estimate_global_extent

advertisement
Estimating the Global Extent of Boreal Forest Burning Using AVHRR
1
R.H. Fraser1, Z. Li2, and J. Cihlar2
Intermap Technologies (research associate at Canada Centre for Remote Sensing), 588 Booth St., Ottawa,
ON, Canada K1A 0Y7; Email: rfraser@ccrs.nrcan.gc.ca; Tel: (613) 947-6613; Fax: (613) 947-1406
2
Canada Centre for Remote Sensing, Ottawa, ON, Canada
Abstract—In this analysis we synergistically combined an
AVHRR-based fire product, 10-day Normalised Difference
Vegetation Index (NDVI) composites, and global land cover
classification to produce the first satellite-based estimate of
global boreal forest burned area. The approach taken was to
eliminate spurious hotspots using an NDVI composite and
forest mask, then use remaining hotspots to “seed” an
iterative burn-growing algorithm. The global extent of boreal
burning in 1993 was estimated to be 6.0 million ha, a
relatively modest level of burning compared to a 1987
estimate for eastern Asia.
2.
INTRODUCTION
The boreal biome, comprising about 25% of the world’s
forest area, undergoes extensive periodic burning [1]. In
order to estimate the quantity of boreal fire emissions (smoke
aerosol, greenhouse gases) and their potential influence on
the earth’s radiation budget, techniques must be developed to
accurately measure the extent of burning.
Satellite-based strategies for assessing boreal burning have
relied on hotspot detection [2-3] and NDVI analysis [3-5].
Hotspot detection involves sensing thermal emissions from
active fires using the mid-infrared channel aboard NOAAAVHRR, ERS-ATSR, or GOES-VAS instruments. NDVI
analysis detects vegetation damage following fire, typically
by comparing pre- and post-burn vegetation conditions. Each
strategy has unique advantages; however, both are prone to
producing significant commission error [3-5].
In this
analysis, we synergistically combine these approaches to
produce a global estimate of boreal forest burned area for
1993.
3.
wide range of environments. We obtained lat/lon coordinates of all hotspots detected between April 1-Sept 1,
1993 from a preliminary release of the Fire Product.
Hotspots lying north of 45 were extracted, then
resampled to a 1-km grid registered to the equal-area,
Interrupted Goode Homolosine projection.
10-day NDVI composites. AVHRR 10-day NDVI
composites were obtained from the USGS EROS
Distributed
Active
Archive
Center
(http://edcwww.cr.usgs.gov/landdaac/).
Composites
were created by applying the maximum NDVI criterion
to AVHRR images collected over 10 day periods from a
network of 29 HRPT stations. Images underwent
standardised calibration, atmospheric correction, and
geometric registration to Goode Homosoline. Individual
10-day composites were found to be severely affected by
cloud contamination over boreal regions. The maximum
NDVI from four 10-day composites were thus used to
create a composite for the period Aug 21-Sept 30, 1993.
This post-burn composite precedes seasonal vegetation
senescence, yet includes most of the 1993 boreal fire
season.
NOAA/NASA Pathfinder 8-km resolution
NDVI composites were also examined, but were found to
show little or no signal from boreal burns due to cloud
contamination and spatial aggregation.
Land cover classification.
A Global Land Cover
Characteristics data set was produced by multitemporal,
unsupervised clustering of the 10-day NDVI composites
described above [7]. A mask of forested pixels lying
north of 45 was created by separating the five forest
types from 17 IGBP categories. Although the mask
included some non-boreal forest, the 1993 burned areas
almost entirely consisted of boreal forest.
SATELLITE DATA
BURNED AREA MAPPING
Three global-scale, 1-km resolution products derived from
NOAA-AVHRR were used in the burned area analysis:
1.
Fire algorithm product. The IGBP-DIS Global Fire
Product, produced by the Joint Research Centre of the
European Commission, is the first global-scale analysis
of fire (hotspot) distribution, covering an 18 month
period beginning April, 1992 [6]. The IGBP fire
algorithm detects potential hotspots using a mid-infrared
(3.7 m) threshold, then confirms them using contextual
information from background pixels. The contextual
tests are designed to allow the algorithm to adapt to a
The burn mapping procedure synergistically combines the
three AVHRR products described above. It borrows from the
HANDS (Hotspot and NDVI Differencing Synergy)
algorithm [8], with modifications designed to cope with
severe cloud contamination and noise inherent in the global
NDVI composites. For example, only a post-burn NDVI
composite is used rather than the difference between
anniversary date composites. The general strategy is to first
eliminate false hotspots using the forest mask, a filter, and
post-burn NDVI composite (steps 1-3 below). Remaining
confirmed hotspots are then used as “seeds” to initiate an
iterative, burn-growing algorithm that is supervised by the
post-burn NDVI (step 4). The processing steps are described
below and illustrated in Figure 1. The effect that each step
had on the computed burned area is shown in Table 1.
Thresholds were derived by examining selected burns
throughout the study region.
Step 1. Apply forest mask. In the Global Fire Product, the
IGBP algorithm was applied to all land cover types except
water and desert. The IGBP land cover classification was
therefore used to mask all other non-forest types from
consideration.
Step 2. Eliminate single hotspots. Single, non-connecting
hotspot pixels are eliminated since they often represent noise
in boreal environments (e.g., cloud edges or glint from small
lakes). In addition, unlike tropical forest, small burns (<1
km2) are responsible for only a fraction of boreal burned area.
If real hotspots are erroneously removed in this step, they are
likely recovered in step 4.
Step 3. Eliminate hotspots having high post-burn NDVI.
Hotspot pixels having a post-burn NDVI greater than 0.5 are
removed. Almost all real burned pixels were observed to
have late summer NDVI values less than 0.5.
Step 4. Apply burn-growing algorithm. After steps 1-3
eliminate most false hotspots, remaining hotspots typically
exhibit a scattered, patchy distribution within burns (Fig.1b).
To identify gaps missed by the hotspot algorithm, hotspots
are used to “seed” an iterative burn-growing algorithm. At
each iteration, a pixel is considered burned and is added to
the hotspot mask if it satisfies three conditions:
1. The pixel is connected either directly or diagonally to at
least one hotspot or previously identified burned pixel;
2. The pixel has an NDVI less than 0.5 in the post-burn
composite; and
3. The pixel’s post-burn NDVI is not larger by more than
0.05 from the mean post-burn NDVI of neighbouring
hotspots and previously identified burned pixels. This
condition is more stringent than condition 2, and ensures
that the growth algorithm does not extend beyond the
true burn perimeter. It also provides a spatially variable
threshold that can adapt to the unique vegetation and
burning conditions within each burn.
Based on the typical gap size between interspersed hotspots
(e.g., Fig.1b), five iterations of the growth algorithm were
found sufficient.
RESULTS AND DISCUSSION
The IGBP hotspot algorithm was previously found capable
of identifying the majority of burns within Canadian boreal
forest [5]. However, it also detected a large number of false
fires, appearing as either single pixels or clusters of pixels
that had burned the previous year. The 1993 IGBP hotspot
mask similarly includes a significant proportion of false
hotspots over boreal regions (Table 1). The burned area
algorithm used in the present analysis (steps 1-3) was
Table 1. Affect of each algorithm step on burned area
Algorithm step
1. All IGBP hotspots > 45N
2. Mask non-forest
3. Eliminate single hotspots
4. Post-burn, hotspot NDVI < 0.5
5. Apply growth algorithm (5)
Burned area (ha)
17,225,000
4,930,700
3,392,500
1,582,500
6,015,500
effective in eliminating most of these false fires, leaving a
patchy distribution of hotspots contained within real burns. A
burn-growing algorithm (step 4) was then applied to identify
any gaps that were not detected as hotspots. In most cases,
the procedure results in an accurate depiction of burned area
(Fig.1c) by comparison to the burned areas visible in the
post-burn NDVI composite (Fig.1a) and to the spatial
distribution of hotspots (Fig.1b).
World-wide boreal forest burned area in 1993 was
calculated to be 6.0 million ha, which is much larger than a
previous estimate of 2.4 million ha for 1992 [9], but
significantly smaller than a 1987 estimate (14 million ha) for
east-Asian boreal forest [10]. Burned area within Canada
was calculated to be 1.75 million ha, which agrees well with
the official figure of 1.84 million ha from the Canada Forest
Service [11]. Burning in Alaska affected 383,500 ha of
forest, an area 33 percent larger (288,589 ha) than that
recorded by the Alaska Fire Service [12]. Notwithstanding
the overestimate for Alaska, the global estimate is expected to
be conservative since the post-fire composite (Aug 21-Sept
30) preceded the end of fire season, and some burns
undoubtedly contained no hotspots.
While the burn mapping procedure was effective in
eliminating false hotspots and identifying missed burned
patches, its accuracy is ultimately constrained by the input
data sets. For example, burned patches were not consistently
visible in the post-burn NDVI composite due to cloud
contamination or weak NDVI contrast with non-burned
forest. In addition, geometric distortions caused by pixel
overlap, off-nadir pixel growth, and misregistration may blur
the burn signal [13]. In some regions, especially Alaska,
these problems caused the burn-growing algorithm to extend
beyond the true burn perimeter and overestimate burned area.
Conversely, some large burned patches appeared to be
underestimated where there was a small proportion of
hotspots detected due to cloud cover.
Despite these
limitations, the synergy created by combining the three
AVHRR data sets provides a more accurate representation of
burned areas than is possible from hotspot detection alone.
SUMMARY AND CONCLUSIONS
Three global-scale NOAA-AVHRR products were
combined to produce an estimate of worldwide boreal forest
burned area in 1993. Cloud contamination in 10-day NDVI
composites posed a significant problem, obscuring the burn
signal in several areas. A comparison of the burned area (6.0
million ha) to estimates for previous years underscores the
large inter-annual variability in boreal forest burning.
classification,” Int. J. Remote. Sens., vol. 15, pp. 3,4733,491, 1994.
ACKNOWLEDGMENTS
We thank the Global Vegetation Monitoring Unit (Space
Applications Institute, Joint Research Centre, Ispra, Italy) for
providing access to the preliminary version of the Global Fire
Product.
REFERENCES
[1] M.G. Weber and B.J. Stocks, “Forest fires and
sustainability in the boreal forests of Canada,” Ambio,
vol. 27, pp. 545-550, 1998.
[2] M.D. Flannigan, and T.H. Vonder Haar, “Forest fire
monitoring using NOAA satellite AVHRR,” Can. J. For.
Res., vol. 16, pp. 975-982, 1986.
[3] Z. Li, J. Cihlar, L. Moreau, F. Huang, and B. Lee,
“Monitoring fire activities in the boreal ecosystem,” J.
Geophys. Res., vol. 102, pp. 29,611-29,624, 1997.
[4] E.S. Kasischke, and N.H. French, “Locating and
estimating the areal extent of wildfires in Alaskan boreal
forests using multiple-season AVHRR NDVI composite
data,” Remote Sens. Environ., vol. 51, pp. 263-275,
1995.
[5] Z. Li, S. Nadon, B. Stocks, and J. Cihlar, “Satellite
detection of Canadian boreal forest fires part 2:
algorithm validation and comparison,” unpublished.
[6] E. Dwyer, J-M. Gregoire, and J-P. Malingreau, “Global
analysis of vegetation fire using satellite images: spatial
and temporal dynamics,” Ambio, vol. 27, pp. 175-181,
1998.
[7] J.C. Eidenshink, and J.L. Faundeen, “The 1-km AVHRR
global land data set: first stages in implementation,” Int.
J. Remote Sens., vol. 15, pp. 3,443-3,462, 1994.
[8] R.H. Fraser, Z. Li, and J. Cihlar, “Hotspot and NDVI
differencing synergy (HANDS): a new technique for
burned area mapping,” unpublished.
[9] D.R. Cahoon, B.J. Stocks, J.S. Levine, W.R. Cofer, and
J.A. Barber, “Monitoring the 1992 forest fires in the
boreal ecosystem using NOAA AVHRR satellite
imagery,” in Biomass Burning and Global Change. J.S.
Levine, Ed. Cambridge:MIT Press, 1996, pp.795-801.
[10] D.R. Cahoon, B.J. Stocks, J.S. Levine, W.R. Cofer, and
J.M. Pierson, “Satellite analysis of the severe 1987
forest fires in northern China and southeastern Siberia,”
J. Geophys. Res., vol. 99, pp. 18,627-18,638, 1994.
[11] Canadian Council of Forest Ministers. National Forestry
Database Program. Canadian Forest Service, Natural
Resources Canada.
[12] Personal communication, Alaska Fire Service, Bureau
of Land Management, US Department of the Interior.
[13] A. Moody, and A.H. Strahler, “Characteristics of
composited AVHRR data and problems in their
Figure 1 (a) post-burn NDVI composite where lighter greyshades represent smaller NDVI; (b) confirmed fire algorithm
hotspot pixels (after applying steps 1-3); (c) final burned area
mask after applying five iterations of burn-growing algorithm
(step 4). The burns are located in eastern Russia.
Download