Micrometeorological and Canopy Controls of Fire Susceptibility in a Forested Amazon Landscape

Ecological Applications, 15(5), 2005, pp. 1664±1678 ᭧ 2005 by the Ecological Society of America

MICROMETEOROLOGICAL AND CANOPY CONTROLS OF FIRE SUSCEPTIBILITY IN A FORESTED AMAZON LANDSCAPE

DAVID RAY,1,4 DANIEL NEPSTAD,1,2,3,5 AND PAULO MOUTINHO2 1Woods Hole Research Center, P.O. Box 296, Woods Hole, Massachusetts 02543 USA 2Instituto de Pesquisa Ambiental da Amazonia, Av. Nazare 669, 66035-170, BeleÂm, PA, Brazil 3Universidade Federal Do ParaÂ (UFPa), Nucleo de Altos Estudos Amazonicos, Av. Augusto Corre, Campus da Universidade-Guama, CEP 66.059, BeleÂm, ParaÂ, Brazil

Abstract. Fire is playing an increasing role in shaping the structure, composition, and function of vast areas of moist tropical forest. Within the Brazilian Amazon, cattle ranching and swidden agriculture provide abundant sources of ignition to forests that become susceptible to ®re through selective logging, severe drought and, perhaps, fragmentation. Our understanding of the biophysical factors that control ®re spread through Amazon forests remains largely anecdotal, however, restricting our ability to model the Amazon ®re regime, and to simulate the effects of trends in climate and land-use activities on future regimes. We used experimental ®res together with measurements of micrometeorology (rainfall, vapor pressure de®cit [VPD], wind velocity), canopy attributes (leaf area index [LAI], canopy height), and fuel characteristics (litter moisture content [LMC] and mass) to identify the variables most closely associated with ®re susceptibility in the east-central Amazon. Fire spread rates (FSR, m/min) were measured in three common forest types: an 8-yr-old regrowth forest, a recently logged/burned forest, and a mature forest. One hundred ®res were set in each study area during the last two months of the 2002 dry season. VPD, recent precipitation history, wind velocity, and LAI explained 57% of the variability in FSR. In combination, LAI, canopy height, and recent precipitation history accounted for ϳ65% of the variability in VPD, the single most important predictor of FSR, and approximately half of the total observed variability in FSR. Using logistic regression we were able to predict whether a ®re would spread or die 72% of the time based on LAI, canopy height, and recent precipitation history. An approximate threshold in ®re susceptibility was associated with a LMC of ϳ23%, somewhat higher than previously reported (15%). Fire susceptibility was highest under low, sparse canopies, which permitted greater coupling of relatively hot, dry air above the canopy with the otherwise cool, moist air near the forest ¯oor. Fire susceptibility increased over time after rain events as the forest ¯oor gradually dried. The most important determinants of ®re susceptibility can be captured in ecosystem and climate models and through remotely sensed estimates of canopy structure, canopy water content, and microclimatic variables. Key words: Amazon; canopy structure; drought; experimental ®re; rain forest ®re.

INTRODUCTION et al. 2004). Drought in these regions may grow in severity and frequency in the future through changes Fire impoverishes vast areas of tropical forest in the in the ENSO regime associated with global warming Amazon Basin (Cochrane et al. 1999, Nepstad et al. (Trenberth and Hoar 1997, Timmermann et al. 1999), 1999a, MendoncËa et al. 2004) and Southeast Asia (Sie- deforestation-induced inhibition of rainfall (Nobre et gert et al. 2001, Page et al. 2002) through the inter- al. 1991, Silva Dias et al. 2002), and through greater action of drought and logging. These ®res are espe- evapotranspiration associated with rising temperatures cially severe during El NinÄo Southern Oscillation (White et al. 1999). Hence, while ®res have occurred (ENSO), when these regions experience severe in the Amazon for at least 6000 years (Sanford et al. drought. During the 1997±1998 ENSO, approximately 1985), the observation that their frequency is increas- 40 000 and 14 000 km2 of forest experienced understory ing from 400±700-yr return intervals that appear to ®res in the Amazon Basin and Borneo, respectively, have characterized the region in pre-Columbian times releasing 0.2±0.4 and 0.9 Pg of carbon to the atmo- (Meggers 1994) to less than 25-yr return intervals in sphere (Page et al. 2002, Alencar et al. 2004, MendoncËa affected areas today (Cochrane et al. 1999) is of great concern. Manuscript received 10 March 2005; accepted 29 March 2005. The ®rst time a mature Amazon forest burns, the ®re Corresponding Editor: A. R. Townsend. ϳ 4 Present address: Department of Forest Ecosystem Sci- tends to be low ( 10 cm ¯ame height) and slow moving ence, University of Maine, Orono, Maine 04469 USA. (0.25 m/min), yet can cause high tree mortality rates 5 Corresponding author; e-mail: [email protected] even among large trees (23±44% of trees Ͼ10 cm dbh) 1664 October 2005 RAIN FOREST FIRE SUSCEPTIBILITY 1665 because of the long contact time between the ¯ame and the eastern and southern portions of the region (Nep- the base of the tree (Holdsworth and Uhl 1997, Coch- stad et al. 1999a, b), and may expand into the region's rane et al. 1999, Cochrane and Schulze 1999, Gerwing interior as new highways are paved (Nepstad et al. 2002, Barlow et al. 2003). Unlike grasslands, savannas, 2001). Approximately 200 000 to 300 000 km2 of land and many coniferous forests, in which periodic ®res (between 30% and 50% of the previously deforested lower susceptibility to additional ®re by reducing fuel area) in the Brazilian Amazon is in some stage of forest loads (Pyne et al. 1996), wildland ®res in moist tropical recovery (Fearnside and GuimaraÄes 1996, Houghton et forests appear to increase the likelihood of further burn- al. 2000, Zarin et al. 2001), while 10 000±15 000 km2/ ing. Subsequent ®res are more intense (higher ¯ame yr of forest are selectively logged (Nepstad et al. heights and faster spread rates), perhaps because of the 1999b). Simulation and prediction of both short-term reduced canopy density and increased fuel loads as- trends in ®re risk and ®re occurrence, and long-term sociated with the tree mortality from the ®rst ®re, and changes in the Amazon ®re regime that may result from can kill additional and larger trees (Cochrane and climate change and frontier expansion, will require a Schulze 1999). In the most extreme scenario, recurrent mechanistic understanding of the controls on forest sus- ®res may promote the conversion of high forest to ®re- ceptibility to ®re. Information derived from controlled adapted scrub vegetation in a process termed savani- ®re experiments is necessary to elucidate the under- zation (Cochrane and Schulze 1998, Kinnaird and lying mechanisms. Past work has pointed to linkages O'Brien 1998). The effects of understory ®res on forest between forest structure, understory microclimate and fauna (mammals and large-bodied birds) appear also the moisture content of ®ne fuels (Uhl et al. 1988, Uhl to be large (Kinnaird and O'Brien 1998, Peres et al. and Kaufmann 1990, Holdsworth and Uhl 1997). How- 2003, Barlow and Peres 2004). ever, rigorous testing of how these variables contribute It is hypothesized that canopy density (e.g., leaf area to ®re susceptibility has not been carried out within the index, LAI) is an important determinant of forest sus- seasonally dry Amazon, where these ®res represent the ceptibility to ®re in the Amazon (Nepstad et al. 1995), greatest threat. Fire science remains a ¯edgling disci- although this hypothesis remains to be rigorously test- pline in the moist tropics, and basic information is re- ed. Large areas of mature, moist tropical forest in the quired to determine the appropriateness of extending Amazon remain resistant to ®re even during dry sea- existing ®re modeling frameworks developed for tem- sons of Ͼ4 mo duration by maintaining a dense leaf perate forests to these ecosystems (Cochrane 2003). canopy that prevents the solar heating of the forest Therefore, we conducted a large number of experi- understory that is necessary to dry the forest ¯oor mental ®res in forests with markedly different struc- (Nepstad et al. 1995). Canopy maintenance during pro- tures in order to (1) determine the physical and envi- longed drought is possible through water absorption by ronmental factors regulating the behavior of understory root systems that penetrate Ͼ8 m into the soil (Nepstad ®res, and (2) assess the potential for applying the re- et al. 1994, 1995, 2004, Jipp et al. 1998). Forest sus- sulting relationships to the prediction of ®re suscep- ceptibility to ®re increases when canopy damage tibility in the eastern Brazilian Amazon. caused by selective logging (Woods 1989, Uhl and Vi- eira 1989, Uhl and Kauffman 1990, Holdsworth and METHODS Uhl 1997, Alencar et al. 2004), or resulting from tree Study system death associated with previous understory ®res (Coch- rane and Schulze 1999, Cochrane et al. 1999), creates The study sites were located 100 km south of the canopy gaps and reduces LAI, allowing more desic- city of SantareÂm in ParaÂ State, Brazil (ϳ3Њ S, 55Њ W). cating solar radiation to penetrate the forest interior, Measurements were carried out within three forest while adding to the fuel load on the forest ¯oor. Re- structure types with known disturbance histories. In- growing forests may be more susceptible to ®re than cluded were (1) a mature forest (Mfor) that was lightly mature forests (Uhl et al. 1988) because of their rel- logged (Ͻ10 m3/ha harvested) 14 yr prior in 1988; (2) atively low LAI and, perhaps, their low stature, which a forest that was selectively logged two to three times allows tighter coupling of the canopy and forest interior between 1993 and 1997 (ϳ30 m3/ha harvested), and air. Severe drought may also increase forest suscepti- burned soon after in 1997 (L/Bfor); and, (3) a young bility to ®re. As deep soil moisture is depleted, canopy regrowing forest (Rfor) on a former swidden agriculture thinning begins to take place, allowing greater pene- site, abandoned 8 yr prior. The L/Bfor and Rfor were tration of solar radiation to the forest ¯oor as recently located approximately 500 m apart; the Mfor was sit- documented in a large throughfall exclusion experi- uated ϳ10 km southwest of the other areas. ment (Nepstad et al. 2002). Forest edges may be more Rainfall in the region is strongly seasonal (Fig. 1). susceptible to ®re because of drying and tree mortality Between 1999 and 2002, the average annual precipi- associated with these environments (Kapos 1989, Laur- tation was 2200 mm measured at a nearby research site ance et al. 2001) (Nepstad et al. 2002), and 1843 mm/yr based on a 25 In the Amazon, severe seasonal drought, selective yr record from a long-term met station in Belterra, logging, and regrowing forests are concentrated along located ϳ50 km to the north. Approximately 75% of 1666 DAVID RAY ET AL. Ecological Applications Vol. 15, No. 5

lowing the event. For example, ®ve days after a 10- mm rain event the value of the weighted precipitation would be 2 mm/d (10 mm/5 d). When two or more rainfall events occurred close enough in time that the value of the ®rst had not dropped below 1 mm/d, this remainder was added to the new rainfall total before dividing by the updated rainless days count. The objective was to derive a measure of precipitation that would incorporate differences in event size and their diminishing in¯uence over time. Because pluviometers had not been set up prior to the initial experimental ®re campaign we had to rely on rainfall estimates obtained from other sources for initializing our weighted precipitation variable. These included (1) the moisture FIG. 1. Daily rainfall for 2002 from a nearby study site content of the leaf litter when we ®rst arrived at the located ϳ36 km from the areas where experimental ®res were conducted. Diagonal crosshatching indicates the pe- sites to take measurements, as determined in the lab, riod during the dry season when the experimental ®res took (2) rainfall measured in a wedge shaped rain gage lo- Ͻ place. cated at the entrance to the Mfor ( 3 km away), and (3) conversations with local residents. All of these factors pointed to no measurable rainfall for at least 1- this annual total comes during the 6-mo wet season, week prior. We then identi®ed a median event size of between January and June, at both locations (Fig. 1). 2 mm based on the previous months rainfall record Soils in the area are deeply weathered clays (Oxisol, from a pair of pluviometers located at a nearby study Haplustox), interspersed with pockets of coarser tex- site. Thus, the value of the weighted precipitation var- tured materials (Silver et al. 2000). Our study sites were iable was initialized at 0.29 mm/d (2 mm/7 d). located on the ®ner textured soils. These soils have a On 18 September 2002, we established a monitoring highly aggregated structure, and drain well. station for measuring temperature (T, ЊC) and relative humidity (RH, %) within an open area located close to Sampling design Ͻ the Mfor ( 3 km away). A pair of Hobo RH-T sensors A sampling grid was established within each area. (Onset Computer Corporation, Pocasset, Massachu-

At the Mfor and L/Bfor sites, four transect lines were setts, USA), mounted within a U.S. Weather Service established within ϳ25-ha blocks. The parallel transect approved cabinet and ®xed to a wooden post 1.3 m lines were 250 m long with 250 m between them, each above the ground, were used to log readings at half- consisting of 10 evenly spaced (25-m interval) grid hour intervals for the duration of the study period (18 ϭ points (n 40 points/area). The sample grid at the Rfor September to 27 November 2002). Similar measure- was smaller, covering ϳ2.5 ha in total area; here, four ments were also available from a long-term meteoro- 100 m long transects were established at 50-m inter- logical station located ϳ36 km away (Nepstad et al. vals, each having 10 grid points separated by 10 m (n 2002). Average midday values for T and RH, deter- ϭ 40 points). Narrow trails (0.5 m wide) were created mined between 13:00 and 15:00 local time, were used to provide access within each area; care was taken to to calculate vapor pressure de®cit (VPD, kPa). avoid cutting of larger stems that might thin the canopy. Sampling points were established approximately 3 m Forest structure from the access trail on both sides of each grid point, Aboveground forest biomass was estimated for each for a total of 80 sample points per forest. Twenty ad- area using allometric equations developed for primary ditional sample points were located at random in each (Chambers et al. 2001) and secondary (Nelson et al. forest type for a total of 100 sample points per area. 1999) forest species in the Brazilian Amazon. At the Ն Mfor, we measured all trees 10 cm dbh (diameter at Rainfall and understory microclimate breast height) within eight 20 ϫ 50 m plots centered Daily rainfall estimates for the study sites were pro- at randomly selected locations along our grid lines, and vided by automated pluviometers (RainWise 20 cm di- all trees Ն30 cm in 30 ϫ 50 m plots at these same Ն Ն ameter tipping-bucket; RainWise, Inc., Bar Harbor, locations. At the L/Bfor trees 10 cm dbh and 30 cm Maine, USA) beginning on 25 October 2002. Two sam- dbh were measured along ®ve continuous vegetation pling locations were chosen, one in an open ®eld at the transects of 4 ϫ 500 m and 10 ϫ 500 m, respectively. Ն entrance to disturbed sites (L/Bfor and Rfor), and within And at the Rfor, live trees 5 cm dbh were sampled ϫ a large canopy gap along the access road to the Mfor. along four 5 50 m transects centered on the grid This information was used to derive a weighted pre- lines; stems between 1 and 5 cm dbh were measured cipitation variable, obtained by dividing the size of the on 1 m radius circular plots at 40 randomly selected most recent event by the number of rainless days fol- grid points. An average canopy height was determined October 2005 RAIN FOREST FIRE SUSCEPTIBILITY 1667 for each study area based on clinometer calibrated oc- tance between sample points, the number of ®res com- ular estimation of the height of the dominant vegetation pleted in each area on each day was not always the within each area. same. In total, 100 experimental ®res were set in each An estimate of leaf area index (LAI) was obtained area (300 total) over the course of 13 measurement for each grid point using a LI-COR LAI-2000 plant days. canopy analyzer (LI-COR, Inc., Lincoln, Nebraska, Crews of three to four people were responsible for USA). Paired optical sensors were used to take simul- carrying out the experimental ®res in each area, allow- taneous measurement of above and below canopy light ing nearly simultaneous measurement of microclimatic conditions. Above-canopy measurements were ob- and ®re behavior variables at each grid point, and tained in large open areas nearby the study sites, while across areas. At each location, we ®rst measured air ϳ two below-canopy readings were taken 1 m above RH and T, and collected a ®xed area sample of the leaf Њ the ground at each sample point. A 45 view cap was litter for mass determination. Wind speed at the time used to help isolate the individual point estimates of of the ®re was estimated by hanging a 1 m long piece LAI; information from the ®fth concentric ring of the of ¯agging tape at arm's length for a period of 1 min Њ optical sensor (62.3±74.1 view angle) was dropped and noting the maximum degree of displacement from from the analysis to address this same issue. These the vertical using three categories: zero to indicate no measurements were made at two times during the study displacement, one if the tape was between 0Њ and 45Њ period; sample points on the right-hand side of the from the vertical, and two if it went above 45Њ. access trails were done ®rst (on 15±16 October), and Experimental ®res were carried out by ®rst placing points on the left-hand side were measured approxi- a 20 cm radius circular metal hoop on the forest ¯oor mately one month later. We measured LAI early in the at the grid point. Steel pins were then pushed into the morning (6:00±7:30 local time) to minimize the in¯u- soil 30 cm outside of the hoop in each cardinal direc- ence of heterogeneous sky conditions and glare. Litter biomass was determined by collecting all tion. Approximately 10 mL of kerosene was then ap- leaves and small twigs (Ͻ1 cm diameter) above the plied to the litter within the metal hoop to ensure initial mineral soil within 40 cm diameter circular plots (0.13 ignition. Fires were allowed to burn for a distance of m2/sample), at the times and locations when experi- upto1minanydirection from the plot center, or for mental ®res were carried out. Litter samples were a period of up to 5 min, whichever came ®rst. The 1- weighed at the time of collection, and biomass and m maximum distance criterion was established because moisture contents were determined after drying sam- of the risks associated with allowing the ®res to burn ples for 48 h at 65ЊC in the lab. for up to 5 min regardless of spread distance. Observers noted the time it took for the ¯ames to reach each of Experimental ®re measurements the four pins, and the 1-m distance for ®res that spread that far. Flame heights, measured perpendicular to the Five experimental ®re campaigns were carried out ground surface, were recorded when the ®re reached 1 approximately biweekly between 25 September and 27 m, or had burned for 5 min; ¯ame heights were not November 2002. A typical campaign lasted three to measured for ®res that burned out. After each ®re was four days. The ®rst day consisted of measuring under- extinguished, we measured the distance of spread in story T and RH using a battery-powered, ventilated, the direction of each pin, and also along the radius of wet-bulb-dry-bulb psychrometer, and litter mass and maximum ®re spread when it was different. To allow moisture content as described earlier. Psychrometer the calculation of spread rates in cases where the ®re readings were taken just above (ϳ5 cm) the forest ¯oor at each sample point. All measurements were made did not burn for at least 5 min or arrive at the 1 m simultaneously by two-person crews working indepen- distance threshold, we also recorded the time that each dently in each area; these data were gathered between of these ®res died. Fire spread rate (FSR) was deter- ϳ13:00±14:00 local time. The purpose of these mea- mined by dividing the distance along the axis of max- surements was to quantify variability in understory imum spread (less the 20 cm distance to the center of conditions across the entire study areas within a short the circle where kerosene was used to ignite the ®re) timeframe, something that was not possible to do when by the amount of time the ®re was actively burning. carrying out the experimental ®res. The ®nal campaign Finally, each ®re was classi®ed in regards to the did not include this initial day of area-wide measure- likelihood that it would have continued burning if not ments. extinguished. A binary response variable (spreading/ During the subsequent days of each campaign, we dying) was assigned to each ®re based on agreement set between ®ve and 10 experimental ®res per day at among team members. Fire intensity (visual estimation randomly selected sample points within each area. of increasing/decreasing) and continuity of surround- These ®res commenced at ϳ13:00 and were always ing fuels provided the primary criteria for making these completed before 16:00. As a result of differences be- assessments. Fires that did not require manual extinc- tween forests in terms of ease of movement and dis- tion were assigned to the dying category. 1668 DAVID RAY ET AL. Ecological Applications Vol. 15, No. 5

Statistical analysis most extreme conditions (ambient met station VPD ϭ 2.64 kPa), the average mid-day VPD was 4.7 and 3.6 The analyses view the three study areas as repre- times higher at the R and L/B forest sites than within senting a continuum of forest structures. A combination for for the M understory, respectively. On average, midday of regression techniques (nonlinear, stepwise linear, for VPD at the M remained between 2.7 and 2.1 times and logistic) was used to identify relationships among for lower than those at the R and L/B , respectively. The the measured variables. Our approach was to construct for for maximum VPD recorded at individual sampling points a set of models that would allow prediction of ®re in each forest structure type was 4.6, 3.2, and 2.4 kPa behavior and risk based on (1) all signi®cant measured for the R , L/B , and M , respectively. In sum, the variables, and (2) the subset of more easily measured for for for drying environment measured at the disturbed forests variables that describe forest canopy structure (average was more similar to that encountered at the met station canopy height and LAI) and recent precipitation history than within the M understory. (weighted precipitation [mm/d]). Our estimate of wind for speed was treated as a class variable because it was Forest canopy and biomass not measured on a continuous scale. Degrees of free- dom were provided by the individual experimental ®res Forest canopies ranged from the short stature and carried out in each area (n ϭ 300). A log-transformation low LAI of the Rfor, to the relatively tall canopy height was applied to the FSR variable in order to satisfy and high LAI of the Mfor; the L/Bfor was intermediate assumptions of the analysis. All signi®cance tests were in these characteristics (Table 1). Canopy gap fraction carried out at the ␣ϭ0.05 level. Analyses were con- was four and 6.5 times higher at the L/Bfor and Rfor than ducted using SyStat software (SPSS, Chicago, Illinois, for the Mfor (Table 1). Standing aboveground biomass

USA). increased dramatically from the regrowth to the Mfor;

again, the L/Bfor was intermediate (Table 1). Small di- RESULTS ameter lianas were abundant at Rfor and L/Bfor, although Rainfall and understory micrometeorology they were not inventoried. Large diameter coarse woody debris and standing dead trees were most abun- The study took place during the latter part of the dant at the L/Bfor. Litter mass was similar across all 2002 dry season (Fig. 1). When the experimental ®re sites (Table 1). campaigns began on 24 September, the region was in the midst of a severe seasonal drought. A total of 27.3 Litter moisture dynamics mm of rainfall was recorded over the preceding 54-d period at the nearby reference site, with an average of Average mid-day litter moisture contents (LMC) ϳ0.5 mm/d, compared to 108.8 Ϯ 20.3 mm during the were relatively low at all three sites during the extended equivalent period over the previous three years. The dry periods between rainfall events (Fig. 2). Excluding number of rainless days preceding the experimental ®re the sampling dates that fell within a week of substantial campaigns ranged from 2 to 24 (mean ϭ 10 d, median rains (25 October, 5, 12, and 13 November), average ϭ 7 d). One especially rainy period occurred during mid-day LMC during the dry season remained close to the study, extending over ®ve days between 5 and 9 14% at Rfor (range of 5±34%), 16% at L/Bfor (range of November; total inputs ranged from 54 mm at the ma- 8±42%), and 23% (range of 12±40%) at the Mfor. This ture forest site (M ) to 136 mm at the disturbed forests trend became reversed, however, immediately follow- for ing the heavy rains, when higher LMC was recorded (Rfor and L/Bfor) (see Fig. 2). Approximately half of the difference in precipitation between sites (38 mm) was at the disturbed than the Mfor (Fig. 2). This may have associated with a single rain event on 8 November, been result of differences in the size of the rain events affecting each area, and perhaps also due to differences which did not even register at the Mfor. The next to last ®re campaign began three days after these rains ended, in canopy interception among sites. The condition did on 12 November. The amount of rainfall recorded dur- not persist however, because declines in LMC were also ing this period at the nearby reference site was con- more rapid at the disturbed forests. siderably higher (Ͼ30%) than at either of the study Forest canopy controls of LMC and VPD areas. Average mid-day temperatures (T) at the forest ¯oor In combination, canopy height and LAI could ex- were consistently higher at the disturbed forests than plain ϳ60% of the variability in understory VPD across within the Mfor (Fig. 2). The reverse was true for relative the range of conditions represented in this study (Fig. humidity (RH), where the moisture saturation of the 3). The ability of the tall dense canopy of the Mfor to air in the understory of the Mfor remained much higher buffer microclimatic conditions in the forest understory than at either of the disturbed forests (Fig. 2). Vapor contrasts sharply with that of the disturbed forests. The pressure de®cit (VPD), which integrates the variation strength of the relationship between understory VPD, in evaporative demand of the atmosphere associated canopy height, and LAI was increased slightly (by 5%, with both T and RH, was consistently higher at the R2 ϭ 0.65) when the weighted precipitation variable disturbed forests than at the Mfor (Fig. 2). Under the was included. October 2005 RAIN FOREST FIRE SUSCEPTIBILITY 1669

FIG. 2. Summary of micrometeorological measurements taken in the three forest types and in an open area during the period when experimental ®res were being conducted. Panels show mean and standard error of midday (13:00±15:00) temperature (T), relative humidity (RH), and vapor pressure de®cit (VPD) of the forest understory at the regrowth (triangles), logged/burned (circles), and mature (squares) forest sites, and at 1.3 m height in the open area (line and crosses), average moisture content of the leaf litter (LMC) at midday, and daily precipitation for the study sites and at a nearby reference location. 1670 DAVID RAY ET AL. Ecological Applications Vol. 15, No. 5

TABLE 1. Structural features of the three forests where experimental ®res were conducted, SantareÂm, ParaÂ, Brazil.

Average Standing live canopy Leaf area index Gap fraction biomass Leaf litter Forest height (m) (m2/m2) (%) (Mg/ha) mass (Mg/ha) Mature 30 6.07 Ϯ 0.07 2.0 Ϯ 0.2 319 4.2 Ϯ 0.2 Logged/burned 20 4.14 Ϯ 0.10 8.2 Ϯ 0.6 165 4.6 Ϯ 0.2 Regrowth 8 3.87 Ϯ 0.14 13.0 Ϯ 1.5 42 4.2 Ϯ 0.2 Note: Biomass estimates for the mature and logged/burned forests are based on stems Ն10 cm dbh, and on stems Ն1 cm dbh in the regrowth forest. Values are mean Ϯ SE.

In contrast to the strong relationship we found be- 2 kPa) at the disturbed forests at times when the weight- tween VPD and canopy height, and LAI, the combi- ed precipitation variable was relatively high (Ͼ7 mm/ nation of average canopy height and LAI were of lim- d). These ®ndings suggest that the understory micro- ited use for predicting understory LMC across sites (R2 climate at the disturbed forests was more tightly cou- ϭ 0.15). The strongest predictor of LMC was short- pled with the micrometeorological conditions above the term precipitation history, summarized by our weighted canopy than at the intact Mfor. Absorption and re¯ection precipitation variable, which, in combination with can- of incoming direct solar radiation by leaves and other opy structure explained 72% of the variability in LMC plant surfaces takes place closer to the ground in the across sites. During extended periods of dry weather disturbed forests, on average, heating the forest un-

LMC declined with increasing VPD at all study sites. derstory air more than in the Mfor. Moreover, the However, this relationship broke down in the disturbed amount of direct solar radiation that reaches the forest forests during times when the leaf litter was still wet ¯oor is also presumably higher in the disturbed forests but drying rapidly following a recent rain event due to given the lower LAI and correspondingly higher can- the reduced buffering capacity of the canopy in those opy gap fraction (Table 1). areas (Fig. 4). Relatively high VPDs, Ͼ1 kPa for ex- In addition to regulating the amount of solar radia- ample, were not associated with LMC Ͼ 20% in the tion making its way to the forest understory, canopy

Mfor. By contrast, LMC between 60% and 80% were structure also strongly in¯uenced the movement of air often associated with much higher VPD (between 1 and within that environment. Our measurements indicate

that the relatively taller and denser canopy of the Mfor,

and to a lesser extent the L/Bfor, maintained a more effective boundary between the above- and below-can-

opy environments than at the short-stature Rfor. Average wind speeds at mid-day most often fell into the medium

category across these areas (Rfor, 58%; L/Bfor, 47%; and

FIG. 3. The relationship between midday vapor pressure de®cit (VPD), average canopy height (HT), and leaf area index (LAI), in the forest understory of the three forest struc- FIG. 4. The relationship between litter moisture content ture types at times when experimental ®res were conducted (LMC) and vapor pressure de®cit (VPD) in the three forest (13:00±16:00). Also plotted is the response surface resulting structure types at times when experimental ®res were con- from a multiple linear regression analysis. Solid symbols in- ducted (13:00±16:00). Symbols containing a cross indicate dicate spreading ®res, and open symbols are dying ®res. VPD measurements made in close proximity to a precipitation ϭ 3.470 Ϫ 0.053(HT) Ϫ 0.207(LAI); adjusted R2 ϭ 0.59, P event (weighted precipitation Ͼ7 mm/d). Solid symbols in- Ͻ 0.01. dicate spreading ®res, and open symbols are dying ®res. October 2005 RAIN FOREST FIRE SUSCEPTIBILITY 1671

Mfor, 54% of total). There was, however, a marked difference among them in regards to the frequency at which high winds were recorded (Rfor, 26%; L/Bfor, 7%; and Mfor, 3% of total). Fire spread rates and ¯ame heights Fire spread rates (FSR) differed signi®cantly among ϭ Ͻ the three areas (ANOVA, F2,35 13.181, P 0.001). On average, FSR was more than three times higher at Ϯ Ϯ the Rfor (0.257 0.034 m/min) than Mfor (0.067 0.019 Ϯ m/min). FSR was intermediate at the L/Bfor (0.185 0.023 m/min), yet still more than two times higher than at the Mfor. According to a Bonferoni means separation, FSR was signi®cantly higher at the disturbed forests Ͻ than at the Mfor (P 0.01); there was no difference ϭ between the Rfor and L/Bfor in this regard (P 0.163). Maximum FSR for individual experimental ®res ranged from 0.93 m/min at Rfor to 0.43 m/min in the Mfor.

Again, these values were intermediate at the L/Bfor (to 0.63 m/min). The average ¯ame heights associated with spreading ®res (®res that needed to be extinguished manually) also differed substantially among these sites (ANOVA, ϭ Ͻ F2,27 10.629, P 0.001). However, unlike for FSR, Ϯ ¯ame heights were higher at the L/Bfor (38.1 3.4 cm), Ϯ ϭ than at the Rfor (27.9 1.5 cm) (Bonferoni P 0.016). Ϯ And, while ¯ame heights measured at the Mfor (19.6 2.8 cm) were considerably lower than those at the L/ Ͻ Bfor (P 0.001), they were not dramatically different ϭ from those at the Rfor (P 0.146). Consistent with expectation, a relatively strong positive correlation was found between ¯ame height and FSR among spreading ®res (R ϭ 0.45, P Ͻ 0.01). In addition, FSR exhibited a weak but signi®cant negative correlation with forest ¯oor mass (R ϭϪ0.20, P ϭ 0.04), however, no relationship was detected between ¯ame height and forest ¯oor mass (R ϭ 0.12, P ϭ 0.44). FIG. 5. The relationship between ®re spread rate (FSR) and (A) vapor pressure de®cit (VPD) (ln[FSR ϩ 1] ϭ 0.018 ϩ 2 ϭ Ͻ Correlates of understory micrometeorology 0.087[VPD]; adjusted R 0.50, P 0.01), and (B) litter moisture content (LMC) in the understory of the three forest and ®re spread structure types when experimental ®res were carried out ϭ ϩ Ϫ Understory VPD was the single strongest predictor (13:00±16:00) (FSR 0.043 0.838 exp[ 0.107(LMC)]; adjusted R2 ϭ 0.45, P Ͻ 0.01). (C) Important thresholds as- of FSR measured in this study. Using a linear regres- sociated with spreading ®res during extended periods of dry sion model, this variable explained 50% of the vari- weather (weighted precipitation variable Ͻ2.5 mm/d). Dashed ability in FSR (Fig. 5a). We detected a break point in horizontal and vertical lines correspond to 23% LMC and the behavior of our experimental ®res at a VPD of 0.75 kPa VPD thresholds, respectively. The crosshatched area ϳ0.75 kPa, below which spreading ®res were strongly represents the intersection of the two environmental thresholds. Solid symbols indicate spreading ®res, and open sym- inhibited (Fig. 5a). For example, 95% (145/152 ®res) bols are dying ®res. of all experimental ®res that required manual extinction took place at VPD above this apparent threshold. There was also a high proportion of ®res that did not require FSR, measured across the three forests, was strongly manual extinction (47%, 70/148 ®res), yet also took related to LMC; 45% of the variability in FSR could place above this threshold. However, of these, 40% (28/ be explained by LMC using a negative exponential 70 ®res) took place in close proximity to a rain event regression model (Fig. 5b). A threshold in ®re suscep- (weighted precipitation variable Ͼ7 mm/d), and there- tibility, corresponding to a LMC of ϳ23%, was iden- fore may have been limited by high fuel moisture. Pre- ti®ed by solving for the tangent line with slope ϭϪ1 dicted FSR was four times higher for ®res taking place in relation to the ®rst derivative of the equation de- at VPD between 0.76 and 3.0 kPa (0.20 m/min) than scribing that relationship. In addition, 93% (141/152 for VPD between 0 and 0.75 kPa (0.05 m/min). ®res) of all ®res requiring manual extinction took place 1672 DAVID RAY ET AL. Ecological Applications Vol. 15, No. 5

TABLE 2. Reduced regression model of ®re spread rate (FSR; Eq. 1) on canopy structure variables (average canopy height and leaf area index [LAI]) and weighted precipitation (mm/ d) developed across the three forest types based on n ϭ 300 experimental ®res.

Parameter Estimate SE tP Constant 0.361 0.016 21.977 Ͻ0.001 Weighted precipitation (mm/d) Ϫ0.009 0.001 Ϫ8.134 Ͻ0.001 Canopy height (m) Ϫ0.005 0.001 Ϫ7.077 Ͻ0.001 LAI Ϫ0.021 0.004 Ϫ5.226 Ͻ0.001

at LMC below that apparent threshold. Similar to what ing ®res was closer to neutral, with 57% (90/159) we found for VPD, there were also a substantial number spreading. of ®res that did not require manual extinction (49%, Models of ®re behavior 73/148 ®res), yet took place when LMC was below the 23% threshold. Unlike what we found for VPD though, The most parsimonious description of FSR was ar- only a very small proportion of these ®res were as- rived at by combining information about understory sociated with proximate rain events (3%, 2/59 ®res), VPD, mm/d, wind speed, and LAI. This model param- as low LMC was seldom associated with high values eterization accounted for 57% of the variability in FSR of the weighted precipitation variable. However, a rel- measured across the three areas. The fact that VPD atively high proportion (32%, 19/59 ®res) of these self- alone could explain 50% of the variation in FSR (Fig. extinguishing ®res took place when VPD was below 5a) suggests the other variables made modest, yet sig- the 0.75 kPa threshold (Fig. 5b). Predicted FSR was ni®cant (P Ͻ 0.001 for all), contributions to the de- ®ve times higher for LMC between 5% and 23% (0.26 velopment of the full model. The in¯uence of ``forest m/min) than for LMC between 24% and 100% (0.05 type'' was negligible (P ϭ 0.803) after variability in m/min). understory microclimate and canopy structure had been Further insights were gained by examining the joint accounted for. The highest degree of correlation ob- in¯uence of VPD and LMC on FSR during periods of served between any two independent variables was R very dry weather (Fig. 5c). For this purpose we omitted ϭ 0.64 for VPD and LAI, re¯ecting the link between the experimental ®res that took place in close proximity canopy density and understory microclimate. Interpre- to a substantial rain event (weighted precipitation var- tation of the independent variable effects was intuitive; iable Ͻ2.5 mm/d), thereby retaining 78% of the original FSR was positively correlated with VPD and wind sample (234/300 ®res). This had the immediate effect speed, and negatively correlated with mm/d and LAI. of removing most of the high-VPD±high-LMC com- Because wind speed entered the model as a categorical binations observed in the disturbed forests (Fig. 4). variable the full solution consists of three subtly dif- Over 90% (130/143 ®res) of the spreading ®res that ferent regressions, each scaled to the appropriate level took place under these conditions conformed to both of the wind speed variable. When the analysis was of the proposed ®re susceptibility thresholds. The restricted to independent variables representing forest spreading ®res that took place outside of this zone (rep- structure (average canopy height and LAI), and short- 2 resented by crosshatching in the bottom right, Fig. 5c), term precipitation history, model R declined from 0.57 were either associated with high VPD (Ͼ0.75 kPa) or to 0.46. Even so, this approach was still able to capture low LMC (Ͻ23%); in no instance did a spreading ®re 84% of the variability accounted for in the full model exceed both of these thresholds. On the other hand, using fewer more easily obtained variables (Eq. 1, Ta- 44% (40/91) of the ®res that failed to spread shared ble 2): these same characteristics. Nonspreading ®res may ln(FSR ϩ 1) ϭ 0.361 Ϫ 0.009(precip.) Ϫ 0.021(LAI) have been limited by factors other than fuel moisture Ϫ or air VPD. 0.005(HT) (1) Wind speeds at the time the experimental ®res were where HT is average canopy height (m) and all other conducted in¯uenced ®re behavior, as indicated by the variables are as de®ned previously. The weighted pre- positive association between spreading ®res and in- cipitation variable turned out to be the most important creasing wind speeds. For example, over 70% (26/36 predictor of FSR in the reduced model; average canopy ®res) of all ®res associated with the high wind speed height and LAI both contributed signi®cantly (Table category required manual extinction; most of these 2). The fact that canopy structure variables had proven took place at the Rfor. By contrast, there were approx- such strong predictors of understory VPD (Fig. 3), imately twice as many self-extinguishing as spreading which was tightly linked to FSR (Fig. 5a), suggested ®res associated with the low wind speed category (69 they would also be useful for predicting FSR. Average vs. 36 ®res, respectively). The association between the canopy height and LAI were positively correlated (R intermediate wind speed category and spreading or dy- ϭ 0.60). October 2005 RAIN FOREST FIRE SUSCEPTIBILITY 1673

TABLE 3. Reduced logistic regression model of understory ®re susceptibility (pFire; Eq. 2) on canopy structure variables (average canopy height and leaf area index [LAI]) and weighted precipitation developed across the three forest types.

Parameter Estimate SE tPOdds ratio Upper Lower Constant 5.235 0.656 7.974 Ͻ0.001 Weighted precipitation (mm/d) Ϫ0.300 0.050 Ϫ6.000 Ͻ0.001 0.741 0.817 0.671 Canopy height (m) Ϫ0.135 0.023 Ϫ5.837 Ͻ0.001 0.874 0.914 0.835 LAI Ϫ0.360 0.139 Ϫ2.590 0.01 0.698 0.916 0.531 Note: The model was highly signi®cant (chi-square P Ͻ 0.001; McFadden's rho-squared ϭ 0.36) based on information from 152 spreading and 148 dying ®res.

Models of ®re risk signi®cantly to the model's ability to distinguish be- We assessed ®re riskÐthe probability of a ®re tween spreading and dying ®res. The three-variable model spreading under a given set of conditionsÐusing lo- (Eq. 2, Table 3) successfully predicted the outcome of 73% gistic regression. This approach treats the experimental of the experimental ®res, only a slight (5%) decrease in ®res as a binary variable (spreading/dying), and pro- performance relative to the full model: vides information about the conditions under which the pFire ϭ 1 Ϫ [1/{1 ϩ exp[5.235 Ϫ 0.300(precip.) forests are likely to burn. The binary status of each ®re Ϫ Ϫ was assigned conservatively (see Methods). 0.135(HT) 0.360(LAI)]}] The full logistic model, arrived at through stepwise (2) selection, included VPD (P ϭ 0.002), LMC (P Ͻ where pFire is the probability of a spreading ®re and 0.001), litter mass (P ϭ 0.001), and average canopy the other variables are as de®ned previously. height (P ϭ 0.013). Of these, litter mass, LMC, and average canopy height were not included in the linear Predicting susceptibility to understory ®re FSR model using FSR, while recent precipitation history, wind speed, and LAI were included in the full The most powerful ecological models capture the linear model, but not the corresponding logistic model. key mechanisms that regulate system behavior using a Although litter mass exhibited relatively weak corre- small number of easily measured or simulated vari- spondence with the measured physical characteristics ables. We therefore assessed ®re susceptibility using of the experimental ®res, it turned out to be an im- the reduced logistic model (LAI, average canopy portant predictor of ®re success in the logistic analysis. height, and weighted precipitation). These variables Because forest ¯oor mass provides a surrogate for have the advantage of being estimable using remotely available fuels, this variable can also be thought of as sensed data (canopy structure ϭ average height and describing their continuity, which was a consideration LAI), and the network of met stations distributed across in our classi®cation scheme for spreading ®res. The the Amazon Basin (precipitation), and could be pro- possibility that wind speed could have the effect of jected using stand dynamics and climate models. increasing FSR without necessarily increasing the like- The canopy structure variables in our data set de- lihood that an experimental ®re would be classi®ed as scribed a fairly wide range of conditions: average can- successful or not, provides a possible explanation for opy height ranged from 8 to 30 m, while average LAI its being dropped from the logistic model. For example, extended from ϳ4to6m2/m2 (point measurements of a ®re that burned rapidly but only for a short time, LAI ranged from 0.5 to 7.2). The weighted precipitation perhaps due to some unique characteristics of the fuels variable ranged between ϳ0 and 17 mm/d over the in close proximity to the point of ignition, would have period when experimental ®res were conducted. To ex- had a high FSR based on our calculation method (dis- plore ®re susceptibility as a function of rainfall, we tance/time), yet would also have been classi®ed as not divided the canopy structure variables into discrete spreading because it burned out. The canopy height classes and treated short-term precipitation history as and LAI variables contain similar information, and giv- the continuous predictor variable. Average canopy en the high proportion of the total variability already heights were represented by three levels: short, 10 m; accounted for by the inclusion of VPD and LMC, it medium, 20 m; tall, 30 m, while low, medium, and high makes sense that only one of these canopy structure levels of LAI were set at 3, 4.5, and 6, respectively. variables would be included in the model. This four- The weighted precipitation variable was allowed to variable parameterization successfully predicted the range from 0 to 20 mm/d. All of these conditions ap- outcome of 78% of the experimental ®res (McFadden's proximate observed conditions at the study sites and rho-squared ϭ 0.480). during the study period. Of the nine possible combi- A reduced logistic model was ®t to the same canopy nations of canopy variables (three LAI ϫ three canopy structure and precipitation variables used in the second heights), we did not include either the tall-canopy±low- stage of the FSR analysis. Average canopy height, LAI, LAI, or the low-canopy±high-LAI forest structures, be- and the weighted precipitation variable all contributed cause they were not represented in our data, and do 1674 DAVID RAY ET AL. Ecological Applications Vol. 15, No. 5

high levels of the precipitation variable (e.g., 10 mm/d; Fig. 6).

DISCUSSION Rain forest ®res in the Brazilian Amazon are inex- tricably linked to land-use change and associated anthropogenic ignition sources (Nepstad et al. 2001, Alencar et al. 2004). Therefore, to understand the physical and environmental controls on understory ®re behavior our research focused on a mosaic of vegetation types spanning a range of disturbed and intact forests. Our approach to quantifying these relationships differed in some fundamental ways from those taken by past workers. To our knowledge the present study is the ®rst to correlate measurements of forest structure FIG. 6. The probability of an understory ®re not spreading in association with different canopy structures and recent pre- and understory microclimate with large numbers of ex- cipitation history during the dry season. Canopy structure perimental ®res carried out simultaneously over a wide variables are represented by three levels of average canopy range of forest structures. By using a large systematic height (short canopy [SC] ϭ 10 m, medium canopy [MC] ϭ sampling grid we were able to capture the broadest ϭ 20 m, and tall [TC] 30 m), and three levels of leaf area possible range of conditions. Thus, instead of charac- index (low LAI [LL] ϭ 3, medium LAI [ML] ϭ 4.5, and high LAI [HL] ϭ 6). Each regression relates a combination of the terizing conditions at the harvested site (L/Bfor)onthe two canopy structure variables to the weighted precipitation basis of measurements taken only within canopy gaps, variable (e.g., MC-LL refers to the medium average canopy as others have done (Uhl and Kaufmann 1990), we height and low leaf area combination). The horizontal dashed focused on the disturbed forest matrix as a whole, pro- line indicates the break point in the probability of a ®re spreading or dying (pFire ϭ 0.5). viding a more objective representation of the effect such forest structures have on ®re susceptibility. This is an important distinction because conventional har- not re¯ect likely combinations of these variables across vests in the region typically leave Ͼ50% of the forest the landscape. canopy intact (Uhl and Vieira 1989, Asner et al. 2004). If we de®ne the threshold of forest ®re susceptibility The ability to predict ®re behavior was greatly im- as the conditions under which the probability of ®re proved by viewing the studied forests as a continuum success is 0.5 (from the logistic model, scale 0±1), then of forest height and canopy density (Fig. 5a, b). Only the tall-canopied ecosystems remain resistant to ®re by taking this approach were we able to demonstrate even during periods of extremely low precipitation a strong connection between the measured independent (Fig. 6). Also, the tall average height and medium LAI variables and ®re behavior. This ®nding may indicate forest structure combination (TC-ML) just barely con- either that the range of conditions presented at each tacts the risk threshold at the minimum level of the study site was not broad enough to account for the precipitation variable (Fig. 6). The medium canopy difference in the behavior of the individual experi- height high LAI combination (MC-HL) was susceptible mental ®res, or that larger scale structural features (i.e., to ®re at the lowest levels of the precipitation variable. differences in canopy height and the abundance of can- By contrast, canopies of medium or short average opy gaps) at each site had an overriding in¯uence. We height having either medium or low LAI became sus- found some evidence in support of the latter hypoth- ceptible at increasingly higher values of the precipi- eses; for example, over 85% (26/30 ®res) of the ®res tation variable (Fig. 6). Consistent with the selection that took place under dry conditions (weighted precip- of average height as the single descriptor of canopy itation Ͻ2.5mm/d) but at points on the sample grid structure in the full logistic model, it emerges as the with relatively high LAI (Ͼ4.5) were observed to most important factor conferring ®re resistance (Table spread in the Rfor, yet only one of ®ve ®res behaved

3). Preliminary evidence of this is provided by the fact similarly at the Mfor under those same conditions. There that taller canopies with reduced LAI still appear more was only a slight tendency for experimental ®res to be resistant than the next shorter canopy category com- more likely to spread at the L/Bfor when LAI was below bined with the next higher LAI category (e.g., TC-ML 4.5 (77%, 40/52 ®res) than when they were above this Ͼ MC-HL and MC-LL Ͼ LC-ML, where Ͼ indicates level (68%, 17/25 ®res). So, while the incidence of more ®re resistance) (Fig. 6). Even so, recent precip- spreading ®res did increase with the overall level of itation history is the ®nal arbiter of ®re resistance for canopy disturbance, they did not exhibit the tight cor- these disturbed forests; soon after a soaking rain the respondence with point level estimates of canopy den- probability of a ®re spreading is very low, and largely sity that might be expected if those conditions were independent of forest structure or ignition source. All the sole drivers of ®re susceptibility. Unfortunately, the modeled forest types exhibited high ®re resistance at small numbers of experimental ®res that were con- October 2005 RAIN FOREST FIRE SUSCEPTIBILITY 1675 ducted in the other studies of this type did not span a resemble the external conditions, allowing fuels to dry wide enough range of conditions to provide an inde- more rapidly, and to lower moisture contents. These pendent test of this assertion. conditions correspond to a higher probability of ®re Our ®ndings provide additional evidence that rates occurrence (Fig. 6), and, more intense ®re behavior of moisture loss from understory leaf litter are greatly (greater spread rates and ¯ame heights). Conversely, accelerated in disturbed forests (Uhl et al. 1988, Uhl as recovering forests increase in height and canopy and Kauffman 1990), and for vegetation having shorter, density, their resistance to ®re is expected to increase. less dense canopies (Uhl et al. 1988), than for intact The establishment of robust ®re susceptibility thresh- mature forests. Uhl and Kauffman (1990) monitored olds requires a large number of observations relating leaf litter moisture within four different ecosystem the key variables and carried out over a broad range types in the eastern Amazon (mature forest, logged of conditions. Based on observations from two loca- forest gaps, regrowing forest, and pasture) over a pe- tions in Paragominas, Uhl and Kaufmann (1990) and riod of 14 days following a 2-cm rain event. During Holdsworth and Uhl (1997), estimated that undisturbed that time LMC only fell below 15% (their ¯ammability moist tropical forest vegetation remained resistant to threshold) at the pasture site, and within canopy gaps ®re for at least two weeks following 2- and 1-cm rain at a logged forest. By contrast, seventeen days follow- events, respectively, based on their 15% litter moisture ing a substantially larger rainfall input, our disturbed threshold. These same circumstances resulted in ®re forests had average mid-day LMC that was equal to susceptibility in as few as four to ®ve days within large

(Rfor), or only slightly above (L/Bfor) the 15% level. A canopy gaps, while regrowing forest was still consid- possible explanation for this discrepancy lies in the fact ered resistant for up to 14 days. By contrast, their re- that they took their measurements at the beginning of growth forest would have been considered ¯ammable the dry season, when ambient RH is generally higher within approximately one week of the 2-cm rain if this and would act to limit drying, and when LAI may have threshold were expanded to the ϳ23% LMC level de- been higher because of greater soil moisture (Nepstad tected in this study. Another factor complicating the et al. 1994. establishment of any absolute ®re susceptibility thresh- Aspects of ®re behavior documented in this study olds is the possibility that they may vary as a function are consistent with limited observations reported in the of ambient climatic conditions. For example, if the literature, namely that ®re movement within the leaf background RH is relatively high it will take a longer litter layer is typically slow, and associated ¯ame period of time for fuels to arrive at a moisture content heights are rather short. Cochrane et al. (1999) reported where combustion is possible, assuming the RH drops spread rates of between 0.25 and 0.52 m/min for nat- low enough for this to happen. Likewise, it is plausible urally occurring understory ®res they observed in the that sustained combustion can take place at higher eastern Amazon in forests that had suffered one to three LMC if VPD is also relatively high. Some evidence previous burns, respectively. Experimental ®res con- supporting this contention is provided by the obser- ducted in the moister Venezuelan Amazon exhibited vation that all successful ®res that took place at LMC average spread rates of between 0.17 Ϯ 0.04 and 0.24 Ͼ23% were also associated with VPD in excess of our Ϯ 0.04 m/min for a regrowing forest (n ϭ 8 ®res) and 0.75-kPa threshold (Fig. 5c). And in fact, Uhl and within treefall gaps in upland rain forests (n ϭ 7 ®res), Kaufmann (1990) established their LMC threshold ear- respectively (Uhl et al. 1988); spreading ®res at the ly in the dry season (June), while our work was con- disturbed forests in this study (Rfor, L/Bfor) averaged ducted later in the dry season (September±November). 0.27 Ϯ 0.01 m/min (n ϭ 137). Further, those authors While our study areas do represent a wide range of were unsuccessful at starting ®res under undisturbed forest structure conditions, they do not incorporate the rain forest conditions, which is largely consistent with full spectrum of conditions present across the land- our experience. Reports of ¯ame heights on the order scape. However, by selecting two approximate end- of tens of centimeters were attributed to initial ®res points (Rfor and Mfor), and an intermediate case (L/Bfor), moving through leaf litter in the logged forest under- we were able to span a considerable range. The re- story (Cochrane 2003), and similarly as 21 Ϯ 5to24 sulting ®re susceptibility thresholds will require vali- Ϯ 4 cm tall for experimental ®res in regrowing and dation across a broader range of forest structures. Find- upland rain forest treefall gaps, respectively (Uhl et al. ings from the present work indicate that the moisture 1988). Flame heights associated with spreading ®res at content of ®ne fuels can vary widely between nearby the disturbed forests were somewhat higher in this areas as a function of forest structural controls over the ϭ Ϯ ϭ Ϯ study (Rfor 29 2; L/Bfor 38 4 cm). understory microclimate as modi®ed by recent precip- Such ®ndings support the notion that the tall, dense itation history (Fig. 4). The ability to discriminate canopies of intact forests act to buffer the understory among areas on this basis is of critical importance for microclimate from the high temperatures and vapor quantifying ®re risk within this structurally diverse pressure de®cits acting on the forest canopy. As this landscape. Other attempts to forecast ®re susceptibility buffering capacity is lost through disturbance (or severe in the Amazon have relied on historical relationships drought), however, understory microclimate begins to between landscape features and ®re occurrence ob- 1676 DAVID RAY ET AL. Ecological Applications Vol. 15, No. 5 tained from satellite images of anthropogenic land- underestimate the ¯ammability of forest edges, since scapes and interviews of landholders (Alencar et al. our measurements were concentrated in forest interiors. 2004). This approach is useful for identifying the key Climatic factors provide another key input for reli- features of land use most commonly associated with ably predicting forest ®re behavior across large areas forest ®res in the moist tropics, yet is not guided by like the Brazilian Amazon. Topographical variation, any underlying mechanisms per se. Cochrane and which represents a signi®cant consideration in moun- Schulze (1999) derived susceptibility estimates for tainous terrain, is modest within this landscape, and is their study sites using LMC thresholds from Uhl and therefore expected to have limited bearing on ®re be- Kaufmann, (1990) as they relate to canopy openness havior here. In regions like this where precipitation is (Holdsworth and Uhl 1997). The water balance-leaf strongly seasonal, conditions favorable to ®re propa- area approach employed by Nepstad et al. (2004) is gation are concentrated during the dry season, when best suited to assessing vulnerability of mature forest, soaking rains are much less frequent, cloud cover is and therefore may overestimate resistance of disturbed reduced, and the moisture content of the air is lower. forests. And other ®re models are derived from maps Uhl et al. (1988) encountered conditions that allowed of ®re occurrence determined using thermal thresholds ®re propagation within an aseasonal rain forest envi- measured by satellites (Cardoso et al. 2003), but apply ronment in the Venezuelan Amazon, but this was only primarily to nonforest ®res. The utility of these ap- true for shorter and less dense vegetation types and proaches is largely dependent on the assumption that within canopy gaps in the tall dense-canopied forest. the observed relationships are causal in nature, and Further, the climatic conditions that gave rise to those unlikely to change over time. A mechanistic under- ``windows of opportunity'' were relatively infrequent standing of ®re behavior provides the necessary foun- based on examination of a long-term climate record. dation for building ¯exible models of ®re risk (e.g., Accurate predictions of ®re risk depend on an ap- Deeming et al. 1978). Ultimately, we believe that most propriate characterization of the fuels available for useful models will incorporate the best features of both combustion (Burgan et al. 1998). The present research approaches, because observational information will was focused on one key component of the rain forest still be required to address the critical issues of ignition fuel matrixÐthe spatially contiguous leaf litter layer. source in this study system. A justi®cation for having taken this approach is pro- Although not measured as a part of this study, other vided by the observation that understory ®res need not important characteristics of the ®ne fuel matrix may be intense to cause severe damage to the moist tropical have contributed to the observed variability in the be- forest vegetation because of the thin bark of many trop- havior of these experimental ®res. For example, spe- ical forest trees and the relatively long residence times cies-level differences in the physical structure of the associated with slow moving ground ®res (Uhl and leaf litter (large vs. small leaves, curled vs. ¯at edges) Kaufmann 1990, Cochrane et al. 1999, Barlow et al. in¯uence the amount of air that is able to move across 2003). Coarser fuel categories appear to play an im- the fuel surface, and thus speed or slow the drying portant role in maintaining smoldering ®res during pe- process. Fuel chemical composition may also have a riods when climatic conditions are unfavorable for ac- substantial in¯uence on ¯ammability characteristics of tive ®re spread through the leaf litter (e.g., at night understory fuels (Whelan 1995). An extreme example when RH is higher) (Cochrane et al. 1999). The abun- of this is provided by laboratory tests of moisture of dance of coarse fuels in the forest understory increases extinction, de®ned as the moisture content of a fuel with logging intensity (Cochrane and Schulze 1999, beyond which combustion is not sustained, carried out Gerwing 2002), and these may dry out to the point of on the foliage of 24 Mediterranean species that indi- supporting more severe ®res as a consequence of the cated an extremely broad range of between 40±140% associated reductions in canopy cover. LMC (Dimitrakopoulos and Papaioannou 2001). Our The amount of time during the 2002 dry season experimental ®res appeared more intense where the leaf (July±December, 184 days) that each of the critical pre- litter of certain species was abundant (e.g., Breu resina, cipitation levels identi®ed in our logistic model of ®re Tetragastris sp., Burseraceae). risk were present increased exponentially as canopy Further, this study does not directly address the ef- height and LAI declined. So, while the tall dense can- fects of potential edge-related drying on forest suscep- opied forest structures maintained a high level of ®re tibility to ®re. Warm, relatively dry air that moves into resistance throughout the dry season, the shorter less forest edges from neighboring agricultural lands in- dense canopies were vulnerable most of the time. For crease understory VPD and, hence, fuel moisture dry- example, a 20 m tall forest canopy having LAI ϭ 4.5 ing. However, forest edges in the Amazon tend to ®ll (MC-ML) was vulnerable during ϳ70% of the dry sea- quickly with regrowing vegetation, which diminishes son, and forests represented by the most disturbed can- this in¯uence (Kapos 1989). Increased mortality among opy structure combination (SC-LL) were vulnerable for large trees in close proximity to edges, by reducing Ͼ80% of the time. Providing a source of ignition is canopy cover, may also act to increase forest suscep- available, a common feature of actively developing tibility to ®re (Laurance 2001). Thus, our results may landscapes, it appears that even modest alterations to October 2005 RAIN FOREST FIRE SUSCEPTIBILITY 1677

more susceptible to ®re than tall forests with high LAI because of the critical role of the canopy in buffering the forest interior from the hot, dry conditions of the external environment. Recent precipitation events, however rare during the dry season, confer ®re resistance to all forest types, but for periods of time that are inversely related to the degree of structural alter- ation. The likelihood of ®re-susceptible forests igniting is highest in landscapes where swidden agriculture and cattle ranching are prevalent (Nepstad et al. 2001). For- est ®res are largely accidental in the Amazon, and pro- grams to prevent them are being tested. By quantifying the important relationships between forest structure and understory microclimate this study provides a basis for improving ®re susceptibility modeling in the Amazon Basin.

ACKNOWLEDGMENTS We are grateful to the Instituto de Pesquisa Ambiental da Amazonia (IPAM) for institutional support, in particular the ®eld crew at the ``Seca Floresta'' study site, and to the stu- dents from our annual ®eld course in Amazon Ecology (2002) for assistance in carrying out the experimental ®res; to Os- waldo de Carvalho, Jr., Sanae Hayashi, and Ane Alencar for help with data, ®gures, and logistics; and to The Woods Hole Research Center for institutional support. The comments of two anonymous reviewers helped improve an earlier version of this manuscript. Funding for this research was provided by NASA's LBA-ECO program (CD-05 and LC-14), the Gor- don and Betty Moore Foundation, NSF (DEB-0213011), the FIG. 7. Mature forest canopies exert strong control over Ford Foundation, and a CAPES fellowship. understory microclimate, conferring ®re resistance in seasonally dry rain forest environments. When these controls LITERATURE CITED (oval-shaped boxes) are compromised by external factors Alencar, A., L. SoloÂrzano, and D. Nepstad. 2004. Modeling (rectangular-shaped boxes) like extreme drought (e.g., El forest understory ®re in an eastern Amazon landscape. Eco- NinÄo cycles), selective logging, and ®re, and perhaps through logical Applications 14:S139±S149. forest fragmentation, this capacity becomes diminished. Dis- Asner, G. P., M. Keller, R. Periera, J. C. Zweede, and J. N. turbed canopies allow conditions in the forest understory to M. Silva. 2004. Canopy damage and recovery after selec- become coupled with the external environment, and the leaf tive logging in Amazonia: ®eld and satellite studies. Eco- litter and air can dry to the point at which sustained com- logical Applications 14:S280±S298. bustion is possible. Once this internal regulation mechanism Barlow, J., and C. A. Peres. 2004. Ecological responses to is lost, short-term precipitation history and ignition sources El NinÂo-induced surface ®res in central Brazilian Ama- become the primary determinants of ®re susceptibility. zonia: management implications for ¯ammable tropical forests. Proceedings of the Royal Society of London B 359(1443):367±380. Barlow, J., C. A. Peres, B. O. Lagan, and T. Haugaasen. 2003. canopy structure can put these forests at high risk of Large tree mortality and the decline of forest biomass fol- burning under moderately severe dry season condi- lowing Amazonian wild®res. Ecology Letters 6:6±8. tions. Burgan, R. E., R. W. Klaver, and J. M. Klaver. 1988. Fuel The ¯ammability of forest understories in the east- models and ®re potential from satellite and surface obser- central Amazon is regulated by long- and short-term vations. International Journal of Wildland Fire 8:159±170. Cardoso, M. F., G. C. Hurtt, B. Moore, C. A. Nobre, and E. rainfall patterns, forest height, and canopy density M. Prins. 2003. Projecting future ®re activity in Amazonia. (LAI), as summarized in Fig. 7. Long-term precipita- Global Change Biology 9:656±669. tion history determines deep soil water content, which Chambers, J. Q., J. D. Santos, R. J. Ribeiro, and N. Higuchi. regulates LAI and, hence, the amount of radiation that 2001. Tree damage, allometric relationships, and aboveground net primary production in central Amazon forest. reaches the forest interior. Seasonal recharge of the Forest Ecology and Management 5348:1±12. deep soil water that these forests depend on to maintain Cochrane, M. A. 2003. Fire science for rainforests. Nature dense canopies during the dry season is prevented dur- 421:913±919. ing periods of severe drought (e.g., ENSO years) (Nep- Cochrane, M. A., A. Alencar, M. D. Schulze, C. M. Souza,Jr., stad et al. 2004). The ¯ammability of disturbed forests D. C. Nepstad, P. A. Lefebvre, and E. A. Davidson. 1999. Positive feedbacks in the ®re dynamic of closed canopy is the dynamic outcome of processes that reduce LAI tropical forests. Science 284:1832±1835. and forest height (logging, ®re) and those that restore Cochrane, M. A., and M. D. Schulze. 1998. Forest ®res in LAI and height (Rfor). Short forests with low LAI are the Brazilian Amazon. Conservation Biology 12:948±950. 1678 DAVID RAY ET AL. Ecological Applications Vol. 15, No. 5

Cochrane, M. A., and M. D. Schulze. 1999. Fire as a recurrent B. Guerrero. 2004. Amazon drought and its implications event in tropical forests of the eastern Amazon: effects on for forest ¯ammability and tree growth: a basin-wide anal- forest structure, biomass, and species composition. Biotro- ysis. Global Change Biology 10(5):704±717. pica 31:2±16. Nepstad, D. C., A. G. Moreira, and A. Alencar. 1999a. Flames Deeming, J. E., R. E. Burgan, and J. D. Cohen. 1978. The in the rain forest: origins, impacts and alternatives to Am- national ®re danger rating systemÐ1978. USDA Forest azonian ®re: chapter 5. The pilot program to conserve the Service. General Technical Report INT-39. Brazilian rain forest. World Bank, BrasõÂlia, Brazil. Dimitrakopoulos, A. P., and K. K. Papaioannou. 2001. Flam- Nepstad, D. C., et al. 2002. The effects of rainfall exclusion mability assessment of Mediterranean forest fuels. Fire on canopy processes and biogeochemistry of an Amazon Technology 37:143±152. forest. Journal of Geophysical Research 107(53):1±18. Fearnside, P. M., and W. M. GuimaraÄes. 1996. Carbon uptake Nepstad, D. C., A. VerõÂssimo, A. Alencar, C. A. Nobre, E. by secondary forests in Brazilian Amazonia. Forest Ecol- Lima, P. A. Lefebvre, P.Schlesinger, C. Potter, P.Moutinho, ogy and Management 80:35±46. E. Mendoza, M. A. Cochrane, and V. Brooks. 1999b. Gerwing, J. J. 2002. Degradation of forests through logging Large-scale impoverishment of Amazonian forests by log- and ®re in the eastern Brazilian Amazon. Forest Ecology ging and ®re. Nature 398:505±508. and Management 157:131±141. Nobre, C. A., P. J. Sellers, and J. Shukla. 1991. Amazonian Holdsworth, A. R., and C. Uhl. 1997. Fire in Amazonian deforestation and regional climate change. Journal of Cli- selectively logged rain forest and the potential for ®re re- mate 4:957±988. duction. Ecological Applications 7:713±725. Page, S. E., F. Siegert, J. O. Rieley, H. D. V. Boehm, A. Jaya, Houghton, R. A., D. L. Skole, C. A. Nobre, J. L. Hackler, K. and S. Limin. 2002. The amount of carbon released during T. Lawrence, and W. H. Chomentowski. 2000. Annual ¯ux- peat and forest ®res in Indonesia during 1997. Nature 420: es of carbon from deforestation and regrowth in the Bra- 61±65. zilian Amazon. Nature 403:301±304. Peres, C. A., J. Barlow, and T. Haugaasen. 2003. Vertebrate Jipp, P., D. C. Nepstad, K. Cassel, and C. R. de Carvalho. responses to surface ®res in a central Amazonian forest. 1998. Deep soil moisture storage and transpiration in for- Oryx 37:97±109. ests and pastures of seasonally-dry Amazonia. Climatic Pyne, S. J., P. L. Andrews, and R. D. Laven. 1996. Intro- Change 39:395±412. duction to wildland ®re: ®re management in the United Kapos, V. 1989. Effects of isolation on the water status of States. Wiley and Sons, New York, New York, USA. forest patches in the Brazilian Amazon. Journal of Tropical Sanford, R. L., J. Saldarriaga, K. Clark, C. Uhl, and R. Her- Ecology 5:173±185. rera. 1985. Amazon rain-forest ®res. Science 227:53±55. Kinnaird, M. F., and T. G. O'Brien. 1998. Ecological effects Siegert, F., G. Ruecker, A. Hinrichs, and A. A. Hoffmann. of wild®re on lowland rainforest in Sumatra. Conservation 2001. Increased damage from ®res in logged forests during Biology 12:954±956. droughts caused by El NinÄo. Nature 414:437±440. Laurance, W. F., T. E. Lovejoy, H. L. Vasconcelos, E. M. Silva Dias, M. A. F., et al. 2002. Cloud and rain processes Bruna, R. K. Didham, P. C. Stouffer, C. Gascon, R. O. in biosphere-atmosphere interaction context in the Amazon Bierregaard, S. G. Laurance, and E. Sampaio. 2001. Eco- region. Journal of Geophysical Research 107:8072. system decay of Amazonian forest fragments: a 22-year Silver, W. L., J. Neff, M. McGroddy, E. Veldkamp, M. Keller, investigation. Conservation Biology 16(3):605±618. and R. Cosme. 2000. Effects of soil texture on below- Meggers, B. J. 1994. Archeological evidence for the impact ground carbon and nutrient storage in a lowland Amazonian of Mega-NinÄo events of Amazonia during the past two forest ecosystem. Ecosystems 3:193±209. millennia. Climatic Change 28:321±338. Timmermann, A., J. Oberhuber, A. Bacher, M. Esch, M. Latif, MendoncËa, M. J. C. de, M. del C. V. Diaz, D. C. Nepstad, and E. Roeckner. 1999. Increased El NinÄo frequency in a climate model forced by future greenhouse warming. Na- R. Seroa, A. Alencar, J. C. Gomes, and R. A. Ortiz. 2004. ture 395:694±697. The economic cost of the use of ®re in the Brazilian Am- Trenberth, K. E., and T. J. Hoar. 1997. El NinÄo and climate azon. Ecological Economics 49:89±105. change. Geophysical Research Letters 24:3057±3060. Nelson, B. W., R. C. G. Mesquita, J. L. G. Pereira, S. Souza, Uhl, C., and J. B. Kauffman. 1990. Deforestation, ®re sus- and L. B. e Couto. 1999. Allometric regressions for im- ceptibility and potential tree responses to ®re in the eastern proved estimate of secondary forest biomass in the central Amazon. Ecology 71:437±449. Amazon. Forest Ecology and Management 117:149±167. Uhl, C., J. B. Kauffman, and D. L. Cummings. 1988. Fire Nepstad, D. C., G. O. Carvalho, A. C. Barros, A. Alencar, J. in the Venezuelan Amazon 2: environmental conditions P. Capobianco, J. Bishop, P. Moutinho, P. A. Lefebvre, U. necessary for forest ®res in the evergreen rainforest of Ven- L. Silva, and E. Prins. 2001. Road paving, ®re regime ezuela. Oikos 53:176±184. feedbacks, and the future of Amazon forests. Forest Ecol- Uhl, C., and I. C. G. Vieira. 1989. Ecological impacts of ogy and Management 154:395±407. selective logging in the Brazilian Amazon. Biotropica 21: Nepstad, D. C., C. J. R. de Carvalho, E. A. Davidson, P. Jipp, 98±106. P. A. Lefebvre, G. H. Negreiros, E. D. da Silva, T. A. Stone, Whelan, R. J. 1995. The ecology of ®re. Cambridge Uni- S. E. Trumbore, and S. Vieira. 1994. The role of deep roots versity Press, Cambridge, UK. in the hydrological and carbon cycles of Amazonian forests White, A., M. G. R. Cannell, and A. D. Friend. 1999. Climate and pastures. Nature 372:666±669. change impacts on ecosystems in the terrestrial carbon sink: Nepstad, D. C., P. Jipp, P. Moutinho, G. H. Negreiros, and a new assessment. Global Environmental Change 9:S21± S. Vieira. 1995. Forest recovery following pasture aban- S30. donment in Amazonia: Canopy seasonality, ®re resistance Woods, P. 1989. Effects of logging, drought and ®re on struc- and ants. Pages 333±349 in D. Rapport, editor. Evaluating ture and composition of tropical forests in Sabah, Malaysia. and monitoring the health of large-scale ecosystems. Biotropica 21:290±298. Springer-Verlag, New York, New York, USA. Zarin, D. J., M. J. Ducey, J. M. Tucker, and W. A. Salas. Nepstad, D. C., P. Lefebvre, U. L. Silva,Jr., J. Tomasella, P. 2001. Potential biomass accumulation in Amazonian Re- Schlesinger, L. Solorzano, P. Moutinho, D. G. Ray, and J. growth Forests. Ecosystems 4:658±668.