高级检索

留言板

尊敬的读者、作者、审稿人, 关于本刊的投稿、审稿、编辑和出版的任何问题, 您可以本页添加留言。我们将尽快给您答复。谢谢您的支持!

姓名
邮箱
手机号码
标题
留言内容
验证码

基于贝叶斯模型平均法的森林火灾预测模型构建研究

白海峰 刘晓东 牛树奎 何亚东

白海峰, 刘晓东, 牛树奎, 何亚东. 基于贝叶斯模型平均法的森林火灾预测模型构建研究——以云南省大理州为例[J]. 北京林业大学学报, 2021, 43(5): 44-52. doi: 10.12171/j.1000-1522.20200173
引用本文: 白海峰, 刘晓东, 牛树奎, 何亚东. 基于贝叶斯模型平均法的森林火灾预测模型构建研究——以云南省大理州为例[J]. 北京林业大学学报, 2021, 43(5): 44-52. doi: 10.12171/j.1000-1522.20200173
Bai Haifeng, Liu Xiaodong, Niu Shukui, He Yadong. Construction of forest fire prediction model based on Bayesian model averaging method: taking Dali Prefecture, Yunnan Province of southwestern China as an example[J]. Journal of Beijing Forestry University, 2021, 43(5): 44-52. doi: 10.12171/j.1000-1522.20200173
Citation: Bai Haifeng, Liu Xiaodong, Niu Shukui, He Yadong. Construction of forest fire prediction model based on Bayesian model averaging method: taking Dali Prefecture, Yunnan Province of southwestern China as an example[J]. Journal of Beijing Forestry University, 2021, 43(5): 44-52. doi: 10.12171/j.1000-1522.20200173

基于贝叶斯模型平均法的森林火灾预测模型构建研究

——以云南省大理州为例

doi: 10.12171/j.1000-1522.20200173
基金项目: 国家自然科学基金项目(31770696)
详细信息
    作者简介:

    白海峰。主要研究方向:林火生态。Email:haifengbai@sina.com 地址:100083 北京市海淀区清华东路35号北京林业大学生态与自然保护学院

    责任作者:

    刘晓东,博士,教授。主要研究方向:林火生态。Email:xd_liu@bjfu.edu.cn 地址:同上

  • 中图分类号: S762.2

Construction of forest fire prediction model based on Bayesian model averaging method: taking Dali Prefecture, Yunnan Province of southwestern China as an example

  • 摘要:   目的  本文基于贝叶斯模型平均法,结合二项逻辑斯蒂回归模型,构建云南省大理州森林火灾发生预测模型,以期提高林火预测精度,为研究地区林火管理提供技术支持。  方法  利用2000—2013年大理州林火数据及对应的气象数据,分别运用二项逻辑斯蒂回归模型和贝叶斯模型平均法,对该地区森林火灾对气象因子的响应进行实证分析。二项逻辑斯蒂回归模型为单一模型,建模前通过对各解释变量进行多重共线性检验,剔除有显著共线性的解释变量,然后通过逐步回归法,筛选最终变量并进行参数拟合。贝叶斯平均模型为组合模型,基于贝叶斯模型平均法建模时,采用奥卡姆窗的方法来适当调整模型空间,并以5个最优模型的后验概率作为权重进行加权建模。将全样本数据随机分成80%的训练样本和20%的测试样本,基于训练样本建立模型,对测试样本进行预测,通过对比观测值和预测值计算模型的准确率。  结果  通过二项逻辑斯蒂模型拟合,优度为0.783,预测精度为0.718。通过贝叶斯平均模型拟合,优度为0.868,预测精度为0.807。2个模型预测结果对比显示,在训练集中,贝叶斯平均模型的预测准确率比二项逻辑斯蒂回归模型高9.3%;在测试集中,贝叶斯平均模型的预测准确率比二项逻辑斯蒂回归模型高8.9%。  结论  在基于气象因子的大理州林火发生预测模型构建研究中,贝叶斯平均模型的拟合优度和预测精度均高于二项逻辑斯蒂模型,表明贝叶斯模型平均法具有一定的现实应用意义,可用于提高研究地区林火预测精度,有利于森林火灾的决策管理。

     

  • 图  1  研究区2000—2013年火点分布

    Figure  1.  Distribution of fire points in the study area from 2000 to 2013

    图  2  Step_LR模型的ROC曲线图

    AUC为曲线下面积。下同。AUC, area under curve. The same below.

    Figure  2.  ROC curve of Step_LR model

    图  3  BMA模型可视化

    图示根据奥卡姆窗被选中的98个模型以及每个模型各自选中的变量。横轴为模型编号,宽度表示该模型的后验概率大小,纵轴为解释变量代码。红色表示该变量与被解释变量存在正相关关系,蓝色表示存在负相关关系,无颜色即表示该变量没有被选入该模型。The figure shows the 98 models selected according to the Occam’s window and the variables selected by each model. The x-axis refers to the model No., the width represents the posterior probability of the model, and the y-axis is equidistant, showing the code of each explanatory variable. The red indicates that the variable has a positive correlation with the explained variable, the blue indicates that there is a negative correlation, and the variable without colour is not selected into the model.

    Figure  3.  BMA model visualization

    图  4  BMA_LR模型的ROC曲线图

    Figure  4.  ROC curve of BMA_LR model

    表  1  模型变量的基本统计描述

    Table  1.   Basic statistical description of model variables

    模型变量 Model variable变量代码 Variable code最小值 Min. value最大值 Max. value均值 Mean标准差 SD
    日平均风速 Daily average wind speed/(m·s−1) WIN_avg 0.80 10.80 3.64 1.38
    日最大风速 Daily maximum wind speed/(m·s−1) WIN_max 3.10 20.60 9.21 2.35
    日照时数 Sunshine hour/h SSD 2.20 12.20 9.24 1.80
    日平均气压 Daily average pressure/kPa PRS_avg 79.33 80.75 80.05 0.22
    日平均气温 Daily average temperature/℃ Tavg 4.20 24.10 15.88 3.34
    日最高气温 Daily maximum temperature/℃ Tmax 12.10 31.00 23.34 3.17
    日最低气温 Daily minimum temperature/℃ Tmin −0.80 18.20 8.59 3.92
    日平均水汽压 Daily average water vapor pressure/kPa VP_avg 0.27 1.68 0.71 0.21
    日平均相对湿度 Daily average relative humidity/% RH_avg 21.00 72.00 41.46 8.27
    日最小相对湿度 Daily minimum relative humidity/% RH_min 6.00 46.00 18.78 5.71
    前一日20:00—20:00降雨量
    20:00 the day before−20:00 precipitation/mm
    Pre 0 3.00 0.03 0.23
    细小可燃物湿度码 Fine fuel moisture code FFMC 79.33 97.56 94.6 1.70
    粗腐殖质湿度码 Duff moisture code DMC 18.18 342.68 113.29 52.55
    干旱码 Drought code DC 61.91 660.71 373.32 98.28
    初始蔓延指数 Initial spread index ISI 1.44 15.63 10.01 1.99
    累积指数 Build-up index BUI 23.92 339.76 128.54 48.63
    火险天气指数 Fire weather index FWI 5.35 49.05 33.47 7.14
    火点 Fire point Fire 0 1 0.50 0.50
    注:各模型变量样本数为1 102。Note: sample number of each model variable is 1 102.
    下载: 导出CSV

    表  2  变量的多重共线性检验

    Table  2.   Multicollinearity test of variables

    变量 VariableWIN_avgWIN_maxSSDPRS_avgTmaxTminRH_avgRH_minFFMCISI
    VIF值 VIF value8.671.409.137.686.678.128.113.791.959.35
    下载: 导出CSV

    表  3  Step_LR模型参数拟合

    Table  3.   Parameter estimation of Step_LR model

    变量
    Variable
    估计系数
    Estimated coefficient
    标准误差
    Std error
    Z
    Z value
    P
    P value
    截距 Intercept −13.006 2.225 −2.923 0.003
    WIN_max 0.013 0.003 3.932 0.000
    Tmax 0.049 0.004 11.281 0.000
    Tmin −0.030 0.004 −8.448 0.000
    RH_avg −0.129 0.013 −9.903 0.000
    RH_min 0.055 0.020 2.733 0.006
    FFMC 0.077 0.023 2.091 0.037
    下载: 导出CSV

    表  4  基于贝叶斯后验概率的模型平均

    Table  4.   Model average based on Bayesian posterior probability

    Variablep! = 0SDModel 1Model 2Model 3Model 4Model 5
    Intercept 100 45.070 −6.028 82.170 −5.490 86.100 −3.648
    WIN_avg 1.3 0.002
    WIN_max 11.5 0.003
    SSD 5.9 0.004
    PRS_avg 23.4 0.005 −0.011 −0.012
    Tavg 26.8 0.021 0.048
    Tmax 98.0 0.015 0.054 0.054 0.028 0.063 0.046
    Tmin 26.6 0.011 −0.024
    VP_avg 95.6 0.023 −0.081 −0.082 −0.077 −0.102 −0.064
    RH_avg 19.7 0.036 −0.038
    RH_min 33.9 0.033 0.061
    Pre 0.0 0
    FFMC 0.3 0.004
    DMC 16.6 0.005
    DC 6.0 0.001
    ISI 0.0 0
    BUI 70.0 0.006 0.008 0.007 0.007 0.007 0.007
    FWI 8.3 0.013
    nVar 3 4 5 5 4
    BIC −7 039 −7 038 −7 038 −7 038 −7 037
    post prob 0.062 0.057 0.056 0.039 0.035
    注:本表为程序输出表格,其中Variable表示变量,p!=0为变量回归系数不为零的后验概率,SD为标准差,model 1 ~ model 5为BMA筛选的后验概率最大的5个模型,Intercept为截距项,从WIN_avg至FWI为各变量代码,参考表1,nVar为模型选中的变量数,BIC为贝叶斯信息量,post prob为模型后验概率。Notes: this table is the program output table, where Variables represents the model variables, P!=0 is the posterior probability that the regression coefficient of the variable is not zero; SD is the standard deviation; model 1−model 5 are the 5 models with the largest posterior probability screened by BMA; Intercept is the intercept item, and from Win_avg to FWI is the variable code, as shown in Tab. 1. nVar is the number of variables selected by the model, BIC is the Bayesian information criterion, and post prob is the posterior probability of the model.
    下载: 导出CSV

    表  5  Step_LR模型和BMA_LR模型中最终指标体系及预测准确率

    Table  5.   Final indicator system and prediction accuracy in the Step_LR and BMA_LR model

    模型 Model模型指标体系 Model index system预测准确率 Prediction accuracy/%
    训练集 Training sample (80%)测试集 Test sample (20%)
    Step_LR WIN_max, Tmax, Tmin, RH_min, RH_avg, FFMC73.371.8
    BMA_LR PRS_avg, Tavg, Tmax, Tmin, VP_avg, RH_avg, RH_min, BUI82.680.7
    下载: 导出CSV
  • [1] Rigo D D, Giorgio L, Durrant T H, et al. Forest fire danger extremes in Europe under climate change: variability and uncertainty[M]. Luxembourg: Publications Office of the European Union, 2017.
    [2] 田晓瑞, 宗学政, 舒立福, 等. ENSO事件对中国森林火险天气的影响[J]. 应用生态学报, 2020, 31(5):65−73.

    Tian X R, Zong X Z, Shu L F, et al. Impacts of ENSO events on forest fire weather of China[J]. Chinese Journal of Applied Ecology, 2020, 31(5): 65−73.
    [3] 白夜, 武英达, 贾宜松, 等. 2019—2020年澳大利亚气候异常与山火爆发的关系分析及应对策略[J]. 中国应急救援, 2020(2):23−27. doi: 10.3969/j.issn.1673-5579.2020.02.006

    Bai Y, Wu Y D, Jia Y S, et al. Link between climate anomaly and Australia bushfires in 2019−2020[J]. China Emergency Rescue, 2020(2): 23−27. doi: 10.3969/j.issn.1673-5579.2020.02.006
    [4] 赵凤君, 舒立福. 森林草原火灾扑救安全学[M]. 北京: 中国林业出版社, 2015.

    Zhao F J, Shu L F. Forest and grassland fire fighting safety[M]. Beijing: China Forestry Publishing House, 2015.
    [5] 岳超, 罗彩访, 舒立福, 等. 全球变化背景下野火研究进展[J]. 生态学报, 2020, 40(2):385−401.

    Yue C, Luo C F, Shu L F, et al. A review on wildfire studies in the context of global change[J]. Acta Ecologica Sinica, 2020, 40(2): 385−401.
    [6] Marlon J R, Bartlein P J, Gavin D G, et al. Long-term perspective on wildfires in the western USA[J]. Proceedings of the National Academy of Sciences of the United States of America, 2012, 109(9): 3203−3204.
    [7] Westerling A L. Increasing western US forest wildfire activity: sensitivity to changes in the timing of spring[J]. Philosophical Transactions of the Royal Society of London (Series B): Biological Sciences, 2016, 371: 1−10.
    [8] 潘登, 郁培义, 吴强. 基于气象因子的随机森林算法在湘中丘陵区林火预测中的应用[J]. 西北林学院学报, 2018, 33(3):175−183.

    Pan D, Yu P Y, Wu Q. Application of random forest algorithm on the forest fire prediction based on meteorological factors in the hilly area, central Hunan Province[J]. Journal of Northwest Forestry University, 2018, 33(3): 175−183.
    [9] North M P, Stephens S L, Collins B M, et al. Reform forest fire management[J]. Science, 2015, 349: 1280−1281. doi: 10.1126/science.aab2356
    [10] Fischer A P, Spies T A, Steelman T A, et al. Wildfire risk as a socioecological pathology[J]. Frontiers in Ecology and the Environment, 2016, 14(5): 276−284. doi: 10.1002/fee.1283
    [11] Zhang G, Wang M, Liu K. Forest fire susceptibility modeling using a convolutional neural network for Yunnan Province of China[J]. International Journal of Disaster Risk Science, 2019, 10(3): 386−403. doi: 10.1007/s13753-019-00233-1
    [12] Murphy T E, Tsang S W, Leo L S, et al. Bayesian model averaging for selection of a risk prediction model for death within thirty days of discharge: the silver-ami study[J]. International Journal of Statistics in Medical Research, 2019, 8: 1−7. doi: 10.6000/1929-6029.2019.08.01
    [13] Huang H, Liang Z, Li B, et al. Combination of multiple data-driven models for long-term monthly runoff predictions based on Bayesian model averaging[J]. Water Resources Management, 2019, 33(9): 3321−3338. doi: 10.1007/s11269-019-02305-9
    [14] 王倩, 师鹏飞, 宋培兵, 等. 基于贝叶斯模型平均法的洪水集合概率预报[J]. 水电能源科学, 2016(6):64−66.

    Wang Q, Shi P F, Song P B, et al. Multi-model ensemble flood probability forecasting based on BMA[J]. Water Resources and Power, 2016(6): 64−66.
    [15] 张畅, 陈新军. 海洋环境因子对澳洲鲐亲体补充量关系的影响: 基于贝叶斯模型平均法的研究[J]. 海洋学报, 2019, 41(2):99−106.

    Zhang C, Chen X J. The effect of environmental factors on stock-recruitment relationship of spotted mackerel-based on Bayesian model averaging method[J]. Haiyang Xuebao, 2019, 41(2): 99−106.
    [16] 李丽琴. 云南省森林火灾发生与气象因子之间的关系研究[D]. 北京: 北京林业大学, 2010.

    Li L Q. Study on the relationship between forest fires and the meteorological factors in Yunnan[D]. Beijing: Beijing Forestry University, 2010.
    [17] 周明昆, 王永平, 高月忠. 气象因子对云南大理森林火灾的影响[J]. 四川林业科技, 2012, 33(6):96−99. doi: 10.3969/j.issn.1003-5508.2012.06.022

    Zhou M K, Wang Y P, Gao Y Z. Effects of meteorological factors on forest fires in Dali, Yunnan[J]. Journal of Sichuan Forestry Science and Technology, 2012, 33(6): 96−99. doi: 10.3969/j.issn.1003-5508.2012.06.022
    [18] Martell D L, Otukol S, Stocks B J. A logistic model for predicting daily people-caused forest fire occurrence in Ontario[J]. Canadian Journal of Forest Research, 1987, 17(5): 394−401. doi: 10.1139/x87-068
    [19] 苏漳文, 刘爱琴, 郭福涛, 等. 福建林火发生的驱动因子及空间格局分析[J]. 自然灾害学报, 2016, 25(2):110−119.

    Su Z W, Liu A Q, Guo F T, et al. Driving factors and spatial distribution pattern of forest fire in Fujian Province[J]. Journal of Natural Disasters, 2016, 25(2): 110−119.
    [20] 于建龙, 刘乃安. 我国大兴安岭地区森林雷击火发生的火险天气等级研究[J]. 火灾科学, 2010, 19(3):131−137. doi: 10.3969/j.issn.1004-5309.2010.03.004

    Yu J L, Liu N A. Lightning-caused wildland fire weather danger rating in Daxing’anling region[J]. Fire Safety Science, 2010, 19(3): 131−137. doi: 10.3969/j.issn.1004-5309.2010.03.004
    [21] Bisquert M, Caselles E, Sánchez J M, et al. Application of artificial neural networks and logistic regression to the prediction of forest fire danger in Galicia using MODIS data[J]. International Journal of Wildland Fire, 2012, 21(8): 1025−1029. doi: 10.1071/WF11105
    [22] Oliveira S, Oehler F, San-Miguel-Ayanz J, et al. Modeling spatial patterns of fire occurrence in Mediterranean Europe using multiple regression and random forest[J]. Forest Ecology and Management, 2012, 275(4): 117−129.
    [23] 陈岱. 基于Logistic回归模型的大兴安岭林火预测研究[J]. 林业资源管理, 2019(1):116−122.

    Chen D. Prediction of forest fire occurrence in Daxing’an Mountains based on logistic regression model[J]. Forest Resources Management, 2019(1): 116−122.
    [24] Raftery A E, Gneiting T, Balabdaoui F, et al. Using Bayesian model averaging to calibrate forecast ensembles[J]. Monthly Weather Review, 2005, 133(5): 1155−1174. doi: 10.1175/MWR2906.1
    [25] 梁慧玲, 林玉蕊, 杨光, 等. 基于气象因子的随机森林算法在塔河地区林火预测中的应用[J]. 林业科学, 2016, 52(1):89−98.

    Liang H L, Lin Y R, Yang G, et al. Application of random forest algorithm on the forest fire prediction in Tahe Area based on meteorological factors[J]. Scientia Silvae Sinicae, 2016, 52(1): 89−98.
    [26] 顾先丽, 吴志伟, 张宇婧, 等. 气候变化背景下江西省林火空间预测[J]. 生态学报, 2020, 40(2):667−677.

    Gu X L, Wu Z W, Zhang Y J, et al. Prediction research of the forest fire in Jiangxi Province in the background of climate change[J]. Acta Ecological Sinica, 2020, 40(2): 667−677.
    [27] Chang Y, Zhu Z L, Bu R C, et al. Predicting fire occurrence patterns with logistic regression in Heilongjiang Province, China[J]. Landscape Ecology, 2013, 28(10): 1989−2004. doi: 10.1007/s10980-013-9935-4
    [28] Guo F T, Su Z W, Wang G Y, et al. Understanding fire drivers and relative impacts in different Chinese forest ecosystems[J]. Science of the Total Environment, 2017, 605: 411−425.
    [29] Flannigan M D, Krawchuk M A, Groot W J D, et al. Implications of changing climate for global wildland fire[J]. International Journal of Wildland Fire, 2009, 18(5): 483−507. doi: 10.1071/WF08187
    [30] Loepfe L, Rodrigo A, Lloret F. Two thresholds determine climatic control of forest fire size in Europe and northern Africa[J]. Regional Environmental Change, 2014, 14(4): 1395−1404. doi: 10.1007/s10113-013-0583-7
    [31] 蔡奇均, 曾爱聪, 苏漳文, 等. 基于Logistic回归模型的浙江省林火发生驱动因子分析[J]. 西北农林科技大学学报, 2020, 48(2):108−115.

    Cai Q J, Zeng A C, Su Z W, et al. Driving factors of forest fire in Zhejiang Province based on logistic regression model[J]. Journal of Northwest A&F University, 2020, 48(2): 108−115.
  • 加载中
图(4) / 表(5)
计量
  • 文章访问数:  444
  • HTML全文浏览量:  162
  • PDF下载量:  52
  • 被引次数: 0
出版历程
  • 收稿日期:  2020-06-19
  • 修回日期:  2021-01-07
  • 网络出版日期:  2021-04-21
  • 刊出日期:  2021-05-27

目录

    /

    返回文章
    返回