GLOBAL BIOCHEMICAL CYCLES, VOL. 16, NO. 1, 10.1029/2000GB001357, 2002
[7] GAC data are constructed by taking data from every third AVHRR scan, averaging the data from every four along-scan LAC pixels, and skipping the fifth [Kidwell, 1995; Belward et al., 1994]. Clearly, the along-track, scan line skipping procedure means that any fire signals present in the discarded scan lines will be totally absent from the resultant GAC data set. However, since the spatial distribution of fires should not be confined to a regular pattern of AVHRR scan lines, the GAC data can be treated as a simple one-third sample in this regard, and the equivalent LAC along-track statistical distribution of fire pixels can easily be reconstructed [Belward et al., 1994]. What is more critical is how the along-scan averaging procedure affects the ability of GAC data to discriminate fire-affected pixels. To investigate this, a simple model of the AVHRR measurement process was used to study the effect of the GAC along-scan spatial subsampling. The model simulated AVHRR LAC spectral radiances (L3.7 and L11) given an observed surface with a particular temperature and the presence or absence of a subpixel fire hot spot, again with particular size and temperature characteristics. For simplicity, no atmospheric or emissivity effects were included in the simulation. Numerical equations used were based on the two-component, area-weighted Planck functions originally described by Matson and Dozier [1981] ((1) and (2)). As outlined by Robinson [1991], these equations have previously been inverted with real AVHRR data and used to estimate the size and temperature of subpixel fires.
where R3.7 is the modeled spectral radiance in the AVHRR 3.7 µm channel (W m-2 sr-1µm-1), R11 is the modeled spectral radiance in the AVHRR 11 µm channel (W m-2 sr-1 µm-1), Sh is the planimetric area of the fire (m2), A is the AVHRR pixel area (m2), L3.7(T) is the spectral radiance emitted in the AVHRR 3.7 µm channel by a body at T Kelvin (W m-2 sr-1µm-1), L11(T) is the spectral radiance emitted in the AVHRR 11 µm channel by a body at T Kelvin (W m-2sr-1 µm-1), Th is the fire temperature (K), and Tc is the ambient background surface temperature (K).
[8] Spectral radiances from four simulated “neighboring” LAC pixels were averaged to obtain a single GAC pixel measurement. The simulation was carried out repeatedly, with the number of LAC pixels containing active fires varied between one and four to determine the effect on the resultant GAC pixel measurement. The LAC and GAC data sets were then converted from spectral radiance into brightness temperature units (T3.7 and T11) using the inverse Planck function procedure outlined by Wooster et al. [1995]. The difference between the corresponding 3.7 and 11 µm brightness temperatures (T3.7-11) was then calculated since this is the most important discriminatory parameter used in AVHRR fire detection techniques. As outlined previously, the differential response of the AVHRR 3.7 and 11 µm channels to subpixel hot spots causes the value of T3.7-11 to rise sharply in the presence of a fire [Matson and Dozier, 1981; Robinson, 1991]. At nonfire “background” pixel locations, (1) and (2) indicate T3.7-11 to be zero; however, in real AVHRR data the spatially and spectrally varying atmospheric transmissivity and surface emissivity typically mean background T3.7-11 values of 0.5–5 K. Thus, for a pixel containing an active fire the value of the T3.7-11 parameter over and above the natural background variation is directly related to the capability for discriminating that pixel as “fire affected.” The larger the T3.7-11 increase, the more unambiguous the hot spot identification.
|
|
[9] Results from the modeling are shown in Figure 5. As would be expected, when all four contributing LAC pixels include an active fire, the resultant GAC pixel signal is identical to the individual LAC pixel measures. However, while lowering the number of LAC pixels containing active fires does reduce the corresponding GAC T3.7-11 measure, because of the nonlinear and logarithmic relationships involved in (1) and (2) the variables are not reduced in direct proportion. For example, when the LAC fire pixel T3.7-11 measure is 40 K the equivalent GAC measure would still be 17 K even if only one fire pixel were included in the four contributing LAC pixels. Increasing the number of LAC fire pixels included in the GAC pixel measure increases the GAC T3.7-11 measure further (Figure 5). The relationships are relatively insensitive to either the background or fire temperature, providing evidence that fire analysis based on AVHRR GAC data could be a relatively reliable tool in many circumstances. However, inspection of the model results also confirmed that because of the four-pixel averaging used to derive GAC data, for any particular fire temperature and T3.7-11 detection threshold, the smallest detectable fire in a GAC pixel is still four times larger than the smallest detectable fire in a LAC pixel.
[10] The model indicates that AVHRR GAC data are clearly theoretically capable of identifying active fires, perhaps more so than initially might be thought given the extreme subsampling involved. However, it was noted that the simple modeling performed here did not incorporate simulation of the AVHRR point spread function or spatial autocorrelation, which are important when analyzing real data [Breaker, 1990; Wooster et al., 1998b]. Thus, having proved that GAC data are theoretically capable of effectively detecting active fires under a number of scenarios, the capacity of the data for this purpose was further tested using a series of real data intercomparisons.
[11] This analysis of AVHRR data concentrated on nighttime scenes only since daytime fire detection presents more complex issues due to the additional component of 3.7 µm solar radiation being reflected from clouds, water bodies, and other image features [Robinson, 1991]. Many previous AVHRR LAC fire studies have concentrated solely on nighttime data [e.g., Lee and Tag, 1990; Langaas, 1992, 1993; Legg and Laumonier, 1999], but daytime LAC detection methods are now also well established. Justice and Dowty [1994] and Giglio et al. [1999] review many of these procedures. For the current study it was felt that use of daytime data would add a further and unnecessary complication since the primary purpose was concerned with evaluating the effect of the GAC subsampling procedures on fire detection capability. Daytime GAC fire detection would be prone to increased error from the effect of solar-reflected radiation increasing the T3.7-11 signal at certain nonfire pixels. Thus the current study was implemented using the simplest nighttime case. Use of daytime data may, of course, be important in discriminating the shorter-lived agricultural fires that may only be active during the day and which may even dominate burning in non-El Niño years [Eva and Lambin, 1998]. However, the 1997 Borneo fires are reported to have been generally large, long-lived events, and following the reasoning of Legg and Laumonier [1999], the use of nighttime data was thus considered appropriate.
[12] Using the Satellite Active Archive, three dates in October 1997 were identified where fires existed on Borneo under relatively cloud-free conditions and where both GAC and LAC versions of the same nighttime data existed. Cloud-contaminated pixels were first screened out using a simple 11 µm temperature threshold [Robinson, 1991]. Following most previous AVHRR fire studies, the fire pixel detection tests were based on thresholding of the T3.7-11 data values, and two specific methods were tested. The first approach was based on use of a single, image-wide T3.7-11 threshold, following, for example, Kaufman et al. [1990], Robinson [1991], Kennedy et al. [1994], and Buongiorno et al. [1997]. The second “contextual” approach involved “potential” fire pixels being identified using a single, image-wide T3.7-11 threshold, but then these pixels were confirmed or rejected as “true” fire pixels based on comparison of their actual T3.7-11 value with those of the immediately surrounding background pixels [Justice and Dowty, 1994]. The specific contextual algorithm implemented was that of Flasse and Ceccato [1996], adapted to the particular conditions encountered on Borneo following Nakayama et al. [1999]. Though apparently successful when applied to the LAC data sets, the contextual method provided poor results when tested on the corresponding GAC imagery, failing to detect many fire-affected pixels that could clearly be discriminated by visual inspection. Investigation showed that this failure was due to the contextual aspect of the procedure where, as outlined by Flasse and Ceccato [1996], the T3.7-11 threshold is altered as a function of the standard deviation of the T3.7-11 background pixel values. The background pixel T3.7-11 standard deviation is noted to be much higher for a set number of GAC pixels than for the same number of LAC pixels owing to the averaging of four LAC pixels into every GAC measurement. This makes the GAC T3.7-11 threshold returned by the contextual filter anomalously high, and this precludes many obviously fire-affected GAC pixels being confirmed as such by the second stage of the contextual technique. Consequently, the contextual algorithm performs poorly in this case.
|
|
[13] The spatially varying thresholds of the contextual approach are more obviously required during analysis of daytime imagery, where solar heating and solar-reflected radiation typically lead to a higher spatial variability in the 3.7 µm brightness temperature background. The apparent deficiency of the contextual approach for GAC analysis and the use of solely nighttime data led our study to adopt the simple, image-wide thresholding approach for all further processing. Each LAC and GAC scene was thus subject to a fire detection test based on thresholding of the T3.7-11 parameter. Analysis of the T3.7-11 histograms confirmed the relative similarity of this parameter in the LAC and GAC data sets (Figure 6) and indicated that the appropriate threshold for identifying fire-affected pixels was somewhere above 5 K. Detection of hot spot (fire) pixels within each image was attempted five times, varying the T3.7-11 threshold between 6 and 10 K in order to optimize the method. On each run, AVHRR pixels having T3.7-11 values above the chosen threshold were classified as containing active fires (i.e., fire affected). Fire count statistics were then extracted, and following Belward et al. [1994], the inverse of the GAC subsampling procedure was applied to the GAC fire pixel counts so that results from contemporaneous GAC and LAC scenes could be meaningfully compared. This simply involved correcting the GAC fire counts for the 27% sampling factor used when deriving GAC data and multiplying the resultant fire counts by 4 to take into account the fact that each GAC pixel is originally derived from four LAC measurements [Belward et al., 1994].
|
|
[14] Results from these procedures are shown in Figure 7 and indicate a strong correspondence between the number of fire pixels identified in the corresponding LAC and GAC imagery. The relationship is reasonably consistent over all image dates and detection thresholds tested. Analysis also confirmed that the spatial distribution of the LAC fire pixels was effectively reproduced by the GAC imagery, though obviously with a loss of spatial detail. As expected, when using the highest T3.7-11 threshold (10 K), the number of identified fire pixels is at a minimum. At this threshold, there is also only a 3% variation in the number of LAC fire pixels detected between the individual image dates, and importantly, the difference between the number of fire pixels detected in corresponding LAC and GAC images consistently lies within a small -3 to +3% range. However, visual inspection of the imagery confirms that use of the 10 K threshold precludes detection of many fire pixels. When using a lower T3.7-11 threshold of 6 K, similar inspection confirms a much stronger agreement between visual and automated fire pixel identification. At this threshold, fire pixel counts are at a maximum, and there is a 20% variation between counts on different image dates. Crucially, however, there is still only a relatively small difference between the fire pixel counts extracted from temporally coincident LAC and GAC data, ranging from 0.15 to 13%. GAC data therefore appear to provide a reasonably reliable record of fire pixel number under these conditions. A further decrease in the T3.7-11 threshold to 5 K results in the incorrect classification of many obviously nonfire pixels, particularly on the 21 October 1997 image. As previously stated, in the high humidity and smog-affected atmosphere of Borneo a top-of-atmosphere T3.7-11 value of this magnitude can be generated by atmospheric effects alone. This accounts for the errors of commission when using a 5 K threshold, and a T3.7-11 threshold of 6 K was therefore chosen for the remainder of the study.
|
|
[15] It is worth considering why the agreement between the LAC and GAC fire statistics of Borneo appear so strong. Specifically, why are the data in much better agreement than are corresponding data sets for West Africa, where Belward et al. [1994] found that GAC data simulated from daytime LAC scenes severely underestimated fire pixel number? To help answer this, we can consider conditions in two different ecosystems; ecosystem 1, which favors small fire sizes, and ecosystem 2, where conditions favor the generation of much larger fires (Figure 8). Ecosystem 1 might correspond broadly to the West African ecosystems analyzed by Belward et al. [1994] during relatively normal climatic conditions, while ecosystem 2 is likely to more closely resemble the degraded Indonesian forest under the El Niño-related drought conditions encountered during the current study. Assuming a constant fire and background temperature, application of any particular T3.7-11 threshold to AVHRR data of either ecosystem will directly translate to the same minimum detectable fire size. As outlined in section 3.1, this minimum detectable fire size will be 4 times larger for GAC than for LAC data. For example, results from the model used in section 3.1 indicate that use of a 10 K T3.7-11 threshold results in a minimum detectable fire size of 50 m2 for LAC data and 200 m2 for GAC data in either ecosystem. This assumes an 800°C fire, a 30°C background, and the absence of atmospheric and surface emissivity effects, so these minimum detectable fire sizes would be somewhat increased in real data because of the presence of these effects. Comparing theoretical fire detection rates in ecosystem 1 and 2, Figure 8 indicates that in both ecosystems the majority of the fires are >50 m2 and are thus detectable using AVHRR LAC data with the 10 K T3.7-11 threshold. However, in ecosystem 1 the majority of fires are <200 m2 and so are missed when using AVHRR GAC data with this detection threshold. This is not the case for ecosystem 2, which favors the development of fires larger than the 200 m2 minimum detectable fire size of GAC data, and thus the GAC fire detection rates are much closer to those obtained with LAC data in this ecosystem. The fact that the drought-stricken forest of Borneo favored the development of very large fires, whereas the West African ecosystems studied by Belward et al. [1994] generally favored much smaller fires, is one explanation for the much stronger agreement between LAC and GAC fire counts found in the current study.
[16] A further reason for the comparatively poor fire detection performance reported by Belward et al. [1994] for GAC data of West Africa is the use in that study of a hot spot identification method incorporating a requirement that the 3.7 µm brightness temperature itself must exceed 320 K. This test was introduced largely so that any GAC pixels identified as fire affected would be directly comparable across the sahelian grassland, bush savanna, and Guineo-Congoilan forest zones present in that study region [Belward et al., 1994, Figure 3]. However, use of a 320 K 3.7 µm threshold effectively means that all four in-line LAC pixels that go into constructing any particular GAC pixel must themselves all be fire pixels if the resulting GAC pixel is also to be classed as a fire. Any GAC pixel containing fewer than four individual LAC fire pixels will thus not be classified as a hot spot using this methodology. In West Africa the occurrence of four in-line LAC fire pixels is much more likely in the savannah environment, where fires extend over an often relatively long and linear fire front. The West African forest fires noted by Belward et al. [1994] tended to be characterized more by individual LAC hot spot pixels. Thus Belward et al. [1994] found agreement between LAC and GAC fire counts to be worst in the forested regions and best in the savanna. For Borneo, however, the climate and ecosystem type is far less varied than that of West Africa, and no 3.7 µm threshold test was required for effective hot spot detection. Detection of fire pixels was based solely on thresholding of the T3.7-11 data, and Figure 5 indicates that with this technique many GAC pixels comprised of less than four fire affected LAC pixels will still themselves be identified as hot spots when using an appropriately low threshold. This is a further factor that explains the increased proportion GAC fire pixels correctly identified as hot spots in the current study when compared to that of Belward et al. [1994].
[17] The disadvantage of the approach used here is that without use of the 320 K 3.7 µm threshold test of Belward et al. [1994], we do not know exactly how many of the original LAC pixels comprising the resultant GAC measure actually contained a fire. By applying the inverse GAC data averaging procedure of Belward et al. [1994] to compare the LAC and GAC fire statistics we are inherently assuming that for every GAC hot spot pixel detected, there were, on average, four LAC hot spot pixels. Detailed analysis of an AVHRR LAC scene (12 October 1997) confirms that the majority grouping of LAC fire pixels is, in fact, an along-scan group of four, thus confirming the preponderance of larger fires under this ecoclimatic condition. However, owing to the hot spot sensitivity of the 6 K T3.7-11 threshold many of the GAC fire pixels clearly cannot have had all four contributing LAC fire pixels as fire affected. Nevertheless, the strong agreement between the 1997 LAC and GAC fire counts (Figure 7) indicates that in these conditions the overestimate introduced by the applying the times 4 factor to the GAC fire counts is broadly balanced by the underestimate introduced by the failure to identify GAC fire pixels having a fire size smaller than the minimum detectable threshold (Figure 8). When analyzing longer multiyear time series, it would be necessary to consider whether the fire size distribution is constant or whether it changes from year to year. It seems likely that fire size distribution will be very different between El Niño and non-El Niño years, but between the various El Niño years covered by the GAC archive a more constant distribution might be expected. Multiyear LAC-to-GAC comparisons or analysis of high spatial resolution imagery would be required to confirm this situation.
[18] Finally, it should be noted that despite the early findings of quantitatively poor GAC performance in West African environments by Belward et al. [1994], subsequent studies have gone onto extract fire count statistics from African GAC data and have used these to parameterize atmospheric chemistry models [e.g., Cooke et al., 1996]. Thus, even in these highly variable ecosystems, GAC-derived statistics have been shown to have significant value.
Citation: Study of the 1997 Borneo fires: Quantitative analysis using global area coverage (GAC) satellite data, Global Biogeochem. Cycles, 16(1), 10.1029/2000GB001357, 2002.
