Building three-dimensional modeling method and system combining deep learning and satellite image

Abstract

The application discloses a building three-dimensional modeling method and system combining deep learning and satellite images, relates to the technical field of building modeling, and comprises the following steps: collecting open-source satellite images of a target building, and performing image preprocessing to obtain standard satellite images; adopting a deep learning model to extract contour features and boundary features as building contour data, and to extract building detail features as building detail data; calling a historical building height prediction model library to integrate and construct an adaptive height prediction channel; using the adaptive height prediction channel to perform building height prediction according to image spectral features and image texture features, and outputting predicted building heights; and performing three-dimensional simulation modeling on the target building according to the building contour data, the building detail data and the predicted building heights to obtain a building three-dimensional model. The application solves the technical problems of poor building height prediction accuracy and high training cost in the prior art.

Description

Technical Field

[0001] This invention relates to the field of building modeling technology, and more specifically to a method and system for 3D building modeling that combines deep learning and satellite imagery. Background Technology

[0002] In the field of 3D building modeling, traditional methods mainly rely on techniques such as manual on-site measurement. While these methods can obtain relatively accurate modeling data, they suffer from significant limitations, including high costs, low efficiency, and long update cycles. This is particularly true for large-scale modeling needs, where traditional methods are difficult to implement and uneconomical. In recent years, with the development of remote sensing technology, satellite imagery-based modeling methods have gradually been applied. However, traditional processing methods have limited accuracy in building outline extraction and detailed feature recognition, especially in recognizing complex building structures, which can easily lead to model distortion. Particularly in building height prediction, existing methods often rely on simple height regression models, resulting in low prediction accuracy and failing to meet the demands of high-precision 3D modeling. Furthermore, they cannot dynamically adjust prediction strategies based on different building features, leading to high training costs and poor prediction accuracy. Summary of the Invention

[0003] This application provides a method and system for 3D building modeling that combines deep learning and satellite imagery, in order to address the technical problems of poor accuracy in building height prediction and high training costs in the prior art.

[0004] In view of the above problems, this application provides a method and system for 3D modeling of buildings that combines deep learning and satellite imagery.

[0005] In a first aspect, this application provides a method for 3D modeling of buildings combining deep learning and satellite imagery, the method comprising: Acquire open-source satellite imagery of the target building and perform image preprocessing to obtain standard satellite imagery; Using a deep learning model, the contour and boundary features of the standard satellite imagery are extracted as building contour data, and the building detail features are extracted as building detail data. Based on the image spectral and texture features of the standard satellite imagery, as well as the building outline and building detail data, the historical building height prediction model library is invoked to integrate and construct an adaptive height prediction channel. Using the adaptive height prediction channel, the building height is predicted based on the image spectral features and image texture features, and the predicted building height is output. Based on the building outline data, building detail data, and predicted building height, a three-dimensional simulation model of the target building is performed to obtain the building's three-dimensional model.

[0006] Secondly, this application provides a 3D building modeling system that combines deep learning and satellite imagery, including: The image acquisition module is used to acquire open-source satellite imagery of the target building and perform image preprocessing to obtain standard satellite imagery. The feature extraction module is used to extract the contour features and boundary features of the standard satellite image as building contour data and extract the building detail features as building detail data using a deep learning model. The prediction channel construction module is used to integrate and construct an adaptive height prediction channel based on the image spectral and texture features of the standard satellite image, as well as the building outline and building detail data, by calling the historical building height prediction model library. The height prediction module is used to predict building height based on the image spectral features and image texture features using the adapted height prediction channel, and output the predicted building height. The 3D modeling module is used to perform 3D simulation modeling of the target building based on the building outline data, building detail data, and predicted building height, to obtain a 3D building model.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application proposes a method and system for 3D building modeling that combines deep learning and satellite imagery. Through automatic extraction of building contours and detailed features based on deep learning, intelligent fusion of multi-source image features and historical prediction models, and adaptive height prediction channel construction, the automation, accuracy, and reliability of building height prediction are significantly improved. Compared with traditional methods, the technical solution provided in this application significantly overcomes the shortcomings of traditional methods in terms of insufficient height prediction accuracy, achieving efficient, accurate, and adaptive building height prediction, and providing an accurate data foundation for precise and reliable 3D building modeling. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 This is a flowchart illustrating the method for 3D building modeling that combines deep learning and satellite imagery, as provided in an embodiment of this application.

[0010] Figure 2 A schematic diagram of the structure of a 3D building modeling system combining deep learning and satellite imagery provided in an embodiment of this application.

[0011] The components represented by each mark in the attached diagram are explained below: Image acquisition module 100, feature extraction module 200, prediction channel construction module 300, height prediction module 400, and 3D modeling module 500. Detailed Implementation

[0012] This application provides a method and system for 3D building modeling that combines deep learning and satellite imagery, in order to address the technical problems of poor accuracy in predicting building height and high training costs in existing technologies.

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0014] It should be noted that the terms "comprising" and "having" are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to these processes, methods, products, or devices.

[0015] Example 1, as Figure 1 As shown, this application provides a method for 3D modeling of buildings combining deep learning and satellite imagery, wherein the method includes: S10: Acquire open-source satellite imagery of the target building and perform image preprocessing to obtain standard satellite imagery.

[0016] In the process of 3D building modeling based on satellite imagery, the source satellite imagery usually has problems such as radiometric distortion, geometric distortion, inconsistent resolution and noise interference. Directly using these original images will seriously affect the accuracy of subsequent feature extraction and model construction.

[0017] Step S10 in the method provided in this application embodiment includes: According to the preset image processing scheme, the open-source satellite imagery is subjected to radiometric correction, geometric correction, resolution normalization, and image fusion enhancement to obtain standard satellite imagery.

[0018] In this embodiment, open-source satellite imagery of the target building is acquired, and preprocessed according to a preset image processing scheme to improve image quality. Preprocessing includes radiometric correction, geometric correction, resolution normalization, and image fusion enhancement. Radiometric correction aims to eliminate radiometric distortion caused by the sensor itself and atmospheric scattering; for example, using the radiometric calibration tool in ENVI software to convert the raw digital quantization values ​​of image pixels into accurate surface reflectance values. Geometric correction corrects geometric distortions caused by topographic undulations and other factors in the original imagery by using an RPC model combined with ground control points (GCPs), ensuring accurate geographic coordinates. Resolution normalization resamples images from different sources to a unified spatial resolution, such as a 1-meter pixel scale. Image fusion enhancement improves image quality in blurred or shadowed areas by fusing images from multiple time periods. After preprocessing, a standard satellite imagery is obtained.

[0019] By performing systematic preprocessing operations such as radiometric correction, geometric correction, resolution normalization, and image fusion enhancement on open-source satellite imagery, the original images were successfully transformed into high-quality standard satellite imagery. This eliminated the distortion and noise of the original images, unified data standards and scales, and provided a clear, accurate, and consistent image data foundation for subsequent feature extraction and altitude prediction.

[0020] S20: Using a deep learning model, extract the contour and boundary features of the standard satellite imagery as building contour data, and extract the building detail features as building detail data.

[0021] Traditional image processing methods rely on hand-designed feature extraction algorithms, which are poorly adapted to complex architectural structures and easily affected by shadows, occlusions, and background interference, resulting in incomplete contours, blurred boundaries, and loss of detailed features. This inaccuracy in feature extraction directly impacts subsequent processes, ultimately causing distortion and structural errors in the 3D model.

[0022] Step S20 in the method provided in this application embodiment includes: Collect sample satellite image sets, perform contour and boundary annotations on buildings in different sample satellite image sets to obtain sample image contour label sets, and perform detailed feature annotations on buildings in different sample satellite image sets to obtain sample image detail label sets; Using the sample satellite image set as input and the sample image contour label set as supervision, a convolutional neural network is trained until convergence to obtain a building contour recognition model. Using the sample satellite image set as input and the sample image detail label set as supervision, a convolutional neural network is trained until convergence to obtain a building detail recognition model. The building outline recognition model and the building detail recognition model are used to extract features from the standard satellite imagery to obtain building outline data and building detail data, respectively.

[0023] In this embodiment, a sample satellite image set is acquired, and the buildings in different sample satellite images are annotated with outlines and boundaries to obtain a sample image outline label set. Detail feature annotations are also performed on the buildings in different sample satellite images to obtain a sample image detail label set. For example, multiple satellite images are acquired as a sample satellite image set, and polygonal outlines are drawn along the outermost edge of each building in the satellite image set. The coordinate data of the outlines are extracted to obtain sample image outline labels, which are then integrated to obtain a sample image outline label set. Detail annotations are performed on the acquired satellite images, for example, annotating the number of windows in the building images, and roof structures such as flat roofs, pitched roofs, and curved roofs, as image detail labels, which are then integrated to obtain a sample image detail label set.

[0024] A building outline recognition model is constructed based on the U-Net++ deep learning network. U-Net++ is a convolutional neural network used for image semantic segmentation tasks. An exemplary four-layer structure is adopted: the input layer receives satellite imagery, the convolutional layers use 32 3×3 convolutional kernels, the pooling layers use 2×2 max pooling, the fully connected layers have 32 nodes activated using the ReLU function, and the output layer outputs the recognized building outline labels. Training is performed using a sample satellite imagery set as input and a sample image outline label set as supervision. Images from the sample satellite imagery set are input into the building outline recognition model to obtain building outline labels. The obtained building outline labels are compared with the true sample image outline labels, and the cross-entropy loss function is used to calculate the prediction error. The Adam optimizer is used to automatically adjust the model's internal parameters through backpropagation to minimize this loss value. Iterative optimization continues until the loss value no longer decreases significantly. For example, if the accuracy of the output building outline labels is above 90% when inputting satellite imagery, the building outline recognition model is considered successfully trained.

[0025] Using the same method, a building detail recognition model was constructed and trained based on the U-Net++ deep learning network. An exemplary four-layer structure was adopted, where the input layer receives satellite imagery, the convolutional layers use 32 3×3 convolutional kernels, the pooling layers use 2×2 max pooling, the fully connected layers have 32 nodes activated using the ReLU function, and the output layer outputs the recognized building detail labels. Using a sample satellite imagery set as input and a sample image detail label set as supervision, the convolutional neural network was trained until convergence using the same training method as the building outline recognition model, resulting in the building detail recognition model.

[0026] By using building outline recognition models and building detail recognition models, features are extracted from standard satellite images, and the extracted features are integrated into feature vectors to obtain building outline data and building detail data.

[0027] By employing a deep learning-based convolutional neural network model, intelligent and high-precision extraction of building outline features, boundary features, and detail features is achieved. This accurately identifies complex building structures, resists interference from light and shadow and occlusion, and extracts complete and accurate building outline data and rich detail feature data, providing reliable feature input for subsequent height prediction and 3D reconstruction.

[0028] S30: Based on the image spectral and texture features of the standard satellite imagery, as well as the building outline and building detail data, call the historical building height prediction model library and integrate to construct an adaptive height prediction channel.

[0029] Existing building height prediction technologies typically rely on independent models trained according to architectural style, which are prone to failure when faced with unknown or mixed-style buildings, resulting in a significant drop in prediction accuracy. At the same time, each style requires separate training and maintenance of the model, leading to high training costs, poor scalability, and difficulty in fully utilizing shared architectural features between different styles, thus limiting the generalization ability of height prediction.

[0030] Step S30 in the method provided in this application embodiment includes: A library of historical building height prediction models is constructed, comprising several building height prediction models. These models are built based on deep learning, with image spectral features and image texture features as inputs and building height as output. Obtain several sample training sets and several sample data feature sets for the several building height prediction models. Based on the image spectral features, image texture features, building outline data and building detail data, perform feature association analysis on the several sample training sets and several sample data feature sets respectively, and output several feature association degrees. This involves acquiring several sample training sets and several sample data feature sets for the several building height prediction models. Based on the image spectral features, image texture features, building outline data, and building detail data, feature association analysis is performed on the several sample training sets and several sample data feature sets, including: Obtain several sample data feature sets for the several building height prediction models, wherein the sample data features are the sample building outline data and sample building detail data of the sample building corresponding to the sample training data; Randomly select the first building height prediction model and obtain the first sample training set and the first sample data feature set; Based on the image spectral features and image texture features, feature association analysis is performed on multiple first sample training data in the first sample training set to calculate the average association degree of the first training data. Based on the building outline data and building detail data, feature association analysis is performed on multiple first sample data features in the first sample data feature set to calculate the average first data feature association degree. The first feature correlation degree is calculated based on the average correlation degree of the first training data and the average correlation degree of the first data features, and several feature correlation degrees are obtained by sequential analysis. The first feature correlation degree is calculated based on the mean correlation degree of the first training data and the mean correlation degree of the first data features, including: Feature analysis is performed on multiple first sample training data in the first sample training set to construct multiple first feature vectors; Distribute the plurality of first feature vectors into a three-dimensional space and calculate the first distribution discreteness; The ratio of the first distribution dispersion to the preset standard distribution dispersion of the model training is set as the training weight adjustment coefficient, and multiplied by the initial training weight to obtain the adaptive training weight, wherein the initial training weight is 0.7, and the adaptive training weight is greater than or equal to 0.5 and less than or equal to 0.9. The adaptive data feature weights are obtained by subtracting the adaptive training weights from 1. Based on the adaptive training weights and the adaptive data feature weights, the mean correlation of the first training data and the mean correlation of the first data features are weighted and calculated to obtain the first feature correlation. Based on the aforementioned feature correlation, a model integration strategy is set, and the historical building height prediction model library is called according to the model integration strategy to construct an adapted height prediction channel. The model integration strategy based on the aforementioned feature correlations includes: Select the minimum feature correlation degree among the several feature correlation degrees, and round down the ratio of the feature correlation degree to the minimum feature correlation degree to set it as the model selection number, thus obtaining several model selection numbers; A model integration strategy is constructed based on the number of models selected and the average fusion mechanism.

[0031] In this embodiment, a historical building height prediction model library is constructed. This library includes several building height prediction models, each used to identify buildings of different architectural styles, such as Chinese, European, and Italian architecture. The height prediction models are built using deep learning, with image spectral features and image texture features as inputs and building height as the output. For example, a four-layer building height prediction model is constructed: the input layer receives image spectral features and image texture features; the first hidden layer uses 64 nodes activated by the ReLU function; the second hidden layer uses 32 nodes activated by the ReLU function; and the output layer outputs the predicted building height.

[0032] Specifically, firstly, several sample data feature sets of several building height prediction models are obtained, where the sample data features are the sample building outline data and sample building detail data of the sample buildings corresponding to the sample training data.

[0033] A first building height prediction model is randomly selected, and a first sample training set and a first sample data feature set are randomly obtained as the current processing object.

[0034] Several training sets of samples for several building height prediction models and several sample data feature sets are obtained. Based on image spectral features, image texture features, building outline data and building detail data, feature correlation analysis is performed on several training sets of samples and several sample data feature sets respectively, and several feature correlation degrees are output.

[0035] Specifically, based on image spectral features and image texture features, feature association analysis is performed on multiple first-sample training data in the first-sample training set to calculate the average association degree of the first-sample training data. Image spectral features refer to the energy values ​​of ground objects reflected or radiated by satellite sensors in multiple different electromagnetic wave bands. Each pixel in a multispectral satellite image contains a sequence of reflectance values ​​in different bands, which can characterize the material composition and surface properties of ground objects. For example, through principal component analysis, features such as reflectance from multiple highly correlated bands are integrated into a feature vector as image spectral features. Image texture features are calculated by analyzing the spatial distribution and statistical characteristics of grayscale values ​​within a pixel's neighborhood, such as a 5×5 window, and characterize the surface roughness of ground objects in the image. For example, the frequency of pixel pairs with a certain distance and direction in the image is calculated using a gray-level co-occurrence matrix to quantify image texture features, such as contrast and entropy, and integrated into a feature vector as image texture features.

[0036] Calculate the cosine similarity between the image spectral features and image texture features of the current target building and the features of each sample in the first sample training set. Then, calculate the arithmetic mean of all similarity values ​​to obtain the mean of the correlation degree of the first training data.

[0037] Based on image spectral features and image texture features, feature association analysis is performed on multiple first sample training data in the first sample training set. For example, the cosine similarity between the image spectral features and image texture features of multiple first sample training data and other sample training data is calculated, and the arithmetic mean is obtained to get the mean correlation degree of the first training data.

[0038] Based on building outline data and building detail data, feature association analysis is performed on multiple features of the first sample data feature set. For example, the cosine similarity between the building outline data and building detail data of multiple first sample training data and other sample training data is calculated, and the arithmetic mean is obtained to obtain the mean value of the first data feature association degree.

[0039] Perform feature analysis on multiple first sample training data in the first sample training set to construct multiple first feature vectors; Multiple first feature vectors are distributed in a three-dimensional space. For example, principal component analysis can be used to reduce the dimensionality of the multiple first feature vectors. After dimensionality reduction, feature points are obtained in multiple three-dimensional spaces. The average Euclidean distance between the feature points is calculated to obtain the first distribution dispersion. The higher the dispersion of the training data distribution, the better the generalization ability of the model, and the greater the weight of the correlation between training data should be.

[0040] The ratio of the first distribution dispersion to the preset standard distribution dispersion of the model training is set as the training weight adjustment coefficient, which is multiplied by the initial training weight to obtain the adaptive training weight. The initial training weight is 0.7, and the adaptive training weight is greater than or equal to 0.5 and less than or equal to 0.9. When the calculated adaptive training weight is less than 0.5, it is set to 0.5; when the calculated adaptive training weight is greater than 0.9, it is set to 0.9. The weights focus on the model's generalization ability. The higher the dispersion of the training data distribution, the better the generalization ability of the sample model, and the larger the weight ratio of the training data correlation should be, thus the smaller the adaptive data feature weight should be. Conversely, the lower the dispersion of the training data distribution, the worse the generalization ability of the sample model, and the smaller the weight ratio of the training data correlation should be, thus the larger the adaptive data feature weight should be. By setting dynamic weights, analysis and dynamic adjustments can be made based on the scenario and data, improving the accuracy and reliability of feature correlation calculation.

[0041] The adaptive data feature weights are obtained by subtracting the adaptation training weights from 1, i.e., Adapted data feature weights = 1 - Adapted training weights.

[0042] Based on the adapted training weights and adapted data feature weights, the mean correlation degree of the first training data and the mean correlation degree of the first data feature are weighted. The first feature correlation degree is calculated as follows: First feature correlation degree = (adapted training weights × mean correlation degree of the first training data) + (adapted data feature weights × mean correlation degree of the first data feature).

[0043] Among several feature correlation degrees, the minimum feature correlation degree is selected. The ratio of the feature correlation degree to the minimum feature correlation degree is rounded up and set as the model selection quantity. This yields several model selection quantities. For example, if a feature correlation degree is 0.8 and the minimum feature correlation degree is 0.2, then the model selection quantity = feature correlation degree ÷ minimum feature correlation degree = 0.8 ÷ 0.2 = 4. If the result is not an integer, it is rounded up, so the model selection quantity is 4.

[0044] A model integration strategy is constructed based on the number of models selected and an average fusion mechanism. Specifically, models are selected from the model library according to the number of models selected for each model for integration. The output of the integrated model is the average fusion of the outputs of the multiple integrated models. Following the model integration strategy, a historical building height prediction model library is invoked to construct an adapted height prediction channel.

[0045] Preferably, the constructed adaptive height prediction channel can be further optimized through transfer learning to further improve the model's generalization ability. Specifically, feature extraction layer parameters are extracted from the multiple models included in the adaptive height prediction channel to construct a shared feature extraction foundation. Then, a style adaptive prediction layer is built on top of this foundation. This layer contains two fully connected layers: the first fully connected layer uses 32 nodes and ReLU activation, the second fully connected layer uses 16 nodes and ReLU activation, and the output layer uses 1 node and linear activation. First, the shared feature extraction layer parameters are frozen, and only the style adaptive prediction layer is trained with a learning rate of 0.001. After training, all parameters are unfrozen, and fine-tuning is performed with a smaller learning rate, such as 0.0001, using the Adam optimizer and Huber loss function to make it more adaptable to height prediction for various architectural styles.

[0046] By fusing image spectral features, texture features, and building outline and detail data, the building height prediction model is intelligently filtered and integrated, constructing a height prediction channel that is highly adapted to the current target building features. This achieves accurate matching between the prediction model and building features, providing accurate height prediction solutions for different buildings while significantly reducing training costs.

[0047] S40: Using the adaptive height prediction channel, predict the building height based on the image spectral features and image texture features, and output the predicted building height.

[0048] In this embodiment, the acquired image spectral features and image texture features are input into the adaptive height prediction channel to predict the building height and output the predicted building height.

[0049] S50: Based on the building outline data, building detail data and predicted building height, perform three-dimensional simulation modeling of the target building to obtain a three-dimensional building model.

[0050] In this embodiment, a 3D simulation model of the target building is performed based on building outline data, building detail data, and predicted building height to obtain a 3D building model. By fusing the building outline data, building detail data, and predicted building height, high-quality 3D building simulation modeling is achieved, ensuring the accuracy of the 3D model in terms of geometry, structural details, and spatial dimensions, ultimately producing a high-quality 3D building model.

[0051] Example 2, as Figure 2 As shown, based on the same inventive concept as the method for 3D building modeling combining deep learning and satellite imagery provided in Embodiment 1, this embodiment of the invention also provides a 3D building modeling system combining deep learning and satellite imagery, including: Image acquisition module 100 is used to acquire open-source satellite imagery of the target building and perform image preprocessing to obtain standard satellite imagery; The feature extraction module 200 is used to extract the contour features and boundary features of the standard satellite image as building contour data and extract building detail features as building detail data using a deep learning model. The prediction channel construction module 300 is used to integrate and construct an adaptive height prediction channel by calling the historical building height prediction model library based on the image spectral features and image texture features of the standard satellite image, as well as the building outline data and building detail data. The height prediction module 400 is used to predict building height based on the image spectral features and image texture features using the adapted height prediction channel, and output the predicted building height. The 3D modeling module 500 is used to perform 3D simulation modeling of the target building based on the building outline data, building detail data and predicted building height, to obtain a 3D building model.

[0052] In one embodiment, the image acquisition module 100 is further configured to: According to the preset image processing scheme, the open-source satellite imagery is subjected to radiometric correction, geometric correction, resolution normalization, and image fusion enhancement to obtain standard satellite imagery.

[0053] In one embodiment, the feature extraction module 200 is further configured to: Collect sample satellite image sets, perform contour and boundary annotations on buildings in different sample satellite image sets to obtain sample image contour label sets, and perform detailed feature annotations on buildings in different sample satellite image sets to obtain sample image detail label sets; Using the sample satellite image set as input and the sample image contour label set as supervision, a convolutional neural network is trained until convergence to obtain a building contour recognition model. Using the sample satellite image set as input and the sample image detail label set as supervision, a convolutional neural network is trained until convergence to obtain a building detail recognition model. The building outline recognition model and the building detail recognition model are used to extract features from the standard satellite imagery to obtain building outline data and building detail data, respectively.

[0054] In one embodiment, the prediction channel building module 300 is further configured to: A library of historical building height prediction models is constructed, comprising several building height prediction models. These models are built based on deep learning, with image spectral features and image texture features as inputs and building height as output. Obtain several sample training sets and several sample data feature sets for the several building height prediction models. Based on the image spectral features, image texture features, building outline data and building detail data, perform feature association analysis on the several sample training sets and several sample data feature sets respectively, and output several feature association degrees. This involves acquiring several sample training sets and several sample data feature sets for the several building height prediction models. Based on the image spectral features, image texture features, building outline data, and building detail data, feature association analysis is performed on the several sample training sets and several sample data feature sets, including: Obtain several sample data feature sets for the several building height prediction models, wherein the sample data features are the sample building outline data and sample building detail data of the sample building corresponding to the sample training data; Randomly select the first building height prediction model and obtain the first sample training set and the first sample data feature set; Based on the image spectral features and image texture features, feature association analysis is performed on multiple first sample training data in the first sample training set to calculate the average association degree of the first training data. Based on the building outline data and building detail data, feature association analysis is performed on multiple first sample data features in the first sample data feature set to calculate the average first data feature association degree. The first feature correlation degree is calculated based on the average correlation degree of the first training data and the average correlation degree of the first data features, and several feature correlation degrees are obtained by sequential analysis. The first feature correlation degree is calculated based on the mean correlation degree of the first training data and the mean correlation degree of the first data features, including: Feature analysis is performed on multiple first sample training data in the first sample training set to construct multiple first feature vectors; Distribute the plurality of first feature vectors into a three-dimensional space and calculate the first distribution discreteness; The ratio of the first distribution dispersion to the preset standard distribution dispersion of the model training is set as the training weight adjustment coefficient, and multiplied by the initial training weight to obtain the adaptive training weight, wherein the initial training weight is 0.7, and the adaptive training weight is greater than or equal to 0.5 and less than or equal to 0.9. The adaptive data feature weights are obtained by subtracting the adaptive training weights from 1. Based on the adaptive training weights and the adaptive data feature weights, the mean correlation of the first training data and the mean correlation of the first data features are weighted and calculated to obtain the first feature correlation. Based on the aforementioned feature correlation, a model integration strategy is set, and the historical building height prediction model library is called according to the model integration strategy to construct an adapted height prediction channel. The model integration strategy based on the aforementioned feature correlations includes: Select the minimum feature correlation degree among the several feature correlation degrees, and round down the ratio of the feature correlation degree to the minimum feature correlation degree to set it as the model selection number, thus obtaining several model selection numbers; A model integration strategy is constructed based on the number of models selected and the average fusion mechanism.

[0055] In summary, the embodiments of this application have at least the following technical effects: This application proposes a method and system for 3D building modeling that combines deep learning and satellite imagery. Through automatic extraction of building contours and details based on deep learning, intelligent fusion of multi-source image features and historical prediction models, and adaptive height prediction channel construction, the automation, accuracy, and reliability of building height prediction are significantly improved. Specifically, the technical solution provided in this application significantly overcomes the shortcomings of traditional methods, such as strong reliance on manual intervention, limited feature extraction capabilities, and weak model generalization ability, achieving accurate processing from satellite imagery to building height prediction. By deeply fusing the spectral and texture features of the imagery with the building contours and details extracted by deep learning methods, and integrating and optimizing using a historical building height prediction model library, it can adapt to the modeling needs of different building types, effectively improving the accuracy of building height prediction. Compared with traditional methods, the technical solution provided in this application significantly overcomes the shortcomings of traditional methods in terms of insufficient height prediction accuracy, achieving efficient, accurate, and adaptive building height prediction, providing an accurate data foundation for precise and reliable 3D building modeling.

[0056] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0057] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0058] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A method for 3D building modeling combining deep learning and satellite imagery, characterized in that, The methods include: Acquire open-source satellite imagery of the target building and perform image preprocessing to obtain standard satellite imagery; Using a deep learning model, the contour and boundary features of the standard satellite imagery are extracted as building contour data, and the building detail features are extracted as building detail data. Based on the spectral and textural features of the standard satellite imagery, as well as the building outline and detail data, a historical building height prediction model library is invoked to integrate and construct an adaptive height prediction channel, including: A library of historical building height prediction models is constructed, comprising several building height prediction models. These models are built based on deep learning, with image spectral features and image texture features as inputs and building height as output. Obtain several sample training sets and several sample data feature sets for the several building height prediction models. Based on the image spectral features, image texture features, building outline data, and building detail data, perform feature association analysis on the several sample training sets and several sample data feature sets respectively, and output several feature association degrees, including: Obtain several sample data feature sets for the several building height prediction models, wherein the sample data features are the sample building outline data and sample building detail data of the sample building corresponding to the sample training data; Randomly select the first building height prediction model and obtain the first sample training set and the first sample data feature set; Based on the image spectral features and image texture features, feature association analysis is performed on multiple first sample training data in the first sample training set to calculate the average association degree of the first training data. Based on the building outline data and building detail data, feature association analysis is performed on multiple first sample data features in the first sample data feature set to calculate the average first data feature association degree. The first feature correlation degree is calculated based on the mean correlation degree of the first training data and the mean correlation degree of the first data features, and several feature correlation degrees are obtained by sequential analysis, including: Feature analysis is performed on multiple first sample training data in the first sample training set to construct multiple first feature vectors; Distribute the plurality of first feature vectors into a three-dimensional space and calculate the first distribution discreteness; The ratio of the first distribution dispersion to the preset standard distribution dispersion of the model training is set as the training weight adjustment coefficient, and multiplied by the initial training weight to obtain the adaptive training weight, wherein the initial training weight is 0.7, and the adaptive training weight is greater than or equal to 0.5 and less than or equal to 0.

9. The adaptive data feature weights are obtained by subtracting the adaptive training weights from 1. Based on the adaptive training weights and the adaptive data feature weights, the mean correlation of the first training data and the mean correlation of the first data features are weighted and calculated to obtain the first feature correlation. Based on the aforementioned feature correlation, a model integration strategy is set, and the historical building height prediction model library is called according to the model integration strategy to construct an adapted height prediction channel; The model integration strategy based on the aforementioned feature correlations includes: Select the minimum feature correlation degree among the several feature correlation degrees, and round down the ratio of the feature correlation degree to the minimum feature correlation degree to set it as the model selection number, thus obtaining several model selection numbers; A model integration strategy is constructed based on the number of models selected and the average fusion mechanism. Using the adaptive height prediction channel, the building height is predicted based on the image spectral features and image texture features, and the predicted building height is output. Based on the building outline data, building detail data, and predicted building height, a three-dimensional simulation model of the target building is performed to obtain the building's three-dimensional model.

2. The method for 3D building modeling combining deep learning and satellite imagery according to claim 1, characterized in that, According to the preset image processing scheme, the open-source satellite imagery is subjected to radiometric correction, geometric correction, resolution normalization, and image fusion enhancement to obtain standard satellite imagery.

3. The method for 3D building modeling combining deep learning and satellite imagery according to claim 1, characterized in that, Using a deep learning model, the contour and boundary features of the standard satellite imagery are extracted as building contour data, and building detail features are extracted as building detail data, including: Collect sample satellite image sets, perform contour and boundary annotations on buildings in different sample satellite image sets to obtain sample image contour label sets, and perform detailed feature annotations on buildings in different sample satellite image sets to obtain sample image detail label sets; Using the sample satellite image set as input and the sample image contour label set as supervision, a convolutional neural network is trained until convergence to obtain a building contour recognition model. Using the sample satellite image set as input and the sample image detail label set as supervision, a convolutional neural network is trained until convergence to obtain a building detail recognition model. The building outline recognition model and the building detail recognition model are used to extract features from the standard satellite imagery to obtain building outline data and building detail data, respectively.

4. A 3D building modeling system combining deep learning and satellite imagery, characterized in that: The system for implementing the building 3D modeling method combining deep learning and satellite imagery as described in any one of claims 1-3, the system comprising: The image acquisition module is used to acquire open-source satellite imagery of the target building and perform image preprocessing to obtain standard satellite imagery. The feature extraction module is used to extract the contour features and boundary features of the standard satellite image as building contour data and extract the building detail features as building detail data using a deep learning model. The prediction channel construction module is used to integrate and construct an adaptive height prediction channel based on the image spectral and texture features of the standard satellite image, as well as the building outline and building detail data, by calling the historical building height prediction model library. The height prediction module is used to predict building height based on the image spectral features and image texture features using the adapted height prediction channel, and output the predicted building height. The 3D modeling module is used to perform 3D simulation modeling of the target building based on the building outline data, building detail data and predicted building height, to obtain a 3D building model.

Images