Main

Wildfire currently burns 3.94–5.19 × 106 km2 of land annually1. Fire has always modified the terrestrial carbon cycle, ecosystem function and biodiversity, from the tropics to high latitudes2,3. The incidence and severity of wildfires around the world are increasingly affected by human endeavour, including associated climate change, with catastrophic impacts of severe fires now occurring in many parts of the world4,5,6,7.

Fire is also one of the earliest tools appropriated by humans2, with the earliest deliberate use of fire by hominins dating from ~ 1 million years ago8. Habitual use of fire potentially allowed manipulation of fire frequency and fire seasonality from as early as 250 ka to 350 ka (refs. 9,10). As human populations grew, and developed both agriculture and pastoralism into the Holocene, our species also achieved an increasing measure of control over fuel loads and fuel connectivity, and therefore fire regime—that is, fire incidence and intensity9.

Humans have modified both modern landscapes and patterns of biodiversity that were considered ‘natural’ for millennia before the transformations that have occurred since the Industrial Revolution11,12,13,14. Fire was one of several tools used by pre-industrial societies to modify and manage land and resources9. The cessation of Indigenous fire management in many parts of the world has contributed to biodiversity decline and, through a build-up of connected fuel loads, an increase in catastrophic fires5,14,15,16.

There is abundant evidence for human use and manipulation of their environment extending back at least tens of millennia17; however, separating anthropogenic and climatic impacts on fire regime in prehistory is challenging2,18,19. Clear and diverse human impacts, including those on fire regime, are uncommon in the early Holocene20, but become progressively more common from the mid-Holocene in regions where societies transitioned to agriculture, pastoralism and urbanization21.

Human manipulation of fire regimes in the large areas occupied by low-density hunter-gatherer populations has been difficult to identify2,19. Where detection has been claimed, the evidence is generally derived from the appearance of, or abrupt increase in, the abundance of micro-charcoal particles in sedimentary records—often in conjunction with pollen evidence of the modification of local vegetation. Most studies claim the detection of a hunter-gatherer influence on fire from the mid- to late Holocene in Australia22, southern Africa23 and North America24. It is not yet known whether the many well-documented impacts of Indigenous fire management on many aspects of biodiversity could have been developed over only a few millennia12,19.

Detecting prehistoric human influence on fire regime is especially problematic in tropical savannahs, where fire has always had a short (generally sub-decadal) return interval and vegetation is adapted to fire25. Charcoal is therefore continuously detectable in savannah environmental records regardless of the source of ignition. In Australia, as elsewhere, anthropogenic climate change exacerbated by the cessation of Indigenous fire management is driving a trend towards more extreme fires14,15,16,26. While the catastrophic fires in southern Australia in 2019–2020 received much attention5,6,27, most fires occur in the tropical savannahs of northern Australia (70% of total burnt area 1997–200128). While fire is a natural driver of ecosystem dynamics in savannahs, the cessation of Indigenous land management over much of northern Australia following European arrival11,19,29,30 has changed vegetation structure31,32 and fuel characteristics33,34. This has decreased biodiversity and altered the carbon balance19,29,35,36,37.

Fire has been shaping Australia’s terrestrial ecosystems for millions of years38. Humans arrived on the continent ~65,000 years ago39, radiating across the continent over the next 5,000 to 10,000 years (refs. 40,41). The prevalence of Indigenous use of fire through ‘firestick farming’42 is now thought to have been essential to maintaining the landscapes and biodiversity observed across northern Australia and considered natural at European arrival11,29,30. An early claim of evidence of human impact identified through an abrupt increase in micro-charcoal abundance 40,000 years ago at Lynch’s Crater in north-eastern Australia43 has not been supported by subsequent meta-analyses of compilations of larger sets of charcoal records around the country44,45. However, few of these records are long enough to assess the characteristics of a natural fire regime. We used multiple fire and climate indicators in a continuous savannah sediment record spanning the past 150,000 years to identify whether (and if so, when) Indigenous fire management became a dominant control on fire regime and ecosystem dynamics in the savannahs of northern Australia.

A 150-ka record of climate and vegetation in northern Australia

Girraween lagoon (12° 31′ 3.6″ S, 131° 04′ 50.7″ E, 25 m above sea level; Fig. 1) is a perennial waterbody located on the outskirts of Darwin, Northern Territory, Australia. The climate is monsoonal with a mean annual regional rainfall of 1,731 mm (Bureau of Meteorology Station 014015, Darwin Airport), 90% of which falls between November and April. The modern vegetation around the lagoon is Eucalyptus-dominated tropical open-forest savannah and/or savannah woodland (dominant overstory: Eucalyptus tetrodonta, Eucalyptus miniata and Corymbia polycarpa) with a grassy understorey. The modern environment and evolution of the lagoon since 26.6 ka has been described in detail46,47.

Fig. 1: Location of Girraween Lagoon.
figure 1

The coastline of modern Australia is shown by the solid white line on the landmass at the maximum extent of lower sea level during the Last Glacial Maximum. a, Modern vegetation. b, Satellite image of the site. The other locations are mentioned in the text, and the modern southern limit of the savannah biome is shown as a dashed white line. Map data from Google, Maxar Technologies.

Full size image

The lagoon formed following collapse into a sinkhole and has been permanently wet since that time. During dry periods, the waterbody was restricted to the sinkhole itself (~0.01 km2); during wet periods (as today) the surface area expands across a shallow depression to an area of 0.45 km2 (Fig. 1). The site is 200 km west of the earliest archaeological site of Madjedbebe39 and, as a permanent water source, the 0.92 km2 catchment is likely to have had some human presence for ~65,000 years.

A 19.4 m core was collected using a floating platform with hydraulic coring rig in 4.5 m of water, with 235 samples taken at 5–10 cm intervals. The record is composed of alternating peats that formed during wet periods with high lake levels, and clays deposited during dry periods with low water levels. The chronology for the record is provided by 12 radiocarbon46,47 dates and 24 optically stimulated luminescence dates (Supplementary Table 1).

Peaks in the total organic carbon (TOC) content corresponding to high lake levels occur mainly in Marine Isotope Stages (MIS) 1 and 5, with the largest of these (>15% TOC) relating to insolation maxima in the region (Fig. 2a). These reflect wet monsoonal conditions similar to today. Rainfall at the site is also modulated by the distance to the coast47—the low offshore bathymetric gradient means that the coast retreats by up to 320 km during glacial periods. Smaller peaks in TOC (<10%), indicative of wet conditions, are related to times of higher sea level and/or times of higher insolation (Fig. 2a). From MIS 4 to MIS 2 there was an extended period of uniformly low lake level regardless of insolation (<2% TOC). This period is due at least in part to the distance to the coast being 150–320 km, because rainfall decreases away from coasts in monsoonal climates downwind of the moisture-source region, becoming progressively drier as previous rainout occurs48. It is also due to comparatively low insolation over this period that weakened monsoon intensity49 (Fig. 2a).

Fig. 2: 150 ka record of environmental change at Girraween Lagoon.
figure 2

a, TOC content (right y axis) and the distance to the coast (left y axis) through time. The distance to the coast was calculated using the average of the bathymetry on transects north and northwest from the coast adjacent to Girraween Lagoon to the continental shelf edge and the LS16 sea level curve73. b, PARchar (right y axis) and MARspac (leftt y axis) both on a logarithmic (base 10) scale (Supplementary Fig. 2 shows the two time series on a linear scale). c, Grass pollen as a percentage of total dryland pollen (left y axis) and δ13CSPAC (right y axis). Measured values (dots) do not always correspond exactly to 250 yr interpolated values (Methods). The alternating background colours correspond to approximate MIS durations, with the peaks of MIS 5 substages also identified (a–e)73. Wet-season (December to March) insolation peaks at 15° S occur at 136, 114, 92, 70, 45, 19 and ~0 ka (vertical red dotted lines, 0 ka not shown; see Supplementary Fig. 2 for insolation curves).

Full size image

A 150 ka record of fire regime in northern Australia

We used three measures of fire derived from the Girraween lagoon record: (1) micro-charcoal particle accumulation rate (PARchar; Fig. 2b), the most commonly applied indicator of fire, for which higher accumulation rates indicate a higher incidence of fire44; (2) the mass accumulation rate (MAR) of stable polycyclic aromatic carbon (SPAC) by hydrogen pyrolysis (MARSPAC; Fig. 2b). This is a geochemical measure of the proportion of carbon in condensed (>7) aromatic rings in the char, indicating temperature of formation22,50,51,52. The combination of these two measures allows inference of relative fire intensity, because the proportion of SPAC in char increases as a function of the temperature at which it is produced. Thus, higher-intensity fires will produce char containing abundant SPAC, while lower-intensity fires will produce char with little SPAC22,50,51,52. (3) We also measured the δ13C of the SPAC (δ13CSPAC; Fig. 2c), which provides a relative measure of the proportion of grass-derived (mainly C4) relative to tree/shrub-derived (exclusively C3) biomass that is not combusted, but is instead preserved as SPAC50,51,52. The δ13CSPAC is broadly determined by vegetation, and therefore climate, but is also modified because high-intensity fires more fully combust fine (grassy) fuels33,50, leading to a relative increase in larger woody (C3)-derived biomass preserved in the charcoal and therefore, a lower δ13CSPAC compared with the δ13C of the biomass.

PARchar (Fig. 2b) varies by five orders of magnitude (4–105,400 particles cm−2 yr−1); it was high during periods of wetter climate (defined herein by high TOC and a generally lower grass pollen contribution to total dryland pollen; %C4 pollen) and low during drier intervals (low TOC and high %C4 pollen). The apparently low PARchar in the clay-rich sediments that dominate during dry periods (<2 % TOC) are an artefact of char–clay mineral interactions53,54. These lead to charcoal particles in the clay sediments, particularly fragile grass-derived charcoal, being comminuted to below the minimum size for counting (<10 µm; Methods) during laboratory disaggregation of the sediment, however gentle. This means that PARchar is only a reliable indicator of relative fire incidence during intervals of higher TOC. The mean PARchar through the Holocene is 42,200 particles cm−2 yr−1, more than double the means for the previous two highest multi-millennial peaks in accumulations in the past 150 ka (16,700 and 10,500 particles cm−2 yr−1), both of which occur at times of high insolation near 114 and 92 ka, similar to the Holocene (Fig. 2b and Supplementary Fig. 2).

As a geochemical measure, MARSPAC (Fig. 2b and Supplementary Fig. 2) is insensitive to clay content and is present throughout the record. Accumulation rates range over three orders of magnitude (1–131 mg cm−2 yr−1); rates are most consistently high through interglacial MIS 5 (mean MARSPAC = 30.2 mg cm−2 yr−1) and consistently low in glacial stage MIS 2 (mean MARSPAC = 6.5 mg cm−2 yr−1; Fig. 2b and Supplementary Fig. 2).

Tree–grass balance as a measure of savannah response to changing rainfall, rainfall distribution and potentially Indigenous management (percentage of C4 pollen; Fig. 2c) ranges widely and varies substantially over centennial timeframes. The major wetter periods are generally characterized by 30–50% (minimum 2.8%) C4 pollen, indicating a more wooded savannah. During drier periods, C4 pollen often comprised >80% (maximum 99.2%) of the total dryland pollen, which is suggestive of an open, grassy savannah. The δ13CSPAC (Fig. 2c) also ranged widely from −27.3 to −13.0‰ and exhibited only limited similarity at the broadest scale to the pollen record.

Separating anthropogenic from natural fire regime

The vegetation structure of tropical savannahs is influenced by both fire incidence and intensity, which in turn vary with climate, climate variability, land management and the frequency of ignition31,55. Tree cover responds to changes in fire regime on a decadal timescale32. In Australia’s tropical savannahs, fire regime is influenced by precipitation and precipitation variability56,57 in combination with seasonal temperature changes57. Land management affects grass biomass58, and the flammability of both grass biomass59,60 and litter60,61. Higher fuel loads produce hotter fires34, so the time since the last fire affects intensity62. Fire size and intensity increase in the late dry season63, while early dry-season fires have a lower intensity64.

The attributes of an Indigenous fire regime in northern Australia are well documented. They include the use of frequent burning (1–3 yr return interval) of small areas (hectares to tens of hectares) to maintain low fuel loads and connectivity beginning in, but not confined to, the early dry season30,36,65,66. These are low-intensity fires limited to the ground layer, thereby reducing the risk of large, high-intensity fires while promoting a mosaic of vegetation and attendant biodiversity. This regime produces a low proportion of SPAC in the char created and a relatively high δ13CSPAC, because a larger component of grassy fuel is more likely to be preserved as charcoal instead of being fully combusted.

The characteristics of a natural fire regime in Australia’s tropical savannahs cannot be observed directly today, but can be inferred on the basis of lightning being the only source of ignition, and climate being the dominant control of fire regime. The average modern density of lightning strikes is 1.21 km−2 yr−1, with more lightning striking grasslands than wooded vegetation classes67. Almost no strikes occur in the region from April to September (early dry season), but they peak in November and December during the transition to wet season conditions. With considerably fewer sources of ignition, and successful ignitions likely to result in fires concentrated in the late dry season68, a natural fire regime probably has fewer fires. Those that do occur are higher intensity due to greater, more fully cured fuel loads, and burn larger areas due to greater fuel connectivity. A higher-intensity fire will more completely combust biomass, particularly fine fuels such as grass and litter, thereby producing less char50,69 and a higher proportion of SPAC in that char. A natural fire regime will produce relatively lower δ13CSPAC from the higher preservation potential of coarse, woody fuels relative to fine, grassy fuels—a phenomenon known as the savannah isotope disequilibrium effect50.

We therefore hypothesize that a natural (more-intense) fire regime will be identifiable by: (1) a relatively lower PARchar containing relatively more SPAC than under an anthropogenic fire regime; and (2) a lower δ13CSPAC relative to the δ13C of the biomass as measured by the percentage of grass pollen in the total dryland pollen count. An anthropogenic (less-intense) fire regime will be identifiable by: (1) a relatively higher PARchar containing relatively less SPAC than under a natural fire regime; and (2) a higher δ13CSPAC relative to the δ13C of the biomass as inferred from the percentage of C4 pollen. Under both natural and anthropogenic fire regimes, micro-charcoal is dispersed by wind and water70,71, with grass-derived char being dispersed more widely than char derived from woody materials71. This difference in dispersal potential means that charcoal (and pollen) in terrestrial records such as Girraween, close to the area being burnt, tends to be biased towards larger, less aerodynamic size classes, often wood-derived charcoal. In contrast, marine records remote from sites of burning will be biased towards smaller, more aerodynamic size classes, often grass-derived charcoal71.

The combined records (Fig. 2) can be assessed from the perspective that the record before 65 ka, and for some unknown period after, must reflect a natural fire regime. There are clear qualitative differences. Comparing the glacial stages, for example, the MARSPAC in MIS 2 (post-human arrival) was one-third of that in MIS 6 (pre-human arrival). Comparing interglacial stages, PARchar in MIS 1 (post-human arrival) was more than double that observed at any time in MIS 5 before human arrival (Fig. 2b and Supplementary Fig. 2).

There is considerable intersample variation throughout the record, and it is of course also possible that low-intensity fires always resulted stochastically from lightning ignitions, and that climate extremes might promote periods of intense fires. We therefore tested for statistical dissimilarity in the records (Methods) using 30 ka as the arbitrary time after which there was a transition to an identifiable anthropogenic fire regime. We compared the two measures of fire (PARCHAR and MARSPAC) and the two measures of vegetation (percentage of C4 pollen and δ13CSPAC). To control for the impact of the particle accumulation rate of micro-charcoal in the high clay content in parts of the core, we tested whether the choice of a TOC threshold modified our results (Methods). When TOC is >5.1%, the attributes of the record after 30 ka are different to the natural fire regime in place for the preceding 120 ka and consistent with the establishment of an anthropogenic fire regime at some time after 30 ka (Figs. 3 and 4). The decline in time-series length (sample size n) for the youngest (anthropogenic fire-regime) period does not affect variance in any of the measures until <3 ka (Supplementary Fig. 3), demonstrating that the shift from natural to anthropogenic fire regimes is not an artefact of diminishing statistical power.

Fig. 3: Relationships between fire and vegetation indices at Girraween lagoon.
figure 3

a, Relationship between MARSPAC and PARCHAR. We hypothesized a more pronounced positive relationship under the pre-human natural fire regime than under the lower-intensity fires present in the anthropogenic fire regime. b, Relationship between the percentage of C4 pollen and δ13CSPAC. We hypothesized a more pronounced positive relationship under the lower-intensity fires present in the anthropogenic fire regime than under the pre-human natural fire regime. In both cases, the relationships pertain only to periods where the TOC is >5.1%, and the hypothesized temporal split between the natural and anthropogenic periods is 30 ka. Red dashed lines indicate the best-fit relationships with uncertainty for periods dominated by a natural fire regime, and solid blue lines indicate best-fit relationships with uncertainty during periods dominated by anthropogenic fire regimes. We provide the relationship for all samples in Supplementary Fig. 1.

Full size image
Fig. 4: Histograms of Spearman’s correlation coefficient for fire and vegetation indices at Girraween lagoon.
figure 4

a, Relationship between MARSPAC and PARCHAR resampled for the natural fire regime (>30 ka; red bars) relative to the observed Spearman’s correlation coefficient (ρ) for the anthropogenic fire regime (<30 ka; blue vertical dashed line). b, Relationship between the percentage of C4 pollen and δ13CSPAC resampled for the natural fire regime (>30 ka; red bars) relative to the observed Spearman’s ρ for the anthropogenic fire regime (<30 ka; blue vertical dashed line). The relationships pertain only to samples where the TOC is >5.1%. The probabilities of generating the result randomly are 0.0071 (a) and <0.0002 (b) for this combination of thresholds. The regions shaded grey indicate ρ < 0.05. In both panels, x axes indicate the binned value of ρ for the relationship (indicated by ~) between each variable under the natural fire regime across all iterations. Solid black lines indicate probability density function describing the Gaussian shape of each histogram.

Full size image

Evolving fire regimes into the Holocene

The temporal depth of the multiple proxies in the Girraween record enables us to determine the attributes of a natural fire regime as expressed in the proxies before human arrival. This then allows us to attribute influence on fire regime in the period younger than ~11 ka to human agency with statistical confidence, because the probabilities of a randomly generated difference between the natural versus anthropogenic fire regime using all indicators quickly rise to high values (>0.05) when TOC exceeds 2% (Fig. 5a) and after 11 ka (Fig. 5b). It is unlikely that any anthropogenic influence on fire regime was imposed instantaneously, so some measure of control was probably exercised before 11 ka. The high δ13CSPAC around 15 ka when the percentage of C4 pollen was low, as well as the low MARSPAC in glacial MIS 2 compared with glacial MIS 6, provides qualitative evidence that this was the case. A trend towards increased δ13CSPAC (Fig. 2c) over a period when vegetation was stable that began before ~30 ka could also indicate an early, progressive implementation of Indigenous fire management. However, the clay-rich sediments during these periods make the PARchar unsuitable for direct comparison to MARSPAC.

Fig. 5: Sensitivity of hypothesis to fire and vegetation indices at Girraween lagoon.
figure 5

a,b, Sensitivity to variation in the choice of the TOC threshold (a; temporal split constant at 30 ka) and the temporal split defining the change from natural from anthropogenic fire regimes (TOC threshold constant at >2%). The probability that the relationship measured as Spearman’s ρ between MARSPAC and PARCHAR is greater for the natural fire regime than the anthropogenic fire regime is shown by the green dashed line. The probability that the relationship measured as Spearman’s ρ between the percentage of C4 pollen and δ13CSPAC is greater for the anthropogenic fire regime than the natural fire regime is shown by the solid black line. Both hypothesis test outcomes are insensitive to the choice of either threshold at TOC thresholds >2% and temporal splits older than 10 ka. See the Methods for explanation of our approach.

Full size image

It is possible that anthropogenic influence increased progressively through the post-glacial period of sea-level rise following the end of MIS 2, with small human populations in the interior of Sahul45 in glacial MIS 2 limiting peoples’ ability to affect fire regime. It is clear that as population densities increased in the Holocene45, the savannahs around Girraween Lagoon, and by implication, across northern Australia, were modified from their natural state. This change might have been initiated in post-glacial times by monsoon reactivation, leading to the return of mesic savannahs and higher fuel loads than had been seen in the Girraween Lagoon area since human arrival.

While climate ultimately dominates the variability in fire regime, Indigenous management of fire observed ethnographically across northern Australia was firmly in place from the earliest Holocene. It is plausible that Indigenous fire management was developed over millennia in the wetter coastal regions on the now-flooded Sahul shelf, introduced to the Girraween lagoon region as populations retreated landwards in the face of post-glacial sea-level rise (Fig. 2a). It seems unlikely that the Australian case is unique, and low-density hunter-gatherer populations in other parts of the world might similarly have established control over fire regimes earlier than is currently recognized.

Our results suggest that the mesic savannahs at Girraween Lagoon were established over several millennia under a fire regime influenced by Indigenous fire management from at least the early Holocene. This fire regime was fundamentally different to previous times when mesic savannahs were present at the site during MIS 5. It is therefore unsurprising that the rapid change to a European fire regime, with an increased incidence of larger, more-intense, late dry-season fires reminiscent of the natural fire regime has massively altered biodiversity19,35 and increased greenhouse gas emissions65. Reversing these trends in Australia’s tropical savannahs requires the re-implementation of an Indigenous fire regime30,36, with additional benefits including an increase over time in tree biomass of ~0.11 t dry matter ha−1 (ref. 72). By implication, the reintroduction of Indigenous land-management strategies in other parts of the world could assist in mitigating the impact of catastrophic fires, and increase carbon sequestration in the future14,15.

Methods

Laboratory analyses

Initial handling

We split the core sections in half, and took samples of known volume at 5 or 10 cm intervals (depending on the changing nature of sediments) for pollen and micro-charcoal analysis and geochemistry (see below). We weighed, freeze-dried and reweighed samples to calculate the dry bulk densities.

Pollen and PARCHAR

We processed 2 cm3 sediment samples for pollen and micro-charcoal analysis. Sample preparation followed standard techniques74,75. We selected chemical preparations to disperse sedimentary materials initially, then progressively remove humic acids, calcium carbonates, bulk (in)organics and cellulose and silicates, as well as render pollen ornamentations more visible (including Na4P2O7, KOH, HCL, acetolysis and ethanol washes). We sieved at 125 µm and 7 µm. We added a Lycopodium spike during laboratory preparations to determine the relative concentrations of pollen and micro-charcoal particles.

We did pollen identifications at James Cook University based on regionally appropriate image and slide libraries built from collections at the lagoon and online resources such as the Australasian Pollen and Spore Atlas (apsa.anu.edu.au). Pollen sums averaged 300 grains. We present the proportion of total grass pollen as a percentage of total dryland pollen taxa in each sample, with the 97 taxa included in the dryland pollen sum detailed in Supplementary Table 2. The dryland pollen sum excludes all wetland-associated plant types, including riverine, floodplain, seasonal or permanent swamps, and lagoons of the Northern Territory. We also excluded spores identified as algal or Pteridophyte.

We counted micro-charcoal incorporated within the final pollen concentrate (black, opaque, angular particles >10 µm) simultaneously with pollen. We used the dry bulk density and the sedimentation rate derived from the age model (see below) to convert micro-charcoal particle counts to a PARCHAR.

MARSPAC and δ 13CSPAC

We first decarbonated samples for TOC and SPAC, carbon isotope analysis and carbon isotope composition in 2 M HCl for 3 h. We isolated the SPAC component by hydrogen pyrolysis, now an established method for quantifying pyrogenic carbon76. We mixed approximately 250 mg of the sediment with a Mo catalyst using an aqueous methanol solution of ammonium dioxydithiomolybdate, sonicated and dried at 60 °C overnight. We placed the sample–catalyst mixture in a reactor, pressurized with H2 to 150 bar with a purge gas flow of 5 l min−1 and then heated at 300 °C min−1 to 250 °C, and then at 8 °C min−1 to 550 °C, which was held for 5 min. Labile carbon is removed during the hydrogen pyrolysis reaction, and the remaining carbon is composed of a stable form of pyrogenic carbon with more than seven condensed aromatic rings—this is defined as SPAC76.

We measured carbon abundance and carbon isotope composition by elemental analysis isotope ratio mass spectrometry using a ThermoScientific Flash EA with Smart EA option coupled with a Conflo IV to a Delta VPlus at James Cook University’s Advanced Analytical Centre. We determined carbon abundances using a thermal conductivity detector. We report carbon isotope measurements (in ‰) as deviations from the Vienna Pee Dee Belemnite (VPDB) reference standard scale for δ13C. We used USGS40 and two internal laboratory reference materials (Taipan, Chitin) within each analytical sequence for three-point calibrations of isotope delta-scale anchored by the VPDB. We calibrated our internal standards using USGS40 and USGS41 international reference materials and results have an uncertainty of ±0.2‰ or better. SPAC percentages have an average standard error of 2.2% of the value, which we corrected for possible in situ production of SPAC during the hydrogen pyrolysis reaction77,78. We used the dry bulk density and the sedimentation rate derived from the age model (see below) to convert SPAC abundances to MARSPAC.

Chronology

Radiocarbon

We pre-treated samples of bulk sediment for radiocarbon dating using hydrogen pyrolysis to remove labile carbon and decontaminate the charcoal component79. We combusted samples to CO2 and reduced them to a graphite target for measurement at Australia’s Nuclear Science and Technology Organisation79. Details of all radiocarbon dates used to construct the age model for the Girraween Lagoon record have been published previously46,47.

Optically stimulated luminescence dating

We extended the chronology for the sedimentary sequence at Girraween Lagoon beyond the limit of radiocarbon using optically stimulated luminescence dating. This method gives an estimate of the time since grains of quartz were last exposed to sunlight80. The time of sediment deposition is estimated by dividing the equivalent dose (De, the radiation energy absorbed by grains since deposition) by the environmental dose rate (the rate of supply of ionizing radiation to the grains since deposition). De is estimated from laboratory measurements of the optically stimulated luminescence from quartz. The environmental dose rate is estimated from laboratory measurements of environmental radioactivity, plus the small contribution from cosmic rays.

We collected 24 sediment samples from depths of between 467 and 1,670 cm below the present ground surface. We collected samples from one half of the split core in subdued red light. Each sample represents a 10 cm depth interval with the midpoint depth provided in Supplementary Table 1. We first removed light-exposed portions of the samples to measure the prevailing moisture content and radioactivity in the sediment to estimate the environmental dose rate. We then prepared the remaining light-safe core, and extracted quartz grains to measure the optically stimulated luminescence signal and estimate De.

We took single-grain optically stimulated luminescence measurements of De for all samples. We prepared and measured samples using methods and equipment described and tested previously81,82. We measured all single quartz grains with diameters of 180–212 µm using the single-aliquot regenerative dose procedure83,84. Aberrant grains were rejected using a standard set of quality assurance criteria81. We used two different methods (conventional single-aliquot regenerative dose and the weighted-mean natural signal (LnTn) method85) to determine De for age estimates from the same data. We applied the LnTn method to samples that contained a large proportion of ‘saturated’ grains (that is, the optically sensitive traps filled with respect to dose), resulting in a truncation of their conventional single-aliquot regenerative dose De distributions and consequent underestimation of De and age82,86,87. We did all analyses for both methods using the functions implemented in R packages numOSL88,89 and Luminescence90.

The environmental dose rate consists of the external beta, gamma and cosmic ray dose rates and an internal alpha dose rate. We assumed an internal alpha dose rate of 0.031 ± 0.008 Gy kyr−1 for all samples. We measured beta dose rates directly using a Risø GM-25-5 multi-counter system based on published procedures91,92,93. We calculated gamma dose rates from estimates of uranium (U) and thorium (Th) obtained from thick-source alpha counting and potassium (K) from inductively coupled plasma optical emission spectroscopy, except for two small organic-rich samples (GIR3-O-1992 and GIR3-R-1993), for which we measured the U, Th and K contents using instrumental neutron activation analysis. We converted the U, Th and K values to gamma dose rates (Gy kyr−1) using established dose rate conversion factors90. We also took into account the cosmic ray contribution to the total dose rate94, adjusting for site altitude (25.9 m), geomagnetic latitude and the density and thickness of water (1 m) and sediment (variable) overburden. We used an average density of water of 1.025 g cm−3. There is high variation in water content and bulk density of the sediment down-core. These differences in bulk density are important because energy loss per gram per cubic centimetre is greater at lower densities. We constructed a simple model assuming an average water depth of ~1 m over the entire burial period and used present-day sediment depths, moisture content and bulk density estimates.

We corrected the measured beta dose rates for grain-size attenuation, and adjusted the beta, gamma and cosmic ray dose rates for water content (based on the measured field values and bulk densities) and for organic attenuation for organic-rich samples95,96. We divided the sample De by the corresponding total environmental dose rates to calculate the optically stimulated luminescence ages in calendar years before present; the associated uncertainties are at the 68.2% confidence level and include all known and estimated sources of random and systematic error.

Age model

We applied a Bayesian approach for age modelling. We used the R package rbacon94 to calibrate radiocarbon ages using the Southern Hemisphere calibration97 SHCal20 and combined calibrated radiocarbon ages and optically stimulated luminescence ages for determination of depth probability distributions of age in R 4.0.0 (ref. 98). See Supplementary Fig. 4 for the age depth model and Supplementary Table 1 for the data associated with the new optically stimulated luminescence dates that we employed in generating the model. We used the mean output from the model. The average deviation of the median from the mean age at each increment across the whole record is 0.2% of the mean age, the maximum difference is 2.1% for ages older than 400 yr.

Analysis

We hypothesized that the relationship between MARSPAC and PARchar would shift with the onset of landscape-scale moderation of the fire regime by people. Under this hypothesis, the pre-human era would have higher-intensity fires and therefore more MARSPAC relative to PARchar compared with the era dominated by anthropogenic fire regimes. This can be visualized as a bivariate plot of MARSPAC versus PARchar, with the expectation that MARSPAC increases at a faster rate relative to PARchar under the natural regime compared with the anthropogenic fire regime. Likewise, higher-intensity fires characteristic of natural fire regimes should produce lower δ13CSPAC in charcoal given the higher preservation potential of coarse, woody fuels compared with fine, grassy fuels. We therefore expected a more pronounced positive relationship between the proportion of C3 to C4 biomass contributing to the char (as measured by δ13CSPAC) and the vegetation inferred from the percentage of grass pollen in total dryland pollen under an anthropogenic fire regime compared with a natural one.

We compared the relationship between MARSPAC and PARchar, and between the percentage of grass pollen and δ13CSPAC, in the definitive pre-human era with a period when humans had probably established landscape-scale fire control. Choosing the date of this transition is somewhat arbitrary because we do not know precisely when this human signal would have emerged, but this is immaterial for analysis given the necessity of testing the relationship only during periods of relatively wetter climates (as we explain below). Initially we chose 30 ka as the threshold date before which we assumed a natural fire regime, and after which an anthropogenic fire regime. Examining the bivariate relationships for these two periods while ignoring the relative wetness regime (as measured by TOC), there is little indication of a shift in the MARSPAC versus PARchar and percentage of grass pollen versus δ13CSPAC relationships as hypothesized; in fact, the slopes of the two relationships for the natural and anthropogenic fire regimes appear to be opposite to expectations (Supplementary Fig. 1). In the case of MARSPAC versus PARchar, this is an artefact of the physical breakdown of micro-charcoal particles in the clay-rich sections of the core through the development of charcoal–mineral complexes such that the charcoal is comminuted into size classes finer than can be visually recognized (<10 µm) during laboratory disaggregation. We could identify the issue of comminution to finer particle sizes in our record because: (1) we had both clay-rich and clay-poor intervals to compare; and (2) we had a geochemical measure of pyrogenic carbon abundance that did not rely on visually counting char particles in a particular size class.

When we considered the relationships through filtering by the TOC content such that we included only reliable PARchar, measurements, relationships emerged in the hypothesized direction: the trend for MARSPAC versus PARchar was less pronounced under the natural fire regime, and the trend for the percentage of grass pollen versus δ13CSPAC was more pronounced under the anthropogenic fire regime (Fig. 3).

To test whether such relationships were supported statistically, and to examine the sensitivity of the results to threshold choices (that is, minimum organic carbon abundance and the temporal transition point from natural to anthropogenic fire regimes), we implemented the following procedures. We first standardized the data to ensure that we had values at consistent and regular intervals along the time series using the ‘approx’ function in R; here, we took values every 250 years across each time series (that is, 150 ka, 145.75 ka, 145.50 ka … 0 ka). This increment was a trade-off between uniform coverage of the time series and oversampling. The average temporal resolution across the entire record was 550 years.

Next, we split all of the time series at 30 ka; we assigned values >30 ka as the natural fire regime and values ≤30 ka as the anthropogenic fire regime. We then filtered the time series to retain only the values by successively lower percentage TOC contents, starting with a threshold of >5.1% TOC, which was the lowest TOC percentage from 11 ka to the present (Fig. 1a). For the MARSPAC versus PARchar relationship, we first log10-transformed each variable given the relationship’s power-law-like behaviour, and then randomly resampled x, y pairs (with replacement) from the natural-regime period using only the number of x, y pairs from the anthropogenic-regime period. Repeating this process 10,000 times, we calculated the number of iterations where the Spearman’s ρ for the resampled natural-regime relationship was less than Spearman’s ρ for the resampled anthropogenic-regime relationship. Dividing the number of times ρnatural < ρanthropogenic by the number of iterations provided a probability that the observed difference in the relationships could be derived randomly (akin to a type I error probability). For the percentage of grass pollen versus δ13CSPAC relationship, we applied the same process but instead calculated the numbers of times ρnatural > ρanthropogenic to test the hypothesized change in the relationship.

To examine the sensitivity of the results to the choice of the TOC threshold, we repeated the procedure for both MARSPAC versus PARchar and the percentage of grass pollen versus δ13CSPAC for decreasing TOC (>10%, >9%, >8% … >1%). To test the sensitivity of our results to the choice of the temporal threshold below and above which the categories of natural and anthropogenic fire regime applied, we repeated the resampling procedure at a wetness value of >5.1% TOC, but varied the temporal split by values of 1,000 yr either side of 30 ka to the extremes of 65 ka and 1 ka.