Automatic Smoke Detection in MODIS Satellite
Data based on
K
-means Clustering and Fisher
Linear Discrimination
Xiaolian Li, Jing Wang, Weiguo Song, Jian Ma, Luciano Telesca, and Yongming Zhang
Abstract
Satellite-based remote sensing technique provides images to
detect and monitor forest fire smoke. Aiming at automatical-
ly separating smoke plumes from other cover types, several
bands of the Moderate Resolution Imaging Spectroradiometer
(MODIS) onboard the Terra/Aqua satellites were selected.
A smoke identification algorithm that integrates K-means
clustering and Fisher Linear Discrimination was developed.
It’s evaluated that the algorithm can identify more than 98
percent of the smoke pixels by using the k-folds cross-val-
idation technique. Then, the algorithm was validated in:
(a) Daxing’anling area (China) on 29 April 2009, (b) Amur
Region (Russia) on 29 April 2009, (c) Australia on 30 Septem-
ber 2011, and (d) Canada on 19 June 2013, in which several
fires occurred. By comparing the results with the grayscale
images, it can be seen that the algorithm has the capability
to capture heavy smoke as well as part of dispersed smoke.
The results suggest that the proposed algorithm can be
used as an innovative tool for detecting forest fire smoke.
Introduction
During a forest fire a large amount of smoke could be emitted,
which may have a tremendous impact on the regional envi-
ronment and long-term climate. In fact, since forest fire smoke
contains aerosol particles and greenhouse gases, such as CO
2
,
it contributes to the increase of the global concentration of
greenhouse gases in the atmosphere and affects the chemistry
of the troposphere (Crutzen and Andreae, 1990; Kaskaoutis
et al.
, 2011). Furthermore, the emitted smoke affects the local
climate because of the change of scatter and absorption of in-
coming solar radiation (Li, 1998; Randerson
et al.
, 2006). The
fire smoke even does harm to the public health (Chrysoulakis
et al.
, 2007; Li
et al.
, 2009).
Since smoke is the product of early forest fire, smoke detec-
tion can be fruitfully applied in early forest fire detection and
fire behavior analysis (Fromm
et al.
, 2000), as well as in small
and cool fires detection (Wang
et al.
, 2007).
Although satellite remote sensing provides high quality im-
ages for detecting fire smoke and estimating smoke properties,
accurate smoke identification is still a difficult task. This is due
to the reason that smoke and other cover types, clouds for in-
stance, show similar features in spectral bands of the satellite.
So far, one of the most commonly used approaches to
identify smoke was to assign different colors to different
bands or band combinations (Christopher and Chou, 1997;
Chrysoulakis and Opie, 2004; Chung and Le, 1984; Kaufman
et al.
, 1990; Randriambelo
et al.
, 1998), which resulted in a
true or false-color image. The image provides a visual separa-
tion of smoke from other scene types. Christopher
et al
. (1996)
employed false-color images to compute several textural
measures to visually separate smoke aerosols from other scene
types. This kind of approach seems promising to visually
identify smoke, however, it cannot be used in an automatic
smoke detection procedure. Trying to develop an automatic
and accurate smoke detection procedure, multi-threshold
methods were introduced. This method is based on the phys-
ical properties of the different scene types such as clouds,
vegetation, and water. Smoke can thus be automatically
classified by several threshold tests. Baum and Trepte (1999)
used a grouped threshold method for scene identification
with
NOAA
/
AVHRR
measurements. Randriambelo
et al
. (1998)
applied multi-spectral methods for fire smoke plume detec-
tion in south-eastern Africa and Madagascar. Xie
et al
. (2007)
developed a multi-threshold method for detecting smoke
plumes based on the analysis of spectral characteristics of dif-
ferent cover types. Zhao
et al.
(2010) used satellite imagers for
smoke detection based on spectral and spatial threshold tests.
Li
et al.
(2001) applied
AVHRR
imagery to automatic smoke
detection by using neural networks and multi-threshold.
All these multi-threshold methods are effective in smoke
detection, but present the difficulty in finding fixed thresholds
for images acquired in different seasons and regions. Chry-
soulakis
et al.
(2003 and 2007) proposed a multi-temporal
change detection approach using two images in the same area
acquired at different times, and used the Normalized Dif-
ference Vegetation Index
(
NDVI
) and the infrared radiance to
detect the core of plume; by enlarging such core, the complete
area covered by the plume was detected. But the limited
spatial resolution and the lack of a sufficient number of
AVHRR
channels (Gong
et al.
, 2006) limited the retrieving of plume
spectral characteristics (Li
et al.
, 2001; Chrysoulakia
et al.
,
2007).
MODIS
sensor, on the contrary, has 36 channels covering
the spectrum from the visible to the far infrared; therefore, an
Xiaolian Li, Weiguo Song and Yongming Zhang are with the
State Key Laboratory of Fire Science, University of Science
and Technology of China, Jinzhai 96, 230027 Hefei, Anhui,
China (
).
Jing Wang is with the China Ship Building Industry Cor-
poration NO.719 Research Institute, Zhongshan Road 450,
Wuchang District, 430064 Wuhan, China.
Jian Ma is with the National United Engineering Laboratory of
Integrated and Intelligent Transportation, School of Transpor-
tation and Logistics, Southwest Jiaotong University, 610031
Chengdu, China.
Luciano Telesca is with the Institute of Methodologies For
Environmental Analysis, National Research Council of Italy,
Tito (PZ), 85050, Italy.
Photogrammetric Engineering & Remote Sensing
Vol. 80, No. 10, October 2014, pp. 971–982.
0099-1112/14/8010–971
© 2014 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.80.10.971
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
September 2014
971
