Scholarlink

<p>Post-disaster building damage assessment using high-resolution optical satellite imagery is an essential component of urban resilience and emergency response. This study evaluates three deep learning models—you only look once version 8 (YOLOv8-seg), U-Net, and deep fully convolutional neural network (DFCNN)—on the xBD (xView2 Building Damage) dataset to detect and segment building damage caused by earthquakes, floods and other natural hazards across diverse geographic regions and urban environments. Apart from the usual performance metrics, the study also employs a feature-based analysis to examine the effectiveness of contextual urban features on model accuracy. The basic features, such as urban density, urban orderliness effect index (UOEI), building structure types, and types of hazards, are used to measure the sensitivity of the predictions made by the models in different types of urban settings. The study concludes that YOLOv8-seg achieved the highest performance, with an F1-score of 0.80, using only post-disaster images and the selected pre-processing techniques. The findings highlight that detection accuracy is strongly influenced by urban density, structural patterns, and hazard characteristics. By incorporating contextual features into the evaluation process, this study improves the interpretability and applicability of deep learning models for post-disaster damage assessment and supports more reliable decision-making in real-world hazard scenarios.</p> Graphical abstract <p></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Feature-based framework for evaluating deep learning models in post-disaster building damage assessment under diverse natural hazards

  • 期刊论文

  • 发表日期：

    2026年5月14日

查看全文
  • Milad Farmahini Farahani,
  • Mohammad Reza Nikoo,
  • Houshang Dolatshahi Pirooz,
  • Alireza Shadmani,
  • Amir H. Gandomi

摘要

Post-disaster building damage assessment using high-resolution optical satellite imagery is an essential component of urban resilience and emergency response. This study evaluates three deep learning models—you only look once version 8 (YOLOv8-seg), U-Net, and deep fully convolutional neural network (DFCNN)—on the xBD (xView2 Building Damage) dataset to detect and segment building damage caused by earthquakes, floods and other natural hazards across diverse geographic regions and urban environments. Apart from the usual performance metrics, the study also employs a feature-based analysis to examine the effectiveness of contextual urban features on model accuracy. The basic features, such as urban density, urban orderliness effect index (UOEI), building structure types, and types of hazards, are used to measure the sensitivity of the predictions made by the models in different types of urban settings. The study concludes that YOLOv8-seg achieved the highest performance, with an F1-score of 0.80, using only post-disaster images and the selected pre-processing techniques. The findings highlight that detection accuracy is strongly influenced by urban density, structural patterns, and hazard characteristics. By incorporating contextual features into the evaluation process, this study improves the interpretability and applicability of deep learning models for post-disaster damage assessment and supports more reliable decision-making in real-world hazard scenarios.

Graphical abstract