基于随机森林算法的中国西南地区林火发生预测模型构建及驱动因子  

Construction of forest fire prediction model and driving factors analysis based on random forests algorithm in Southwest China

作　　者：何锐 陆恒 晋子振 秦艳[5] 杨弘毅 刘治银 杨广睿 徐进滢 龚雪 赵求东 HE Rui;LU Heng;JIN Zizhen;QIN Yan;YANG Hongyi;LIU Zhiyin;YANG Guangrui;XU Jinying;GONG Xue;ZHAO Qiudong(Key Laboratory of Ecohydrology of Inland River Basin,Northwest Institute of Eco-Environment and Resources,Chinese Academy of Sciences,Lanzhou 730000,China;College of Geography and Resources Sciences,Sichuan Normal University,Chengdu 610000,China;University of Chinese Academy of Sciences,Beijing 100049,China;State Key Laboratory of Cryospheric Science,Northwest Institute of Eco-Environment and Resources,Chinese Academy of Sciences,Lanzhou 730000,China;College of Resources and Environment Sciences,Xinjiang University,Urumqi 830046,China;School of Geographical Science,Nanjing Normal University,Nanjing 210023,China)

机构地区：[1]中国科学院西北生态环境资源研究院内陆河流域生态水文重点实验室,兰州730000 [2]四川师范大学地理与资源科学学院,成都610000 [3]中国科学院大学,北京100049 [4]中国科学院西北生态环境资源研究院冰冻圈科学国家重点实验室,兰州730000 [5]新疆大学资源与环境科学学院,乌鲁木齐830046 [6]南京师范大学地理科学学院,南京210023

出　　处：《生态学报》2023年第22期9356-9370,共15页Acta Ecologica Sinica

基　　金：教育部人文社科青年项目(21YJCZH099);四川师范大学地理与资源科学学院创新训练项目(202010636326)。

摘　　要：林火直接破坏森林资源,改变森林的结构与功能,影响局地甚至全球气候状况并威胁人类生命和财产安全,在气候变暖背景下林火将更加频发,因此开展林火预测/预报研究至关重要。利用MODIS(Moderate-resolution Imaging Spectroradiometer)的温度异常/火产品(MOD14A1)获取逐日林火数据,分析了2001—2018年中国西南地区林火时空分布特征;采用随机森林算法,综合考虑气象、地形、可燃物状况及植被等林火驱动因子,构建了中国西南地区干、湿季林火发生预测模型,系统分析了西南地区干湿季林火发生的主要驱动因子。结果表明:(1)中国西南地区林火主要集中分布于云南大部、四川西南部及贵州南部地区,并呈集聚分布特征;林火多发于干季,占林火发生总次数的96.5%,年林火发生次数呈阶段性变化特征,2001—2014年呈现显著增加趋势,随后表现为不显著减少趋势;(2)构建的干、湿季林火发生预测模型能较准确地模拟林火发生状况:训练期模型准确率分别处于82.94%—83.99%与85.12%—90.31%之间,AUC(Area Under Curve)值分别处于0.908—0.914与0.922—0.965之间;测试期模型准确率分别为79.73%和83.27%,AUC值分别为0.886和0.855;(3)海拔是西南地区林火发生最关键的限制因子,林火多集中于中海拔区,而在低海拔和高海拔地区林火不易发生,这与人类活动密切相关。当日的气象条件是干季林火发生次重要的驱动因子,可燃物的温湿度状况则是湿季林火发生次重要的驱动因子。FWI系统指标(Fire Weather Index)在西南地区有较好的适用性且对于区域干湿季林火发生均有重要的影响,因此在西南地区林火预测/预报工作中有必要引入FWI系统指标。Forest fires directly destroy forest resources,change the structure and function of forests,affect local and even global climate conditions and threaten human life and property.And forest fire will be more frequent in the context of global warming,so the study of forest fire prediction/forecast is vital.We obtained the daily forest fire data from the Thermal Anomalies/Fire Daily L3 product(MOD14A1)of Moderate-resolution Imaging Spectroradiometer(MODIS)and analyzed the spatial and temporal distribution characteristics in Southwest China during 2001—2018.The random forests algorithm was adopted to construct the dry and wet season′s forest fire prediction models with full consideration of driving factors(contained meteorological elements,topography,fuel and vegetation).Then,we identified the main driving factors of forest fire occurrences through simulation in dry and wet seasons in Southwest China.The following major conclusions were drawn:(1)The forest fires were mainly concentrated in the most regions of Yunnan Province,southwestern Sichuan Province and southern Guizhou Province,and showed a significant agglomeration distribution.The forest fires mostly occurred in the dry season accounting for 96.5% of the total forest fires.The number of annual forest fires showed a staggered transformation with a significant increase trend during 2001—2014 and non-significant decrease trend after 2014.(2)The developed random forests algorithm-based forest fire prediction model archived good performances.The accuracy of the model in the training period were between 82.94%—83.99% and 85.12%—90.31%,and the Area Under Curve(AUC)values were between 0.908—0.914 and 0.922—0.965,respectively;the accuracy of the model in the testing period were 79.73% and 83.27%,and the AUC values were 0.886 and 0.855,respectively.(3)The elevation was the most important limiting factor to forest fire occurrences for both dry and wet seasons in Southwest China.The forest fires were concentrated in the midaltitude areas,while they were less

关 键 词：林火预测模型 随机森林算法 林火驱动因子 FWI系统 西南地区 

分 类 号：S762[农业科学—森林保护学]