|
|
|
Address for correspondence:
Eric S. Kasischke
P.O. Box 134001
Ann Arbor, MI 48113-4001
However, the total area contained within the fire boundaries mapped
by AVHRR were only 61% of those mapped by the field observers. However,
the AVHRR data used in this study did not span the entire time period during
which fires occurred, and it is believed the areal estimates could be improved
significantly if an expanded AVHRR data set were used.
During the summer of1991, the EROS Data Center (EDC) of the U.S. Geologic Survey produced a set of composite images of the state of Alaska from data collected by NOAA's Advanced Very High Resolution Radiometer (AVHRR) during 1990 and 1991. The images were generated from AVHRR source data of 1.1 km resolution, which were resampled to 1 km x 1 km pixels and composited over 15-16 day intervals to create the bimonthly data sets. This compositing technique utilized the maximum normalized vegetation index (NDVI) "greenness" value as the basis of selection of pixels for each image. This methodology is the same as that used to produce the biweekly composite AVHRR images of the conterminous United States (Sadowski and Watkins, 1991; Eidenshink, 1992), which have been recently used to develop a land-cover characteristics data set for the conterminous U.S. (Loveland et al., 1992). In this article, we examine the NDVI composite imagery produced from source data collected during the summer of 1990.
Figure 1 presents two of the NDVI composite images produced by EDC. One image was generated from data collected between 15 and 30 June 1990, while the other was generated from data collected between 1 and 15 August 1990. These images are from the east central basin of Alaska between the Alaska Range' on the south and the Brooks Range on the north, an area where a high percentage of the wildfires in Alaska occur. It will be shown in this article that the large areas in the August image exhibiting a significant decrease in the NDVI value correspond with the boundaries of the large wildfires that occurred in this region during the 1990 fire season. Although the AVHRR is not directly detecting the fire itself, it is detecting a scar left on the landscape which is the result of severe damage to the vegetation cover. We utilize this scar as a surrogate for estimation of the areal extent of the fire.
The overall goal of the research presented in this article is to evaluate the utility of composite-AVHRR data for determining the locations and mapping the areal extent of wildfires (via detection of burn scars) in boreal forest ecosystems. More specifically, this research focused on answering the following questions:
1. What is the detection accuracy on AVHRR imagery for fires >400 ha in size?
2. How does the total areal extent estimated from AVHRR data compare to that estimated from field records?
3. For larger fires (> 10,000 ha), how do the boundaries mapped from
AVHRR images compare to those mapped from field' records?
A major factor in many ecosystem processes within boreal forests is the widespread occurrence of wildfires (Wein and MacLean, 1983; Dyrness et al., 1986), which play an especially important role in nutrient cycling and the release of trace gases into the atmosphere (Crutzen et al., 1979; Mooney et al., 1987).
The fire regime in the interior forests of Alaska is not dissimilar to forests found in the northern regions Canada and Eurasia. We shall use data on Alaska fires to illustrate some properties of boreal forests. In Figure 2a we present a plot of the total areal extent of the boundaries of wildfires in Alaska since 1940 (Alaska Fire Service, 1991; Gabriel and Tande, 19&3). Over this time period, 17.6 million ha of land were subjected to wildfire, representing approximately 18% of the boreal forest region of Alaska. Over 75% of the area burned in the past half century occurred during fires occurring in only 12 years (Fig. 2a) within this time period (e.g., 1940, 1941, 1946, 1947, 1950, 1954, 1957, 1969, 1977, 1988, 1990, 1991).
The occurrence of fires in Alaska in any given year is highly dependent upon the availability of moisture during the growing season (Dyrness et. al., 1986), with severe fire years occurring during the drier, warmer springs and summers. During severe fire years, large fires are the rule rather than the exception. For example, during the years 1988,1990, and 1991, a total of 2.84 million ha of boreal forest burned in the state of Alaska. Of this total area, 1.6% occurred in fires less than 1,000 ha in size, 98% occurred in fires greater than 1,000 ha in size, 85% greater than 10,000 ha in size, and 66% occurred in fires greater than 25,000 ha in size. The four largest fires during this time period were 117,000 ha, 162,000 ha, 186,000 ha, and 219,000 ha, respectively. Each of these fires represents a larger area than was burned during the entire fire season from 25 other years during the last half century.
From a climate change standpoint, not only is the study of the occurrence of fire in boreal forests important for the cycles of greenhouse gases, but they may also be an indicator of changes in climate patterns. This results from the sensitivity of fire in boreal forests to climate conditions (Wien and MacLean, 1983). Figure 2b illustrates the longer-term trends in the fire history record. This plot was generated by a running average of the area burned for a given year with the areas of the previous two and subsequent two fire seasons. This plot shows there were four major fire periods during the past half century, the result of either extremely high individual fire years (e.g., 1957 and 1969) or several significant fire years grouped together (e.g., 1940 and 1941; and 1988, 1990, and 1991). This plot suggests there is a longer-term periodicity to climatic factors controlling wildfires in boreal forests.
Although adequate historical and annual records are available for Alaska, information concerning the total area of boreal forests burned on a global basis is not easily obtained. Crutzen et al. (1979) estimate that between 1 million ha and 1.5 million ha of boreal forest burn every year. However, given that the boreal forests of Alaska represent approximately 4% of the world's total, extrapolation of the area burned in Alaska each year (340,000 ha per year average since 1940) leads to an estimate of 8.5 million ha per year on a global scale.
The AVHRR system has the capability to map the entire world, including
its boreal forests in an efficient and timely manner. The National Aeronautics
and Space Administration (NASA) is using this capability to coordinate
the collection of a daily global AVHRR for an 18-month period spanning
1992 and 1993 as parts of its Earth Observing System. Because a significant
portion of the burned boreal forest occurs in events which cover large
areas, the 1 km AVHRR data may provide a useful means for estimation of
forest fire extent in northern regions on a global basis. In addition,
analysis of these data may provide additional information on the severity
of the fires, as well as on the patterns of regrowth of vegetation on these
sites. Such information is important in a variety of ecosystem process
studies, and is not readily derived from ground-based records and/or studies.
There are several methods used to generate the fire boundary map. For those fires which are being actively managed, careful records of the boundaries of fires are maintained because of the need to precisely locate areas requiring fire suppression efforts. Upon completion of an event, the final boundary maps are transferred to permanent record. For large fires which are not actively managed, a fire observer maps the boundary of the fires from observations made from a light aircraft.
With respect to comparison with the AVHRR data set, there are several aspects of the accuracy of the AFS fire maps which must be considered. First, there is the question of how well does the outer edge of the fire boundary mapped by visual observation actually match reality. We feel that in most cases, the boundaries match fairly well, to within 1-2 km. We base this assessment on the fact that there is a high degree of terrain relief or the presence of significant land attributes (e.g., lakes, rivers, etc.) to provide benchmarks to the field observers. In addition, most field observers have a large degree of familiarity of the area they are working in.
The second question of accuracy has to do with the characteristics of the fires themselves. Fires do not burn uniformly throughout an area, and thus the resultant scarring to the vegetation, which is detected on AVHRR imagery, will not be uniform. A portion of the fire area mapped by the AFS is likely to have little or no fire scar because of little fire impact on a certain portion of the area. Thus, we would expect the areal extent of the fires estimated by the AFS to be an upper bound. How much of an overestimation these boundaries represent is difficult to estimate without further research.
The fire boundaries were digitized and entered into a computer geographic
information system (GIS) data base. Information on each mapped fire was
added to this database for use in the analysis. This additional information
included the start and stop dates for each fire, fire-boundary size, and
the latitude and longitude of the origin point for each fire (as provided
by the AFS). Using the data base, the boundaries from the different fires
were combined and mapped into an appropriate geographic projection. For
this study, we employed the Albers Equal Area Projection using the same
projection attributes as the EDC-generated AVHRR data sets. This allowed
us to compare readily the mapped fire boundaries provided by the Alaska
Fire Service to the boundaries mapped from the AVHRR data set. Figure
4 presents a map of the points of origin for all fires within the study
area, as well as the boundaries of those fires whose size was greater than
10,000 ha.
Next, any detections which were located above the beeline (> 750 m elevation in interior Alaska) along with any pixels within 1 km of major river systems were eliminated for consideration as potential fires. The latter pixels were eliminated because although the geometric registration appeared to be quite good, some misregistration should be expected, and therefore false alarms may occur due to the low NDVI of a river signature.
The initial threshold (t1) for determining areas of significant change between the two images was an NDVI difference of 23. This threshold level was selected because it resulted in a number of detected fires which was within the same order of magnitude as the number of fires which have occurred in this region in the past. Based upon past history, the upper boundary of fires >400 ha in size expected for this region is 100. A threshold of t1 = 23 resulted in 111 possible fires, while a threshold of t1 = 24 resulted in > 250 possible fires.
The boundaries of all remaining detected polygons with areas were converted
to a vector format and imported into the GIS data base. Figure
5 presents a false-color AVHRR image of the test site, with the differences
detected in the above process displayed in yellow. Comparison of these
areas to the fire locations in Figure 4
shows a clear correlation.
The areal extent of the AVHRR-detected fires was then compared to those
represented by the digitized fire records. Using the initial threshold
of t1 = 23, it was found that the AVHRR-estimated areas
were considerably smaller than the fire record areas. To try to improve
this detection accuracy, a second detection image was generated with a
second threshold (t2= 18) 20% lower than t1.
The boundaries of the fires that were detected from the second threshold
were then converted to a vector format and imported into the GIS data base,
and their areas compared to the areas contained in the fire records.
Table 1 summarizes the detection statistics for six different fire size classes (note: 100 ha= 1 AVHRR composite image pixel): 1) 400-1000 ha; 2) 1000-2000 ha; 3) 2000-5000 ha; 4) 500010,000 ha; 5) 10,000-40,000 ha; and 6) >40,000 ha. From this table, it can be seen that the detection rate improves significantly for fire sizes > 1000 ha. For fires with a size greater than 1000 ha, the rate of detection was 73.2 % (with no false alarms). For fires with a size greater than 2000 ha, the rate of detection of 89.5% (with no false alarms). Review of the start and stop dates for the four missed fires showed that one (fire no. 126) had expired prior to the middle of June. After fires, there can be a significant amount of vegetation regrowth due to sprouting from roots. Since this fire stopped early in the growing season, such regrowth probably occurred in this region; hence no changes in the NDVI signature would be expected.
It is significant that high fire detection rates utilizing multi-date AVHRR images are only achieved when fires larger than 2000 ha are considered. From the data presented in Table 1, it can be seen that fires greater than 2000 ha in size represent some 96% of the total area burned in wildfires in the test area. This ratio is representative of forest fires in boreal regions during significant fire seasons; for example, the total acreage burned occurs in a few larger fires, rather than a larger number of smaller fires.
Table 2 summarizes the estimates of burned area from the two different NDVI thresholds for fires greater than 2000 ha, as well as the areas of the burns from the fire records. These data are summarized for the various fire size classes in Table 1. We can see that using the second threshold improved the percent of the areas mapped by the AVHRR method from 48.1 % to 58.1 % of the total acreage mapped and recorded by the Alaska Fire Service. In general, the areal extent of the larger fires (> 5000 ha) were more accurately mapped than the smaller fires. Using the second threshold (t2) improved the percent of the areas mapped by the AVHRR method from 48.1% to 58.1 % of the total acreage mapped and recorded by the AFS throughout the test area in the summer of 1990. When compared to the areas of those fires detected by the AVHRR, the second method mapped an area which was 60.6% of the acreage mapped by the AFS.
Finally, in Figure 6 we present a comparison of the boundaries from fire records to those generated from AVHRR data using the second threshold. The fires in these examples range in size from 10,000 (Fire 261) to >115,000 ha (Fires 17 and 42). Even though this technique produces boundaries smaller than those from the fire records, in most cases the overall shape of the fire is well represented on the AVHRR change detection image.
There are several possible explanations for the underestimation of fire area on AVHRR imagery. Fires in interior Alaska typically begin in mid-May and continue through the first snowfalls in September and October. The larger fires burn for weeks, and it is not uncommon for fires to burn for 3 months. The-average duration of the test area fires greater than 2000 ha in size was 63 days. A large percent of the fires (74%) in the test site larger than 2000 had a start and / or end date outside of the period when the AVHRR data were collected. Given this fact, it is expected that the areal estimates derived from the AVHRR data would be smaller than the actual boundaries. Improvements in estimating the areal extent of the burns can most likely be achieved by utilizing data collected earlier and later in the growing season, or by using the individual AVHRR channels.
The second reason is that the fire records contain maps of the boundary
of the outer edges of a fire. The areal estimates contained within the
fire records assume that all regions within the boundaries are burned.
In fact, there is considerable patchiness in the intensity of burning throughout
most of large wildfires, with some landscape units more susceptible to
burning than others. Thus, the ground-truth used in this study in many
cases most likely overestimates that actual area burned.
We have shown that two-date AVHRR data can be used to detect the locations of the large-burned areas in the boreal regions of Alaska. The high detection rates for fires greater than 2000 ha in size (which represent a very high portion of the total acreage burned) considered with the true impact of the false alarms (which represent an area < 0.001% of the total area burned in the test site) demonstrates the potential for using AVHRR to monitor boreal forest fires on a global basis. Although the total acreage estimated by the AVHRR data was considerably lower than the areas mapped by the Alaska Fire Service, we believe that these estimates can be improved by using AVHRR data sets collected earlier and later in the growing season. In addition, since the majority of acreage burned is contained within a few large fires, it would not be unreasonable to derive better areal estimates from higher resolution systems such as Landsat, Spot or the ERS-1 SAR [a recent study by Kasischke et al. (1992) shows that fire boundaries are easily mapped on ERS-1 SAR imagery].
In closing, we should note that there are several aspects of biomass burning in boreal regions which contrast sharply with biomass burning in other regions of the world, such as the rainforests of South America or the Savannah regions of Africa. First, in these other regions, the sources of biomass burning are largely anthropogenic. In South America, it is associated with the slash and burn techniques used in recently cleared tracts of rainforest. In Africa, biomass burning is associated with the annual burning of large areas used as pasture or being cultivated. In boreal regions, the larger fires are almost entirely started by lightning. Human-caused fires represent a high percentage of the number of fires, but they represent a very small percent of total area burned.
The second point of contrast is the size and number of fires. Although many fires occur in boreal forests, relatively few large fires result in the vast majority of acreage burned. In South America and Africa, the cumulative effect of biomass burning is the result of a large number of small fires.
These differences lead to alternate methods of using AVHRR for detection
of biomass burning. For the smaller fires in South America and Africa,
the areal extent of burning is indirectly inferred from the number and
duration of thermal signatures or the extent of smoke plumes resulting
from fires (Malingreau et al., 1990). In boreal regions, it is believed
that areal extent of burning can be directly mapped.
Crutzen, P. J., Heidt, L. E., Krasnec, J. P., Pollock, W. H., and Seiler, W. (1979), Biomass burning as a source of atmospheric gases CO, H2, N20, NO, CH3Cl, and COS, Nature 282:253-256.
Dyrness, C. T., Viereck, L. A., and Van Cleve, K. (1986), Fire in taiga communities of interior Alaska, in Forest Ecosystems in the Alaskan Taiga (K. Van Cleve, F. S. Chapin III, P. W. Flanagan, L. A. Viereck, and C. T. Dyrness, Eds.), Ecological Series Vol. 57, Springer-Verlag, New York, pp. 74-86.
Eidenshink, J. C. (1992), The 1990 conterminous U.S. AVHRR data set, Photogramm. Eng. Remote Sens., forthcoming.
Gabriel, H. W., and Tande, G. F. (1983), A Regional Approach to Fire History in Alaska, BLM-Alaska Technical Report 9, Bureau of Land Management, Anchorage, AK, 34 pp.
Kasischke, E. S., Bourgeau-Chavez, L. L., French, N. H. F., Harrell, P., and Christensen, N. L. Jr. (1992), Initial observations on the use of SAR imagery to monitor wildfires in boreal forests, Int. J. Remote Sens., forthcoming.
Loveland, T. R., Merchant, J. W., Ohlen, D. O., and Brown, J. F. (1992), Development of a Land-Cover Characteristics Data Base for the Conterminous U.S., Photogramm. Eng. Remote Sens. 57:1453-1463.
Malingreau, J. P., LaPorte, N., and Gregoire, J. M. (1990), Exceptional fire events in the tropics -- southern Guinee, January 1987, Int. J. Remote Sens. 11(12):2121-2123.
Mooney, H. A., Vitousek, P. M., and Matson, P. A. (1987), Exchange of materials between terrestrial ecosystems and the atmosphere, Science 238:926-932.
Olson, J. S., Pfuderer, H. A., and Chan, Y. H. (1978), Changes in the Global Carbon Cycle and the Biosphere, ORNL / ETS-109, Oak Ridge, TN, 169 pp.
Sadowski, F. G., and Watkins, A. H. (1991), Development of land data sets for studies of global climate change, Proceedings of the ASPRS Convention, Baltimore, MD.
Schlesinger, W. H. (1984), Soil organic matter: a source of atmospheric CO2, in The Role of Terrestrial Vegetation in the Global Carbon Cycle -- Measurement by Remote Sensing (G. M. Woodwell, Ed.), Wiley, Chichester, pp. 111127.
Tans, P. P., Fung, I. Y., and Takahashi, T. (1990), Observational constraints on the global atmospheric CO2 budget, Science 247:1431-1438.
Wein, R. W., and MacLean, D. A. (1983), The Role of Fire in Northern Circumpolar Ecosystems (Scope 18), Wiley, Chichester.