ry High
2 (near
1995, at
)ased on
provided
rer 5000
in such
e of the
' associ-
ally sent
hs Via a
e direc-
ry time
cessing
ATA
NOAA
he fires
AVHRR
1. Geo-
jer day
| of the
tection
em, the
ised.
ot per-
of the
umbers
lese re-
ecutive
coming
pectral
e three
than a
nes are
je geo-
] every
ta and
Satellite
image data
Detection
of error lines
Archiving of
orbit data
Geo-coding
Search of
hot spots
False alarm
screening
Generation Alert mes-
of alert sage
message
DSPZÉEUEEEEEREESEE
HEGRE
Figure 1: Tasks in fire detection.
the timing information extracted from the image data. This
interpolation grid is then inverted and used when the image
data are resampled to a regular pixel grid in the selected map
projection (Gauss-Kriiger).
Resampling is a critical step in image processing if the aim
of the system is to detect small active fires. Usually a fire
affects only one pixel in a NOAA image. If bi-linear interpo-
lation is used in resampling, most of the one-pixel fire signals
are rendered undetectable as a consequence of the averaging
introduced. On the other hand, if cubic convolution is used,
numerous false alarms are generated especially along sharp
edges in connection of boundaries between clouds and cloud
shadows. This is due to the edge enhancement that cubic-
convolution resampling produces. The most suitable standard
resampling technique for fire detection is the nearest neigh-
bour resampling. If the pixel grid of the geo-coded image is
dense enough (the inter-pixel distance of the AVHRR sensor
at nadir is approximately 0.8 km), no fire signals are lost in
the resampling and no false alarms are generated either.
32 Fire Detection Algorithm
As in many other systems, the detection of small fires is based
on the use of band-3 data (wavelength 3.5 pm). All pixels
that have a digital number less than a pre-determined thresh-
old are considered to be potential fire pixels. These pixels are
searched in the image and connected fire pixels are grouped
together to form fire patches or "hot spots”.
In addition to the spectral features used widely in fire detec-
tion, the fire detection system takes into account the imaging
geometry. The imaging geometry is represented by the devi-
EAN
1 UR
= The digital-number scale in bands 3, 4, and 5 of the AVHRR sensor is
versed so that high radiant power corresponds to low digital numbers and
oy radiant power to high digital numbers. That is why low digital numbers
Te sought for in fire detection, not high.
585
ation angle. The deviation angle o is the angle between the
direction of reflected sun light and the line of sight from target
to the sensor (figure 2). The deviation angle is computed:
a marc cos( TR) (1)
Sun
C Sensor
S
R
c
/
^^target %
Figure 2: The deviation angle between the vectors R (re-
flected sun light) and S (target to sensor).
A detected hot spot is considered to be a real fire only if it
fulfils all of the following criteria:
1. the deviation angle must be greater than a pre-
determined threshold,
2. the band-2 average of the hot spot must be smaller
than a pre-determined threshold,
3. the band-4 average of the hot spot must be smaller
than a pre-determined threshold,
4. the hot spot must be further than a pre-determined
threshold away from the borders of the image swath,
5. the number of pixels in the hot spot must be smaller
than a pre-determined threshold.
As in (Pereira and Setzer, 1993), the original digital num-
bers are used as such and are not converted into tempera-
ture units. Conversion to temperature units would make the
thresholds directly comparable with those used in other sys-
tems. It would not make the system more general because the
thresholds must be adapted when moving from one ecosystem
to another (Kennedy et a/ 1994).
Despite the fact that the data are not converted to temper-
ature units, the fire detection algorithm resembles the algo-
rithm described by (Kennedy et a/ 1994). The search of hot
spots based on band-3 data corresponds to criterion i. Crite-
rion 3 corresponds to criterion iii and criterion 2 to criterion
iv. The criterion ii is always fulfilled in hot spots detected
in the Boreal forest zone where extremely hot natural targets
are missing. The criteria of geometric nature (criteria 1, 4,
and 5) in the list above are such elements in the algorithm
that do not have counterparts in the algorithm described by
(Kennedy et al 1994).
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996
