Full text of DOI:10.1080/01431160050144965

This article was downloaded by: [Umeå University Library]
On: 15 November 2014, At: 17:17
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 mapping of
Canadian boreal forest fires:
Evaluation and comparison of
algorithms
Z. Li a , S. Nadon b , J. Cihlar a & B. Stocks c
a Canada Centre for Remote Sensing , 588 Booth,
Ottawa, Ontario, Canada , K1A OY7
b Environment Canada , Ottawa, Ontario,
Canada , K1A OH3
c Canadian Forest Service , Sault Ste Marie,
Ontario, Canada
Published online: 25 Nov 2010.
To cite this article: Z. Li , S. Nadon , J. Cihlar & B. Stocks (2000) Satellite-
based mapping of Canadian boreal forest fires: Evaluation and comparison of
algorithms, International Journal of Remote Sensing, 21:16, 3071-3082, DOI:
10.1080/01431160050144965
To link to this article: http://dx.doi.org/10.1080/01431160050144965
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, 3071–3082
Satellite-based mapping of Canadian boreal forest Ž res: evaluation and
comparison of algorithms
Z. LI†*, S. NADON‡, J. CIHLAR† and B. STOCKS§
†Canada Centre for Remote Sensing, 588 Booth, Ottawa, Ontario,
Canada, K1A OY7
‡Environment Canada, Ottawa, Ontario, Canada, K1A OH3
§Canadian Forest Service, Sault Ste Marie, Ontario, Canada
(Received 27 September 1998; in Ž nal form 4 June 1999)
Abstract. This paper evaluates annual Ž re maps that were produced from
/
NOAA-14 AVHRR imagery using an algorithm described in a companion paper
(Li et al., International Journal of Remote Sensing, 21, 3057–3069, 2000 (this issue)).
Burned area masks covering the Canadian boreal forest were created by composit-
ing the daily maps of Ž re hot spots over the summer and by examining Normalized
DiŒerence Vegetation Index (NDVI) changes after burning. Both masks were
compared with Ž re polygons derived by Canadian Ž re agencies through aerial
surveillance. It was found that the majority of Ž re events were captured by the
satellite-based techniques, but burnt area was generally underestimated. The burn
boundary formed by the Ž re pixels detected by satellite were in good agreement
with the polygons boundaries within which, however, there were some Ž res missed
by the satellite. The presence of clouds and low sampling frequency of satellite
observation are the two major causes for the underestimation. While this problem
is alleviated by taking advantage of NDVI changes, a simple combination of a
hot spot technique with a NDVI method is not an ideal solution due to the
introduction of new sources of uncertainty.
In addition, the performance of the algorithm used in the International
Geosphere–Biosphere Programme (IGBP) Data and Information System (IGBP-
DIS) for global Ž re detection was evaluated by comparing its results with ours
and with the Ž re agency reports. It was found that the IGBP-DIS algorithm is
capable of detecting the majority of Ž res over the boreal forest, but also includes
many false Ž res over old burned scars created by Ž res taking place in previous
years. A step-by-step comparison between the two algorithms revealed the causes
of the problem and recommendations are made to rectify them.
1. Introduction
Forest Ž res can have major environmental and economic impact. In Canada
alone, about 9000 Ž res, on average, burn every year consuming an average of about
2.5 Mha of forests (Stocks et al. 1996, Li et al. 2000). Such average statistics are,
however, of limited signiŽ cance, as the area burned in boreal forests shows a tremend-
ous inter-annual variation. For example, the total burnt area across Canada can
diŒer by a factor of more than 12 within a couple of years, e.g. 5.1 Mha in 1995 and
*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

3072
Z. L i et al.
0.4 Mha in 1997 according to satellite-based estimation (Li et al. 2000). Much of the
area burned was observed in remote regions of Canada.
Information on wild Ž res can be collected by a variety of means. Among those,
the most frequently used are traditional ground-based in situ observation, air-borne
human and photographic surveillance, and space-borne remote sensing. The quality,
density and frequency of ground-based and air-borne Ž re observations vary consider-
ably from one country or region to another. Only satellites oŒer the potential to
acquire global uniform Ž re information on a repetitive basis. However, like many
optical remote sensing techniques, satellite Ž re detection algorithms have some
disadvantages.
The largest problem is cloud cover below which no Ž res can be detected, an
inherent limitation for most optical remote sensing applications. A lack of temporal
sampling and a limited range of radiometric sensitivity are the major shortcomings
of the Advanced Very High Resolution Radiometer (AVHRR), which has been most
widely used for Ž re detection (Justice et al. 1996). The quality of Ž re data inferred
from satellites thus warrants thorough assessment. Unfortunately, only a handful of
studies using satellite sensor data to detect and map Ž res include validation exercises
due to limited ground data on actual Ž res (e.g. Setzer et al. 1994, Li et al. 1997).
This study presents an in-depth evaluation of a Ž re database derived from daily
AVHRR satellite images (Li et al. 2000). The database was created by applying a
Ž re-detection algorithm that was developed at the Canada Centre for Remote Sensing
(CCRS) which is hereafter referred to as the CCRS Ž re-detection algorithm (CFDA).
Section 2 compares yearly accumulated areas of satellite-detected Ž re pixels to annual
burnt area as reported by Canadian Ž re management agencies. The accumulated
areas of ongoing Ž res are also compared with the areas of Ž re scars detected using
the Normalized DiŒerence Vegetation Index (NDVI) in §3.
In addition, §4 of this paper presents a comparison of the CFDA to a global Ž re-
detection algorithm developed and used in the International Geosphere–Biosphere
Programme Data and Information System (IGBP-DIS) (Justice et al. 1996,
Malingreau and Gregoire 1996, IGBP 1997). The algorithm was intended to generate
a global coherent and consistent biomass burning dataset. The major diŒerence
between the CFDA and the IGBP algorithm is that the CFDA is a traditional
threshold algorithm using Ž xed values while the IGBP algorithm is a contextual
algorithm that relies on local background information to dynamically set thresholds.
As a result, the IGBP algorithm was designed to be self-adaptive and consistent
over large areas under diŒerent environmental conditions (Flasse and Ceccato 1996).
Careful evaluation of this algorithm in diŒerent parts of the world is essential to
achieve the goal of the IGBP Ž re project. The objective of §4 is to test if the IGBP
algorithm is self-adaptive enough to be eŒective over Canadian boreal forests.
2. Evaluation of Ž re hot spots
Canada routinely gathers and maintains rather detailed and complete records of
forest Ž res, which enables a comprehensive evaluation of satellite detection results,
thanks to provincial and territorial Ž re management agencies. Fire polygon data
were acquired by means of airborne infrared mapping. Data from all agencies were
compiled to determine the total area burned in Canada. No standard data format
exists among the provinces and territories and the data quality and availability may
vary from one province/territory to another. The Canadian Forest Service (CFS)
established a National Forestry Database Program to gather, process and standardize

Comparing algorithms that detect forest Ž res
3073
Ž re information in digital format and distribute them in a geographic information
system (GIS). Such data in two regions of Northwest Territories and Saskatchewan
in 1995 were employed in this investigation. Each of the two regions contains
1375 1200 1 km pixels.
Ö
Figure 1 presents comparisons of these data to the annual composites of Ž re
pixels detected by satellite over the two regions. The blue polygons outline the
boundaries of burnt area reported by Ž re agencies while red dots are locations of
active Ž re pixels detected during the entire Ž re season. In Saskatchewan, 47 major
Ž re events were reported by Ž re agencies and all of them were detected by the CFDA.
Figure 1. Comparison of the annual composite of Ž re spots with data obtained by Ž re
agencies for two regions in Canada in 1995. The blue polygons outline the boundaries
of burnt area reported by Ž re agencies using conventional means of monitoring while
red dots are locations of active Ž re pixels observed by satellite during an entire
Ž re season.

3074
Z. L i et al.
In the Northwest Territories, there are 24 large Ž res reported by Ž re agencies, six of
which were missed by satellite. However, the missed Ž res covered small areas on the
order of 1000 ha. Combining data from the two regions, we calculated that 7.3% of
the pixels showing Ž re activity fall outside the burnt area boundaries reported by
Ž re agencies. We visually inspected satellite images for all the cases in which satellite
detects Ž re activities which were not reported by Ž re agencies. It was determined
that 20% of these pixels have smoke associated with them. Most of the large Ž re
clusters detected by satellite show some smoke phenomena. These seemingly real
Ž res were not reported by Ž re agencies. Only 5.8% of the pixels that are detected
by satellite as Ž res are not certain. Their sizes are very small (smaller than 300 ha).
Therefore, the probability that pixels detected by satellites as Ž res are true Ž res
(
>
300 ha) is as high as 94%.
Figure 2 is a histogram showing the frequency of pixels detected as Ž res inside
the Ž re polygons. This Ž gure shows that about 60% of the Ž re pixels were detected
only once and 26% twice. This suggests that the success of Ž re monitoring from
space depends very much on the frequency of satellite overpass, the speed of Ž re
spreading and the amount of cloud cover. These factors partially explain why the
yearly hot spot composite is patchy rather than continuous. On the other hand,
however, the area inside the polygons reported by Ž re agencies is not necessarily
completely burned, as is shown in the next section.
3. Comparison against satellite-detected Ž re scars
The potential of the CFDA in mapping burnt area is also evaluated by comparing
areas of Ž re hot spots detected from individual AVHRR images with Ž re scars
detected from AVHRR clear-sky composites. Several studies have shown that Ž re
scars may be visible from the diŒerence in vegetation indices before and after burning
in boreal and temperate forests (Cahoon et al. 1992, Kasischke et al. 1993, Kasischke
and French 1995, Li et al. 1997). Here, we examine changes in NDVI between two
consecutive years. Clear-sky composite data of maximal NDVI over 10-day periods
are available across Canada since 1993 following a series of corrections (Cihlar et al.
1997). Twenty 10-day composites from 20 April to 10 November were processed
every year except for 1994 when the data for the last Ž ve composites were unavailable.
Figure 2. Histogram of the frequency of satellite-detected Ž re pixels inside the polygons
shown in Ž gure 1.

Comparing algorithms that detect forest Ž res
3075
Under normal conditions, NDVI exhibits a clear seasonal cycle. It increases gradually
in spring, reaches a peak value in August, and then declines quickly in autumn. Such
a seasonal cycle may shift earlier or later depending on climate.
Following a Ž re event, NDVI has a sharp drop. The method used here is intended
to detect all new Ž re scars in a particular year. In order to do so, two pairs of NDVI
images were employed, one in spring and another in autumn. Note that the years
of the two consecutive periods are diŒerent. For example, to determine the 1995 Ž re
scars as shown here, the spring pair comes from 1995 and 1996 images, while the
autumn pair consists of 1994 and 1995 images. Comparing NDVI images from the
same period of the year, as opposed to comparing spring and autumn images from
the same year, can overcome the problem associated with the annual NDVI cycle.
Comparisons of NDVI for two periods reduce the noise introduced by the inter-
annual variation of NDVI. The spring and autumn periods were chosen to be 21–31
May and 11–20 September, respectively. Contained within the two periods is the
bulk of a Ž re season without snow cover in the region of interest. Snow cover over
forest diminishes the NDVI drastically.
The diŒerence in NDVI was computed pixel by pixel for each pair of the images.
A threshold was then used to identify pixels with scar signature. After some tests, a
relative NDVI change of 9% was found to separate most Ž re scars from the back-
ground. Pixels that showed a relative NDVI drop greater than 9% in both pairs of
images were marked as Ž re scar pixels.
Figure 3 presents a comparison of the results obtained with the CFDA and the
scar detection method in the same two regions as shown earlier. Three colours are
used to denote Ž re pixels detected by the CFDA alone (red), the scar method alone
(green) and by both methods (yellow), in comparison with the blue Ž re boundaries
reported by Ž re agencies. It is evident that the Ž re scars determined from NDVI
agree well with the areas of the polygons. Fire pixels determined by the CFDA (red
and yellow) occupy about 60% of the area inside the blue outlines. The Ž re scars
(green and yellow) derived from NDVI cover about 63% of the total area of the
polygons. About half of the total Ž re pixels are identiŽ ed by both methods. These
overlapped pixels have the highest probability of being true Ž res. On the contrary,
pixels inside the polygons that are not marked by either algorithm are more likely
to be unburned patches. They occupy 19% of the total areas of the polygons, with
1.4% identiŽ ed as water bodies.
Limitations in the detection methods caused a signiŽ cant portion of the burnt
area to be detected by only one method. For the CFDA, cloud cover and insu cient
temporal sampling were dominant factors. For the NDVI scar detection method,
sub-grid burning and burns of little damage (most likely surface Ž res) were hard to
detect. In addition, a change in NDVI may be caused by something other than a
Ž re, such as drought, tree diseases, insects, etc. The most prominent problem in the
NDVI approach is probably associated with cloud contamination, which can reduce
NDVI value considerably. Although great care was exercised to remove cloud
contaminated pixels in generating clear composite data (Cihlar et al. 1997), there is
no assurance that all pixels are cloud-free, in particular for residual clouds and
persistent clouds. These limitations lead to a considerable number of scattered false
Ž re pixels that are evident in Ž gure 3. Therefore, a simple combination of the two
methods may not be ideal in mapping burned area, as both real and false Ž res
accumulate concurrently. As such, a new synergistic method based on hot spot and
NDVI has been developed that takes advantage of the strengths of each method,
while at the same time avoiding their weakness (Fraser et al. 1999).

3076
Z. L i et al.
Figure 3. Comparison of the annual composite of Ž re spots for 1995 with Ž re scars detected
using NDVI in two regions of Canada. Red dots represent Ž re pixels detected by
single-day AVHRR algorithm only, while green dots are Ž re pixels identiŽ ed using
NDVI only. Yellow represents Ž re pixels that are detected by both techniques.
4. Comparison with IGBP algorithm
The DIS group of the International Geosphere–Biosphere Program (IGBP-DIS)
developed an algorithm for global Ž re moritoring using AVHRR data. To date, this
algorithm has been employed to process NOAA-11 data over most parts of the
tropical regions for a period of 21 months between April 1992 and December 1993
(IGBP-DIS WWW site: http://www.mtv.sai.irc.it/projects/Ž re/gfp/home.html).
The IGBP algorithm was designed to generate a global, uniform Ž re database in
order to provide better information on the spatial and temporal distribution of Ž re
at a global scale. The CFDA algorithm was designed for speciŽ c use in boreal forests.

Comparing algorithms that detect forest Ž res
3077
For the sake of understanding and comparing the performance of the two algorithms,
the IGBP algorithm is described Ž rst with reference to CFDA.
4.1. Comparative description of the IGBP algorithm
In an attempt to avoid the need for adjusting the algorithm over diŒerent regions
or ecosystems, the method employed by IGBP is a contextual approach (Flasse and
Ceccato 1996, Justice et al. 1996, IGBP 1997). The decision to designate a pixel as
a Ž re or not is based on comparison of its radiometric signature with those from
surrounding pixels. The Ž rst step of the method applies absolute thresholds to mark
all pixels that have:
Õ
T
311 K and T
T
8 K
>
(1)
>
3
as potential Ž res, where T and T are the brightness temperatures in AVHRR
3
4
3
4
Õ
channel 3 and 4. The second step consists of contextual tests whereby T and T
T
3
3
4
for potential Ž res are compared with the corresponding average values calculated
from the pixels surrounding the potential Ž re pixels under study. Pixels that are
potential Ž res, water or clouds are excluded from the computation. The remaining
pixels are simply referred to as normal background pixels. Water pixels are identiŽ ed
using a land-water mask while cloud pixels are identiŽ ed following a threshold test.
The contextual tests conŽ rm a potential Ž re pixel as being true if it passes the
following criteria tests :
1
[
T
]
Õ
T
max {T
2 s
1
, 8 K}
(2)
(3)
>
>
3
4
34bg
34bg
and
T
T
2 s
3 K
1
1
3
3bg
3bg
where T
is the mean background brightness temperature in channel 3; s
[
the
]
3bg
3bg
Õ
standard deviation of background T ; T
the mean value of background T
T
;
3
34bg
3
4
[
the standard deviation of background T
]
. The background pixels
used to compute these statistical thresholds are from a window centred at the
Õ
and s
T
34bg
3
4
potential Ž re pixel under testing. The window has an initial dimension of 3 3 pixels
Ö
Ö
and is allowed to grow gradually up to 15 15 pixels, or until at least 25% of the
pixels inside the window qualify as normal background pixels. If the window reaches
the largest size and less than 25% of the pixels inside are deemed as background,
the potential Ž re pixel under study is not conŽ rmed.
It follows that the thresholds are dynamic variables, in contrast to the Ž xed
values used in CFDA. To gain a sense of their magnitude in the boreal forest
application, the values obtained during the execution of the IGPB algorithm over a
scene of 1200 1200 pixels on 24 June 1995 were saved and analysed. Table 1
Ö
delineates the overall statistics of these parameters for the 48 214 pixels identiŽ ed as
potential Ž res.
To put these values in context, they are compared with those used in the CFDA,
in table 2. Table 2 also provides a complete comparison of all the tests used in the
two algorithms. The order of the tests as listed is not necessarily the same as that
used in the code. Similar tests have been grouped together on the same line of the
1
method in which 3 K was used instead of 8 K, which contradicts equation (1). If a pixel is
There appears to be an error in the original document (IGBP 1997) describing the IGBP
Õ
declared as a potential Ž re by (1), T
T must be larger than 8K.
3
4

3078
Z. L i et al.
Table 1. The statistics of the IGBP contextual thresholds obtained over a boreal forest
region.
Parameter
Mean value (K)
SD (K)
T
7.9
1.6
310
2.1
11
2.9
1.8
3
1.3
6
34bg
s
34bg
T
3bg
s
T
T
3bg
1 2s
34bg
1 2s 1 3
3bg
34bg
3bg
317
3
Table 2. A comparison of CFDA and IGBP algorithms.
CFDA NOAA-14 IGBP algorithm
>
Õ
>
T
T
315 K
potential Ž re test: T
311 K
3
317 K (mean)
3
3
>
contextual test: T
potential Ž re test: T
contextual test: T
3
Õ
<
>
T
14 K
T
8K
4
11 (mean)
4
Õ3
>
T
3
4
<
eliminate cropland, grassland and water
<
0.20
eliminate water
R
0.22
R
2
2
Õ
Õ
<
T
–
T
–
–
T
4.1 K and T
T µ 19 K
–
4
4
5
3
4
Õ
sun-glint: |R
cloud: T µ 265 K
cloud: R 1 R # 1.20
cloud: R 1 R # 0.8 and T µ 285 K
R |µ 0.02
1
2
>
260 K
5
1
2
2
1
5
eliminate single pixels
–
table. The listed contextual values are averages and the actual values used in the
test vary from pixel to pixel. The range of variation is indicated by the standard
deviations given in table 1.
The similarities and discrepancies between the two algorithms are evident from
table 2, which helps understand the diŒerences between the results obatined by the
two Ž re-detection algorithms. Some key tests are common but may be in slightly
diŒerent forms. The IGBP algorithm has more tests to eliminate cloud pixels. The
IGBP sun-glint test was evaluated and was found to be ine cient for the current
application. As a result of the contextual tests, execution of the IGBP algorithm is
much more time-consuming than for the CFDA (10 times slower). This poses a
concern for real-time Ž re detection, for which the CFDA is designed.
4.2. Performance of the IGBP algorithm
Figure 4 presents the results of Ž re detection using the IGBP algorithm for the
1995 Ž re season over the same two regions as shown in Ž gure 1. The burnt areas
from Ž re agencies again are depicted by the blue polygons. Fire pixels identiŽ ed by
the algorithm cover 48% of the area inside the Ž re polygons, compared with 60%
for the CFDA. On the other hand, many pixels outside of the boundaries of the
reported Ž res are marked as Ž res by the IGBP algorithm. With reference to Ž gure 1,
the proportion of false Ž res is considerably higher with the IGBP algorithm than
with the CFDA. Figure 5 shows Ž re pixels detected by CFDA for both 1995 and
1994. A comparison between Ž gures 4 and 5 indicates that the majority of the Ž re
pixels outside of the polygons identiŽ ed by IGBP algorithm are actually scars from

Comparing algorithms that detect forest Ž res
3079
Figure 4. Same as Ž gure 1 but using the IGBP Ž re-detection algorithm.
1994 Ž res. All major burns shown in Ž gure 5 for 1994 were veriŽ ed against ground
data from Ž re agencies.
The reason that the IGBP algorithm misidentiŽ es old Ž re scars as active Ž res
appears to be that the thresholds for potential Ž re tests are set too low, which tends
to confuse radiometric signals from fresh Ž re scars with on-going Ž res. Fresh scars
have very low albedos and thus heat up more than forest, leading to a high brightness
temperature. As a result, the test of T for potential Ž res can be readily passed. The
3
Õ
second test based on T
T
is presumably insu cient to eliminate such a warm
3
4
background. Since the statistics calculated for background pixels do not include
potential Ž re pixels, many pixels surrounding the potential Ž re under examination
are disqualiŽ ed. The algorithm continues searching for colder background pixels.

3080
Z. L i et al.
Figure 5. The distribution of Ž re pixels detected with the CFDA algorithm in 1995 (red dots)
and in 1994 (grey dots).
At the end, the values of T
ground values, which favours the conŽ rmation of these potential Ž re pixels by the
and T
obtained are lower than the actual back-
3bg
34bg
contextual tests.
These results suggest that the IGBP algorithm is not self-adaptive enough to be
used universally and blindly. After all, the IGBP algorithm relies on many Ž xed
thresholds. Adjustments to these thresholds may be necessary. In this case, increasing
Õ
the thresholds of T and T
T would remedy the problem.
3
3
4
5. Conclusion
In this paper we evaluated the peformance of an algorithm developed to detect
Ž res in the Canadian boreal forest based on NOAA AVHRR data. The evaluation

Comparing algorithms that detect forest Ž res
3081
used satellite sensor data and ground observations over parts of western Canada in
1995. The areas and locations of Ž res detected with a conventional method and from
AVHRR satellite data (both individual daily images and 10-day clear composites)
were compared. In addition, the satellite Ž re-detection algorithm that we developed
(CFDA) and that designed for use by IGBP were compared. The Ž ndings of the
study are summarized as follows.
E
In comparison to the Ž res reported by Canadian Ž re agencies and those
indicated by smoke plumes, the likelihood that a pixel was correctly identiŽ ed
by the CFDA as Ž re is as high as 94%.
The majority of Ž re events are captured by satellite, but the area of burning
E
~
given by an ensemble of Ž re pixels may be underestimated considerably ( 35%
on average) by satellite due to cloud cover and low frequency of satellite
overpass.
E
E
The combination of active Ž re detection and Ž re scar identiŽ cation may consid-
erably reduce the area of missed Ž res, but also concurrently increase the number
of false Ž res.
Relative to CFDA, the IGBP Ž re-detection algorithm developed for global
applications suŒers larger commission and omission errors when applied to
boreal forest. It tends to mark some of the burned scars from previous year as
active Ž res. Re-tuning of certain thresholds would be required to achieve high
accuracy.
The study suggests that Ž re detection in boreal forests using AVHRR data is a
viable strategy for obtaining timely and consistent Ž re information with acceptable
accuracy. The major limitations of this data source are spatial resolution and clouds,
which interfere with obtaining local detail and high temporal frequency, respectively.
Future satellite sensors such as MODIS will reduce the former deŽ ciency, while the
latter requires new techniques such as the synergic combination of hot spots and
NDVI diŒerencing (Fraser et al. 1999). At any rate, the study demonstrates that
satellite sensor data can serve as an important tool for acquiring timely and moder-
ately accurate information on boreal forest Ž res for strategic and management
applications.
Acknowledgements
The authors are grateful to John Mason, Sun Hua, Norm Mair and Bruno Croft
for providing us with ground Ž re data, and Robert Fraser for comments.
References
Cahoon, D. R., Jr, Stocks, B. J., Levine, J. S., Cofer, W. R., III, and Chung, C. C., 1992,
Evaluation of a technique for satellite-derived area estimation of forest Ž res. Journal
of Geophysical Research, 97, 3805–3814.
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.
Flasse, S. P., and Ceccato, P. S., 1996, A contextual algorithm for AVHRR Ž re detection.
International Journal of Remote Sensing, 17, 419–424.
Fraser, R., Li, Z., and Cihlar, J., 1999, A new technique for mapping burns: hot spot and
NDVI diŒerencing synergy (HANDS). Remote Sensing of Environment, in press.
IGBP, 1997, DeŽ nition and implementation of a global Ž re product derived from AVHRR
data. IGBP-DIS Working Paper no. 17.

3082
Comparing algorithms that detect forest Ž res
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.
Kasischke, E. S., and French, N. H. F., 1995, Locating and estimating the areal extent of
windŽ res in Alaskan boreal forests using multiple-season AVHRR NDVI composite
data. Remote Sensing of Environment, 51, 263–275.
Kasischke, E. S., French, N. H. F., Harrell, P., Christensen, N. L., Jr, Ustin, S. L., and
Barry, D., 1993, Monitoring of wildŽ res in boreal forests using large area AVHRR
NDVI composite image data. Remote Sensing of Environment, 45, 61–71.
Li, Z., Cilar, 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., and Cihlar, J., 2000, Satellite-based detection of Canadian boreal forest
Ž res: development and application of the algorithm. International Journal of Remote
Sensing, 21, 3057–3069.
Malingreau, J. P., and Gregoire, J.-M., 1996, Developing a global vegetation Ž re monitoring
system for global change studies: a framework. In Biomass Burning and Global Change,
edited by J. S. Levine (Cambridge, MA: The MIT Press), pp. 14–24.
Setzer, A. W., Pereira, A. C., and Pereira, M. C., 1994, Satellite studies of biomass burning
in Amazonia: some practical aspects. Remote Sensing Reviews, 10, 91–103.
Stocks, B. J., Lee, B. S., and Martell, D. C., 1996, Some potential carbon budget implications
of Ž re management in the boreal forest. In Forest Ecosystems, Forest Management and
the Global Carbon Cycle, edited by M. J. Apps and D. T. Price, NATO ASI Series,
Sub-series 1, vol. 40, Global Environmental Change (Berlin: Springer), pp. 89–96.