Onset of summer aridification and the decline of Homo floresiensis at Liang Bua 61,000 years ago

Abstract

The cause of the disappearance of the primitive hominin Homo floresiensis from the Indonesian island of Flores about 50,000 years ago is a key question in palaeoanthropology. While the potential roles of climate change and human agency continue to be debated, the history of freshwater availability essential for survival at the type locality, Liang Bua, remains poorly understood. Although speleothem δ¹⁸O is widely used to reconstruct monsoon rainfall, variations in summer and winter rainfall, with distinct δ¹⁸O values, can complicate interpretations of mean annual rainfall. Here, we combine speleothem Mg/Ca, a proxy for local rainfall, with δ¹⁸O to determine annual, summer and winter rainfall amounts concurrent with H. floresiensis and Stegodon, one of its primary prey. Geochemical modelling of the Mg-¹⁸O system reveals a sustained decline in mean annual rainfall from ~1560 to 990 mm between 76,000 and 61,000 years ago. Critically, summer rainfall decreased to a record low of ~450 mm at 61,000–55,000 years ago, alongside a marked decline in both the abundance and relative proportion of Stegodon remains in the fossil record. These findings increase the likelihood that progressive landscape aridification, and intensified human-faunal competition for dwindling resources, culminated in abandonment of Liang Bua.

Introduction

In 2003, excavations at the Liang Bua rock shelter uncovered the skeletal remains of a previously unknown small-bodied hominin species, Homo floresiensis^1,2. The H. floresiensis occupation interval was initially dated to 95–12 thousand years (ka) ago^2,3, raising the possibility of interaction with early modern humans (Homo sapiens) in Island Southeast Asia, with implications for the extinction of the species. Following revisions to the site’s stratigraphy and chronology, however, all H. floresiensis skeletal remains now date to about 100–60 ka⁴. The associated stone artefact record and assemblage of large fauna typically found with H. floresiensis [Stegodon florensis insularis (pygmy elephant), Leptoptilos robustus (giant marabou stork), Trigonoceps sp. (vulture)] end at about 50 ka^5,6. The revisions indicate that their disappearance from the record preceded modern human arrival at Liang Bua around 46 ka, marked by a notable shift in the assemblage of raw materials for stone tool making⁶. Definitive evidence of a temporal overlap between modern humans and H. floresiensis elsewhere on Flores also is lacking, leaving compelling reasons to consider the role of climate change in the extirpation.

Freshwater availability is crucial for the survival of humans and large fauna^7,8. The Liang Bua biome would have been shaped in part by the seasonally dry rainfall regime on Flores, and the timing and duration of the Indonesian-Australian summer monsoon. In particular, S. florensis insularis, an important prey for H. floresiensis, weighed ~350–950 kg⁵ and, like modern elephants, may have had limited tolerance for dry-season aridity^9,10. At present, the Indo-Pacific Warm Pool drives deep atmospheric convection in Island Southeast Asia and the seasonal distribution of rainfall is modulated by the latitudinal migration of the Intertropical Convergence Zone. About 60% of the mean annual rainfall (~1400 mm) in the uplands around Liang Bua occurs during the austral summer (December–March) due to the northwest monsoon. The remainder falls during the dry season and two monsoon transition months (April and November), and is related to low-level south-easterly tradewinds and storms generated along the southern periphery of the Intertropical Convergence Zone. Hydroclimate reconstructions covering the H. floresiensis occupation of Liang Bua indicate that past variations in sea level, insolation and inter-hemispheric temperature gradients had profound influences on regional rainfall^11,12,13,14. Thus, the understanding of keystone fossil taxa discontinuities at Liang Bua requires the assessment of past rainfall with emphasis on the shifting balance between wet-season and dry-season rainfall at the site.

Much of our current knowledge of the western Flores hydroclimate pertains to the last 45,000 years, and is derived from precisely dated speleothem oxygen-isotope (δ¹⁸O) records^15,16,17,18. These highly resolved time series document rainfall δ¹⁸O variability, providing insights into broad-scale monsoon circulation. In low-latitude Australasia, where rainfall patterns are bimodal, however, speleothem δ¹⁸O alone generally cannot be used to definitively infer local rainfall amounts^19,20,21,22. For example, on Flores changes in the relative proportions of convective summer rainfall and tradewind-associated winter rainfall, with contrasting δ¹⁸O values^15,23, can affect speleothem δ¹⁸O without altering the mean annual rainfall. Thus, a critical challenge is to elucidate the histories of both rainfall amount and rainfall seasonality specific to Liang Bua.

In optimal karst settings, speleothem Mg/Ca serves as an independent indicator of past changes in local monsoon rainfall^{24,25,26,27,28,29,30}. Accumulating evidence shows that prior calcite precipitation (PCP) in karst void spaces, where CO₂ degassing occurs, can increase Mg/Ca in cave dripwaters due to preferential exclusion of Mg²⁺ during calcite formation^{31,32,33,34,35}. Reduced rainfall and karst infiltration rates create conditions favouring PCP, resulting in higher Mg/Ca in the speleothem ‘downstream’. In principle, tandem measurements of Mg/Ca and δ¹⁸O in speleothems can be used to infer rainfall seasonality. We interpret times of difference between Mg/Ca and δ¹⁸O in terms of the mixing ratio of summer monsoon rainfall (lower δ¹⁸O) and winter tradewind rainfall (higher δ¹⁸O), which dictates the amount-weighted average speleothem δ¹⁸O^36,37. Ideally, the tandem Mg-¹⁸O method should expedite the reconstruction of local summer and winter rainfall amounts, even for tropical settings where annual rainfall and monsoon seasonality change in unexpected ways.

We develop this approach using high-resolution measurements of Mg/Ca and δ¹⁸O for a precisely dated stalagmite from Liang Luar cave, located near Liang Bua (Fig. 1). Our aim was to construct a season-specific record of local rainfall amounts and monsoon seasonality during 91–47 ka, including the critical final phase of H. floresiensis at Liang Bua. The new hydroclimate record reveals a gradual ~37% decline in mean annual rainfall from 76 to 61 ka, with maximum summer aridity (~48% lower than at present) at 61–55 ka. We link the rainfall reconstruction to new measurements of δ¹⁸O in S. florensis insularis tooth enamel recovered from the Pleistocene deposits at Liang Bua, as an indicator of the δ¹⁸O of the animals’ drinking water, using absolute-dated tie-points to synchronise the records. The refined chronology for S. florensis insularis shows that 92% of its remains date to a narrow time-interval equating to 76–62 ka, and that their last appearance at 57 ka coincided with peak aridity. H. floresiensis fossil remains are similarly restricted to 87–61 ka. Together, these findings provide robust evidence for the identification of climate change as a potential driver of the decline of H. floresiensis and S. florensis insularis at Liang Bua after 61 ka.

Fig. 1: Physiographic attributes of the study area.

a Location of the island of Flores, Indonesia. Grey shading shows land exposed above the present-day –80 m isobath at the sea-level lowstand around 65 ka. b Digital elevation map of western Flores with study site locations. c Topographic map showing the proximity of Liang Luar and Liang Bua in the Wae Racang river valley. Symbols indicate auxiliary datasets discussed in the text: archaeological sites with Stegodon tooth enamel isotope data (triangles); rainfall isotope monitoring sites (squares); and modern rainfall analogues for reconstructed changes in monsoon seasonality (circles). Base maps were created using QGIS software and the GEBCO23 grid¹⁰⁸.

Results and Discussion

Liang Luar and stalagmite LR09-K2

The entrance to Liang Luar cave (8° 32’ 10.1” S, 120° 26’ 45.4” E; 570 m.a.s.l.) is only ~660 m from Liang Bua (8° 31’ 50.4” S, 120° 26’ 36.9” E; 560 m.a.s.l.). Both localities are situated within low hills of Late Miocene-Early Pliocene dolomitic limestone³⁸ adjacent to Late Pleistocene alluvial terraces of the Wae Racang river valley (Fig. 1). Stalagmite LR09-K2, the focus of this study, was discovered ~860 m in from the cave entrance as a result of mini-coring and uranium-thorium (U-Th) dating campaigns conducted in 2006 and 2007, and collected in 2009. The straight-line distance between Liang Bua and LR09-K2 is ~1.26 km. LR09-K2 is 1.54 m in length and composed of compact calcite (Supplementary Fig. 1). The chronology for the 50-year resolution Mg-¹⁸O record is well constrained by 32 high-precision U-Th dates, with a median age uncertainty of ±0.29 ky (2σ) (Supplementary Table 1). The record spans 91.4 to 46.8 thousand years before present (ky BP). Mg/Ca and δ¹⁸O were measured on the same precisely dated sample aliquots, ensuring strict temporal alignment of the time series.

Fidelity of the Mg/Ca and δ¹⁸O records

Previous work has shown that Liang Luar is well suited for oxygen-isotope studies^{15,16,17,18,39}. The cave atmosphere is stabilised by narrow constrictions near its entrance and relative humidity near 100% (inactive surfaces are damp) minimises evaporative effects on stalagmite δ¹⁸O. Good replication of contemporaneous stalagmite δ¹⁸O records for Liang Luar demonstrates the lack of karst controls⁴⁰ or in-cave effects, particularly PCP on the cave ceiling^41,42,43, on the δ¹⁸O of infiltrating waters. Supplementary Fig. 2 illustrates the median difference for all 14 coeval stalagmite δ¹⁸O records spanning the last 91 ky (0.09‰; range, 0.00–0.50‰). These differences are small in the context of the 3.5‰ δ¹⁸O range in the K2 record. Further, good reproducibility can only be achieved through consistent recording of the amount-weighted mean annual δ¹⁸O of effective rainfall. No significant difference was observed between the δ¹⁸O values of Liang Luar drip sites (–5.4 to –5.5‰ VSMOW, n = 2) and the amount-weighted averages of pool waters (–5.4 to –5.6‰ VSMOW, n = 5) collected in May 2007 (ref. ⁴⁴). Thus, the dripwater residence times in the ~50–60-m-thick bedrock cap above the deep cave interior appear to be sufficient for homogenisation of the seasonal contrast in rainfall δ¹⁸O along flow pathways.

Figure 2 shows the Mg/Ca and δ¹⁸O time series, corrected for the effect of terrestrial ice volume on seawater δ¹⁸O (δ¹⁸Oiv, see Methods). In general, Mg/Ca and δ¹⁸Oiv increased from 91 to 47 ky BP, with coeval maxima around 58–55 ky BP. At times, Mg/Ca and δ¹⁸Oiv are tightly coupled, especially during abrupt millennial-scale shifts to lower values, and after ~63 ky BP. However, there are swings in δ¹⁸Oiv without Mg/Ca counterparts. Most prominent are divergences to higher values from ~85 to 76 ky BP, and to lower values from ~70 to 62 ky BP. These signals are too large to be attributable to temperature effects on stalagmite Mg/Ca or δ¹⁸Oiv. For example, a 2–3 °C cooling of Indo-Pacific Warm Pool sea-surface temperature (and hence tropical surface air temperature) during glacial times would decrease the stalagmite Mg/Ca by ~0.6 mmol/mol⁴⁵, negating only 4% of the ~15 mmol/mol increase observed from 76 to 55 ky BP. Additionally, the effects of temperature on ¹⁸O/¹⁶O fractionation in tropical rainfall and stalagmite calcite are in opposite directions, and similar in scale, thus they tend to cancel out⁴⁶.

Fig. 2: δ¹⁸O, Mg/Ca, Sr/Ca and PCP results for stalagmite LR09-K2.

a δ¹⁸O adjusted for the influence of ice volume-induced changes in the mean δ¹⁸O of seawater (δ¹⁸Oiv; Methods, Supplementary Fig. 2). Triangles show U-Th dates. b Positive covariation between Mg/Ca (orange) and Sr/Ca (grey). The period of near-constant Mg/Ca and Sr/Ca (91.4–81.8 ky BP, blue line) marks stalagmite calcite deposition from dripwaters derived from dolomitic limestone bedrock near the onset of PCP. The Sr/Ca values for 55.5–46.8 ky BP have been adjusted by +0.0272 mmol/mol to account for a consistent offset between the Mg/Ca and Sr/Ca records (Methods; Supplementary Figs. 3, 4). c Stalagmite Mg/Sr. Sections of the record with near-constant Mg/Sr (296 ± 37, 1σ; grey shading) are indicative of PCP³⁴. The ln(Sr/Ca)/ln(Mg/Ca) slope of 0.85 for the record (excluding 70.3–64.9 ky BP) is diagnostic of PCP³³ (Supplementary Fig. 4). The black curve shows the best fit Mg/Sr profile produced by PCP at a ln/ln slope of 0.85, as determined by Rayleigh fractionation modelling (Supplementary Fig. 5). The blue curve shows the good agreement with Mg/Sr produced by the theoretical PCP ln/ln slope of 0.88 (ref. ³³). d Summary of stalagmite growth rates. The PCP signal is uninterrupted across a broad range of growth rates, but overprinted by anomalously high Mg/Sr around a slow-growth interval centred at 68.6–65.9 ky BP.

Robust evidence for PCP in the stalagmite K2 record is an essential prerequisite for interpretation of the relationship between Mg/Ca and δ¹⁸Oiv. The PCP signature is governed by Rayleigh fractionation where Ca²⁺ decreases exponentially in solution relative to Mg²⁺ and Sr²⁺due to Mg-Ca and Sr-Ca partition coefficients less than one⁴⁷. The associated positive correlation between Mg/Ca and Sr/Ca is a fundamental property of PCP³³. We used two diagnostic criteria to test for PCP: (1) the presence of positive covariation between Mg/Ca and Sr/Ca and near-constant Mg/Sr³⁴ and (2) linear correlation between ln(Sr/Ca) and ln(Mg/Ca) within a slope range of 0.88 ± 0.13 (ref. ³³).

The K2 record shows positive covariation and near-constant Mg/Sr, with the exception of a brief interval of slow growth with anomalously high Mg/Sr (and negative correlation) at 70.3–64.9 ky BP (Fig. 2). The good agreement of the observed ln(Sr/Ca) versus ln(Mg/Ca) slope for the K2 record (0.85, R² = 0.89) with the theoretical slope of 0.88 ± 0.13 confirms that the time series (without 70.3–64.9 ky BP) is governed by PCP (Supplementary Figs. 3, 4). Rayleigh fractionation modelling, based on the algorithm of Wassenburg et al.³⁵, further confirms that PCP is well developed in the K2 record; progressive removal of ~80% of Ca²⁺ from solution by PCP is required to produce the observed range in stalagmite Mg/Ca and Sr/Ca (Supplementary Note 1, Supplementary Fig. 5).

Empirical relationship between Mg/Ca and δ¹⁸Oiv

A key step in the interpretation of the stalagmite K2 Mg-¹⁸O record is to establish the relationship between Mg/Ca, δ¹⁸Oiv and rainfall. At Liang Luar, there is a marked difference of ~7‰ between the amount-weighted average δ¹⁸O for convective summer monsoon rainfall (–9‰ VSMOW) and tradewind-associated dry-season rainfall (–2‰ VSMOW; Supplementary Figs. 6, 7). As dry-season rainfall in southern Indonesia has uniformly high δ¹⁸O (ref. ²³, Supplementary Note 2), we interpret periods of positive covariation between stalagmite Mg/Ca and δ¹⁸Oiv in terms of changes in the fraction of low-δ¹⁸O summer rainfall, which would drive PCP, Mg/Ca and δ¹⁸Oiv in the same direction.

We calibrated the relationship between Mg/Ca and δ¹⁸Oiv in the K2 record over three intervals with consistent positive covariation: 91.4–89.6, 75.5–70.3 and 60.3–46.7 ky BP (Fig. 3). These intervals encompass a broad range of climates, including the Marine Isotope Stage 5 (MIS 5) interglacial, the MIS 5/4 transition with abrupt millennial-scale signals, and the relatively dry MIS 3. The best-fit log-linear relationship for these periods is δ¹⁸Oiv = 4.822*log₁₀(1000Mg/Ca) – 10.56 (R² = 0.90). This Mg-¹⁸O relationship is specific to stalagmite K2 given the likelihood of spatial differences in bedrock Mg/Ca and Sr/Ca and flow paths with unique dripwater residence times, CO₂ degassing histories and solute compositions.

Fig. 3: Mg/Ca and δ¹⁸Oiv systematics for stalagmite LR09-K2.

a Calibration of the exponential relationship between stalagmite Mg/Ca and δ¹⁸Oiv. The best fit is constrained by sections of the records with consistent positive covariation between Mg/Ca and δ¹⁸Oiv (Sections 1–3 in panel c) driven by the common effect of summer rainfall amount on PCP and δ¹⁸Oiv. b Log-linear transformation of the best fit for Mg/Ca and δ¹⁸Oiv. Colour-coded symbols show the good agreement of the best fits for Sections 1–2 (millennial-scale variability, black line) and Sections 1–3 (full data range; R² = 0.90, grey shading shows 95% confidence interval). c Best fit of log₁₀(1000Mg/Ca) and δ¹⁸Oiv through time. The records have been aligned using the log-linear relationship: δ¹⁸Oiv = 4.822*log₁₀(1000Mg/Ca) – 10.56. The grey curve shows calculated Mg/Ca values for the slow-growth interval (70.3–64.9 ky BP) overprinted by non-PCP processes (see Methods).

A notable finding from this analysis of the Mg-¹⁸O system is that δ¹⁸Oiv can be a reliable indicator of local rainfall. Thus, interpretations involving potential competing controls on rainfall δ¹⁸O can be ruled out. Notably, the δ¹⁸Oiv variability in the K2 record cannot be attributed to ‘upstream’ changes in evaporative source regions or the rain-out of ¹⁸O from air masses en route to the cave site, as proposed for the Asian summer monsoon^48,49,50. Instead, rainfall at Liang Luar is sustained primarily by moisture evaporated nearby from the exceptionally warm Java-Flores and Timor-Savu seas^15,18, rather than from distant sources. If rain-out were a dominant factor, the measured δ¹⁸Oiv would reflect the integrated rainfall amount upstream of Liang Luar, with no clear counterpart signal in Mg/Ca.

The systematic relationships among Mg/Ca, δ¹⁸Oiv and local rainfall at Liang Luar require a more direct mechanism related to the so-called amount effect, where the δ¹⁸O of low-latitude rainfall is inversely related to rainfall amount⁵¹. Recent studies have confirmed that the δ¹⁸O of rainfall in tropical Australasia is significantly correlated with the intensity of monsoon convection on monthly to annual timescales^{52,53,54,55,56,57}. In this context, stronger monsoon moisture convergence and convective rainfall lead to higher cloud formations, and thus lower mean condensation temperatures and lower δ¹⁸O for rainfall. Conversely, in winter, orographic rainfall and storms associated with low-level southeast tradewinds and local land breezes form at lower altitudes and higher condensation temperatures, resulting in less rainfall with higher δ¹⁸O^54,55,57. In southern Indonesia, the narrower range of cloud heights in winter makes the δ¹⁸O of winter rainfall less variable than in summer²³ (Supplementary Fig. 7). Thus, the observed connection between δ¹⁸Oiv and the amount effect at Liang Luar likely reflects a lowering of rainfall δ¹⁸O in summer with the development of convective systems associated with the Intertropical Convergence Zone.

Rainfall reconstruction

Figure 4 shows the separation of the K2 Mg-¹⁸O record into summer and winter rainfall in-mixing components. The Mg/Ca profile is dominated by a long-term increase indicating a decline in rainfall from MIS 5 to MIS 3. Initial ²³⁴U/²³⁸U activity ratios [²³⁴U/²³⁸U]i calculated for the U-Th age determinations on stalagmite K2 also increase through time. During drier periods, with reduced karst infiltration rates and longer seepage water residence times, [²³⁴U/²³⁸U]i can increase due to preferential leaching of ²³⁴U from damaged crystal lattice sites in carbonate bedrock⁵⁸. The positive covariation between stalagmite Mg/Ca and [²³⁴U/²³⁸U]i in the K2 record corroborates the general drying trend. Periods of consistent positive covariation between Mg/Ca and δ¹⁸Oiv are driven by the common effect of summer rainfall amount on PCP and δ¹⁸Oiv. Important insights into monsoon seasonality can be obtained at times when the Mg/Ca and δ¹⁸Oiv proxies differ due to gains or losses of winter rainfall, with higher δ¹⁸O. Two distinct periods of divergence are evident; one centred at ~80 ky BP during the MIS 5a interstadial and another at ~66 ky BP during the MIS 4 glacial.

Fig. 4: Effects of summer and winter rainfall on the LR09-K2 Mg-¹⁸O record.

a Stalagmite K2 initial ²³⁴U/²³⁸U activity ratios (with 2σ error bars) as indicators of groundwater residence time and karst dryness at Liang Luar (grey curve is the 3-pt running mean). b The K2 Mg-¹⁸O record. Differences between the Mg/Ca and δ¹⁸Oiv curves mark the opposing effects of changes in winter rainfall. Blue lines show the average δ¹⁸Oiv (–6.2 ± 0.2‰, 2 SE) for the Common Era¹⁸. c δ¹⁸O of the summer rainfall in-mixing component (δ¹⁸Osi) after accounting for the effects of winter rainfall on Mg/Ca and δ¹⁸Oiv (see Supplementary Note 3). d The winter rainfall component of the δ¹⁸Oiv record (Δ¹⁸Ow = δ¹⁸Oiv – δ¹⁸Osi). e-g Schematic examples of the effects of summer rainfall (solid arrows) and winter rainfall (dashed arrows) on Mg-¹⁸O signals: positive covariation with wetter summer (e); negative covariation with wetter winter (MIS 5a, f); and cancelling effect on Mg/Ca with combined effect on δ¹⁸Oiv for equally wetter summer and drier winter (MIS 4, g). Orange and blue shading shows the net effects of winter rainfall on Mg/Ca and δ¹⁸Oiv, respectively.

To translate the stalagmite K2 Mg-¹⁸O record to rainfall amount, we developed a cave dripwater oxygen-isotope mixing model to determine the contributions of summer (December–March) and winter (April–November) rainfall to the amount-weighted average δ¹⁸Oiv (see Supplementary Note 3). The mass balance calculations show that the full range of δ¹⁸Oiv in the K2 record can be described by altering the ratio of summer to winter rainfall in the cave dripwaters. The modelled δ¹⁸O of convective summer rainfall is allowed to vary systematically in accordance with the regional rainfall amount effect²³. To scale the results, the wetter interglacial end of the mixing model is set at the modern average δ¹⁸Oiv at Liang Luar (–6.2‰), corresponding to an annual summer rainfall contribution of 62%. The δ¹⁸O of the non-convective rainfall contribution in winter is set at the regional average value of –2‰^15,23.

With these boundary conditions in place, periods of positive covariation of Mg/Ca and δ¹⁸Oiv in the K2 record can be described by in-mixing of variable amounts of summer rainfall with a fixed amount of winter rainfall. In this case, the observed 3.5‰ increase in δ¹⁸Oiv from MIS 5b to MIS 3 is accounted for by a conservative decrease in summer rainfall (from 68% to 44% of mean annual rainfall, Supplementary Fig. 8). The stability of the relationship between the δ¹⁸O of summer rainfall and rainfall amount through time is evident in lengthy stalagmite Mg-¹⁸O records available for south-western Sulawesi, near Flores. In Sulawesi, stalagmite Mg/Ca and δ¹⁸Oiv are both dictated largely by summer rainfall and define a consistent shift in δ¹⁸Oiv across each glacial-interglacial climate transition (2.7–3.0‰, n = 3)^29,59. The mixing model also accounts for periods when the Mg/Ca and δ¹⁸Oiv records differ due to the opposing effects of winter rainfall on stalagmite δ¹⁸O (Supplementary Fig. 9). Figure 4 shows the calculated summer rainfall in-mixing component for the full record, after accounting for the effects of winter.

Supplementary Note 3 describes the scaling of the Mg/Ca and δ¹⁸Oiv records to conservative estimates of annual, summer and winter rainfall. Briefly, the modern average stalagmite δ¹⁸Oiv (–6.2‰) is set to the modern rainfall seasonality at Liang Luar with summer and winter rainfall at 870 mm (62%) and 530 mm (38%), respectively. The lower limit for summer rainfall in the model (265 mm at δ¹⁸Oiv = –2‰) occurs under conditions of no convective rainfall. The relationship between K2 stalagmite log₁₀(1000Mg/Ca) and δ¹⁸Oiv (defined in Fig. 3) is used to scale Mg/Ca to mean annual rainfall. Importantly, the calibrated relationship between Mg/Ca and rainfall is independent of oxygen-isotope seasonality, whereas δ¹⁸Oiv is sensitive to changes in the proportions of summer and winter rainfall. Thus, the two relationships can be used to determine the combination of summer and winter rainfall that is uniquely compatible with each Mg-¹⁸O data pair in the K2 record (Supplementary Fig. 10).

Modelled mean annual rainfall at Liang Luar declined by ~37% from 1560 mm in MIS 5a (83–76 ky BP ave.) to 990 mm at the onset of MIS 3 (61–55 ky BP ave., Fig. 5). These amounts are 11% higher and 29% lower than at present. Summer rainfall declined by 56% from a maximum of 1030 mm in MIS 5b (91.4–89 ky BP ave.) to a record low of 450 mm at 61–55 ky BP during MIS 3 (48% lower than at present). This long-term decline was punctuated by abrupt increases in summer rainfall during cold Greenland stadials associated with the MIS 5/4 interglacial-glacial transition (~75–70 ky BP), and by significantly wetter summers during MIS 4 (68–65 ky BP). Winter rainfall varied from 810 mm in MIS 5a (53% above modern at 83–76 ky BP) to 340 mm in MIS 4 (36% below modern at 68–65 ky BP).

Fig. 5: History of rainfall amount and seasonality at Liang Luar.

a Modelled mean annual rainfall based on the LR09-K2 Mg/Ca record (see Supplementary Note 3). b Modelled mean monthly rainfall rate for summer (December–March) after accounting for the effects of winter rainfall on Mg/Ca and δ¹⁸Oiv. c Modelled mean monthly rainfall rate for winter (April–November). Dashed lines indicate modern average rainfall values at Liang Luar (1998–2019 CE). GS and H indicate summer rainfall responses to Greenland stadials and Heinrich events. d–g Modern analogue rainfall records for past hydroclimate extremes at Liang Luar (grey shading in panels a–c). d Wet tropical Buranga, south-eastern Sulawesi (4.875° S, 122.875° E) with low seasonality (MIS 5a). e Tropical savanna at Cape Wessel, east Arnhem Land, northern Australia (11.125° S, 136.625° E) with high seasonality (MIS 4). f Subtropical uplands west of Brisbane, eastern Australia (27.625° S, 153.125° E) with low summer rainfall (onset of MIS 3). g Modern rainfall at Liang Luar for reference (8.54° S, 12.45° E). Modern analogue averages for mean annual rainfall (orange lines), summer rainfall (green lines) and winter rainfall (dark-orange lines) are plotted against the proxy records on left (average absolute difference <5%). The best fits with the proxy records were determined using the Global Precipitation Climatology Center (GPCC) monthly rainfall data visualizer (version 2020) based on quality-controlled rain-gauge station records interpolated onto a 0.25° x 0.25° latitudinal-longitudinal grid¹⁰⁹.

To put these results into context, the relatively wet MIS 5a, with low seasonality, is akin to the modern-day rainfall pattern in wet, tropical south-eastern Sulawesi, north of Flores, with a relatively high proportion of winter rainfall (Fig. 5). In contrast, the distinctly seasonal climate during MIS 4 is akin to monsoonal savanna settings in northern Australia with prolonged dry seasons. The driest interval, 61–55 ky BP, with exceptionally low summer rainfall, is similar to subtropical settings adjacent to the east coast of Australia at the southernmost margin of the Australian summer monsoon.

The accuracy of the rainfall estimates can be assessed with climate model simulations available for more recent periods of lowered sea levels (Last Glacial Maximum)⁶⁰ and abrupt climate change (Heinrich Stadial 1)⁶¹. During the MIS 4 glacial (70–60 ky BP), stalagmite K2 indicates an average reduction in mean annual rainfall of 330 mm, relative to modern, which is comparable to the coupled Community Earth System Model version 1.2 (CESM1) result for the Last Glacial Maximum (–360 to –550 mm for the 2° grid including Flores)⁶⁰. The increased reduction in winter rainfall relative to summer during MIS 4 (–0.8 mm d^-1 versus –0.5 mm d^-1) also is consistent with CESM1 (–0.75 to –1.25 mm d^-1 for April–November versus –0.25 to –0.5 mm d^-1 for December–February)⁶⁰. Also, a recent transient CESM1 simulation for Heinrich Stadial 1 (17.5–14.7 ky BP) shows that the rainfall response on Flores was almost entirely restricted to summer⁶¹. The modelled summer rainfall increased by 2–3 mm d^-1 in CESM1, which is essentially the same as the calculated summer increases for the earlier Greenland stadials 20 and 19.2 in the K2 record (2.4–2.6 mm d^-1). The calculated increase in summer rainfall for Heinrich Event 5 is somewhat smaller (1.6 mm d^-1).

Season-specific monsoon forcing mechanisms

We use the difference between the mean monthly rainfall rates for summer and winter (divided by the annual mean monthly rainfall rate)⁶² as a simple index of past rainfall seasonality at Liang Luar. The index reflects the relative intensity of summer rainfall compared with the mean annual rainfall. The modern-day seasonality value for Liang Luar is 1.3. The stalagmite K2 record yields a wide range of seasonality values, with the lowest during the relatively wet MIS 5a (0.7 at 83–76 ky BP, 46% below modern) and the highest during MIS 4 (1.7 at 68–65 ky BP, 31% above modern). Figure 6 shows the similarity of the seasonality index and stalagmite K2 δ¹⁸Oiv. As modelled, the δ¹⁸Oiv record is controlled by the relative contribution of summer rainfall to the amount-weighted average δ¹⁸O of cave dripwaters. In this respect, it represents an approximate measure of the relative intensity of local summer monsoon rainfall through time.

Fig. 6: The stalagmite K2 rainfall records in the context of season-specific monsoon forcing mechanisms.

a Comparison of the reconstructed monsoon seasonality index and stalagmite K2 δ¹⁸Oiv. Both records provide a measure of the relative intensity of local summer monsoon rainfall. Dashed line indicates the modern average seasonality (1998–2019 CE). b Phasing of the modelled mean annual rainfall at Liang Luar with sea level and sea-surface temperature (SST). For reference, the global mean sea-level record and associated 2σ confidence interval, as determined by Waelbroeck et al.¹⁰⁴, and the marine sediment Mg/Ca-SST reconstruction for ODP Hole 806b in the western equatorial Pacific (159° 22’ E, 0° 19’ N)⁴⁵. c Millennial-scale phasing of the modelled summer rainfall (December–March) with the Greenland (NGRIP) ice-core δ¹⁸O record⁷⁴, plotted on the GICC05modelext time scale¹¹⁰, and the U-Th-dated composite speleothem δ¹⁸O records for Sanbao–Hulu China^36,72,73 and northern Borneo^12,75, adjusted to account for the effect of ice volume. The pink bar indicates the northward shift of the monsoon rain belt near the onset of MIS 3, and summer aridification on Flores. Tan bars mark Greenland stadials¹¹¹ with measurable effects on the Australasian monsoon. Green dots correspond with increases in summer rainfall on Flores beginning mid-way through strong NGRIP interstadial-stadial transitions (see text). The lowermost curve shows orbital precession-scale phasing of the modelled winter rainfall (April–November) with northern summer insolation (June–August at 30° N, black curves), supporting the East Asian summer monsoon strength^36,72,73 as a potential driver of winter tradewind rainfall on Flores.

When viewed in the context of previous work, the season-specific rainfall reconstruction for Liang Luar helps clarify large-scale mechanisms that influenced the Australasian monsoon on glacial-interglacial, precessional and millennial scales. As an approximation, Fig. 6 shows the likely role of lower sea levels in the reduction of mean annual rainfall across the MIS 5/4 transition, consistent with regional hydroclimate reconstructions^14,29 and model studies of the effect of Sunda–Sahul shelf exposure on atmospheric convection in Island Southeast Asia^60,63. Also, as a principal source of water vapour to the atmosphere, the concurrent ~2 °C decrease in Indo-Pacific Warm Pool sea-surface temperatures⁴⁵ and associated reduction in atmospheric humidity may have steepened the decline in rainfall.

At the orbital precession scale, a key finding is that rainfall seasonality on Flores was driven primarily by changes in winter rainfall rather than summer rainfall (Fig. 6). The two distinct periods of anomalous winter rainfall centred at ~80 ky BP and ~66 ky BP are opposite in direction and separated by about half of a 23-ky precession cycle. This phasing reduced rainfall seasonality at Liang Luar during MIS 5a and amplified it during MIS 4. Previous studies have shown that precession-scale variations of the Australian–Indonesian winter monsoon are in phase with the northern summer monsoon, owing to their linkage via cross-equatorial flow of surface winds^64,65,66. In this framework, the increased winter rainfall in Flores during MIS 5a likely reflects enhanced moisture transport en route from the Australian high-pressure belt in response to higher northern summer insolation over the Asian landmass. Conversely, during MIS 4, lower northern summer insolation would have weakened the interhemispheric thermal gradient and cross-equatorial flow. Regional continental shelf exposure during the –80-m sea-level lowstand in MIS 4 (ref. ⁶⁷) may have further reduced northward moisture advection from the Timor Sea and winter rainfall in Flores^63,68.

The proxy evidence for latitudinal displacement of the tropical monsoon rain belt on millennial scales is well established^69,70,71. The season-specific rainfall record for Liang Luar helps confirm the observed antiphase behaviour of the Indonesian–Australian^17,18 and East Asian summer monsoons^36,72,73. A prime example is the coordinated northward displacement of the Australasian monsoon rain belt at the MIS 4/3 transition ~60 ky BP evident in speleothem records (Fig. 6). Palaeoclimate records and model experiments suggest that the expansion of ice sheets and sea-ice at high northern latitudes during MIS 4 would have increased the interhemispheric temperature gradient and displaced the Australasian monsoon rain belt to the south^70,71. The north-to-south coherence of the North Greenland ice-core δ¹⁸O record⁷⁴ and the speleothem δ¹⁸O records for China^36,72,73, Borneo^12,75 and Flores supports this mechanism. In contrast, the records reveal a coordinated northward displacement at the onset of MIS 3 when the Northern Hemisphere warmed abruptly, resulting in wetter summer monsoons in China and summer aridification in Flores.

On closer inspection, the Liang Luar rainfall reconstruction supports the idea of northern winter cooling, and strengthening of the Asian winter monsoon, as a season-specific mechanism driving increases in summer rainfall in Flores^76,77 (Fig. 6). The three coherent reductions in both Mg/Ca and δ¹⁸Oiv during the prominent stadial events around the MIS 5/4 transition (events 21, 20, 19.2) are due entirely to increases in summer rainfall over Flores. A key observation is that the summer rainfall responses in Flores began about mid-way through the strong interstadial–stadial transitions in the Greenland ice-core δ¹⁸O record, and led the weakening of the summer monsoon in China. The result is consistent with interpretation of the ice-core isotopic signature as largely an expression of cooling in northern winter⁷⁶, leading to exceptionally cold stadial winters across the Eurasian landmass⁷⁷. In this case, the summer rainfall signal in Flores can be viewed as an incipient southward expansion of the Australasian monsoon rain belt triggered by a cross-equatorial ‘pressure-push’ from the Asian winter monsoon^13,78. From this perspective, the subsequent weakening of the summer monsoon in China at peak stadial conditions signals year-round cooling in eastern Asia and wholesale southward displacement of the Australasian monsoon.

Decline of Homo floresiensis and Stegodon alongside climate change

Homo floresiensis and Stegodon coexisted on Flores for at least one million years based on the fossil records from Liang Bua and the So’a Basin, ~80 km east of Liang Bua^79,80,81,82. As obligate drinkers, like modern humans and proboscideans^9,10, both species would have been critically dependent on surface freshwater to survive the high-amplitude climate changes that characterise glacial-interglacial cycles. Previous work has identified a pattern of drier climate conditions accompanying the last appearances of Stegodon sondaari at Tangi Talo (~1.4–1.27 Ma), the oldest fossil site in the So’a Basin, and Stegodon florensis at Mata Menge (~0.77–0.7 Ma)^79,81,83, based on analysis of δ¹⁸O in the carbonate phase of Stegodon fossil tooth enamel from the two sites (Fig. 7). Tooth enamel δ¹⁸O reflects the mean annual δ¹⁸O of the animals’ drinking water, which in turn reflects the site-specific δ¹⁸O of rainfall, modified by surface evaporation^84,85,86. The Stegodon enamel from the upper occupation levels at Tangi Talo and Mata Menge shows significant increases in δ¹⁸O (1.3–1.9‰) relative to stratigraphically older material. Tandem measurements of δ¹³C in the same tooth samples reveal no significant flexibility in Stegodon feeding behaviour in response to climate change^79,81,83. At Tangi Talo, S. sondaari maintained its preference for a mixture of leafy C₃ plants and C₄ grasses, and at Mata Menge S. florensis continued to specialise in C₄ grazing. The cause of the extinction of S. sondaari between ~1.27 Ma and 0.9 Ma remains unknown⁸¹, but S. florensis ultimately led to its dwarfed insular descendent, Stegodon florensis insularis, which represents the last known occurrence of Stegodon on Flores at ~50 ka^4,5,6.

Fig. 7: δ¹⁸O and δ¹³C in fossil Stegodon tooth enamel from the So’a Basin and Liang Bua.

a Partial stratigraphic summary of Liang Bua with approximate ages (after Sutikna et al.^4,6). The presence of S. florensis insularis and H. floresiensis is indicated with blue bars (dotted where inferred). b Stratigraphic trends in the δ¹⁸O of Stegodon tooth enamel. Boxplots (right) show offsets in δ¹⁸O between the older and younger stratigraphic phases of S. sondaari at Tangi Talo (~1.4 Ma versus 1.27 Ma) and S. florensis at Mata Menge (~0.77 Ma versus 0.7 Ma)^79,81,83 in the So’a Basin. Increases in mean δ¹⁸O (1.3‰ and 1.9‰) preceding the disappearances of Stegodon are defined by 95% confidence intervals (±2 SE, coloured bars); black lines indicate 2σ ranges. On the left, S. florensis insularis tooth enamel δ¹⁸O versus excavation depth in Sectors XI (orange), XV (mauve) and XVI (yellow) at Liang Bua⁸³. The linear fit of the tooth δ¹⁸O data with the U-Th dated stalagmite K2 δ¹⁸O record (blue, not corrected for ice-volume effects) provides an absolute time scale for the fossil-bearing strata below volcanic tephra T1^4,5,6 (see Methods). Vertical black bars indicate δ¹⁸O ranges for three excavation intervals containing 2–3 tooth samples: grey diamonds mark three outliers. The estimated age for tephra T1 (dashed vertical line at 56 ky BP) lies between the previously assigned age (60 ka, black triangle)⁴ and the U-Th date for the oldest overlying flowstone (54.4 ± 0.3 ka, 2σ; red bar)⁴. The age-depth relationships for two U-Th dated H. floresiensis ulnae (LB1, LB6) from Sector XI (orange squares, weighted means ±2 SE)⁴ and one thermoluminescence date for Sector XV (mauve circle, mean ±2σ)⁴ are shown for reference. c The δ¹³C of Stegodon tooth enamel as an indicator of dietary preferences.

New δ¹⁸O and δ¹³C measurements on S. florensis insularis tooth enamel from the Pleistocene deposits at Liang Bua⁸³, together with the stalagmite K2 Mg-¹⁸O record, provide a precisely dated chronological framework for examining the disappearance of Stegodon and H. floresiensis alongside major climate changes (Figs. 7, 8). The analysis is based on 29 tooth samples collected from the fossil-bearing colluvial deposits within stratigraphic Unit IB, dated to approximately 120–60 ka⁴ (Methods). The majority of the samples (27) are from overlapping depositional intervals across three adjacent excavation sectors (Supplementary Fig. 11, Supplementary Table 2). Figure 7 shows the marked similarity of the tooth enamel δ¹⁸O depth profile and the precisely dated stalagmite K2 δ¹⁸O record. Both datasets share three diagnostic features: low δ¹⁸O values during MIS 5b (92–85 ky BP); a gradual increase during MIS 5a (85–71 ky BP); and an abrupt return to lower δ¹⁸O during MIS 4 (71–62 ky BP). The best-fit linear age-depth relationship indicates that the teeth in this depositional interval span ~91 to 61 ky BP, within the reported age-range for Unit 1B⁴. With the exception of two individuals, the δ¹³C results show that, like its predecessor, S. florensis insularis maintained a dietary preference for C₄ grasses regardless of climate change.

Fig. 8: Chronology of climate change, Stegodon florensis insularis and Homo floresiensis at Liang Bua.

a–c Reconstructed mean annual rainfall, mean monthly rainfall rate in summer (December–March) and rainfall seasonality at Liang Bua. Dashed lines indicate modern average climatic values (1998–2019 CE). Coloured bars mark three distinct hydroclimate phases: (1) wetter with low seasonality (91–76 ky BP); (2) variable with high seasonality (76–61 ky BP); and (3) year-round dryness (61–47 ky BP). d Stalagmite K2 [²³⁴U/²³⁸U]i values (with 2σ error bars) as indicators of groundwater residence time and karst dryness (grey curve is the 3-pt running mean). e Distribution of S. florensis insularis fossil remains in excavation Sector XI (grey bars) and Sector IV (red curve) at Liang Bua⁵ scaled to time using the oxygen-isotope age-depth relationship in Fig. 7: age = 9.11*depth (m) + 28.2. Sector IV is ~11 m from Sector XI and has been adjusted by –0.2 m depth (1.8 ky to younger) to align the fossil-rich deposits (see Supplementary Figs. 11, 12). The age-depth relationship indicates that 92% of the Stegodon remains in these sectors occur between 76 and 62 ky BP (Phase 2), with the youngest dating to 57 ky BP. The pink bar represents volcanic tephra T3 (ref. ⁴). f Distribution of associated H. floresiensis remains in Sectors XI (grey) and IV (red). Significant specimens are identified by published “LB” numbers^88,112. g Schematic summary of dry-season water stress at Liang Bua (blue, green, yellow, orange, red indicate lowest to highest water stress). Line C1 marks the onset of the driest climate on record (61–55 ky BP) with a significant increase in [²³⁴U/²³⁸U]i indicating drying of the karst during the final phase of S. florensis insularis and H. floresiensis.

The similarity of the K2 stalagmite δ¹⁸O and S. florensis insularis tooth δ¹⁸O records could only occur if the isotopic composition of Stegodon’s drinking water was concordant with the infiltrating waters feeding-in to the stalagmite. Only three tooth samples were isotopic outliers, indicating that these individuals obtained most of their body water from divergent sources (Fig. 7). The results suggest that the majority of individuals were reliant on the Wae Racang river as the primary source of isotopically homogenised drinking water near Liang Bua. In this case, the mean δ¹⁸O of tooth enamel would be set by the δ¹⁸O of groundwater-fed river baseflow for most of the year⁸⁷, and periodic runoff in summer. Analysis of numerous S. florensis insularis remains at Liang Bua indicates that H. floresiensis subsisted in part by hunting or scavenging primarily subadult individuals, which make-up 94% of the fossil assemblage⁵. Thus, young individuals unable to wander far in search of alternative freshwater sources during dry periods may have been particularly vulnerable to predation. Even in the present-day climate, baseflow in the small Wae Racang rainfall catchment does not persist year-round during drier years.

The refined U-Th age-depth relationship for Unit 1B clarifies the timing of S. florensis insularis and H. floresiensis at Liang Bua. Figure 8 shows results for two well-documented, fossil-rich excavation sectors: Sector XI, located near the east wall of the cave⁸⁸, and Sector IV, approximately 11 m westward near the cave centre² (see Supplementary Figs. 11, 12). A key finding is that 92% of the 716 Stegodon remains within the age-depth. calibration window occur within a narrow depth interval (5.25–3.65 m) equating to 76–62 ky BP (Supplementary Table 3). In contrast, only 10 Stegodon specimens post-date 62 ky BP, with the youngest, at 3.25–3.15 m in Sector XI, dating to 57 ky BP. H. floresiensis skeletal remains, while much less common, are last recorded at 61 ka. The distribution of Stegodon in sectors XI and IV is consistent with their broader spatio-temporal distribution in 12 excavation sectors at Liang Bua⁶. Of the 10,656 skeletal elements attributed to Stegodon in Units 1B and 2, only 23 occur in Unit 2 between volcanic tephra layers T1 and T3 ( ~60–50 ka). Also, Stegodon constitutes 6.8% of the Unit B fossil assemblage, a markedly higher proportion than the 0.6% identified in Unit 2.

The preservation of subadult S. florensis insularis remains in association with stone tools has been attributed to H. floresiensis as the primary agent responsible for their accumulation^2,5. Hence, the concentration of their remains reflects the intensity of H. floresiensis occupation activity at Liang Bua⁸⁹. The decline in Stegodon abundance underway by ~64 ky BP in Sectors XI and IV, along with their near-absence in Unit 2, may signal the onset of a gradual abandonment of the site (Fig. 8). This decline began well before the deposition of the prominent volcanic tephra layer T3 dating to ~50–47 ka⁴. Thus, while local volcanism and possible water supply contamination may have contributed to the final abandonment of Liang Bua, the adaptive challenge posed by persistent aridification and reduced freshwater availability after 61 ky BP may have taken precedence.

The stalagmite K2 palaeomonsoon reconstruction identifies three distinct season-specific hydroclimate phases, each with implications for dry-season water stress, and thus the potential for enhanced human-faunal interactions around dwindling river baseflows (Fig. 8). During Phase 1 (92–76 ky BP), mean annual rainfall exceeded present-day levels and seasonality was low. Using the current climate as a guide, the favourable MIS 5 rainfall regime would have sustained perennial baseflow in the Wae Racang river, thus freshwater availability was likely unconstrained. Phase 2 (76–61 ky BP) was characterised by drier conditions punctuated by abrupt increases in summer rainfall at 74, 71 and ~68–65 ky BP during Greenland stadials and MIS 4. Higher rainfall seasonality and dry-season aridity at ~70–61 ky BP may have promoted the development of ‘sub-tropical’ open-grazing habitats supportive of S. florensis insularis. This hydroclimate scenario aligns with the fossil evidence indicating increased interaction between H. floresiensis and S. florensis insularis, perhaps as they aggregated near increasingly restricted, but still reliable, riverbed water resources replenished by high summer rainfall. Phase 3 (61–47 ky BP) includes the onset of the relatively dry MIS 3, marked by persistently low mean annual rainfall and minimum values in summer at 61–55 ky BP. However, unlike Phase 2, the low summer rainfall likely led to critically low river baseflows. The ensuing rise in stalagmite [²³⁴U/²³⁸U]i beginning at ~55 ky BP reflects the impact of aridity on groundwater residence times within the surrounding karst (Fig. 8).

Our findings point to climate change, potentially in combination with predator-prey dynamics, as a contributor to the decline of H. floresiensis and S. florensis insularis prior to their disappearance at Liang Bua. A plausible scenario is that increasing dry-season water stress forced stegodonts to migrate to better-watered habitats⁹⁰, followed by H. floresiensis in search of their preferred prey. In this case, the near absence of both species at Liang Bua after ~61 ky BP may simply reflect their relocation to more favourable habitats. On the other hand, if S. florensis insularis was compromised by limited freshwater availability, even low intensity predation of the subadults could have triggered a critical population decline^8,91, ultimately leading to abandonment of the site by H. floresiensis. Whether H. floresiensis and S. florensis insularis interacted with modern humans (Homo sapiens) elsewhere on Flores remains uncertain. Fossil evidence places modern humans on an eastward dispersal trajectory from continental Asia (Sunda) by ~75–63 ka⁹², with initial arrival in Sahul by ~70–60 ka⁹³. The southern dispersal route is aligned with the Lesser Sunda Islands, including Flores^94,95. Thus, the onset of the apparent abandonment of Liang Bua beginning early in hydroclimate Phase 3 allows for potential overlap with migrating modern humans. Our precisely dated rainfall reconstruction provides a firm chronological framework for future studies to investigate the extinctions of these iconic species in the context of major climate change.

Methods

U-Th dating and age model

Reconnaissance U-Th dating of mini-cores extracted discretely from the bases and tops of Liang Luar stalagmites in situ was performed at the University of Queensland, Australia, before we collected LR09-K2. This was done to optimise the age-range of the specimens while minimising the number collected. The LR09-K2 chronology is based on 32 U-Th dates determined by multi-collector inductively coupled plasma mass spectrometry (MC-ICP-MS) at the University of Minnesota, USA and the University of Melbourne, Australia (Supplementary Table 1). Solid samples for dating were extracted adjacent to the Mg-¹⁸O sampling track along the main growth axis of LR09-K2 using a small, diamond-coated circular-saw blade fitted to an air drill. The calcite blocks, weighing 190–300 mg, were dissolved and equilibrated with a mixed ²²⁹Th–²³³U–²³⁶U tracer before U and Th were chemically separated and purified⁹⁶. The MC-ICP-MS analytical procedures performed in the two laboratories are broadly similar^97,98 and yielded consistent results. U-Th ages were calculated using the half-lives for ²³⁴U (245,620 ± 70 years) and ²³⁰Th (75,584 ± 30 years) determined by Cheng et al.⁹⁸, and converted to ky BP (with ‘present’ being 1950 CE).

As all of the samples have high ²³⁰Th/²³²Th values, the effect of detrital ²³⁰Th on the calculated ages is very small. We determined corrected U-Th ages using the bulk Earth ²³⁰Th/²³²Th activity ratio of 0.82, with an arbitrary uncertainty of 50%. The median age adjustment is 10 years to younger. Application of a large ²³⁰Th/²³²Th activity ratio of 7 ± 2, the maximum used in previous studies of Liang Luar speleothem material^15,17,18, shifted the median age by only 87 years to younger. The corrected U-Th ages from both laboratories have a fully propagated 2σ median uncertainty of ±287 years.

The age-depth model for the LR09-K2 record was calculated using the OxCal (version 4.3) module, a Bayesian chronological modelling tool that employs Markov chain Monte Carlo^99,100. Three of the 35 available U-Th ages (LR09-K2-05, LR09-K2-M9, LR09-K2-M19) were slightly inconsistent with the age-depth trend defined by surrounding dates, and were excluded from the age model (Supplementary Fig. 1, Supplementary Table 1). The ages of five sample pairs analysed in the two laboratories agree within measurement uncertainty and were averaged into single age determinations, leaving a total of 27 precise control points to anchor the age model.

OxCal accounts for the non-uniform nature of stalagmite growth between U-Th age control points using a P-sequence Poisson-process depositional model. The model was run with a prior outlier probability of 0.05 using ko = 1 mm^-1, k = 0.06 mm^-1 and U(2,2). The interpolation rate value (ko) determines the rigidity of the model and enables trade-off between less interpolated age uncertainty (high ko, high rigidity) and stricter adherence to U-Th ages (low ko, low rigidity). As a ko value of 1 mm^-1 gives a relatively flexible age model, we consider the confidence envelope to be highly conservative. The age-depth model was interpolated linearly to the depths of individual isotope measurements.

Oxygen isotope analysis

In preparation for δ¹⁸O analysis, the 1.54-m-long stalagmite LR09-K2 was cut into manageable 300–400 mm sections at 45° angles perpendicular to the intended sampling slab (Supplementary Fig. 1). Contiguous 20-mm-thick slabs were then cut using a water-cooled diamond saw. The slabs were polished, and sampling transects were positioned on the central growth axis and offset slightly across each 45° cut to ensure temporal overlap and time series continuity. Powder samples for δ¹⁸O analysis were then milled in a narrow pre-cleaned groove (1-mm depth and width). Sampling was continuous to simplify recovery of the milled powder and prevent aliasing in the record.

The average growth rate for LR09-K2 from 91.4 to 46.8 ky BP is 3.45 mm/100 years, based on the OxCal age model. The sampling interval was set at 1.5 mm, which yielded an average age resolution of 43 years (10–90% quantile range = 10–72 years) for a total of 1026 δ¹⁸O determinations. Sub-samples of the calcite powders were weighed to 180–220 μg and analysed for δ¹⁸O at the Research School of Earth Sciences, ANU on a Finnigan MAT-251 dual-inlet mass spectrometer equipped with an automated Kiel I carbonate reaction device. CO₂ was liberated for isotopic analysis by reaction under vacuum with anhydrous 105% H₃PO₄ at 90 °C. The results were normalised to the Vienna Peedee Belemnite (VPDB) scale such that in-run measurements on the NBS-19 and NBS-18 standards yielded values of −2.20‰ and −23.0‰ for δ¹⁸O, respectively. The original delta values for NBS-19 and NBS-18 were used to maintain the consistency of results through time in the, RSES, ANU Stable Isotope Facility. The analytical precision (2σ) for 232 in-run measurements of δ¹⁸O on the primary standard, NBS-19, was ±0.08‰.

Thirty-eight additional samples were milled continuously at higher resolution (0.4–0.5 mm) to obtain an age resolution of ~83 years (versus ~276 years) across a slow-growing (0.37 mm/100 years) section of LR09-K2 spanning 68.79–65.64 ky BP. A sideways clear-cut was made at each increment to remove overlap caused by the curvature of the 1-mm-diameter end-mill bit. These sub-samples were weighed to 110–160 μg and analysed for δ¹⁸O at RSES, ANU on a Thermo MAT-253 mass spectrometer coupled to a Kiel IV carbonate device operating at 70 °C. The analytical precision (2σ) for 12 in-run measurements of δ¹⁸O on NBS-19 was ±0.10‰. The mean δ¹⁸O value for coeval data produced by the MAT-251 and MAT-253 is in excellent agreement (MAT-253 is only 0.03‰ lower). Based on this result, the 38 high-resolution data points were substituted into the time series at 68.79–65.64 ky BP. Due to sample-size limits, 11 measurements of Mg/Ca and Sr/Ca made at 1.5 mm resolution from 68.79 to 65.64 ky BP were linearly interpolated to 38 data-points to match the δ¹⁸O.

Mg/Ca and Sr/Ca analysis

Mg/Ca and Sr/Ca were measured on splits of the sample powders analysed for δ¹⁸O using simultaneous inductively coupled plasma atomic emission spectroscopy (ICP-AES) at RSES, ANU and the Australian Nuclear Science and Technology Organisation (ANSTO). At ANU, 257 initial measurements on every fourth sample in the record were made following the reference solution data normalisation procedure of Schrag¹⁰¹. Then, 639 measurements were made at ANSTO following the intensity ratio calibration method of de Villiers et al.¹⁰². The resulting dataset provides full resolution of the millennial variability from 84 to 46.8 ky BP. Every second sample was measured in the relatively fast-growing section of LR09-K2 predating 84 ky BP.

At ANU, sample aliquots weighing ~0.5 mg were transferred to acid-washed vials and dissolved in 5 mL of 2% (v/v) HNO₃ to yield a consistent Ca concentration of 39 ± 5 ppm (2σ). One blank vial with 2% HNO₃ was analysed at the start of each sample batch. Measurement of a Ca matrix-matched ‘bracketing’ reference solution in 2% HNO₃ was performed to standardise the Mg/Ca and Sr/Ca of each sample sequentially¹⁰¹. The elemental ratios of the reference (Mg/Ca = 5.754 mmol/mol, Sr/Ca = 8.983 mmol/mol) were determined by thermal ionisation mass spectrometry at RSES, ANU. The median standard error of the means for the reference solution pairs (n = 258) bracketing the samples was 0.004 mmol/mol for Mg/Ca and 0.0033 mmol/mol for Sr/Ca. Ten replicate measurements were made on each sample solution and the normalised (blank-corrected) averages are reported for the result.

At ANSTO, ~1 mg sample aliquots were dissolved in 5 mL of 3% (v/v) HNO₃ to yield a consistent Ca concentration of 76 ± 6 ppm (2σ). Four blank vials with 3% HNO₃ were analysed at the start of each sample batch, and a Ca matrix-matched reference solution was run between blocks of five samples to correct for any instrument drift. The in-run precision (1σ) for 132 corrected measurements of the reference Mg/Ca (19.992 mmol/mol) and Sr/Ca (0.1025 mmol/mol) was 0.040 mmol/mol (0.2%, RSD) and 0.0003 mmol/mol (0.3%, RSD), respectively. To ensure accuracy, in-run measurements were performed on three Ca matrix-matched standard solutions prepared gravimetrically from certified single-element National Institute of Standards and Technology (NIST) reference materials. The solutions covered the range of values in stalagmite LR09-K2 (Mg/Ca, 4–30 mmol/mol; Sr/Ca, 0.07–0.5 mmol/mol). The results were calculated by regressing the measured ICP-AES Mg/Ca and Sr/Ca intensity ratios (corrected for blank intensity) against the Mg/Ca and Sr/Ca (mmol/mol) of the standard solutions¹⁰².

Agreement between the two laboratories is excellent for Mg/Ca (the primary element of interest): in-run measurements on the JCp-1 standard yielded average Mg/Ca values of 4.248 mmol/mol (ANU) and 4.246 mmol/mol (ANSTO), both within the 1σ uncertainty of the certified value (4.199 ± 0.065 mmol/mol)¹⁰³. However, the average Sr/Ca value of the ANU data is consistently offset by +0.0307 mmol/mol relative to the fully calibrated ANSTO data. The Sr concentration of the ANU ‘coral’ reference solution bracketing each sample is ~180 times higher than that for the stalagmite K2 calcite samples, thus the instrument rinse time was probably too brief to clear the ICP spray chamber of Sr prior to the onset of each low-Sr sample measurement. Nevertheless, the offset between the two datasets is remarkably consistent, and subtraction of 0.0307 mmol/mol from the ANU dataset produced excellent fine-scale agreement.

Data refinement

To compare the stalagmite trace element and δ¹⁸O time series, both datasets were binned at 50-year resolution. Prior to binning, the data were interpolated at 1-year resolution in MATLAB using a piece-wise cubic Hermite spline, creating a weighted average throughout each 50-year bin to ensure equal representation of data points. The 1-year data were averaged to the mid-point of each 50-year bin and then adjusted to account for changes in the mean δ¹⁸O of seawater due to changes in terrestrial ice-volume. We use the global ocean δ¹⁸O reconstruction based on a composite benthic foraminiferal δ¹⁸O record corrected for changes in deep-ocean temperature¹⁰⁴, with a 0.085‰ change in seawater δ¹⁸O equating to 10 m sea level. The 1500-year resolution ocean δ¹⁸O record was interpolated into 50-year bins and smoothed with a 1,550-year moving average (50-year steps) prior to the adjustment.

In Fig. 2, the Sr/Ca data for 55.5–46.8 ky BP have been adjusted by +0.0272 mmol/mol to account for a consistent offset relative to Mg/Ca. This interval shows good evidence for PCP (near-constant Mg/Sr, ln/ln slope 0.84; Supplementary Figs. 3, 4), but the cause of the offset is unknown. A switch to prior aragonite precipitation due to aridity (and high Mg/Ca in dripwaters) around 55.5 ky BP would decrease Sr/Ca, but with negative covariation with Mg/Ca¹⁰⁵. Laboratory experiments have shown that the Sr-Ca distribution coefficient in calcite is growth-rate dependent¹⁰⁶, but the growth of stalagmite K2 did not change notably around 55.5 ky BP. Evaluation of potential changes in non-carbonate sources of Sr due to changes in vegetation productivity and Sr availability from soil¹⁰⁷, or perhaps volcanic activity near Liang Luar after ~60 ka⁴,  is beyond the scope of this study.

Mg/Ca data points have been reconstructed across the interval of anomalously high Mg/Sr spanning the first half of MIS 4 (70.3–64.9 ky BP), where the underlying PCP signal has been overprinted by non-PCP processes. However, the Mg/Ca and δ¹⁸Oiv are well correlated before 70.3 ky BP, and also during the second half of MIS 4 (64.9–60 ky BP), when the PCP signal is uninterrupted. Furthermore, the pattern of change in δ¹⁸Oiv across the full length of MIS 4 is symmetrical. Therefore, we applied the best fit relationship between Mg/Ca and δ¹⁸Oiv for the second half of MIS 4 [log₁₀(1000Mg/Ca) = 0.0845*δ¹⁸Oiv + 1.727 (R² = 0.67)] to infill PCP-compatible Mg/Ca data across the first half of MIS 4.

Analysis of modern rainfall and oxygen-isotope climatology

Establishing the relationship between Mg/Ca and δ¹⁸Oiv in stalagmite LR09-K2 and past rainfall requires site-specific information on modern mean annual rainfall, rainfall seasonality and the δ¹⁸O of rainfall. Supplementary Note 2 describes our analysis of the modern rainfall and oxygen-isotope climatology. Briefly, in 2006–2007, the isotopic composition of rainfall was monitored at a collection site established ~700 m from Liang Luar and Liang Bua (Fig. 1). The rainfall amounts and δ¹⁸O and δ²H values determined for 39 days with rainfall are summarised in Griffiths et al.¹⁵. We scaled the short-term dataset to the longer-term rainfall climatology of western Flores using data from the NASA Tropical Rainfall Measuring Mission (TRMM) and rain-gauge data from the Global Historical Climatology Network (GHCN). The mean annual rainfall at Liang Luar was determined to be 1,405 ± 33 mm (2 SE) with a summer (December–March) contribution of 61.7 ± 0.6% (2 SE) (Supplementary Fig. 6). Given the uncertainties in this estimate, we rounded the result to 1,400 mm and allocated 870 mm (62%) to summer and 530 mm to the rest of the year (‘winter’, April–November) to scale the rainfall history for Liang Luar.

Supplementary Fig. 7 shows that the rainfall seasonality and δ¹⁸O seasonality at Liang Luar and the neighbouring islands of Bali and Sulawesi²³ are closely linked. The divisions between summer and winter, as defined here, align with conspicuous shifts in rainfall δ¹⁸O marking the start and end of summer monsoon convection. The cave dripwater oxygen-isotope mixing model described in Supplementary Note 3 was scaled to summer and winter rainfall amounts on the grounds that the good reproducibility of Liang Luar stalagmite δ¹⁸O records (Supplementary Fig. 2) reflects consistent recording of the effect of changes in rainfall seasonality on the amount-weighted mean annual δ¹⁸O of the cave dripwaters.

Stable isotope analysis of Stegodon tooth enamel

The analytical approach for measurements of δ¹⁸O and δ¹³C in the carbonate phase of Stegodon fossil tooth enamel samples from the So’a Basin is described elsewhere^79,81,83. The new measurements of δ¹⁸O and δ¹³C reported here for S. florensis insularis at Liang Bua were conducted on 29 tooth samples from five excavation sectors within the cave (Supplementary Table 2)⁸³. Prior to sampling for isotopic analysis, the outer tooth surfaces were cleaned using a rotary hand-held device to minimise contamination. Approximately 100 mg of powdered enamel was then obtained using a diamond-coated burr on a Dremel hand drill. The enamel was drilled across the tooth (or tooth fragment) from top to bottom to obtain an annual average isotope signal. The resulting powder was pre-treated with 2% sodium hypochlorite to oxidise organic residues and 0.1 M acetic acid to remove exogenous carbonates, and rinsed with deionised water and dried following each chemical treatment. Samples were analysed for δ¹⁸O and δ¹³C at the School of Earth, Atmospheric and Life Sciences Geochemical Laboratory at the University of Wollongong, Australia. Approximately 5 mg of pre-treated powder was reacted under vacuum with 105% H₃PO₄ at 90 °C in a MultiPrep carbonate device. The liberated CO₂ was analysed using a Micromass PRISM III dual-inlet mass spectrometer. Data were normalised to the VPDB scale via NBS-18 (δ¹⁸O, –23.01‰; δ¹³C, –5.04‰), NBS-19 (δ¹⁸O, –2.20‰; δ¹³C, +1.95‰), NBS-20 (δ¹⁸O, –4.14‰; δ¹³C, –1.06‰) and in-house laboratory standards.

The δ¹⁸O results for 27 tooth samples collected across three adjacent excavation sectors (XI, XV, XVI) near the east wall of Liang Bua were used to establish the relationship between the tooth δ¹⁸O depth-series and the U-Th dated δ¹⁸O record for stalagmite K2 (Fig. 7). These samples span overlapping intervals within a large depth-range (7.05–3.65 m) with clear stratigraphic continuity (Supplementary Fig. 11). A linear age-depth fit linking the tooth δ¹⁸O and stalagmite δ¹⁸O was identified following three adjustments to the tooth dataset: (1) subtract 0.2 m from the excavation depths in Sector XVI to account for the gentle slope of the fossil-bearing strata and align the δ¹⁸O with coeval values in Sector XV; (2) minimise the difference between the tooth and stalagmite records by linear adjustment of excavation depth to U-Th age; and (3) subtract 1.0‰ from the tooth δ¹⁸O values to account for the systematic offset between the records. Three tooth samples were identified as outliers from the age-depth relationship (Fig. 7, Supplementary Table 2). These could be due to either a seasonal bias in the sampling of these tooth fragments or genuine differences in the mean annual δ¹⁸O for Stegodon individuals that derived a significant proportion of their body water from divergent localised sources. The resulting U-Th tie-point ages represent the time of tooth formation. For subadult individuals, which constitute 90% of the tooth analyses, the time-gap between tooth formation, predation and burial at Liang Bua is expected to be minimal. However, the assignments of coeval ages to stratigraphically associated H. floresiensis remains are only accurate to the extent that their burial was contemporaneous with Stegodon.

We note that milk teeth belonging to subadult Stegodon individuals were omitted from the analysis of the So’a Basin fossil assemblage, which contains adult molars, to avoid the inclusion of lactating or weaning individuals. However, subadult teeth necessarily formed the majority of the enamel analyses for Liang Bua, where subadult individuals dominate the fossil assemblage⁵. Nevertheless, the average δ¹⁸O values for the prenatal (n = 9), juvenile (n = 13) and adult (n = 2) groups, excluding the three outliers in Fig. 7, were found to fall within a narrow range (0.33‰), well within the 2.4‰ range covered by the tooth δ¹⁸O dataset (Supplementary Fig. 13). The reproducibility demonstrates that the mixed tooth δ¹⁸O data for Liang Bua can be compared with the stalagmite K2 δ¹⁸O time series. Details on the dental formula used to assign each S. florensis insularis tooth to an ontogenetic stage can be found in Puspaningrum (Appendix II)⁸³.