PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
January 2017
19
Fire Risk Prediction Using Remote Sensed
Products: A Case of Cambodia
Bo Yu, Fang Chen, Bin Li, Li Wang, and Mingquan Wu
Abstract
Forest fire is threatening human life in monsoon countries,
such as Cambodia, which suffers from forest fire frequently.
Developing an efficient method to predict fire risk for large
areas is becoming significantly important. However, the
methods used in fire risk prediction are mostly based on field-
based meteorological data, and the coefficients are hard-de-
fined, heavily depending on user experience. We propose
to use a user-friendly machine learning method, Random
Forest™, to train a regression model by synthesizing publicly
available remote sensed products to predict fire risk ratings at
pixel-level in eight-day advance. The structure of our model
synthesizes features in three-time intervals T1, T2, and T3
to predict fire occurrence probability in T4. The experiment
demonstrates the efficiency of such model in predicting fire
occurrence with a correlation coefficient of 0.987 and mean
square error being 0.00285. It results in a practical way to
predict fire risk and prevent fires.
Introduction
Fire risk influences human life, ecosystem health, and climate
change in terms of forest degradation (Kandya
et al.
,1998),
greenhouse gasses release (Flannigan
et al.
, 2009) and firefight-
er safety hazard (Kales
et al.
, 2007). Large fires damage human
settlements (Venevsky
et al.
, 2002) and depress both adults
and children psychologically (Jones
et al.
, 2002). Fire risk has
become a threat for economics for having caused 2.4 billion
USD
per year between years 2002 and 2012 (Chatenoux and
Peduzzi, 2013). The smoke, particulate matter (
PM
) and carbon
monoxide (CO) released from fire also bring about deadly
diseases, such as lung cancer, heart and breathing problems
(Chowdhury and Hassan, 2015; Forsberg
et al.
, 2012; Mangan,
2007; Stefanidou
et al.
, 2008). Besides the negative impacts,
forest fire has some positive impacts as well. Wildfires occur
naturally to promote forest regeneration and control disease
and insect infestations (Pausas and Paula, 2012). Excessive
suppression of fire causes forest overgrowth, which in turn
increase the risk, intensity, and duration of a wildfire (Henion
and Luo, 2013 and 2016). Moreover, because of human ex-
pansion and global warming, the frequency of fire occurrence
is increasing (Loehman
et al.,
2011). Therefore, developing a
reliable and robust fire risk management system is necessary
to maximize the advantages and suppress the disadvantages of
fires, by understanding the behavior and impact of a fire.
To the best of our knowledge, fire risk monitoring meth-
ods can be point-wise meteorological data-based and remote
sensed data-based methods (Chowdhury and Hassan, 2015).
The meteorological data-based methods have been widely
used in operational systems in Canada, Russia, the United
States, and Australia (Burgan
et al.
1997; Deeming
et al.
1977; Luke and McArthur, 1978; Nesterov, 1949; Stocks
et
al.
, 1989). The remote sensed data-based methods are used to
describe large-scale fires and their locations which are often
unreachable for human beings.
Generally, there are four main operating systems for fire
risk monitoring. They are Fire Weather Index (
FWI
) System in
Canada (Stocks
et al
., 1989), McArthur’s Forest Fire Danger
Rating System (
FFDRS
) (Luke and McArthur, 1978), Russian
Nesterov Index (Nesterov, 1949) and the Wildland Fire Assess-
ment System (
WFAS
) in the United States (Burgan
et al
., 1997).
All of them have been widely used for quite a long time.
FWI
is
the best established for having been adapted in many coun-
tries (Alexander and Cole, 2001, Lee
et al.
, 2002), such as New
Zealand (Alexander and Fogarty, 2002), Alaska (Alexander
and Cole, 2001), Mexico (Lee, et al., 2002), Argentina (Taylor,
2001), Sweden (Granstrom and Schimmel, 1998), Portugal
(San-Miguel-Ayanz
et al.
,2003), Spain (Viegas
et al., 2000
), and
Indonesia (de Groot
et al.
, 2007). The strong adaptive ability
of
FWI
owes to its simple input features: humidity, tempera-
ture, precipitation, and wind speed (Stocks
et al
.,1989).
FFDRS
predicts the spread rate of fires by using the record of rainfall,
evaporation, wind speed, temperature, and humidity as input
variables (Dowdy and Scientific, 2009). The Russian Nesterov
Index is a simple index for evaluating fire risk (Nesterov, 1949),
and has been widely used in Russia. The index requires sever-
al meteorological variables, such as precipitation, temperature,
wind speed, and relative humidity.
WFAS
is developed from
the National Digital Forecast Database (
NDFD
). Its fire risk rating
level considers current and antecedent weather, fuel types, and
both live and dead fuel moisture (Deeming
et al
., 1977).
The point-wise meteorological data-based operating systems
above are well established, but suffer from several disadvan-
tages which hinder the development of fire risk rating system:
(a) The input meteorological data are limited by the distribu-
tion and positions of meteorological stations, especially for
high-risk locations and less-developed countries (Chowdhury
and Hassan, 2015; Hijmans
et al.
, 2005); and (2) Data from
different meteorological stations need to be interpolated to
form a spatial map for further analysis, but different interpola-
tion methods would generate different spatial maps (Longley,
2005). The corresponding fire risk ratings from one operating
system would be different, limiting the adaptive ability of the
Bo Yu and Bin Li are with the Key Laboratory of Digital Earth
Science, Institute of Remote Sensing and Digital Earth, Chi-
nese Academy of Sciences, Beijing 100101, China.
Fang Chen is with the Key Laboratory of Digital Earth Science,
Institute of Remote Sensing and Digital Earth, Chinese Acade-
my of Sciences, Beijing 100101, China. Email: chenfang@radi.
ac.cn; Hainan Key Laboratory of Earth Observation, Institute of
Remote Sensing and Digital Earth, Chinese Academy of Scienc-
es, Sanya 572029, China; and University of Chinese Academy
of Sciences, Beijing 100049, China
Li Wang and Mingquan Wu are with the State Key Laboratory of
Remote Sensing Science, Institute of Remote Sensing and Digital
Earth, Chinese Academy of Sciences, Beijing 100101, China.
Photogrammetric Engineering & Remote Sensing
Vol. 83, No. 1, January 2017, pp. 19–25.
0099-1112/17/19–25
© 2017 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.83.1.19
