precision stereo observation dataset.

[0064] The digital surface model generation module integrates the dual-line array stereo images and laser altimetry data from the high-precision stereo observation dataset to generate a digital surface model with sub-meter accuracy.

[0065] The building outline automatic extraction module takes the digital surface model and related multispectral images as input, uses an optimized deep learning semantic segmentation model to automatically identify and extract the building roof outline, and outputs vector polygon data.

[0066] The building elevation extraction module performs spatial analysis on the digital surface model for each building vector profile, automatically calculates its top and bottom elevations, and associates the elevation values ​​with the vector profile data.

[0067] The 3D white model automatic construction module automatically generates a 3D block model based on the building vector outline with elevation attributes, and integrates it into a 3D white model dataset of buildings in the target area.

[0068] Compared with the prior art, the present invention has the following advantages:

[0069] This invention addresses the core technical bottleneck of insufficient elevation accuracy in urban areas through joint adjustment and system correction of dual-line array stereo imagery and laser altimetry data. It employs a deep learning model that integrates spectral, texture, and elevation features, along with an anti-interference statistical spatial analysis method, to achieve automated, high-precision, and robust extraction of building outlines and heights. Finally, based on accurate geometric and attribute data, it automatically generates structurally complete and semantically rich city-level 3D white models. This overcomes the shortcomings of traditional aerial surveying, such as high cost and long cycle, as well as the insufficient automation and accuracy of existing satellite technology. It enables the production of high-quality 3D data foundations that meet the application needs of low-altitude economy and digital cities at a lower cost and higher efficiency, achieving rapid, accurate, and large-scale acquisition of urban 3D information over a wide area. Attached Figure Description

[0070] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0071] Figure 1 A flowchart of an automatic method for generating 3D white models of buildings based on multi-source satellites, provided in an embodiment of the present invention;

[0072] Figure 2 A block diagram of an automatic generation system for 3D white models of buildings based on multi-source satellites, provided in an embodiment of the present invention. Detailed Implementation

[0073] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0074] like Figure 1 As shown, this invention provides a method for automatically generating 3D white models of buildings based on multi-source satellites, including the following steps:

[0075] Acquire dual-line array stereo image data and laser altimetry data of the target area, and perform radiometric correction, geometric correction and joint adjustment on the acquired raw data to form a high-precision stereo observation dataset with unified geographic reference;

[0076] The fused dual-line array stereo image data and the laser altimetry data are used to generate a digital surface model covering the target area through stereo matching and laser point cloud interpolation. The laser altimetry data is used to correct and enhance the elevation accuracy of the initial DSM generated from the stereo image.

[0077] Using the digital surface model and related multispectral images as input, a deep learning semantic segmentation model is used to automatically identify and refine the roof outline of each building in the target area, and output as vector polygon data. The deep learning semantic segmentation model is optimized for the spectral, texture and elevation features of the building tops in the satellite images.

[0078] In this embodiment, by deeply fusing multi-source satellite data such as dual-line array stereo images and laser altimetry, the absolute elevation accuracy of the laser data is fully utilized to systematically correct and enhance the optical stereo model, so that the final digital surface model achieves sub-meter accuracy in urban areas, providing a reliable foundation for building height extraction.

[0079] Extract the vector outline of each building, perform spatial analysis on the digital surface model, calculate the statistical characteristic value of the elevation value of the digital surface model within the outline of the building, determine the elevation value representing the main body of the roof as the top elevation of the building, and determine the base elevation of the building in contact with the ground by analyzing the elevation of the area adjacent to the edge of the building outline, which is used as the bottom elevation of the building.

[0080] In this embodiment, a deep learning semantic segmentation model optimized for the spectral, texture, and elevation features of buildings is adopted, which significantly improves the completeness and boundary accuracy of building outline extraction in complex urban environments.

[0081] Using the vector outline of each building as the base, and the difference between its corresponding top elevation and bottom elevation as the building height, vertical stretching is performed to generate a three-dimensional block model with a real planar position and height, thus forming a three-dimensional white model dataset of buildings in the target area.

[0082] In this embodiment, the elevation of the main roof structure is extracted by statistical analysis, and the base elevation is intelligently determined by combining buffer analysis. This effectively avoids interference from skylights, vegetation, shadows, etc., and ensures the accuracy and robustness of the height inversion of individual buildings.

[0083] In this embodiment, the entire process is based on satellite remote sensing data, eliminating the reliance on aerial photogrammetry and airborne lidar. It achieves full-chain automation from data acquisition and processing to 3D model generation, greatly shortening the production cycle and enabling rapid response to the needs of large-scale urban 3D data acquisition and updates. The entire process from raw data to white model dataset is completed automatically by the algorithm, without the need for extensive manual editing and adjustments, significantly reducing labor and time costs and enabling large-scale production.

[0084] Acquire dual-line array stereo imagery and laser altimetry data for the target area. Perform radiometric correction, geometric correction, and joint adjustment on the acquired raw data to form a stereo observation dataset with a unified georeferenced reference, including:

[0085] The constellation of stereo mapping satellites is scheduled to conduct quasi-synchronous observations of the target area, acquire a pair of stereo images consisting of a dual-line array of forward-looking, backward-looking, and front-looking cameras, as well as discrete laser point cloud data acquired by the onboard laser altimeter;

[0086] In this embodiment, by scheduling the stereo mapping satellite constellation for quasi-synchronous observation, the acquisition time of the dual-line array stereo image and the laser altimetry data is highly similar, effectively reducing the matching error caused by changes in the ground surface or differences in lighting conditions.

[0087] Based on the satellite's precise orbit and attitude data, the dual-line array stereo image pairs are initially geo-registered with the laser point cloud to ensure that they cover the same geographical area.

[0088] Radiometric correction is performed on the dual-line array stereo image pair, and the dual-line array stereo image and laser point cloud are integrated for geometric correction.

[0089] In this embodiment, an integrated geometric correction strategy is adopted, which places optical images and laser point clouds in a unified adjustment model for processing. This avoids the error accumulation caused by step-by-step processing. The laser point cloud is used as a strong constraint and the ground control points are used as an absolute reference for joint regional network adjustment. This fully utilizes the absolute elevation accuracy of the laser altimetry data to check and optimize the systematic errors of the optical stereo model, which significantly improves the overall geometric accuracy and reliability of the stereo observation dataset. In particular, it achieves an accuracy level that is difficult to achieve in the elevation dimension by traditional satellite stereo measurement.

[0090] Using ground control points and the laser point cloud as high-precision control benchmarks, the interior and exterior orientation elements of the image are jointly optimized using the regional network adjustment method to obtain a stereo observation dataset.

[0091] In this embodiment, the entire preprocessing process, from data acquisition and registration to joint adjustment, is highly automated, reducing reliance on manual intervention. The processing flow is standardized and repeatable, making it suitable for large-scale operational production. Furthermore, the resulting stereo observation dataset has a unified geographic reference and a high degree of internal consistency.

[0092] The regional network adjustment method is used to jointly optimize the interior and exterior orientation elements of the image, including:

[0093] A joint adjustment model based on bundle adjustment is constructed, in which each image point in the dual-linear array stereo image and each laser foot point in the laser point cloud are regarded as observations originating from the spatial object point.

[0094] This invention designs a joint adjustment model:

[0095] The coordinates (x, y) of each image point in the dual-line array stereo image and the three-dimensional coordinates (X, Y, Z) of each laser foot point in the laser point cloud are used to construct the observation value vector.

[0096] Establish a set of parameters to be optimized, including the following parameters: the number of image interior orientation element corrections, the number of image exterior orientation element corrections, and the object space coordinate corrections for all connection points and laser footpoints;

[0097] Implement a hierarchical constraint mechanism:

[0098] First layer of constraint: Laser point cloud as a strong elevation constraint

[0099] The planar coordinates (X,Y) of the laser footpoint are treated as a terrain reference with appropriate weights; the elevation (Z) of the laser footpoint is assigned a high weight as a reference source for elevation control; and by setting a differentiated variance-covariance matrix, the laser elevation information occupies a dominant weight in the normal equation.

[0100] Second layer of constraint: Ground control points serve as absolute references.

[0101] At known ground control points sparsely distributed in the target area, their three-dimensional coordinates (X,Y,Z) are used as fixed observations or assigned the highest weights to ensure that the entire adjustment system is strictly consistent with the geodetic datum.

[0102] In this embodiment, a unified optical-laser joint adjustment model was constructed, which treats optical image points and laser foot points as equivalent observations for overall calculation, thereby achieving deep geometric fusion of the two types of heterogeneous observation data.

[0103] In the joint adjustment model, the three-dimensional coordinates of the laser point cloud are used as strong constraints, and sparsely distributed ground control points are used as absolute control references to establish a unified error equation, the expression of which is:

[0104]

[0105] in, Represents the residual vector of observations. This represents the design matrix corresponding to the image point observations, where the unknowns are the refined values ​​of the camera parameters and the error corrections for the satellite attitude and orbit system. This represents the correction vector for the camera and orbit attitude parameters. The design matrix represents the coordinates of the corresponding object points. This represents the correction vector indicating the coordinates of the connection points and the laser point cloud data. This represents a vector of constant terms representing the difference between the observed values ​​and the corresponding calculated values.

[0106] In this embodiment, the high-precision three-dimensional coordinates of the laser point cloud are used as strong constraints to effectively correct the system distortion of the optical stereo image caused by errors in orbit, attitude and camera model. This significantly improves the absolute accuracy of the stereo model, especially in the elevation direction, and solves the core bottleneck of insufficient elevation accuracy in urban areas of satellite optical stereo mapping.

[0107] Based on the aforementioned error equation and weight matrix, an iterative solution is performed using the weighted least squares criterion, the expression of which is:

[0108] ;

[0109] in, This represents the weight matrix determined based on the prior accuracy of the observed values. The setting principle is: assign higher weights to the three-dimensional coordinate observation values ​​of laser point clouds and assign the highest fixed weights to the coordinate observation values ​​of ground control points;

[0110] The camera parameter refinement values, satellite attitude and orbit system error corrections, and optimized three-dimensional coordinates of all connection points of the stereo image are iteratively calculated to geometrically align the multi-source observation data with high precision.

[0111] In this embodiment, the prior accuracy and reliability of different observations are reflected by setting a differentiated weight matrix. The laser point cloud is given a higher weight to reflect its elevation benchmark role, and the ground control point is given the highest fixed weight to ensure the absolute position benchmark. This dual control mechanism of strong laser constraint and absolute benchmark of control point ensures that the adjustment system is both robust and accurate.

[0112] In this embodiment, the weighted least squares criterion is used for iterative solution, which can optimally balance the contributions of various observations. The adjustment result is statistically optimal, extracting the effective information from the observation data to the maximum extent and suppressing the influence of random errors.

[0113] The fused dual-line array stereo image data and the laser altimetry data are used to generate a digital surface model covering the target area through stereo matching and laser point cloud interpolation. The laser altimetry data is used to correct and enhance the elevation accuracy of the initial DSM generated from the stereo image, including:

[0114] For the dual-linear array stereo image pair, a semi-global stereo matching algorithm is used to calculate the disparity pixel by pixel to generate an initial digital surface model. ;

[0115] The laser point cloud in the stereo observation dataset is filtered, and the filtered laser point cloud is gridded to generate a laser elevation control surface in the form of a regular grid. ;

[0116] In this embodiment, by constructing a laser elevation control surface as an absolute benchmark, the regional elevation deviations caused by factors such as optical stereo imaging system errors and atmospheric refraction are systematically revealed.

[0117] The initial digital surface model With the laser elevation control surface Accurate spatial registration is performed, and the elevation difference between two DSMs at the same plane location is calculated to form an elevation residual field.

[0118] Based on the elevation residual field, a trend surface fitting method is used to establish the initial... Elevation system error model within the target area ;

[0119] This invention employs a trend surface fitting method to perform spatial autocorrelation analysis on the generated elevation residual field, determine the spatial distribution law of the systematic error, and establish a mathematical surface model based on a polynomial function, the expression of which is:

[0120]

[0121] in, For planar coordinates, The residual value, , , , , For undetermined coefficients, This is the random error term;

[0122] A multi-level fitting strategy is adopted: the first-level fitting uses a low-order polynomial to fit the regional systematic deviation trend; the second-level fitting improves the ability to describe complex systematic error patterns in local areas with large residuals by increasing the order of the polynomial or using a piecewise fitting method; based on the least squares principle, the optimal fitting coefficients are obtained by solving the normal equations.

[0123] In this embodiment, the trend surface fitting method is used to model the elevation residual field, which can effectively distinguish and eliminate systematic deviations while preserving the true details of the ground surface and avoiding the local distortion that may be introduced by simple difference replacement.

[0124] Using the elevation system error model For the initial Perform global correction on the corrected digital surface model With laser elevation control surface The data is then fused together, and interpolation and smoothing are performed on the entire area to generate a digital surface model.

[0125] In this embodiment, a fusion strategy is adopted, which not only ensures the highest absolute accuracy in the laser point coverage area, but also utilizes the continuity of optical images to ensure the integrity and seamlessness of the global model, thus solving the problem of coverage holes caused by sparse or uneven spatial distribution of laser point clouds.

[0126] The corrected digital surface model is compared with the laser elevation control surface. The fusion is performed, and the fusion strategy is as follows:

[0127] In the area covered by the laser point cloud, the laser elevation control surface is used. The elevation value is the primary factor in the correction of the digital surface model. The elevation value of the corresponding location in the middle The value is adaptively approximated to preserve the high precision of the laser and smooth the local noise of the optical DSM;

[0128] In areas not covered by the laser point cloud, the corrected digital surface model is used directly. Fill in the elevation values.

[0129] Using the digital surface model and related multispectral imagery as input, a deep learning semantic segmentation model is employed to automatically identify and refine the roof outline of each building within the target area, outputting vector polygon data. The deep learning semantic segmentation model is optimized for the spectral, texture, and elevation features of building rooftops in the satellite imagery, including:

[0130] Create a multi-channel input feature image for model training and inference. The multi-channel input feature image includes red, green, blue and near-infrared bands from multispectral images, providing spectral and vegetation index information, as well as normalized elevation and slope channels derived from the digital surface model.

[0131] In this embodiment, normalized elevation and slope channels derived from DSM are introduced, enabling the model to directly learn the three-dimensional morphological features of buildings in three-dimensional space, fundamentally different from existing methods that rely solely on two-dimensional spectral textures.

[0132] The multi-channel input feature image is input into the deep learning semantic segmentation model to learn and construct the mapping relationship from elevation features to buildings, save the optimal model weights, and obtain the initial binary segmentation mask for buildings.

[0133] The binary segmentation mask is post-processed, and the building boundaries are optimized using an edge thinning algorithm to obtain the optimized binary segmentation mask;

[0134] In this embodiment, by combining the post-processing algorithm of edge refinement, the initial segmentation result output by the depth model is geometrically optimized, which effectively eliminates jagged boundaries, makes the extracted building outlines smoother and more regular, and better conforms to the geometric characteristics of real buildings, thus greatly improving the geometric quality of the output data.

[0135] In this embodiment, the data connection processing steps of the deep learning semantic segmentation model include two parts: a training phase and an inference phase, specifically as follows:

[0136] In the data preprocessing process, the digital surface model of the target area and related multispectral images are used as feature images, and overlapping blocks are cropped according to a size of 512×512 pixels to ensure full coverage of the area.

[0137] The data from each channel is standardized to map the numerical range to the distribution interval during training, using a mean of 0 and a variance of 1.

[0138] The R, G, B, and NIR multispectral channels are stacked with the DSM-derived channels of nDSM and Slope in a preset order to form an 8-channel input tensor.

[0139] The input image is flipped horizontally and vertically to generate multiple enhanced versions for ensemble inference, thereby obtaining enhanced data.

[0140] The deep learning semantic segmentation model used in this invention is based on an encoder-decoder architecture, and its data connection processing flow is as follows:

[0141] A deep convolutional neural network is used as the backbone network. Multi-scale features are extracted by downsampling layer by layer. The input tensor of 512×512×8 is convolved and pooled to obtain multi-level feature maps of 256×256×64, 128×128×128, 64×64×256, and 32×32×512.

[0142] A feature pyramid structure is introduced, and the high-level semantic features of 32×32×512 are upsampled and concatenated with the mid-level features of 64×64×256. Then, they are upsampled and concatenated with the low-level features of 128×128×128 to obtain semantic features that are connected across levels with the low-level detailed features.

[0143] Spatial resolution is gradually restored by transposing convolution, generating a feature map with the same size as the input. Then, a 1×1 convolution and a Softmax activation function are used to map the final 512×512×64 feature map into a pixel-level 512×512×2 classification probability map, where 2 represents the two classes: building and non-building.

[0144] In this embodiment, the post-processing optimization process is as follows:

[0145] For the output classification probability map, the threshold is set to 0.5. Pixels with probability values ​​greater than the threshold are marked as building 1, and the rest are background 0, thus generating an initial binary segmentation mask.

[0146] Conditional random field (CRF) optimization was used to refine the boundary of the binary mask, guided by the original multispectral image.

[0147] Perform connected component analysis on the binary mask, calculate the area of ​​each connected region, and remove noise regions with an area smaller than a preset threshold of 20 square meters.

[0148] Morphological filling of the holes inside the building ensures the integrity and continuity of the building area, and the building boundaries are smoothed.

[0149] The optimized binary mask is vectorized into polygons to generate an independent vector outline for each building, which is then assigned a unique identifier. The output is structured vector data.

[0150] In this embodiment, the vectorization process is as follows:

[0151] For the optimized binary mask, extract the contour point sequence of each connected region, simplify the contour point sequence, remove redundant points, and at the same time maintain the main features of the geometric shape.

[0152] Based on the prior knowledge that most buildings are right-angled rectangles, the outline is regularized by right angles to make the edges more consistent with the actual building shape.

[0153] Assign a unique ID to each extracted building polygon, and check and fix topological errors such as polygon self-intersection and overlap.

[0154] In this embodiment, structured vector data with unique identifiers treats each building as an independent polygonal object, providing a direct and standardized input for subsequent attribute assignment and 3D modeling.

[0155] In this embodiment, the final output of the deep learning semantic segmentation model is a structured, semantic building vector dataset, which mainly includes:

[0156] Building binary segmentation mask: A raster file with the same resolution as the input image, where each pixel is labeled as either building 1 or non-building 0;

[0157] Optimized binary segmentation mask: refined segmentation mask after edge thinning, hole filling, and small patch removal;

[0158] Building vector outline data: Each building has an independent vector polygon containing the coordinate sequence of closed loops, and a unique ID associated with each building polygon;

[0159] Spatial Index: Create a spatial index to support fast spatial query and analysis.

[0160] Extract the vector contour of each building, perform spatial analysis on the digital surface model, and calculate the statistical characteristic values ​​of the elevation values ​​of the digital surface model within the building contour range, including:

[0161] Each extracted building vector profile is spatially analyzed on the digital surface model, and the extraction of top and bottom elevations is divided into two parallel sub-processes.

[0162] The calculated top elevation of each building With bottom elevation It is stored as an attribute field in the corresponding building vector outline data.

[0163] In this embodiment, statistical methods are used to extract the main elevation of the roof, rather than simply using the maximum or average value. This effectively filters out local low-value interference caused by roof skylights, equipment, vegetation, etc., and accurately reflects the general height of the building roof, making the results more representative.

[0164] In this embodiment, the base elevation is obtained by creating a ring-shaped buffer and filtering reliable ground points, which avoids the defects of directly using building edge points that are easily affected by walls, shadows or attachments, and ensures the accuracy of the elevation of the actual contact surface between the building and the ground.

[0165] The elevation value representing the main body of the roof is determined as the top elevation of the building. The base elevation, where the building contacts the ground, is determined by analyzing the elevation of the area adjacent to the building's outline edge, and is used as the bottom elevation of the building. This includes:

[0166] The building top elevation is automatically extracted, specifically as follows:

[0167] For the current building vector outline, extract the elevation values ​​of all pixels inside the digital surface model to form the roof elevation sample set of the building;

[0168] Calculate the statistical distribution of the roof elevation sample set, and based on the statistical distribution, identify and remove obvious abnormal low-value points in the sample. The abnormal low-value points correspond to depressions caused by skylights, vents or vegetation on the roof.

[0169] For the filtered set of valid roof elevation samples, calculate its statistical characteristic value, and determine this characteristic value as the roof elevation of the building. ;

[0170] Automatic extraction of the building's base elevation, specifically:

[0171] Create an annular buffer zone with a width of d along the boundary of the current building vector outline. The width d is adaptively set according to the building size to ensure that the buffer zone falls into the ground area around the building.

[0172] Within the annular buffer, the elevation values ​​of all pixels are extracted from the digital surface model to form a candidate ground elevation sample set.

[0173] Calculate the statistical characteristics of the candidate ground elevation sample set, and determine all points in the candidate sample set that are lower than the statistical characteristic value plus a threshold Th as reliable ground points;

[0174] Calculate the average elevation of all reliable ground points and determine it as the base elevation where the building contacts the ground. .

[0175] In this embodiment, dynamic judgment is made based on the statistical distribution characteristics of the data itself, such as outlier identification and ground point screening, so that it can automatically adapt to diverse buildings with different densities, different roof structures and different surrounding environments, ensuring the universal applicability of the method to buildings of different scales, from detached houses to large building complexes, and reliable results can be obtained.

[0176] Using the vector outline of each building as its base, and the difference between its corresponding top and bottom elevations as its building height, a vertical stretching process is performed to generate a three-dimensional block model with a true planar position and height, forming a three-dimensional white model dataset of buildings in the target area, including:

[0177] Perform geometric checks and repairs on each building vector profile with attached top elevation H_top and bottom elevation H_base attributes to ensure that it is a closed, non-self-intersecting planar polygon.

[0178] The vector contour polygon is used as the base surface, and its geometric shape is defined by its planar coordinates. The bottom elevation H_base is used as the reference elevation surface where the base surface is located.

[0179] Based on top elevation With bottom elevation The difference is used as the stretching height, and the stretching is performed in the vertical direction to generate a three-dimensional prism with the bottom surface as the base and the top surface parallel to the bottom surface, which serves as the basic white model of the building.

[0180] For each generated 3D prism, the projected coordinates of its base polygon and its base elevation are used as the basis for the calculation. Precisely placed in a three-dimensional coordinate system;

[0181] The unique identifier, outline area, and calculated height value of the building are used as attributes and stored in the three-dimensional block model.

[0182] In this embodiment, precise spatial positioning based on the bottom projection coordinates and bottom elevation ensures the accuracy of the absolute and relative positions of each building model in the 3D scene, laying a reliable foundation for spatial analysis and application.

[0183] The three-dimensional block models of all buildings within the target area are spatially integrated, and geometric conflicts or overlaps are eliminated to generate a continuous three-dimensional white model dataset of buildings covering the entire target area.

[0184] In this embodiment, the process of generating a three-dimensional block model from a two-dimensional vector outline is fully automated, requiring no manual modeling or interactive editing. It can generate large-scale urban building white models with one click, and the production efficiency is increased by orders of magnitude compared with traditional methods.

[0185] In this embodiment, the model generation is based entirely on the high-precision bottom contour and accurate building height extracted in the aforementioned steps, so that the three-dimensional white mold has a real planar position and vertical height, and the spatial accuracy is fully guaranteed by the source data.

[0186] An automatic 3D white model generation system for buildings based on multi-source satellites, used to implement the automatic 3D white model generation method for buildings based on multi-source satellites as described in any one of claims 1-9, the system comprising:

[0187] The data acquisition and preprocessing module performs radiometric correction, geometric correction and joint adjustment on multi-source satellite data to form a high-precision stereo observation dataset.

[0188] The digital surface model generation module integrates the dual-line array stereo images and laser altimetry data from the high-precision stereo observation dataset to generate a digital surface model with sub-meter accuracy.

[0189] The building outline automatic extraction module takes the digital surface model and related multispectral images as input, uses an optimized deep learning semantic segmentation model to automatically identify and extract the building roof outline, and outputs vector polygon data.

[0190] The building elevation extraction module performs spatial analysis on the digital surface model for each building vector profile, automatically calculates its top and bottom elevations, and associates the elevation values ​​with the vector profile data.

[0191] The 3D white model automatic construction module automatically generates a 3D block model based on the building vector outline with elevation attributes, and integrates it into a 3D white model dataset of buildings in the target area.

[0192] In this embodiment, modules such as multi-source data processing, deep learning extraction, and spatial analysis modeling are highly integrated, forming a complete technology chain from raw satellite data input to the final output of 3D white model dataset. The data interfaces between the modules within the system are standardized and the processing flow is seamlessly connected, realizing true end-to-end fully automated processing. This completely changes the traditional 3D modeling operation mode that relies on switching between multiple software and frequent manual intervention, resulting in a revolutionary improvement in production efficiency.

[0193] This invention addresses the core technical bottleneck of insufficient elevation accuracy in urban areas through joint adjustment and system correction of dual-line array stereo imagery and laser altimetry data. It employs a deep learning model that integrates spectral, texture, and elevation features, along with an anti-interference statistical spatial analysis method, to achieve automated, high-precision, and robust extraction of building outlines and heights. Finally, based on accurate geometric and attribute data, it automatically generates structurally complete and semantically rich city-level 3D white models. This overcomes the shortcomings of traditional aerial surveying, such as high cost and long cycle, as well as the insufficient automation and accuracy of existing satellite technology. It enables the production of high-quality 3D data foundations that meet the application needs of low-altitude economy and digital cities at a lower cost and higher efficiency, achieving rapid, accurate, and large-scale acquisition of urban 3D information over a wide area.

[0194] The above embodiments are merely exemplary embodiments of this application and are not intended to limit this application. The scope of protection of this application is defined by the claims. Those skilled in the art can make various modifications or equivalent substitutions to this application within its substance and scope of protection, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of this application.

Claims

1. A method for automatically generating 3D white models of buildings based on multi-source satellites, characterized in that, Includes the following steps: Acquire dual-line array stereo image data and laser altimetry data of the target area, and perform radiometric correction, geometric correction and joint adjustment on the acquired raw data to form a high-precision stereo observation dataset with unified geographic reference; The fused dual-line array stereo image data and the laser altimetry data are used to generate a digital surface model covering the target area through stereo matching and laser point cloud interpolation. The laser altimetry data is used to correct and enhance the elevation accuracy of the initial DSM generated from the stereo image. Using the digital surface model and related multispectral images as input, a deep learning semantic segmentation model is used to automatically identify and refine the roof outline of each building in the target area, and output as vector polygon data. The deep learning semantic segmentation model is optimized for the spectral, texture and elevation features of the building tops in the satellite images. Extract the vector outline of each building, perform spatial analysis on the digital surface model, calculate the statistical characteristic value of the elevation value of the digital surface model within the outline of the building, determine the elevation value representing the main body of the roof as the top elevation of the building, and determine the base elevation of the building in contact with the ground by analyzing the elevation of the area adjacent to the edge of the building outline, which is used as the bottom elevation of the building. Using the vector outline of each building as the base, and the difference between its corresponding top elevation and bottom elevation as the building height, vertical stretching is performed to generate a three-dimensional block model with a real planar position and height, thus forming a three-dimensional white model dataset of buildings in the target area.

2. The method for automatically generating a 3D white model of a building based on multi-source satellites according to claim 1, characterized in that, Acquire dual-line array stereo imagery and laser altimetry data for the target area. Perform radiometric correction, geometric correction, and joint adjustment on the acquired raw data to form a stereo observation dataset with a unified georeferenced reference, including: The constellation of stereo mapping satellites is scheduled to conduct quasi-synchronous observations of the target area, acquire a pair of stereo images consisting of a dual-line array of forward-looking, backward-looking, and front-looking cameras, as well as discrete laser point cloud data acquired by the onboard laser altimeter; Based on the satellite's precise orbit and attitude data, the dual-line array stereo image pairs are initially geo-registered with the laser point cloud to ensure that they cover the same geographical area. Radiometric correction is performed on the dual-line array stereo image pair, and the dual-line array stereo image and laser point cloud are integrated for geometric correction. Using ground control points and the laser point cloud as high-precision control benchmarks, the interior and exterior orientation elements of the image are jointly optimized using the regional network adjustment method to obtain a stereo observation dataset.

3. The method for automatically generating a 3D white model of a building based on multi-source satellites according to claim 2, characterized in that, The regional network adjustment method is used to jointly optimize the interior and exterior orientation elements of the image, including: A joint adjustment model based on bundle adjustment is constructed, in which each image point in the dual-linear array stereo image and each laser foot point in the laser point cloud are regarded as observations originating from the spatial object point. In the joint adjustment model, the three-dimensional coordinates of the laser point cloud are used as strong constraints, and sparsely distributed ground control points are used as absolute control references to establish a unified error equation, the expression of which is: ； in, Represents the residual vector of observations. This represents the design matrix corresponding to the image point observations, where the unknowns are the refined values ​​of the camera parameters and the error corrections for the satellite attitude and orbit system. This represents the correction vector for the camera and orbit attitude parameters. The design matrix represents the coordinates of the corresponding object points. This represents the correction vector indicating the coordinates of the connection points and the laser point cloud data. This represents a vector of constant terms representing the difference between the observed values ​​and the corresponding calculated values. Based on the aforementioned error equation and weight matrix, an iterative solution is performed using the weighted least squares criterion, the expression of which is: ； in, This represents the weight matrix determined based on the prior accuracy of the observed values. The setting principle is: assign higher weights to the three-dimensional coordinate observation values ​​of laser point clouds and assign the highest fixed weights to the coordinate observation values ​​of ground control points; The camera parameter refinement values, satellite attitude and orbit system error corrections, and optimized three-dimensional coordinates of all connection points of the stereo image are iteratively calculated to geometrically align the multi-source observation data with high precision.

4. The method for automatically generating a 3D white model of a building based on multi-source satellites according to claim 3, characterized in that, The fused dual-line array stereo image data and the laser altimetry data are used to generate a digital surface model covering the target area through stereo matching and laser point cloud interpolation. The laser altimetry data is used to correct and enhance the elevation accuracy of the initial DSM generated from the stereo image, including: For the dual-linear array stereo image pair, a semi-global stereo matching algorithm is used to calculate the disparity pixel by pixel to generate an initial digital surface model. ; The laser point cloud in the stereo observation dataset is filtered, and the filtered laser point cloud is gridded to generate a laser elevation control surface in the form of a regular grid. ; The initial digital surface model With the laser elevation control surface Accurate spatial registration is performed, and the elevation difference between two DSMs at the same plane location is calculated to form an elevation residual field. Based on the elevation residual field, a trend surface fitting method is used to establish the initial... Elevation system error model within the target area ; Using the elevation system error model For the initial Perform global correction on the corrected digital surface model With laser elevation control surface The data is then fused together, and interpolation and smoothing are performed on the entire area to generate a digital surface model.

5. The method for automatically generating a 3D white model of a building based on multi-source satellites according to claim 4, characterized in that, The corrected digital surface model is compared with the laser elevation control surface. The fusion is performed, and the fusion strategy is as follows: In the area covered by the laser point cloud, the laser elevation control surface is used. The elevation value is the primary factor in the correction of the digital surface model. The elevation value of the corresponding location in the middle The value is adaptively approximated to preserve the high precision of the laser and smooth the local noise of the optical DSM; In areas not covered by the laser point cloud, the corrected digital surface model is used directly. Fill in the elevation values.

6. The method for automatically generating a 3D white model of a building based on multi-source satellites according to claim 5, characterized in that, Using the digital surface model and related multispectral imagery as input, a deep learning semantic segmentation model is employed to automatically identify and refine the roof outline of each building within the target area, outputting vector polygon data. The deep learning semantic segmentation model is optimized for the spectral, texture, and elevation features of building rooftops in the satellite imagery, including: Create a multi-channel input feature image for model training and inference. The multi-channel input feature image includes red, green, blue and near-infrared bands from multispectral images, providing spectral and vegetation index information, as well as normalized elevation and slope channels derived from the digital surface model. The multi-channel input feature image is input into the deep learning semantic segmentation model to learn and construct the mapping relationship from elevation features to buildings, save the optimal model weights, and obtain the initial binary segmentation mask for buildings. The binary segmentation mask is post-processed, and the building boundaries are optimized using an edge thinning algorithm to obtain the optimized binary segmentation mask; The optimized binary mask is vectorized into polygons to generate an independent vector outline for each building, which is then assigned a unique identifier. The output is structured vector data.

7. The method for automatically generating a 3D white model of a building based on multi-source satellites according to claim 6, characterized in that, Extract the vector contour of each building, perform spatial analysis on the digital surface model, and calculate the statistical characteristic values ​​of the elevation values ​​of the digital surface model within the building contour range, including: Each extracted building vector profile is spatially analyzed on the digital surface model, and the extraction of top and bottom elevations is divided into two parallel sub-processes. The calculated top elevation of each building With bottom elevation It is stored as an attribute field in the corresponding building vector outline data.

8. The method for automatically generating a 3D white model of a building based on multi-source satellites according to claim 7, characterized in that, The elevation value representing the main body of the roof is determined as the top elevation of the building. The base elevation, where the building contacts the ground, is determined by analyzing the elevation of the area adjacent to the building's outline edge, and is used as the bottom elevation of the building. This includes: The building top elevation is automatically extracted, specifically as follows: For the current building vector outline, extract the elevation values ​​of all pixels inside the digital surface model to form the roof elevation sample set of the building; Calculate the statistical distribution of the roof elevation sample set, and based on the statistical distribution, identify and remove obvious abnormal low-value points in the sample. The abnormal low-value points correspond to depressions caused by skylights, vents or vegetation on the roof. For the filtered set of valid roof elevation samples, calculate its statistical characteristic value, and determine this characteristic value as the roof elevation of the building. ; Automatic extraction of the building's base elevation, specifically: Create an annular buffer zone with a width of d along the boundary of the current building vector outline. The width d is adaptively set according to the building size to ensure that the buffer zone falls into the ground area around the building. Within the annular buffer, the elevation values ​​of all pixels are extracted from the digital surface model to form a candidate ground elevation sample set. Calculate the statistical characteristics of the candidate ground elevation sample set, and determine all points in the candidate sample set that are lower than the statistical characteristic value plus a threshold Th as reliable ground points; Calculate the average elevation of all reliable ground points and determine it as the base elevation where the building contacts the ground. .

9. The method for automatically generating a 3D white model of a building based on multi-source satellites according to claim 8, characterized in that, Using the vector outline of each building as its base, and the difference between its corresponding top and bottom elevations as its building height, a vertical stretching process is performed to generate a three-dimensional block model with a true planar position and height, forming a three-dimensional white model dataset of buildings in the target area, including: Perform geometric checks and repairs on each building vector profile with attached top elevation H_top and bottom elevation H_base attributes to ensure that it is a closed, non-self-intersecting planar polygon. The vector contour polygon is used as the base surface, and its geometric shape is defined by its planar coordinates. The bottom elevation H_base is used as the reference elevation surface where the base surface is located. Based on top elevation With bottom elevation The difference is used as the stretching height, and the stretching is performed in the vertical direction to generate a three-dimensional prism with the bottom surface as the base and the top surface parallel to the bottom surface, which serves as the basic white model of the building. For each generated 3D prism, the projected coordinates of its base polygon and its base elevation are used as the basis for the calculation. Precisely placed in a three-dimensional coordinate system; The unique identifier, outline area, and calculated height value of the building are used as attributes and stored in the three-dimensional block model. The three-dimensional block models of all buildings within the target area are spatially integrated, and geometric conflicts or overlaps are eliminated to generate a continuous three-dimensional white model dataset of buildings covering the entire target area.

10. An automatic generation system for 3D white models of buildings based on multi-source satellites, characterized in that, The system for implementing the automatic generation method of 3D white model of a building based on multi-source satellites as described in any one of claims 1-9 includes: The data acquisition and preprocessing module performs radiometric correction, geometric correction and joint adjustment on multi-source satellite data to form a high-precision stereo observation dataset. The digital surface model generation module integrates the dual-line array stereo images and laser altimetry data from the high-precision stereo observation dataset to generate a digital surface model with sub-meter accuracy. The building outline automatic extraction module takes the digital surface model and related multispectral images as input, uses an optimized deep learning semantic segmentation model to automatically identify and extract the building roof outline, and outputs vector polygon data. The building elevation extraction module performs spatial analysis on the digital surface model for each building vector profile, automatically calculates its top and bottom elevations, and associates the elevation values ​​with the vector profile data. The 3D white model automatic construction module automatically generates a 3D block model based on the building vector outline with elevation attributes, and integrates it into a 3D white model dataset of buildings in the target area.

Images