This file was created by scanning the printed publication. Errors identified by the software have been corrected; however, some errors may remain. The Use of AVHRR Satellite Imagery to Monitor Boreal Ecosystem Forest Fires Donald R o Cahoon, Jr.' and Brian Jo Stocks2 Abstract - Forest fires are an integral part of the natural forces that shape the composition and evolution of the boreal forest. These fires, often very intense, commonly grow to sizes large enough for satellite monitoring using the National Oceanic and Atmospheric Administration (NOAA) Advanced Very High Resolution Radiometer (AVHRR) instrument. The AVHRR instrument acquires l-km and 4-km resolution (at nadir) satellite imagery in five spectral bands. Multispectral methods have been developed, using the AVHRR imagery, to estimate the size and growth rate of bumed forest. The AVHRR bumed area estimates have been evaluated and the error determined for both the l-km and 4-km imagery. Applying these methods, the spatial distribution and extent of forest fires has been mapped for the 1987 and 1992 Russian boreal forest fire season. Due to a strong contrast in the total area bumed in each of the 2 years, the large interannual variability of bumed forest is demonstrated. INTRODUCTION The boreal ecosystem is a forested belt, mostly comprised of conifer trees, that stretches across the Northern Hemisphere's circumpolar countries. The boreal forest covers over 12 million km2 of the Earth's surface which is about 10% of the total land surface area. In terms of areal extent, the boreal forest ranks about fifth compared to all terrestrial biomes, but ranks second only to tropical forests in total plant mass (Whittaker and Likens 1975). Given the ecosystem's size and total available plant mass it is not surprising that the World's boreal forests contain almost 40% of all terrestrial carbon (Kasischke et al. 1995a). This means that the boreal forest must play an extremely important role in the global carbon budget and it is important to characterize any disturbance and process that perturbs the carbon cycle within this ecosystem. Fire is one such disturbance. Because of the enormous geographical scale for which the boreal forest extends, a trade-off between satellite resolution, accuracy of estimating the acreage burned, and frequency of satellite coverage has been made. In reality, there are three satellite systems that could potentially be used to monitor forest fires. Landsat and SPOT quickly come to mind because of their high spatial resolution, but do not offer the repeat coverage needed to compensate for cloud cover Scientist,NASAILangley Research Center, Harnpton, VA. Senior Research Scientist, Canadian Forest Service, Sault. Ste. Marie, ONT, Canada. 1 Research obscuration of the land surface. Also, the amount of data and cost of enough imagery to cover the entire boreal region would be unrealistic. Rather, we have chosen to use AVHRR imagery because it offers a reasonable resolution at a low cost, and it provides multiple overpasses each day. After an evaluation of the AVHRR imagery, we have found that estimates of the surface area can be within a few percent. This error is considered reasonable for the large geographical scale we are evaluating as long as consistent results can be obtained. To develop consistent surface area estimation results, we have developed a methodology for processing the AVHRR imagery. The following sections will describe the AVHRR instrument and archived imagery, the burned area estimation methodology and error analysis, and describe some of our ongoing research on boreal forest fires. AVHRR INSTRUMENT The AVHRR instrument has flown on several National Oceanic and Atmospheric Administration (NOAA) sun-synchronous polar-orbiting meteorological satellites since 1978 and provides a nearly continuous archived data set. These satellites have flown in a morning and afternoon configuration which provides four overpasses each day in the low to mid-latitudes, and upwards toward 8 overpasses in the higher latitudes of the boreal forest. The AVHRR instrument is a scanning imager which scans in the across track direction and acquires data in five spectral bands (table 1). The swath width is just over 2800 km given a nominal orbital altitude of 830 km and a total scan angle of 110.8". Each cross track scan provides 2048 observations (pixels) with an instantaneous field-of-view (IFOV) of approximately 1.4 mrad. The precise IFOV for each channel varies slightly and is included in table 1. Table 1. - AVHRR channel characteristics. Channel Wavelength (microns) Cross track IFOV (mrad) Along track IFOV (mrad) Channel selection of the AVHRR instrument was driven by operational meteorological and oceanographic requirements. However, over time the fire community has found ways to apply the channel selection to its own research. In summary, channel 1 and channel 2 can be used to differentiate smoke from clouds (Chung and Le 1984). Channel 2, because of its sensitivity to the reflectance fi-om vegetated surfaces, provides tremendous contrast between bum and unburned areas (Cahoon et al. 1992; Cahoon et al. 1994; Kasischke et al. 1995a). Channel 3 has long been demonstrated to be sensitive to hot temperature sources (Matson and Dozier 1981; Scorer 1987) and is useful for mapping active forest fires (Cahoon et al. 1994; Stocks et al. 1996). Channels 4 and 5 are both similar in that they are low in contrast when calibrated, but provide a basis for distinguishing between burned areas and water in mid-afternoon imagery (Cahoon et al. 1994). The imagery is archived in either of two formats, local area coverage (LAC) or global area coverage (GAC). The LAC format is the full 1-krn (at nadir) resolution product and is recorded onboard the spacecraft and downlinked to specific sites. The AVHRR 1-km data are also continuously broadcast to any ground-station within the satellite's horizon; which is referred to as high resolution picture transmission format (HRPT) data. In the archives HRPT and LAC data are the same. Due to a limited amount of onboard memory for recording LAC data, a 4/15 subsampled product is continuously recorded for broadcast to selected ground stations. This format is the global area coverage (GAC) and is frequently referred the 4 km imagery. The GAC subsampling scheme is shown in figure 1. I Averaged and Saved Discarded Figure 1. - GAC subsampling scheme. Each row is a portion of a scan line. Each box is one image pixel. The dark box represents a GAC "4 km" pixel. BURNED AREA ESTIMATION METHODOLOGY The surface area of bum scars is estimated using either LAC or GAC archived imagery. In a previous study, Cahoon et al. (1992) demonstrated that, given the large fire sizes in the boreal system that GAC imagery is sufficient for estimating the area. In the following section we will examine the surface area estimation error using both the LAC and GAC imagery. This section will lay out the methodology through which we process the AVHRR imagery. Because of the intermittent availability of clear-sky imagery, which negates the possibility of observing the entire study region at one time, a clear-sky mosaic is created. From this mosaic, all bum scars can be observed in a single image, making the area-burned estimate simpler and alleviating the concerns of double counting or missing any burned area. The development of the clear-sky mosaic is a several-step process which will be described with the use of matrix notation. An individual multispectral(3-banded) AVHRR scene is represented by matrix Dn and the mosaic of all the individual AVHRR scenes is matrix M, where n is the total number of AVHRR scenes; i and j define the number of elements, d and m, in the matrices. Each pixel in the AVHRR image is corrected for brightness changes due to the solar zenith angle. Using the NOAA-supplied navigation data, each element dii is the result of resampling the original NOAA scene into a cylindrical-equidistant map projection. Since each AVHRR scene is resampled, for every Dn the elements dii represent the same geographical region and are a function of latitude (q) and longitude (A). The mosaic image (M) is the same size and spatially represents the same geographical area as the AVHRR scenes Dn. Then each element of both M and Dn can be compared, with the results of the comparison composited into a final scene (M'). The comparison, M'ij = MIN (Dnii. M i j ) (2) tests for the minimum between the mosaicked image (M) and the AVHRR scene (Dn) in the AVHRR channel 2 band, and saves the elements for all channels of the minimum-valued image in the new mosaic M'. The testing is iterative, where M' is substituted for M, and continues n times through all of the scenes. Since clouds are highly reflective and the burned area is of low reflectance in the AVHRR channel 2, the burned areas are retained while the clouds are removed. In estimating the total regional burned area, it was first necessary to classify which image pixels contained bumed area. An unsupervised minimum distance classification on AVHRR channels 1, 2, and 4 was than used iteratively to isolate the scar pixels and statistically determine the edge pixels that are counted as having a burn scar. The iterative process continues until bum scars, that have been correlated with known fire activity determined from the AVHRR imagery, are isolated and falsely classified pixels have been minimized. The resulting scar map is subjected to another processing step to determine the area of each pixel. Since a cylindrical-equidistant projection is used and each side of a pixel is parallel to both meridians and parallels, the distance along each pixel in terms of degrees of latitude and longitude is easy to determine as well as the surface area. The area of every pixel classified as a scar is then integrated to derive the total area burned. BURNED AREA ERROR ANALYSIS An extensive analysis has been conducted to evaluate the error of estimating the burned surface area using both the LAC and GAC imagery. The details are outlined in Cahoon et al. (1992). In brief, surface targets of known area are delineated in both the LAC and GAC imagery. These targets are a series of lakes in Canada that often contain several islands, sometimes of considerable size, and thus are not homogeneous. The heterogeneous nature of the lake targets is similar to that of bum scars which contain unburned areas within a its perimeter. The results from this analysis are in figure 2. The mean surface area estimation error associated with using LAC imagery is lower than that of GAC imagery across the entire size range, even though for larger targets the difference is practically negligible. This comes as no surprise since we expect the analysis of higher resolution imagery to yield better results when working with small areas. However, it is interesting to note that for larger targets, the error variance for the GAC-derived estimates is less than that of the LAC analysis. This likely has a lot to do with the overlap of the AVHRR pixels (Cahoon et al. 1992). The overlap is considerable for the 1-km data and the edge of a target is defined by a gradual gradient, leaving room for interpretation as to where to define the target's edge. Much of the surface area estimate errors are due to blurred edges of targets. At least for the GAC imagery, some of this overlap has been eliminated by subsampling, leaving a better defined edge and fewer perimeter pixels to skew the analysis one way or another. Overall, the results show that for LAC imagery a mean error of about 2% can be expected, and that 90% of all targets will fall within 4% error for targets down to 100 km2. For the GAC imagery, the mean error is about 2% worse than that of the LAC imagery for targets that are 100 km2, and drops off to being negligible when evaluating targets that are 10,000krn2 or more in size. Burned area estimate errors, from both the 1988 Yellowstone Fires (Cahoon et al. 1992) and the 1987 Great Dragon Fire in China (Cahoon et al. 1994), are consistent with these results. RESEARCH APPLICATION: RUSSIAN BOREAL FOREST FIRES Forest fires have been a major disturbance regime affecting the world's boreal forests for millennia. Increased development within the world's boreal zone has necessitated fire management programs designed to protect human interest and forest investment. While these programs have been largely effective in reducing unwanted fires in selected regions, forest fires continue to exert enormous influence in the world's boreal forests. With the rising concern that potential global warming will lead to an increase of fire activity within the boreal forests (Flannigan and Van Wagner 1991, Wotton and Flannigan 1993, Stocks 1993, Fosberg et al. 1996),much attention as recently been focused on the important role the boreal forest plays in the Earth's carbon budget (Kasischke et al. 1995b), and how an increase in fire activity will impact that budget and potentially provide . I - a m . . I 10j . . . ... Area (km2) ..I 104 . . a . m . Figure 2. AVHRR-derived surface area estimation error for various target sizes. LAC errors shown by dashed line; GAC errors shown by solid line. The lower line, for both LAC and GAC, represents mean error and the upper line shows the upper error bound for 90% of all targets. positive feedback to global warming (Kurz et al. 1994). Given these concerns and the need for a quantitative assessment and a historical perspective of boreal fire activity, we have begun to use the satellite techniques described in this paper to assess the total area burned in the boreal forest during the last 15 years for which AVHRR data are archived. Leading into this multiyear study, as test cases, we have completed the burned area analysis for the Russian boreal forest for the years 1987 (Cahoon et al. 1994) and 1992 (Cahoon et al. 1996). The 2 years were totally dissimilar in terms of area burned, with 1987 being a severe fire year for both Russia and the northern Chinese provinces. It was during 1987 that one of the largest recorded fires in China's history occurred (about 1.3 million ha) and large fires throughout eastern Russia were very prevalent (Cahoon et al. 1994). The 1987 analysis was conducted using over 200 GAC images. As demonstrated in the previous section, the burned surface area estimation using GAC imagery is not much greater than 6% for fire sizes down to 100 krn2 in size. Fortunately, in the boreal forest, fire sizes are very large with 95% of the area burned by 5% of the fires. The estimate of the total area burned during the 1987 fire season is 14.446 million ha. In contrast, even though there was widespread fire activity in 1992, the total area burned was 1.5 million hectares. This is a relatively low annual total of burned acreage which was concurrently experienced in Alaska and Canada. The 1992 analysis used over 100 LAC images that cover the region from 30' to 180' east longitude. The boreal region was split into 30' longitude blocks. Each block was processed one at a time (Cahoon et al. 1996). There was an obvious similarity between 1987 and 1992, widespread fire activity is primary east of the Yenisey River (about 90°N latitude). The reason for the higher amount of burned area to the east is likely due to the increase in population toward the west and the low marshy areas west of the Yenisey River. With the population increase to the west comes additional fire management pressures and better access to fires for fire suppression purposes. CONCLUSIONS With limited resources and the constant need to assess spatial changes within our natural resource regimes, satellite and aircraft remote sensing is being drawn upon to provide spatially dependant data over various scales. The need to monitor and assess the post-burned size of forest fires is one such role that satellite-based observation can play. Even though it is more difficult to used space-based resources to contribute directly to fue fighting, it is not out of the question to utilize this resource to monitor fire activity and map post-burned fires over large geographical scales and remote regions. The remoteness of the boreal ecosystem, a region where considerable travel and fire reconnaissance is accomplished by air, almost dictates the need in a costcutting era to assess fire activity using satellite imagery. The remoteness, combined with the current interests in developing a consistent fire record throughout the entire boreal forest, has lead us to develop the AMIRR techniques used in our two trial studies, which will be implemented as we embark on a 15year analysis of the interannual variations of burned forest. Further evaluations will be made regarding the successfulness of detecting and estimating the size of smaller fires. Further, a test case has also been developed which demonstrates the potential use of the AVHRR imagery for monitoring fire growth rates (Stocks et a1 1996). To conduct this research, customized computer code has been developed over a number of years to implement our approach. Each segment of our software system has been carefully tested. We have begun to evaluate the potential of commercial packages for registering and mapping AVHRR images. In one prominent GIs/ image processing package, we have had very limited success with registration and would not trust the resulting surface area estimates. This does lead to our concern that any commercial tools that can map fires and estimate the size of burned areas using AVHRR satellite imagery should carefully be scrutinized for reliability. REFERENCES Cahoon, Jr., D. R., B. J. Stocks, J. S. Levine, W. R. Cofer 111, and C. C. Chung, 1992, Evaluation of a technique for satellite-derived estimation of biomass burning, J. Geophvs. Res, 97(D4): 3805-38 14. Cahoon, Jr., D. R., B. J. Stocks, J. S. Levine, W. R. Cofer III, and J. M. Pierson, 1994, Satellite analysis of the severe 1987 forest fires in northern China and southeastern Siberia, J. Geophvs. Res, 99@9): 18,627-18,638. Cahoon, Jr., D. R., B. J. Stocks, J. S. Levine, W. R. Cofer, and J. A. Barber, 1996, Monitoring the 1992 Forest Fires in the Boreal Ecosystem using NOAA AVHRR Satellite Imagery: Biomass Burning and Global Change, MIT Press, Cambridge, MA. (in press) Chung, Y. S. and H. V. Le, 1984, Detection of forest-fire smoke plumes by satellite imagery, Atmos. Environ., 18(10): 2143-215 1. Flannigan, M. D., and C. E. Van Wagner, 1991, Climate change and wildfire in Canada, Can. J. For. Res,, 21: 66-72. Fosberg, M. A,, B. J. Stocks, and T. J. Lynham, 1996, Risk analysis in strategic planning: fire and climate change in the boreal forest: Fire in Ecosystems of Boreal Eurasia, Blackwell Press, Oxford, UK. (in press) Kasischke, E. S., N. H. F. French, L. L. Bourgeau-Chavez, and N. L. Christensen, Jr., 1995, Estimating release of carbon from 1990 and 1991 forest fires in Alaska, J. Geophys. Res, 100: 2941-2951. Kasischke, E. S., N. L. Christensen, and B. J. Stocks, 1995, Fire, global warming, and the carbon balance of boreal forests, Ecol. Ap~l.,5(2): 437-45 1. Kurz, W. A., M. J. Apps, B. J. Stocks, and W. J. A. Volney, 1994, Global climate change: disturbance regimes and biospheric feedbacks of temperate and boreal forests: Biotic Feedbacks in the Global Climate System: Will the Warming Speed the Warming?: Oxford Univ. Press, Oxford, UK. Matson, M., and J. Dozier, 1981, Identification of subresolution high temperature sources using a thermal IR sensor, Phot. Eng. Remote Sensing, 47(9): 13111318. Scorer, R. S., 1987, Hot Spots and Plumes: Observation by Meteorological Satellite, Atmos. Environ., 2 1 (6): 1427-1435. Stocks, B. J., 1993, Global warming and forest fires in Canada, For. Chron., 69: 290-293. Stocks, B. J., D. R. Cahoon, Jr., W. R. Cofer 111, and J. S. Levine, 1996, Monitoring large scale fire behavior in northeastern Siberia using NOAAAVHRR satellite imagery, Biomass Burning and Global Change, MIT Press, Cambridge, MA. (in press) Whittaker, R. H., and G. E. Likens, 1975, Primary Production: The biosphere and man: Primary Productivity of the biosphere, Ecological Studies 14, Springer Verlag, New York. Wotton, B. M., and M. D. Flannigan, 1993, Length of fire season in a changing climate, For. Chron., 69: 187-192. BIOGRAPICAL SKETCH Donald R. Cahoon, Jr. is a research scientist with the National Aeronautics and Space Administration, Langley Research Center, Hampton, Virginia. He has worked as an analyst of satellite, aircraft, and meteorological data sets for over 11 years. During the last 7 years he has specialized in satellite remote sensing. His primary applications have been forest fire monitoring and surface radiation budget studies. Brian J. Stocks is a senior forest fire scientist with the Canadian Forest Service in Sault Ste. Marie, Ontario. Over the past 28 years he has specialized in the development of forest fire danger rating and fire behavior prediction systems, and is currently involved in the investigation of cause-and-effect relationships between global change and forest fires.