CO2 and fire influence tropical ecosystem stability in response to climate change

Interactions between climate, fire and CO2 are believed to play a crucial role in controlling the distributions of tropical woodlands and savannas, but our understanding of these processes is limited by the paucity of data from undisturbed tropical ecosystems. Here we use a 28,000-year integrated record of vegetation, climate and fire from West Africa to examine the role of these interactions on tropical ecosystem stability. We find that increased aridity between 28–15 kyr B.P. led to the widespread expansion of tropical grasslands, but that frequent fires and low CO2 played a crucial role in stabilizing these ecosystems, even as humidity changed. This resulted in an unstable ecosystem state, which transitioned abruptly from grassland to woodlands as gradual changes in CO2 and fire shifted the balance in favor of woody plants. Since then, high atmospheric CO2 has stabilized tropical forests by promoting woody plant growth, despite increased aridity. Our results indicate that the interactions between climate, CO2 and fire can make tropical ecosystems more resilient to change, but that these systems are dynamically unstable and potentially susceptible to abrupt shifts between woodland and grassland dominated states in the future.

modern-day savanna-forest boundary (Fig. 1). We reconstruct changes in vegetation and hydrology using the carbon and hydrogen isotope composition of sedimentary leaf wax n-alkanes (δ 13 C wax , δ D wax ). Changes in δ 13 C are interpreted as reflecting changes in the relative proportions of plants using the C 3 and C 4 photosynthetic pathways 13 . C 3 vegetation includes most woody plants and winter season grasses with average δ 13 C values that are as low as − 36‰, whereas most summer or dry season C 4 grasses have δ 13 C values of around − 20‰ 14 . We simultaneously reconstruct changes in precipitation using hydrogen isotope analysis of the same compounds (δ D wax ) 15 , which allows us to assess the relative phasing of vegetation and precipitation changes independent of any potential complications associated with organic matter transport pathways or residence time.
Changes in fire frequency are estimated from layer counts of distinct charcoal laminae within the annual sediment laminations (Fig. 2). Because the charcoal occurs as layers and is composed mostly of relatively intact fragments, we interpret the charcoal layers as reflecting the seasonal transport of charcoal to the lake from dry season fires with little or no time lag. We then use the frequency of charcoal layers as a proxy for annual fire frequency. Reconstructions of fire frequency are supported by fire reconstructions from concentrations of levoglucosan, a byproduct of cellulose combustion, and sediment particulate charcoal counts in pollen slides. Previous studies have shown that particulate charcoal 16 and levoglucosan 17 can provide reliable indicators of paleofire in some sedimentary archives and Lake Bosumtwi is an ideal location for preservation of the fire marker levoglucosan because of its permanent bottom water anoxia and well-preserved organic matter 18 . While all three approaches provide proxies of changes in fire regime, processes such as transport and deposition, fire intensity/temperature and fuel source will affect them differently, leading to differences in fire reconstructions [19][20][21][22][23] (Fig. 3). However, the most significant changes are consistent between the fire proxy indicators, indicating that they do preserve consistent large scale variations in the fire regime (SOM).

Results
Evidence for a nonlinear tropical vegetation response to changes in hydroclimate. The Lake Bosumtwi record shows that the climate, fire and ecological regimes of tropical West Africa changed significantly over the past 28,000 years. During the early part of the record ( Fig. 3A; 15,000-28,000 yr BP), δ D wax values were positive, indicating more arid conditions, fire was more frequent and δ 13 C wax values were higher (− 17.3 to − 20.6‰) indicating that the landscape was dominated by drought-tolerant C 4 summer grasses with only minor contributions from C 3 woody plants that burned almost annually. In comparison with the rest of the record, δ 13 C wax values show little variability, despite significant variations in δ D wax and fire at this time. After ca. 15,100 yr BP (Fig. 3B, blue dashed line), more depleted δ D wax values indicate a gradual return to more humid conditions. These are accompanied by a decline in the frequency of fires, which disappear almost entirely for much of the most humid portion of the Holocene (i.e., 10,000-3,200 yr BP). However, in contrast to the gradual nature of these changes in precipitation and fire, the δ 13 C wax record suggests that the vegetation response was abrupt, with a ~25% change in the proportion of C 3 woody plants in < 150 years. The rapid expansion of woody plants also occurs ca. 300 years after the onset of wetter conditions (at 14,800 yr BP), at the time when fire frequency had declined from nearly annual to one fire every five years (Fig. 3B, green dashed line). Over much of the remainder of the record, the landscape around Lake Bosumtwi was dominated by a variable mix of C 3 and C 4 plants suggestive of an open woodland or forest environment similar to today, though the data also suggest a trend towards an increasing proportion of C 3 plants over the past 14,500 years. Coherent variations in δ 13 C wax and δ D wax , during this period indicate that the dominant control over changes in the relative proportions of C 3 /C 4 plants was precipitation, as expected in the absence of fire disturbance. After ca. 3200 yr BP, δ D wax values suggest that conditions became significantly more arid, and were accompanied by an increase in fire. Nevertheless, fire frequency during the late Holocene was much lower than during the late Pleistocene (27 to 15 ka), and the magnitude of the fire response recorded in the levoglucosan and particulate charcoal records differ, suggesting a distinctly different fire regime. Unlike the earlier part of the record, severe late Holocene aridity (∆ δ D wax = ~17‰) was associated with only minor changes in δ 13 C wax , (∆ δ 13 C wax = ~2-3‰) indicating that woody plant-dominated ecosystems persisted despite these environmental changes.
The changing relationships between δ 13 C wax and δ D wax suggest that reconstructed vegetation changes over the past 28,000 years cannot be explained solely by precipitation but instead reflect the competing influences of climate, fire and atmospheric CO 2 , as suggested by modeling and modern ecosystem studies 24,25 . The invariant nature of the early portion of the δ 13 C wax record, when compared with changes in δ D wax, shows that these grasslands were insensitive to temporarily increased precipitation. The fire proxies also show that at this time, significant landscape-scale fires occurred at annual to interannual timescales. Our evidence for extensive grasslands across tropical Africa at this time with a high frequency of burning is supported by both vegetation synthesis and modeling studies 26,27 and lower resolution fire reconstructions from marine sediment cores, which indicate that glacial conditions were accompanied by substantial export of fire-derived elemental carbon 28 . We hypothesize that these fires were sufficiently frequent and widespread to prevent woody plants from becoming established, as can occur in modern savannah ecosystems 7,8 . Prior to ~15,000 yr BP, this effect would have been exacerbated by low CO 2 levels, which would have slowed the rate of woody plant germination, growth and recovery, making it difficult for C 3 plants to recover from fire disturbance even as humidity increased 4 . A critical role for fire and CO2 in controlling tropical vegetation stability. We propose that the high fire frequency early in the record is also the proximate cause of the abrupt and delayed expansion of C 3 plants at 14,800 yr BP (Fig. 4). The rapid shift in vegetation occurred when precipitation was similar to earlier (> 18,000 yr BP) and later (< 3000 yr BP) periods when no systematic vegetation change happened, demonstrating that vegetation was not responding to changes in precipitation alone (Fig. 3). Instead, the abrupt vegetation change reflects a threshold response to the interactions between gradually declining fire frequency, increasing CO 2 and increasing precipitation. The abruptness of this transition likely reflects a combination of the equilibration of C 3 plant composition with available moisture and a feedback between woody plant expansion and a further reduction in the spread of subsequent grassland fires. Support for a threshold mechanism in controlling the vegetation response is also consistent with the ~300 year lag between the initial changes in δ D wax and δ 13 C wax , and the transition from a system where vegetation is unresponsive to changes in precipitation (> 14,800 yr BP) to one where the dominant control on vegetation is precipitation (14,800-3,000 yr BP) (Fig. 3). Although other abrupt changes in vegetation occur in the record at ca. 11,500 yr BP and 5700 yr BP, these lack the characteristics indicative of a threshold vegetation response ( Supplementary Information Fig. S4). For example, neither show clear evidence for a lag in the timing of the vegetation response to climate, nor is there evidence of changing sensitivity of the vegetation to environmental conditions. Both of these are evident in the shift at 14,800 yr BP.
Following the initial expansion of C 3 plants and for much of the remaining record, the frequency and intensity of fire remained low, CO 2 was high and increasing, and vegetation responded on centennial to millennial timescales to changes in precipitation, with greater proportions of C 3 plants accompanying wetter conditions (Fig. 3). However, this interval is also characterized by an overall trend towards greater C 3 plant abundance and a decrease in the sensitivity of vegetation to changes in precipitation. We hypothesize that this reflects the influence of increasing CO 2 , which favors the growth of C 3 plants relative to C 4 grasses 29 . These effects would also have been enhanced by biogeophysical feedbacks; well-developed forests can act to suppress the growth and colonization of grasses by reducing ground light availability, maintaining higher sub-canopy humidity and keeping temperatures lower, all of which promote the growth of C 3 plants and are capable of maintaining forest stability 30 , even as environmental conditions change.

Discussion
The Lake Bosumtwi record highlights the competing roles of climate, fire and CO 2 on the long-term stability and evolution of grasslands and forests of tropical West Africa. In agreement with many studies of modern tropical grassland ecosystems, frequent fires played a crucial role in excluding woody plants during glacial to late glacial times, even as precipitation and CO 2 increased 4-6,9,10,31 (Fig. 4). However, this also resulted in a state of disequilibrium between climate and vegetation, leading to an abrupt vegetation response to changes in climate. In contrast, dramatically increased aridity and fire in the late Holocene did not cause a substantial increase in drought-tolerant C 4 grasses, suggesting that tropical forests became more resilient to changes in climate and disturbance. We propose that this is a consequence of high CO 2 and biogeophysical feedbacks, which resulted in conditions more favorable for C 3 plant growth and recovery, stabilizing existing forest cover despite the increased aridity (Fig. 4). We hypothesize that, as with the late glacial grasslands, this has resulted in a dynamically unstable climate-vegetation state that is susceptible to rapid transitions between C 3 and C 4 plant-dominated landscapes 32 . Although elevated CO 2 should continue to act as a stabilizing element, at higher CO 2 the advantages for C 3 plants decrease 33 , limiting its ability to stabilize these systems and increasing the potential for an abrupt collapse of tropical forest ecosystems in response to increased aridity. The sensitivity of these ecosystems to slight perturbations in climate today is evident in the Dahomey Gap, just to the east of Lake Bosumtwi, where precipitation is only 200-400 mm yr −1 lower than Bosumtwi but the natural landscape is dominated by grasslands with only isolated patches of forest 34 . Here, vegetation records indicate that small increases in rainfall resulted in wholesale return of forest cover briefly in the late Holocene 35 . In the future, much of West Africa could become more arid, provided that projected increases in precipitation are insufficient to overcome a net drying of soils driven by increased temperature 36 . Such gradual changes in hydroclimate and fire could potentially overwhelm the natural buffering capacity of high CO 2 and allow current climate-vegetation stability to give way to rapid forest-savanna transformation in a region that is already a mosaic of these two land-cover types.

Methods
Study site. Past variations in the climate, vegetation and fire history of West Africa were reconstructed using the sediment geochemical record from Lake Bosumtwi, Ghana. Lake Bosumtwi is a small (~8 km diameter), internally draining lake occupying a meteorite impact crater in southern Ghana (Supplementary Fig. S1). The lake is deep (~75 m depth asl) and sheltering by the crater walls inhibit mixing, leading to stratification and permanently anoxic bottom waters. This limits bioturbation and allows for the preservation of mm-scale laminations that have been demonstrated to be annual 18 . The majority of the catchment is forested except for the flat lying terraces and drainages, which have been converted to agriculture 37 . The regional climate is controlled primarily by the West African monsoon, which brings moisture to southern Ghana in May-September when the Intertropical Convergence Zone (ITCZ) migrates to the north (Supplementary Fig. S1). In winter, as the ITCZ shifts southward, the climate of the region is dominated by dry northwesterly winds, which inhibit rainfall.

Age model development.
A detailed description of the age model used in this study is published elsewhere 38 . It is based on a combination of radiocarbon dating, Bayesian age depth modeling and varve counting. The chronology consists of 107 radiocarbon dates on bulk organic matter and macrofossils covering the upper 21.4 meters of sediment. Dates from different cores were correlated using marker laminae matches and Bayesian age depth modeling was performed using the R software package BACON 39 and using the IntCal09 radiocarbon calibration curve 40 . Over the deglacial section of the core, we revised the age model by wiggle matching the uncalibrated radiocarbon ages to the IntCal09 calibration curve using the varves as a constraint on the time elapsed between radiocarbon age. An optimal fit was determined using Χ 2 minimization 38 . Lipid and compound-specific stable isotope analysis of n-alkanes. Sediment samples (1-5 g) were freeze dried, homogenized and solvent-extracted by either accelerated solvent extraction (Dionex ASE 200) or a microwave extraction (MARS) using dichloromethane:methanol (9:1; v/v). n-alkanes and levoglucosan were isolated from the total lipid extract by silica gel column chromatography. n-alkanes were additionally isolated by silver nitrate silica gel chromatography and molecular sieve or urea adduction. Stable carbon and hydrogen isotope analysis of the long chain n-alkanes was performed by gas chromatography isotope ratio mass spectrometry (GC-IR-MS). Description of the methods for δ D wax analysis and corrections for changes in vegetation and ice volume are described in the online Supplementary Materials. δ 13 C wax values were measured against internal calibrated propane reference gas standards and are reported in % VPDB (Vienna Pee Dee Belemnite). The precision of the analysis, based on repeated analysis of the standard mix was better than 0.3%. Each sample was measured in triplicate and the mean precision for the long chain n-alkanes (C 27 , C 29 , C 31 , C 33 ) based on this replicate analysis was better than 0.4%. Levoglucosan analysis was performed by gas chromatography mass spectrometry. Identification of the trimethylsilyl ether derivative of levoglucosan was identified in the mass scan by its characteristic fragments (m/z 363, 333, 217, 204) and its retention time determined with an authentic standard. Quantification was performed using an internal standard (androstanol). Levoglucosan concentrations (ppm) were normalized to time (ppm/yr) using varve counts over the laminated intervals of the core and using the Bacon age-depth model over the unlaminated sections.
Reconstruction of fire frequency from sedimentary charcoal. Paleofire was reconstructed from a combination of microscopic particulate charcoal counts and frequency analysis of charcoal layers visually identified in sediment thin sections under transmitted light (Supplementary Fig. S2). Particulate charcoal fragments above 10 μ m were counted on slides prepared for palynological analysis. A minimum of 2000 charcoal particles were counted per sample and abundance relative to an exotic marker, Lycopodium, was calculated. Charcoal counts were normalized to time (particles/cm 2 /yr) using the same approach as for levoglucosan. Fire frequency was estimated using visual identification of annual charcoal layers in sediment thin sections under transmitted light. Previous work on the laminations from Lake Bosumtwi has demonstrated that they are annual in nature and visible charcoal occurs as a distinct layer within many of the annual laminations, suggesting that the charcoal is deposited after seasonal fire events, rather than being remobilized material from fires occurring in previous years . Charcoal layer frequency analysis was performed on a single continuous set of thin sections spanning the record. Reproducibility was assessed via replication over a several short, overlapping intervals.