Abstract:

Rapid and accurate localization of individual buildings and identification of earthquake-induced damage are essential for post-earthquake loss assessment and the efficient allocation of rescue resources. To overcome the limitations of conventional seismic damage detection methods, including strong dependence on training samples, blurred building boundaries, and interference from non-building objects, this study proposes a damaged-building change detection model based on the fusion of texture features(TFs) from unmanned aerial vehicle(UAV)imagery and principal component analysis(PCA). By integrating pre-earthquake satellite imagery with post-earthquake UAV data, the proposed model substantially improves the accuracy and efficiency of seismic damage identification through multi-feature fusion and dimensionality reduction. Case studies of the 2023 Jishishan M_S6.2 earthquake in Gansu Province and the 2021 Yangbi M_S6.4 earthquake in Yunnan Province were conducted to systematically validate the applicability and technical advantages of the model.

A normalized digital surface model(nDSM)was generated from UAV-derived digital surface model(DSM)data using triangulated irregular network(TIN)iterative filtering and point-cloud thinning, effectively removing vegetation and terrain effects and enabling precise extraction of individual building outlines. Building targets in pre- and post-earthquake single-phase(SP)images were segmented using region of interest(ROI)technology, while high-precision image registration was performed to ensure spatiotemporal consistency. For change detection, a multi-feature fusion framework was developed, including three models: texture-feature PCA change detection(PCA+TF+ROI+CD), texture-feature intensity-difference change detection(TF+DC+ROI+CD), and intensity-difference change detection(DC+ROI+CD). These models were designed to address the limitations of traditional single-phase machine-learning methods, such as maximum likelihood(ML) and Mahalanobis distance(MD), and were compared with ML+SP, ROI+ML+SP, MD+SP, and ROI+MD+SP methods.

Experimental results show that the PCA+TF+ROI+CD model achieved an overall accuracy(OA) of 89.4% and a Kappa coefficient of 0.85 in the Yangbi earthquake case, substantially outperforming conventional single-phase methods, including ML with an OA of 78.2% and MD with an OA of 75.6%. The model also achieved post-earthquake image matching accuracies above 88% in both case studies, with values of 88%±4% in Gansu and 91%±3% in Yunnan. Full processing of the disaster-affected areas was completed within 7h, satisfying the 24-h emergency response requirement. By integrating nDSM and ROI techniques, the model reduced the false-detection rate by 30% and improved boundary clarity by 20%. Moreover, the combination of large-scale pre-earthquake satellite imagery and high-resolution post-earthquake UAV data enables multi-scale analysis, ranging from regional damage assessment to individual building-level inspection. In the Gansu case, for example, multi-texture PCA fusion incorporating contrast, dissimilarity, and mean features suppressed redundancy caused by inter-feature correlations, reduced speckle noise, and enabled visualization of damage distribution at both pixel and building scales.

Nevertheless, the model performance remains sensitive to image quality, with accuracy decreasing under cloud cover, shadows, or low-resolution conditions. In complex urban environments, blurred boundaries between buildings and adjacent objects may require supplementary LiDAR data. In addition, limitations in computational efficiency for large-scale data processing may restrict real-time emergency applications. Future work should focus on multi-source data fusion, including optical, radar, and LiDAR data, to improve robustness under adverse weather and complex terrain conditions; transfer learning and adaptive algorithms to enhance model generalization; and optimized parallel computing architectures to improve processing efficiency in operational disaster response. Overall, this study provides an efficient and transferable solution for post-earthquake building damage detection, with both methodological innovation and practical value, while further improvements are needed in multi-source data compatibility, real-time processing, and adaptability to complex scenarios.

Key words: Jishishan M_S6.2 earthquake in Gansu Province, Yangbi M_S6.4 earthquake in Yunnan Province, multi-feature fusion, principal component analysis, pre- and post-earthquake change detection, earthquake damage detection

摘要：

在地震灾害应急响应中, 快速精准识别单体建筑震害对灾损评估和救援决策至关重要。针对传统震害检测方法存在的样本依赖性高、边界模糊和非建筑干扰等问题, 文中提出基于无人机影像纹理特征主成分分析(PCA)的破坏建筑物变化检测模型, 结合震前卫星影像与震后无人机数据, 通过多特征融合与降维处理提升检测效能。以2023年甘肃积石山 M_S6.2 和2021年云南漾濞 M_S6.4 地震为例, 模型利用三角网迭代滤波和点云抽稀方法, 从无人机数字表面模型(DSM)生成归一化数字表面模型(nDSM), 有效剔除植被与地形干扰, 精准提取建筑物轮廓; 基于感兴趣区域(ROI)技术分割震前、震后单时相影像中的建筑目标, 并通过高精度配准(甘肃匹配精度为88%±4%, 云南为91%±3%)确保时空一致性。通过主成分分析方法融合协同性、对比度、相异性等8类纹理特征, 筛选前4项主成分构建变化检测框架, 并与纹理特征强度差值法(TF+DC+ROI+CD)、强度差值法(DC+ROI+CD)、震后单时相最大似然法(ML)及马氏距离法(MD)的结果进行对比。结果表明, PCA融合模型在漾濞地震中的总体精度达89.4%、Kappa系数为0.85, 显著优于单时相方法(ML OA, 78.2%; MD OA, 75.6%)。在研究区0.62km²(甘肃)和0.49km²(云南)范围内实现7h全流程处理, 能够满足应急响应时效需求。

关键词: 甘肃积石山M_S6.2地震, 云南漾濞M_S6.4地震, 多特征融合, 主成分分析, 震前震后变化检测, 震害检测