AI building classification method based on multi-modal remote sensing and GIS data fusion

Abstract

The invention discloses an AI building classification method based on multi-mode remote sensing and GIS data fusion, and relates to the technical field of building classification. The whole process is highly integrated and automatically operated, the links and time cost of manual intervention are greatly reduced, the data processing efficiency is remarkably improved, and the method is particularly suitable for rapid building general survey and updating tasks in a large-scale urban area; the intelligent conversion from original multi-source data to a high-value spatial information product is successfully realized through a modular design and a cooperative operation mechanism, and efficient, reliable and sustainable technical support is provided for accurate city perception, dynamic monitoring and scientific decision making; and in the aspect of disaster response, a high-precision building distribution diagram can be quickly generated, the building vulnerability is evaluated in an auxiliary manner by combining three-dimensional structure characteristics, a high-risk area is accurately positioned, and reliable spatial intelligence support is provided for emergency resource allocation and evacuation planning.

Description

Technical Field

[0001] This invention relates to the field of building classification technology, specifically to an AI building classification method based on the fusion of multimodal remote sensing and GIS data. Background Technology

[0002] Buildings are mainly classified according to their purpose, number of floors / height, and scale, as follows: I. By Purpose: Civil buildings (residential buildings: houses, dormitories, apartments, etc., for daily living; public buildings: social service buildings such as schools, hospitals, libraries, and office buildings, for education, medical care, commerce, and transportation); industrial buildings: production workshops, warehouses, power rooms, and other industrial production-related buildings; agricultural buildings: greenhouses, livestock farms, agricultural machinery stations, and other agricultural production and processing facilities); II. By Number of Floors / Height: Low-rise / Multi-story buildings (residential: ≤27 meters; public buildings: ≤24 meters; single-story public buildings: <24 meters); High-rise buildings (residential: <27 meters; public buildings: <24 meters and ≤100 meters); Super high-rise buildings (residential / public buildings: <100 meters); III. By Scale: Large-scale buildings (closely related to daily life and numerous, such as houses, schools, hospitals, and shops); Large-scale buildings (large in scale and costly, such as large stadiums, airports, and museums).

[0003] The main function of building classification is to identify the building density or building type in different areas of a city. In the event of a disaster, it can help determine the distribution of severely affected areas and provide a basis for disaster relief decision-making.

[0004] Existing building classification methods mainly include traditional image processing techniques based on a single data source, statistical models, and early machine learning algorithms. These methods typically rely on optical remote sensing images, identifying building outlines through feature extraction (such as edge detection and texture analysis) and threshold segmentation, and then classifying them. For example, Support Vector Machines (SVM) or decision trees are used to distinguish building types, but these methods are susceptible to changes in lighting, occlusion, and noise, resulting in limited classification accuracy. Furthermore, some methods attempt to combine GIS data (such as land use maps or topographic information), but the data fusion mechanism is relatively simple and struggles to effectively handle the spatiotemporal inconsistencies between multimodal data. In disaster response scenarios, existing technologies often cannot update classification results in real time and have poor adaptability to complex urban environments (such as high-rise building clusters), limiting their application in emergency decision-making. Therefore, we propose an AI building classification method based on the fusion of multimodal remote sensing and GIS data. Summary of the Invention

[0005] The purpose of this invention is to provide an AI building classification method based on the fusion of multimodal remote sensing and GIS data to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention specifically adopts the following technical solution:

[0007] An AI-based building classification method based on the fusion of multimodal remote sensing and GIS data, the method comprising:

[0008] The multimodal data acquisition and preprocessing module is used to acquire multi-source remote sensing image data covering the target area, simultaneously acquire GIS vector data of the target area, and preprocess the collected data.

[0009] The feature engineering and fusion module is used to extract features from remote sensing images and GIS data, and to perform multimodal feature fusion on the data after feature extraction.

[0010] The AI ​​classification model building and training module uses a deep neural network model as the core classifier to build a classification model from the data after feature fusion, and then trains the built model.

[0011] The intelligent building classification and output module is used to input preprocessed and feature-fused target area data into a trained AI classification model. The model infers from the input data and outputs a fine-grained classification probability vector for each building unit, determines the final type of each building, and generates a thematic map and structured database containing the building's spatial location, outline, and detailed classification information.

[0012] The model update and feedback module is used to establish a continuous evaluation system for model performance, and to evaluate model accuracy using newly acquired data or expert validation results.

[0013] Furthermore, the multi-source remote sensing image data acquired by the multimodal data acquisition and preprocessing module includes, but is not limited to: high-resolution optical images, synthetic aperture radar images, lidar point cloud data and their derived digital surface models and digital elevation models.

[0014] Furthermore, the GIS vector data acquired by the multimodal data acquisition and preprocessing module includes: building outlines, land use types, road networks, administrative divisions, and points of interest information.

[0015] Furthermore, the preprocessing in the multimodal data acquisition and preprocessing module includes: radiometric calibration, atmospheric correction, geometric fine correction, image registration, point cloud filtering and classification, vector data topology checking and repair, and coordinate system transformation.

[0016] Furthermore, the remote sensing image feature extraction in the feature engineering and fusion module includes: extracting spectral features, texture features, morphological features, and target segmentation contours from optical images; extracting backscattering intensity, interferometric coherence, and polarization decomposition features from synthetic aperture radar images; and extracting geometric and structural features such as height, volume, shape, slope, and roughness from lidar point cloud data and its derived digital surface models.

[0017] Furthermore, the GIS data features in the feature engineering and fusion module include: extracting the building's own attributes, spatial relationship features, and contextual features based on GIS vector data.

[0018] Furthermore, the multimodal feature fusion in the feature engineering and fusion module adopts adaptive weighted fusion, feature concatenation, graph neural network-based association modeling, or attention mechanism-based feature selection and fusion strategies to deeply fuse heterogeneous features extracted from different modalities and construct a joint feature vector that can comprehensively characterize the physical attributes, spatial structure, and environmental context of a building.

[0019] Furthermore, the AI ​​classification model building and training module uses a historical sample dataset with accurate building type labels to train the model, and the samples need to cover various typical building types and complex scenes within the target area.

[0020] Furthermore, the model update and feedback module, based on an online learning strategy, allows the model to update its parameters without forgetting old knowledge after acquiring new labeled samples, in order to adapt to changes in urban development and data sources and maintain the timeliness and accuracy of classification results.

[0021] Furthermore, it includes the following steps:

[0022] Step 1: Multimodal Data Acquisition and Preprocessing: Acquire multi-source remote sensing image data covering the target area, including high-resolution optical images, synthetic aperture radar images, lidar point cloud data, and their derived digital surface models and digital elevation models; simultaneously acquire GIS vector data of the target area, including building outlines, land use types, road networks, administrative divisions, and points of interest information; preprocess the collected data, including radiometric calibration, atmospheric correction, geometric fine correction, image registration, point cloud filtering and classification, vector data topology checking and repair, and coordinate system transformation, to ensure data quality and consistency.

[0023] Step 2, Feature Engineering and Fusion: Features are extracted from the preprocessed remote sensing images, including spectral features, texture features, morphological features, and target segmentation contours from optical images; backscattering intensity, interferometric coherence, and polarization decomposition features from synthetic aperture radar images; and geometric and structural features such as height, volume, shape, slope, and roughness from lidar point cloud data and its derived digital surface models. Simultaneously, building attributes, spatial relationship features, and contextual features are extracted from GIS vector data. Adaptive weighted fusion, feature concatenation, graph neural network-based association modeling, or attention-based feature selection and fusion strategies are employed to deeply fuse heterogeneous features from different modalities, constructing a joint feature vector that comprehensively represents the physical attributes, spatial structure, and environmental context of buildings.

[0024] Step 3, AI classification model construction and training: Based on a deep neural network model as the core classifier, the model is trained using a historical sample dataset with accurate building type labels. The samples cover various typical building types and complex scenes within the target area. The training process includes model architecture design, parameter initialization, loss function optimization, and iterative training to build a high-precision building classification model.

[0025] Step 4, Intelligent Classification and Output of Buildings: Input the preprocessed and feature-fused target area data into the trained AI classification model. The model infers from the input data, calculates the fine-grained classification probability vector for each building unit, and determines the final type of each building based on probability thresholds or decision rules. Generate a thematic map and structured database containing the spatial location, outline, and detailed classification information of buildings, and output it in an interactive geographic information system format.

[0026] Step 5, Model Update and Feedback: Establish a continuous model performance evaluation system, and use newly acquired data or expert validation results to evaluate the accuracy and analyze errors of the model; based on the online learning strategy, after acquiring new labeled samples, incrementally update the model parameters to avoid forgetting old knowledge, adapt to changes in the urban environment and data source updates, and ensure the timeliness and accuracy of classification results; regularly generate model performance reports and feed them back to the system for subsequent optimization.

[0027] The beneficial effects of this invention are as follows:

[0028] 1. This invention integrates and automates the entire process from automatic collection and standardized preprocessing of multi-source heterogeneous data to feature fusion, model inference, and result output. This not only significantly reduces the number of manual intervention steps and time costs, but also significantly improves data processing efficiency. It is especially suitable for rapid building surveys and updates in large urban areas.

[0029] 2. This invention integrates the rich texture details of high-resolution optical images, the all-weather imaging capability of radar data, the precise three-dimensional structural information of lidar, and the spatial relationships and semantic attributes contained in GIS vector data to systematically construct a multi-dimensional, multi-scale joint feature representation. This complementarity effectively overcomes the limitations of a single data source and greatly improves the accuracy of building identification and fine-grained classification in complex urban environments. It is particularly advantageous in distinguishing between building types with similar forms but different functions (such as high-rise residential buildings and office buildings).

[0030] 3. The dynamic optimization mechanism of the model update and feedback module in this invention enables the system to continuously adapt to changes in data distribution brought about by urban development (such as new areas and building renovations) and the introduction of new remote sensing data sources. Through online learning and expert feedback loops, model parameters are iteratively optimized, effectively alleviating data drift problems, ensuring that classification performance remains stable and reliable in long-term operation, and reducing the risk of misclassification caused by environmental changes.

[0031] 4. The modular design and collaborative operation mechanism of this invention successfully realizes the intelligent transformation from raw multi-source data to high-value spatial information products, providing efficient, reliable and sustainable technical support for precise urban perception, dynamic monitoring and scientific decision-making.

[0032] 5. In terms of disaster response, this invention can quickly generate high-precision building distribution maps, combine three-dimensional structural features to assist in assessing building vulnerability, accurately locate high-risk areas, and provide reliable spatial intelligence support for emergency resource allocation and evacuation planning. Attached Figure Description

[0033] Figure 1 This is a block diagram of the present invention;

[0034] Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0036] Please see Figure 1 - Figure 2 This invention provides an AI building classification method based on the fusion of multimodal remote sensing and GIS data, the method comprising:

[0037] The multimodal data acquisition and preprocessing module is used to acquire multi-source remote sensing image data covering the target area, simultaneously acquire GIS vector data of the target area, and preprocess the collected data; the preprocessed data has standard format and high quality characteristics, which facilitates subsequent processing.

[0038] The feature engineering and fusion module is used to extract features from remote sensing images and GIS data, and to perform multimodal feature fusion on the extracted data; the data after multimodal feature fusion has a unified feature expression form that is compatible with spatiotemporal characteristics.

[0039] The AI ​​classification model building and training module uses a deep neural network model as the core classifier to build a classification model from the data after feature fusion, and then trains the built model. The trained model can efficiently identify and classify urban buildings and output accurate classification result images.

[0040] The intelligent building classification and output module is used to input pre-processed and feature-fused target area data into a trained AI classification model. The model infers from the input data and outputs a fine-grained classification probability vector for each building unit, determining the final type of each building and generating a thematic map and structured database containing the building's spatial location, outline, and detailed classification information. This operation enables automated and high-precision building classification tasks, greatly improving data processing efficiency and recognition accuracy. The generated thematic maps and structured databases can be directly integrated into a GIS platform, identifying building density or building types in different urban areas. In the event of a disaster, it can assist in determining the distribution of severely affected areas, providing a basis for disaster relief decision-making.

[0041] The model update and feedback module is used to establish a continuous evaluation system for model performance, using newly acquired data or expert validation results to evaluate model accuracy. Based on the accuracy evaluation results, this module can automatically optimize model parameters or trigger a retraining process, ensuring the classification system continuously adapts to changes in the environment and data distribution shifts, thereby significantly reducing the misclassification rate and enhancing the model's generalization ability. Simultaneously, this module supports a feedback loop mechanism, allowing expert knowledge to be integrated into the model iteration process, improving the robustness and reliability of long-term applications, ultimately maintaining high-precision building classification performance, and providing more stable and reliable data support for scenarios such as urban planning and disaster response.

[0042] In this embodiment, preferably, the multi-source remote sensing image data acquired by the multimodal data acquisition and preprocessing module includes, but is not limited to: high-resolution optical images, synthetic aperture radar images, lidar point cloud data and their derived digital surface models and digital elevation models; by acquiring high-resolution optical images, the surface texture, color and geometric details of buildings can be clearly captured, especially providing rich visual information in urban areas; synthetic aperture radar images have the ability to penetrate clouds and image in all weather conditions, and are suitable for data acquisition under different meteorological conditions; lidar point cloud data and their derived digital surface models and digital elevation models accurately depict the three-dimensional structural features of buildings, such as height, slope and volume information.

[0043] In this embodiment, preferably, the GIS vector data acquired by the multimodal data acquisition and preprocessing module includes: building outlines, land use types, road networks, administrative divisions, and points of interest (POIs). By collecting building outline data, the spatial location and base morphology of buildings can be accurately located, providing a basic geographic framework for subsequent classification. Land use type information helps to understand the functional zoning background of buildings and assists in distinguishing different categories such as residential, commercial, and industrial. Road network data can reveal the correlation between buildings and transportation facilities, reflecting their accessibility and spatial distribution patterns. Administrative division information provides management unit boundaries for building classification and supports regional statistics and analysis. POI information supplements the functional attributes of buildings, such as the labeling of specific places like schools, hospitals, and shopping malls, significantly improving the semantic richness of fine-grained classification.

[0044] In this embodiment, the preferred preprocessing steps in the multimodal data acquisition and preprocessing module include: radiometric calibration, atmospheric correction, geometric fine correction, image registration, point cloud filtering and classification, vector data topology checking and repair, and coordinate system transformation. Radiometric calibration eliminates inconsistencies in sensor responses, ensuring that the radiometric values ​​in the image data accurately reflect the true reflectivity or radiance of ground features. Atmospheric correction removes the effects of atmospheric scattering and absorption, restoring the true reflectivity of the ground surface. Geometric fine correction eliminates geometric distortions in the image through control point matching, ensuring spatial accuracy. Image registration ensures spatial alignment of images from different sources or time phases, facilitating subsequent multimodal data fusion. Point cloud filtering and classification remove noise points and identify ground feature categories, improving the quality of point cloud data. Vector data topology checking and repair corrects geometric errors and maintains the integrity of data topology relationships. Coordinate system transformation unifies all data to the same reference system, achieving spatial consistency. The synergistic effect of these preprocessing steps greatly improves the reliability and comparability of multi-source data, providing a high-quality input foundation for the feature engineering module.

[0045] In this embodiment, preferably, the remote sensing image feature extraction in the feature engineering and fusion module includes: extracting spectral features, texture features, morphological features, and target segmentation contours from optical images; extracting backscattering intensity, interferometric coherence, and polarization decomposition features from synthetic aperture radar images; and extracting geometric and structural features such as height, volume, shape, slope, and roughness from lidar point cloud data and its derived digital surface models. Through the extraction of the above image features, the physical attributes, structural details, and functional differences of buildings can be comprehensively characterized, thereby laying a data foundation for multimodal feature fusion.

[0046] In this embodiment, preferably, the GIS data features in the feature engineering and fusion module include: extracting the building's own attributes, spatial relationship features, and contextual features based on GIS vector data; through the above operations, a multi-dimensional feature system that combines individual attributes and group distribution patterns can be constructed, fully exploring the spatial correlation and semantic information contained in the GIS data, and thus forming a strong complement to the physical features extracted from remote sensing images, jointly supporting the expression of highly discriminative fusion features.

[0047] In this embodiment, preferably, the multimodal feature fusion in the feature engineering and fusion module adopts adaptive weighted fusion, feature concatenation, graph neural network-based association modeling, or attention-based feature selection and fusion strategies to deeply fuse heterogeneous features extracted from different modalities, constructing a joint feature vector that can comprehensively characterize the physical attributes, spatial structure, and environmental context of buildings. This greatly enhances the discriminative ability of the classification model, effectively captures subtle differences in buildings in complex urban environments, and improves the recognition accuracy of different functional categories. Through adaptive weight allocation or attention mechanisms, the spatial consistency and semantic relevance of feature representations are optimized, reducing inconsistencies between modalities and ensuring that the joint feature vector efficiently supports the inference process in subsequent AI classification models, ultimately achieving more robust and accurate building classification results.

[0048] In this embodiment, preferably, the AI ​​classification model construction and training module uses a historical sample dataset with accurate building type labels to train the model, and the samples need to cover various typical building types and complex scenes in the target area; thereby effectively improving the model's generalization ability and robustness, and ensuring accurate identification of various types of buildings in a changing urban environment.

[0049] In this embodiment, preferably, the model update and feedback module, based on an online learning strategy, allows the model to update its parameters without forgetting old knowledge after acquiring new labeled samples. This adapts to changes in urban development and data sources, maintaining the timeliness and accuracy of classification results. Furthermore, it enables real-time monitoring of data distribution changes after deployment, automatically triggering an incremental learning mechanism to efficiently integrate new samples with historical knowledge, avoiding catastrophic forgetting problems, ensuring high-precision classification capabilities throughout long-term urban development, and providing real-time and reliable data services and decision support for applications such as smart city governance and disaster risk assessment.

[0050] The coordinated operation of the above modules enables an end-to-end intelligent building classification process, significantly improving the automation level and accuracy of urban information extraction. Specifically, the collaborative advantages of this system are reflected in the following key aspects:

[0051] Specifically, this manifests in several ways: from the automatic acquisition and standardized preprocessing of multi-source heterogeneous data to feature fusion, model inference, and result output, the entire process is highly integrated and automated. This significantly reduces the need for manual intervention and time costs, while also substantially improving data processing efficiency, making it particularly suitable for rapid building surveys and updates in large urban areas. By deeply fusing the rich texture details of high-resolution optical images, the all-weather imaging capabilities of radar data, the precise three-dimensional structural information of lidar, and the spatial relationships and semantic attributes inherent in GIS vector data, the system constructs a multi-dimensional, multi-scale joint feature representation. This complementarity effectively overcomes the limitations of a single data source, greatly improving the accuracy of building identification and fine-grained classification in complex urban environments, especially when distinguishing between building types with similar forms but different functions (such as high-rise residential buildings and office buildings). The dynamic optimization mechanism of the model update and feedback module enables the system to continuously adapt to changes in data distribution brought about by urban development (such as newly built areas and building renovations) and the introduction of new remote sensing data sources. Through online learning and expert feedback loops, model parameters are iteratively optimized, effectively mitigating data drift problems and ensuring accurate data transmission. The system ensures stable and reliable classification performance over long-term operation, reducing the risk of misclassification due to environmental changes. The output structured database and thematic maps strictly adhere to GIS platform standard formats, supporting seamless integration into business systems such as urban planning, disaster emergency response, and smart city management. The spatial location, outline, type, and probability information of buildings included not only provides accurate base maps for macro-level analysis (such as building density distribution and functional zoning) but also supports refined decision-making at the micro-level. For example, during disasters, the system can quickly identify the spatial clustering of buildings of different structural types, combine their three-dimensional features to assist in assessing structural vulnerability, accurately locate high-risk areas, and provide crucial spatial intelligence support for emergency resource allocation and evacuation route planning. The feature engineering and fusion modules explicitly extract and fuse physical attributes (spectral, texture, geometry) and spatial semantic relationships (adjacent roads, functional area background), and expert knowledge is integrated into model iteration through a feedback mechanism, enhancing the interpretability of classification results. Combined with understandable physical features and spatial logic, the classification results are more easily understood and trusted by professional users such as planners and emergency managers, improving the usability and adoption rate of the results in practical business.

[0052] In summary, the modular design and collaborative operation mechanism of this method successfully realizes the intelligent transformation from raw multi-source data to high-value spatial information products, providing efficient, reliable and sustainable technical support for precise urban perception, dynamic monitoring and scientific decision-making.

[0053] An AI-based building classification method based on the fusion of multimodal remote sensing and GIS data includes the following steps:

[0054] Step 1: Multimodal Data Acquisition and Preprocessing: Acquire multi-source remote sensing image data covering the target area, including high-resolution optical images, synthetic aperture radar images, lidar point cloud data, and their derived digital surface models and digital elevation models; simultaneously acquire GIS vector data of the target area, including building outlines, land use types, road networks, administrative divisions, and points of interest information; preprocess the collected data, including radiometric calibration, atmospheric correction, geometric fine correction, image registration, point cloud filtering and classification, vector data topology checking and repair, and coordinate system transformation, to ensure data quality and consistency.

[0055] Step 2, Feature Engineering and Fusion: Features are extracted from the preprocessed remote sensing images, including spectral features, texture features, morphological features, and target segmentation contours from optical images; backscattering intensity, interferometric coherence, and polarization decomposition features from synthetic aperture radar images; and geometric and structural features such as height, volume, shape, slope, and roughness from lidar point cloud data and its derived digital surface models. Simultaneously, building attributes, spatial relationship features, and contextual features are extracted from GIS vector data. Adaptive weighted fusion, feature concatenation, graph neural network-based association modeling, or attention-based feature selection and fusion strategies are employed to deeply fuse heterogeneous features from different modalities, constructing a joint feature vector that comprehensively represents the physical attributes, spatial structure, and environmental context of buildings.

[0056] Step 3, AI classification model construction and training: Based on a deep neural network model as the core classifier, the model is trained using a historical sample dataset with accurate building type labels. The samples cover various typical building types and complex scenes within the target area. The training process includes model architecture design, parameter initialization, loss function optimization, and iterative training to build a high-precision building classification model.

[0057] Step 4, Intelligent Classification and Output of Buildings: Input the preprocessed and feature-fused target area data into the trained AI classification model. The model infers from the input data, calculates the fine-grained classification probability vector for each building unit, and determines the final type of each building based on probability thresholds or decision rules. Generate a thematic map and structured database containing the spatial location, outline, and detailed classification information of buildings, and output it in an interactive geographic information system format.

[0058] Step 5, Model Update and Feedback: Establish a continuous model performance evaluation system, and use newly acquired data or expert validation results to evaluate the accuracy and analyze errors of the model; based on the online learning strategy, after acquiring new labeled samples, incrementally update the model parameters to avoid forgetting old knowledge, adapt to changes in the urban environment and data source updates, and ensure the timeliness and accuracy of classification results; regularly generate model performance reports and feed them back to the system for subsequent optimization.

[0059] The above-described methods significantly improve the overall accuracy and efficiency of building classification, enabling end-to-end intelligent processing and playing a crucial role in multiple application scenarios. These steps automatically integrate multi-source data, reducing manual intervention and data processing time, ensuring the robustness and adaptability of the classification model in complex urban environments. Simultaneously, they support real-time monitoring and data updates, effectively addressing data distribution drift caused by urban development and environmental changes, and maintaining long-term stability of classification performance. In disaster response, these methods can rapidly generate high-precision building distribution maps, combining 3D structural features to assist in assessing building vulnerability, accurately locating high-risk areas, and providing reliable spatial intelligence support for emergency resource allocation and evacuation planning.

[0060] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An AI building classification method based on multi-modal remote sensing and GIS data fusion, characterized in that, The method comprises: a multi-modal data acquisition and preprocessing module for acquiring multi-source remote sensing image data covering a target area, synchronously acquiring GIS vector data of the target area, and preprocessing the collected data; a feature engineering and fusion module for extracting remote sensing image and GIS data features, and performing multi-modal feature fusion on the data after feature extraction; an AI classification model construction and training module for constructing a classification model on the data after feature fusion by taking a deep neural network model as a core classifier, and training the constructed model; a building intelligent classification and output module for inputting the target area data after preprocessing and feature fusion into the trained AI classification model, performing inference on the input data by the model, outputting a fine-grained classification probability vector of each building unit, determining the final type of each building, and generating a thematic map and a structured database containing building spatial position, contour, and detailed classification information; a model updating and feedback module for establishing a model performance continuous evaluation system, and performing model precision evaluation by using newly acquired data or expert verification results.

2. The AI building classification method based on multi-modal remote sensing and GIS data fusion according to claim 1, characterized in that, The multi-source remote sensing image data acquired by the multi-modal data acquisition and preprocessing module includes but is not limited to high-resolution optical images, synthetic aperture radar images, laser radar point cloud data, and derived digital surface models and digital elevation models. 3.The AI building classification method based on multi-modal remote sensing and GIS data fusion of claim 1, wherein, The GIS vector data acquired by the multi-modal data acquisition and preprocessing module includes building contours, land use types, road networks, administrative divisions, and point of interest information.

4. The AI building classification method based on multi-modal remote sensing and GIS data fusion according to claim 1, characterized in that, The preprocessing in the multi-modal data acquisition and preprocessing module includes radiation calibration, atmospheric correction, geometric precision correction, image registration, point cloud filtering and classification, vector data topology inspection and repair, and coordinate system conversion.

5. The AI building classification method based on multi-modal remote sensing and GIS data fusion according to claim 1, characterized in that, The remote sensing image feature extraction in the feature engineering and fusion module includes extracting spectral features, texture features, morphological features, and target segmentation contours from optical images, extracting backscattering intensity, interference coherence, and polarization decomposition features from synthetic aperture radar images, and extracting height, volume, shape, slope, and roughness from laser radar point cloud data and derived digital surface models.

6. The AI building classification method based on multi-modal remote sensing and GIS data fusion according to claim 1, characterized in that, The GIS data features in the feature engineering and fusion module include extracting building itself attributes, spatial relationship features, and context features based on GIS vector data.

7. The AI building classification method based on multi-modal remote sensing and GIS data fusion according to claim 1, characterized in that, The multi-modal feature fusion in the feature engineering and fusion module adopts adaptive weighted fusion, feature concatenation, graph neural network-based association modeling, or attention mechanism-based feature selection and fusion strategies to deeply fuse heterogeneous features from different modalities, and construct a joint feature vector capable of comprehensively representing building physical attributes, spatial structure, and environmental context.

8. The AI building classification method based on multi-modal remote sensing and GIS data fusion according to claim 1, characterized in that, The AI classification model construction and training module trains the model by using a historical sample data set with accurate building type labels, and the samples need to cover various typical building types and complex scenes in the target area.

9. The AI building classification method based on multi-modal remote sensing and GIS data fusion according to claim 1, characterized in that, The model updating and feedback module allows the model to update parameters without forgetting old knowledge after obtaining new labeled samples, to adapt to urban development and changes in data sources, and to maintain the timeliness and accuracy of classification results based on an online learning strategy.

10. The AI building classification method based on multi-modal remote sensing and GIS data fusion according to claim 1, characterized in that, The method comprises the following steps: Step 1, multi-modal data acquisition and preprocessing: obtaining multi-source remote sensing image data covering the target area, including high-resolution optical images, synthetic aperture radar images, laser radar point cloud data, and derived digital surface model and digital elevation model; synchronously obtaining GIS vector data of the target area, including building outlines, land use types, road networks, administrative divisions and point of interest information; preprocessing the collected data, including radiation calibration, atmospheric correction, geometric precision correction, image registration, point cloud filtering and classification, vector data topology checking and repair, coordinate system conversion, to ensure data quality and consistency; Step 2, feature engineering and fusion: extracting features from the preprocessed remote sensing images, including extracting spectral features, texture features, morphological features and target segmentation contours from optical images, extracting backscattering intensity, interference coherence and polarization decomposition features from synthetic aperture radar images, and extracting height, volume, shape, slope and roughness from laser radar point cloud data and derived digital surface model; at the same time, extracting building attributes, spatial relationship features and context features from GIS vector data; Adopting adaptive weighted fusion, feature concatenation, graph neural network-based association modeling or attention mechanism-based feature selection and fusion strategy, heterogeneous features from different modalities are deeply fused to construct a joint feature vector that can comprehensively represent the physical properties, spatial structure and environmental context of buildings; Step 3, AI classification model construction and training: based on a deep neural network model as the core classifier, the model is trained using a historical sample dataset with accurate building type labels, covering various typical building types and complex scenes in the target area; The training process includes model architecture design, parameter initialization, loss function optimization and iterative training to build a high-precision building classification model; Step 4, building intelligent classification and output: input the preprocessed and feature-fused target area data into the trained AI classification model, the model infers the input data, calculates the fine-grained classification probability vector of each building unit, and determines the final type of each building based on the probability threshold or decision rule; Generate a thematic map and structured database containing building spatial location, outline and detailed classification information, output in an interactive geographic information system format; Step 5, model updating and feedback: Establish a model performance continuous evaluation system, use newly acquired data or expert verification results to evaluate the accuracy of the model and analyze the errors; based on the online learning strategy, after obtaining new labeled samples, incrementally update the model parameters to avoid forgetting old knowledge, adapt to changes in urban environment and data source updates, and ensure the timeliness and accuracy of the classification results; generate model performance reports regularly and feed back to the system for subsequent optimization.

Images