House earthquake hidden danger deep learning identification and grade evaluation method based on multi-scale remote sensing image fusion
By combining multi-scale remote sensing image fusion with deep learning, the problems of low efficiency and inaccurate results in the assessment of building earthquake hazards have been solved, realizing fully automatic and accurate hazard identification and assessment, and improving the coverage and reliability of the assessment.
Patent Information
- Application Number
- CN202511758625.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-27
- Publication Date
- 2026-01-09
AI Technical Summary
Existing technologies for assessing seismic risks in buildings suffer from problems such as low efficiency of manual on-site inspections, strong subjectivity in assessment results, and insufficient information from single-scale remote sensing images, leading to inaccurate assessment results and difficulty in large-scale and rapid screening.
A method combining multi-scale remote sensing image fusion and deep learning was adopted. Multi-source remote sensing data was acquired and preprocessed, a multi-branch deep learning model was constructed, multi-scale features were extracted and adaptively weighted fusion was performed, and an ensemble learning classifier was used to assess the level of earthquake hazards and generate a thematic map of earthquake hazard levels.
It enables fully automated, accurate identification and scientific assessment of potential earthquake hazards in buildings, improving the coverage, efficiency and reliability of assessment results, and providing a basis for rapid screening and disaster prevention and mitigation decisions.
Smart Images
Figure CN121305307A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of remote sensing science and technology, specifically a deep learning method for identifying and assessing the level of seismic hazards in buildings based on multi-scale remote sensing image fusion. Background Technology
[0002] Multi-scale remote sensing image fusion refers to the technical process of collaboratively utilizing remote sensing image data from the same region but with different spatial resolutions, spectral characteristics, and temporal phases, and integrating the complementary information of each source data through a specific algorithm model to generate a new image with greater application value. Its core principle lies in breaking through the information limitations of single-scale data by fusing medium- and low-resolution images that contain macroscopic layout context information and high-resolution images that preserve fine ground structure textures, thereby providing a more complete and feature-rich data foundation for subsequent deep feature extraction and accurate target identification.
[0003] Currently, earthquake hazard assessment of buildings mainly relies on two traditional technical approaches: one is the assessment method based on manual on-site inspection, which requires professional technicians to conduct on-site visits with simple measuring tools, visually observing the structural type, material condition, crack distribution, foundation settlement, and other superficial characteristics of the buildings, and making preliminary judgments based on personal experience; the other is the visual interpretation method based on single-scale remote sensing imagery, which involves manually delineating and performing simple quantitative analysis by comparing pre-earthquake low-to-medium resolution satellite images. However, these two traditional technical approaches have significant technical drawbacks: First, manual on-site inspection requires a large number of professional personnel, resulting in extremely low work efficiency. A single technician can only complete a limited number of building assessments per day, which cannot meet the timeliness requirements for rapid screening of large areas; second, manual judgment is highly subjective, and differences in the experience of different technicians can lead to inconsistent assessment standards, seriously affecting the comparability and reliability of the assessment results; third, the method based on single-scale remote sensing imagery is limited by the completeness of the data source information and cannot simultaneously acquire macroscopic environmental data. The information and microstructural characteristics make it difficult to guarantee the accuracy of assessment results. More seriously, these technical defects will bring a series of chain problems in practical applications: when faced with large-scale regional assessment tasks, traditional methods cannot complete a comprehensive survey under limited time and manpower conditions, inevitably missing a large number of potential risk points; the lack of unified quantitative assessment standards makes it difficult to effectively compare and analyze assessment results from different regions; the low efficiency and high cost of the assessment process seriously restrict the routine implementation of geological disaster prevention and control work. These problems directly affect the accuracy and reliability of regional earthquake risk assessment, posing serious hidden dangers to people's lives and property. In recent years, with the development of remote sensing technology, some studies have attempted to apply machine learning methods to building identification, but most of these methods are limited to using a single data source or simple feature combination, failing to effectively solve the problems of multi-scale information fusion and deep feature mining. Especially in the professional field of building earthquake hazard assessment, existing technologies have not yet established a complete technical chain from multi-source data fusion to hazard level assessment. Summary of the Invention
[0004] The purpose of this invention is to provide a deep learning-based method for identifying and assessing the level of seismic hazards in buildings based on multi-scale remote sensing image fusion, in order to solve the problems mentioned in the background art.
[0005] To achieve the above objectives, the present invention provides the following technical solution: a deep learning method for identifying and assessing the level of seismic hazards in buildings based on multi-scale remote sensing image fusion, comprising the following steps:
[0006] Step 1: Data Acquisition and Preprocessing
[0007] Acquire multi-source remote sensing image data of the target area, which includes at least high-resolution optical images, lidar point cloud data, and synthetic aperture radar data; perform radiometric calibration and atmospheric correction on the optical images, filter and classify the lidar point cloud data to generate a digital surface model, and perform geometric fine registration on all multi-source data to form a multi-source data layer with unified spatial reference.
[0008] Step 2: Multi-scale feature fusion modeling
[0009] A multi-branch deep learning model is constructed, which includes a feature extraction branch for extracting macro-contextual features and micro-structural features from the multi-source data layers of different scales or sources, a feature fusion module for adaptively weighted fusion of the multi-scale features, and a prediction head for multi-task learning based on the fused feature map.
[0010] Step 3: Hazard Identification and Feature Quantification
[0011] Using the multi-task learning prediction head, based on the fused features, the accurate segmentation of building objects is completed simultaneously, and the building binary mask and the intrinsic features related to earthquake hazards are quantified, outputting a quantized feature vector including structural irregularity index, roof material type, building height and building spacing.
[0012] Step 4: Level Assessment
[0013] Based on the quantized intrinsic feature vectors, and using a pre-defined evaluation rule, an integrated machine learning classifier is used to assess the seismic hazard level of each segmented building object and output discrete risk level labels.
[0014] Step 5: Output Results
[0015] Generate and output thematic maps of building seismic hazard levels, in which the spatial outline of each building object is associated with its assessed hazard level and quantitative feature vector, and stored in the standard format of geographic information system.
[0016] Preferably, the multi-source remote sensing image data includes sub-meter high-resolution optical images, lidar point cloud data, and synthetic aperture radar data.
[0017] Preferably, the preprocessing includes radiometric calibration and atmospheric correction of the optical image, filtering and classification of the lidar point cloud to generate a digital surface model, and strict geometric registration of all data sources.
[0018] Preferably, the multi-scale feature fusion includes feature extraction and fusion at the macro, meso, and micro scales, where the macro scale corresponds to street-level features, the meso scale corresponds to building-level features, and the micro scale corresponds to component-level features.
[0019] Preferably, the multi-scale feature fusion deep learning model adopts a multi-branch neural network architecture, including a feature extraction branch, a feature fusion module, and a multi-task learning head.
[0020] Preferably, the feature fusion module uses a channel attention mechanism to achieve adaptive feature weighted fusion, and the multi-task learning head simultaneously outputs a house segmentation mask and multiple quantified hazard feature values.
[0021] Preferably, the quantitative hazard characteristics include a combination of at least three characteristics among structural irregularity index, roof material type, building height, and building spacing.
[0022] Preferably, the preset evaluation rule adopts an ensemble learning method based on random forest, which uses quantified hazard features as input features and outputs discrete hazard levels.
[0023] Preferably, the thematic map of hazard levels adopts the standard format of a geographic information system, and each building patch is associated with an attribute table, recording its location coordinates, quantitative feature values, and hazard level.
[0024] The beneficial effects of this invention are as follows:
[0025] By organically combining multi-scale remote sensing image fusion with deep learning technology, fully automated and accurate extraction and quantitative assessment of seismic hazards in buildings, from macroscopic layout to microscopic structural features, has been achieved. Its beneficial effects lie in constructing an assessment system with wide coverage, high identification efficiency, and objective and reliable assessment results. This significantly improves the comprehensiveness of hazard identification and the scientific nature of assessment results, providing efficient technical means and reliable decision-making basis for rapid screening, precise prevention and control of seismic risks in buildings over a wide area, as well as the optimal allocation of disaster prevention and mitigation resources. Attached Figure Description
[0026] Figure 1 This is a flowchart of the present invention. Detailed Implementation
[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0028] like Figure 1As shown, this embodiment of the invention provides a deep learning-based method for identifying and assessing the level of seismic hazards in buildings based on multi-scale remote sensing image fusion, including the following steps:
[0029] Step 1: Data Acquisition and Preprocessing
[0030] Acquire multi-source remote sensing image data of the target area, which includes at least high-resolution optical images, lidar point cloud data, and synthetic aperture radar data; perform radiometric calibration and atmospheric correction on the optical images, filter and classify the lidar point cloud data to generate a digital surface model, and perform geometric fine registration on all multi-source data to form a multi-source data layer with unified spatial reference.
[0031] Step 2: Multi-scale feature fusion modeling
[0032] Construct a multi-branch deep learning model, which includes a feature extraction branch for extracting macro-contextual features and micro-structural features from multi-source data layers of different scales or sources, a feature fusion module for adaptively weighted fusion of multi-scale features, and a prediction head for multi-task learning based on the fused feature map.
[0033] Step 3: Hazard Identification and Feature Quantification
[0034] Using a multi-task learning prediction head, based on the fused features, the system simultaneously completes accurate segmentation of building objects, outputs a binary mask of the building, and quantifies the intrinsic features related to earthquake hazards, outputting a quantized feature vector including structural irregularity index, roof material type, building height, and building spacing.
[0035] Step 4: Level Assessment
[0036] Based on the quantized intrinsic feature vectors, and using a pre-defined evaluation rule, an integrated machine learning classifier is used to assess the seismic hazard level of each segmented building object and output discrete risk level labels.
[0037] Step 5: Output Results
[0038] Generate and output thematic maps of building seismic hazard levels, in which the spatial outline of each building object is associated with its assessed hazard level and quantitative feature vector, and stored in the standard format of geographic information system.
[0039] This paper presents an end-to-end automated technical process that integrates multi-source data acquisition with thematic map output of hazard levels. This process enables accurate identification and scientific assessment of seismic hazards in buildings. The core principle lies in the organic combination of multi-scale remote sensing image fusion technology and deep learning models. First, a multi-source data layer with a unified spatial benchmark is constructed through rigorous geometric registration, radiometric correction, and point cloud processing to eliminate data heterogeneity and ensure the accuracy of information fusion. Then, a multi-branch deep learning architecture is used to extract macroscopic features reflecting the regional building layout and microscopic features representing the details of building structures. Next, a feature fusion module achieves adaptive weighted fusion of cross-scale information to form a complete feature representation. Subsequently, based on a multi-task learning mechanism, accurate segmentation of building objects and quantitative extraction of multi-dimensional intrinsic features such as structural irregularities, roof materials, building height, and spacing are completed simultaneously. Finally, an ensemble learning classifier maps the quantitative features to objective hazard levels and generates a thematic map with clear spatial location and attribute information. This complete technical chain effectively solves the inherent defects of traditional methods, such as reliance on human experience, inconsistent assessment standards, and difficulty in achieving large-scale rapid surveys. It significantly improves the coverage, assessment efficiency, and reliability of hazard identification.
[0040] The multi-source remote sensing image data includes sub-meter high-resolution optical images, lidar point cloud data, and synthetic aperture radar data.
[0041] Sub-meter level optical imaging provides rich texture details, lidar data provides accurate three-dimensional geometric information, and synthetic aperture radar data provides deformation monitoring capabilities. By complementing the advantages of multiple data sources, an information foundation that comprehensively reflects the structural characteristics of a building is constructed.
[0042] The preprocessing includes radiometric calibration and atmospheric correction of optical images, filtering and classification of lidar point clouds to generate digital surface models, and rigorous geometric registration of all data sources.
[0043] Radiation processing eliminates atmospheric effects to ensure image color authenticity, point cloud processing extracts accurate 3D building models, and geometric registration ensures the consistency of spatial location of multi-source data, providing an accurate data foundation for subsequent feature fusion.
[0044] Multi-scale feature fusion includes feature extraction and fusion at the macro, meso, and micro scales. The macro scale corresponds to street-level features, the meso scale to building-level features, and the micro scale to component-level features.
[0045] By extracting macroscopic features reflecting the regional architectural pattern, mesoscopic features reflecting the overall form of the buildings, and microscopic features reflecting structural details, a comprehensive characterization from the environment to the building's own attributes is achieved, providing complete feature support for accurate assessment.
[0046] The multi-scale feature fusion deep learning model adopts a multi-branch neural network architecture, including a feature extraction branch, a feature fusion module, and a multi-task learning head.
[0047] By using a dedicated feature extraction branch to process information at different scales, a feature fusion module to achieve cross-scale information interaction, and a multi-task learning head to simultaneously optimize house segmentation and feature quantization tasks, the learning efficiency and feature quality of the model are improved.
[0048] Among them, the feature fusion module adopts the channel attention mechanism to achieve adaptive feature weighted fusion, and the multi-task learning head simultaneously outputs the house segmentation mask and multiple quantified hazard feature values.
[0049] By utilizing the channel attention mechanism, the fusion weights are automatically adjusted according to the importance of features to highlight key information. Through multi-task output, the location information and structural attribute parameters of the house are obtained simultaneously, ensuring that all kinds of information required for the assessment can be extracted completely at once.
[0050] Among them, the quantitative hazard characteristics include at least three combinations of structural irregularity index, roof material type, building height and building spacing.
[0051] Based on the principles of earthquake engineering, the combination of geometric, material, and environmental characteristics that have the most significant impact on the seismic performance of buildings is selected. These indicators with clear physical meaning ensure the scientific nature and interpretability of the hazard assessment.
[0052] The preset evaluation rules adopt an ensemble learning method based on random forest, which takes the quantitative hazard features as input features and outputs discrete hazard levels.
[0053] By leveraging the advantages of random forest models in handling high-dimensional features and their resistance to overfitting, stable and reliable ranking is achieved by integrating the voting results of multiple decision trees, ensuring the accuracy and robustness of the evaluation results.
[0054] The hazard level thematic map adopts the standard format of geographic information system, and each building patch is associated with an attribute table, recording the location coordinates, quantitative characteristic values and hazard level.
[0055] By adopting a common geographic information data format, the exchangeability and integrability of assessment results are ensured. Complete assessment information is recorded through structured attribute tables, which facilitates subsequent query analysis and decision-making applications.
[0056] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.
[0057] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A deep learning-based method for identifying and assessing the level of seismic hazards in buildings based on multi-scale remote sensing image fusion, characterized in that, Includes the following steps: Step 1: Data Acquisition and Preprocessing Acquire multi-source remote sensing image data of the target area, which includes at least high-resolution optical images, lidar point cloud data, and synthetic aperture radar data; perform radiometric calibration and atmospheric correction on the optical images, filter and classify the lidar point cloud data to generate a digital surface model, and perform geometric fine registration on all multi-source data to form a multi-source data layer with unified spatial reference. Step 2: Multi-scale feature fusion modeling A multi-branch deep learning model is constructed, which includes a feature extraction branch for extracting macro-contextual features and micro-structural features from the multi-source data layers of different scales or sources, a feature fusion module for adaptively weighted fusion of the multi-scale features, and a prediction head for multi-task learning based on the fused feature map. Step 3: Hazard Identification and Feature Quantification Using the multi-task learning prediction head, based on the fused features, the accurate segmentation of building objects is completed simultaneously, and the building binary mask and the intrinsic features related to earthquake hazards are quantified, outputting a quantized feature vector including structural irregularity index, roof material type, building height and building spacing. Step 4: Level Assessment Based on the quantized intrinsic feature vectors, and using a pre-defined evaluation rule, an integrated machine learning classifier is used to assess the seismic hazard level of each segmented building object and output discrete risk level labels. Step 5: Output Results Generate and output thematic maps of building seismic hazard levels, in which the spatial outline of each building object is associated with its assessed hazard level and quantitative feature vector, and stored in the standard format of geographic information system.
2. The method for deep learning-based identification and level assessment of building seismic hazards based on multi-scale remote sensing image fusion according to claim 1, characterized in that, The multi-source remote sensing image data includes sub-meter high-resolution optical images, lidar point cloud data, and synthetic aperture radar data.
3. The method for deep learning-based identification and level assessment of building seismic hazards based on multi-scale remote sensing image fusion according to claim 1, characterized in that, The preprocessing includes radiometric calibration and atmospheric correction of optical images, filtering and classification of lidar point clouds to generate digital surface models, and rigorous geometric registration of all data sources.
4. The method for deep learning-based identification and level assessment of building seismic hazards based on multi-scale remote sensing image fusion according to claim 1, characterized in that, The multi-scale feature fusion includes feature extraction and fusion at the macro, meso, and micro scales, where the macro scale corresponds to street-level features, the meso scale corresponds to building-level features, and the micro scale corresponds to component-level features.
5. The method for deep learning-based identification and level assessment of building seismic hazards based on multi-scale remote sensing image fusion according to claim 1, characterized in that, The multi-scale feature fusion deep learning model adopts a multi-branch neural network architecture. It includes a feature extraction branch, a feature fusion module, and a multi-task learning head.
6. The method for deep learning-based identification and level assessment of building seismic hazards based on multi-scale remote sensing image fusion according to claim 5, characterized in that, The feature fusion module uses a channel attention mechanism to achieve adaptive feature weighted fusion, and the multi-task learning head simultaneously outputs a house segmentation mask and multiple quantified hazard feature values.
7. The method for deep learning-based identification and grade assessment of building seismic hazards based on multi-scale remote sensing image fusion according to claim 1, characterized in that, The quantitative hazard characteristics include a combination of at least three of the following: structural irregularity index, roof material type, building height, and building spacing.
8. The method for deep learning-based identification and grade assessment of building seismic hazards based on multi-scale remote sensing image fusion according to claim 1, characterized in that, The preset evaluation rules adopt an ensemble learning method based on random forest, taking quantified hazard features as input features and outputting discrete hazard levels.
9. The method for deep learning-based identification and grade assessment of building seismic hazards based on multi-scale remote sensing image fusion according to claim 1, characterized in that, The thematic map of the hazard level adopts the standard format of geographic information system. Each building patch is associated with an attribute table, which records the location coordinates, quantitative feature values and hazard level.