Full text of DOI:10.1080/01431160050144956

This article was downloaded by: [Florida State University]
On: 17 October 2014, At: 18:51
Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number:
1072954 Registered office: Mortimer House, 37-41 Mortimer Street,
London W1T 3JH, UK
International Journal of
Remote Sensing
Publication details, including instructions for
authors and subscription information:
http://www.tandfonline.com/loi/tres20
Satellite-based detection of
Canadian boreal forest fires:
Development and application
of the algorithm
Z. Li a , S. Nadon b & J. Cihlar a
a Canada Centre for Remote Sensing , 588 Booth,
Ottawa, Ontario, Canada , K1A OY7
b Environment Canada , Ottawa, Ontario,
Canada , K1A OH3
Published online: 25 Nov 2010.
To cite this article: Z. Li , S. Nadon & J. Cihlar (2000) Satellite-based
detection of Canadian boreal forest fires: Development and application of the
algorithm, International Journal of Remote Sensing, 21:16, 3057-3069, DOI:
10.1080/01431160050144956
To link to this article: http://dx.doi.org/10.1080/01431160050144956
PLEASE SCROLL DOWN FOR ARTICLE
Taylor & Francis makes every effort to ensure the accuracy of all
the information (the “Content”) contained in the publications on our
platform. However, Taylor & Francis, our agents, and our licensors
make no representations or warranties whatsoever as to the accuracy,
completeness, or suitability for any purpose of the Content. Any opinions
and views expressed in this publication are the opinions and views of
the authors, and are not the views of or endorsed by Taylor & Francis.
The accuracy of the Content should not be relied upon and should be
independently verified with primary sources of information. Taylor and
Francis shall not be liable for any losses, actions, claims, proceedings,

demands, costs, expenses, damages, and other liabilities whatsoever
or howsoever caused arising directly or indirectly in connection with, in
relation to or arising out of the use of the Content.
This article may be used for research, teaching, and private study
purposes. Any substantial or systematic reproduction, redistribution,
reselling, loan, sub-licensing, systematic supply, or distribution in any
form to anyone is expressly forbidden. Terms & Conditions of access
and use can be found at http://www.tandfonline.com/page/terms-and-
conditions

int. j. remote sensing, 2000, vol. 21, no. 16, 3057–3069
Satellite-based detection of Canadian boreal forest Ž res: development
and application of the algorithm
Z. LI†*, S. NADON‡ and J. CIHLAR†
†Canada Centre for Remote Sensing, 588 Booth, Ottawa, Ontario, Canada,
K1A OY7
‡Environment Canada, Ottawa, Ontario, Canada, K1A OH3
(Received 27 July 1998; in Ž nal form 4 June 1999)
Abstract. This study presents a comprehensive investigation of Ž res across the
Canadian boreal forest zone by means of satellite-based remote sensing. A Ž re-
detection algorithm was designed to monitor Ž res using daily Advanced Very
High Resolution Radiometer (AVHRR) images. It exploits information from
multichannel AVHRR measurements to determine the locations of Ž res on satellite
2
pixels of about 1 km under clear sky or thin smoke cloud conditions. Daily Ž re
maps were obtained showing most of the active Ž res across Canada (except those
obscured by thick clouds). This was achieved by Ž rst compositing AVHRR scenes
acquired over Canada on a given day and then applying the Ž re-detection algo-
/
rithm. For the Ž re seasons of 1994–1998, about 800 NOAA AVHRR daily mosaics
were processed. The results provide valuable nation-wide information on Ž re
activities in terms of their locations, burned area, starting and ending dates, as
well as development. The total burned area as detected by satellite across Canada
is estimated to be approximately 3.9, 4.9, 1.3, 0.4 and 2.4 million hectares in 1994,
1995, 1996, 1997 and 1998, respectively. The peak month of burning varies
considerably from one year to another between June and August, as does the
spatial distribution of Ž res. In general, conifer forests appear to be more vulnerable
to burning and Ž res tend to grow larger than in deciduous forests.
1. Introduction
Quantitative information about the spatial and temporal distribution of Ž res is
important for forest protection and in the management of forest resources. It is also
indispensable to such disciplines as ecology, wildlife management and atmospheric
chemistry (Levine 1991). Emissions of greenhouse gases and aerosols from Ž res are
of great interest to climate change studies (Crutzen and Andreae 1990, Levine 1991).
Beside changes in land cover as a result of burning, Ž res also aŒect terrestrial–
atmospheric interactions (Levine et al. 1995, Scholes 1995).
In the last decade, images from the Advanced Very High Resolution Radiometer
(AVHRR) on board the National Oceanic and Atmospheric Administration (NOAA)
series satellites were often employed to detect wild Ž res (Robinson 1991, Arino and
Mellnotte 1995, Illera et al. 1995). While NOAA-AVHRR instruments were designed
mainly for weather surveillance and sea surface temperature management, their
*Corresponding author; e-mail: li@ccrs.nrcan.gc.ca
International Journal of Remote Sensing
ISSN 0143-1161 print/ISSN 1366-5901 online © 2000 Government of Canada
http://www.tandf.co.uk/journals

3058
Z. L i et al.
capability for detecting Ž res has been exploited as well. Studies on the detection and
environmental and climatic impact of tropical Ž res have made the most extensive
use of AVHRR data (Kaufman et al. 1990, Malingreau 1990, Justice et al. 1996 ). It
was estimated that tropical Ž res around the world destroyed about 15 10 km of
Ö
6
2
forests in the last decade (Singh 1993). Only a few handful of studies were reported
on the detection of boreal forest Ž res using AVHRR data (Flannigan and Vonder
Haar 1986, Cahoon et al. 1994, Li et al. 1997).
The total burned area in the world’s boreal forests remains largely unknown
(Setzer and Malingreau 1996). The lack of information hinders our ability to close
the global carbon budget because the boreal biome covers an immense area where
Ž re is a signiŽ cant, if not a dominant, mechanism for releasing surface carbon stocks
to the atmosphere (Sellers et al. 1995). Therefore, monitoring boreal forest Ž res is
critical to evaluating their contribution to the atmospheric stock of greenhouse gases
and aerosols, which are two major factors altering the course of anthropologic
climate change. On the other hand, Ž re occurrence may serve as an indicator of
climate change. While an individual Ž re event is random, the trend of Ž re activity
over a large area in the long run is strongly aŒected by the climatic regime.
Therefore, our study on forest Ž res has both practical and research objectives:
near real-time Ž re monitoring and acquiring a consistent Ž re database in Canada;
and estimation of greenhouse and aerosol emissions and studying the relation/
interaction between climate change and Ž re activity. This study is an extension of a
preliminary investigation by Li et al. (1997) on monitoring boreal forest Ž res using
AVHRR satellite data. In that study, a Ž re-detection algorithm for use with NOAA-11
AVHRR data was developed and validated. The previous study area was limited to
the Boreal Ecosystems Atmosphere Study (BOREAS) (Sellers et al. 1995) region of
1200 1200 km . The total area of Ž res detected by satellite agreed well with a
Ö
2
ground-based estimate by Ž re agencies. In the present study, the same algorithm is
applied to process all single-day AVHRR scenes acquired by NOAA-11 throughout
the Ž re season of 1994 covering the entire Canadian landmass. The algorithm is,
however, found to generate too many false Ž res when applied to the AVHRR images
acquired by NOAA-14. Considerable eŒorts were therefore devoted to develop a
new algorithm for use with NOAA-14 AVHRR data. The new algorithm was used
to process AVHRR images obtained from 1995–1998 across Canada. The results of
Ž re detection are presented in this paper. In a companion paper (Li et al. 2000), the
performance of the algorithm in mapping burnt area is evaluated in comparison to
Ž re polygon data acquired by Ž re agencies.
The following section describes the data used in the study. Section 3 is concerned
with the development and evaluation of a satellite Ž re-detection algorithm. An
analysis of Ž re statistics obtained across Canada from 1994 to 1998 is presented in
§4, and §5 summarizes the results.
2. Data
This study makes use of AVHRR images from NOAA-11 in 1994 and from
NOAA-14 from 1995 to 1998. The local overpass time of NOAA-11 in Canada was
around 15:45 with a viewing geometry close to the principal plane, i.e. the AVHRR
sensor scanned the earth in either backward or forward scattering direction (Li et al.
1997). NOAA-14 has an earlier overpass time (13:00 LT in 1995 and somewhat later
in subsequent years) with a viewing plane 45ß relative to the principal plane. After
receiving the AVHRR data at the Prince Albert station receiving station in

An algorithm to detect forest Ž res
3059
Saskatchewan, they were radiometrically calibrated and geometrically referenced
using the geocoding and compositing (GEOCOMP) AVHRR data processing system
(Robertson et al. 1992). The calibration for visible (channel 1) and near-infrared
(channel 2) measurements of AVHRR/NOAA-11 employs time-dependent oŒset and
gain values derived by Cihlar and Teillet (1995), while those for NOAA-14 were
based on values recommended by Rao and Chen (1996). The thermal AVHRR
channels (3–5) were calibrated using on-board blackbody reference. Pixel locations
were Ž rst computed by an orbit model that takes into account spacecraft orbit,
velocity, attitude and altitude, earth rotation and curvature, and then Ž ne-tuned
with reference to the ground control points determined from high resolution Landsat
image chips. An overall registration accuracy of less than 1 km can be achieved with
a su cient number of ground control points available under clear conditions (Cihlar
et al. 1997).
A small amount of AVHRR data from NOAA-14 was used for the development
of the Ž re-detection algorithm. They consist of single-day AVHRR scenes over a
region of 1200 1200 km in northern Manitoba and Saskatchewan around the study
Ö
2
region of the BOREAS where many forest Ž res occurred (Li et al. 1997). Besides,
this region is representative of the boreal biome. Through human inspection of the
satellite images, Ž res can be visually identiŽ ed by co-location of the hot spots on
channel 3 images and the associated smoke plume patterns on channel 1 (Setzer and
Malingreau 1996, Li et al. 1997). A smoke plume usually shows conical or bending
shape with a vortex over its origin. The inspection was facilitated by exploiting
various functions of image enhancement such as linear/nonlinear stretching available
from an image processing software (PCI). We have examined a large number of
AVHRR scenes to obtain Ž re masks as a training database for the development of
a Ž re-detection algorithm.
3. Algorithm
The Ž re-detection algorithm is based on the framework of Kaufman et al. (1990)
with modiŽ cations for the NOAA-14/AVHRR sensor. In addition, new tests are
introduced to cope with the special environment and burning conditions of the
boreal ecosystem. The algorithm consists of two major steps: marking potential Ž res
and removing false Ž res. Both steps encompass threshold tests. All the tests are
optimized to both detect real Ž res and eliminate as many false Ž res as possible.
While the majority of tests were proposed previously, the threshold values were
chosen following a trial-and-error approach based on the Ž re training dataset.
Histogram analyses of re ectance and brightness temperature corresponding
to burning and nonburning pixels proved to be an eŒective means of obtaining
optimized threshold values.
3.1. Identifying potential Ž re pixels
Following the pioneering work of Flannigan and Vonder Haar (1986) on auto-
matic detection of Ž res using AVHRR data, numerous investigations have been
conducted. Most have employed a threshold test based on the brightness temperature
of AVHRR channel 3. This marks all potential Ž res not obscured by thick clouds
by identifying the hot spots, i.e. pixels with brightness temperature (T ) higher
than a certain value. According to the Planck’s Law, AVHRR channel 3, with a
3
central wavelength around 3.7 mm, receives maximum radiative energy from objects
emitting at temperatures around 800 K. This temperature is close to the temperature

3060
Z. L i et al.
for burning biomass (Kennedy et al. 1994). However, since the AVHRR sensor was
not designed for Ž re detection, it loses sensitivity at such high temperatures. For the
AVHRR onboard NOAA-14, channel 3 becomes saturated at T
320 K. Never-
#
3
theless, it still proves to be the most useful AVHRR channel for Ž re detection as the
brightness temperature for most non-Ž re pixels is usually signiŽ cantly lower. In
practice, thresholds less than the saturation temperature are often used for two
reasons. First, wild Ž res have a large range of burning temperature ranging from
lower than 500 K to higher than 1000 K and they also have a variable fraction of
burning area within a pixel. To allow for the detection of all Ž res, a lower threshold
value is needed. Second, Setzer and Verstraete (1994) identiŽ ed an engineering design
problem in the on-board processing of channel 3 output signal for NOAA-11. They
found that signals greatly exceeding the saturation limit are assigned values below
the saturation limit. As a result, targets that are extremely hot can have a brightness
temperature lower than 320 K.
A brightness temperature of T
315 K was chosen in this study. This threshold
5
3
captures nearly all real Ž res in clear or thin-cloud covered regions, as conŽ rmed by
the Ž re training database. On the other hand, it also leads to far too many false Ž re
pixels. These false identiŽ cations stem from the limitations of channel 3 such as
sunglint eŒect, re ective soils, pixel overlap, sensor degradation, etc. (Setzer and
Malingreau 1996). Sunglint is the most serious problem encountered in our analysis
due to the presence of many lakes and rivers distributed throughout the Canadian
Shield. For NOAA-11 that has view geometry near the principal plane, the problem
can readily be resolved by avoiding the measurements made in forward scattering
directions (Li et al. 1997). For NOAA-14, however, this simple solution is no longer
valid since its viewing plane is far away from the principal plane and because sunglint
also occurs from cloud side in the backscattering direction. The subsequent tests are
thus introduced to remove the false Ž re pixels.
3.2. Removing false Ž res
Kaufman et al. (1990) introduced three tests to eliminate false Ž res. We adopted
these tests but tuned the threshold values. The Ž rst one uses the diŒerence between
Õ
channel 3 and channel 4 brightness temperatures (T
T ) to identify false Ž re pixels
3
4
caused by a warm background. Some surface types (e.g. bare soil) can become warm
enough to saturate channel 3. Since the spectral window of channel 4 is located in
the electromagnetic spectrum that has maximum radiative emission for the ordinary
Õ
earth temperatures, T
T is instrumental in discriminating these false Ž re pixels.
In the case of biomass burning, channel 3 receives much more radiant energy than
3
4
Õ
channel 4 and thus the value of T
For the boreal forests, the threshold value for T
T is high (Kennedy et al. 1994, Dowty 1996).
Õ
3
4
T was set to 14 deg K. All pixels
T values lower than 14 deg K are considered false Ž re pixels caused by a
3
4
Õ
with T
warm background.
3
4
The second test employing channel 4 alone deletes false Ž res caused by highly
re ective clouds. Radiance measured by channel 3 originates from both solar
re ection and terrestrial emission. Re ection of the solar radiation by clouds can be
large enough to also saturate channel 3. Since clouds usually have cold tops and
thus low brightness temperatures, this test rejects marked Ž re pixels of T less
than 260 K.
4
The third test is intended to eliminate more general bright-scene objects including
both clouds and surface pixels. Note that Ž re hot spots have relatively low re ectance

An algorithm to detect forest Ž res
3061
in channel 2 (R ) due to spewing ash and biomass consumption (Kennedy et al.
1994). Therefore, all Ž re pixels with R 0.22 are considered as false Ž res.
2
>
2
An additional threshold test is introduced to eliminate false Ž re pixels caused by
thin cirrus clouds. In some cases, a combination of warm background and thin
clouds can saturate channel 3 and negate the second test. Thin cirrus clouds have
Õ
low T and high T because of re ection, leading to large values of T and T
T .
The additional test makes use of the diŒerence between the two thermal channels of
4
3
3
3
4
Õ
AVHRR, channels 4 and 5. The diŒerence T
cirrus clouds, which is often referred to as the split window technique (Inoue 1987).
T has been used to identify thin
4
5
Õ
This test is implemented in combination with a relaxed test using T
Õ
T . All
3
4
Õ
T
thresholds have been established using the training database.
hot spots with T
T
4.1K and T
19 K are removed as false Ž res; these
<
>
4
5
3
4
In addition to the above threshold tests, two special screenings are made. The
Ž rst is designed in accordance with our objective of detecting forest Ž res. In Canada,
forest Ž res are far more extensive and persistent than other types of Ž res. By restricting
to forest Ž res, we can prevent certain problems. In particular, a large number of false
Ž res caused by sunglint and soil re ection occur over water bodies, barren land,
cropland and grassland. This test resorts to the use of a land cover classiŽ cation at
1 km resolution derived from AVHRR data (Pokrant 1991). It includes 10 coarse
land cover classes, namely, water, forests of four types (mixed wood, deciduous,
conifer and transitional), Arctic/alpine tundra, barren lands, agricultural cropland,
rangeland–pasture and cities. Only forest pixels are considered in the test, similar to
Chuvieco and Martin (1994).
The second screening eliminates all single hot spot pixels, i.e. no neighbour pixels
marked as Ž res. It copes with the problems caused by sub-pixel contamination. Due
to the coarse pixel resolution and land cover classiŽ cation, a designated forest pixel
may contain a fraction of water body such as the edge of a lake, rendering a sunglint
problem. Although NOAA-14’s viewing geometry is away from the principal plane,
sunglint can still be observed over a rough water surface. Also, sunglint from cloud
side tends to aŒect isolated pixels where the Sun’s position, the satellite viewing
angle and the cloud surface conditions combine in the right geometry. This screening
takes advantage of the fact that boreal forest Ž res usually spread much larger than
a 1-km AVHRR pixel size. Most single-tagged hot spots turn out to be false Ž res.
However, the screening has an adverse eŒect of eliminating small and initial-stage
Ž res, which may be recovered by visual inspection of Ž re smokes though. Figure 1
presents a  owchart of the Ž re-detection algorithm.
The e ciency of each step of the algorithm is shown in table 1 that was computed
from 24 training scenes of 1200 km 1200 km around the BOREAS region during
Ö
1995 Ž re season. For this training dataset, Ž res were identiŽ ed by visual inspection
of hot spots and smoke plumes in the absence of thick clouds. The results of the
tests are given in terms of the number of true Ž res (NTF) retained and the number
of false Ž res (NFF) removed in each test. Note that the e ciency of a speciŽ c test
may depend on the sequence of the tests.
It follows from table 1 that after applying all the tests, the algorithm missed 11%
of the true Ž re pixels and eliminated 99% of the false Ž res. In terms of Ž re events,
the ratio of real Ž res missed would be smaller than the one given in table 1, since a
Ž re event usually consists of a cluster of Ž re pixels over a period of time. The test
to remove single Ž re pixels causes very few real Ž res to be missed, yet it eliminates
a large number of false Ž res. The most eŒective threshold test for eliminating false

3062
Z. L i et al.
Figure 1. The  owchart of the Ž re-detection algorithm for use with NOAA-14 AVHRR data.
Table 1. Summary of the Ž re-detection tests and the statistics of their e ciency.
Test no.
Description
Threshold
NTF
NFF
1
2
3
4
5
6
7
Initial test
T
T
12 569 168168
3Õ
3
Eliminate warm background
Eliminate nonforest scenes
Eliminate bright scenes
Eliminate cloud edges or thin clouds
Eliminate cold clouds
Eliminate single Ž re pixels
T
12 569
12 569
12 442
11 307
11 307
11 160
48 855
30 511
5665
2673
2673
4
land cover
R
T
T
2Õ
4
4
Õ
T
and T
T
5
3
4
1828
Ž res appears to be the diŒerence between brightness temperatures in channels 3 and
4, followed by a test based on channel 2 re ectance. While 14% of the ‘Ž re pixels’
that passed all the tests remain false Ž res, further reduction of these would lead to
a larger number of real Ž res being missed.
Given the limitations in the signal content of AVHRR measurements pertaining

An algorithm to detect forest Ž res
3063
to burning, a trade-oΠis necessary between the NTF retained and the NFF removed.
This study has two goals: to develop a monitoring tool with a high probability for
the detected pixels to be real active Ž res and to be able to use this tool to get an
approximate overall estimation of Ž re statistics across Canada. It is thus a sound
choice to use an algorithm leading to comparable rates of missing real Ž res and
retaining false Ž res.
4. Application
The algorithm described above was applied to Canada-wide AVHRR mosaic
images obtained every day from NOAA-14 in 1995, 1996, 1997 and 1998. The
algorithm for processing NOAA-11 AVHRR data in 1994 was presented by Li
et al. (1997).
A coast-to-coast AVHRR image mosaic is created by compositing about Ž ve
scenes. Daily Canada-wide AVHRR mosaic is available from 1994 to 1998. Each
mosaic contains data for the Ž ve AVHRR channels together with satellite view zenith
(VZ), solar zenith (SZ) and relative azimuth (RA) angles and NDVI.
Figure 2 is an example of a Ž re mask derived from AVHRR images on 25 June
1995. Fire pixels are marked as red spots and the background is a false colour
composite denoting land (green), blue (ocean), cloud (white) and smoke (yellow).
Figure 2. Forest Ž res detected on 25 June 1995 across Canada. The red dots are Ž re hot
spots, superimposed on an AVHRR false colour composite image (green for vegetation,
orange for smoke and white for clouds, blue for thin cloud over land).

3064
Z. L i et al.
Broadly speaking, there are three major Ž re episodes, one located in Northwest
Territories, one in a border region between Saskatchewan and Manitoba, and one
in Quebec. Almost all of the detected Ž res are accompanied by smoke plumes,
conŽ rming their occurrence. Cloud cover is the primary limitation of the AVHRR-
based Ž re detection technique, especially for real-time Ž re surveillance. It poses a
less serious problem for obtaining seasonal Ž re statistics, since cloud cover changes
quickly from day to day.
From the daily Ž re masks, annual composites of burned areas can be obtained.
Figure 3 presents a composite of Ž re pixels detected between 1994 and 1998. Fire
activity appears to be widespread and exhibits strong inter-annual and spatial
variation. It is interesting to note that Ž res in remote areas, such as in the Northwest
Territories and northern parts of western provinces, tend to grow larger than those
in more populated areas such as southern Ontario and Quebec. This may be related
to Ž re management practices, forest types and weather conditions. The northern
forests consist primarily of conifers, while the southern ones are mainly deciduous.
The conifer boreal forest stands are usually much drier than deciduous forests,
facilitating Ž re growth. In addition, forest Ž re is a natural phenomenon in low-value
northern forests. Therefore, little human intervention is normally exercised.
The spatial distribution of Ž res diŒers signiŽ cantly among the years studied. In
1994, most Ž res occurred along the northern border of the boreal forest in western
Figure 3. The distribution of Ž re spots across Canada in 1994–1998.

An algorithm to detect forest Ž res
3065
Canada. In 1995, Ž res were widespread across the country. In 1996, fewer Ž res
occurred generally conŽ ned to Ontario and Quebec. Even less forested area was
burned in 1997, mostly in Quebec and Manitoba. Very few Ž re pixels from diŒerent
years overlap. Browsing through the sequence of single-day Ž re images in 1995
indicates that Ž res usually travelled from west to east, as weather systems do.
The cumulative area of all Ž res across Canada is approximately 3.9, 4.9, 1.3, 0.4
and 2.4 million hectares for 1994, 1995, 1996, 1997 and 1998, respectively. In compar-
ison, total burned area estimated by the Canadian Forest Service is 6.2, 6.6, 1.9, 0.6
and 4.7 million hectares (IFFN 1999), respectively for the above years. In comparing
these numbers, one has to keep in mind that the areas shown in Ž gure 3 represent
the accumulated areas of the active Ž res detected when the satellite overpasses and
when no thick clouds are present. Therefore, the underestimation of burnt area
(about 35% on average) is caused primarily by cloud cover and by the frequency of
satellite observation (one per day). Another reason is that Ž re agencies may tend to
overestimate the burnt area by omitting to record unburned islands lying within
large burnt regions (Fraser et al. 1999). The higher level of underestimation in 1994
can also be partially caused by the NOAA-11 malfunction in early September, thus
not allowing subsequent Ž res to be detected. A more detailed validation of this
algorithm against ground-based observations is given Li et al. (2000).
Figure 4 shows the daily variation in total burned area from 1994 to 1997. It
indicates that the period of active burning generally occurs from June to August
with a peak in late July and early August in 1994, late June in 1995, middle June in
1996 and early June in 1997. Superimposed upon the seasonal trend is a substantial
day-to-day variation. To a large extent, the daily  uctuations are caused primarily
by the presence of clouds and secondarily by missing images. This conclusion is
drawn from an extensive visual inspection of single-day images. The inspection helps
determine the causes of the  uctuations that are classiŽ ed into four categories: regular
points (no signiŽ cant artiŽ cial interruption), interruption by cloud cover, missing
satellite scenes and the combination of the latter two. The majority of the dips do
not correspond to real declines in Ž re activities. On the other hand, consecutive low
Ž re counts are more likely real. Most of the false low Ž re counts appear to last for
only 1 day. The seasonal trend of Ž re activities is more clearly evident in Ž gure 5,
which shows the monthly counts of Ž re pixels from 1994 to 1997. The most
widespread burning occurred in June 1995.
In addition to the temporal statistics of all Ž res across Canada, information on
the spatial spread of regional Ž res is also clearly conveyed by assigning diŒerent
colours to the Ž re pixels detected during various periods of time. Figure 6 is an
example of such an image for Northwest Territories in 1995 in which the Ž re
occurrence is separated into 10-day periods. It is found that the largest Ž re was Ž rst
detected between 1 and 10 June and expanded rapidly in all directions during the
following 10 days and did not extinguish until the end of August. Other Ž res in this
region have much shorter lifetimes.
5. Summary
This study provides the Ž rst comprehensive satellite surveillance of boreal forest
Ž res across the Canadian landmass during a 5-year period. An algorithm designed
to detect boreal forest Ž res was developed, and 5 years of AVHRR daily imagery
data were processed to track Ž re development and to obtain Ž re statistics.
The Ž re-detection algorithm was designed on the basis of previous investigations

3066
Z. L i et al.
Figure 4. The daily variation of total number of Ž res from 1994 to 1997 with explanation
for the cause of sudden drops.
Figure 5. The monthly variation of total number of Ž res across Canada.

An algorithm to detect forest Ž res
3067
1995
April-May
1-10 June
11-20 June
21-30 June
1-10 July
11-20 July
21-31 July
1-10 August
11-20 August
21 August - 31 October
Figure 6. Fire development in Northwest Territories.
with special considerations to cope with the unique boreal environment. The algo-
rithm consists of two basic steps, namely, marking all potential Ž re pixels and
removing false Ž re pixels. This is done through a series of threshold tests, a land
cover type test, and a test for the spatial contingency of Ž re pixels. The threshold
values for the tests were optimized using a small volume of training dataset obtained
by a visual analysis of AVHRR images. The algorithm is capable of capturing the
majority of real Ž res in the absence of thick clouds.
The algorithm was applied to hundreds of Canada-wide AVHRR images from
1994 to 1998. Daily Ž re maps were obtained for the 5 years, from which monthly
and seasonal Ž re statistics were computed. It was found that forest Ž res in Canada
show strong inter- and intra-annual variations and that they are distributed
nonuniformly. The total burned area detected in the 5 years diŒered by a factor of

3068
Z. L i et al.
over 12, ranging from 0.4 million hectares in 1997 to 4.9 million hectares in 1995.
These statistics are less than those obtained by conventional means of Ž re surveillance
with an underestimation on the order of 35%. The underestimation is mainly caused
by cloud cover and limited satellite revisit frequency. The peak Ž re activity in a
burning season occurs between June and August. Large Ž re clusters were found in
the north-western conifer forests, while small Ž res scattered over the south-eastern
deciduous forests.
This study lays a foundation for an operational satellite Ž re monitoring system
to monitor Ž res on near real-time basis and to obtain a timely Ž re statistic across
Canada. The system has the advantages of automatic operation, consistent data
quality, cost-eŒective use, and rapid response over the vast territory of the Canadian
boreal forest. The system was run very successfully during the Ž re season of 1998.
Daily Ž re images were generated and delivered to Ž re agencies, media and general
public within 24 h of the reception of satellite data. The quality of near real-time
detection is as good as shown in this paper. More and updated information
can be found at the Canada Centre for Remote Sensing web page at
http://www.ccrs.nrcan.gc.ca/ccrs/tekrd/rd/apps/em/cchange/canŽ re.html.
Acknowledgments
We would like to acknowledge the assistance of Mr Andrew MacPherson, Stephen
Boles, and Robert Fraser in the development and application of the Ž re-detection
algorithm. Devel-Tech Inc. in Saskatoon provided AVHRR mosaic images under a
contract to CCRS.
References
Arino, O., and Mellnotte, J. M., 1995, Fire index atlas. Earth Observation, European Space
Agency Quarterly, December.
Cahoon, D. R., Jr, Stocks, B. J., Levine, J. S., Cofer, W. R., III, and Pierson, J. M., 1994,
Satellite analysis of the severe 1987 forests Ž res in Northern China and Southeastern
Siberia. Journal of Geophysical Research, 99, 18 627–18 638.
Chuvieco, E., and Martin, M. P., 1994, A simple method for Ž re growth mapping using
AVHRR channel 3 data. International Journal of Remote Sensing, 15, 3141–3146.
Cihlar, J., and Teillet, P. M., 1995, Forward piecewise linear calibration model for quasi-
real-time processing of AVHRR data. Canadian Journal of Remote Sensing, 21, 22–27.
Cihlar, J., Chen, J., and Li, Z., 1997, Seasonal AVHRR multichannel data sets and products
for studies of surface–atmosphere interactions. Journal of Geophysical Research, 102,
29 625–29 640.
Crutzen, P. J., and Andreae, M. O., 1990, Biomass burning in the tropics: impact on
atmospheric chemistry and biogeochemical cycles. Science, 250, 1669–1678.
Dowty, P., 1996, The simulation of AVHRR data for the evaluation of Ž re-detection tech-
niques. In Biomass Burning and Global Change, edited by J. S. Levine (Cambridge,
Massachusetts: MIT Press), pp. 25–39.
Flannigan, M. D., and Vonder Haar, T. H., 1986, Forest Ž re monitoring using NOAA
satellite AVHRR, Canadian Journal of Forest Research, 16, 975–982.
Fraser, R., Z. Li, and Cihlar, J., 1999, A new technique for mapping burns: hot spot and
NDVI diŒerencing synergy (HANDS). Remote Sensing of Environment, in press.
IFFN, 1999, International forest Ž re news: Canada Report 1998, no. 20, 8 pp. (available from
//
/
/
http: www.ruf.uni-freiburg.de Ž reglobe ).
Illera, P., Fernandez, A., and Casanova, J. L., 1995, Automatic algorithm for the detection
and analysis of Ž res by means of NOAA AVHRR images. Advances in Remote Sensing,
4, 1–6.
Inoue, T., 1987, A cloud type classiŽ cation with NOAA-7 split window measurements. Journal
of Geophysical Research, 92, 3991–4000.

An algorithm to detect forest Ž res
3069
Justice, C. O., Kendall, J. D., Dowty, P. R., and Scholes, R. J., 1996, Satellite remote
sensing of Ž res during the SAFARI campaign using NOAA advanced very high
resolution radiometer data. Journal of Geophysical Research, 101, 23 851–23 863.
Kaufman, Y. J., Tucker, C. J., and Fung, I., 1990, Remote sensing of biomass burning in
the tropics. Journal of Geophysical Research, 95, 9927–9939.
Kennedy, P. J., Belward, A. S., and Gregoire, J.-M., 1994, An improved approach to Ž re
monitoring in West Africa using AVHRR data. International Journal of Remote Sensing,
15, 2235–2255.
Levine, J. S. (ed.), 1991, Global Biomass Burning: Atmospheric, Climatic, and Biospheric
Implications (Cambridge, Massachusetts: MIT Press).
Levine, J. S., Cofer, W. R., III, Cahoon, D. R., Jr, and Winstead, E. L., 1995, Biomass
burning: a driver for global change. Environmental Science and T echnology, 29, 120–125.
Li, Z., Cihlar, J., Moreau, L., Huang, F., and Lee, B., 1997, Monitoring Ž re activities in
the boreal ecosystem. Journal of Geophysical Research, 102, 29 611–29 624.
Li, Z., Nadon, S., Cihlar, J., and Stocks, B., 2000, Satellite mapping of Canadian boreal
forest Ž res: evaluation and comparison of algorithms. International Journal of Remote
Sensing, 21, 3071–3082 (this issue).
Malingreau, J. P., 1990, The contribution of remote sensing to the global monitoring of Ž res
in tropical and subtropical ecosystems. In Fire in the T ropical Biota, edited by J. G.
Goldammer, Ecosystem Processes and Global Challenges, Ecological Studies (Berlin:
Springer), pp. 337–370.
Pokrant, H., 1991, Land cover map of Canada derived from AVHRR images. Manitoba
Remote Sensing Centre, Winnipeg, Manitoba, Canada.
Rao, C. R. N., and Chen, J., 1996, Post-launch calibration of the visible and near-infrared
channels of the Advanced Very High Resolution Radiometer on NOAA-14 spacecraft.
International Journal of Remote Sensing, 17, 2743–2747.
Robertson, B., Erickson, A., Friedel, J., Guindon, B., Fisher, T., Brown, R., Teillet, P.,
D’Iorio, M., Cihlar, J., and Sancz, A., 1992, GEOCOMP, a NOAA AVHRR
geocoding and compositing system. Proceedings of the ISPRS Conference, Commission
2, Washington, DC (Washington: ISPRS), pp. 223–228.
Robinson, J. M., 1991, Fire from space: global Ž re evaluation using infrared remote sensing.
International Journal of Remote Sensing, 12, 3–24.
Scholes, R. J., 1995, Greenhouse gas emissions from vegetation Ž res in Southern Africa.
Environmental Monitoring and Assessment, 38, 169–179.
Sellers, P. J., Hall, F. G., Margolis, H., Kelly, B., Baldocchi, D., denHartog, J.,
Cihlar, J., Ryan, M., Goodison, B., Crill, P., Ranson, J., Lettenmaier, D., and
Wickland, D. E., 1995, The Boreal Ecosystem-Atmosphere Study (BOREAS): an
overview and early results from 1994 Ž eld year. Bulletin of American Meteorological
Society, 76, 1549–1577.
Setzer, A. W., and Malingreau, J. P., 1996, AVHRR monitoring of vegetation Ž res in the
tropics: toward the development of a global product. Biomass Burning and Global
Change, edited by J. S. Levine (MIT Press: Cambridge, Massachusetts), pp. 25–39.
Setzer, A. W., and Verstraete, M. M., 1994, Fire and glint in AVHRR’s channel 3: a possible
reason for the non-saturation mystery. International Journal of Remote Sensing, 15,
711–718.
Singh, K. D., 1993, The 1990 tropical forest resource assessment, Unasyva/FAO, 44, 10–19.