A method for rapid updating of two-dimensional and three-dimensional general plans of industrial parks

Abstract

The application discloses a kind of industrial park two-dimensional and three-dimensional general plan fast updating method.The application is first to industrial park multi-temporal optical remote sensing image pre-processing, based on classic deep learning change detection model completes global change area preliminary screening, output suspected change point dataset;Subsequently, combined with the complex scene features of industrial park, construct change point credibility evaluation model, eliminate the pseudo change area generated by model detection;For the real suspected change area after screening, high-precision data is obtained by artificial field measurement;Finally, based on the measured data iterative update park two-dimensional general plan and three-dimensional real scene model, complete the accurate iteration of general plan data.This method combines optical remote sensing image deep learning change detection technology and artificial field retest calibration technology, realizes the efficient, high-precision iterative update of two-dimensional and three-dimensional general plan of park, solves the industry problem that pure artificial detection cannot meet the high-precision surveying and mapping standard of industrial park at present.

Description

Technical Field

[0001] This invention relates to the intersection of surveying and mapping geographic information technology and artificial intelligence technology, specifically to a method for rapidly updating two-dimensional and three-dimensional master maps of industrial parks. Background Technology

[0002] As the core carrier of industrial clusters, industrial parks rely on their two-dimensional and three-dimensional master plans as fundamental data for park planning and management, safety supervision, land use verification, and infrastructure operation and maintenance. Two-dimensional master plans contain planar information such as park boundaries, road networks, structure distribution, and functional zoning, while three-dimensional master plans recreate three-dimensional information such as the elevation, shape, and spatial layout of park structures. The real-time performance and accuracy of both directly determine the quality of digital management of the park. With the accelerated industrial iteration within industrial parks, frequent dynamic changes occur, including factory renovations and expansions, temporary building construction, pipeline upgrades, and greening and facility updates. Traditional master plan update methods are no longer adequate to meet the management needs of the current rapid development of industrial parks.

[0003] There are two main methods for updating 2D and 3D master maps of existing industrial parks: The first is a fully manual on-site survey and update, which involves surveyors conducting full-area on-site surveys, point measurements, and data collection to complete the master map update. Its advantages are high data accuracy and reliability, and it can adapt to complex industrial scenarios; however, its drawbacks are extremely prominent. The full-area manual survey involves a large workload, a long cycle, and high labor costs. It also suffers from severe update lag in large-area industrial parks, making it impossible to achieve routine and rapid updates.

[0004] The second method is a pure deep learning-based automatic update of remote sensing images. Relying on remote sensing optical imagery, a deep learning model automatically identifies changed areas within the park and directly completes iterative updates of the overall map. This method significantly improves update efficiency and reduces labor costs, making it suitable for ordinary urban scenarios with simple terrain and minimal structures. However, in complex industrial park scenarios, the pure deep learning detection and update solution has unavoidable technical flaws, which is the core reason why the industry cannot currently achieve high-precision overall map updates solely through AI.

[0005] Firstly, industrial park scenarios are extremely complex, featuring irregularly shaped steel structure workshops, structures of varying heights, dense underground and overhead pipelines, reflective roof materials, equipment stacking and obstruction, vegetation interspersed with buildings, and interference from temporary construction sites. Classic deep learning models are prone to problems such as false change detection, missed detection of small targets, boundary contour offset, and missing elevation information.

[0006] Secondly, deep learning models can only make judgments based on the pixel features of two-dimensional images, and cannot obtain the real geographical elevation, the actual structural parameters of structures, and the information on changes in hidden works. However, the three-dimensional master map of the industrial park has extremely high requirements for elevation accuracy and spatial structure accuracy, and pure image detection cannot meet the standards of three-dimensional modeling.

[0007] Third, changes in lighting and shadows in industrial scenes, temporary placement of equipment, vehicle stacking, and seasonal vegetation changes are easily misjudged by the model as changes in structures, generating a large amount of invalid change data. If the general plan is updated directly based on the model results, it will lead to distorted drawing data and substandard accuracy, failing to meet the surveying and mapping specifications for park planning approval and safety supervision. Summary of the Invention

[0008] The purpose of this invention is to overcome the technical pain points of existing industrial park master map update methods, such as low efficiency, insufficient detection accuracy of pure deep learning, and poor adaptability to complex scenarios. It provides a rapid and high-precision update method for two-dimensional and three-dimensional master maps of industrial parks based on intelligent interpretation of remote sensing images. This method relies on AI to quickly locate the changed areas in the park, greatly reducing the scope of manual surveying. At the same time, it supplements high-precision spatial data through manual on-site resurveying, realizing efficient and accurate iterative updates of two-dimensional and three-dimensional master maps in complex industrial park scenarios.

[0009] To achieve the above-mentioned technical objectives, this invention provides a method for rapidly updating two-dimensional and three-dimensional site plans of an industrial park, specifically including the following steps:

[0010] S1. Multi-temporal optical image preprocessing: Collect historical baseline optical remote sensing images and real-time optical remote sensing images of the industrial park to be detected, and perform preprocessing operations on the original images, including radiometric correction, geometric correction, image registration and noise reduction enhancement, to eliminate detection errors caused by illumination differences, atmospheric interference and geometric distortion, and obtain preprocessed multi-temporal image pairs.

[0011] S2. Image change detection based on deep learning model; The classic UNet deep learning change detection model is used as the basic feature extraction network. The multi-temporal images preprocessed in step S1 are input into the trained UNet model. Through feature extraction, feature fusion, upsampling and decoding, a pixel-level change detection heatmap is output to distinguish changed pixels from unchanged pixels and determine the dataset of suspected change points.

[0012] S3. Screening of suspected change points and elimination of false changes; Construct a confidence scoring formula for change areas, quantify and score all suspected change areas, set a confidence threshold, compare the comprehensive confidence score of each change area with the confidence threshold, eliminate false change areas with confidence scores less than the confidence threshold, and save real suspected change areas with confidence scores greater than the confidence threshold.

[0013] S4. High-precision manual field measurement: Using total station, RTK equipment, and 3D laser scanning equipment, conduct full-area field survey of the real suspected change areas screened in step S3, collect high-precision measured data of the real suspected change areas and different change conditions, verify the change type, investigate hidden changes that cannot be identified by images, and calibrate the model to detect boundary offset, size error, and missed small target problems.

[0014] S5. Iterative updates of 2D and 3D site plan data; Based on high-precision data obtained through manual measurement, iterative updates of the 2D and 3D site plans are completed respectively.

[0015] S6. Verify and archive the accuracy of updated results; conduct full-area accuracy verification of the updated two-dimensional and three-dimensional general maps, randomly select points in the changed areas for re-measurement and verification, statistically analyze the update error, and archive the updated two-dimensional and three-dimensional general map data, change detection reports, and measured data ledgers in a unified manner to form a regularly updated database, providing data support for subsequent park management.

[0016] The preferred technical solution of this invention: the preprocessing in step S1 must ensure global registration error. Pixels are used to ensure the effectiveness of subsequent change detection. The image registration error correction formula is as follows:

[0017] (1)

[0018] In the formula: Image registration error; The reference image pixel coordinates; These are the pixel coordinates of the image to be detected.

[0019] The preferred technical solution of this invention is as follows: In step S2, the UNet model possesses pixel-level semantic segmentation capabilities, adapting to refined change recognition in remote sensing images; the model training employs a binary classification cross-entropy loss function to ensure the accuracy of change detection, and the loss function formula is as follows:

[0020] (2)

[0021] In the formula: This represents the cross-entropy loss value. This represents the total number of pixels in the image. The pixel is the true label, where 1 represents a changed pixel and 0 represents an unchanged pixel; Predict the probability of pixel changes for the model;

[0022] After the model training converges, the weight parameters are fixed and used for change detection across the entire industrial park.

[0023] Simultaneously, a formula for calculating image pixel differences is constructed to quantify the feature differences between two images, aiding in change identification:

[0024] (3)

[0025] In the formula: coordinates Pixel difference value; The grayscale value of the reference image pixel; For real-time image pixel grayscale values; when If the value exceeds a preset threshold, it is identified as a suspected changed pixel.

[0026] The preferred technical solution of this invention is as follows: The formula for scoring the credibility of the changed region in step S3 is as follows:

[0027] (4)

[0028] In the formula: A comprehensive credibility score is given for the area of ​​change. The pixel continuity score represents the connectivity integrity of changing pixels; Region area score, removing scattered noise pixels; To determine the timing stability score, transient changes caused by temporary disturbances are eliminated. The weighting coefficients are preset with fixed values ​​based on the characteristics of the industrial park scenario, and .

[0029] Set a credibility threshold ,when When an area is identified as a genuine suspected area of ​​change, it is included in the measured list; when If the area is identified as a pseudo-change region, it will be directly removed without the need for manual retesting.

[0030] The preferred technical solution of this invention is as follows: the high-precision measured data in step S4 includes the boundary plane coordinates of the structure, ground elevation, building floor height, building roof height, actual shape parameters of the structure, and pipeline routing; different changing working conditions include distinguishing between new construction, demolition, reconstruction and expansion, renovation, and pipeline modification; hidden changes that cannot be identified by images include underground pipeline alterations, internal structural adjustments of the factory, and equipment foundation modifications.

[0031] The preferred technical solution of this invention is as follows: In step S4, a three-dimensional laser scanning device is used to collect point cloud data of the structure, restore the real morphological parameters of irregular structures, slopes, and cantilever components, and control the point cloud registration error to be less than 1 cm; the high-density point cloud data obtained by the three-dimensional laser scanning is used to carry out three-dimensional modeling of the structure, pipe gallery and ancillary facilities. It is necessary to quantitatively control the modeling error and define the single-point spatial modeling error of the three-dimensional model:

[0032] (7)

[0033] In the formula: For single-point 3D modeling errors; These are the original coordinates of the measured point cloud; Given the coordinates of the corresponding points in the 3D model, the single-point modeling error is required. ≤2cm.

[0034] For the park's curved roof, cantilevered steel structure, and inclined supports, surface fitting error is further introduced:

[0035] (8)

[0036] In the formula: The average fitting error of the surface; This represents the number of sampling points on the curved surface. The normal deviation from a single sampling point to the model-fitted surface controls the surface fitting error. ≤1.5cm.

[0037] The preferred technical solution of this invention is as follows: In step S4, on-site surveys are conducted on overhead and underground pipeline networks to calculate concealed parameters such as the azimuth angle, burial depth, and pipe diameter of the pipeline network; during the actual measurement process, the coordinates of key nodes in the pipeline network, support height, and pipe outer diameter are collected one by one to accurately calculate the horizontal azimuth angle of the pipeline network.

[0038] (9)

[0039] In the formula: The azimuth angle of the horizontal direction of the pipeline network. The coordinates of the starting node of the pipeline network are in plane. These are the planar coordinates of the pipeline termination node.

[0040] Accurately reconstruct the pipeline layout direction; for underground concealed pipe networks, obtain the pipe top elevation and corresponding ground elevation through on-site detection, and use the burial depth calculation formula to obtain the actual burial depth, accurately capturing hidden changes in minor pipe network modifications, relocations, and newly added buried pipelines; the burial depth calculation formula is as follows:

[0041] (10)

[0042] In the formula: Measured burial depth of underground pipelines The measured ground elevation at the location corresponding to the pipeline. The measured elevation of the top of the underground pipeline.

[0043] The preferred technical solution of this invention: In step S4, the plane coordinates of the structure boundary are collected collaboratively by RTK and total station, and the point position is calculated based on the polar coordinate formula. The plane point position error control formula is as follows:

[0044] (5),

[0045] In the formula: For planar point position error, for Shaft error, for Shaft error.

[0046] Measured ground, floor, and roof elevation data, elevation error The overall measured error satisfies:

[0047] (6)

[0048] In the formula: To comprehensively measure the positional errors, for Shaft horizontal error, for Shaft horizontal error, This represents elevation error.

[0049] The preferred technical solution of the present invention is as follows: the two-dimensional master plan update in step S5 includes: matching the original coordinate system and layer structure of the park's two-dimensional master plan, correcting the original building boundaries, plot range, and pipeline lines according to the measured planar coordinate data, adding new building layers, deleting demolished facility layers, and updating planar information such as functional zones, road boundaries, and green areas to ensure the accuracy of the planar position and the integrity of the layers in the two-dimensional master plan;

[0050] The 3D master plan update includes: reconstructing the 3D model of the changed area based on measured elevation data, structural morphology parameters, and spatial location data; correcting the elevation deviation, morphological distortion, and positional offset of the original model; replacing the old model data; achieving accurate iteration of the park's 3D real-scene model; and ensuring that the 3D model is completely consistent with the actual working conditions on site.

[0051] The preferred technical solution of the present invention is as follows: if the error in step S6 is within the allowable range of the specification, the update is deemed qualified; if there is a local deviation, the measured data is recalibrated and updated a second time.

[0052] Beneficial effects of the present invention

[0053] (1) Balancing update efficiency and high accuracy, adapting to complex industrial scenarios. This invention utilizes a deep learning model to complete rapid initial screening of changes across the entire domain, replacing traditional manual full-domain screening, significantly reducing the amount of invalid surveying work and improving update efficiency; at the same time, it abandons the defects of pure AI direct updates, and solves the problem of insufficient AI detection accuracy caused by irregular buildings, occlusion, reflection, and hidden projects in industrial parks by manually re-testing and calibrating the model error, ensuring that the two-dimensional and three-dimensional general maps meet the statutory surveying accuracy standards.

[0054] (2) Reduce labor costs and achieve routine updates. Compared with the traditional manual survey method, this invention uses AI to screen and narrow down the measurement range, reducing more than 80% of the ineffective manual work, greatly reducing labor and time costs, and enabling monthly and quarterly routine updates of the 2D and 3D master plans of industrial parks, ensuring the real-time nature of the drawing data. Attached Figure Description

[0055] Figure 1 This is a flowchart of the method of the present invention;

[0056] Figure 2 This is a schematic diagram of the UNet network structure used for remote sensing image change detection in this invention;

[0057] Figure 3 This is an updated two-dimensional and three-dimensional overall map of the changed area of ​​the industrial park in an embodiment of the present invention. Detailed Implementation

[0058] 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.

[0059] The embodiment presents a method for rapid updating of two-dimensional and three-dimensional site maps of an industrial park. The overall technical process is shown in Figure 1. Specifically, it includes six core steps: multi-temporal optical image preprocessing, image change detection based on a deep learning model, screening of suspected change points and removal of false changes, high-precision manual on-site measurement, iterative updating of two-dimensional and three-dimensional site map data, and verification and archiving of the updated results. The specific technical solution is as follows:

[0060] S1. Multi-temporal optical image preprocessing: Acquire historical baseline optical remote sensing images and real-time optical remote sensing images of the industrial park during the period to be inspected. Perform preprocessing operations on the raw images, including radiometric correction, geometric correction, image registration, noise reduction and enhancement, to eliminate detection errors caused by illumination differences, atmospheric interference, and geometric distortion. Preprocessing must ensure full-area registration accuracy. Pixels are used to ensure the effectiveness of subsequent change detection. The image registration error correction formula is as follows:

[0061] (1),

[0062] In the formula: Image registration error; The reference image pixel coordinates; These are the pixel coordinates of the image to be detected.

[0063] S2. Image Change Detection Based on Deep Learning Model; This invention uses the classic UNet deep learning change detection model as the basic feature extraction network. The UNet model has pixel-level semantic segmentation capability, is adapted to fine-grained change recognition of remote sensing images, and is suitable for detecting changes in small-scale structures in parks; The preprocessed multi-temporal images are input into the trained UNet model, and through feature extraction, feature fusion, and upsampling decoding, a pixel-level change detection heatmap is output to distinguish changed pixels from unchanged pixels.

[0064] The model training employs a binary classification cross-entropy loss function to ensure the accuracy of change detection. The loss function formula is as follows:

[0065] (2),

[0066] In the formula: This represents the cross-entropy loss value. This represents the total number of pixels in the image. The pixel is the true label, where 1 represents a changed pixel and 0 represents an unchanged pixel; Predict the probability of pixel changes for the model;

[0067] After the model training converges, the weight parameters are fixed and used for change detection across the entire industrial park.

[0068] Simultaneously, a formula for calculating image pixel differences is constructed to quantify the feature differences between two images, aiding in change identification:

[0069] (3),

[0070] In the formula: coordinates Pixel difference value; The grayscale value of the reference image pixel; For real-time image pixel grayscale values; when If the value exceeds a preset threshold, it is identified as a suspected changed pixel.

[0071] S3. Screening of Suspected Change Points and Removal of False Changes: The change regions directly output by the deep learning model contain a large amount of false change data, mainly originating from light and shadow shifts in industrial parks, temporary storage of vehicles and equipment, seasonal changes in vegetation, and rooftop reflection interference. These need to be filtered and denoised using a credibility assessment model. A credibility scoring formula for change regions is constructed to quantitatively score all suspected change regions.

[0072] (4),

[0073] In the formula: A comprehensive credibility score is given for the area of ​​change. The pixel continuity score represents the connectivity integrity of changing pixels; Region area score, removing scattered noise pixels; To determine the timing stability score, transient changes caused by temporary disturbances are eliminated. The weighting coefficients are preset with fixed values ​​based on the characteristics of the industrial park scenario, and .

[0074] Set a credibility threshold ,when When an area is identified as a genuine suspected area of ​​change, it is included in the measured list; when When a region is identified as a spurious change, it is directly removed without requiring manual retesting. This step significantly reduces the amount of invalid measurements and ensures efficient updates.

[0075] S4. Manual on-site high-precision measurement; This step is the core key to ensuring the accuracy of the overall map update in this invention, and it is also the core feature that distinguishes it from existing pure AI automatic update technology. The suspected change areas after screening in step three are only visual change features at the image level. Due to the limitations of the complex scene of the industrial park, the model detection results have inherent accuracy defects and cannot be directly used for overall map update. Manual on-site re-measurement must be carried out.

[0076] The specific content of manual field measurement includes: using total stations, RTK equipment, and 3D laser scanning equipment to conduct a full-area on-site survey of suspected change areas, collecting high-precision measured data such as the plane coordinates of structure boundaries, ground elevation, building floor height, structure morphological parameters, and pipeline routing; verifying the type of change, distinguishing between different change conditions such as new construction, demolition, expansion, renovation, and pipeline modification; investigating hidden changes that cannot be identified by images, including underground pipeline alterations, internal structural adjustments of factory buildings, and equipment foundation modifications—areas where AI detection is blind; and calibrating the model to detect issues such as boundary offsets, dimensional errors, and missed small targets.

[0077] 1. The plane coordinates of the structure boundary are collected through RTK and total station collaboration. The point position is calculated using polar coordinate formulas. The plane point position error control formula is as follows:

[0078] (5),

[0079] In the formula: For planar point position error, for Shaft error, for Shaft error.

[0080] Measured ground, floor, and roof elevation data, elevation error The overall measured error satisfies:

[0081] (6)

[0082] In the formula: To comprehensively measure the positional errors, for Shaft horizontal error, for Shaft horizontal error, This represents elevation error.

[0083] 2. Use 3D laser scanning equipment to collect point cloud data of structures, reconstruct the true morphological parameters of irregular structures, slopes, and cantilever components, and control the point cloud registration error to be less than 1cm; use the high-density point cloud data obtained by 3D laser scanning to carry out 3D modeling of structures, pipe corridors, and ancillary facilities, and quantitatively control the modeling error; define the single-point spatial modeling error of the 3D model:

[0084] (7),

[0085] In the formula: For single-point 3D modeling errors; These are the original coordinates of the measured point cloud; Given the coordinates of the corresponding points in the 3D model, the single-point modeling error is required. ≤2cm.

[0086] For irregularly shaped components such as the park's arched roof, cantilevered steel structure, and inclined supports, surface fitting error is further introduced:

[0087] (8),

[0088] In the formula: The average fitting error of the surface; This represents the number of sampling points on the curved surface. The normal deviation from a single sampling point to the model-fitted surface controls the surface fitting error. ≤1.5cm.

[0089] 3. Conduct on-site surveys of overhead and underground pipeline networks, calculating concealed parameters such as pipeline azimuth, burial depth, and pipe diameter. During the survey, collect coordinates of key pipeline nodes, support heights, and pipe outer diameters one by one to accurately calculate the horizontal azimuth of the pipeline network.

[0090] (9),

[0091] In the formula: The azimuth angle of the horizontal direction of the pipeline network. The coordinates of the starting node of the pipeline network are in plane. These are the planar coordinates of the pipeline termination node.

[0092] Accurately restore the pipeline layout direction; for hidden underground pipe networks, obtain the pipe top elevation and the corresponding ground elevation through on-site detection, use the buried depth calculation formula to obtain the real buried depth, and accurately capture hidden changes such as minor pipe network renovation, displacement, and newly added pre-buried pipelines.

[0093] (10),

[0094] where: is the measured buried depth of the underground pipeline, is the measured ground elevation at the position corresponding to the pipeline, is the measured elevation of the top of the underground pipeline.

[0095] Through this step, the hidden parameters of the pipe network that cannot be obtained by optical images and intelligent detection are fully supplemented, ensuring that the routing, pipe diameter parameters and buried depth attributes of the two-dimensional pipe network layer fully match the on-site actual working conditions with the spatial posture of the three-dimensional pipe gallery model, and meet the high-precision specifications of the two-dimensional and three-dimensional general surveying and mapping of the industrial park.

[0096] S5. Iterative update of two-dimensional and three-dimensional general map data; based on the high-precision data obtained from manual actual measurement, the iterative update of the two-dimensional general map and the three-dimensional general map is completed respectively.

[0097] 1. Two-dimensional general map update; match the coordinate system and layer structure of the original park two-dimensional general map, correct the original structure boundaries, plot ranges, and pipe network lines according to the measured plane coordinate data, add new structure layers, delete layers of demolished facilities, update plane information such as functional partitions, road boundaries, and greening ranges, so as to ensure the plane position accuracy and layer integrity of the two-dimensional general map.

[0098] 2. Three-dimensional general map update; relying on the measured elevation data, structural form parameters, and spatial position data, reconstruct the three-dimensional model of the changed area, correct the problems of elevation deviation, form distortion and position offset of the original model, replace the original old model data, realize the accurate iteration of the park three-dimensional reality model, and ensure that the three-dimensional model is completely consistent with the on-site actual working conditions

[0099] S6. Accuracy verification and archiving of updated results; perform full-domain accuracy verification on the updated two-dimensional and three-dimensional general maps, randomly select points in the changed area for re-measurement verification, and count the update errors. If the error is within the allowable range of the specification, the update is judged to be qualified; if there is a local deviation, re-calibrate the measured data and perform a second update. Finally, the updated two-dimensional and three-dimensional general map data, change detection reports, and measured data ledgers are archived uniformly to form a normalized update database, which provides data support for subsequent park management

[0100] The present invention is further described below in combination with specific application examples. The application example is a rapid update method for two-dimensional and three-dimensional general maps of a metallurgical plant, which specifically includes the following steps:

[0101] I. Two high-resolution optical remote sensing images of a metallurgical plant area were collected, one from the historical baseline period and the other from the current monitoring period. The spatial resolution of the images was 0.5m, covering the entire area of ​​the plant. First, radiometric correction was performed on the two original images to eliminate grayscale deviations caused by atmospheric scattering and uneven illumination. Then, geometric correction was performed to correct the geometric distortion of the images based on the coordinates of existing control points in the plant area. Subpixel image registration was carried out, and the two images were aligned using feature point matching technology. The registration error was calculated according to formula (1). Finally, the global registration error was controlled within 0.3 pixels, meeting the requirements of subsequent processing techniques.

[0102] II. Selection of Appendices Figure 2 The classic UNet deep learning model shown is used as the feature extraction network. The encoder part of the model extracts deep features of the image through 4 layers of convolutional pooling, and the decoder part restores pixel-level details through 4 layers of upsampling and feature fusion, which is suitable for the identification of small-scale structural changes in the park. A training dataset containing 2000 pairs of multi-temporal remote sensing images was constructed, and the changed pixels (newly built, demolished, and renovated areas) and unchanged pixels were labeled. The binary classification cross-entropy loss function of formula (2) was used to train the model. The batch size was set to 16 and the learning rate was 0.001. After 50 rounds of iterative training, the model converged and the network weights were fixed. The preprocessed images of the two periods were input into the trained model, and the pixel-level change detection heatmap was output. At the same time, the pixel difference value was calculated by formula (3), and the gray level difference threshold was set to 25 to screen out suspected changed pixels and initially delineate the suspected change areas in the park.

[0103] Third, based on the characteristics of the metallurgical park scenario, a credibility assessment model was constructed, and quantitative scoring was carried out according to formula (4). Based on experience, the weight coefficients α=0.4, β=0.3, and γ=0.3 were set to satisfy α+β+γ=1. The credibility threshold S0=0.6 was set. After scoring and screening, pseudo-change areas were eliminated, and the remaining real suspected change areas were included in the actual measurement list, which greatly reduced the workload of subsequent manual measurement.

[0104] IV. Establish a professional surveying team to conduct on-site surveys of suspected real changes. Use RTK equipment to collect the plane coordinates of the boundary of the structure, use a total station to assist in calibrating the accuracy of the points, and control the plane point error according to formula (5); use a level to measure the elevation of the ground, floors and roof to ensure that the comprehensive measurement point error meets the requirements. Use three-dimensional laser scanning equipment to collect point clouds of irregular components such as curved roofs and cantilevered steel structures. The point cloud registration error is controlled within 0.8cm. Use formulas (7) and (8) to control the single-point modeling error and surface fitting error. Conduct on-site detection of overhead pipe corridors and underground pipelines. Calculate the azimuth, burial depth and pipe diameter of the pipeline network according to formulas (9) and (10), identify the displacement of underground pipelines and the renovation of the equipment foundation inside the factory, and comprehensively supplement the data of hidden works.

[0105] V. Based on high-precision measured data, the overall plan was updated. In the two-dimensional plane, the plan was matched to the park's original 2000 national geodetic coordinate system, correcting boundary offsets of structures, adding new factory building layers, deleting demolished facility layers, and simultaneously updating planar information such as roads, pipelines, functional zones, and green areas to ensure the integrity and accuracy of the two-dimensional overall plan layers. In the three-dimensional plane, a real-world model of the changed areas was reconstructed, focusing on correcting the changed areas marked in Appendix 3, replacing old model data, and ensuring that the three-dimensional model is completely consistent with the spatial posture of the on-site structures and utility tunnels, restoring the true three-dimensional scene of the park.

[0106] VI. Fifty points were randomly selected for re-measurement and verification. Statistical analysis showed that the average error in planar position was 0.8cm, the average error in elevation was 1.2cm, and the average error of a single point in the 3D model was 1.5cm, all meeting the requirements of surveying and mapping specifications. The update was deemed qualified. The updated 2D and 3D general maps, change detection reports, and measured data ledgers were uniformly archived into the park's digital management platform to form a regularly updated database.

[0107] The above description is merely one embodiment of the present invention, and while it is detailed and specific, it should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A method for rapidly updating two-dimensional and three-dimensional site maps of an industrial park, characterized in that, Specifically, the following steps are included: S1. Multi-temporal optical image preprocessing: Collect historical baseline optical remote sensing images and real-time optical remote sensing images of the industrial park to be detected, and perform preprocessing operations on the original images, including radiometric correction, geometric correction, image registration and noise reduction enhancement, to eliminate detection errors caused by illumination differences, atmospheric interference and geometric distortion, and obtain preprocessed multi-temporal image pairs. S2. Image change detection based on deep learning model; The classic UNet deep learning change detection model is used as the basic feature extraction network. The multi-temporal images preprocessed in step S1 are input into the trained UNet model. Through feature extraction, feature fusion, upsampling and decoding, a pixel-level change detection heatmap is output to distinguish changed pixels from unchanged pixels and determine the dataset of suspected change points. S3. Screening of suspected change points and elimination of false changes; Construct a confidence scoring formula for change areas, quantify and score all suspected change areas, set a confidence threshold, compare the comprehensive confidence score of each change area with the confidence threshold, eliminate false change areas with confidence scores less than the confidence threshold, and save real suspected change areas with confidence scores greater than the confidence threshold. S4. High-precision on-site manual measurement; Using total station, RTK equipment, and 3D laser scanning equipment, a full-area field survey was conducted on the real suspected change areas screened in step S3. High-precision measured data of the real suspected change areas and different change conditions were collected to verify the change type, investigate hidden changes that could not be identified by images, and calibrate the model to detect boundary offsets, size errors, and missed small targets. S5. Iterative updates of 2D and 3D site plan data; Based on high-precision data obtained through manual measurement, iterative updates of the 2D and 3D site plans are completed respectively. S6. Verify and archive the accuracy of updated results; conduct full-area accuracy verification of the updated two-dimensional and three-dimensional general maps, randomly select points in the changed areas for re-measurement and verification, statistically analyze the update error, and archive the updated two-dimensional and three-dimensional general map data, change detection reports, and measured data ledgers in a unified manner to form a regularly updated database, providing data support for subsequent park management.

2. The method for rapidly updating two-dimensional and three-dimensional site maps of an industrial park according to claim 1, characterized in that: In step S1, preprocessing must ensure global registration error. Pixels are used to ensure the effectiveness of subsequent change detection. The image registration error correction formula is as follows: （1）， In the formula: Image registration error; The reference image pixel coordinates; These are the pixel coordinates of the image to be detected.

3. The method for rapidly updating two-dimensional and three-dimensional site maps of an industrial park according to claim 1, characterized in that: In step S2, the UNet model possesses pixel-level semantic segmentation capabilities, adapting to refined change recognition in remote sensing images. The model training employs a binary classification cross-entropy loss function to ensure the accuracy of change detection. The loss function formula is as follows: （2）， In the formula: This represents the cross-entropy loss value. This represents the total number of pixels in the image. The pixel is the true label, where 1 represents a changed pixel and 0 represents an unchanged pixel; Predict pixel change probabilities for the model; After the model training converges, the weight parameters are fixed and used for change detection across the entire industrial park. Simultaneously, a formula for calculating image pixel differences is constructed to quantify the feature differences between two images, aiding in change identification: （3）， In the formula: coordinates Pixel difference value; The grayscale value of the reference image pixel; For real-time image pixel grayscale values; when If the value exceeds a preset threshold, it is identified as a suspected changed pixel.

4. The method for rapidly updating two-dimensional and three-dimensional site maps of an industrial park according to claim 1, characterized in that: The formula for scoring the credibility of the changed region in step S3 is as follows: （4）， In the formula: A comprehensive credibility score is given for the area of ​​change. The pixel continuity score represents the connectivity integrity of changing pixels; Region area score, removing scattered noise pixels; To determine the timing stability score, transient changes caused by temporary disturbances are eliminated. The weighting coefficients are preset with fixed values ​​based on the characteristics of the industrial park scenario, and ; Set a credibility threshold ,when When an area is identified as a genuine suspected area of ​​change, it is included in the measured list; when If the area is identified as a pseudo-change region, it will be directly removed without the need for manual retesting.

5. The method for rapidly updating two-dimensional and three-dimensional site plans of an industrial park according to claim 1, characterized in that: The high-precision measured data in step S4 includes the boundary plane coordinates of the structure, ground elevation, building floor height, building roof height, actual shape parameters of the structure, and pipeline routing; different changing conditions include distinguishing between new construction, demolition, expansion and renovation, and pipeline modification; hidden changes that cannot be identified by images include underground pipeline alterations, internal structural adjustments of the factory, and equipment foundation modifications.

6. A method for rapidly updating two-dimensional and three-dimensional site plans of an industrial park according to claim 1 or 5, characterized in that: In step S4, point cloud data of the structure is collected using 3D laser scanning equipment to restore the true morphological parameters of irregular structures, slopes, and cantilever components, controlling the point cloud registration error to be less than 1cm. High-density point cloud data obtained from 3D laser scanning is used to perform 3D modeling of the structure, pipe gallery, and ancillary facilities. Modeling errors need to be quantitatively controlled, and the single-point spatial modeling error of the 3D model is defined as follows: （7）， In the formula: For single-point 3D modeling errors; These are the original coordinates of the measured point cloud; Given the coordinates of the corresponding points in the 3D model, the single-point modeling error is required. ≤2cm; For the park's curved roof, cantilevered steel structure, and inclined supports, surface fitting error is further introduced: （8）， In the formula: The average fitting error of the surface; This represents the number of sampling points on the curved surface. The normal deviation from a single sampling point to the model-fitted surface controls the surface fitting error. ≤1.5cm.

7. A method for rapidly updating two-dimensional and three-dimensional site maps of an industrial park according to claim 1 or 5, characterized in that: In step S4, on-site surveys are conducted of overhead and underground pipeline networks to calculate concealed parameters such as pipeline azimuth, burial depth, and pipe diameter. During the survey, coordinates of key pipeline nodes, support height, and pipe outer diameter are collected one by one to accurately calculate the horizontal azimuth of the pipeline network. （9）， In the formula: The azimuth angle of the horizontal direction of the pipeline network. The coordinates of the starting node of the pipeline network are in plane. The coordinates of the pipeline termination node in planar coordinates; Accurately reconstruct the pipeline layout direction; for underground concealed pipe networks, obtain the pipe top elevation and corresponding ground elevation through on-site detection, and use the burial depth calculation formula to obtain the actual burial depth, accurately capturing hidden changes in minor pipe network modifications, relocations, and newly added buried pipelines; the burial depth calculation formula is as follows: （10）， In the formula: Measured burial depth of underground pipelines The measured ground elevation at the location corresponding to the pipeline. The measured elevation of the top of the underground pipeline.

8. A method for rapidly updating two-dimensional and three-dimensional site plans of an industrial park according to claim 5, characterized in that: In step S4, the plane coordinates of the structure boundary are collected collaboratively using RTK and total station. Point position calculations are then performed based on polar coordinate formulas. The plane position error control formula is as follows: （5）， In the formula: For planar point position error, for Shaft error, for Shaft error; Measured ground, floor, and roof elevation data, elevation error The overall measured error satisfies: （6） In the formula: To comprehensively measure the positional errors, for Shaft horizontal error, for Shaft horizontal error, This represents elevation error.

9. The method for rapidly updating two-dimensional and three-dimensional site plans of an industrial park according to claim 1, characterized in that: The S5 step of updating the 2D master plan includes: matching the original coordinate system and layer structure of the park's 2D master plan; correcting the original building boundaries, plot range, and pipeline routes based on the measured plan coordinate data; adding new building layers; deleting demolished facility layers; and updating plan information such as functional zones, road boundaries, and green areas to ensure the accuracy of the 2D master plan's plan position and the integrity of its layers. The 3D master plan update includes: reconstructing the 3D model of the changed area based on measured elevation data, structural morphology parameters, and spatial location data; correcting the elevation deviation, morphological distortion, and positional offset of the original model; replacing the original old model data; achieving accurate iteration of the park's 3D real-scene model; and ensuring that the 3D model is completely consistent with the actual working conditions on site.

10. A method for rapidly updating two-dimensional and three-dimensional site maps of an industrial park according to claim 1, characterized in that: If the error is within the allowable range of the specification in step S6, the update is deemed qualified; if there is a local deviation, the measured data is recalibrated and updated a second time.

Images