Scholarlink

<p>Accurate extraction of built structures from satellite imagery is crucial for effective post-disaster response, particularly when rapid identification of blocked roads and debris can mean the difference between life and death. In this study, we systematically evaluate and compare three state-of-the-art deep learning models—YOLO, Mask R-CNN, and Roboflow instance segmentation—for their ability to detect and segment debris from satellite images captured after the devastating Mw 7.8 earthquake in Turkey on February 6, 2023. Despite challenges such as variable image resolutions and the scarcity of annotated disaster-specific data, each model demonstrated a&#xa0;baseline level of performance; notably, Roboflow instance segmentation achieved 53.5%, Mask R-CNN (Detectron2) obtained 39.8% at IoU = 0.50, YOLOv5 achieved 57.1% mean Average Precision (mAP), and YOLOv8 scored 52.0%. This work compares four state-of-the-art models in detail and uses publicly available satellite data in resource-constrained environments. It also incorporates an offline graphical user interface (GUI) that overlays traffic maps with identified debris to identify blocked highways for emergency routing. To overcome the limitations of a modest dataset—initially comprising only 510 satellite images—we employed extensive augmentation techniques, effectively doubling the dataset and underscoring the viability of deep learning approaches even in resource-constrained environments. Future work should focus on debris detection with multi-class, integration of higher-resolution satellite data, and validation of real-time system deployment to overcome existing operational readiness and accuracy constraints.</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

Deep Learning for Post-earthquake Road Blockage and Debris Detection in Resource-Limited Contexts

  • 期刊论文

  • 发表日期：

    2025年12月29日

查看全文
  • Zeinab Bayat,
  • Serdar Dindar,
  • Burcu Güldür Erkal

摘要

Accurate extraction of built structures from satellite imagery is crucial for effective post-disaster response, particularly when rapid identification of blocked roads and debris can mean the difference between life and death. In this study, we systematically evaluate and compare three state-of-the-art deep learning models—YOLO, Mask R-CNN, and Roboflow instance segmentation—for their ability to detect and segment debris from satellite images captured after the devastating Mw 7.8 earthquake in Turkey on February 6, 2023. Despite challenges such as variable image resolutions and the scarcity of annotated disaster-specific data, each model demonstrated a baseline level of performance; notably, Roboflow instance segmentation achieved 53.5%, Mask R-CNN (Detectron2) obtained 39.8% at IoU = 0.50, YOLOv5 achieved 57.1% mean Average Precision (mAP), and YOLOv8 scored 52.0%. This work compares four state-of-the-art models in detail and uses publicly available satellite data in resource-constrained environments. It also incorporates an offline graphical user interface (GUI) that overlays traffic maps with identified debris to identify blocked highways for emergency routing. To overcome the limitations of a modest dataset—initially comprising only 510 satellite images—we employed extensive augmentation techniques, effectively doubling the dataset and underscoring the viability of deep learning approaches even in resource-constrained environments. Future work should focus on debris detection with multi-class, integration of higher-resolution satellite data, and validation of real-time system deployment to overcome existing operational readiness and accuracy constraints.