Scholarlink

<p>Rapid and accurate identification of collapsed buildings after natural disasters is critical for effective emergency response and resource allocation. Remote sensing (RS) imagery provides rapid situational awareness; however, most existing approaches remain limited to pixel-wise segmentation, which does not directly translate into operational building-level decisions. This study proposes a decision-level collapsed building identification framework that explicitly bridges pixel-wise semantic segmentation and building-level classification using pre-disaster building footprint polygons. The framework is evaluated using a novel UAV-based dataset acquired after the 2023 Kahramanmaraş earthquake sequence in Turkey, together with the publicly available xView2 Building Damage Assessment (xBD) dataset to examine robustness across different disaster scenarios and sensing modalities. Collapsed building identification is performed in two stages using the DeepLabV3+ segmentation pipeline with multiple backbone architectures. Pixel-wise segmentation outputs are aggregated within pre-disaster building polygons, and a threshold sensitivity analysis (τ-sweep) is introduced to systematically evaluate the impact of decision thresholds on building-level classification outcomes. Experimental results show that while the ResNet50 backbone achieves the best overall segmentation performance, the MobileNetV2 backbone yields superior performance for collapsed buildings, achieving a DSC of 0.853 and an IoU of 0.755. At the decision level, polygon-based classification achieves consistently high accuracy across architectures on the xBD dataset, reaching up to 93.5%. A computational complexity analysis indicates that per-image inference times range from 58.8 to 96.2&#xa0;ms, demonstrating that the proposed decision-level formulation is computationally efficient and well-suited for scalable, time-critical collapsed building identification over large RS datasets.</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

Multi-task building damage assessment via deep semantic segmentation and pre-disaster polygons

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月5日

查看全文
  • Serhat Alpergin,
  • Hasan Polat,
  • Mehmet Siraç Özerdem

摘要

Rapid and accurate identification of collapsed buildings after natural disasters is critical for effective emergency response and resource allocation. Remote sensing (RS) imagery provides rapid situational awareness; however, most existing approaches remain limited to pixel-wise segmentation, which does not directly translate into operational building-level decisions. This study proposes a decision-level collapsed building identification framework that explicitly bridges pixel-wise semantic segmentation and building-level classification using pre-disaster building footprint polygons. The framework is evaluated using a novel UAV-based dataset acquired after the 2023 Kahramanmaraş earthquake sequence in Turkey, together with the publicly available xView2 Building Damage Assessment (xBD) dataset to examine robustness across different disaster scenarios and sensing modalities. Collapsed building identification is performed in two stages using the DeepLabV3+ segmentation pipeline with multiple backbone architectures. Pixel-wise segmentation outputs are aggregated within pre-disaster building polygons, and a threshold sensitivity analysis (τ-sweep) is introduced to systematically evaluate the impact of decision thresholds on building-level classification outcomes. Experimental results show that while the ResNet50 backbone achieves the best overall segmentation performance, the MobileNetV2 backbone yields superior performance for collapsed buildings, achieving a DSC of 0.853 and an IoU of 0.755. At the decision level, polygon-based classification achieves consistently high accuracy across architectures on the xBD dataset, reaching up to 93.5%. A computational complexity analysis indicates that per-image inference times range from 58.8 to 96.2 ms, demonstrating that the proposed decision-level formulation is computationally efficient and well-suited for scalable, time-critical collapsed building identification over large RS datasets.