THE ENSO-FIRE DYNAMIC IN INSULAR SOUTHEAST ASIA DOUGLAS O. FULLER1 and KEVIN MURPHY2 1
Department of Geography and Regional Studies, University of Miami, Coral Gables, Florida 33124, U.S.A. E-mail:
[email protected] 2 Global Science & Technology Inc., 10210 Greenbelt Road, Lanham, Maryland 20706, U.S.A.
Abstract. We examined the spatiotemporal patterns of fire in insular Southeast Asia from July 1996 to December 2001 using a set of consistent, nighttime fire observations provided by the Along Track Scanning Radiometer (ATSR) sensor. Monthly ATSR fire counts were analyzed relative to georeferenced climatic and land-cover data from a variety of sources. We found that fires were strongly correlated with Southern Oscillation Index (SOI) (r = −0.75) and Ni˜no 3.4 index (r = 0.72) in forested land-cover types within the equatorial belt (5.5◦ S–5.5◦ N). Cross-correlation analysis revealed that detrended SOI was modestly correlated (r = 0.42) with detrended monthly fire count with a positive lag of four months. However, our analysis also revealed that fire counts reached their maximum 6 months before the absolute maximum of SOI. Annual sums of SOI ( SOI ) and fire counts revealed linearity for SOI ≤ 0. Overall, the results suggest that ENSO indices may have limited predictive utility at a monthly time scale, but that temporal aggregation and additional fire observations may enhance our capacity to forecast fires in different cover types based on ENSO data.
1. Introduction El Ni˜no has been defined as the irregular development of an anomalously warm pool of surface water in the eastern tropical Pacific, which generally lasts between 12 and 18 months (Glantz, 2001). As El Ni˜no warm phases decay, sea surface temperatures (SST) decrease in the tropical Pacific, and a cold phase relative to longterm conditions, or La Ni˜na, may develop. The cycling of warm and cold SST phases in the tropical Pacific influences global rainfall distribution and constitutes part of a climatic oscillation referred to the El Ni˜no-Southern Oscillation or ENSO (Glantz and Nicholls, 1997). Paleoclimatological evidence shows that at least 435 moderate to strong El Ni˜no events have occurred in the Holocene (Riedinger et al., 2002) and considerable speculation exists concerning how warming of the troposphere may increase the frequency and intensity of future El Ni˜no events (Yu and Boer, 2002; Herbert and Dixon, 2003). El Ni˜no–related impacts include increased likelihood of drought in southern Africa and Southeast Asia, reduced hurricane number and intensity in the tropical Atlantic, and increased storm frequency and intensity along the Pacific coast from Peru to southern California. Each of these climatic events produces a set of related impacts such as elevated risk of food shortages in regions dependent upon rainfed agriculture, destructive wildfires in normally humid locations, and coastal-zone Climatic Change (2006) 74: 435–455 DOI: 10.1007/s10584-006-0432-5
c Springer 2006
436
DOUGLAS O. FULLER AND KEVIN MURPHY
erosion where storms are rare or absent during non–El Ni˜no periods. Because El Ni˜no–related risk events carry such high costs to society, forecasters have increased their efforts to predict and disseminate information that may indicate an impending strong event, such as the 1997–1998 El Ni˜no, which is regarded by some researchers as the most intense of the 20th century (Glantz, 2001; Broad et al., 2002). Since this particular event, interest among climatologists, the media and the general public in El Ni˜no–related risks has grown markedly (Glantz, 2001). As data from El Ni˜no events are analyzed, they have helped advance development of dynamical climate models designed to predict the onset, duration and intensity of future ENSO events (Barnston et al., 1999). In addition, various climatic indices are used to characterize the development of warm and cold SST events (Trenbert and Stepaniak, 2001). Satellite and buoy observations, particularly related to the measurement of SST in the tropical Pacific, are used along with indices based on meteorological data, such as the Southern Oscillation Index (SOI), as predictors of El Ni˜no onset and intensity (Nicholls and Beard, 2001; Glantz, 2001). The SOI itself has been used extensively as a diagnostic of various El Ni˜no–related teleconnections (Anyamba and Eastman, 1996; Kitzberger et al., 2001; Lim and Kafatos, 2002) and is based on monthly mean sea surface pressure anomalies at Tahiti (T) and at Darwin, Australia (D), where SOI is generally presented as running mean of daily values given by T − D (Trenberth, 1997). Other commonly employed ENSO indices include Ni˜no Region 1 + 2, Region 3, Region 3.4 and Region 4 SST Indices, which reflect changes in SST in different regions of the Pacific, relative to a base period climatology from 1950 to 1979 (Trenberth, 1997). Although the peak of each El Ni˜no event tends to be in phase with the northern hemisphere winter, studies based on ENSO indices and the SST field observed from satellites suggest that each event is unique in terms of its onset, severity, persistence, and impacts (Glantz, 2001; Trenbert and Stepaniak, 2001). Despite inherent variability of ENSO events (and therefore inherent difficulties of prediction), major wildfires tend to show consistent relationships with different phases of the ENSO cycle in different parts of the globe. For example, Kitzberger et al. (2001) used dendrochronological data and climate records to identify fire patterns in the southwest United States and Patagonia following the transition from El Ni˜no to La Ni˜na. They reasoned that increased precipitation in these areas during El Ni˜no has tended to increase primary production, which has led to increased fuel loads. A subsequent cold phase event (La Ni˜na) generally has produced drought conditions in certain areas on the western side of the Pacific Basin, which in turn increased the risk of wildfires in these areas. Harrison and Meindl (2001) used analysis of variance to establish a statistical relationship between fire size and occurrence in Florida. Their study raised the possibility of using ENSO indices to predict the severity of the fire season in the growing wildland–urban interface of Florida. Successful predictive capacity, they reasoned, would help resource managers design and implement hazard mitigation
THE ENSO-FIRE DYNAMIC IN INSULAR SOUTHEAST ASIA
437
strategies that aim to reduce the costs of major wildfire episodes to human health, livelihoods and property. Unlike the temperate areas where ENSO may determine fire cycles, the lowland tropical areas of the Pacific and Amazon basin tend to experience a drought during the peak of each El Ni˜no event (e.g., Laurance et al., 2001; Kirono et al., 1999). On the islands of Borneo and Sumatra, for example, the very strong El Ni˜no events of 1982–1983 and 1997–1998 produced a large number of wildfires that burned in rainforest vegetation (Stolle and Tomich, 1999; Fuller and Fulk, 2000; Siegert et al., 2001), which normally experiences fire at intervals ranging from hundreds to thousands of years (Cochrane, 2003). These rainforest fires typically consume less fuel than crown fires found in mid-latitude conifer forests, but can severely damage many fire-sensitive trees that lack thick bark and other fire-related adaptations. This leads to development of a positive feedback between fire and deforestation, in which potential fuel load increases after fires, rendering rainforests more vulnerable to destructive fires during future El Ni˜no events (Cochrane, 2003). The study of active fires and burned area has benefited from a number of satellite remote sensing platforms maintained by various governmental bodies around the world. Although research leading to improved fire products is still needed (Justice et al., 2002), sensor systems such as the Advanced Very High Resolution Radiometer (AVHRR) have advanced understanding of fire dynamics from regional-to-global scales (Fuller, 2000). One of the challenges that remains is the development of internally consistent, multi-year time series of remotely sensed fire observations that permit comparison of different fire events and their possible relationships with recent ENSO events. If indeed droughts and wildfires do become more severe as a result of more frequent severe El Ni˜no events, then the need for improved predictive capacity will increase. This is particularly true for humid lowland areas in Indonesia and Amazˆonia, where, until recently, fire has been a relatively rare occurrence over the past 10,000 years (Cochrane, 2003). This last fact underscores the increasing role that humans are playing in setting fires in Indonesia and Amazˆonia to clear forest for pasture and agricultural production (Cochrane, 2003). The recent Indonesian fire events, in particular, were largely blamed on smallholders who used fire to open up forest areas for agricultural exploitation. For example, a recent study identifies swidden agriculturists as largely responsible for the related problems of plantation fires and deforestation during the 1997–1998 El Ni˜no event (Varma, 2003), although some authors believe agribusiness interests were in fact responsible for many ignitions (Barber and Schweithelm, 2000). Whereas smallholders do play a role in setting fires, fieldbased research has revealed that most smallholders were also the principal victims of fires, which burned both industrial plantations and subsistence farm plots (Barber and Schweithelm, 2000). Although ignition sources ultimately may prove difficult to pinpoint, it has become clear that fires present a potentially catastrophic driver of land-cover change in the lowland forests in large parts of the archipelagos of Southeast Asia (Stolle et al., 2003).
438
DOUGLAS O. FULLER AND KEVIN MURPHY
Figure 1. Map of the study region showing the location of the meteorological stations in Kalimantan where rainfall data were analyzed relative to ENSO indices.
In this study, we take advantage of a unique, multi-year fire data set derived from the Along Track Scanning Radiometer (ATSR), to analyze temporal and spatial patterns of fire. The ATSR is carried on board The European Space Agency’s ERS-2 spacecraft and is equipped with a thermal infrared band, from which active fire observations have been derived over the period July 1996–December 2001 (see http://www.atsr.rl.ac.uk/ for more information). We analyzed this fire time series for a window covering most of insular Southeast Asia centered over the island of Borneo (Figure 1). We examined these fire data in light of a set of ENSO indices in order to assess the degree of association between fire occurrence and ENSO over five and a half year period containing the very strong 1997–1998 El Ni˜no and subsequent La Ni˜na from 1999 to 2001. We also examined the distribution of fire points using a set of georeferenced data layers including tree cover, human population density, vegetation cover, and elevation (a correlate of humidity and accessibility in the study region) to factors that may be driving ignitions and fire extent during El Ni˜no and La Ni˜na years. Thus, the purpose of our research has been twofold: (1) To assess the temporal dynamics of fire in the study region as well as the environmental factors that may be associated with fire regimes. (2) To assess whether commonly used ENSO indices possess predictive utility for forecasting future fire events in the study region. 2. Study Area We examined fire and associated factors in the large archipelagic region extending from approximately 100◦ E–140◦ E longitude and 15.5◦ S–7.5◦ N latitude (Figure 1), centered over the island of Borneo. Many of the islands that fall within Figure 1
THE ENSO-FIRE DYNAMIC IN INSULAR SOUTHEAST ASIA
439
are volcanic in origin and have large changes in relief from sea level to the island interiors. Pronounced elevation gradients within the region are associated with a range of vegetation types. Extensive areas of lowland forest occur from sea level to 800 m. Above 800–900 m upland and montane tropical forests are found in most parts of the Indo-Malayan islands (Hill, 2002). In addition, the region contains large areas of wetland vegetation including freshwater swamp forest, mangrove and coastal swamp forest (MacKinnon, 1996). In some areas, changes in land use and land cover have produced a mosaic of secondary forest and grassland, dominated by the invasive grass species, Imperata cylindrica L. (Cogon grass). Rainfall is heavy during the austral summer (December–March), particularly in the equatorial belt where mean annual precipitation generally exceeds 2000 mm per year (Hill, 2002). Seasonal drought generally occurs throughout the study region during the months of July–September, and rainfall seasonality generally increases as one moves south and east from the islands of eastern Indonesia to northern Australia. Even in wet, equatorial locations, a brief dry period normally occurs during August–September each year (Kirono et al., 1999).
3. Data and Methods A variety of methods is available to detect active fires using remotely sensed data. The ATSR fire data that we used were derived from nighttime passes with a single, rule-based threshold applied to band 5 (3.55–3.93 μm), where an active fire is detected if the land surface within a pixel was greater than or equal to 308 K. ATSR band 5 spatial resolution is 1 km, and it is probable that a small fire within a pixel (less than 1/10 of a pixel’s area) is sufficient to produce a positive fire detection (Matson and Dozier, 1981), although we lacked actual fire data to test the sensitivity of the threshold. Single temperature thresholds such as this can produce reliable fire data for the study region (Fuller and Fulk, 2000); however, nighttime observations tend to underestimate fire as burning generally reaches a diurnal maximum in the tropics between 12:00 and 15:00 local time (Justice et al., 2002). Stolle (2003) evaluated the accuracy of ATSR fire data using maps of burned area in western Indonesia and found that the correspondence between ATSR fire detections and burned area was higher than that for fire detections derived from other remotely sensed data available for the region, such as imagery from AVHRR channel 3. However, because the overpass frequency restricts observations for this region to 1 in 4 days, ATSR data likely omit many active fires that burned between successive passes. Moreover, due to cloud cover, smoke and diurnal fire patterns, ATSR data will undersample active fires, which is a problem common to all observations derived from polar-orbiting platforms. Notwithstanding known problems of omission, commission errors for the region were low relative to other sources and the temporal coverage of the ATSR fire product makes this data source unique for understanding interannual fire variability relative to ENSO (Stolle, 2003).
440
DOUGLAS O. FULLER AND KEVIN MURPHY
TABLE I Data used to analyze fire distribution in insular Southeast Asia Data set
Source
Time period
ATSR active fires Percent tree cover Population density ENSO indices Vegetation type
http://www.atsr.rl.ac.uk http://modis.umiacs.umd.edu/vcf500LatLong.htm Landscan 2 data, http://sedac.ciesin.columbia.edu/data.html www.cgd.ucar.edu/cas/catalog/climind/index.html World Conservation Monitoring Centre
July 1996–2001 2001 2001 1995
We analyzed fire distribution in light of several other gridded data sets mapped at comparable spatial resolution (0.5–1 km) including population density from Landscan 2, elevation, land cover type, and percent tree cover. Table I gives the sources of each of these layers including relevant internet locations that describe the technical details of each. Accumulated fire counts from July 1996 to December 2001 were used to sample each of these layers to examine relations between areas where fires were recorded and those areas where fires were absent throughout the ATSR time series. 3.1.
TIME-SERIES ANALYSIS
Monthly fire counts from ATSR over the study period yield n = 66 observations, sufficient to conduct certain types of time-series analysis including computation of the autocorrelation (ACF) and cross-correlation functions (CCF). Thus, all autoand cross-correlation analysis described later refers temporal correlation as opposed to spatial autocorrelation. These data were log transformed to facilitate plotting, analysis, and interpretation. The log10 transformation also has the advantage of reducing non-linear trends in the data (Statsoft, 2003). Monthly fire counts were matched with monthly ENSO indices, including SOI, Ni˜no 1 + 2, Ni˜no 3, Ni˜no 3.4, and Ni˜no 4. These data were obtained from the Climate Dynamics Section of the National Center for Atmospheric Research (NCAR). Each of these data sets has had some smoothing applied to reduce noise and to facilitate their interpretation in ENSO forecasting (see http://www.cgd.ucar.edu/cas/catalog/climind/ for more information). In addition, we obtained a limited set of monthly rain gauge data over the study period from four stations in East Kalimantan, which are shown on Figure 1. These data were used to examine the relationships between rainfall and the ENSO indices in an area that was affected by catastrophic wildfires during the El Ni˜no episodes of 1982–1983 and 1997–1998 (Siegert et al., 2001). ACFs were examined for the ENSO indices and ATSR fire series to assess serial autocorrelation, trends and seasonality. Seasonal trends were evident using the ACF for each series and were removed using a ratios-to-moving average approach (also known as the Census I approach) implemented in Statistica version 6.1 software (Statsoft, 2003).
THE ENSO-FIRE DYNAMIC IN INSULAR SOUTHEAST ASIA
441
We then inspected the detrended time series for evidence of residual autocorrelation by plotting their ACFs. Cross-correlations between detrended fire and ENSO index time series were then examined using the CCF. 4. Results 4.1.
SPATIOTEMPORAL PATTERNS OF FIRE
Figures 2a–2f show the fire distribution for each of the years from 1996 (July– December only) through December 2001. The fires associated with the El Ni˜no
(a)
(b) Figure 2. Annual fire counts from ATSR nighttime satellite observations: (a) 1996 (July–December only); (b) 1997 (strong El Ni˜no); (c) 1998 (strong El Ni˜no to La Ni˜na transition); (d) 1999 (strong La Ni˜na); (e) 2000 (La Ni˜na); (f ) 2001 (La Ni˜na). (Continued on next page)
442
DOUGLAS O. FULLER AND KEVIN MURPHY
(c)
(d)
(e) Figure 2. (Continued )
THE ENSO-FIRE DYNAMIC IN INSULAR SOUTHEAST ASIA
443
(f) Figure 2. (Continued )
event can be seen in Figures 2b and 2c, which show that fire concentration shifted from Sumatra and southern Borneo in 1997 to the eastern side of Borneo in 1998. This pattern is consistent with a bimodal pattern that occurred during the El Ni˜no of 1982–1983, during which rains played a significant role in extinguishing fires during December. Drought then recurred in early 1983, helping to cause the devastating fires in East Kalimantan, Indonesia (Goldammer and Siebert, 1990). La Ni˜na conditions prevailed after June 1998 and fire number diminished significantly in the 3 years following the very strong El Ni˜no of 1997–1998 (Figures 2d–2f ). We used a vegetation map by MacKinnon (1996) to assess the temporal distribution of fire by cover type. The map covers Indonesia only and contains nine classes. We aggregated these classes into three main types: (1) forest, including lowland, upland, and montane types (total area = 489,645 km2 ); (2) non-forest, including agriculture, plantations, non-forest (degraded), and semi-deciduous woodland (total area = 883,037 km2 ); and (3) wetland, including freshwater swamp forest and mangrove formations (total area = 49,709 km2 ). Although much less extensive than the forest and non-forest classes, we analyzed wetlands as a distinct type because they were an important source of smoke, haze, and CO2 emissions, especially in areas of significant peat accumulation (Page et al., 2002). During the peak of the fires in September 1997 (Figure 3), the fires occurred more or less in synchrony within these three cover types; however, as weather conditions shifted from dry to humid (November–December 1997) during the El Ni˜no event itself, the number of fires that occurred within wetland and forest types diminished relative to those in non-forest locations. This pattern can be explained in part by the large amount of swamp forest in southern Borneo and Sumatra that burned in 1997 (Page et al., 2002) and relatively low fire susceptibility of closed-canopy moist forests relative to formations with open or disturbed canopies (Cochrane, 2003). Further, the study
444
DOUGLAS O. FULLER AND KEVIN MURPHY
Figure 3. Monthly time series of ATSR fire counts (log10 transformed) in three land-cover types: swamp-mangrove forests, non-forest (including agriculture, plantations, degraded lands), and forests (including lowland and upland types). Land-cover data provided by MacKinnon (1996).
area contains about twice as much non-forest cover than forest (MacKinnon, 1996), which helps to explain why non-forest locations tended to contain more active fires. Figure 3 reveals a pattern of consistently higher fire numbers in non-forest locations during the subsequent La Ni˜na, during which fires were detected during the wet seasons of 1999–2001. Since we lacked comparable cover data for areas outside Indonesia, we further stratified the study region into two broad latitudinal belts: an equatorial belt from 5.5◦ N to 5.5◦ S latitude, and seasonal belt from 5.5◦ S to 15.5◦ S, which covers the islands of Java, Bali, Nusa Tenggara, and a portion of northern Australia. Figure 4
Figure 4. Fire distribution within two latitudinal belts: −5.5–5.5◦ (equatorial) and −15.5 to −5.5◦ (seasonal).
THE ENSO-FIRE DYNAMIC IN INSULAR SOUTHEAST ASIA
445
shows the time series of monthly fire counts within these ranges and reveals strong, consistent 12-month seasonality of fires in the southern belt and weak seasonality in the equatorial belt. Fire seasonality was largely synchronous within the two latitudinal strata; however, the 1998 fires in Borneo (Figure 2c) appear as an anomalous spike in the time series (Figure 4). The weak seasonality evident in the equatorial belt suggests that even during La Ni˜na years, when wet conditions prevailed, there was a sufficient dry window during August and September to allow for some biomass burning. We used the total accumulated fires for the period 1996–2001 to examine further the distribution of fire. Using a set of randomly selected locations where fire occurred (n = 120) and did not occur (n = 283), we calculated the means and standard deviations (s.d.) of tree cover, human population and elevation (Figures 5a– 5c). The error bars on these plots show the standard error of the mean and suggest large differences in fire distribution relative to these three factors. We also used a t-test for independent samples to test for statistically significant differences for fire and non-fire locations by elevation, population density, and percent tree cover. The mean elevation for fire locations and non-fire locations (Figure 5a) was 116 m (s.d. = 226 m) and 274 m (s.d. = 333 m), which was significant (t-value = −4.73, p < 0.05). The large standard deviation for both fire and non-fire locations suggest that elevation is a poor predictor of fire location, although it is evident from other fire maps that high elevation (>1000 m) were largely spared from fire (Fuller and Fulk, 2001). The distribution of fire with respect to human population density
(a) Figure 5. Means for fire points (n = 120) and non-fire locations (n = 238) mapped from 1996 to 2001 for three variables related to fire: (a) elevation; (b) population density; and (c) percent tree cover. Error bars show the standard error of the mean. (Continued on next page)
446
DOUGLAS O. FULLER AND KEVIN MURPHY
(b)
(c) Figure 5. (Continued )
(Figure 5b) revealed that fires were more prevalent in areas of low population (mean = 16.5 persons km−2 , s.d. = 31.5 persons km−2 ) than in densely populated areas, which was also significant (t-value = −3.73, p < 0.05). This indicates that fires were largely a rural phenomenon or restricted to the peripheries of towns and cities. Field observations made during the 1998 fires in East Kalimantan (Fuller and Fulk, 1998) are consistent with this result. The mean tree cover in areas that experienced
447
THE ENSO-FIRE DYNAMIC IN INSULAR SOUTHEAST ASIA
fire was 42% (s.d. 25%); whereas, non-fire locations had a mean tree cover of 54% (s.d. 27%) (Figure 5c). Again, the t-test revealed statistical significance for this variable (t = −3.73, p < 0.05). This result supports general understanding of fire ecology in tropical regions, in which open canopy formations tend to possess lower humidity and thus higher flammability than areas with closed canopies (Cochrane, 2003). In addition, many of the areas that experienced fire have already burned in previous events and may become dominated by fire-adapted species (Van Nieuwstadt et al., 2001), which is consistent with the positive-feedback hypothesis proposed by Cochrane (2003). 4.2.
ASSOCIATION OF FIRES WITH
ENSO INDICES
ENSO appears to be a primary factor controlling fire distribution and timing, particularly in the equatorial belt of our study region (Figure 4); although strictly speaking, changes in pressure and temperature reflected in ENSO indices translate into changes in precipitation that control flammability of vegetation. Therefore, we examined the ENSO indices using rainfall data from four stations in eastern Borneo (Figure 1) for the period 1990–2002. These data characterize two distinct climatic regions: a seasonal area of southeast Borneo represented by the Balikpapan and Samarinda stations, and a consistently wetter region to the north represented by Tarakan and Tanjung Selor. For each precipitation time series the SOI was also examined, which revealed moderately strong association between ENSO and rainfall in the central part of the study region where many fires occurred in 1997– 1998. We calculated the Pearson correlation coefficient for precipitation at each station versus the five ENSO indices evaluated, including SOI, Ni˜no 1 + 2, Ni˜no 3, Ni˜no 3.4 and Ni˜no 4. Table II gives the results of this analysis and shows that the SOI had the largest absolute correlation coefficient (shown in bold in Table II) for the two southern stations (Balikpapan and Samarinda), whereas Ni˜no 3.4 showed a stronger relationship than other ENSO indices for the northern stations. This result is consistent with the increased use of the Ni˜no 3.4 index in ENSO forecasts (Glantz, 2001); however, the SOI showed a stronger relationship than other indices in two TABLE II Correlation coefficients, rainfall-ENSO indices for selected meteorological stations in East Kalimantan, Indonesia Station rainfall
SOI
Ni˜no 1 + 2
Ni˜no 3
Ni˜no 3.4
Ni˜no 4
Balikpapan Samarinda Tanjung Selor Tarakan
0.326 0.437 0.193 0.199
−0.191 −0.129 0.094 −0.119
−0.283 −0.330 −0.129 −0.186
−0.296 −0.415 −0.245 −0.207
−0.187 −0.323 −0.208 −0.116
Numbers in bold indicate the maximum absolute correlation for the five ENSO indices analyzed.
448
DOUGLAS O. FULLER AND KEVIN MURPHY
locations (Balikpapan and Samarinda) close to devastating fires of 1998 (Siegert et al., 2001). Although both indices were correlated with rainfall in the central part of the study region, the correlation coefficients shown in Table II are modest and suggest that their predictive utility may be limited for forecasting monthly rainfall in these locations. We examined the monthly fire counts (log10 transformed) against the five ENSO indices (Figure 6). The data reveal a strong inverse relationship between fire number and SOI at monthly time resolution. The SST-based indices were, in contrast, positively related to fire counts at this resolution. The zero-lag correlation coefficients for these monthly fire counts revealed relationship for SOI and Ni˜no 3.4 was comparable (−0.68 and 0.70, respectively). We also evaluated the strength of the relationship between fire and ENSO indices by calculating the correlations for El Ni˜no and La Ni˜na periods and we found that the SOI and Ni˜no 3.4 consistently showed higher zero-lag correlations than the other indices. Moreover, although we expected to find higher correlation coefficients for El Ni˜no periods than either La Ni˜na or El Ni˜no/La Ni˜na together, the strongest relationships were generally found using the complete time series of fire observations for the entire study region (July 1996–December 2001). Further, the strength of the relationship between ENSO
Figure 6. Time series of monthly ENSO indices, including SOI, Ni˜no 1 + 2, Ni˜no 3, Ni˜no 4, and Ni˜no 3.4, plotted against monthly fire counts from ATSR (log10 transformed).
THE ENSO-FIRE DYNAMIC IN INSULAR SOUTHEAST ASIA
449
Figure 7. Scatter plot showing the relationship between the SOI and ATSR fire counts (log10 transformed) in lowland tropical forest.
indices and fire was generally not improved when we calculated correlation coefficients by vegetation type or latitudinal strata (as described earlier). However, the correlations between forest fires and SOI and forest fires and Ni˜no 3.4 yielded slightly stronger relationships of −0.75 and 0.72 respectively. Figure 7 shows the relationship between the SOI and the log-transformed ATSR fire counts in lowland forest locations. A similar negative, linear relationship was found between log10 fire counts and SOI in non-forest (r = −0.69) and wetland locations (r = −0.55), although these correlation coefficients were slightly weaker than for lowland forest. Overall, these results suggest that the ENSO indices reflect broadly the regional patterns of climate and fire and may provide moderate utility for predicting fire patterns within forested regions of the study area, even though fires in upland areas appear to be rare. Subsequent to this analysis, we examined the autocorrelation and crosscorrelation functions for fire and SOI and Ni˜no 3.4, the two indices that showed consistently stronger relationships than other SST indices with rainfall and fire data. To reduce autocorrelation due to inclusion of areas with high seasonality where fires are common during La Ni˜na and El Ni˜no years (Figure 4), we focused subsequent time-series analysis on equatorial fires. Linear fits of fire, SOI, and Ni˜no 3.4 revealed strong negative (fire and Ni˜no 3.4) and positive (SOI) trends in the July 1996–December 2001 period. Therefore, we performed Census I–type detrending, which removes seasonal components as well as linear trends (Statsoft, 2003). Removal of linear trends components should be carried out precursor to CCF analysis. In addition, it is customary to remove the seasonal component after removal of linear trends to further reduce confounding effects of temporal autocorrelation (Vandaele, 1983).
450
DOUGLAS O. FULLER AND KEVIN MURPHY
The temporal autocorrelation for all fire and ENSO time series generally exhibited behavior in which the correlation coefficient dropped exponentially within 3–6 time lags or months in this case. Autocorrelation dropped more sharply for SOI than the other indices, indicating the presence of greater noise in this particular series. It should be noted that the Ni˜no 3.4 index appears more highly smoothed than the SOI (Figure 6), which proved problematic for transforming these series to remove residual autocorrelation. Inspection of the autocorrelograms (not shown) indicated that the Census I technique removed linear trends from the Ni˜no 3.4, but did not remove autocorrelation from the time series. However, both detrending techniques removed linear trends and autocorrelation from the noisier SOI data. Figure 8 shows the plot of detrended SOI versus detrended equatorial fires. Analysis of the cross-correlogram of these detrended series, in which equatorial fires were lagged by SOI, shows two spikes at lags −3 and +9. The positive correlation of 0.42 at lag +3 indicates that SOI may serve as a weak predictor of the number of equatorial fires 3 months in advance. The moderate correlation coefficient at lag +9 (r = 0.60) suggests that major fire events peak well before the SOI reaches a minimum. Although this result is possibly unique to this particular ENSOfire cycle, data compiled by the US National Weather Service Climate Prediction Center (http://www.cpc.ncep.noaa.gov/data/indices/soi) indicate that SOI reached
Figure 8. Seasonally detrended, stationary time series for monthly fire counts (log10 transformed) in the equatorial belt (−5.5 to 5.5◦ ) and SOI.
THE ENSO-FIRE DYNAMIC IN INSULAR SOUTHEAST ASIA
451
Figure 9. Annual sums of SOI and fire counts for 1997–2001.
a minimum value comparable to the minimum reached during the 1982–1983 El Ni˜no event in February. Unfortunately, ATSR fire counts were not available for the 1982–1983 event. We aggregated the monthly data to an annual scale by summing SOI and fire counts (for the entire region) for each year to reveal an intriguing pattern shown in Figure 9. This reveals that the sum of SOI for 1997 was highly negative (−37.8), while the number of fire detections approached 30,000. This figure suggests that for negative annual SOI, annual fire counts and annual SOI are proportional and that SOI may prove to be a useful predictor of fire numbers in the study region. Once the annual sum of SOI becomes positive, it appears that fire numbers remain low (below 5000 in this case). 5. Discussion and Conclusions The analysis of CCF suggests that the number of fires at the monthly time scale is a better predictor of SOI than SOI is of fire. However, more data are probably needed to test this hypothesis in a more rigorous fashion. At an annual time scale there appears to be linearity between SOI and total annual fire counts if the sum of annual SOI ≤0, although again more fire observations are needed to ascertain whether a linear function fits the data or whether the relationship follows a negative exponential function. In any event, the aggregation of data to an annual time scale reveals some clarity in the relationship between fire and ENSO, similar to what other researchers have found when monthly observations are aggregated using sums or averages (Goward and Prince, 1995). The aggregation of data from the monthly to annual scale shows that time essentially acts as a filter that removes noise that is present at finer temporal scales. A key consideration to further our understanding of the ENSO-fire dynamic is to extend the length of the fire time series using comparable satellite observations.
452
DOUGLAS O. FULLER AND KEVIN MURPHY
In this analysis, we only possessed enough fire data to cover one ENSO cycle that incorporated a very strong El Ni˜no event and the subsequent strong La Ni˜na. It remains to be seen whether inclusion of more moderate El Ni˜no and La Ni˜na episodes in the fire time series will improve the predictive capacity of ENSO indices at either the monthly or annual time scales. Although other fire data from sensors such as the AVHRR and the Moderate Resolution Imaging Spectrometer (MODIS) exist, it remains a challenge to compare these because they are derived from very different algorithms and portions of the diurnal fire cycle. At present, the longest data set available is the 1-km ATSR monthly nighttime fire counts used in this study. To extend this time series forward would require some temporal overlap (∼3 months) between AVHRR, ATSR, and MODIS fire products and possible finetuning of nighttime algorithms (i.e., threshold adjustment) to reduce omission and commission errors between the ATSR and the three other data sets. In this way, it may be possible to reduce discontinuities (i.e., gaps) and create a comparable time series of fire counts that can be used to evaluate any trends over the last two decades. The Advanced ATSR (AATSR), launched in March 2002, may also be a good source for extending the time series forward, and we intend to evaluate data as they become available. Construction of the longer time series will also help to validate different algorithms and fire products and allow us to derive new metrics related to long-term fire frequency, such as number of fire occurrences per pixel per decade, maximum number of fire occurrences within the annual cycle, and descriptive statistics such as the mean annual fires per land-cover class presented here. In addition, because much research has emphasized the Kalimantan fires associated with recent severe El Ni˜no events (Fuller and Fulk, 2000, 2001; Siegert et al., 2001; Stolle et al., 2003) we know relatively little about the fire distribution during relatively wet (La Ni˜na) periods. A long time series of fire observations will allow us to capture a more reliable estimate of the baseline level of seasonal burning during non El Ni˜no events and determine whether this baseline is increasing as land-cover change continues in the region. Although desirable, a long, internally consistent fire time series from satellite observations may not prove to be an unbiased estimate of interannual fire variability. For example, Kasischke et al. (2003) showed that ATSR fire detections in a boreal region were positively biased during periods of major burning. This suggests that there will always be trends in fire time series, although it remains to be seen whether the same problem applies to humid tropical regions such as Southeast Asia. Unfortunately, reliable burn scar data, which Kasischke and colleagues (2003) used for assessing ATSR observations of Canadian wildfires, are not available for parts of Indonesia, Malaysia, and the Philippines analyzed in this study. Other multivariate statistical methods such as logistic regression may help to advance prediction of El Ni˜no–related fire events similar to those experienced in 1982–1983 and 1997–1998. For example, Stolle (2003) used this method successfully to predict fire occurrence within certain land-cover types in southern Sumatra.
THE ENSO-FIRE DYNAMIC IN INSULAR SOUTHEAST ASIA
453
Canonical correlation analysis and similar data-reduction techniques have been applied to long-time series of annual data from paleoclimate reconstructions to successfully model area burned in the western United States (e.g. Westerling and Swetnam, 2003). Such work, which relies upon long time series in which subsequent observations are largely independent (i.e., show little autocorrelation), shows the importance of ENSO and climatic change, in general, in producing favorable fire regimes. If indeed the El Ni˜no phase of ENSO does become stronger as global warming continues into the next century, then we can assume that fires will grow more destructive and continue to be a major factor in the massive deforestation found in our study region. Although fuel loads will decrease as fires remove remaining standing biomass, the climatic effects of deforestation are likely to include reduction of annual rainfall and heightened seasonality as noted by Shukla et al. (1990) in their study of Amazonian deforestation. This will likely reinforce the fire dynamic such that episodic, El Ni˜no fires will become replaced by seasonal burning of highly disturbed pyrophytic vegetation, which characterizes large areas of insular Southeast Asia today. Acknowledgments We wish to thank Dan Griffith for his suggestions regarding the use of time-series statistics. Kenny Broad and Amy Clement deserve special thanks for their constructive comments on an early version of this work. In addition, three anonymous reviewers provided many helpful suggestions on the original manuscript. References Anyamba, A. and Eastman, J. R.: 1996, ‘Interannual variability of NDVI over Africa and its relation to El Nino Southern Oscillation’, Int. J. Remote Sens. 17, 2533–2548. Barber, C. V. and Schweithelm, J.: 2000, Trial by Fire, World Resources Institute, Washington, DC. Barnston, A. G., Glantz, M. H., and He, Y.: 1999, ‘Predictive skill of statistical and dynamical climate models in SST forecasts during the 1997–1998 El Ni˜no episode and the 1998 La Ni˜na onset’, Bull. Am. Meterol. Soc. 80, 217–243. Broad, K., Pfaff, A. S. P., and Glantz, M. H.: 2002, ‘Effective and equitable dissemination of seasonalto-interannual climate forecasts: Policy implications from the Peruvian fishery during El Nino 1997–1998’, Clim. Change 54, 415–438. Cochrane, M. A.: 2003, ‘Fire science for rainforests’, Nature 421, 913–919. Fuller, D. O. and Fulk, M.: 1998, ‘Satellite remote sensing of the 1997–1998 fires in Indonesia: Data, methods, and future perspectives’, Unpublished report submitted to The World Wide Fund for Nature – Indonesia Programme, Jakarta, Indonesia. Fuller, D. O.: 2000, ‘Satellite remote sensing of biomass burning using optical and thermal sensors’, Prog. Phys. Geogr. 24, 543–561. Fuller, D. O. and Fulk, M.: 2000, ‘Comparison of NOAA-AVHRR and DMSP-OLS for operational fire monitoring in Kalimantan, Indonesia’, Int. J. Remote Sens. 21, 181–187. Fuller, D. O. and Fulk, M.: 2001, ‘Burned area in Kalimantan, Indonesia mapped with NOAA-AVHRR and Landsat TM imagery’, Int. J. Remote Sens. 22, 691–697.
454
DOUGLAS O. FULLER AND KEVIN MURPHY
Glantz, M.: 2001. Currents of Change: Impacts of El Ni˜no and La Ni˜na on Climate and Society, 2nd edn., Cambridge University Press, Cambridge, UK. Glantz, M. and Nicholls, N.: 1997, ‘Appendix: Chronology of interest in El Ni˜no’, in Glantz, M. (ed.), Currents of Change: Impacts of El Ni˜no and La Ni˜na on Climate and Society, 2nd edn., Cambridge University Press, Cambridge, UK. Goldammer, J. G. and Siebert, B.: 1990, ‘The impact of droughts and forest fires on tropical lowland rain forest of East Kalimantan’, in Goldammer, J. G. (ed.), Fire in the Tropical Biota, Ecological Processes and Global Challenges, Berlin, Springer-Verlag, Berlin, pp. 11–31. Goward, S. N. and Prince, S. D.: 1995, ‘Transient effects of climate on vegetation dynamics: Satellite observations,’ J. Biogeogr. 22, 549–564. Hansen, M., DeFries, R., Townshend, J. R. G., Sohlberg, R., Dimiceli, C., and Carroll, M.: 2002, ‘Towards an operational MODIS continuous field of % tree cover algorithm: Examples using AVHRR and MODIS data’, Remote Sens. Environ. 83, 303–319. Harrison, M. and Meindl, C. F.: 2001, ‘A statistical relationship between El Nino-Southern Oscillation and Florida wildfire occurrence’, Phys. Geogr. 22, 187–203. Herbert, J. M. and Dixon, R. W.: 2003, ‘Is the ENSO phenomenon changing as a result of global warming?’, Phys. Geogr. 23, 196–211. Hill, R.: 2002, Southeast Asia: People, Land and Economy, Allen and Unwin, Crows Nest, NSW, Australia, 310 pp. Justice, C. O., Giglio, L., Korontzi, S., Owens, J. J., Morisette, T., Roy, D., Descloitres, J., Alleaume, J. S., Petitcolin, F., and Kaufman, Y.: 2002, ‘The MODIS fire products’, Remote Sens. Environ. 83, 244–262. Kasischke, E. S., Hewson, J. H., Stocks, B., van der Werf, G., and Randerson, J.: 2003, ‘The use of ATSR active fire counts for estimating relative patterns of biomass burning – A study from the boreal forest region’, Geophys. Res. Lett. 30 (18): article number 1969. Kirono, D. G. C., Tapper, N. J., and McBride, J. L.: 1999, ‘Documenting Indonesian rainfall in the 1997/1998 El Ni˜no event’, Phys. Geogr. 20, 422–435. Kitzberger, T., Swetnam, T. W., and Veblen, T. T.: 2001, ‘Inter-hemispheric synchrony of forest fires and the El Ni˜no-Southern Oscillation’, Global Ecol. Biogeogr. 10, 315–326. Lau, K. M. and Wu, H. T.: 1999, ‘Assessment of the impacts of the 1997–1998 El Nino on the Asian–Australia monsoon’, Geophys. Res. Lett. 26, 1747–1750. Laurance, W. F., Williamson, G. B., Delamonica, P., Oliveira, A., Lovejoy, T. E., Gascon, C., and Pohl, L.: 2001, ‘Effects of a strong drought on Amazonian forest fragments and edges’, J. Trop. Ecol. 17, 771–785. Lim, C. and Kafatos, M.: 2002, ‘Frequency analysis of natural vegetation distribution using NDVI/AVHRR data from 1981 to 2000 for North America: Correlations with SOI’, Int. J. Remote Sens. 23, 3347–3383. MacKinnon, J.: 1996, Review of the Protected Areas System of the Indomalayan Realm: Remaining Habitat Map, based on 1980–1990’s AVHRR and LANDSAT images. Scale 1:1,000,000. Distributed in digital form by the World Conservation Monitoring Centre, Cambridge, UK. Matson, M. and Dozier, J.: 1981, ‘Identification of subresolution high temperature sources using a thermal IR sensor, Photogramm. Eng. Remote Sens. 47, 1311–1318. Nicholls, N. and Beard, G.: 2001, ‘The application of El Ni˜no-Southern Oscillation information to seasonal forecasts in Australia’, in White, D. A. (ed.), Drought: A Global Assessment, Vol. 1., Routledge, London and New York, pp. 41–51. Page, S. E., Siegert, F., Rieley, J. O., Boehm, H.-D. V., Jaya, A., and Limin, S.: 2002, The amount of carbon released from peat and forest fires in Indonesia during 1997’, Nature 420, 61– 65. Riedinger, M. A., Steinitz-Kannan, M., Last, W. M., and Brenner, M.: 2002, ‘A ∼6100 14 C yr record of El Ni˜no activity from the Galapagos Islands’, J. Paleolimnol. 27, 1–7.
THE ENSO-FIRE DYNAMIC IN INSULAR SOUTHEAST ASIA
455
Shukla, J., Nobre, C., and Sellers, P.: 1990, ‘Amazon deforestation and climate change’, Science 247, 1322–1325. Siegert, F., Ruecker, G., Hinrichs, A., and Hoffmann, A. A.: 2001, ‘Increased damage from fires in logged forests during droughts caused by El Ni˜no’, Nature 414, 437–440. Statsoft: 2003, Statistica 6.1 Software, Statsoft Inc., Tulsa, OK. Stolle, F. and Tomich, T. P.: 1999, ‘The 1997–1998 fire event’, Nat. Resources 35, 22–30. Stolle, F.: 2003, Land Use as Driver of Vegetation Fires in Indonesia: A Spatial Modelling Approach to Identify and Quantify Causes of Fire, Ph.D. Dissertation, Universit´e catholique de Louvain, Belgium. Stolle, F., Chomitz, K. M., Lambin E. F., and Tomich, T. P.: 2003, ‘Land use and vegetation fires in Jambi Province, Sumatra, Indonesia’, Forest Ecol. Manag. 179, 277–292. Trenberth, K. E.: 1997, ‘The definition of El Ni˜no’, Bull. Am. Meteorol. Soc. 78, 2771–2777. Trenbert, K. E. and Stepaniak, D. P.: 2001, ‘Indices of El Ni˜no evolution’, J. Climate 14, 1697–1701. Van Nieuwstadt, M. G. L., Sheil, D., and Kartawinata, K.: 2001, ‘The ecological consequences of logging in the burned forests of East Kalimantan, Indonesia’, Conserv. Biol. 15, 1183–1186. Vandaele, W.: 1983, Applied Time Series and Box-Jenkins Models, Academic Press, New York. Varma, A.: 2003, ‘The economics of slash and burn: A case study of the 1997–1998 Indonesian forest fires’, Ecol. Econ. 46, 159–171. Westerling, A. L. and Swetnam, T. W.: 2003, ‘Interannual to decadal drought and wildfire in the western United States’, Eos: Trans. Am. Geophys. Union 84, 545, 554–555. Yu, B. and Boer, G. J.: 2002, ‘The roles of radiation and dynamical processes in the El Nino–like response to global warming’, Clim. Dyn. 19, 539–553. (Received 30 March 2004; in revised form 29 November 2004)