Building structure identification method and model based on machine learning

Abstract

The invention provides a building structure identification method and model based on machine learning, and relates to the technical field of machine learning, and the method comprises the steps: obtaining the geometric contour data of a building through a GIS, and obtaining the POI data of the building from a network map; the method comprises the following steps: processing GIS data, carrying out manual annotation on a building to obtain a training data set, and matching POI data with the building to generate a complete shp format data set; extracting feature values of the building from the training data set, wherein the feature values comprise geometric features, shape features and POI features; performing classification training on the training data set by adopting a pre-clustering active learning method to obtain a building structure prediction model; and based on the prediction model, constructing an urban-scale building structure information resource database. Through the multi-source data fusion and machine learning technology, automatic recognition of the building structure is achieved, the classification accuracy is remarkably improved, the recognition cost is reduced, and the method is particularly suitable for the urban environment containing a large number of old buildings.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning, and in particular to a building structure recognition method and model based on machine learning. Background Art

[0002] With the acceleration of urbanization, the number of older buildings is increasing, making them a key target for future building renovation and resource management. However, due to their age, incomplete documentation and structural information, these buildings pose significant challenges for structural assessment and management. Scientifically identifying and classifying building structures not only helps improve the efficiency and accuracy of older building renovations but also provides crucial data support for urban planning, resource scheduling, building safety testing, and carbon footprint analysis.

[0003] Current methods for building structural identification primarily include visual inspection, ultrasonic testing, infrared thermal imaging, laser scanning, and digital modeling. Visual inspection relies on manual experience and, while low-cost and simple to use, can only detect surface defects and cannot effectively identify underlying structural problems. Technologies such as ultrasonic testing, infrared thermal imaging, and vibration analysis can identify certain internal structural issues, but typically require specialized equipment and highly skilled operators, resulting in high costs and time commitments, and are significantly affected by environmental conditions. Digital modeling constructs three-dimensional models and integrates multiple inspection data to visually display the structural condition of a building. However, this technology requires high hardware performance, modeling standards, and specialized expertise, hindering widespread adoption. Furthermore, existing technologies have limited applicability for older buildings, particularly historic structures. Many traditional methods rely on complex sensors or LiDAR data, which are expensive and time-consuming to acquire, making them particularly inadequate for large-scale, urban-scale applications.

[0004] Although machine learning techniques have been widely used to classify features such as building function, age, and style, research on the classification of building structures remains relatively limited. Compared to other classification tasks, building structure classification is more complex because its inherent attributes are more implicit and diverse. Most existing studies rely on high-dimensional imagery or LiDAR data, which are expensive to acquire and require high data processing capabilities, limiting their feasibility in resource-limited scenarios. In addition, current machine learning algorithms and feature selection schemes for building structure classification are not comprehensive enough and fail to fully explore potential discriminative features, such as the combination of building geometry and functional data. More importantly, current building structure recognition technologies have not yet fully utilized modern data analysis tools, such as machine learning and geographic information systems (GIS), to provide more efficient and low-cost solutions.

[0005] In summary, the current building structure recognition technology has significant deficiencies in economy, applicability and ease of operation. Especially for old buildings and urban application scenarios, it is difficult to meet the needs of high efficiency and low cost. Summary of the Invention

[0006] The purpose of this invention is to provide a building structure recognition method and model based on machine learning, which can better integrate building geometry data, point of interest data and other related information, realize the accurate recognition of old building structures at low cost and high efficiency, and provide reliable data support for building renovation and resource management.

[0007] To achieve the above object, the present invention is implemented through the following technical solutions:

[0008] A building structure recognition method based on machine learning includes the following steps:

[0009] S1: Obtaining geometric outline data of a building through a geographic information system (GIS), and obtaining point of interest (POI) data of the building from a web map;

[0010] S2: Process the GIS data, manually annotate the buildings to obtain a training dataset, and match the POI data with the buildings to generate a complete shp format dataset;

[0011] S3: Extracting feature values ​​of buildings based on the training data set, where the feature values ​​include geometric features, shape features, and POI features;

[0012] S4: using a pre-clustering active learning method to perform classification training on the training data set to obtain a building structure prediction model;

[0013] S5: Based on the building structure prediction model, construct a city-scale building structure information resource database.

[0014] Further: in step S2, the manual labeling includes labeling the building as one of the following structural types: masonry structure, concrete frame, concrete frame-shear wall, concrete shear wall and steel structure.

[0015] Furthermore: the geometric features include: building area, perimeter, number of floors, floor height, length-to-width ratio and height-to-width ratio.

[0016] Furthermore: the shape features include: shape irregularity and the number of contour vertices, wherein the shape irregularity is calculated by the following formula: SI = Ab / AMBR, wherein SI is the shape irregularity of the building contour, Ab is the building contour area, and AMBR is the minimum bounding rectangle area surrounding the building contour.

[0017] Furthermore, the POI features include: the number of POIs, the number of POI types and the main POI types.

[0018] Furthermore: Step S4 specifically includes:

[0019] S401: Use the DBSCAN method to cluster the training data and distinguish noise samples, core samples and non-core samples;

[0020] S402: Perform stratified sampling on the core samples to construct an initial training sample set and an unlabeled sample pool;

[0021] S403: Select a machine learning algorithm to train the initial training sample set;

[0022] S404: Iteratively train the classifier through active learning methods to query and label the most valuable samples.

[0023] Furthermore, the machine learning algorithm includes one or more of the following: gradient boosting decision tree GBDT, decision tree DT, random forest RF, support vector machine SVM, K nearest neighbor KNN and multi-layer perceptron MLP.

[0024] Furthermore, the step S4 further includes:

[0025] S405: Selecting features using feature optimization technology, wherein the feature optimization technology includes a screening method, a packaging method, and an embedding method;

[0026] S406: Calculate the macro-average precision, macro-average recall, and macro-average F1 value to evaluate the classification results.

[0027] The present invention also provides a building structure recognition model based on machine learning, comprising:

[0028] A data acquisition module is used to acquire geometric outline data of buildings through a geographic information system (GIS) and to acquire point of interest (POI) data of the buildings from a web map;

[0029] A data processing module is used to process the GIS data, manually mark the buildings to obtain a training data set, and match the POI data with the buildings to generate a complete shp format data set;

[0030] A feature extraction module is used to extract feature values ​​of buildings from the training data set, wherein the feature values ​​include geometric features, shape features and POI features;

[0031] A prediction modeling module is used to perform classification training on the training data set using a pre-clustering active learning method to obtain a building structure prediction model;

[0032] A database construction module is used to construct a city-scale building structure information resource database based on the building structure prediction model

[0033] Furthermore: the prediction modeling module includes:

[0034] Clustering unit, used to cluster the training data using the DBSCAN method to distinguish noise samples, core samples and non-core samples;

[0035] The sampling unit is used to perform stratified sampling on the core samples and construct the initial training sample set and the unlabeled sample pool;

[0036] A training unit, used to select a machine learning algorithm to train an initial training sample set;

[0037] The iterative optimization unit is used to iteratively train the classifier through active learning methods, and query and label the most valuable samples.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] First, this method significantly improves the accuracy of building structure classification. By introducing a multidimensional feature system consisting of geometric, shape, and POI features, combined with a pre-clustering active learning model, it can fully capture the discriminative characteristics of building structures, achieving a macro-average F1 value of over 90% for building structure classification, providing reliable data support for building renovation and resource management.

[0040] Second, this invention effectively reduces the cost of building structure identification. By using a geographic information system and web map API to acquire building data, it avoids the reliance on expensive equipment and professionals required by traditional methods. Furthermore, through pre-clustering active learning, the number of labeled samples required is significantly reduced while ensuring classification accuracy, significantly reducing manual labeling costs.

[0041] Third, this method automates urban-scale building structure identification. By integrating multi-source data and machine learning techniques, this method efficiently constructs and updates a city-scale database of building structure information resources, overcoming the limitations of traditional technologies in large-scale applications. It is particularly suitable for urban environments with a large number of older buildings. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 Schematic diagram of a flow chart of a building structure recognition method based on machine learning in one embodiment of the present invention. DETAILED DESCRIPTION

[0043] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0044] In the description of the present invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings and are intended solely to facilitate and simplify the description of the present invention. They are not intended to indicate or imply that the devices or components referred to must have, be constructed, or operate in a specific orientation, and therefore should not be construed as limitations on the present invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0045] like Figure 1 As shown, the building structure recognition method based on machine learning provided by the present invention includes the following specific steps:

[0046] First, we use a geographic information system (GIS) to obtain the building's geometric outline data and retrieve the building's point of interest (POI) data from a web map. Specifically, we use the GIS system to obtain information such as the building's footprint, outline, and attributes, and retrieve the POI data from the AutoNavi Map API. POI data is represented as a quad-tuple, consisting of the POI ID, x-coordinate, y-coordinate, and POI type code. These coordinates are converted to the same WGS1984 Mercator coordinate system as the .shp file to match the POIs to the buildings.

[0047] Next, the GIS data was processed and manually annotated based on building codes and technical standards, relevant publications, and online satellite and street-level maps. The present invention categorizes buildings into five common structural types: masonry, concrete frame, concrete frame-shear wall, concrete shear wall, and steel. To ensure the accuracy of the annotations, the present invention makes three assumptions: all masonry buildings have fewer than six floors; public buildings have fewer than 10 floors for concrete frame structures, while concrete frame-shear wall structures have 10 or more floors; and residential buildings have 20 or more floors for concrete shear wall structures, while concrete frame-shear wall structures have 7 to 19 floors, with concrete frame structures having fewer than 7 floors.

[0048] In the feature extraction stage, the present invention extracts building features from three dimensions: geometric features, shape features, and POI features. Geometric features include building area, perimeter, number of floors, floor height, aspect ratio, and height-to-width ratio. Among them, the floor-to-floor height is calculated using the following formula:

[0049]

[0050] Among them H f is the height between floors, H b is the building height, N f is the number of floors. Building area and perimeter were calculated using ArcGIS's geometry calculation functions in the PCS-World1984 Mercator coordinate system. Building length and width were obtained by simplifying the building outline to a minimum bounding rectangle (MBR).

[0051] Shape features include shape irregularity and the number of contour vertices. Shape irregularity is calculated using the following formula:

[0052]

[0053] Among them A b is the building outline area, A MBR It is the area of ​​the minimum bounding rectangle that encloses the building outline. Since the MBR area is always greater than or equal to the building area, the irregularity SI value ranges from (0, 1.0], with lower values ​​indicating higher irregularity. The number of outline vertices is calculated using the "FeatureVerticesToPoints" function of ArcGIS.

[0054] POI features were obtained by counting the total number of POIs and the total number of POI types for each building. The main POI types were identified using a manually developed decision tree (maximum depth 9) and converted into numerical features using binary one-hot encoding. A total of 23 main POI types were identified, of which 16 building-related POI types were selected as features.

[0055] During the modeling phase, this paper employs a pre-clustering active learning approach. First, the training data is clustered using the DBSCAN algorithm to distinguish between noise samples, core samples, and non-core samples. Stratified sampling is performed on the core samples to construct an initial training sample set and an unlabeled sample pool. A variety of machine learning algorithms can be selected for training, including GBDT, DT, RF, SVM, KNN, and MLP. The classifier is iteratively trained using active learning methods to query and annotate the most valuable samples.

[0056] To further improve the performance of the model, the present invention adopts feature optimization techniques, including screening methods (such as F test), packaging methods (such as recursive feature elimination RFE) and embedding methods (such as LASSO). The classification results are evaluated using macro-average precision, recall rate and F1 value. Among them, the macro-average precision is calculated by the following formula

[0057]

[0058] Among them, P m is the macro average precision, C is the total number of categories, and Pi is the precision of each category. Similarly, the macro average recall and F1 value are calculated.

[0059] Finally, based on the trained building structure prediction model, a city-scale building structure information resource database was constructed. Practice has shown that this method achieves over 90% classification accuracy while reducing costs, providing strong support for urban building renovation and resource management.

[0060] The above is a brief introduction to the present invention. The present invention will be described in detail below through specific embodiments.

[0061] A building structure recognition method based on machine learning includes the following steps:

[0062] S1: Obtaining geometric outline data of a building through a geographic information system (GIS), and obtaining point of interest (POI) data of the building from a web map;

[0063] S2: Process the GIS data, manually annotate the buildings to obtain a training dataset, and match the POI data with the buildings to generate a complete shp format dataset;

[0064] In some embodiments, in step S2, the manual labeling includes labeling the building as one of the following structural types: masonry structure, concrete frame, concrete frame-shear wall, concrete shear wall, and steel structure.

[0065] The matching rules and criteria for POI data and buildings are as follows:

[0066] Get the POI data matching the selected building from the AutoNavi map and express it in a four-tuple format: {POIID, x-coordinate, y-coordinate, POI type code}.

[0067] Convert the coordinates in the AutoNavi map to the same WGS1984 World Mercator coordinate system as the .shp file. POIs in Amap are grouped as follows: "120000" refers to "Commercial Housing" POIs, "120200" refers to "Architecture-Related" POIs, and "120201" refers to "Commercial Office Building" POIs. Amap contains 23 major POI types. Use the ArcGIS SpatialJoin function to map POIs to buildings, setting the parameters JoinONE_TO_MANY, MatchOption=Within_A_Distance, and SearchRadius=decimaldegrees.

[0068] POIs with the following seven main type codes were excluded: 040000, 110000, 150000, 180000, 220000, 970000, and 990000, as they are not associated with buildings.

[0069] The remaining 16 major POI types are used for feature definition.

[0070] S3: Extracting feature values ​​of buildings based on the training data set, where the feature values ​​include geometric features, shape features, and POI features;

[0071] The geometric features include: building area, perimeter, number of floors, floor height, length-to-width ratio and height-to-width ratio;

[0072] The shape features include: shape irregularity and the number of contour vertices, wherein the shape irregularity is calculated by the following formula: SI=Ab / AMBR, wherein SI is the shape irregularity of the building contour, Ab is the area of ​​the building contour, and AMBR is the area of ​​the minimum bounding rectangle surrounding the building contour;

[0073] The POI features include: the number of POIs, the number of POI types and the main POI types.

[0074] S4: using a pre-clustering active learning method to perform classification training on the training data set to obtain a building structure prediction model;

[0075] In some other embodiments, step S4 specifically includes:

[0076] S401: Use the DBSCAN method to cluster the training data and distinguish noise samples, core samples and non-core samples;

[0077] The DBSCAN algorithm was chosen because it can classify training data into three categories: core, non-core, and noise samples. This is essential for forming the initial training samples and unlabeled sample pool required for pre-clustering active learning (AL) methods. Compared to some other clustering algorithms, DBSCAN does not assume the distribution shape of clustered samples (such as k-means and meanshift, which assume a convex distribution) nor does it require a predefined number of clusters (such as k-means, spectral clustering, and agglomerative clustering). It clusters samples based on their density, providing more flexibility in shape and number of clusters, allowing for more accurate clustering.

[0078] In terms of parameter selection:

[0079] DBSCAN uses two important parameters to calculate the density of samples:

[0080] Epsilon (neighborhood radius): This is the radius of a circle centered on a sample, used to calculate the number of neighbors of that sample (i.e., other samples within the circle). For a given epsilon radius, a greater number of neighbors indicates a higher density. In this study, an epsilon value of 0.16 was chosen.

[0081] min_samples (minimum number of samples): This refers to the minimum number of neighbors within the epsilon neighborhood, serving as the minimum threshold for determining whether an area is dense. If a sample has at least min_samples of neighbors within its epsilon neighborhood, it is considered a core sample. The min_samples value chosen for this study is 48.

[0082] Sample differentiation rules:

[0083] Core sample: A sample is defined as a core sample if it has at least min_samples neighbors within its neighborhood with a radius of epsilon.

[0084] Non-core sample: A sample is a neighbor of a core sample, but the number of its neighbors in its epsilon neighborhood is less than min_samples, then the sample is a non-core sample.

[0085] Noise samples: The remaining samples in the training data except the core samples and non-core samples are regarded as noise samples.

[0086] S402: Perform stratified sampling on the core samples to construct an initial training sample set and an unlabeled sample pool;

[0087] S403: Select a machine learning algorithm to train the initial training sample set;

[0088] S404: Iteratively train the classifier through active learning methods to query and label the most valuable samples.

[0089] Among them, the machine learning algorithm includes one or more of the following: gradient boosting decision tree GBDT, decision tree DT, random forest RF, support vector machine SVM, K nearest neighbor KNN and multi-layer perceptron MLP.

[0090] In some other embodiments, the step S4 further includes:

[0091] S405: Selecting features using feature optimization technology, wherein the feature optimization technology includes a screening method, a packaging method, and an embedding method;

[0092] Among them, the screening method (Filter) is introduced:

[0093] Screening methods use statistical tests to score, rank, and select features, filtering out features with low scores. Their feature selection process is independent of the machine learning algorithm.

[0094] Specific methods and parameter examples

[0095] F-test

[0096] Methods: Feature importance was assessed by calculating the ratio of between-group variance to within-group variance.

[0097] Parameters: No specific additional parameters, its calculation is based on the variance characteristics of the data itself.

[0098] Mutual Information

[0099] Method: It measures the interdependence between two variables and is used to assess the degree of association between a feature and a target variable (e.g. building structure type).

[0100] Parameters: No specific additional parameters are used. The mutual information value is determined by calculating the probability distribution of the features and the target variable.

[0101] Wrapper Introduction

[0102] Wrapper methods use greedy search to exhaustively evaluate all possible feature combinations and select the combination that gives the best performance for a particular machine learning algorithm.

[0103] Specific methods and parameter examples

[0104] Recursive Feature Elimination (RFE)

[0105] Method: First, use a machine learning model (such as support vector machine, decision tree, etc.) to train all features, and then gradually delete unimportant features based on the model's importance evaluation indicators (such as feature importance scores in decision trees). Each time one or more features are deleted, the model is retrained until a certain stopping condition is reached (such as only a specified number of features are left).

[0106] Parameters: This may involve selecting an initial machine learning model (for example, in some implementations, different classifiers can be selected as base models), as well as determining the number of features to remove each time or the stopping condition (such as stopping when the number of remaining features reaches a certain preset value).

[0107] Introduction to Embedded Method

[0108] Embedding methods use feature selection as a component of machine learning algorithms, selecting features during the algorithm training process.

[0109] Specific methods and parameter examples

[0110] Least Absolute Shrinkage and Selection Operator (LASSO)

[0111] Method: Add an L1 regularization term to the objective function so that the coefficients of some features shrink to 0 during the optimization process, thereby achieving feature selection.

[0112] Parameters: The main parameter is the regularization parameter lambda, which controls the strength of regularization. Smaller lambda values ​​allow more features to be retained, while larger lambda values ​​shrink more feature coefficients to 0, thereby selecting fewer features.

[0113] ElasticNet

[0114] Method: It combines L1 and L2 regularization, which can perform feature selection like LASSO and handle the correlation between features like ridge regression.

[0115] Parameters: These include the L1 and L2 regularization weights alpha and lambda. Alpha controls the ratio of L1 to L2 regularization, while lambda controls the strength of regularization. By adjusting these parameters, you can control the effectiveness of feature selection and the complexity of the model.

[0116] S406: Calculate the macro-average precision, macro-average recall, and macro-average F1 value to evaluate the classification results.

[0117] S5: Based on the building structure prediction model, construct a city-scale building structure information resource database.

[0118] In another embodiment, the present invention further provides a building structure recognition model based on machine learning, comprising:

[0119] A data acquisition module is used to acquire geometric outline data of buildings through a geographic information system (GIS) and to acquire point of interest (POI) data of the buildings from a web map;

[0120] A data processing module is used to process the GIS data, manually mark the buildings to obtain a training data set, and match the POI data with the buildings to generate a complete shp format data set;

[0121] A feature extraction module is used to extract feature values ​​of buildings from the training data set, wherein the feature values ​​include geometric features, shape features and POI features;

[0122] A prediction modeling module is used to perform classification training on the training data set using a pre-clustering active learning method to obtain a building structure prediction model;

[0123] A database construction module is used to construct a city-scale building structure information resource database based on the building structure prediction model

[0124] Wherein, the prediction modeling module includes:

[0125] Clustering unit, used to cluster the training data using the DBSCAN method to distinguish noise samples, core samples and non-core samples;

[0126] The sampling unit is used to perform stratified sampling on the core samples and construct the initial training sample set and the unlabeled sample pool;

[0127] A training unit, used to select a machine learning algorithm to train an initial training sample set;

[0128] The iterative optimization unit is used to iteratively train the classifier through active learning methods, and query and label the most valuable samples.

[0129] The above embodiments are intended only to illustrate the technical concepts and features of the present invention. Their purpose is to enable those skilled in the art to understand the contents of the present invention and implement them accordingly. They are not intended to limit the scope of protection of the present invention. Any equivalent changes or modifications made in accordance with the spirit of the present invention are intended to be covered by the scope of protection of the present invention.

Claims

1. A building structure recognition method based on machine learning, characterized in that: The following steps are involved: S1: obtaining geometric outline data of a building through a geographic information system (GIS), and obtaining point of interest (POI) data of the building from a network map; S2: Process the GIS data, manually annotate the buildings to obtain a training data set, and match the POI data with the buildings to generate a complete shp format data set; S3: extracting feature values ​​of buildings based on the training data set, wherein the feature values ​​include geometric features, shape features, and POI features; S4: using a pre-clustering active learning method to perform classification training on the training data set to obtain a building structure prediction model; S5: Based on the building structure prediction model, construct a city-scale building structure information resource database.

2. The method for identifying building structures based on machine learning according to claim 1, characterized in that: In the step S2, manual labeling includes labeling the building as one of the following structural types: masonry structure, concrete frame, concrete frame-shear wall, concrete shear wall and steel structure.

3. The method for identifying building structures based on machine learning according to claim 1, characterized in that: The geometric features include: building area, perimeter, number of floors, floor height, length-to-width ratio and height-to-width ratio.

4. The method for identifying building structures based on machine learning according to claim 1, characterized in that: The shape features include: shape irregularity and the number of contour vertices, wherein the shape irregularity is calculated by the following formula: SI = Ab / AMBR, wherein SI is the shape irregularity of the building contour, A b is the building outline area, A MBR It is the minimum bounding rectangle area surrounding the building outline.

5. The method for identifying building structures based on machine learning according to claim 1, characterized in that: The POI features include: the number of POIs, the number of POI types and the main POI types.

6. The method for identifying building structures based on machine learning according to claim 1, characterized in that: The step S4 specifically includes: S401: Use DBSCAN method to cluster the training data and distinguish noise samples, core samples and non-core samples; S402: Perform stratified sampling on the core samples to construct an initial training sample set and an unlabeled sample pool; S403: Select a machine learning algorithm to train the initial training sample set; S404: Iteratively train the classifier through active learning methods to query and label the most valuable samples.

7. The method for identifying building structures based on machine learning according to claim 6, characterized in that: The machine learning algorithm includes one or more of the following: gradient boosting decision tree GBDT, decision tree DT, random forest RF, support vector machine SVM, K nearest neighbor KNN and multi-layer perceptron MLP.

8. The method for identifying building structures based on machine learning according to claim 6, characterized in that: The step S4 further comprises: S405: Selecting features using feature optimization technology, wherein the feature optimization technology includes a screening method, a packaging method, and an embedding method; S406: Calculate the macro-average precision, macro-average recall, and macro-average F1 value to evaluate the classification results.

9. A building structure recognition model based on machine learning, characterized in that: include: A data acquisition module is used to acquire geometric contour data of buildings through a geographic information system (GIS), and to acquire point of interest (POI) data of the buildings from a network map; A data processing module is used to process the GIS data, manually annotate the buildings to obtain a training data set, and match the POI data with the buildings to generate a complete shp format data set; A feature extraction module, used to extract feature values ​​of buildings from the training data set, wherein the feature values ​​include geometric features, shape features and POI features; A prediction modeling module, used for performing classification training on the training data set by using a pre-clustering active learning method to obtain a building structure prediction model; The database construction module is used to construct a city-scale building structure information resource database based on the building structure prediction model.

10. The building structure recognition model according to claim 9, characterized in that: The prediction modeling module includes: The clustering unit is used to cluster the training data using the DBSCAN method to distinguish noise samples, core samples, and non-core samples; The sampling unit is used to perform stratified sampling on the core samples and construct the initial training sample set and the unlabeled sample pool; A training unit, used for selecting a machine learning algorithm to train an initial training sample set; The iterative optimization unit is used to iteratively train the classifier through active learning methods, and query and label the most valuable samples.

Images