Abstract:

Earthquakes, as sudden-onset natural disasters with high destructive potential, not only result in significant casualties but also cause severe damage to infrastructure—particularly buildings—posing major challenges to post-disaster rescue and reconstruction efforts. In emergency response scenarios, the rapid and accurate assessment of building damage is a critical prerequisite for formulating effective rescue strategies and allocating resources efficiently. Traditional manual on-site investigation methods, however, present notable limitations. Disaster-affected areas often experience traffic disruptions and harsh environmental conditions, hindering timely access for investigators. Moreover, manual assessments are time-consuming and generally incapable of meeting the urgent demands of rescue operations within the critical 72-hour post-disaster window. Large-scale manual surveys also involve safety risks, potentially leading to secondary casualties. Therefore, the development of rapid, efficient, and accurate building damage assessment methods holds significant practical and strategic importance.

In response to this need, the this study we proposes an innovative rapid assessment method for earthquake-induced building damage using unmanned aerial vehicle(UAV)imagery combined with machine learning algorithms. This method leverages the advantages of UAV remote sensing—such as high mobility, flexibility, and high spatial resolution—together with advanced image processing and machine learning techniques to enable intelligent identification and assessment of building damage. The study focuses on the M_S6.8 earthquake that struck Dingri County, Xizang, using it as a case study to validate the proposed methodology through a structured technical workflow. The assessment framework comprises three key stages. First, an object-oriented remote sensing classification(OORSC) approach was used to extract individual building features from UAV imagery. By employing rule-based classification strategies, this method effectively eliminates background noise such as trees and roads. After morphological filtering, the completeness of building boundary extraction exceeded 95%, and hole-filling performance was markedly improved, ensuring high-quality input data for subsequent analyses. Second, the study focused on the extraction and optimization of surface texture features. Using algorithms such as the Gray-Level Co-occurrence Matrix(GLCM) and Local Binary Pattern(LBP), critical parameters—including contrast, entropy, and variance—were derived. Experimental data show that, on average, the contrast of collapsed buildings is 26%lower than that of intact buildings, while entropy and variance increase by 32% and 41%, respectively. These features provide robust quantitative indicators for identifying structural damage. Lastly, the study implemented a comparative experimental design incorporating four technical routes, systematically evaluating the performance of classification algorithms such as Support Vector Machines(SVM) and Neural Networks(NN).

The results demonstrate that the neural network model integrating optimized texture features yields the best performance, achieving an overall accuracy(OA)of 91% and a Kappa coefficient of 0.8. Compared to models excluding texture features, the improvement is significant: the neural network model without texture features achieved an OA of 85% and a Kappa of 0.6, while the SVM-based approach achieved an OA of 82% and a Kappa of 0.6. The recognition accuracy by damage level further reveals that severely damaged buildings are most accurately identified(94%)due to their distinctive visual characteristics, followed by collapsed(87%)and moderately damaged buildings(80%). Misclassification of collapsed structures mainly stems from blurred textures in the imagery. These findings underscore the critical role of texture features in building damage identification and validate the proposed method’s effectiveness in supporting post-disaster emergency response.

Despite the promising results, the proposed method has several limitations in practical application. At the technical level, complex environmental backgrounds—such as similar roof materials and shadow effects—can interfere with detection accuracy and demand high image quality. At the data level, a lack of sufficient real-time ground truth data may compromise model training accuracy. At the application level, the method’s capacity to detect complex damage types—such as internal structural failures—remains limited. To address these challenges, future research will focus on several directions. From a technical innovation perspective, advanced methods such as deep learning will be explored, particularly the use of three-dimensional convolutional neural networks(3D-CNNs)for capturing volumetric building features. In terms of data integration, the fusion of multi-source data—such as LiDAR point clouds, digital surface models(DSM), and thermal infrared imagery—will be pursued to build a multimodal feature fusion framework. Methodologically, transfer learning and data augmentation will be applied to enhance model generalizability, and adaptive algorithms will be developed to manage complex and dynamic disaster scenarios. On the application front, the establishment of a standardized sample library and evaluation system is proposed to support the broader deployment and engineering application of the method.

The significance of this study is multidimensional. Theoretically, it introduces a novel approach to building damage identification based on texture features and machine learning, enriching the theoretical framework of remote sensing-based disaster assessment. Technologically, it develops a comprehensive UAV image processing and analysis pipeline, offering a replicable technical route for related research. Practically, the established system can be directly applied to post-earthquake emergency response, enhancing the efficiency and effectiveness of rescue operations. With continued technological advancement, the method holds potential for adaptation to other disaster scenarios, such as typhoons and floods, thereby contributing to integrated disaster risk reduction. Future work will continue to advance research in this area, targeting breakthroughs in key challenges such as multi-source data fusion and intelligent algorithm optimization, with the goal of advancing disaster assessment technologies toward greater intelligence and precision, ultimately contributing to the protection of lives and property and the promotion of sustainable development.

Key words: Dingri M_S6.8 earthquake, UAV imagery, texture features, building damage identification, support vector machine(SVM), neural network

摘要：

地震后及时获取建筑物破坏信息对于应急救援和灾害损失评估至关重要。文中基于震后无人机影像数据, 提出了一种结合面向对象、 支持向量机(SVM)和神经网络(NN)的单体建筑物纹理特征损伤检测方法, 并以2025年西藏定日县 M_S6.8 地震为例进行验证。该方法通过面向对象方法提取单体建筑物信息, 消除非建筑物干扰; 采用灰度共生矩阵(GLCM)提取对比度、 熵和方差等纹理特征, 优化窗口大小至7×7以提升特征区分度。通过对比4种方法发现: 融合最优纹理特征后, 神经网络分类算法(单体+纹理特征+神经网络)的总体精度达91%, Kappa系数为0.8, 较未融合纹理特征的单体+神经网络方法(精度85%、 Kappa 0.6)分别提升6%和0.2; 与支持向量机方法相比, 单体+纹理特征+支持向量机(精度89%、 Kappa 0.7)较单体+支持向量机(精度82%、 Kappa 0.6)提升7%和0.1。实验表明, 纹理特征可显著增强对损伤的识别能力, 倒塌建筑物的对比度均值较完好建筑降低26%, 熵和方差分别增加32%和41%。该方法有效解决了非建筑信息干扰的问题, 经形态学滤波处理后孔洞填充率 >95%。文中研究为震后快速评估提供了高精度、 可量化的技术支撑, 验证了多特征融合与算法协同优化的有效性。

关键词: 定日M_S6.8地震, 无人机影像, 纹理特征, 建筑物震害识别, 支持向量机, 神经网络