A Method for Enhancing the Modeling of LOD2-Level Urban Building Models Driven by Crowdsourced Data

Abstract

The present invention discloses a method for enhancing the modeling of LOD2-level urban building models driven by crowdsourced data. The main steps include: extracting building outline and height information, clustering analysis of building landscapes, collecting crowdsourced data, AI intelligent interpretation and analysis, spatial association mapping of texture information, and three-dimensional enhanced modeling. Aiming at multiple problems commonly existing in existing three-dimensional modeling methods, such as high modeling cost, slow speed, and low restoration degree, the present invention designs a method for enhancing the modeling of LOD2-level urban building models driven by crowdsourced data. Driven by crowdsourced data including remote sensing images, street view pictures, OSM, and POI, by integrating multiple deep learning modules, it automatically extracts building outlines, heights, and texture information. By constructing the spatial mapping relationship between building surfaces and texture information, it effectively enriches the level of detail of buildings, enhances the authenticity and restoration degree of building models, realizes the rapid and automatic construction of LOD2-level urban twin three-dimensional geographical scenes, can significantly reduce the cost of constructing urban-level building models, accelerate the construction rate of building models within the urban area, and take into account the basic authenticity of textures of LOD2-level building models. The restoration accuracy of the building model can reach 85%. This technical invention has good application value in the display of LOD2-level urban building models in multiple fields including spatial planning, natural resources, and smart cities.

Description

Technical Field

[0001] The present invention belongs to the technical field of three-dimensional urban building model modeling, and particularly relates to a method for crowd-sourced data-driven LOD2-level urban three-dimensional enhanced modeling. Background Art

[0002] To meet different application requirements, three-dimensional building models usually adopt multi-scale expression methods. According to the definition of CityGML, all models can be divided into 5 different coherent levels of detail (LOD), where LOD0 refers to a 2.5D digital topographic map, LOD1 refers to a "building block model" without a roof structure, LOD2 refers to a rough model including a roof structure and real building textures, LOD3 refers to a building model with more details, and LOD4 refers to a model including the internal structure of the building.

[0003] Existing real-scene three-dimensional modeling technologies mainly include: manual modeling technology, oblique photography modeling technology, three-dimensional modeling technology based on laser point cloud, and parametric modeling technology based on vector data. Among them, manual modeling and three-dimensional modeling technology based on laser point cloud have the outstanding advantage of fine and accurate model detail effects, but their modeling speed is slow and the labor cost is high, and they are mostly applicable to the refined modeling of building monomers in key areas. Oblique photography modeling technology has significant advantages in modeling speed, texture authenticity, and automated calculation, but still faces multiple difficulties including high aerial flight and calculation costs and large airspace restrictions, and is mostly applicable to the modeling of the main urban areas in some cities with good economic conditions. The parametric three-dimensional modeling based on vector data has low economic cost and time cost and can meet the demand for large-scale urban three-dimensional data simulation and expression, but its three-dimensional display effect of data is insufficient and the texture authenticity is poor, and it cannot intuitively reflect the urban real scene.

[0004] Therefore, how to reduce the cost of urban-level building model construction, accelerate the construction rate of urban LOD2-level building models, and balance the basic authenticity of textures of urban-level models under the premise of limited basic data and modeling funds is a real challenge faced by the construction of multiple fields such as real-scene three-dimensional, smart city, and three-dimensional natural resources. Summary of the Invention

[0005] Objective of the Invention: This design focuses on how to reduce the construction cost of urban-level building models, accelerate the construction rate of urban-level building models, improve the technical efficiency of urban-level model construction, and thus expand the coverage of urban three-dimensional models. A new three-dimensional data construction and service method is developed, integrating and fusing multiple technologies including building contour and height information extraction based on deep learning, building landscape clustering analysis, street view data collection, intelligent interpretation and analysis of building types, design of building texture mapping rules, and three-dimensional enhanced modeling, to promote the realization of rapid, batch, and automated construction of urban-level real-scene three-dimensional models, and provide more real, accurate, and rich urban building landscapes.

[0006] Technical Solution: A crowdsourced data-driven LOD2-level urban three-dimensional enhanced modeling method of the present invention meets the requirements of the second level of multi-level details of the model, that is, the building model has different types of building roofs and real surface textures. The present invention specifically includes the following steps:

[0007] Step 1, extraction of building contour and height information: Construct a deep learning image segmentation model, automatically extract building surfaces and building shadows based on high-resolution remote sensing images, and match building surfaces and their shadows according to multiple principles including adjacency and direction consistency. Use mathematical morphology to characterize the spatial structure characteristics of buildings, and optimize the building contour through a series of morphological opening and closing operations. Based on the relationship between the solar altitude angle and the satellite altitude angle, calculate the corresponding building height according to the length of the building shadow.

[0008] Step 2, building landscape clustering analysis: Integrate the geometric features of building surface contours, building spatial relationships, and building real texture information, and achieve similar building clustering through building block division based on road and spatial proximity, building feature clustering based on self-organizing mapping (SOM) neural networks, and building clustering based on the semantic information of points of interest (POIs), to obtain building clustering surfaces.

[0009] Step 3, street view data collection: Based on the public transportation road network data of crowdsourced data, extract road points at fixed intervals, and obtain street view data within the adjacent range based on the spatial information of the road points.

[0010] Step 4, Intelligent Interpretation and Analysis of Building Categories: Considering multiple factors including different regional characteristics, architectural styles, building uses, and building densities, a building classification system is constructed with building structure, building coatings, and component distribution as specific features. Based on this system, a building sample library and a texture library are established. The data of the building type sample library is input into a deep learning convolutional neural network to train an AI recognition and detection model for building types and evaluate its accuracy. The street view data of the modeling area obtained in Step 2 is input into this model, and the buildings and their types in the picture can be automatically predicted.

[0011] Step 5, Spatial Mapping of Texture Information: Regarding the spatial positions of the street view pictures and the building clustering surfaces, the street view pictures are associated with the building clustering surfaces. Taking the building clustering surfaces as units, their building type labels are counted and sorted by quantity, and the corresponding textures are retrieved according to the highest-level building type, and the texture information is assigned to the corresponding building clustering surfaces to complete the spatial mapping of the building surfaces and the texture information.

[0012] Step 6, 3D Enhanced Modeling: Based on the GIS desktop platform, according to the building height, structure, and texture information, a 3D model of the building is automatically generated to quickly achieve LOD2-level urban 3D enhanced modeling.

[0013] Furthermore, in the above Step 1, a deep learning method is used to extract the building contour and height information from the remote sensing image, and the specific steps are as follows:

[0014] Step 1.1, Extracting Building and Shadow Information by Semantic Segmentation Method: A sample library is constructed by manually vectorizing the buildings and shadows in the high-resolution image, and then a deep learning convolutional neural network is used to train an image semantic segmentation model to automatically extract the building contours and shadow information of the high resolution, and the quadratic cross-entropy is used as the loss function to solve and evaluate the model.

[0015] Step 1.2, Building Surface Optimization: A morphological profile is constructed through a series of morphological opening and closing operations, and the average values of the opening profile (OP) and closing profile (CP) of the pixels within the building object are taken to obtain the average value of the morphological opening profile and the average value of the morphological closing profile of the building object, and the morphological features of the building are extracted.

[0016] Step 1.3, Height Feature Extraction: The building surfaces and shadows are matched following the principles of adjacency, consistent direction, and at least one shadow. A number of parallel lines are generated according to the solar azimuth direction, and the parallel lines are cut by the shadows and the average value of the line segment lengths is calculated as the building shadow length. And the building height is calculated based on the spatial relationship between the solar altitude angle, satellite altitude angle and the building. The specific method is as follows:

[0017] When the sun and the satellite are on the same side of the building, the visible formula (1) for calculating the building height is

[0018]

[0019] When the sun and the satellite are on opposite sides of the building, the visible formula for calculating the building height is formula (2),

[0020] H = L_B×tanω = L_S×tanω Formula (2)

[0021] where L_S is the length of the shadow on the image, L_B is the actual shadow length of the building, σ is the solar altitude angle, and ω is the satellite altitude angle.

[0022] Furthermore, in step 2, by integrating the geometric features of the building surface contour, the building spatial relationship, and the building real texture information, through building block division based on road and spatial proximity, building feature clustering based on SOM neural network, and building clustering based on POI semantic information, similar building clustering is realized to obtain the building clustering surface. The specific steps are as follows:

[0023] Step 2.1, Building block division based on road and spatial proximity: Describe the building proximity relationship by constructing a Delaunay triangulation network, construct a minimum spanning tree and a road network with the minimum distance between buildings as the metric, identify the building clusters with spatial proximity, and construct the building block spatial relationship.

[0024] Step 2.2, Building feature clustering based on SOM neural network: Select multiple factors including building height, direction, shape, and area to describe the geometric features of the building, input into the SOM neural network to extract the characteristics of the map buildings. Through the expression of different factors, generate a geometric feature map of the building spatial feature relationship, and cluster the block buildings.

[0025] Step 2.3, Building aggregation based on POI: Based on the results of step 2.1 and step 2.2, by establishing the matching relationship between POI, building texture and buildings, associate the semantic information of POI and the building texture features with buildings, and realize the aggregation of buildings based on semantic and texture features.

[0026] Furthermore, in step 3, in order to facilitate the construction of the spatial relationship between the street view picture and the building clustering surface, it is necessary to simulate the vehicle movement when collecting street view data. According to the preset forward direction, divide the street view data of each point into the left street view pic_1 and the right street view pic_2.

[0027] Furthermore, in step 4, the specific steps of the intelligent interpretation and analysis of building categories are as follows:

[0028] Step 4.1. Construct a building classification system and texture library: Taking into account multiple factors including regional characteristics, architectural styles, building uses, and building density, a building classification system is constructed based on the specific characteristics of building structure, building coatings, and component distribution. Textures are then produced separately to construct a building texture library.

[0029] Step 4.2: Construct a building classification sample library: Obtain a large amount of street view images from across the country, eliminate invalid images where buildings are too far away, too small, or obscured, and manually label valid street view data based on the building classification system to construct a building classification sample set. Perform data enhancement on the sample dataset to enrich the sample types and improve the robustness of the model.

[0030] Step 4.3, model accuracy evaluation: Building type AI recognition detection model training and accuracy evaluation: Use open source framework to build and train the target detection model. Measure the accuracy of the target detection model based on deep learning, and evaluate the accuracy of the model detection results from multiple aspects including test sets and actual test results. Furthermore, in step 5, when constructing the spatial position of the street view image and the building cluster surface, it is necessary to take the road axis as the center and radiate to both sides with the street view recognition range n as the radius to calculate the accuracy of any point P on the road. c (x c ,y c ) of the radiation boundary point P l (x l ,y l ), P r (x r ,y r ), x c Represents the longitude of the waypoint, y c Represents the latitude of the road point, x l Represents the radiation boundary point P l Longitude, y l Represents the radiation boundary point P l Latitude, x r Represents the radiation boundary point P r Longitude, y r Represents the radiation boundary point P r Latitude, and P l 、P r Do an intersection analysis with the adjacent building cluster surface and put point P c The street view pictures corresponding to P are matched to the building cluster faces where the radiation boundary points are located. l 、P r The coordinate calculation formula is as follows:

[0031]

[0032] Among them, k is the slope of the straight line where the road is located, n is the radiation radius on both sides, and xc represents the longitude of the road point, y c represents the latitude of the road point, x l,r represents P l and P r 's longitude, y l,r represents P l and P r 's latitude.

[0033] Then, taking the building clustering surface as a unit, count its building type labels and sort them by quantity, retrieve the corresponding building texture according to the highest-level building type, assign the texture information to the corresponding building clustering surface, and automatically splice and generate a suitable texture map according to the building size to complete the spatial mapping of the building texture information.

[0034] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0035] (1) The present invention integrates a number of technologies including building contour and height calculation based on deep learning and AI intelligent recognition of building facade textures, realizes the rapid and automated extraction of model information, significantly enhances the level of modeling details and the authenticity of building models, promotes the automated, intelligent, and rapid development of urban 3D construction, and significantly improves the efficiency of urban 3D modeling.

[0036] (2) The present invention is driven by crowdsourced data, including multi-source heterogeneous data such as remote sensing images, street view data, OSM data, and POI information, and intelligently extracts building type, building structure, and building facade texture information, which can effectively enrich the details of building models, enhance the authenticity and restoration degree of urban landscapes, significantly improve the quality and visual effects of urban modeling, and overcome the deficiencies of the lack of building white film texture information and low restoration degree of virtual simulation technology. Description of the Drawings

[0037] Figure 1 is the technical flow chart of the present invention [[ID=3,5]]

[0038] Figure 2 is the overall flow chart of landscape aggregation and zoning

[0039] Figure 3 is the technical flow of intelligent detection of building types based on deep learning

[0040] Figure 4 is an example diagram of the recognition result of building types based on deep learning

[0041] Figure 5 is a schematic diagram of the building texture mapping rule

[0042] Figure 6 is an example diagram of the 3D model of Nanjing generated by the present invention

[0043] Figure 7 Comparison diagram between the example diagram of the 3D model generated by the present invention and the street view data at the same location

[0044] Figure 8 Comparison diagram between the example diagram of the 3D model generated by the present invention and the white film data and fine model data at the same location Specific implementation manner

[0045] The technical solution of the present invention will be described in detail below, but the protection scope of the present invention is not limited to the described embodiments.

[0046] In this embodiment, Nanjing City, Jiangsu Province is selected as the experimental area, and the technical flow chart is as Figure 1 shown, and the specific steps are as follows:

[0047] Step 1. Extraction of building outline and height information: Construct a U-Net neural network model, automatically extract building surfaces and building shadows based on high-resolution remote sensing images, and match building surfaces and their shadows according to multiple principles including adjacency and direction consistency. Use mathematical morphology to characterize the spatial structure characteristics of buildings, and optimize the building outline through a series of morphological opening operations and closing operations. Calculate the corresponding building height through the relationship between the solar altitude angle and the satellite altitude angle and the size of the building shadow. The specific steps are as follows:

[0048] Step 1.1. Extract building and shadow information by semantic segmentation method: Construct a sample library by manually vectorizing buildings and shadows in high-resolution images, and then use a deep learning convolutional neural network to train the model to extract high-resolution building and shadow information. The total area of the processed entire region is 6587.02 km 2 , the total number of predicted buildings is 164,110, and the quadratic cross-entropy is used as the loss function to solve and evaluate the model.

[0049] Step 1.2. Building surface optimization: Construct a morphological profile through a series of morphological opening operations and closing operations, take the average of the opening profile (OP) and closing profile (CP) of the pixels within the building object to obtain the morphological opening profile mean and morphological closing profile mean of the building object, and extract the morphological characteristics of the building.

[0050] Step 1.3. Height feature extraction: Match building surfaces and shadows according to the principles of adjacency, direction consistency, and at least one shadow. Generate several parallel lines according to the direction of the solar azimuth angle, cut the parallel lines through the shadow and calculate the mean value of the line segment length as the building shadow length. And calculate the building height based on the spatial relationship between the solar altitude angle, the satellite altitude angle and the building. The specific method is as follows:

[0051] When the sun and the satellite are on the same side of the building, the building height calculation can be seen in formula (1),

[0052]

[0053] When the sun and the satellite are on the opposite sides of the building, the visible formula for calculating the building height is formula (2),

[0054] H = L_B × tanω = L_S × tanω Formula (2)

[0055] where L_S is the length of the shadow on the image, L_B is the actual shadow length of the building, σ is the solar altitude angle, and ω is the satellite altitude angle.

[0056] Step 2. Building landscape clustering analysis: By integrating the geometric features of the building surface contour, the building spatial relationship, and the building real texture information, through building block division based on road and spatial proximity, building feature clustering based on SOM neural network, and building clustering based on POI semantic information, similar building clustering is realized to obtain the building clustering surface. As Figure 2 shown in the overall flowchart of the landscape aggregation partition, the specific steps are as follows:

[0057] Step 2.1. Building block division based on road and spatial proximity: By constructing a Delaunay triangulation to describe the proximity relationship between buildings, a minimum spanning tree and a road network are constructed with the minimum distance between buildings as the metric, and the building clusters with spatial proximity are identified to construct the spatial relationship of the building blocks.

[0058] Step 2.2. Building feature clustering based on SOM neural network: Select multiple factors including building height, direction, shape, and area to describe the geometric features of the buildings, and input them into the SOM neural network to extract the characteristics of the buildings on the map. Through the expression of different factors, a geometric feature map of the spatial feature relationship of the buildings is generated to classify the buildings in the block.

[0059] Step 2.3. Building aggregation based on POI semantic information: Based on the results of Step 2.1 and Step 2.2, by establishing the matching relationship between POI, building texture, and buildings, the semantic information of POI and the building texture features are associated with the buildings to realize the aggregation of buildings based on semantic and texture features.

[0060] Step 3. Street view data collection: In order to facilitate the construction of the spatial relationship between the street view pictures and the building clustering surface, it is necessary to simulate the vehicle movement during the collection of street view data. A sampling point is set every 20 meters according to the preset forward direction, and the street view data of each point is divided into the left street view pic_1 and the right street view pic_2, and a total of 401,620 street view pictures of Nanjing are obtained.

[0061] Step 4. Intelligent interpretation and analysis of building categories: Considering multiple factors including different regional characteristics, architectural styles, building uses, and building densities, and taking building structures, building coatings, and component distributions as specific features, a building classification system is constructed. Based on this system, a building sample library and a texture library are built. Based on the building type sample library, through a deep learning convolutional neural network, a building type AI recognition and detection model is trained and its accuracy is evaluated. Inputting the street view data of the modeling area obtained in Step 2 into this model can automatically predict the building types in the pictures. As Figure 3 shown is the intelligent detection technical process of building types based on deep learning. The specific steps are as follows:

[0062] Step 4.1. Construction of a building classification system and a texture library: Considering multiple factors including different regional characteristics, architectural styles, building uses, and building densities, and taking building structures, building coatings, and component distributions as specific features, a building classification system is constructed, and textures are made respectively to build a building texture library.

[0063] Step 4.2. Construction of a building classification sample library: A large number of street view pictures are obtained from the whole country, and invalid pictures with buildings being too far, too small, or blocked are removed. Based on the building classification system, the effective street view data is manually marked to construct a building classification sample set. The total number of samples drawn reaches 300,000. Data augmentation is performed on the sample data set to enrich the sample types and improve the robustness of the model.

[0064] Step 4.3. Training and accuracy evaluation of the building type AI recognition and detection model: The Detectron2 framework open-sourced by Facebook is used to construct and train the Faster R-CNN model. In terms of image feature extraction, ResNet50 is used as the feature extraction layer, and the FPN mechanism is added to construct the Faster R-CNN model. Among them, the batch size (Batch_Size) of the model is 8, and a total of 15 different types of building categories need to be distinguished. The base learning rate (BASE_LR) of the model is 0.001, and the number of iterative training times of the model is 400,000 times.

[0065] In this design, four parameter indicators, namely precision, recall, AP (Average Precision), and mAP (mean Average Precision), are used to evaluate the accuracy of the building facade target detection model. The accuracy rate of the model detection reaches 79.3%, the recall rate reaches 80.14%, and the mAP reaches 83.44%. At the same time, combined with the actual detection effect, most of the building facades that conform to the definition of this design in the test images can be effectively detected, and some recognition results are as Figure 4 shown.

[0066] Step 4.4, Intelligent Detection of Building Types: Input the picture into the trained object detection model to predict the building type.

[0067] Step 5, Spatial Mapping of Texture Information: As Figure 5 shown in the schematic diagram of the building texture mapping rule. When constructing the spatial position of the street view picture and the building clustering surface for the first time, with the road axis as the center, radiate to both sides with the street view recognition range n as the radius, and calculate any point P c (x c , y c ) of the radiation boundary points P l (x l , y l ), P r (x r , y r ), where x c represents the longitude of the road point, y c represents the latitude of the road point, x l represents the longitude of the radiation boundary point P l , y l represents the latitude of the radiation boundary point P l , x r represents the longitude of the radiation boundary point P r , y r represents the latitude of the radiation boundary point P r , and perform an intersection analysis of P l , P r with the adjacent building clustering surface, and match the street view pictures corresponding to the point P c to the building clustering surface where its radiation boundary point is located. The coordinate calculation formulas for P l , P r are as follows:

[0068]

[0069] where k is the slope of the line where the road is located, n is the radiation radius on both sides, x c represents the longitude of the road point, y c represents the latitude of the road point, x l,r represents the longitude of P l , P r , and y l,r represents the latitude of P l , P r .

[0070] Then, take the building clustering surface as a unit, count its building type labels and sort them by quantity, retrieve the corresponding building texture according to the highest-level building type, assign the texture information to the corresponding building clustering surface, and automatically splice and generate a suitable texture map according to the building size to complete the spatial mapping of the building texture information.

[0071] Step 6, 3D enhanced modeling: Using the 3D enhanced modeling technology based on SuperMap IDesktop software, according to multiple information such as floor height, number of floors, and texture type in the building data, automatically construct street view enhanced 3D data to achieve a realistic, fast, and 3D display of the regional street view.

[0072] Experimental verification shows that it is feasible to construct a city-level 3D model based on the research method of this paper. The modeling effect is as Figure 6 shown. The total modeling time is 5 days, and the total number of building models reaches 1,641,10. Taking the 3D modeling of the entire area of Nanjing as an example, comparing the costs of this invention with the automatic modeling based on oblique photogrammetry technology and the 3D construction method based on laser point cloud, which are commonly used in current urban modeling, as shown in Table 1, it can be seen that this invention has great advantages in economic cost and time cost and can meet the environment of the rapid change of urban form in today's society.

[0073] Table 1 Estimation Table of the Whole-region 3D Construction Cost in Nanjing

[0074]

Claims

1. A method for enhancing the modeling of LOD2-level urban building models driven by crowdsourced data, characterized in that: This method uses remote sensing images to extract building area and height information, and comprehensively uses multi-element information for building area landscape clustering. It collects street view picture data, trains a building facade object detection model, automatically predicts building texture information, and constructs a spatial mapping relationship between building information and building clusters to achieve the construction of LOD2-level urban building models in the urban area, meeting the requirements of level 2 of multi-level details of the model; This method includes the following steps: Step 1, extraction of building outline and height information: Construct a deep learning image segmentation model, and based on high-resolution remote sensing images, automatically extract building areas and building shadows, and match building areas and their shadows according to multiple principles including adjacency and direction consistency; Use mathematical morphology to characterize the spatial structure characteristics of buildings, and optimize the building outline through a series of morphological opening operations and closing operations; Based on the relationship between the solar altitude angle and the satellite altitude angle, calculate the corresponding building height according to the length of the building shadow; Step 2, building landscape clustering analysis: Comprehensively consider the geometric characteristics of building area outlines, building spatial relationships, and building real texture information. Through building block division based on road and spatial proximity, building feature clustering based on the Self-Organizing Mapping (SOM) neural network, and building clustering based on the semantic information of Points of Interest (POIs), realize the clustering of similar buildings and obtain building clustering surfaces; Step 3, street view data collection: Based on the public transportation road network data of crowdsourced data, extract road points at fixed intervals, and obtain street view data within the adjacent range based on the spatial information of the road points; Step 4, intelligent interpretation and analysis of building categories: Comprehensively consider multiple factors including different regional characteristics, building styles, building uses, and building densities, and construct a building classification system with building structure, building coatings, and component distribution as specific features; build a building sample library and a texture library according to this system; input the data of the building type sample library into a deep learning convolutional neural network, train the building type AI recognition and detection model and conduct accuracy evaluation; input the street view data of the modeling area obtained in Step 2 into this model, and the buildings and their types in the pictures can be automatically predicted; Step 5, spatial mapping of texture information: For the spatial positions of street view pictures and building clustering surfaces, associate the street view pictures with the building clustering surfaces. Take the building clustering surface as a unit, count its building type labels and sort them by quantity, and retrieve the corresponding texture according to the highest-level building type, and assign the texture information to the corresponding building clustering surface to complete the spatial mapping of the building area and texture information; Step 6, three-dimensional enhanced modeling: Based on the GIS desktop platform, automatically generate a three-dimensional model of the building according to the building height, structure, and texture information, and quickly realize LOD2-level urban three-dimensional enhanced modeling.

2. A method for enhancing the modeling of LOD2-level urban building models driven by crowdsourced data according to claim 1, characterized in that: In the above Step 1, the specific steps for extracting building outline and height information based on high-resolution remote sensing image data are as follows: Step 2.

1. Extract building and shadow information using semantic segmentation method: Construct a sample library by manually vectorizing buildings and shadows in high-resolution images, and then use a deep learning convolutional neural network to train an image semantic segmentation model to automatically extract high-resolution building outlines and shadow information, and use quadratic cross-entropy as the loss function to solve and evaluate the model; Step 2.

2. Optimize building surfaces: Construct a morphological profile through a series of morphological opening and closing operations, take the average of the opening profile (OP) and closing profile (CP) of the pixels within the building object to obtain the morphological opening profile mean and morphological closing profile mean of the building object, and extract the morphological features of the building; Step 2.

3. Extract height features: Match building surfaces and shadows according to the principles of adjacency, consistent direction, and at least one shadow; Generate several parallel lines according to the solar azimuth direction, cut the parallel lines by the shadows and calculate the mean value of the line segment lengths as the building shadow length; and calculate the building height based on the spatial relationship between the solar altitude angle, satellite altitude angle and the building. The specific method is as follows: When the sun and the satellite are on the same side of the building, the building height calculation can be seen in formula (1). When the sun and the satellite are on different sides of the building, the building height calculation can be seen in formula (2). H = L_B × tanω = L_S × tanω Formula (2) Where, L_S is the length of the shadow on the image, L_B is the actual shadow length of the building, σ is the solar altitude angle, and ω is the satellite altitude angle.

3. A crowdsourced data-driven LOD2-level urban three-dimensional enhanced modeling method according to claim 1, characterized in that: In step 2, by integrating the geometric features of the building surface contour, the building spatial relationship, and the building real texture information, through building block division based on road and spatial proximity, building feature clustering based on SOM neural network, and building clustering based on POI semantic information, similar building clustering is realized to obtain a building clustering surface. The specific steps are as follows: Step 3.

1. Building block division based on road and spatial proximity: Describe the building proximity relationship by constructing a Delaunay triangulation network, construct a minimum spanning tree and a road network with the minimum distance between buildings as the metric, identify spatially adjacent building clusters, and build the building block spatial relationship; Step 3.

2. Building feature clustering based on SOM neural network: Select multiple factors including building height, direction, shape, and area to describe the geometric features of the building, input the SOM neural network to extract the characteristics of the buildings on the map; Through the expression of different factors, generate a geometric feature map of the building spatial feature relationship, and classify the buildings in the block; Step 3.

3. Building aggregation based on POI semantic information: Based on the results of step 2.1 and step 2.2, by establishing the matching relationship between POIs, building textures and buildings, associate the semantic information of POIs and the building texture features with buildings to achieve the aggregation of buildings based on semantic and texture features.

4. A method for LOD2-level urban three-dimensional enhanced modeling driven by crowdsourced data according to claim 1, characterized in that: [[ID= 5. A method for LOD2-level urban three-dimensional enhanced modeling driven by crowdsourced data according to claim 1, characterized in that: In step 4, the specific steps of intelligent interpretation and analysis of building categories are as follows: Step 5.1: Constructing a building classification system and texture library: Taking into account multiple factors including regional characteristics, architectural styles, building uses, and building density, and using building structure, architectural coatings, and component distribution as specific features, we construct a building classification system, create textures for each, and build a building texture library. Step 5.2: Construct a building classification sample library: Obtain a massive amount of street view images from across China, remove invalid images where buildings are too far away, too small, or obscured, and manually label valid street view data based on the building classification system to construct a building classification sample set. This sample dataset is then augmented to enrich the sample types and improve model robustness. Step 5.3: Training and Accuracy Evaluation of the AI Building Type Recognition Detection Model: Use an open-source framework to build and train the object detection model. Measure the accuracy of the deep learning-based object detection model and evaluate the accuracy of the model detection results from multiple perspectives, including test sets and actual inspection results. Step 5.4: Intelligent detection of building types: Input the image into the trained object detection model to automatically predict the building and its type.

6. A crowdsourced data-driven LOD2-level urban three-dimensional enhanced modeling method according to claim 1, characterized in that: In step 5 described above, when constructing the spatial position of the street view image and the building clustering surface, it is necessary to take the road axis as the center and radiate to both sides with the street view recognition range n as the radius, and calculate any point P on the road c (x c ,y c )'s radiation boundary points P l (x l ,y l ), P r (x r ,y r ), where x c represents the longitude of the road point, y c represents the latitude of the road point, x l represents the longitude of the radiation boundary point P l ; y l represents the latitude of the radiation boundary point P l ; x r represents the longitude of the radiation boundary point P r ; y r represents the latitude of the radiation boundary point P r . And perform an intersection analysis on P l , P r and the adjacent building clustering surface, and match the street view images corresponding to the point P c to the building clustering surface where their radiation boundary points are located respectively; The coordinate calculation formulas for P l , P r are as follows: where k is the slope of the straight line where the road is located, n is the radiation radius on both sides, x c represents the longitude of the road point, y c represents the latitude of the road point, x l,r represents the longitude of P l and P r , y l,r represents the latitude of P l and P r ; Then, the building type labels are counted and sorted by number based on the building cluster surface. The corresponding building texture is retrieved according to the highest-level building type, and the texture information is assigned to the corresponding building cluster surface. The texture is automatically spliced to generate a suitable map according to the building size to complete the spatial mapping of the building texture information.

Images