中国安全科学学报 ›› 2022, Vol. 32 ›› Issue (9): 152-157.doi: 10.16265/j.cnki.issn1003-3033.2022.09.2729

• 防灾减灾技术与工程 • 上一篇    下一篇

基于机器学习的森林火险预测模型

朱馨1,2(), 李建微1,**(), 郭伟3, 毕胜1,2, 伍跃飞1,2   

  1. 1 福州大学 物理与信息工程学院,福建 福州 350116
    2 福州大学 数字中国研究院(福建), 福建 福州 350116
    3 福建省气象局 气象服务中心,福建 福州 350028
  • 收稿日期:2022-03-20 修回日期:2022-07-11 出版日期:2022-10-19 发布日期:2023-03-28
  • 通讯作者: 李建微
  • 作者简介:

    朱 馨 (1997—),男,福建莆田人,硕士研究生,主要研究方向为数字地球与智慧社会。E-mail:

    李建微 副研究员

  • 基金资助:
    国家自然科学基金资助(32071776); 国家自然科学基金资助(41571490); 福建省自然科学基金资助(2020J01465); 中国博士后基金资助(2018M640597)

Forest fire risk prediction model based on machine learning

ZHU Xin1,2(), LI Jianwei1,**(), GUO Wei3, BI Sheng1,2, WU Yuefei1,2   

  1. 1 Academy of Digital China(Fujian),Fuzhou University, Fuzhou Fujian 350116, China
    2 College of Physics and Information Engineering, Fuzhou University, Fuzhou Fujian 350116, China
    3 Fujian Meteorological Service Center, Fuzhou Fujian 350028, China
  • Received:2022-03-20 Revised:2022-07-11 Online:2022-10-19 Published:2023-03-28
  • Contact: LI Jianwei

摘要:

为减少森林火灾带来的损害,通过文献回顾,对森林火险进行建模和预测预报。归纳基于机器学习方法的森林火险预测研究现状,并从森林火灾影响因子的选取、选择合适的火险预测模型以及模型检验方法3个主要方面进行分析阐述。结果表明:森林火险的主要影响因素包括可燃物特征、气象因子、地形、人类活动等;在森林火险预测模型中,反向传播(BP)神经网络方法需要改进后运用,支持向量机(SVM)方法对数据要求高,随机森林(RF)方法通用性强且精度较高,深度学习方法的研究较少,但精度都很高;模型常用的检验方法是准确度、受试者工作特征(ROC)曲线和曲线下的面积(AUC)值等。

关键词: 机器学习, 森林火险预测, 森林火灾, 气象因子, 支持向量机(SVM)

Abstract:

In order to reduce the damage caused by forest fires, the forest fire risk was modeled and predicted through a literature review. This paper summarizes the research status of forest fire risk prediction based on the machine learning method, and analyzes and expounds on the selection of forest fire impact factors, the selection of appropriate fire risk prediction models and the model test methods. The results show that the main influencing factors of forest fire risk include the characteristics of combustibles, meteorological factors, terrain, human activities, etc. In the forest fire risk prediction model, the back propagation(BP) neural network method needs to be improved and applied. The SVM method requires high data, the random forest(RF) method has strong universality and high accuracy, and the deep learning method has less research, but the accuracy is high. The commonly used test methods of the model are accuracy, receiver operating characteristic(ROC) curve and area under curve(AUC) value.

Key words: machine learning, forest fire risk prediction, forest fire, meteorological factors, support vector machine(SVM)