Scholarlink

<p>Accurately modeling wildfire spread can guide fire management decisions. Current deep learning-based research on wildfire prediction focuses on large-scale fires at low spatio-temporal resolution, with high-resolution prediction largely remaining theoretical. For the same study area, more detailed fire parameter inputs typically lead to a significant increase in computational memory demands. To address this challenge, this paper proposes a high-resolution burn map prediction framework based on a divide-and-conquer strategy, consisting of three stages: burn map segmentation, segmented sub-burn map prediction, and predicted result stitching. The framework uses a multi-size segmentation method to mitigate the negative effects on prediction accuracy caused by inappropriate fire occurrences within the segmented sub-burn maps. In addition, it uses an Unet fire margin correction network to reduce predicted error accumulation during the stitching process. The framework is trained and tested using two independent datasets: FARSITE-simulated wildfires and field-prescribed fire experiments. Quantitative results demonstrate the proposed framework’s advantages in both predictive performance and memory efficiency compared to other deep learning models. These results show the framework’s potential for managing larger areas and higher-resolution wildfire scenarios, and can provide a theoretical and data-driven foundation for applying modular and distributed computing approaches to wildfire prediction.</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

A Divide-and-Conquer Deep Learning Framework for High-Resolution Burn Map Prediction Under Low Computational Memory Condition

  • 期刊论文

  • 发表日期：

    2026年2月24日

查看全文
  • Xinyu Wang,
  • Yunchuan Ling,
  • YiFan Fu,
  • Xingdong Li,
  • Hongyang Zhao,
  • Dandan Li,
  • Shufa Sun

摘要

Accurately modeling wildfire spread can guide fire management decisions. Current deep learning-based research on wildfire prediction focuses on large-scale fires at low spatio-temporal resolution, with high-resolution prediction largely remaining theoretical. For the same study area, more detailed fire parameter inputs typically lead to a significant increase in computational memory demands. To address this challenge, this paper proposes a high-resolution burn map prediction framework based on a divide-and-conquer strategy, consisting of three stages: burn map segmentation, segmented sub-burn map prediction, and predicted result stitching. The framework uses a multi-size segmentation method to mitigate the negative effects on prediction accuracy caused by inappropriate fire occurrences within the segmented sub-burn maps. In addition, it uses an Unet fire margin correction network to reduce predicted error accumulation during the stitching process. The framework is trained and tested using two independent datasets: FARSITE-simulated wildfires and field-prescribed fire experiments. Quantitative results demonstrate the proposed framework’s advantages in both predictive performance and memory efficiency compared to other deep learning models. These results show the framework’s potential for managing larger areas and higher-resolution wildfire scenarios, and can provide a theoretical and data-driven foundation for applying modular and distributed computing approaches to wildfire prediction.