Introduction

Mean annual temperature (MAT) in the Arctic is increasing at nearly four times the rate of the global average1, leading to a lengthened snow-free season2,3, which affects polar species, diversity, and ecosystem function4,5,6,7,8; hydrology9; global carbon cycling10,11; and sediment transport12. However, Arctic warming in response to rising levels of CO2 is temporally asymmetric, i.e., winters have warmed significantly more than summers13.

Paleoclimate reconstructions of past Arctic climate from ice cores have mainly focused on determining mean annual conditions, and MAT in particular. However, mean annual conditions do not provide a unique or complete window into climate. For example, San Francisco, USA and Beijing, China share the same MAT and mean annual precipitation (MAP), but have very different seasonal (summer versus winter) climates. Furthermore, sites with similar summer temperatures can exhibit very different winter temperatures (e.g., Chicago, USA and Rome, Italy). As for the Arctic, one of the effects of the highly seasonal insolation (continuously dark winters, followed by summers with continuous light) is a large difference between winter versus summer temperatures. For example, during the year 2020, mean July temperature in Verhkjoyansk (Eastern Siberia) was 16 °C, while mean January temperature was − 43 °C [Ref.14].

Reconstructions of Arctic seasonality are lacking compared to other Holocene climate records, and Holocene records of winter conditions are completely absent for Arctic Siberia [e.g.,15]. The Siberian Arctic houses a significant store of permafrost carbon; therefore, defining baseline conditions for this region is crucial to interpreting and contextualizing the effects of modern climate change. For example, the PAGES Arctic 2k database of Holocene climate reconstructions16 consists of fifty-six records, only five of which come from sites located between Scandinavia (30 °E) and Alaska (162 °W) (and all are summer temperature reconstructions)17,18,19,20, a region that represents ~ 50% of the Arctic. In order to augment this gap in our knowledge, we performed high-resolution stable carbon (δ13C) and oxygen (δ18O) isotope analysis on the annual growth rings of 3000 year-old fossil wood to provide the first proxy reconstruction of whole-year seasonal precipitation and temperature within Arctic Siberia.

Field site

We collected mummified fossil wood from an outcrop located along the Kolyma River in eastern Siberia (68.63°N, 159.15°E; Duvanny Yar, Sakha Republic, Russia; Fig. 1A). This outcrop is considered a model stratigraphic deposit for interpreting pre-industrial Arctic environmental change21. Mummified wood samples (cf. petrified) were excavated at this site from incised channel deposits above the organic-rich Pleistocene yedoma silt (Fig. 1B,C), a vast loess deposit containing ~ 500 Gt of carbon22. In order to obtain reliable ages for the exceptionally well-preserved mummified wood, we extracted α-cellulose from bulk wood using the modified Brendel method (see “Methods”). The clean, chemically pure α-cellulose was then radiocarbon dated to 2840 (sample DY2) to 2890 (sample DY4) 14C BP (± 20 years) at the National Ocean Sciences Accelerator Mass Spectrometry Facility at the Woods Hole Oceanographic Institution (calibrated ages with 95% probability: DY2 = 1077 to 920 cal BCE, DY4 = 1192 to 1005 cal BCE; OxCal online, Version 4.4), indicating that the plants were growing after loess deposition of the yedoma silt. The fragmentary specimens and limited number of rings in each sample precludes cross-correlation between samples.

Figure 1
figure 1

(A) Location of Duvanny Yar (yellow square) and Cherskiy, Russia (green circle). Base map was generated using "pscoast," part of the freely available Generic Mapping Tools package, version 4, previously at http://woodshole.er.usgs.gov/mapit/. (B) Outcrop view of Duvanny Yar showing the yedoma sediments slumping down towards the Kolyma River. (C) Photograph of cut cross-section of mummified wood (DY2) showing distinct earlywood (light brown) and latewood (dark brown) anatomy. (D) Average monthly precipitation (teal bars, 1980–2020) and temperature (red curve, 1900–2020). See Supplementary Information for climate data.

Full size image

Modern environmental conditions

Today, the site experiences a continental, Arctic climate with open Larix (Larch) forest and shrub vegetation (e.g., Pinus pumila; dwarf Siberian pine). Mean annual temperature (MAT, 1900–2021) determined using CRU TS v.4.06 reanalysis data is -12.7 ± 1.1 °C [Ref.14] (Supplementary Information) (Fig. 1D). A short (2006–2022), unpublished instrumental temperature record is also available for the nearby village of Kolymskoye, located 17 km to the northwest of the field site at Duvanny Yar. Comparison of the CRU TS v.4.06 reanalysis data for the period of overlap (16 years) reveals nearly identical mean monthly temperatures (mean difference = 0.1 °C, Pearson’s r = 0.99, n = 192 months), demonstrating reliability in the reanalysis data. Mean annual precipitation (MAP) from a weather station in Cherskiy, Sakha Republic, Russia (68.74°N, 161.40°E) is 236 ± 54 mm (1980–2020), and correlates well (Pearson’s r = 0.74, n = 8) with an overlapping, but short precipitation record from Kolymskoye (2013–2022). The majority of precipitation falls during summer months (Ps, defined as May through October = 156 ± 45 mm) with lesser precipitation during winter (Pw, November through April = 80 ± 27 mm). Snowfall generally occurs through May, and reaches maximum depth in late April21. The warm month mean (Tmax) and cold month mean (Tmin) temperatures are 9.5 ± 1.2 °C and − 35.3 ± 2.4 °C, respectively (median ± 1σ). The growing season is short; only 3 months (June to August) experience average temperatures > 5 °C.

Methods

Consecutive growth rings from four wood pieces (DY1, DY2, DY3, and DY4) were subdivided by hand using a razor blade, parallel to the growth rings, as per our previous studies on modern and fossil wood23,24,25,26,27,28,29,30. Fossil specimens showed distinctive rings in hand sample with discernable earlywood and latewood anatomy (Fig. 1C); therefore, the direction of growth was known. Each of the fossil rings was divided into 8–30 subsamples (average slice thickness: 55 to 187 µm) for δ13C analysis (n = 16 rings and 247 δ13C measurements) and 8–24 subsamples (average thickness of the subsamples ranged from 73 to 182 µm) for δ18O analysis (n = 17 rings and 216 δ18O measurements). Ring numbers and subsamples do not correspond across δ13C and δ18O analyses and cannot be matched within each fossil wood sample (δ13C and δ18O measurements were determined on different sections of the same fossils). All δ13C measurements were made on whole wood and δ18O measurements were made on α-cellulose31,32.

Cellulose was isolated from each subsample for δ18O analysis using a method modified from [Refs.33,34]. The subsamples were placed into 1500 µl polypropylene micro centrifuge tubes and a 10:1 ratio of 80% acetic acid to 70% nitric acid was added to the samples. The subsamples were then agitated using a standard vortex mixer (Fisher Scientific), centrifuged for a minimum of 5 min at 12,000 rpm, and heated in an aluminum block at 120 °C for 30 min before being allowed to cool to room temperature. The liquid was decanted and the subsamples were rinsed twice with 1000 µl of 99% ethanol and once with 1000 µl of deionized water. Next, 300 µl of a 17% sodium hydroxide (NaOH) solution was added, and after 10 min, the subsamples were rinsed with 1000 µl of deionized water. A solution containing 36 µl acetic acid and 260 µl deionized water was then added, followed by a final rinse with 300 µl of 99% ethanol and 300 µl of acetone in successive steps. The subsamples were then dried overnight at 50 °C in a drying oven (Thermo Scientific Precision). Percent cellulose, measured in two samples, averaged 20% (DY2) and 16% (DY4), approximately half the amount of modern wood31.

Each subsample was weighed singly into tin capsules for δ13C analysis of whole wood and into silver capsules for δ18O analysis of α-cellulose (DY1 and DY3 in singlicate; DY2 and DY4 in duplicate). Stable isotope values were determined using a Delta V Advantage Isotope Ratio Mass Spectrometer (IMRS) joined with a Thermo Finnigan Elemental Analyzer (Flash EA1112 Series, Bremen, Germany) (δ13C analysis) or a high temperature conversion elemental analyzer (TC/EA) (Thermo Fisher, Bremen, Germany) configured with a zero-blank autosampler (Costech Analytical, Valencia, CA, USA) (δ18O analysis). Quality control samples (δ13C: JGLUC; δ18O: JCELL2) were analyzed as unknowns within each batch run. Over the course of the analyses, the JGLUC and JCELL2 quality control sample averaged (± 1σ) − 10.64 ± 0.22 ‰ (n = 15) and 20.54 ± 0.10‰ (n = 27), respectively [calibrated reference values are − 10.52‰ (JGLUC) and 20.44‰ (JCELL2)]. All isotope values were expressed in delta notation in units per mille (‰).

Data analysis was preformed in R Studio version 4.3.0 [Ref.35] (Supplementary Information). Uncertainty in all equations was propagated using a Monte Carlo resampling approach, which is detailed in the Supplementary Information.

Results

All samples showed a seasonal change in δ13Cwood and δ18Ocell value through each growth year (Figs. 2, 3, Tables S1, S2), similar in shape to the intra-ring patterns observed in modern wood from other high-latitude and Arctic sites24,25.

Figure 2
figure 2

High-resolution carbon isotope measurements on fossil wood. Vertical dashed lines mark ring boundaries; growth is from left to right. Analytical precision (± 1σ) for each δ13C value was ± 0.2‰. For illustration, δ13Cmax and δ13Cmin values for DY1 are indicated; all data are reported in Tables S1 and S3.

Full size image
Figure 3
figure 3

High-resolution oxygen isotope measurements on fossil cellulose. Vertical dashed lines mark ring boundaries; growth is from left to right. Error bars in (B) and (D) indicate the ± 1σ uncertainty based on duplicate analyses; analytical precision (± 1σ) was ± 0.1‰ for individual measurements. For illustration, δ18Ocell(max) and δ18Ocell(min) values for DY1 are indicated; all data are reported in Tables S2 and S4.

Full size image

Determination of paleoseasonality

We used δ13Cwood data collected across annual growth rings to quantify summer and winter precipitation (Ps: May through October, and Pw: November through April) using a published relationship between intra-ring carbon isotope analyses and precipitation24,25. This relationship was calibrated via an analysis of 33 high-resolution δ13C profiles and 286 annual growth rings from evergreen trees growing across a wide range of environments (including high latitude sites in Sweden and Siberia)24, and has been used to reconstruct precipitation seasonality from living trees28 and using significantly older fossils (Miocene to Eocene in age) collected from Arctic26,27, Antarctic36, and monsoon environments30.

The δ13C values measured within annual growth rings are related to seasonal precipitation via the following (R2 = 0.96) [Ref.24]:

$$\Delta \left( {\updelta^{{{13}}} {\text{C}}_{{{\text{wood}}}} } \right) = 0.0{1}L - 0.{82}\left[ {{\text{ln}}\left( {{\text{P}}_{{\text{s}}} /{\text{P}}_{{\text{w}}} } \right)} \right] + 0.{86}$$
(1)

Within the above, Ps/Pw is the ratio of summer to winter precipitation as a measure of seasonality; L (latitude, in degrees) = 68.63 and Δ(δ13Cwood) is the difference between the maximum δ13C value of a given ring (δ13Cmax) and the preceding minimum δ13C value of the annual cycle (δ13Cmin) [Δ(δ13C) = δ13Cmax – δ13Cmin]. Using Eq. (1) and the δ13Cwood data gained from 16 fossil tree-rings analyzed at high resolution (Fig. 2) we calculate median Ps/Pw = 1.6, suggesting a summer-dominated precipitation regime 3000 years ago, but with a larger fraction of wintertime precipitation compared to the modern meteorological records (Ps/Pw = 2.1; 1981–2019) (One-sided Wilcoxon rank sum test, p < 0.001).

Although Eq. (1) relates Δ(δ13Cwood) to Ps/Pw, Ref.24 noted that reconstructed Ps/Pw ratios can be used to quantify Ps and Pw directly, provided independent constraints on mean annual precipitation (MAP), given that:

$${\text{MAP }} = {\text{ P}}_{{\text{s}}} + {\text{ P}}_{{\text{w}}}$$
(2)

Modeling work by [Ref.37] suggest the site experienced 0.125 mm/day more precipitation 3000 years ago than present (0.125 mm/day = 46 mm/year), yielding a reconstructed Holocene MAP (MAP3ka) for the site 3000 years ago = 282 ± 55 mm. The MAP3ka resamples range from 63 to 516 mm, encompassing all published hydroclimate estimates for this region during this time window15,38. An estimate for maximum MAP3ka using the upper limit of the [Ref.37] estimate (0.25 mm/day wetter than modern) would equate to 307 mm, or approximately the 68th percentile of the MAP3ka resamples, providing a conservative range of values with high uncertainty to account for relatively few independent paleo-MAP constraints for the region.

By solving Eqs. (1, 2) simultaneously, we determined consensus median (± 1σ) Ps = 170 ± 56 mm and Pw = 103 ± 50 mm for the site based on 10,000 solutions that incorporate uncertainty in MAP, Ps and Pw (Fig. 4) (see Supplementary Information). The estimations suggest that the site experienced an absolute increase in precipitation year-round 3000 years ago relative to present (Ps = + 14 mm, Pw = + 23 mm), which equates to ~ 10% greater precipitation in the summer months and ~ 30% increase in winter precipitation (Fig. 4). Because the Ps/Pw ratios directly quantified from the Δ(δ13Cwood) proxy are lower than modern conditions, and independent modeling efforts suggest a slight increase in MAP values 3000 years ago, the estimated Ps and Pw values each are clearly higher than late twentieth century meteorological observations (One-sided Wilcoxon rank sum tests, p < 0.001 for each).

Figure 4
figure 4

Results of annually resolved estimates of summer (Ps, A) and winter (Pw, B) precipitation using 3000-year-old wood fossils. Kernel density plots (right) show the distribution of consensus proxy estimates compared to modern values. Proxy estimates, determined from δ13Cwood and Eqs. (1, 2), are shown as median values (colored symbols) and 95% confidence interval (error bar) calculated from Monte Carlo resampling. Symbols and annual bars along the top of each plot are colored according to precipitation anomaly, calculated as the difference between the proxy value and Cherskiy weather station (1981–2000 C.E.) (positive values = more precipitation 3000 years ago).

Full size image

We used δ18Ocell data collected across annual growth rings to quantify warm-month mean temperature (Tmax) using a published empirical relationship between intra-ring oxygen isotope analyses and temperature. This relationship was calibrated from a dataset of intra-ring oxygen isotope measurements on modern wood cellulose from 33 globally distributed sites, including 10 sites with below-freezing cold-month mean temperatures25. The δ18Ocell values measured within annual growth rings are related to changes in seasonal temperature via the following equation (R2 = 0.82) (see Supplementary Information) [Ref.25]:

$$\Delta \left( {\updelta^{{{18}}} {\text{O}}_{{{\text{cell}}}} } \right) = \left[ {\left( {{17}.{\text{57T}}_{{{\text{max}}}} } \right)/\left( {{39}.{89 } + 0.{\text{44T}}_{{{\text{max}}}} } \right) - \left( {{17}.{\text{57T}}_{{{\text{grow}}}} } \right)/\left( {{39}.{89 } + 0.{\text{44T}}_{{{\text{grow}}}} } \right) + 0.00{82}\left( {{\text{P}}_{{\text{w}}} - {\text{ P}}_{{\text{s}}} } \right)} \right] /{ 1}.{13}$$
(3)

Within the above, Tmax and Tgrow are the average warm month temperature and the average minimum growing season temperature (°C); ∆(δ18Ocell) is the difference between the maximum and minimum δ18Ocell value within each ring [∆(δ18Ocell) = δ18Ocell(max) – δ18Ocell(min)]; Pw and Ps were calculated using Eq. (2), as reported above. For the analysis above, Tgrow was set to be a uniform distribution between 1 and 5 °C [after Refs.8,39,40] in order to account for uncertainty in growth strategy and timing of the start of the growing season. Using Eqs. (1–3), we reconstructed a consensus median Tmax estimate across all 17 rings = 8.5 ± 4.4 °C (Eq. 4) (Fig. 5A) (Supplementary Information), which is significantly cooler than median Tmax of reanalysis temperatures from 1900–2021 (9.5 ± 1.2 °C) (One-sided Wilcoxon rank sum test, p < 0.001).

Figure 5
figure 5

Results of annually resolved estimates of warm month (Tmax, A) and cold month (Tmin, B) average temperature using 3000-year-old wood fossils. Kernel density plots (right) show the distribution of proxy estimates across all samples compared to modern values. Proxy estimates, determined using δ13Cwood, δ18Ocell, and Eqs. (1–4), are shown as median values (colored symbols) and 95% confidence interval (error bar) calculated from Monte Carlo resampling. Symbols and annual bars along the top of each plot are colored according to temperature anomaly, calculated as the difference between proxy value and the CRU TS v.4.06 reanalysis data (1971–2000 C.E.) (positive values = warmer temperatures 3000 years ago).

Full size image

We can use calculated Tmax values to quantify Tmin, provided an independent estimate of MAT because MAT can be approximated as an average of Tmax and Tmin, as confirmed across a global dataset at 10' spatial resolution25. Equation 4, below, was verified for all land areas, including the CRU TS v.4.06 reanalysis data used here (1900–2021, mean difference = 0.2 °C), by comparison with MAT calculated using all 12 months of the year (R2 > 0.99, n = 566,262) [Ref.25], where:

$${\text{MAT }} = \left( {{\text{T}}_{{{\text{max}}}} + {\text{ T}}_{{{\text{min}}}} } \right) /{ 2}$$
(4)

We determined pre-industrial MAT using reanalysis temperatures from 1900–1925 C.E. (− 13.6 ± 0.5 °C) and then adjusted the temperatures upward for MAT 3000 years ago assuming warming of 0.5 ± 0.5 °C, based on Holocene long-term warming trends proposed by [Ref.41], and consistent with latitudinal temperature gradients42,43. The 10,000 resamples for MAT 3000 years ago (MAT3ka) have a median ± 1σ = − 13.1 ± 0.7 °C and a wide range (− 16.7 to − 8.4 °C) to account for the lack of independent constraints for the region [e.g.,15,44,45]. Using our estimates of Tmax determined above, we quantified median Tmin = − 34.7 ± 4.6 °C (Eq. 4) (Fig. 5B) (Supplementary Information), which is 3 °C warmer than Tmin values from the early twentieth century temperatures (1900–1925 C.E.; Tmin = − 37.7 ± 2.5 °C) and 0.6 °C warmer than median Tmin across the full temperature record (1900–2021 C.E.; − 35.3 ± 2.4 °C), consistent with a significant shift in central tendency towards higher Tmin values 3000 years ago (One-sided Wilcoxon rank sum test, p < 0.001).

Discussion

Our determination of increased winter and summer precipitation 3000 years ago, relative to pre-industrial values for the region (Fig. 4), is similar to modern changes observed between 1981 and 2020 (Fig. 6). Across the instrumental record, Ps and Pw have both increased (Ps = 1.4 mm/year, Pw = 1.2 mm/year; F-test, p < 0.05 for each), resulting in an increase in precipitation of 31% (Ps) and 58% (Pw) when comparing the 2010s to the 1980s. Across the years 2011–2020, values for Ps (180 ± 54 mm) and Pw (105 ± 39 mm) are now similar to those determined here for 3000 years ago (Ps = 170 ± 57 mm; Pw = 103 ± 50 mm).

Figure 6
figure 6

Historical precipitation (A–C) and temperature (D–F) data for 1981–2020 C.E. Stripes represent anomalies, calculated as the difference between each annual value and the 1971–2000 C.E. mean value from CRU TS v.4.06 reanalysis data (temperature) and the 1981–2000 C.E. mean value from the Cherskiy weather station records (precipitation). These data reveal a significant increase in precipitation across all seasons and clear bias in seasonal warming towards winter.

Full size image

As for temperature, our results show that the relatively slight increase in mean annual temperature in the Siberian Arctic 3000 years ago relative to pre-industrial values15,41,42,43 was due to notably increased winter temperatures, despite cooler summers (Fig. 5). Within modern records, winter temperatures (November through January) have increased at triple the rate (0.95 °C/decade versus 0.35 °C/decade, 1981–2020) compared to summer temperatures (June through August), with MAT increasing at an intermediate rate of 0.65 °C/decade (Fig. 6). This work demonstrates greater sensitivity of winter versus summer temperatures in the Siberian Arctic region to climate change and the importance of winter temperatures in driving changes in MAT, both 3000 years ago and at present.

Models of future warming project amplification of this asymmetric seasonal warming through CO2 forcing46, with winter temperatures in the Arctic increasing a further 7 to 13 °C, while summer temperatures warm a more modest 2 to 5 °C by the end of this century47. Although the current rate of temperature increase in the Arctic is unprecedented within the historical record48, our novel high-resolution data provide key insight on the full effects of industrialization on Arctic seasonality, including warmer and wetter winters.

Expressions of winter temperature and precipitation within the earth system are myriad, and include increased primary production49, reduced surface albedo due to earlier leaf-out50,51, and enhanced ecosystem respiration resulting in lower net carbon uptake52. All of these have the potential to alter permafrost and the associated carbon stores within. The IPCC estimates that Arctic and boreal permafrost soils contain 1460 to 1600 Gt of organic carbon53 with approximately one-third (~ 500 Gt) contained within yedoma sediments22. There is much uncertainty in how warming will affect soil carbon loss54, but recent model predictions indicate winter warming will increase CO2 loss by 17–41% within permafrost regions55. This will more than offset any potential increases in plant productivity driven by moderate summer warming56, particularly if permafrost thaw is abrupt57.

Much of our baseline climate data for the Arctic are limited spatially (e.g., Arctic Canada and Greenland) or to only the most recent decades16,58,59. Ice-core data from the Greenland Ice Sheet provide millennial-scale records of mean annual climate60, while other longer-term records (e.g., marine foraminifera), may contain seasonal biases that hamper interpretations of long-term warming versus cooling trends44,61,62. Moreover, such analyses focus on measurements of sea surface temperature within low latitude regions due to the scarcity of applicable data from high-latitudes45. Limited, high-latitude terrestrial proxies, such as stable isotope analyses of relict pore ice in permafrost (summer temperature proxy) and ice wedges (winter temperature proxy) indicate millennial-scale trends in seasonal warming in the western Arctic (e.g., western Canada)63,64, while tree-ring width data from western Siberia show a long-term summer cooling trend prior to industrialization65. As we work to further resolve the history of carbon-cycling throughout the Holocene, knowledge of seasonality will be critical to evaluating the dynamic nature of this large carbon pool.

Conclusions

Globally, more than 75% of proxy temperature data for the Holocene represent mean annual conditions, and less than 2% represent winter temperatures62; thus we have very little insight into the seasonality of the past. Within the Arctic, especially, existing data are disproportionately focused across northern parts of Europe and North America, and, like records from lower-latitudes, limited to estimates of long-term, mean annual conditions, or only partial-year seasonality16. Here we used a high-resolution method to demonstrate warmer winter temperatures and higher winter precipitation 3000 years ago, within a severely understudied region of the Arctic: eastern Siberia. Our findings support trends observed today of amplified winter warming under elevated MAT: the fossil plants we analyzed experienced warmer cold-month temperatures 3000 years ago relative to twentieth century temperatures. However, in contrast to expectations, warm-month temperatures were cooler than present, despite a warmer MAT. Today, the climate of eastern Siberia is changing in a similar fashion: increases of + 2.9 °C and + 60% in winter temperature and precipitation, respectively, over the last 30 years. From this we suggest that high MAT values in the Arctic accompanied by increased precipitation and decreased temperature seasonality (i.e., winter temperatures are closer to summer temperatures) are fundamental features of high-latitude environments.

In past intervals of extreme global warmth, the Arctic experienced similar, proportional increases in precipitation across all seasons27 and asymmetric seasonal warming [e.g.,66], but these results are for periods with very high CO2, stable hot greenhouse climates, and ice-free conditions. The record of whole-year seasonality produced here highlights the importance of seasonal climate (temperature and precipitation) on mean annual trends and represents a critical advance in our understanding of changes in climate seasonality in the Arctic under a warmer climate comparable to present and near-future climate predictions.

Warming of Arctic regions is a periodic feature of the geologic record, at times leading to profound changes in the distribution of life on our planet (e.g., the lush Arctic forests of the middle Eocene67). The twentieth century onset of today’s Arctic warming was not the result of uniform changes throughout the year13; our results show that temporal heterogeneity during climate change, with pronounced warming during winter, was also in place 3000 years ago.