Forest mosaics, not savanna corridors, dominated in Southeast Asia during the Last Glacial Maximum

Significance We present new qualitative and statistical analyses of 59 palaeoecological records across Southeast Asia to show that, instead of swings between open savanna and dense rainforest ecosystems, the climatic changes of the Last Glacial Period (119–11.7 ka) and particularly the Last Glacial Maximum (conventionally ~23–19 ka) involved fluid transitions between lowland rainforest, more open canopy seasonally dry forest, and tropical montane forest. This “hybrid” open forest biome provides an alternative to the currently accepted binary ecologies for the region and yields new insights into ecological resilience for tropical forests in Southeast Asia and beyond. Additionally, it points to diversified rather than overturned resource availability for humans that were occupying and migrating through the region.

The exposed Sunda shelf during the LGM was covered with humid, rainforest.A marshy vegetation (sedges, reeds, bamboo) fringed with palms and ferns developed in the valley along the North Sunda River.
The forest contribution is dominated by lowland types.Lower montane rainforest groups peak in the early LGM (SO18300).There is some evidence for distal, downslope migration of upper montane forest types.An increase in grass (and sedge) pollen during the LGM is thought to be more closely linked to the expansion LGM dominated by forest pollen with a lower fernland and swampland contribution than for the slope flat cores analyzed in Wang et al. (2009).LGM dominated by forest pollen.During this time, upper montane forest expanded and compressed the living space of lower montane and lowland forests.This record is interpreted as capturing a fluctuating signal between Borneo (MIS 1 and MIS 3) and the Sunda shelf (MIS 2).During MIS 2, including the LGM δ 13 C org record (which captures a terrestrial and marine signal), suggests that the most prominent vegetation type on the Sunda Shelf during the LGM was C 3 plants.The pollen record suggests that this vegetation comprised lowland forest.
Pollen: Montane forest δ 13 C org:NA C 4 grasslands expanded during the LGM.C 3 vegetation increases between 18 ka and 7-5.5 ka (with a slight reversal during the Bølling-Allerød period).Last 2 k years characterized by increasing C 4 plant abundance.
The location of core 69-3 in the Savu Basin, which is bordered only by narrow shelves, means that higher C 4 herbs input during the last glacial stages cannot solely be from grasslands colonizing the nearby exposed shelves.Rather, tropical rainforest on the islands are thought to be replaced with C 4 herbs.The period of the record between 30.8 ka and 12.7 ka is missing or highly compressed.This is attributed to reduced lake levels under a drying climate and, likely, the formation of a shallow wetland around the lake during that time.This means that the vegetation signal from MIS 2 captured in the core sediments is dominated by a local, herbaceous assemblage.Regional vegetation change on either side of the LGM is difficult to gauge due to the hiatus and local signal being recorded by the lake sediments.However, the regional vegetation appears dominated by a montane forest at around 32 ka to 30 ka, and by a relatively open/seasonal lowland forest signal after 12.7 ka.This record likely captures a strong signal from the north-western Australia, with some contribution from maritime southeast Asia.The period between 40 ka and 11.7 ka, including the LGM, is characterized by the expansion of grassland taxa.

Hordorli (swamp)
Montane forest grew continuously around the site with an increase in higher altitude forest between 25 ka and 10.5 ka.(Birdsell, 1977).
The down-core sampling distribution and age uncertainty windows for the 'canopy openness' time series used for comparative analysis are shown on Fig. S2 (δ 13 C records) and Fig. S3 (pollen records).For the records that permitted comparative analysis (see Material & Methods in the main text and the supplementary dataset), we used the presence or absence of a curve inflection (grey shading) between MIS 3 and MIS 2, and between MIS 2 and MIS 1, to infer whether there was canopy opening, canopy closing, or no canopy change in response to the onset and termination of the LGM.This is summarized using arrows on Fig. 4 in the main text.The resampled curves (grey shading) serve as data inputs for the correlation and comparative analyses summarized on Fig. 5 and Fig. 6 in the main text.) and sediment plant wax (row 2) records we obtained (black points).Grey shading shows the range boundaries (95% confidence interval) for datasets that we resampled at 2000-year intervals between 2 ka and 34 ka for correlation and volatility analysis, using the minimum and maximum ages calculated from age-depth modelling of the sequences (see Supplementary Text Fig. S3: Stratigraphic plot of select pollen data (black dots) used to infer landscape openness through time.Grey shading shows the range boundaries (95% confidence interval) for datasets that we resampled at 2000-year intervals between 2 ka and 34 ka for correlation and volatility analysis, using the minimum and maximum ages calculated from age-depth modelling of the sequences (see Supplementary Text 3).Data plotted against climate anomalies (Hadley CM3) calculated for the region (20º N-11º S; 98º E-141º E) across the same period.Supplementary Text 3: Methods for extracting and synthesizing time-transgressive pollen and isotopic records.

Mainland Sunda
NPK.2 -Nong Pa Kho (Thailand) (Penny, 2001) Chronology: We remodeled five radiocarbon ages from analysis of bulk sediment samples reported in Penny (2001) using Bacon 2.5.0 (Blaauw and Christen, 2011) in R (R Core Team, 2023) (Fig. 4).We ran the model in 48 sections using a sediment accumulation rate (acc.mean) of 200 years cm -1 estimated from linear interpolation between adjacent date samples.We set an assumed age of -44 cal years before present (1994 CE) for the core top (0 cm).We set the basal depth for the core to 230 cm (d.max = 230).We used the Southern Hemisphere calibration curve (SH20) (Hogg et al., 2020), offset by −21 ± 6 yrs, in this analysis (see Hamilton et al. 2019a for justification of this curve and offset).We used the median age-depth model age output to estimate ages for the core sediments at 1-cm intervals.
Pollen data preparation: The author DP provided raw pollen data from the original study (Penny, 2001).We converted select data (grouped montane taxa and Poaceae) into dryland percentage (including Poaceae).Because this site comprises a large herbaceous wetland, much of the Poaceae data are likely derived from a local wetland source (Penny, 2001).This overemphasizes the 'openness' of the terrestrial vegetation signal.
Sunda -Insular Southeast Asia SO18300 -Sunda shelf (South China Sea) (Wang et al., 2009) Chronology: We remodeled five radiocarbon ages from analysis of organic samples (Hanebuth and Stattegger, 2004;Hanebuth et al., 2003) using Bacon 2.5.0 (Blaauw and Christen, 2011) in R (R Core Team, 2023) (Fig.. 5).We ran We ran the model in 179 sections using a sediment accumulation rate (acc.mean) of 50 years cm -1 estimated from linear interpolation between adjacent date samples.We set the basal depth for the core at 885 cm (d.max = 885) and the top of the core at d.min = 0. To maximise the probability of the dates overlapping with the age-depth model (60%), we did not assign a modern age to the upper-core sediments.We used the marine calibration curve (Marine20) (Heaton et al., 2020) in this analysis.There are large uncertainties in the age-depth model produced for this core and the model rejected two out of the five dates (Fig. S5).
Pollen data preparation: We extracted grass pollen data (% terrestrial land seed plants) from graphics presented in Wang et al. (2009) using WebPlotDigitizer (Rohatgi, 2021).SO18323 -Sunda shelf (South China Sea) (Wang et al., 2009) Chronology: We remodeled three radiocarbon ages, reported from analysis of mixed samples (Hanebuth and Stattegger, 2004;Hanebuth et al., 2003) for this study using Bacon 2.5.0 (Blaauw and Christen, 2011) in R (R Core Team, 2023) (Fig. S6).We ran the model in 110 sections using a sediment accumulation rate (acc.mean) of 50 years per cm estimated from linear interpolation between adjacent date samples.We set an assumed age of -47 cal yrs BP (1997 CE) was set for the core top (0 cm).The basal depth for the core was set at 540 cm (d.max = 540).We used the marine calibration curve (Marine20) (Heaton et al., 2020) in this analysis.
There are large uncertainties in the age-depth model produced for this core and the model rejected one out of the three dates used in the analyses (Fig. S6).Fig. S6: Age-depth model produced for SO18323 (Wang et al., 2009).

Pollen data preparation:
We extracted grass pollen data (% terrestrial land seed plants) from graphics presented in Wang et al. (2009) using WebPlotDigitizer (Rohatgi, 2021).SO18302 -Sunda shelf (flat) (South China Sea) (Wang et al., 2009) Chronology: We remodeled three radiocarbon ages, reported from analysis of mixed samples (Hanebuth and Stattegger, 2004;Hanebuth et al., 2003) using Bacon 2.5.0 (Blaauw and Christen, 2011) in R (R Core Team, 2023) (Fig. S 7).We ran the model in 121 sections using a sediment accumulation rate (acc.mean) of 20 years per cm estimated from linear interpolation between adjacent date samples.We set the basal depth for the core at 598 cm (d.max = 598) and the top of the core at d.min=0.To maximise the probability of the dates overlapping with the age-depth model (66%), the upper core sediments were not assumed to be modern.We used the marine calibration curve (Marine20) (Heaton et al., 2020) in this analysis.
There are large uncertainties in the age-depth model produced for this core and the model rejected one out of the three dates used in the analysis (Fig. S7).

Pollen data preparation:
We extracted grass pollen data (% terrestrial land seed plants) from graphics presented in Wang et al. (2009) using WebPlotDigitizer (Rohatgi, 2021).17964 -Sunda shelf (slope) (South China Sea) (Sun et al., 2000) Chronology: We remodeled eight radiocarbon ages, reported from analysis of samples reported in Sun et al. (2000) for this study using Bacon 2.5.0 (Blaauw and Christen, 2011) in R (R Core Team, 2023) (Fig. S8).We ran the model in 131 sections using a sediment accumulation rate (acc.mean) of 20 years per cm estimated from linear interpolation between adjacent date samples.We used an assumed age of -45 cal yrs BP (1995 CE) for the core top (0 cm) and set 1303 cm as the basal depth for the core (d.max = 1303).We used the marine calibration curve (Marine20) (Heaton et al., 2020) fpr this analysis.Fig. S8: Age-depth model produced for 17964 (Sun et al., 2000) Pollen data preparation: Herb pollen (combined wetland and dryland herbs) and grouped arboreal pollen (% land seed plants) were extracted from graphics presented in (Sun et al., 2000) using WebPlotDigitizer.
CB19 -Sunda shelf (South China Sea) (Yang et al., 2021) Chronology: We remodeled eight radiocarbon ages, reported from analysis of seven mixed planktonic foraminifera/ Neogloboquadrina dutertrei samples in Yang et al. (2021) using Bacon 2.5.0 (Blaauw and Christen, 2011) in R (R Core Team, 2023) (Fig. S9).We ran the model in 70 sections using a sediment accumulation rate (acc.mean) of 100 years per cm estimated from linear interpolation between adjacent date samples.We set an assumed age of -62 cal yrs BP (2012 CE) for the core top (0 cm) and set the basal depth for the core at 342 cm (d.max = 342).We used the marine calibration curve (Marine20) (Heaton et al., 2020) in this analysis.
Pollen data preparation: We extracted grouped montane forest pollen (i.e., the sum of upper and lower montane types) and grass pollen (% terrestrial plants) from the graphics presented in Yang et al. (2021) using WebPlotDigitizer (Rohatgi, 2021).NS07-25 -South China Sea (Nansha Trough) (Xiang et al., 2009) Chronology: We remodeled eleven radiocarbon ages reported from planktonic foraminifera in Xiang et al. (2009) using Bacon 2.5.0 (Blaauw and Christen, 2011) in R (R Core Team, 2023) (Fig. S10).We ran the model in 113 sections using the marine calibration curve (Marine20) (Heaton et al., 2020).We used a sediment accumulation rate (acc.mean) of 50 years per cm estimated from linear interpolation between adjacent date samples.We set the basal depth for the core at 556 cm (d.max = 556), the minimum core depth at 0cm (d.min = 0) and used an assumed age of -57 cal.BP (2007 CE) for the core top (0 cm).Fig. S10: Age-depth model produced for NS07-25 (Xiang et al., 2009) Pollen data preparation: Thilakanayaka et al. (2019) provided the raw pollen data from the original study on request.We calculated grass and grouped montane forest pollen as a percent of the dryland pollen sum for this study.
PemCoreB -Lake Pemerak Core B (West Kalimantan, Indonesia [Borneo]) (Anshari et al., 2004) Chronology: We remodeled eight radiocarbon ages for Lake Pemerak Core B (West Kalimantan, Borneo) reported from 7 x pollen concentrates and 1 x bulk sediment sample (Anshari et al., 2004) using Bacon 2.5.0 (Blaauw and Christen, 2011) in R (R Core Team, 2023) (Fig. S11).An hiatus was inserted at the sediment boundary at 69cm, where extrapolation between dated points indicate a ~10,000 year time gap over a depth interval of 4 cm.We set the sedimentation rate between 0 and 69 cm at 50 years per cm, and the sedimentation rate between 69 cm and the basal depth (120 cm) at 200 years per cm based on linear extrapolation between dated points on either side of the sediment boundary.We ran the model in 25 sections using the Southern Hemisphere calibration curve (SHCal20) (Hogg et al., 2020).We set basal depth for the core at 120 cm (d.max = 120), and set the minimum core depth at 0cm (d.min = 0).Fig. S11: Age-depth model produced for PemB (Anshari et al., 2004).
Pollen data preparation: We extracted percentage pollen data grouped into major plant functional types from the original publication using WebPlotDigitizer (Rohatgi, 2021).The plots for herbs (%) and montane forest (%) were used in the analysis.
PemCoreC -Lake Pemerak Core C (West Kalimantan, Indonesia [Borneo]) (Anshari et al., 2004) Chronology: We remodeled three radiocarbon ages for Lake Pemerak Core C (West Kalimantan, Borneo) reported from pollen concentrates (Anshari et al., 2004) using Bacon 2.5.0 (Blaauw and Christen, 2011) in R (R Core Team, 2023) (Fig. S12).We ran the model in 62 sections using the Southern Hemisphere calibration curve (SHCal20) (Hogg et al., 2020) and a sediment accumulation rate (acc.mean) of 500 years per cm estimated from linear interpolation between adjacent date samples.We set the basal depth for the core at 300 cm (d.max = 300), and the minimum core depth at 0cm (d.min = 0), noting that the core surface sediments are unlikely to be modern.Because the oldest dated point was 81 cm depth, data from this core is only presented to 80 cm depth (modelled at 44901 cal.BP).

Pollen data preparation:
We extracted percentage pollen data grouped into major plant functional types from graphics presented in Anshari et al. (2004) using WebPlotDigitizer (Rohatgi, 2021).We used the plots for herbs (%) and montane forest (%) in this analysis.
Chronology: We used the original chronology (van der Kaars et al., 2010) in this analysis as extractable data are only plotted against age rather than depth.
Pollen data preparation: We extracted dryland percentage data (grouped lowland forest, montane forest, and herb pollen) from van der Kaars et al. (2010) using WebPlotDigitizer (Rohatgi, 2021) for use in this study.SO189-144KL -Nias Basin (Indian Ocean off Southwest Sumatra) (Niedermeyer et al., 2014a) Chronology: We remodeled thirty-six radiocarbon ages for SO189-144KL produced from analysis of mixed planktonic, G. sacculifer and G. ruber samples (Mohtadi et al., 2014) for this study using Bacon 2.5.0 (Blaauw and Christen, 2011) in R (R Core Team, 2023) (Fig. S14).We ran the model in 166 sections using the marine calibration curve (Marine20) (Heaton et al., 2020) and a sediment accumulation rate (acc.mean) of 20 years per cm estimated from linear interpolation between adjacent date samples.We set an assumed age of -56 cal yrs BP (2006 CE) for the core top (0 cm) and set the basal depth at 823 cm (d.max = 823).
Pea-Sim-sim (PSS) (North Sumatra, Indonesia) (Maloney, 1980) Chronology: We remodeled thirteen radiocarbon ages for Pea-Sim-sim Swamp for this study using Bacon 2.5.0 (Blaauw and Christen, 2011) in R (R Core Team, 2023) (Fig. S15).We estimated depths from the pollen diagram in Maloney (1980) (Table S2).We ran the model in 197 sections using the Southern Hemisphere calibration curve (SHCal20) (Hogg et al., 2020) and a sediment accumulation rate (acc.mean) of 20 years per cm estimated from linear interpolation between adjacent date samples.We set an assumed age of -30 cal yrs BP (1980 CE) for the core top (0 cm) and set the basal depth at 975 cm (d.max = 975).Fig. S17: Age-depth model produced for DDA (Newsome and Flenley, 1988).
PeaBullok A (PB-A) -Pea Bullok (Sumatra) (Maloney and McCormac, 1995) Chronology: We remodeled seven radiocarbon ages from PeaBullokA bulk sediment samples (Maloney and McCormac, 1995) using Bacon 2.5.0 (Blaauw and Christen, 2011) in R (R Core Team, 2023) (Fig. S18).We ran the model in 162 sections using the Southern Hemisphere calibration curve (SHCal20) (Hogg et al., 2020) and a sediment accumulation rate (acc.mean) of 50 years per cm estimated from linear interpolation between adjacent date samples.We set an assumed age of -40 cal yrs BP (1990 CE) for the core top (0 cm) and set the basal depth for the core at 800 cm (d.max = 800).
Isotopic data: We obtained δ 13 C data from the PANGEA database (Ruan et al., 2018b).
Pollen data preparation: We obtained grass, montane forest and lowland forest pollen percentage data from the PANGEA database (Ruan et al., 2018b).
Pollen data preparation: We extracted grouped montane forest, mangrove forest, lowland forest, and grassland pollen percentage data from Bian et al. (2011) using WebPlotDigitizer (Rohatgi, 2021).These data were recalculated to reflect percentage of the dryland sum by removing the mangrove data.

Wallacea
GeoB10069-3 -Suvu Sea (Sumba) (Dubois et al., 2014) Chronology: We remodeled eighteen radiocarbon ages and their offset values reported from mixed planktonic samples (Dubois, 2014) using Bacon 2.5.0 (Blaauw and Christen, 2011) in R (R Core Team, 2023) (Fig. S22).We ran the model in 192 sections with a boundary set at 725cm.We used a sediment accumulation rate (acc.mean) of 20 years per cm for sediments shallower than 725cm, and 100 years per cm for sediments deeper than 725cm based off linear interpolation between adjacent date samples.We set the core top sediments were set to -50 cal.BP and used the marine calibration curve (Marine20) (Heaton et al., 2020).

Pollen data preparation:
We extracted grouped percentage pollen data (monsoonal (seasonal) forest, montane forest, and C4 herbs) from Dubois et al. (2014) using WebPlotDigitizer (Rohatgi, 2021).TOW10-9B (TOW9) -Towuti (Sulawesi) (Russell et al., 2014) Chronology: We remodeled twenty-three reservoir-corrected radiocarbon ages reported from analysis of 20 bulk organic carbon and three terrestrial macrofossil samples and one Pb 210 age (Russell et al., 2014) using Bacon 2.5.0 (Blaauw and Christen, 2011) in R (R Core Team, 2023) (Fig. S23).We ran the model in 123 sections using a sediment accumulation rate (acc.mean) of 50 years per cm and used the Southern Hemisphere calibration curve (SHCal20) (Hogg et al., 2020).We set the core top sediments to -64cal.BP, and the basal sediments were set to 1156cm (d.max=1156).Pollen data preparation: We extracted grass pollen percentage data as a percentage of the dryland count from Dam et al. (2001) using WebPlotDigitizer (Rohatgi, 2021).G4K12P1 (K12P1) -Molucca Sea Core (Maluku) (van der Kaars, 1991) Chronology: We remodeled two radiocarbon ages for the core, reported from analysis of Pteropods in van der Kaars (1991) using Bacon 2.5.0 (Blaauw and Christen, 2011) in R (R Core Team, 2023) (Fig. S26).We ran the model in 109 sections using the marine calibration curve (Marine20) (Heaton et al., 2020) and a sediment accumulation rate (acc.mean) of 50 years per cm estimated from linear interpolation between adjacent date samples.We set an assumed age of -40 cal yrs BP for the core top (0 cm) and set the basal depth at 535 cm (d.max = 535).Pollen data preparation: We extracted grouped montane forest, mangrove forest lowland forest, and woodland and grass pollen percentage data from van der Kaars (1991) using WebPlotDigitizer (Rohatgi, 2021).BJ8-03-91GGC (91GGC) -Celebes Sea - (Dubois et al., 2014) Chronology: We remodeled six radiocarbon ages for the core from the analysis of mixed planktonics and G. sacculifer (Dubois, 2014) using Bacon 2.5.0 (Blaauw and Christen, 2011) in R (R Core Team, 2023) (Fig. S27).We ran the model in 77 sections using the marine calibration curve (Marine20) (Heaton et al., 2020) and a sediment accumulation rate (acc.mean) of 50 years per cm estimated from linear interpolation between adjacent date samples.We set an assumed age of -41 cal yrs BP for the core top (0 cm), and set the basal depth at 375 cm (d.max = 375).
Pollen data preparation: We extracted percentage grass data from the Hope and Tulip (1994) using WebPlotDigitizer (Rohatgi, 2021).Chronology for grass pollen: We remodeled nine radiocarbon ages for the core (0 to 185cm) from analysis of planktonic foraminifera (van der Kaars et al., 2000), and nine additional tie points from sediments greater than 297cm determined from δ 18 O benthic foraminifera tuning (van der Kaars et al., 2000) and assigned at error of +/-1000, using Bacon 2.5.0 (Blaauw and Christen, 2011) in R (R Core Team, 2023) (Table S3) (Fig. S29).We ran the model in 154 sections using the marine calibration curve Marine20 (Heaton et al., 2020) and a sediment accumulation rate (acc.mean) of 200 years per cm estimated from linear interpolation between adjacent date samples.We set an assumed age of -40 cal yrs BP for the core top (0 cm) and set the basal depth at 764 cm (d.max = 764).Pollen data preparation: We extracted grouped montane forest, mangrove forest, lowland forest, grassland and woodland pollen percentage data from van der Kaars (1991) using WebPlotDigitizer (Rohatgi, 2021), and recalculated them to reflect percentage of the dryland sum by removing the mangrove data.We also removed data from depths greater than 900 cm due to the lack of chronological resolution for the deeper sediments (Fig. S30).
G6-4 -Sahul-Wallacea-Sunda mixed signal (Lombok Ridge) (van der Kaars, 1991) Chronology: We remodeled five radiocarbon ages reported from analysis of Pteropods (van der Kaars, 1991) using Bacon 2.5.0 (Blaauw and Christen, 2011) in R (R Core Team, 2023) (Fig. S31).We ran the model in 182 sections using the marine calibration curve (Marine20) (Heaton et al., 2020) and a sediment accumulation rate (acc.mean) of 200 years per cm estimated from linear interpolation between adjacent date samples.We set an assumed age of -40 cal yrs BP (1990 CE) was set for the core top, and the basal depth for the core was set at 900 cm (d.max = 900).
forest (Nothofagus/ Castanopsis/ Phyllocladus) forest during the LGM.NA Supplementary Text 2: A visualization of the individual records we analyzed that show the inputs used to reconstruct habitat types and transitions in canopy openness, and the data inputs for correlation and comparative analyses.An overview of broad-scale habitat classifications for each Marine Isotope Stage (MIS) we included, as well as the LGM (conventionally classified as ~ 23 ka to 19 ka) (6) are in Fig. S1.These data feed into the simplified habitat classification made on Fig. 3 and Fig. 4 in the main text.

Fig
Fig.S2: Stratigraphic plot of δ 13 C records showing guano (row 1) and sediment plant wax (row 2) records we obtained (black points).Grey shading shows the range boundaries (95% confidence interval) for datasets that we resampled at 2000-year intervals between 2 ka and 34 ka for correlation and volatility analysis, using the minimum and maximum ages calculated from age-depth modelling of the sequences (see Supplementary Text
More positive leaf wax δ 13 C during the LGM indicate an expansion of more drought-tolerant C 4 vegetation.δ 13C n-29 suggest the expansion of montane rainforest taxa during the LGM.δ 13 C n-33 and δ 13 C n-31 suggest the expansion of more open forest or savanna in the lowlands.High grass pollen (40%) during the LGM probably reflects a more open-canopy vegetation type locally in lowland East Java.Due to the blocking effect of the ~ 3000 km-long volcanic arc mountain chains from Sumatra to Java, the authors assumed the contribution of pollen taxa from the exposed Sunda Shelf during the LGM were minor.Australian contributions to the overall pollen counts are relatively low but consistent.
This record likely captures vegetation changes from New Guinea.The end of the LGM and millennia following (i.e., 17 ka to 10 ka) is characterized by a higher prevalence of woodland and grassland taxa.This suggests a more open forest in New Guinea during the peak of the LGM, though the record does not extend back into this period.This record likely captures a mixed signal derived from Timor, northern Australia and the Sahul shelf.Pollen data from the LGM suggest the expansion of an open woodland-scrubland vegetation with forested rims along rivers.
enrichment (to between −26 and −18‰) suggest replacement with savanna (C 4 ) vegetation.Rainforest was again present in the cave area by 13.5 ka. 4 grasslands were a large component of regional vegetation from at least 35 ka until 16 ka, remaining > −22.6‰ until the end of the LGM.After the LGM, an initial decline in δ 13 C occurred at ∼14.7 ka, with an increase in δ 13 C to −23.3‰ between 13.4 ka and 12.5 ka.C 3 vegetation (forest) is evident after 10.5 ka and persisted until the present.

Table S3 :
14C and age-tie points used to remodel the chronology of SH1-2014