﻿ 基于机器学习对森林火灾的预测分析 Prediction and Analysis of Forest Fire Based on Machine Learning

Statistics and Application
Vol.05 No.02(2016), Article ID:17920,9 pages
10.12677/SA.2016.52016

Prediction and Analysis of Forest Fire Based on Machine Learning

Dan Liu

College of Statistics and Mathematics, Yunnan University of Finance and Economics, Kunming Yunnan

Received: Jun. 8th, 2016; accepted: Jun. 27th, 2016; published: Jun. 30th, 2016

ABSTRACT

Forest fire is a kind of destructive and huge disaster, which causes irreparable damage in the ecological environment and brings great harm to human survival and life. Especially since the 1980s, the global warming has continued, and forest fires occur more frequently, leading to huge economic losses to the world each year. So how to predict, prevent or reduce the hazards of forest fires become the common concern of many science disciplines. Rapid detection is an effective way to predict forest fire. To achieve this goal, one approach is to use automated tools based on sensor data, such as the data that meteorological stations offer. The study found that the meteorological conditions (such as temperature, wind speed) are important factors influencing forest fires and some fire indicators (such as forest fire weather index). Therefore, we will explore several machine learning methods to predict forest fire area. Using the data collected from Montesinho National Park in Northeastern Portugal, and a variety of different machine learning techniques, such as support vector machines (SVM) and random forests, four different characteristics (distribution of space, time, climate indicators and FWI system indicator) were analyzed. The best results were obtained using support vector machines and four basic meteorological inputs (such as temperature, relative humidity, wind speed and precipitation), which could accurately predict the damage area of small-scale and frequent fires. The above prediction methods are of great significance for improving the management and allocation of fire-fighting resources.

Keywords:Forest Fire, Machine Learning, Support Vector Machine, Random Forest

1. 引言

2. 数据分析

2.1. 数据介绍

2.2. 变量解释

FWI系统是由6个部分组成：3个代表可燃物湿度的基本子指数，分别为细小可燃物湿度码(FFMC, fine fuel moisture code)，粗腐殖质湿度码(DMC, duff moisture code)和干旱码(DC, drought code)；2个代表可燃物扩散速率和消耗率的中间子指数，分别为初始蔓延速度(ISI, initial spread)和累积指数(BUI, build up)；1个代表火强烈程度的最终指数，FWI。火险气候指数系统中所涉及的元素由每天测量的气温、相对湿度、风速和降水量的气象数据中计算得到。

2.2.1. 细小可燃物湿度码FFMC

FFMC代表的是森林中地被物干质量为0.25 kg∙m−2，厚度为1.2 cm的枯枝落叶和其他的已经固化的细小燃料的含水率。FFMC是代表细小可燃物的可燃性和易燃性的指标，它受温度、降水、相对湿度和风速的影响，值随着燃料含水率的变化而改变，其核心是一个简单的水分交换的指数模型：。其中为前一天的燃料含水率。

2.2.2. 粗腐殖质湿度码DMC

DMC代表的是森林地被物最上层厚度约为7 cm，干质量为5.00 kg∙m−2的有机物质的含水率。DMC用来表明中等下层落叶层和中型木质物质的燃料消耗，DMC模型是一个简单的水分交换的指数模型：

。其中表示前一天的地表可燃物含水率。

2.2.3. 干旱码DC

DC代表的是森林地被物中干质量为25.00 kg∙m−2，厚度为18 cm的深层可燃物和粗死木残体的含水率。干旱码用于衡量季节性干旱对森林燃料以及深层下层落叶层和大型段木的影响指标。DC模型的核心是一个简单的指数模型：。其中表示前一天干旱码的湿度指标。

2.2.4. 初始蔓延指数ISI

ISI代表的是火灾蔓延的潜在等级，由FFMC和风速两个指标决定。ISI一直是表示火灾蔓延等级的很好指标。

3. 模型描述

3.1. 多元线性回归模型

(1)

(2)

3.2. 决策树(DT)模型

3.3. 支持向量机(SVM)模型

3.4. 随机森林模型

4. 实验验证及结果

4.1. 线性相关性分析

4.2. 多元线性模型

Table 1. Table of linear correlation coefficient of meteorological factors

Figure 1. Function diagram of data among variables

Figure 2. Scatter plots of climate variables on FFMC

4.3. 决策树和随机森林

Figure 3. Decision tree

Figure 4. Importance of random forest variables

4.4. 机器学习方法

Logistic回归，决策树，人工神经网络回归，随机森林，支持向量机和K近邻等的实验结果见表3

5. 结论

Table 2. Half off cross validation results

Table 3. Machine learning half off cross validation results

temp起主要作用，RH其次，wind再次之，而rain起作用最小，虽然检测模型的误判率一度达到40%，但我们仍觉得在森林火灾的发生很大程度上取决于temp，控制temp的临界值(或者临界区间)可以很好的预防森林火灾的发生。

Prediction and Analysis of Forest Fire Based on Machine Learning[J]. 统计学与应用, 2016, 05(02): 163-171. http://dx.doi.org/10.12677/SA.2016.52016

1. 1. Cortez. P. and Morais, A. (2007) A Data Mining Approach to Predict Forest Fires Using Meteorological Data. http://www3.dsi.uminho.pt/pcortez/fires.pdf

2. 2. 曲智林, 胡海清. 基于气象因子的森林火灾面积预测模型[J]. 应用生态学报, 2007, 18(12): 2705-2709.

3. 3. 袁建, 江洪, 信晓颖. 基于FWI的浙江省森林火险等级划分[J]. 福建农林大学学报: 自然科学版, 2013, 42(3): 283-288.

4. 4. 田晓瑞, Douglas J. McRae, 舒立福, 赵凤君, 王明玉. 大兴安岭地区森林火险变化及FWI适用性评估[J]. 林业科学, 2010, 46(5): 127-132.

5. 5. 王正旺, 庞转棠, 魏建军, 杨艳萍, 杨梅红. 森林火险天气等级预测及火情检测应用[J]. 自然灾害学报, 2006, 15(5): 154-161.

6. 6. 牛若芸, 翟盘茂, 孙明华. 森林火险气象指数及其构建方法回顾[J]. 气象, 2006, 32(12): 3-9.