Scholarlink

<p>Wildfires occur worldwide, primarily driven by natural factors and human error, with their spread influenced by fuel loads, weather, and topography. Mapping fire potential based on fuel conditions can support targeted management and help reduce wildfire impacts. This study develops an interpretable data-driven remote-sensing framework to classify fuel types and estimate fire probability mapping in savanna areas using multitemporal Sentinel-2 imagery. The method employs a Neural Network (NN) to classify four fuel types, primarily consisting of tree, grass, sparse vegetation, and soil, achieving an accuracy of 98.22%. The classified fuel types complemented with vegetation density and moisture content, was subsequently utilized as input within the Bayesian Belief Network (BBN) process. The approach, NN-BBN was implemented on the savanna landscape of Indonesia’s Mount Bromo region, and the resulting fire probability outputs were validated using a reported wildfire event in late 2023. Validation using Pearson correlation yielded <i>r</i> = 0.80, indicating that the BBN-modeled fire probability is strongly correlated with the observed burned pixels from the actual incident. Also validated across 20 spatial cross-validation folds with per-fold threshold optimization, the model achieved a high and quite stable precision (mean ± SD: 0.841 ± 0.185) but moderate and variable recall (0.551 ± 0.278), resulting in an AUC-PR of 0.835 ± 0.124—indicating reliable positive predictions but inconsistent detection of all true positives. This methodology offers a novel probabilistic instrument for predicting wildfire risk, facilitating early warning and mitigation strategies in tropical savanna ecosystems.</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

Probability to Reality: Neural Network-Enhanced Bayesian Belief Network to Mapping Wildfires Vulnerability in Bromo, Indonesia

  • 期刊论文

  • 发表日期：

    2026年5月9日

查看全文
  • Bayu Satya Adhitama,
  • Ika Siwi Supriyani,
  • Karunika Diwyacitta,
  • Yenni Vetrita,
  • Andy Indradjad,
  • Donny Kushardono,
  • Fauzan Al Ayyubi,
  • Olivia Maftukhaturrizqoh

摘要

Wildfires occur worldwide, primarily driven by natural factors and human error, with their spread influenced by fuel loads, weather, and topography. Mapping fire potential based on fuel conditions can support targeted management and help reduce wildfire impacts. This study develops an interpretable data-driven remote-sensing framework to classify fuel types and estimate fire probability mapping in savanna areas using multitemporal Sentinel-2 imagery. The method employs a Neural Network (NN) to classify four fuel types, primarily consisting of tree, grass, sparse vegetation, and soil, achieving an accuracy of 98.22%. The classified fuel types complemented with vegetation density and moisture content, was subsequently utilized as input within the Bayesian Belief Network (BBN) process. The approach, NN-BBN was implemented on the savanna landscape of Indonesia’s Mount Bromo region, and the resulting fire probability outputs were validated using a reported wildfire event in late 2023. Validation using Pearson correlation yielded r = 0.80, indicating that the BBN-modeled fire probability is strongly correlated with the observed burned pixels from the actual incident. Also validated across 20 spatial cross-validation folds with per-fold threshold optimization, the model achieved a high and quite stable precision (mean ± SD: 0.841 ± 0.185) but moderate and variable recall (0.551 ± 0.278), resulting in an AUC-PR of 0.835 ± 0.124—indicating reliable positive predictions but inconsistent detection of all true positives. This methodology offers a novel probabilistic instrument for predicting wildfire risk, facilitating early warning and mitigation strategies in tropical savanna ecosystems.