belonging probability inverse image approach for forest fire detection
TRANSCRIPT
Belonging probability inverse image approach for forestfire detection
Souleyman Benkraouda, Benabdellah Yagoubi*, Mustapha Rebhi and AhmedBouzianeLaboratoire Electromagn�etisme et Optique guid�ee., Laboratoire Signaux et Syst�emes., D�epartement G�enie Electrique., Facult�e des sciences et sciencesde l’ing�enieur., Universit�e de Abd El Hamid Ibn Badis, Mostaganem, 27000, Algeria
Abstract
We present a method for early forest fire detection from a
satellite image using the belonging probability matrix
image. We have considered each satellite image matrix line
as a realization of a nonstationary random process in the
thermal infra-red (TIR) spectral band and then divided each
line into very small stationary and ergodic intervals to
obtain an adequate mathematical model. Furthermore, the
pixels of the satellite image are considered to be statistically
independent; thus, any small interval of each line behaves,
naturally, as a Gaussian stationary noise. In this work, we
have, therefore, selected the latter as a mathematical model
formodelling these intervals of a satellite imagewithout fire,
and then, we have determined the parameters of this
Gaussian realization. So, when a fire occurs in this forest
zone, we can use these parameters to calculate its belonging
probability to the original image without fire. This proba-
bility should be very small because the fire, in any forest, can
be considered as a rare event. As a consequence, we have
presented a matrix image of the probability inverse of each
interval for a better fire detection observation.
Key words: belonging probability matrix image, forest fire,
satellite image, thermal infra-red spectral band
R�esum�e
Nous pr�esentons une m�ethode de d�etection pr�ecoce des
feux de forets par image satellite en utilisant l’image
matricielle de la probabilit�e d’appartenance. Nous avons
consid�er�e chaque ligne matricielle de l’image satellite
comme une r�ealisation d’un processus al�eatoire non
stationnaire dans la bande spectrale TIR (Thermal Infra-
Rouge), puis divis�e chaque ligne en tr�es petits intervalles
stationnaires et ergodiques, afin d’obtenir un mod�ele
math�ematique ad�equat. Ensuite, les pixels de l’image
satellite sont consid�er�es comme statistiquement ind�epen-
dants, et donc chaque petit intervalle de chaque ligne se
comporte, naturellement, comme un bruit gaussien sta-
tionnaire. Dans ce travail, nous avons donc s�electionn�e ce
dernier comme mod�ele math�ematique pour mod�eliser ces
intervalles d’une image satellite sans feu et nous avons
d�etermin�e les param�etres de cette r�ealisation gaussienne.
Ainsi, lorsqu’un feu survient dans cette zone de foret, nous
pouvons utiliser ces param�etres pour calculer sa probab-
ilit�e d’appartenance �a l’image originale sans feu. Cette
probabilit�e doit etre tr�es faible puisque le feu, dans toute
foret, peut etre consid�er�e comme un �ev�enement rare. Par
cons�equent, nous pr�esentons une image matricielle de
l’inverse de la probabilit�e de chaque intervalle pour une
meilleure observation de d�etection des feux.
Introduction
Forest fire is one of the most important factors that affect the
earth ecosystem and contribute to the global warming. Its
early detection is, therefore, very important to limit any
further damage. Early detection is usually performed by
surveillance either by human observers located at different
places in the forest or by video systems. These kinds of
surveillances are, however, not always efficient. Alternative
surveillances which can be performed by satellite were,
therefore, proposed. In this case, many different methods,
mostly based on radiometric analysis of the satellite image
in the thermal infra-red (TIR) spectral domain (Den Breejen
et al., 1998, pp. 517–532), have been suggested to improve
its early detection. Analysing directly the visual TIR satellite
image does not, however, provide always good results. This*Correspondence: E-mail: [email protected]
© 2013 John Wiley & Sons Ltd, Afr. J. Ecol., 52, 363–369 363
is due to the fact that the image observation accuracy may
be degraded by the clouds and humidity in many damp
regions of the world, whereas in dry regions, such as in the
north of Africa, the accuracy should be less degraded.
To detect forest fires, many algorithms based on statis-
tical methodology and random fields have been suggested.
Among them, worthy of mention are those suggested by
Lafarge et al. (2007a,b).
Despite the interesting results provided by such algo-
rithms, difficulties in their implementation and processing
hampering the performance may be found in each of them.
Therefore, we suggest in this work a method based on the
belonging probability of a rare event to a Gaussian
realization based on simple matrix line segmentation
instead of a Gaussian field realization. As it is well known,
in any problem of modelling, the choice of the model is a
very important task. So it is important for the images to be
characterized by a mathematical model to carry out any
kind of processing. The image can be modelled by a classic
monodimensional model, and consequently, we can use all
the tools and the mathematical analysis techniques of the
monodimensional signals which are well developed in the
literature. This mode of representation has been applied for
the line-by-line processing of the images, in particular for
the coding, the filtering and the storage (Zhong & Sclaro,
2003, pp. 44–50). In the first place, we selected a type of
model with reasonable results that can be applied to a
possible large number of images. After many tests on
several types of models, we have decided to choose the
Gaussian model, mainly because of its simple computing
and the good results obtained. This is basically a conse-
quence of the central limit theorem, which is one of the
most important concepts in statistics and probability. In its
simplest form, this theorem applied to a natural image,
indicating that the pixels’ intensity becomes normally
distributed as more and more pixels are observed. In the
following, we start our discussion by a highlight on
random processes and how they are used for the image
analysis and then apply them to a satellite image, and to
show, particularly, how to detect fires from the image of
the inverse of the belonging probability. In our case, we
apply a Gaussian stationary noise, which is a stationary
random process with noncorrelated samples.
Material and methods
The image shown in Fig. 1 with fire occurred in Californian
forest zone in July 2006 is provided by the Landsat7 polar
orbiting satellite in the TIR spectral band range 10.4–
12.5 lm and is being used for assessing the applicability of
our suggested method to forest fire detection. This image
was acquired at night with a resolution of 60 9 60 m and
showing the wildfires in the short wavelength infra-red
(SWIR) bands, whereas the terrain is visible only in the TIR
band. The software used to perform the image processing is
the image toolbox of Scilab5.3.3, INRIA, Scalay (Paris,
France) which is free software.
The proposed method for forest fire detection is described
as follows:
1 Determine the mean value M of the entire image matrix
of the forest region of interest without fire.
2 Segment each line of this matrix into small stationary
patches and compute their variances using thewholematrix
mean value M to model each interval as a Gaussian model.
3 Consider any pixel with intensity below 70% of the
maximum intensity to be equal 99%M to avoid false alarm.
4 When a fire takes place in this zone, use these previously
estimated Gaussian parameters for the image without a fire
to calculate the inverse of the belonging probability of each
interval to belong to this region of interest with a fire.
5 Arrange the inverses of these probabilities in a matrix
form as their corresponding intervals to observe them as
an image for a quick detection.
Results and discussion
We first survey the segmentation of a random process,
which motivated the use of a Gaussian stationary noise
Fig 1 Thermal infra-red satellite image of forest with fires
(California, U.S.A.). Available at http://landsat.usgs.gov/gal-
lery_view.php?category=nocategory&thesort=pictureId
© 2013 John Wiley & Sons Ltd, Afr. J. Ecol., 52, 363–369
364 S. Benkraouda et al.
model to represent any small interval of each satellite
image matrix line. This Gaussian stationary noise is
deemed a particular case of a random process, which is
reviewed in the following section.
Random process
By repeating a same experiment infinitely in time, we
obtain an infinite number of random curves representing
this experiment. These curves represent a random process
evolution in time. The auto-correlation is, usually, used to
compare two states of the same process; its value indicates
how much two states at different instants of a process are
correlated. For the random process, the auto-correlation is
calculated by the inner product in Hilbert space using the
mathematical expectation (Scharf, 1991; Krabs, 1995;
Hwang & Brown, 1997; Kay, 1998; Yagoubi, 2011) as
follows:
/Xði; jÞ ¼ E½Xi � Xj� (1)
where Xi and Xj are two vectors of the process at two
different instants i and j. If the elements of any parallel
diagonal to the principle of the auto-correlation matrix are
different, then the process is nonstationary. An example of
a nonstationary process may be represented by the curves
(o,., *, and +) of Fig. 2. We can see that the process in the
intervals [0, 25] and [25, 50] do not look the same; hence,
the auto-correlation is not constant for the intervals with
the same length. However, if we focus on a small space or
on a small interval such as [20, 25] and [45, 50] of the
same process, then the curves may not vary too much and
thus the auto-correlation could be considered as approx-
imately constant for the same small intervals. It is possible,
therefore, to divide the process into many small stationary
intervals. Furthermore, the smaller these intervals, the
more stationary and ergodic they will be, because their
mean value calculated vertically can be the same as that
obtained horizontally on any curve in these intervals. This
mean value that can be calculated by the mathematical
expectation is, therefore, almost constant E[Xi] = cte = mx.
If, in addition, the ergodic process samples are indepen-
dent, as those of the satellite image, then each small
interval of any image matrix line behaves as a stationary
noise and can thus be described by the following auto-
correlation relation in Hilbert space:
/wði; jÞ ¼ r2dij (2)
in which r2 is the constant representing the process
variance or its power spectral density and dij is the unit
sample.
Image modelling method
Toobtain themathematicalmodel (Bracewell, 2003; Zang&
Sommer, 2007) for the original image without fire (Fig. 3),
we have segmented each matrix line into small stationary
intervals. To represent each of these intervals by a Gaussian
stationary noise (Rice, 1945; Hida &Hitsuda, 1993; Chan &
Fig 2 Nonstationary random process. Intervals [20, 25] and [45,
50] are smaller enough to be considered as stationary Fig 3 Real satellite image of forest without fires
© 2013 John Wiley & Sons Ltd, Afr. J. Ecol., 52, 363–369
Forest fire detection method 365
Zhou, 2010; Chang & Liu, 2010; Nakamori, 2010) as
described above, we have computed the Gaussian param-
eters the mean min and the variance r2in for each resulting
interval using the following expressions, respectively:
min ¼ 1
L
Xl�1
k¼l�L
xinðkÞ (3)
r2in ¼1
L
Xl�1
k¼l�L
ðxinðkÞ � minÞ2 (4)
where xi (l � L),…,xi (l � 1) are the values of the nth
(n = 1,2…N) interval of the ith (i = 1,2…I) image matrix
line, L is the interval length and l = n.L. In our case, the
intensity of the pixels representing the forest is roughly
uniformly distributed in the image without fire, and hence,
any pixel value, sufficiently deviated from the image
matrix mean value, can be considered as a rare event. This
matrix mean value is deemed the intensity threshold away
from which any pixel behaves as a rare event. Two
symmetrical pixels (hot and cold) with respect to this
matrix mean value have, however, the same Gaussian
probability. It is, therefore, important to fix a threshold
70% of the maximum grey level, below which all the pixels
are considered to be equal to 99% of the matrix mean
value to filter (illuminate) the cold zones and to avoid false
alarm. So, instead of computing the average for each
segment using Eq. 3, we calculate the whole matrix mean
value using the following expression of the image mean
value and consider it as the average for every segment.
M ¼ 1
I:J
XIi¼1
XJj¼1
xði; jÞ (5)
where x(i,j) denotes the (i,j) pixel intensity. Substituting (5)
for min in the expression (4) of the estimated variance for
each segment, we obtain
r2in ¼1
L
Xl�1
k¼l�L
ðxinðkÞ �MÞ2 (6)
The choice of the number N of intervals of each matrix line
depends on the quality of the reconstructed image (Fig. 4);
for higher N, we obtain a good representation for the
original image.
An original satellite image without fire is shown in Fig. 3,
and its corresponding reconstructed image using the
Gaussian model in Fig. 5. To show the accuracy of this
model, we have plotted three arbitrary original lines and
their corresponding reconstructed versions in Fig. 4. Once
the adequate Gaussian model of the image without fire has
been obtained, we save its parameters; variances andmeans
for each interval. These parameters, which characterize
Fig 4 Original (solid line) and reconstructed lines 3, 250 and 500, from top to bottom
© 2013 John Wiley & Sons Ltd, Afr. J. Ecol., 52, 363–369
366 S. Benkraouda et al.
the image, will be used in the following to calculate the
Gaussian belonging probability of any rare event to,
eventually, belong to this image.
The belonging probability method for the fire detection
The same forest zone (California (U.S.A.) [12]), as that of
the image given in Fig. 3 but with fire, is shown in Fig. 1,
and the image of the inverse of the belonging probability
image of these fires corresponding to the Gaussian reali-
zation is given in Fig. 6. It is shown in the latter that the
isolated spot on the right of the belonging probability
image and corresponding to the smallest fire detected by
our method is about two pixels. Because the landsat7
resolution is 60 9 60 m, the size of the smallest forest fire
that can be detected, using our algorithm, could be
<3600 9 2 = 7200 m2.
Because any starting fire is deemed a rare event in any
forest, its probability is typically smaller than that of the
background of the satellite image of a forest. The matrix X,
with I = 1200 lines and J = N.L = 1200 columns of the
image with fires, may be represented as follows:
X ¼
x1ð0Þx1ð1Þ � � � x1ðL� 1Þ|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}Lsamples
� � � x1ðN:L� 1Þ
x2ð0Þx2ð1Þ � � � x2ðN:L� 1Þ...
xIð0ÞxIð1Þ � � � xIðN:L� 1Þ
2666664
3777775
So, when a fire takes place in the same location of the
image without a fire, we can use the Gaussian parameters
previously estimated from the image without fire to
calculate the belonging probability of the intervals of the
above image matrix X to detect this fire. The Gaussian
probability pin of each stationary and ergodic interval with
statistically independent samples is given by the following
expression, which is, usually, applied to random and
statistically independent Gaussian variables:
pinðxiðl� LÞ; ; � � � ; xiðl� 1Þ; r2in;MÞ
¼ 1
rinffiffiffiffiffiffi2p
p� �L Yl�1
k¼l�L
exp �ðxiðkÞ �MÞ22r2in
!(7)
where r2in and M are, respectively, the variance and the
image matrix mean value already estimated from the
image without fires. So, instead of the probability pin, we
have rather calculated its inverse 1pin
using Eq. 7 of the
belonging probability pin of each interval n and arranged
them in a matrix form as their corresponding intervals of
the image matrix X with fires, as follows:
1
P¼
1p11
1p12
� � � 1p1N
1p21
1p22
� � � 1p2N
..
.
1pI1
1pI2
� � � 1pIN
266664
377775
This procedure, for visualizing early forest fire detection,
allows the smaller values of the probabilities to appear with
Fig 5 Reconstructed satellite image model without fires
Fig 6 Image of the inverse of the belonging probability
© 2013 John Wiley & Sons Ltd, Afr. J. Ecol., 52, 363–369
Forest fire detection method 367
higher intensity in the inverse of the belonging probability
matrix 1P image as shown in Fig. 6.
The good quality of the reconstructed image, obtained in
Fig. 5 with N = 300 intervals in each matrix line, after
many tests, indicates that the inverse of the belonging
probability shown in Fig. 6 is more accurate, and thus, it
has, approximately, the same spatial intensity distribution
as the real image in Fig. 1. This accuracy can be, further,
clarified by the mean square error (MSE) results between
the original and the reconstructed lines. The MSE values
are about 0.5 9 10�4, 0.8 9 10�4 and 10�4 for lines 3,
250 and 500, respectively. The plot of these three
reconstructed lines and their corresponding original ones
shown in Fig. 4, indicating that they are almost discon-
certed, which is a fairly firm indication that the Gaussian
parameters are highly well estimated. This belonging
probability procedure allows only the fires to appear in
Fig. 6, and the rest of the background remains dark. In
comparing the real image with fires (Fig. 1) to that of the
belonging probability (Fig. 6), we can say that the
advantage of using the belonging probability for fire
detection is that we can easily identify early fires in the
dark background as indicated by the small spot on the
right of the image (Fig. 6), while it is slightly confusing
and difficult to distinguish them from the background of
the real image (Fig. 1). Both the real satellite image and
the belonging probability image can be, furthermore, used
to reduce the probability of false alarm (Ollero et al., 1999,
pp. 123–131).
Although this study concerns Californian zone due to
the availability of the satellite image, we believe that our
forest fire detection method is well suitable to the North
African environment because it is very much similar to the
Californian environment. To cope with more diverse
vegetation communities, the Gaussian model used for
our forest fire detection method requires parameters
updating as well as adjustments in the thresholds to
accommodate other grasslands which are not, eventually,
similar to the Californian zone studied in this work.
Conclusion
We have estimated the belonging probability inverses of
the Gaussian stationary processes representing the real
satellite image and set them in an image form so that the
rare events, such as the fires, appear with higher intensity.
This procedure, for observing early forest fire detection,
can be, therefore, a good support to the direct observation
of the real satellite image. For forest fire detection, we
usually compare the image of a region of interest without
fire to the same one after a fire takes place. However, a
particular advantage of our method over this one is that
instead of restoring the whole image (1200 9 1200)
without fire, we restore only 301 values, N = 300 and M
(the image mean value), of the Gaussian model for
comparison when a fire takes place. This method can be
used for a real-time detection of a fire and it can be
incorporated into a surveillance system monitoring a zone
of interest in a forest for early fire detection.
Acknowledgements
The authors would like to thank the Algerian Government
for the financial support for this national project.
References
Bracewell, R. (2003) Fourier analysis and Imaging. Kluwer
Academic/Plenum Publishers, New York.
Chan, S.C. & Zhou, Y. (2010) On the performance analysis of a
class of transform-domain NLMS algorithms with gaussian
inputs and mixture gaussian additive noise environment. J.
Signal Process. Syst. 61, 1–17.
Chang, K.-M. & Liu, S.-H. (2010) Gaussian noise filtering from ECG
by Wiener filter and ensemble EMD. J. Signal Process. Syst. 61,
249–264.
Den Breejen, E., Roos, M., Schutte, K., De Vries, J.S. & Winkel, H.
(1998) Infrared measures of energy release and flame temperatures
of forest fires, Proceedings of third International Conference on
Forest Fire Research, Luso (Portugal) pp. 517–532.
Hida, T. & Hitsuda, M. (1993) Gaussian processes. AMS,
Providence.
Hwang, P.Y.C. & Brown, R.G. (1997) Introduction to Random
Signals and APPLIED KALMAN FILTERING: with Matlab
Exercises and Solutions. 3rd edn. John Wiley sons, Inc., New
York, NY.
Kay, S.M. (1998) Fundamentals of Statistical Signal Processing,
Detection Theory, vol. 2, Prentice–Hall, Englewood Cliffs, NJ.
Krabs, W. (1995) Mathematical Foundations of Signal Theory.
Heldermann Verlag, Berlin.
Lafarge, F., Descombes, X., Zerubia, J. & Mathieu, S. (2007a) Forest
fire detection by statistical analysis of rare events from thermical
infrared images. Traitement du Signal 24, 1–12.
Lafarge, F., Descombes, X., Zerubia, J. & Mathieu, S. (2007b)
Forest Fire Detection Based on Gaussian Field Analysis.
European Signal Processing Conference (EUSIPCO), Poznan,
Poland.
Nakamori, S. (2010) Design of RLS Wiener FIR filter using
covariance information in linear discrete-time stochastic
systems. Digital Signal Process. 20, 1310.
© 2013 John Wiley & Sons Ltd, Afr. J. Ecol., 52, 363–369
368 S. Benkraouda et al.
Ollero, A., Arru′e, B.C., Martı′nez, J.R. & Murillo, J.J. (1999)
Techniques for reducing false alarms in infrared
forest-fire automatic detection systems, Control Eng. Pract. 7,
123–131.
Rice, S.O. (1945) Mathematical analysis of random noise. Bell Sys.
Tech. J., 24, 46–156.
Scharf, L.L. (1991) Statistical Signal Processing: Detection,
Estimation, and Times Series Analysis, Addison–Wesley, New
York, NY.
Yagoubi, B. (2011) A geometric approach to a non stationary process.
Proceeding of the 2nd international conference on
mathematical models for engineering science (MMES’11).
Zang, D. & Sommer, G. (2007) Signal modeling for two-
dimensional image structures. J. Vis. Commun. Image Represent.
18, 81–99.
Zhong, J. & Sclaro, S. (2003) Segmenting foreground objects from a
dynamic textured background via a robust kalman flter. In
Proceedings of the International Conference on Computer
Vision (ICCV), pp. 44–50.
(Manuscript accepted 23 September 2013)
doi: 10.1111/aje.12128
© 2013 John Wiley & Sons Ltd, Afr. J. Ecol., 52, 363–369
Forest fire detection method 369