Fire detection technology in Mongolia
The Methodology of fire detection
1. Fire detection methodology in practice at ICC
Using the image processing system IVAS, IDRISI, PCI the ICC created a methodology to detect fire
sources and burnt area, and estimated the threshold values. Most important information source is
channel 3 of NOAA/AVHRR.. This channel is very sensitive to the high temperature target on the ground
and can be used to detect active fire with the following algorithm:
I. Active fire:
a) CH3 > 45°C
b) CH1(or CH2) = 6 - 12
As channel 1 and channel 2 are sensitive to the vegetation, waterbody and clouds, these 2 channels
have the ability to detect burnt area and smoke plums.
II. Burnt area:
a) CH3 > 35 - 45°C
b) CH1(or CH2) = 3 - 6
Channel 4 or 5 are used for cloud masking.
This data processing including a simple geo-referencing can be carried out very quickly. In the final
image product, active fires are identified by visual interpretation and plausibility check.
By above mentioned methodology using the daily NOAA satellite data we can monitor and produced
daily fire maps and hot spots , as showed on Figures 1 have been illustrated the trends of steppe fire
over Dornod and Khentii aimags (North Eastern part of Mongolia).
2. Fire detection methodology using NOAA/AVHRR data in JRC
Since 1998, the Joint Research Centre of the European Community in Ispra (Italy) developed a global
fire monitoring system. It is being developed in response to a call from scientists and policy makers for
globally consistent information on the distribution and behaviour of fire in the environment
Satellite images (NOAA AVHRR) are acquired by a world- wide network of receiving stations. Each
station operates a data processing chain for detecting fires in the satellite imagery. The data is
processed immediately on-site at the receiving station. The thematic product, in this case co-ordinates of
detected fires, is much reduced in volume compared to the original satellite images, so it can reasonably
be transmitted over internet links between receiving stations. Daily, global fire maps are built up at the
JRC in Italy from this regional data by automatically sharing regional fire maps over the internet. Global
fire information is then available on-line, in near real-time (Figure 2)
Figure 1: Steppe fire in April 2000, Northern Mongolia.In 2000, a regional World Fire Web (WFW) node was set up in Mongolia at the ICC in the framework of a
The local WFW nodes might directly respond to the information “coordinates of active fires”. In any case
they have the ability to transfer the data to GIS systems for more detailed evaluations or preparing
cartographic outputs.
Figure 2: WFW Fire map showing active fires in Mongolia
TACIS Project founded by the European Commission. Since then the processing chain is operational
and used for the fire detection in Mongolia.
3. Fire detection algorithm of WFW
In contrary to teh fire detection method developed at ICC, the WFW used a contextual algorithm. Such
an algorithm can be applied to a global data set without having to be adjusted for different geographical
regions. The algorithm chosen is based on work by Prins and Menzel (1992), Flasse and Ceccato
(1996), and reported in Justice and Dowty (1993).
There are two phases:
- Threshold Fire Test - a selection of pixels that could potentially contain fires, and thus be called "fire
pixels".
- Contextual Fire Test - a confirmation of the fire pixel classification by comparing the pixel with its
immediate neighbourhood.
These two phases are described below. Note that, henceforth, Ch(i) represents the bi-directional
reflectance factor of AVHRR channel i (i = 1, 2), and Tb(i) represents the brightness temperature of
channel i (i = 3, 4, 5).
- Threshold Fire Test:
This phase is intended to select all those pixels that may contain a fire. It uses thresholds that are low
enough to keep any possible fire, but high enough to reject most background pixels.
A pixel is selected as a potential fire if:
Tb(3) > 311K and Tb(3) - Tb(4) > 8K
- Contextual Fire Test:
This second phase confirms whether or not the fire detected in test (1) is definitely a fire. The test is
based on knowledge of the candidate pixel in relation to its neighbours. Firstly, in order to reject pixels
whose radiance in Ch(3) is influenced too much by high reflection, the following test is used:
Ch(2; Top of Atmosphere) < 20%
Secondly, statistical information is calculated about the pixels in a surrounding "background" contextual
window. Any surrounding pixels found to be cloud, potential fires, water, desert or affected by sun- glint,
are excluded from the calculation. Starting from a size of 3 × 3, the window is allowed to grow until at
least 25% of the pixels contained qualify to be included in the calculation of the statistics. If there are still
not enough valid pixels when the window has grown to 15 × 15, then the algorithm is unable to make a
decision and the possible- fire pixel is recorded as "indeterminate" - a so-called "blue-point". Once a valid
context window has been built, the following are calculated:
Tb(3)bg = Mean T b(3) in the background.
s(3)bg = Standard deviation of T b(3) in the background.
Tb(34)bg = Mean value of [T b(3) - Tb(4)] of pixels in the background.
s(34) bg = Standard deviation of [Tb(3) - Tb(4)] of pixels in the background.
A potential fire is then confirmed if:
[Tb(3) - Tb(4)] > Tb(34)bg + 2 s(34)bg and Tb(3) > Tb(3)bg + 2 s(3)bg + 3K.
