Scholarlink

<p>Climate change has led to more frequent and intense droughts and temperature extremes in the western United States, resulting in increasingly severe and more frequent wildfires. The concurrent expansion of residential construction into wildland areas has raised the likelihood of ignitions and the potential scale of damage to infrastructure, homes, and businesses. Structure protection in the wildland-urban interface is challenged by these increased ignitions, altered fuel structures in human-dominated areas, and mixed acceptance of prescribed burning and other prefire mitigation to minimize wildfire damage. Being able to track the multiyear building recovery of entire wildfire-impacted communities is crucial to analyze post-fire resilience in recovering communities and can support local decision-makers in determining what conditions slow down or accelerate the repair and rebuilding efforts. Here, we document post-wildfire building recovery using post-fire damage data and corresponding nanosatellite imagery to assess the monthly recovery rate of buildings over 3&#xa0;years following the 2017 Tubbs Fire and 2018 Camp Fire in California, and the 2020 Echo Mountain Complex and 2020 Almeda Drive Fire in Oregon. The approach facilitates tracking of the building recovery timeline across a community using 3&#xa0;m freely available multispectral imagery. We trained the deep residual network model with a binary classification layer to detect changes in the reflected spectra related to repairing or rebuilding in four study areas damaged by wildfires. The residual network using small property-sized fragments of imagery outperformed a U-Net segmentation network and achieved a validation accuracy of 94.0% and test accuracy of 92.6% for the classification of undamaged and damaged or destroyed buildings. We found different recovery timelines across study communities in recovery with distinct temporal patterns, with the Echo Mountain Complex recovering most rapidly, primarily due to the smaller number of structures initially lost and the use of a large proportion of prefabricated mobile structures in the rebuilding effort. In contrast, communities in the Camp Fire are rebuilding infrastructure much more slowly, partly due to a shift toward fire-hardened infrastructure and increased insurance regulations on building types. Our machine learning approach can be applied to determine the pace and the extent of recovery with high accuracy using an established deep network approach and freely available satellite data providing sufficient spatial and temporal resolution.</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

Long-term tracking of recovery of built infrastructure after wildfires with deep network topologies

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年1月29日

查看全文
  • Andres Schmidt,
  • Lisa Ellsworth,
  • Jenna Tilt,
  • Amanda Thiel,
  • Nancy Hiner

摘要

Climate change has led to more frequent and intense droughts and temperature extremes in the western United States, resulting in increasingly severe and more frequent wildfires. The concurrent expansion of residential construction into wildland areas has raised the likelihood of ignitions and the potential scale of damage to infrastructure, homes, and businesses. Structure protection in the wildland-urban interface is challenged by these increased ignitions, altered fuel structures in human-dominated areas, and mixed acceptance of prescribed burning and other prefire mitigation to minimize wildfire damage. Being able to track the multiyear building recovery of entire wildfire-impacted communities is crucial to analyze post-fire resilience in recovering communities and can support local decision-makers in determining what conditions slow down or accelerate the repair and rebuilding efforts. Here, we document post-wildfire building recovery using post-fire damage data and corresponding nanosatellite imagery to assess the monthly recovery rate of buildings over 3 years following the 2017 Tubbs Fire and 2018 Camp Fire in California, and the 2020 Echo Mountain Complex and 2020 Almeda Drive Fire in Oregon. The approach facilitates tracking of the building recovery timeline across a community using 3 m freely available multispectral imagery. We trained the deep residual network model with a binary classification layer to detect changes in the reflected spectra related to repairing or rebuilding in four study areas damaged by wildfires. The residual network using small property-sized fragments of imagery outperformed a U-Net segmentation network and achieved a validation accuracy of 94.0% and test accuracy of 92.6% for the classification of undamaged and damaged or destroyed buildings. We found different recovery timelines across study communities in recovery with distinct temporal patterns, with the Echo Mountain Complex recovering most rapidly, primarily due to the smaller number of structures initially lost and the use of a large proportion of prefabricated mobile structures in the rebuilding effort. In contrast, communities in the Camp Fire are rebuilding infrastructure much more slowly, partly due to a shift toward fire-hardened infrastructure and increased insurance regulations on building types. Our machine learning approach can be applied to determine the pace and the extent of recovery with high accuracy using an established deep network approach and freely available satellite data providing sufficient spatial and temporal resolution.