=Paper=
{{Paper
|id=Vol-1452/paper23
|storemode=property
|title=Development of Trained Algorithm Detection of Fires for Multispectral Systems Remote Monitoring
|pdfUrl=https://ceur-ws.org/Vol-1452/paper23.pdf
|volume=Vol-1452
|dblpUrl=https://dblp.org/rec/conf/aist/ZraenkoMG15
}}
==Development of Trained Algorithm Detection of Fires for Multispectral Systems Remote Monitoring==
https://ceur-ws.org/Vol-1452/paper23.pdf
Development of Trained Algorithm Detection of
Fires for Multispectral Systems Remote
Monitoring
Sergey Zraenko, Margarita Mymrina, and Vladislav Ganzha
Ural Federal University, Mira str., 19, Yekaterinburg, 620002,
Russian Federation
z sm@mail.ru
Abstract. The paper describes research of a trained threshold algo-
rithm for analysis of the fire situation using multispectral remote sensing
data. The algorithm is based on application of sub-satellite information
about the fire on the territory covered by a satellite image. Thresholds
are selected for each of the spectral bands for the determination of known
fire threshold and then are used to select all thermal anomalies in the
picture. According to study of the algorithm the informative spectral
channels are defined for detecting fires. Combination of the results from
the multiple channels spectroradiometer MODIS for increasing the prob-
ability of correct detection is demonstrated.
Keywords: Forest fires, MODIS spectroradiometer, MOD14 algorithm,
trained threshold algorithm, spectral channels, probability of fire detec-
tion
1 Introduction
Data of the earth remote sensing (ERS) is widely used for forest fire detection.
In this case, pictures of Imaging Spectroradiometer MODIS [1], which is a part
of equipment of American Terra and Aqua satellites, are intensively used now.
An operative receiving of information from MODIS is carried out by personal
receiving ground stations, one of which (UniScan-24 [2]) is set at the Ural Federal
University.
The MODIS data are formed in 36 spectral bands and characterized by a
wide swath (2300 km) and high frequencyof udate (up to several times a day).
However, these images have low spatial resolution (1 km in the most of the
infrared spectral channels) that does not allow one to detect reliably fires in a
fraction of a hectare under various observation conditions (surface fire, lack of
fire intensity, poor weather conditions, etc.).
187
2 Problem formulation and development of trained
threshold algorithm
Complexity of solving the problem of detection of the forest fires using remote
sensing data may be shown by presence of a large number of Russian and foreign
Internet services. This information about the fires differs not only from the
actual, but, also, among the sources. The evidence of this is confirmed [3] by
analysis of the fire in State Nature Reserve ”Denezhkin Kamen” (Fig. 1) in 2010
according to the data of several online fire detection and monitoring services.
Information from these services is based on the MODIS sensor data that are
processed by the MOD14 fire detection algorithm and its modifications [4].
Fig. 1. Daily snapshot MODIS 31.07.2010 combined with mask fires and the
boundary of the reserve ”Denezhkin Kamen”
This algorithm uses the temperature data of the pixels at two infrared spec-
tral channels: 4 µm (21 channel) and 11 µm (channel 31). The brightness tem-
perature of a ”hot” pixel in channel 21 should be higher 312K by day and 305K
by night. In addition, if the temperature difference of a pixel between the 21 and
31 channels and the difference between channel 21 and the environment tem-
perature are larger, the probability that the pixel belongs to the fire is higher.
Pre-image is also evaluated under presence of clouds using masks clouds and
global clouds; and the pixels belonging to the solar flare are eliminated. As a
result, the mask of fires is built that contains the so-called ”thermal points”,
i.e., the pixels exceeding the preset temperature thresholds.
It should be noticed that this algorithm is currently the basis for a number of
specialized software tools for processing remote sensing data, such as the program
product ScanEx Image Processor [5]. In the fact, efficiency of the MOD14 is
insufficient in areas with high forest vegetation, and it can detect only relatively
large fires. This factor is caused by fixed values of temperature thresholds that
do not take into account specific conditions of the observation.
188
To increase the likelihood of a correct detection of temperature anomalies, we
have proposed [6] to train our threshold algorithm on the basis of sub-satellite
information about the current fire. Selection of the threshold level was made
using channel 21 of the spectroradiometer MODIS for pictures dated by July 28
(night) and July 29 (day), 2010. The training was based on the known fire in the
reserve ”Denezhkin Kamen”. For this purpose, a part of the Sverdlovsk region
was cut and converted to a range of brightness from 0 to 255, and after that the
detection thresholds were determined. It is also supposed (Fig. 2) that there is
the area free from fire adjacent with another seat of fire, the area of 1 pixel size
(gray area) separated from the fire at the distance about 0.5 – 1.5 pixels is the
boundary region part, which can be occupied with the fire (white square) round
the seat of fire (a dark oval). In our case (Fig. 2), when the seat of the fire is in
the pixel with coordinates (X = 217, Y = 193), this means that ”hot” pixels are
absent at the points X = 215–219 when Y = 191 and Y = 195 and at the points
Y = 191–195 when X = 215 and X = 219.
Fig. 2. The definition of adjacent area to the fire
The threshold in spectral channel 21 (IPOR ) is defined as the average of the
luminance values of pixels from the known fire hearth (IOI ) and the adjacent
region (ISI )
IPOR = (IOImin + ISImax )/2,
IOImin = min (IOI ), (1)
ISImax = max (ISI ).
Here, IOImin is the minimum brightness of a pixel of the source area with known
fire, ISImax is the maximum brightness of a pixel of the adjacent area. In this
case, (I = 1) are pixels (correlated with the fire) in the image (I) exceeding this
189
threshold (
0, if I < IP OR ,
I= (2)
1, if I ≥ IP OR .
3 Research of the trained algorithm
Study of the proposed algorithm was performed using MODIS imagery territory
of the Sverdlovsk region 28–29 July, 2010 and 01.08.2012. The actual fire situ-
ation was determined by the summary of the Ministry of Emergency situations
posted the Internet. In the night picture, the threshold of detection of fire in
the reserve ”Denezhkin Kamen” in channel 21 MODIS (in relative brightness
values 0–255) was determined in the range B21N IGHT = 59–69. After select-
ing the middle value B21N IGHT = 64, the mask of detected fires (Fig. 3) was
constructed. The fire figures (black dots) are visible in the reserve ”Denezhkin
Kamen”, Garinsky, and Sosvinsky counties.
Fig. 3. Mask fires of the trained algorithm for the night snapshot, 28.07.2010; the
black dots are selected thermal points
Table 1 presents information about known fires and the results of detection
of the thermal points obtained (using the night MODIS image) by the proposed
algorithm and the MOD14 algorithm with selected (from the condition of max-
imum detection of existing fires) value T21N IGHT = 300K.
The chosen threshold value T21N IGHT = 300K = 270C was lower than the
value T21N IGHT = 305K, which is considered as the boundary for the recognition
of the pixel to be the thermal points. It is quite realistic to suppose that this was
190
the night temperature of crowns of trees in the Northern part of the Sverdlovsk
region. Note that the proposed algorithm has demonstrated its efficiency, but it
showed slightly worse results compared to the MOD14 with selected thresholds
in channels 21 and 31.
Table 1. Detection of fire by the night snapshot 28.07.2010; the number of thermal
points in each snapshot is given in parentheses
County Actual MOD14 (selected thresholds) Trained
fires algorithm
reserve ”Denezhkin Kamen” 1 3 (4; 3; 1) 2 (3; 1)
Severouralsky 11 2 (1; 1) no
Sosvinsky 15 6 (2; 3; 4; 2; 2; 12) 3 (3; 1; 1)
Ivdelsky 9 1 (1) no
Garinsky 10 10 (1; 1; 1; 2; 1; 1; 2; 3; 2; 7) 3 (6; 3; 1)
When applying the proposed methodology on a daily snapshot, it took to use
as a training standard the fire in Garinsky county because the creeping fire in
the reserve ”Denezhkin Kamen” failed to be detected due to its masking by the
canopy of trees. In this case, the average value of the threshold B21DAY = 43 for
channel 21 fires in Sosvinsky, Garinsky, Krasnoturinsky, Serovsky, Kachkanarsky,
and Karpinsky counties were detected.
In this experiment, the threshold value T21DAY = 320K was selected by the
condition of maximum detection of existing fires and was slightly above the
recommended (T21DAY = 312K) to reduce the number of false thermal points.
Analysis (Table 2) has determined that the false ”hot” pixels were detected in
Kachkanarsky county, and the fires in the reserve ”Denezhkin Kamen”, Sever-
ouralsky, and Ivdelsky counties were skipped. According to the above data, the
trained algorithm showed approximately the same results as MOD14 with thresh-
olds chosen to the highlight of known fires. In general, the results of a daily
snapshot for both algorithms were slightly worse than for the night ones because
of the influence of surface temperature and possible specular reflections from
water surfaces and edges of the clouds.
Verification of the trained algorithm was carried out using the daily MODIS
image of the Sverdlovsk region dated 01.08.2012. In this case, the fire at Kras-
nouralsky county was selected as ”training”. The comparison of the results ob-
tained here for the trained algorithm with the actual number of fires, as well as,
with the MOD14 algorithm with a threshold chosen for the situation in 2010 is
given in Table 3.
The obtained result demonstrates the effectiveness of the proposed algorithm
(selected only one false fire) compared to the MOD14 algorithm where the thresh-
old was selected for the snapshot of the same territory made two years earlier
(with a precision of 3 days), where lots of false fires were found.
191
Table 2. Detection of fires on a daily snapshot 29.07.2010; the number of thermal
points in each snapshot is given in parentheses
County Actual MOD14 (selected thre- Trained algorithm
fires sholds)
reserve ”Denezhkin Kamen” 1 no no
Severouralsky 12 no no
Sosvinsky 15 2 (1; 1) 1 (1)
Ivdelsky 9 2 (1; 1) no
Garinsky 11 6 (1; 2; 2; 2; 3; 5) 4 (5; 2; 1; 1)
Karpinsky 7 1 (1) 2 (3; 1)
Krasnoturinsky 3 no 2 (8; 1)
Serovsky 1 2 (1; 1) 1 (11)
Kachkanarsky 1 1 (12) 2 (15; 12)
Table 3. Detection of fires on a daily snapshot 01.08.2012
County Actual MOD14 (the thresholds Trained algorithm
fires selected for the 29.07.2010
snapshot)
Krasnouralsky 1 1 1
Kachkanarsky no 2 1
Sosvinsky no 1 no
Ivdelsky no 7 no
Severouralsky no 1 no
Krasnoturinsky no 1 no
Karpinsky no 2 no
Further improvement of accuracy of fires detection using MODIS data can
be done by integration of the detection results obtained by processing a series of
images (especially day and night) received at the same or next day. However, at
the time of the next shooting, some fires can be extinguished, and some new fires
may appear. Shooting conditions can also change. Therefore, the most expedient
way is to extract information about the fire according to the other spectral
channels of a single snapshot. Thus, parameters of the signal passing through
the atmosphere from a thermopoint are various for different spectral channels
(depend on a condition of the atmosphere, temperature of a thermopoint and
temperature sensitivity of the channel). So we conducted research of temperature
sensitivity of different channels.
In contrast to our previous publications [3], [6], this paper shows the results
that were received using pictures 28.07.2010 (night) and 29.07.2010 (day) with
the subsequent association of results in spectral channels. Results for the MODIS
channels that are informative to find thermal anomalies are presented in Table
4. The appearance of the sign ”–” for channel 24 means that it is uninformative
during the day because it could not allocate any points of real fires. In addition,
192
the reserve ”Denezhkin Kamen” is a part of Severouralsky county, and the table
presents aggregated information about fires in these area.
Table 4. Results of study of the spectral channels reliability
The channel number 20 21 22 23 24 25
County Actual Fires found (night/day)
fires
(night/
day)
Severouralsky (re- 11/12 2/0 2/0 2/0 2/0 2/– 2/0
serve ”Denezhkin
Kamen”)
Sosvinsky 15/15 6/0 3/0 3/0 4/0 6/– 5/0
Ivdelsky 9/9 0/6 0/0 0/0 0/0 13/– 0/0
Garinsky 10/11 3/3 3/2 3/3 3/3 3/– 3/2
Karpinsky 0/7 0/0 0/0 0/0 0/0 0/– 0/0
Krasnoturinsky 0/3 0/2 0/0 0/0 0/0 0/– 0/0
Serovsky 0/1 0/0 0/1 0/1 0/1 0/– 0/1
Kachkanarsky 0/1 0/1 0/1 0/2 0/1 0/– 0/0
Taborinsky 0/0 0/0 0/0 0/0 0/0 0/– 0/2
The channel reliability 0.244/ 0.178/ 0.178/ 0.200/ 0.444/– 0.222/
night/day (PRN /PRD ) 0.203 0.068 0.085 0.085 0.051
Analysis of the obtained data allows one to determine the probability of a
false detection of the thermal points as the ratio of the total area of the false
”hot” pixels to the area of the territorial county where they are found. The area
of the MODIS pixel in the selected channels is 1 km2 , so, the probability of false
detection of a fire by daylight shots in Kachkanarsky and Taborinsky districts
was equal to 3 · 10−3 and 2 · 10−4 , respectively. Similarly using the night shot
of four non-existent fires were found (13 discovered and reported 9) at Ivdelsky
district. So, the probability of false detection of fire is 2 · 10−4 .
The probability of correct detection of the thermal point (PR ) in the spec-
tral channel (the reliability of the channel) was determined as the ratio of the
total (around snapshot) number of the detected fires in the channel to their ac-
tual number. So, the actual number of fires on territories at the night snapshot
28.07.2010 was 45, and number of thermal points found at the channel 20 is
11. As for the day shot, 59 points are actual fires and 12 are detected. So, the
reliability of the channel 20 at the night is PRN = 0.244, and at the day its PRD
= 0.203. Similarly, reliability of the remaining (21,...,25) channels (Table 4) was
calculated.
Combining the results of several spectral bands, it is possible to was estimate
the probability of correct fires detection based on remote sensing data and appli-
cation of the proposed algorithm. For example, if one of the pixels on the night
193
image of the Garinsky district exceeded trained 20,...,25 channels thresholds, the
probability that it belongs to the fire is
25
Y
PRN = 1 − (1 − PRN k ) = 0.823. (3)
k=20
Using information of only one channel, you can get the reliability of the
fire detection not more than PRN 24 = 0.444. If the pixel at the day image of
the Garinsky district exceeded the threshold of 20,...,23, and 25 channels, the
probability that it is thermal point is PRD = 0.410. And finally, if it is the same
pixel on both daylight and night shots, the probability that it belongs to the fire
is
PR = 1 − (1 − PRN )(1 − PRD ) = 0.896. (4)
Comparing our algorithm with well-known ones, for example [7], [8], we
should say that in them information about statistical characteristics of fire and
background pixels is used to define a threshold of detection. But we suggest the
approach that doesn’t assume any initial allocation of potentially hot pixels for
definition of distribution function of probability density for brightness of fire-
fighters and the pixels surrounding them. It is based on the fact that detection
of thresholds is performed at spectral channels on the basis of allocation of the
known fire in the picture, according to the subsatellite information. In this case,
it is possible to exclude an uncertainty in temperature of a surface and, partially,
in the atmosphere parameters.
The conducted experimental study of our algorithm, which is based on the
existing fire, allows one to conclude that the results received with its help aren’t
worse than ones of the well-known algorithm MOD14. Similar results are re-
ceived in papers [7], [8] that allows one to judge efficiency of our algorithm in
comparison with ones offered there. We plan to conduct further researches using
more extensive actual material for receiving adequate quantitative estimates.
Combination of our results with ones received by the mentioned algorithms will
allow increasing probability of the forest fires correct detection.
The further increase of probability of the correct fire detection according to
MODIS is possible by exception of reflections from stationary natural (water)
and anthropogenous (industrial and civil constructions) objects, and, also, from
nonstationary objects (edges of clouds).
4 Conclusion
The paper presents description of the trained threshold algorithm that uses the
true ground information about one of the fires and builds an effective procedures
for analysis of the fire situation on the basis of the territory shot obtained by the
spectroradiometer MODIS. Study of the proposed algorithm was implemented
for the forest fire detection using data by 21 spectral channels of one night and
194
two daylight MODIS images. Good coincidence of the results of the fires detec-
tion with the ground true data and results of work of the well-known MOD14
algorithm was obtained. Informativity of the MODIS spectral channels for forest
fire detection using the trained algorithm was analyzed. The example shows that
approximately double increase of the probability of correct detection of fires (up
to 0.896) is possible by combining the results of spectral bands 20,...,25 of the
night and day shots.
5 Acknowledgment
The work was partially supported by the RFBR grants Nos 13-07-12168, and
13-07-00785.
References
1. MODIS specifications. NASA. http://modis.gsfc.nasa.gov/about/specifications.php
2. UniScan-ground station for X–band data reception from polar-orbiting remote sens-
ing satellites. RnD Center ”ScanEx”. http://www.scanex.com/en/stations
3. Zraenko, S. M., Mymrina, M. A., Ganzha, V. V.: Obuchaemyi algoritm ob-
naruzheniya pozharov po dannym distantsionnogo zondirovaniya [Learning algo-
rithm of fire detection from remote sensing data]. Zhurnal radioelektroniki: elek-
tronnyi zhurnal, 11 (2014), URL: http://jre.cplire.ru/jre/nov14/8/text.pdf
4. Algorithm Technical Background Document. MODIS Fire Products.
http://modis.gsfc.nasa.gov/data/atbd/atbd-mod14.pdf
5. SCANEX IMAGE PROCESSOR. In-depth Processing of Remotely Sensed
Data and Creation of Thematic Products. RnD Center ”ScanEx”.
http://www.scanex.com/en/software
6. Zraenko, S. M., Mymrina, M. A., Ganzha, V. V.: Application of remote sensing
data for forest fires detection with the use of a trained thresholding algorithm /
Proc. 24th International Crimean Conference Microwave and Telecommunication
Technology (CriMiCo–2014), 2, 1151–1152 (2014).
7. Ying, Li, Vodacek, A., Kremens, R. L., e.a.: A Hybrid Contextual Approach to Wild-
land Fire Detection Using Multispectral Imagery. IEEE Transactions on Geoscience
and Remote Sensing, 43 (9), 2115–2126 (2005).
8. Chin, S., Liew, A., Lim, L., Keong, K.: New Methods of Active Fire Detection Using
MODIS Data. Proc. 26th Asian Conference on Remote Sensing and 2nd Asian Space
Conference, 1, 411–415 (2005).
195