Introduction

Boreal forests account for about 30% of global forest area and are a dominant contributor to the terrestrial carbon sink1. However, in recent years, some of these typically cold and moist forests have experienced increasingly hot and dry conditions2,3, leading to more destructive wildfires and a weakening carbon sink. While burned area has decreased markedly in most of Western Siberia and the Russian Plain over the past two decades, possibly due to increasingly stringent fire management in these regions4, Eastern Siberia witnessed record-breaking wildfires in 2019 and 2021, with annual burned area fraction exceeding 2% across broad swaths of the region (Supplementary Fig. 1a, b). The frequency and size of wildfires in Eastern Siberia have increased significantly in recent decades4,5. Compared to 2001–2011, the multi-year average burned area fraction over 2011–2021 rose by ~1% of total area, rendering it one of the most severely affected regions in the Northern Hemisphere (Supplementary Fig. 1c). This trend has drastically altered ecosystem composition in the region and increased carbon emissions3. Given the perilous positive feedback loop between wildfires, carbon balance, and warming, it is imperative to understand the causes of escalating wildfire risk in Eastern Siberia.

Recent record-breaking wildfires in Eastern Siberia have been linked to intense summer heat6 and extreme climatic water deficit (i.e., difference between precipitation and potential evapotranspiration)3, but the root causes of these changes remain unclear. Anthropogenic climate change, particularly warming, is widely identified as a driver of increased wildfire activity around the world because it elevates the likelihood and severity of compound dry-hot conditions7,8, but this argument does not work for Eastern Siberia because observations and climate models driven by historical external forcings consistently show a long-term wetting trend in northern high latitudes since the mid-twentieth century9 due to the enhancement of poleward moisture transport that results from a warming-induced increase in atmospheric water vapor and shifts in high-latitude circulation regimes10. Global warming therefore cannot explain the observed decreases in Eastern Siberian precipitation9,10 that exacerbate the climatic water deficit, nor can it fully account for the recent pronounced decadal changes in burned area in Eastern Siberia4,5, which decreases until the late 2000s and then increases dramatically (Supplementary Fig. 1d). While greenhouse gas emissions induce a long-term wetting trend in the northern high latitudes, internal climate variability may be crucial in driving decadal precipitation variations at regional or local scales, potentially counteracting the externally driven wetting trend and amplifying precipitation deficit and wildfire risk. In Eastern Siberia, remote internal phenomena such as the North Atlantic Oscillation11, the Interdecadal Pacific Oscillation (IPO)12, and the Atlantic multidecadal variability13,14 may act as the critical drivers of decadal variations in precipitation. However, the importance of precipitation for driving recent wildfires in Eastern Siberia has been challenged in the literature. A recent study identified another form of climatic water deficit (i.e., difference between reference evapotranspiration and actual evapotranspiration) as the most critical driver of burned area in Siberian boreal forests over the past two decades, over the independent effects of summertime temperature and precipitation (among other climate drivers)15. However, this result is unsatisfying because while the study allowed changes in precipitation to affect wildfire frequency and intensity16,17 via mechanisms not mediated by moisture deficit, it did not consider the indirect impact of precipitation on burned area via its key role in regulating water deficit.

Here, we characterize the causal relationships between the climatic causes of wildfire in Eastern Siberia and identify the critical role of declining precipitation in driving the escalation of burned area in Eastern Siberia over the past two decades. We then examine the underlying drivers of declining precipitation and uncover a teleconnection between decadal variations in subpolar North Atlantic (SPNA) sea surface temperature (SST) and Eastern Siberian fire-season precipitation, revealing a potentially role of unforced SPNA decadal cooling in the recent escalation of wildfires in Eastern Siberia.

Drivers of wildfires in Eastern Siberia

Ignition sources, fuel availability, and fire-conducive weather constitute the essential components of wildfires18. Eastern Siberia has abundant forest litter as fuel, but its availability largely depends on ambient moisture and thermal conditions. Regional climate regulates fuel flammability and weather conditions, thereby acting as a primary driver of wildfire dynamics8. Thus, we examine seven climatic factors related to moisture, temperature, and weather conditions: vapor pressure deficit (VPD), surface soil moisture, precipitation, daily maximum and minimum temperature, frost day frequency, and surface wind speed. Over 99% of the burned area in Eastern Siberia occurs between April and September, aligning with the seasonal cycles of daily maximum temperature and aboveground biomass (Supplementary Fig. 1e). For this reason, the period from April to September is defined as fire season, with the preceding months referred to as pre-fire season.

Over the past two decades, fire-season burned area fraction in Eastern Siberia was negatively correlated with fire-season precipitation (r = −0.71, p < 0.001). Fire-season precipitation increased until 2008 and then declined (with a rate of −16 mm per decade during 2001–2021), reaching a minimum of ~250 mm in years with record-breaking wildfires (Supplementary Fig. 1d). In contrast, fire-season daily maximum temperature rose steadily by 0.74 °C per decade, suggesting that precipitation may exert a stronger impact on burned area than daily maximum temperature in recent years. However, precipitation, temperature, VPD, surface soil moisture, and other climatic factors (e.g., surface wind speed, frost day frequency) are interrelated, complicating the analysis. For instance, precipitation can modulate surface soil moisture and VPD, and all these factors may affect wildfire frequency and size.

To identify the primary drivers of wildfires in Eastern Siberia and guide subsequent analyses, we examine the correlations between the fire-season burned area fraction and the above-mentioned seven key climatic factors during fire and pre-fire seasons (Fig. 1a), assign causal directions based on background knowledge, remove indirect casual links, and validate our proposed causal graph (Fig. 1b) using Pearlian-style causal inference19,20 (Methods). Specifically, graph pruning and validation are conducted following the principle that for pairs of variables characterized by statistically significant unconditional correlations (p < 0.1) that are not directly linked in the causal graph, their conditional correlations must be statistically non-significant (p > 0.1) when controlling for the d-separating set implied by the causal graph.

Fig. 1: Links between wildfire and climatic factors.
Fig. 1: Links between wildfire and climatic factors.
Full size image

a Correlations between annual variations in fire-season burned area fraction (bf_fs) and underlying climatic factors in Eastern Siberia from 2001 to 2021. These factors include vapor pressure deficit (vpd), soil moisture (sm), precipitation (pr), daily maximum and minimum temperatures (tmx and tmn), frost day frequency (frs), and near-surface wind speed (ws). Here, ‘fs’ and ‘pfs’ denote fire and pre-fire seasons, respectively. Statistical significance of correlation is marked by one, two, three, and four stars, corresponding to the levels of 0.1, 0.05, 0.01, and 0.001, respectively. The strengths of correlations between bf_fs and climatic factors (except pr_fs) after controlling for the confounding effects of pr_fs on these factors are shown at the second bottom row (highlighted by the blue box). The bottom row shows partial correlations with bf when conditioning on the set {vpd_fs, pr_fs}. b Causal graph constructed using the Pearlian causal inference framework. Unidirectional arrows between variables represent causal effects, while bidirectional arrows indicate correlation. ‘FS’ denotes fire season.

Examining the correlations (Fig. 1a) of fire-season burned area fraction (red box) with the various climatic factors (listed on the x-axis), we find strong correlations during fire-season (fs, left), but only non-significant correlations during pre-fire season (pfs, right). The correlations between pre-fire-season variables (bottom half) and fire-season climatic factors (left half) are non-significant or small, and cannot explain the strong correlations among the fire-season climatic factors (top left quadrant). Thus, our analysis will focus on the fire-season.

Of the seven fire-season climatic factors, fire-season burned area fraction (red box) exhibits the strongest correlation with fire-season precipitation (r = −0.71, p < 0.001), followed by fire-season VPD (r = 0.68, p < 0.001) and daily maximum temperature (r = 0.41, p < 0.1). Its correlation with surface soil moisture is marginally non-significant (r = −0.34, p = 0.14). Initially, we add single-headed arrows from precipitation, daily maximum temperature, surface soil moisture, and VPD to burned area fraction, as all these factors affect vegetation dryness and flammability16,17,21, which are closely related to burned area fraction. The correlations of burned area fraction with daily minimum temperature, frost day frequency, and surface wind speed are negligible, so we do not connect them directly to burned area fraction.

Focusing on the correlations with precipitation (second row/third column), we observe that in addition to its strong correlation with VPD (r = -0.69, p < 0.001), it correlates significantly with surface soil moisture (r = 0.57, p < 0.01). It is understandable because precipitation directly affects surface soil moisture and VPD through water supply. Additionally, there are significant correlations between VPD (first column) and all the factors except surface wind speed. Increased temperature leads to elevated saturation vapor pressure according to the Clausius–Clapeyron relation, so we add arrows from daily maximum and minimum temperatures to VPD. Soil moisture should influence VPD via modulating land surface evaporation22. Given the low local and regional moisture recycling ratios in Eastern Siberia23,24, we argue that the feedback effects of VPD on precipitation and soil moisture are limited. Accordingly, we add arrows from precipitation to surface soil moisture and from precipitation and surface soil moisture to VPD, but no backward paths from VPD to precipitation and soil moisture are included.

Unlike the other variables, there is no clear mechanism for frost day frequency to affect VPD. Frost day frequency (fifth row) is correlated significantly with daily maximum and minimum temperature (fourth and fifth columns), which are both drivers of VPD, so it is possible that frost day frequency is confounded with VPD by temperature. Given the non-significant correlation between daily maximum temperature and frost day frequency when conditioning on daily minimum temperature (partial r = 0.00, p = 0.99), we consider that daily maximum temperature is not a driver of frost day frequency, but rather be confounded through daily minimum temperature. Furthermore, the correlation of frost day frequency with VPD becomes non-significant (partial r = 0.15, p = 0.52) when conditioning on daily minimum temperature, confirming that the relationship is confounded rather than causal. For the rest of the analysis, we omit surface wind speed because it has no significant correlation with either fire-season burned area fraction or other climatic factors, but consider that surface soil moisture, precipitation, and daily maximum and minimum temperatures may have indirect effects on burned area mediated by VPD.

Next, we observe that daily maximum and minimum temperatures (fourth row/column) are strongly correlated. We interpret this relationship as arising from shared sensitivity to hidden confounding factors (e.g., weather systems, horizontal temperature advection), which we represent with a double-headed arrow in Fig. 1b. We also observe weaker correlations between fire-season daily maximum temperature (third row) and precipitation (r = −0.34, p = 0.13) and surface soil moisture (r = −0.32, p = 0.16), which are only marginally non-significant. The observed negative correlation between daily maximum temperature and precipitation is likely due to the confounding effect of atmospheric circulation. For instance, under high-pressure systems, high-temperature and low-precipitation conditions often co-occur, which aligns with this negative correlation. While elevated daily maximum temperature at high latitudes may increase atmospheric water vapor and thus precipitation in early summer via enhancing snow melt and evaporation25, this proposed positive causal effect contradicts the observed negative correlation. Furthermore, the positive causal effect is typically linked to low-pressure systems, whereas the negative confounding relationship is tied to high-pressure systems. As these two relationships are associated with mutually distinct atmospheric circulation patterns, they are physically incompatible. We thus argue that the negative correlation between daily maximum temperature and precipitation is more likely due to the confounding effects of atmospheric circulation associated with high-pressure systems, and add a bidirectional confounding arrow between them in Fig. 1b. While surface soil moisture can alter surface energy partitioning, potentially influencing daily maximum temperature, we do not include an edge from surface soil moisture to daily maximum temperature. This is because land surface evaporation in this moist and cold region is energy-limited, and the evaporation fraction (the ratio of latent heat flux to available energy) does not depend on soil moisture26. As a result, soil moisture cannot significantly affect daily maximum temperature by altering the division of available energy into latent and sensible heat fluxes26. We thus argue that any correlation between surface soil moisture and daily maximum temperature is confounded by precipitation (partial r = −0.16, p = 0.49).

The correlation between surface soil moisture and VPD is significant (r = −0.51, p < 0.05), but becomes statistically non-significant when conditioning on precipitation (partial r = −0.21, p = 0.38). We understand this to mean that the relation between surface soil moisture and VPD is due to the confounding effect of precipitation, which serves as a critical water source for both the soils and the surface atmosphere. Therefore, we remove the path between surface soil moisture and VPD.

Finally, we hypothesize that the effects of surface soil moisture and daily maximum temperature on burned area are mediated by VPD and/or confounded by precipitation. This is confirmed by the loss of statistical significance in the correlations between burned area fraction and surface soil moisture (partial r = 0.11, p = 0.65 when conditioning on precipitation) and daily maximum temperature (partial r = −0.26, p = 0.28 when conditioning on VPD and precipitation). Given that the model no longer incorporates any dependency between surface soil moisture and burned area fraction, we remove surface soil moisture from the causal graph.

Based on the above reasoning, we propose the causal graph shown in Fig. 1b. In the graph, daily maximum and minimum temperatures, VPD, and precipitation are all identified as drivers of burned area fraction. When controlling for precipitation (blue box), the correlation between burned area fraction and VPD weakens (partial r = 0.38, p = 0.10), and the correlation between burned area fraction and daily maximum temperature becomes non-significant, indicating the crucial role of fire-season precipitation in shaping Eastern Siberian burned area over the past two decades. Detailed interpretations of all causal effects in the graph and the related tests are provided in Supplementary Note 1.

Primary causes of recent escalation of Eastern Siberian wildfires

To quantitatively disentangle the influences of these climatic factors on burned area in Eastern Siberia, we perform a path analysis to account for their internal connections (Methods), with results presented in Fig. 2a. While mode fit assessment is more challenging for path analysis, we prefer this approach to standard multi-regression because it provides causal rather than merely predictive insights. Additionally, it accommodates more complex models, distinguishes between direct and indirect effects, and generalizes better to changing background conditions such as global warming17,27 (Supplementary Note 2). Given that strong collinearity between daily maximum and minimum temperatures could distort model estimates (variance inflation factor exceeds 10; Supplementary Note 3), we apply a regularization method to estimate path coefficients to reduce variance of the estimated causal impact of all predictors. Detailed validation and associated sensitivity analysis results for the assumptions of the path analysis model are provided in Supplementary Note 3.

Fig. 2: Individual roles of climatic factors in Eastern Siberian fire-season wildfire.
Fig. 2: Individual roles of climatic factors in Eastern Siberian fire-season wildfire.
Full size image

a Path analysis diagram illustrating the causal relationships between fire-season burned area fraction and related climatic factors. Here, ‘FS’ denotes fire season. The numbers on the single-headed arrow lines are standardized path regression coefficients, with one to four stars indicating significance levels of 0.1, 0.05, 0.01, and 0.001, respectively. Dashed line represents non-significant effect. Correlation coefficients (r) between fire-season daily maximum and minimum temperatures and between fire-season daily maximum temperature and precipitation are shown with the double-headed arrow lines. The total effect (TE) of each factor on fire-season burned area fraction is shown below its label. b Individual contributions of climatic factors on the observed trend in fire-season burned area fraction in Eastern Siberia during 2001–2021. Purple and green denote the direct and indirect effects, respectively, while gray represents the unknown effect. Note that the total trend in burned area fraction explained by the model equals the sum of the direct effects of precipitation and VPD, with the remaining effects attributed to other unknown factors.

For the interpretation of the path coefficients shown in Fig. 2a, note that these are unitless estimates scaled in relation to the standard deviation of both the cause and response variables over the timeframe of the data used (2001–2021). Larger absolute values indicate stronger effects during the study period. Our analysis identifies that precipitation was the dominant driver of burned area. In addition to its strong direct effect (−0.46, p < 0.05), reduced precipitation increases atmospheric dryness (VPD) (−0.44, p < 0.001), and VPD itself promotes fire activity (VPD→bf: 0.37, p < 0.1). Combining these pathways, the total effect of precipitation on burned area fraction (−0.62) reflects both its direct suppression effect on wildfires and its indirect suppression effect via reducing VPD. Daily maximum temperature acts to amplify wildfire risk through its strong positive effect on VPD (0.44, p < 0.001), generating an indirect effect on burned area. Of the variables included in the path diagram, the magnitude of precipitation’s total effect is the largest, underscoring its primacy in regulating burned area variability in Eastern Siberia during the timeframe analyzed.

To estimate the relative contributions of these factors to the observed trend in fire-season burned area fraction over the past two decades, we apply the coefficients of total effects (Fig. 2a) to the trends in standardized fire-season precipitation (−0.73σ per decade), daily maximum temperature (0.95σ per decade), VPD (1.06σ per decade), and other climatic factors (Supplementary Table 1). Our results show that the decline in fire-season precipitation accounted for ~42% of the observed increase in burned area fraction, followed by the increases in VPD (about 36%) and daily maximum temperature (about 14%) (Fig. 2b). These findings indicate that the reduction in fire-season precipitation was the most critical cause of the escalated wildfires in Eastern Siberia over the past two decades.

Possible origin of Eastern Siberian precipitation change

To investigate the contributions of external forcings and internal climate variability to the observed fire-season precipitation changes, we analyze single-model initial-condition large ensembles28,29, which are sets of model simulations driven by identical external forcings but initialized with varying initial conditions (Methods; Supplementary Table 2). We first examine fire-season precipitation trends in Eastern Siberia and northern mid-to-high latitude land areas from 2001 to 2021 using simulations from the fully coupled climate model FGOALS-g330. We choose FGOALS-g3 because it provides the largest ensemble, consisting of 110 members. A large ensemble can more effectively cancel out independently evolving internal variability via the ensemble mean, thereby isolating the signals induced by external forcings. External forcings result in an increase in simulated fire-season precipitation averaged over northern mid-to-high latitude land areas, consistent with the observations (Fig. 3a), allowing us to examine simulated precipitation in Eastern Siberia specifically. External forcings also induce a wetting trend in Eastern Siberia, but in combination with internal climate variability, a small number of ensemble realizations can reproduce precipitation declines with magnitudes close to the observations (orange circles). Similar behaviors emerge in large-ensemble simulations from other models (Supplementary Fig. 2). These results suggest that the recent decline in Eastern Siberian fire-season precipitation was primarily a result of internal climate variability. We confirm this by showing that the mean differences in precipitation trends between the MIN5 ensemble members (the five realizations with the most pronounced decreasing fire-season precipitation trends in Eastern Siberia during 2001–2021) and the ensemble mean (MEAN) exhibit comparable sign and magnitude to the observed precipitation trends in Eastern Siberia and other regions (e.g., Western Europe and North America; Fig. 3b, c), further indicating the dominant role of internal climate variability in the observed fire-season precipitation changes in Eastern Siberia.

Fig. 3: Critical role of internal climate variability in changes of precipitation and geopotential height.
Fig. 3: Critical role of internal climate variability in changes of precipitation and geopotential height.
Full size image

a Observed and ensemble mean fire-season precipitation trends over northern mid-to-high latitude land areas (NH; 45°N–90°N) and Eastern Siberia during 2001–2021. Black whiskers denote the 95% confidence interval of observed trends and the 5–95% range of trends across ensemble members. The first five members of the FGOALS-g3 super-large ensemble with the largest decreasing (MIN5) and increasing (MAX5) fire-season precipitation trends in Eastern Siberia during 2001–2021 are indicated by orange and blue circles, respectively. b Composite differences in fire-season precipitation trends between MIN5 and the ensemble mean (MEAN). c Observed fire-season precipitation trends from 2001 to 2021 based on the CRU TS dataset. d As in b, but for fire-season eddy geopotential height at 500 hPa (Z500). e Z500 trends during 2001–2021 derived from the National Centers for Environmental Prediction/National Center for Atmospheric Research Reanalysis-I (NNRea). Black dots represent significant trends at the 0.05 level.

To identify the primary source of internal climate variability-related atmospheric circulation anomalies that drove fire-season precipitation variations in Eastern Siberia over the past two decades, we calculate SST trend differences between the MIN5 ensemble members and the MEAN in the FGOALS-g3 super-large ensemble. A pronounced cooling trend in the SPNA (10°W–50°W, 45°N–65°N; solid-border box in Fig. 4a) suggests that SST variability in this region may modulate fire-season precipitation in Eastern Siberia. The pattern of observed SST trends from 2001 to 2021 aligns with the MIN5-MEAN SST trend differences, with a notable cooling trend in the SPNA (Fig. 4b). Regression analysis of observed global SST anomalies against Eastern Siberian fire-season precipitation further identifies the SPNA as a key region associated with the precipitation change (Fig. 4c), possibly along with the South Pacific and the Northeastern Pacific (NEP; 120°W–160°W, 10°N–60°N; dashed-border box in Fig. 4b, c).

Fig. 4: Sea surface temperature origin of decadal change in Eastern Siberian fire-season precipitation.
Fig. 4: Sea surface temperature origin of decadal change in Eastern Siberian fire-season precipitation.
Full size image

a Composite differences in global sea surface temperature (SST) trends between the first five members of the FGOALS-g3 super-large ensemble with the largest decreasing (MIN5) fire-season precipitation (Pr) trends in Eastern Siberia (ES) and the ensemble mean (MEAN). b Spatial pattern of observed SST trends with global mean removed during 2001–2021. Green dots mark locations with statistically significant trends at the 0.05 level. Solid and dashed boxes denote the geographic boundaries of the subpolar North Atlantic and the Northeast Pacific, respectively. c SST anomalies associated with fire-season precipitation variations in Eastern Siberia during 1960–2021. Green dots indicate regression coefficients statistically significant at the 0.05 level. d Time series of observed fire-season precipitation anomalies in Eastern Siberia and observed subpolar North Atlantic SST anomalies after removing the global sea surface warming signal and applying a 9-year running mean during 1960–2021. The green dash line represents the 9-year running mean of Eastern Siberian fire-season precipitation anomalies. The Pearson correlation coefficient (r) and its p-value (p) between Eastern Siberian fire-season precipitation and decadal variations in subpolar North Atlantic SST are also shown.

Though the correlations between Eastern Siberian fire-season precipitation and South Pacific SST are significant in observations (Fig. 4c), these regions exhibit small SST trends over the past two decades (Fig. 4b). Thus, the decline in Eastern Siberian fire-season precipitation over this period is likely unrelated to variations in South Pacific SST. Unlike the South Pacific, the trend in NEP SST is large (Fig. 4b). However, a previous study suggested that there are lagged responses of the North Pacific (i.e., North Pacific Multidecadal Oscillation) to North Atlantic forcing via atmospheric teleconnections, amplified by inherent oceanic dynamics and positive air-sea feedbacks31. We show that this mechanism is consistent with our data by confirming that SPNA SST anomalies systematically precede NEP SST variability by 5–15 years, as exemplified by: cold SPNA phases (1901–1911; 1960–1995) followed by cold NEP phases (1905–1918; 1971–2014) and warm SPNA phases (1912–1959; 1996–2012) followed by warm NEP phases (1919–1970 and 2015–2021) (Supplementary Figs. 3 and 4). Thus, we propose that NEP SST is confounded with Eastern Siberian fire-season precipitation by SPNA SST. The correlation between Eastern Siberian fire-season precipitation and NEP decadal variability over 1960–2021 is rendered statistically non-significant when controlling for SPNA decadal variability (partial r = −0.12, p = 0.37). Therefore, we focus on SPNA SST as the primary driver of recent Eastern Siberian hydroclimate variability.

Multiple prior studies have reported substantial decadal variations in SPNA SST32,33,34. To illustrate the consistent decadal variations in SPNA SST and Eastern Siberian fire-season precipitation, we construct an SPNA SST decadal variability index by applying a 9-year running mean to area-weighted averages of SPNA SST anomalies relative to the long-term global-mean sea surface warming signal (Methods; Fig. 4d). The fluctuations of Eastern Siberian fire-season precipitation align closely with the decadal-to-multidecadal phases of SPNA SST, as evidenced by a statistically significant correlation coefficient of 0.42 (p < 0.05). Notably, the recent decline in Eastern Siberian fire-season precipitation coincides with the transition of SPNA SST from a warm to a cold phase around the late 2000s (Fig. 4d). Furthermore, extending the analysis to a longer period (1901–2021) using non-missing precipitation and SST data reveals a similar statistical relationship between SPNA SST and Eastern Siberian fire-season precipitation (Supplementary Fig. 4). Similar synchronous variations between Eastern Siberian fire-season precipitation and SPNA SST are observed using the GPCC precipitation dataset and the Hadley Centre SST dataset (Supplementary Fig. 5). Collectively, these results indicate that part of the decline in Eastern Siberian fire-season precipitation over the past two decades may stem from concurrent decadal cooling in the SPNA.

Mechanism of Subpolar North Atlantic influence on Eastern Siberian precipitation

To reveal the dynamic mechanism underlying the observed precipitation decline, we examine the model simulations that best captured the drying trend (i.e., the MIN5 ensemble). The spatial pattern of MIN5-MEAN simulated trend differences in precipitation (Fig. 3b) aligns with that of MIN5-MEAN simulated trend differences in 500-hPa eddy geopotential height (Z500, Fig. 3d). The geopotential height pattern features an anomalous eastward-propagating Rossby wave train, characterized by alternating positive and negative anomalies. The eastward-propagating wave train starts with a negative anomaly over eastern North America and the North Atlantic, which is succeeded by a positive anomaly over Western Europe, a negative anomaly over western Russia, and a positive anomaly over Eastern Siberia, finally terminating with a negative-positive anomaly pair over the high-latitude and western North Pacific, respectively (Fig. 3d). The spatial pattern of Z500 trends derived from reanalysis data during 2001–2021 mimics the wave train pattern in MIN5-MEAN Z500 trend differences (Fig. 3e), with a statistically significant pattern correlation coefficient (r = 0.46, p < 0.05). Discrepancies in the pattern and magnitude between the simulations and observations may arise from internal climate variability present in the observations that is not fully captured by the specific ensemble numbers selected for their Eastern Siberian precipitation signal. Consistent results are obtained using alternative precipitation and Z500 datasets (i.e., the Global Precipitation Climatology Centre (GPCC) precipitation dataset and National Centers for Environmental Prediction-Department of Energy Reanalysis-II dataset; Supplementary Fig. 6). Positive Z500 anomalies in Eastern Siberia favor anticyclonic conditions, which enhance atmospheric subsidence and reduce precipitation. These atmospheric circulation anomalies linked to internal climate variability may have directly caused the decline in fire-season precipitation over Eastern Siberia during the past two decades.

To elucidate the fundamental mechanism behind the SPNA origin of Eastern Siberian precipitation change, we regress the fire-season mean Z300 and Z500 anomalies derived from reanalysis data onto the observed decadal variations of SPNA SST. This analysis identifies quasi-stationary Rossby wave trains in the upper and middle troposphere, specifically the North Atlantic-Eurasian teleconnection pattern35 (Fig. 5a, b). The pattern comprises five activity centers located in the subtropical North Atlantic, the eastern region of the SPNA, Eastern Europe, Siberia, and North China. During the warm phase of SPNA SST, negative Z300 and Z500 anomalies are observed over Eastern Siberia (Fig. 5a, b). Associated with such anomalously low pressure, a cyclonic anomaly develops in the lower troposphere (Fig. 5c), which enhances atmospheric ascending motion (Supplementary Fig. 7). Additionally, given the climatological southward moisture gradient, anomalous southerly winds along the southern boundary of Eastern Siberia strengthen the transport of atmospheric moisture into the region (Fig. 5c). This enhanced moisture transport, combined with an influx of cold air from northwestern Eastern Siberia, creates favorable conditions for an increase in fire-season precipitation in Eastern Siberia (Fig. 5d). These dynamic changes explain the positive (negative) fire-season precipitation anomalies in Eastern Siberia during the warm (cold) phases of SPNA decadal variability.

Fig. 5: Physical links between subpolar North Atlantic sea surface temperature and Eastern Siberian fire-season precipitation.
Fig. 5: Physical links between subpolar North Atlantic sea surface temperature and Eastern Siberian fire-season precipitation.
Full size image

a, b Observed 300-hPa and 500-hPa eddy geopotential height anomalies (Z300 and Z500) associated with decadal variations in subpolar North Atlantic (SPNA) sea surface temperature (SST) from 1960 to 2021. Black dots indicate regression coefficients statistically significant at the 0.05 level. c Zonal and meridional wind anomalies at 700 hPa (UV700; curly vectors) related to decadal variations in subpolar North Atlantic SST, overlaid with climatological mean 700-hPa specific humidity (Q700) during 1960–2021. d As in a but for land precipitation anomalies.

To establish the mechanical link between decadal variations in SPNA SST and Eastern Siberian fire-season precipitation, we calculate the 300-hPa wave activity flux (WAF) to trace the propagation characteristics of stationary Rossby waves associated with SPNA SST decadal variability (Methods). The WAF is strengthened over the SPNA and subtropical North Atlantic, propagating eastward into Eastern Europe, Western Russia, and Siberia (Supplementary Fig. 8). This suggests that the SPNA may be a critical origin of the atmospheric wave train over high-latitude Eurasia that affects the atmospheric circulation in Eastern Siberia. To further validate whether and how SPNA decadal variability modulates Eastern Siberian fire-season precipitation, we perform an idealized sensitivity experiment using a linear baroclinic model (Methods). The linear baroclinic model is forced by a positive vorticity anomaly over the SPNA (Supplementary Fig. 9a) and climatological mean flow during the fire season. The prescribed vorticity anomaly triggers an anomalous atmospheric Rossby wave train across high-latitude Eurasia, with a notable cyclonic anomaly in Eastern Siberia (Supplementary Fig. 9b). This cyclonic anomaly enhances ascending motion and increases Eastern Siberian fire-season precipitation, consistent with observational patterns. These results indicate the role of SPNA decadal variability in regulating Eastern Siberian fire-season precipitation over the past two decades.

Discussion

Using causal inference, path analysis, and large ensemble simulations, we investigate the underlying climatic causes of the escalation of wildfires in Eastern Siberia over the past two decades. By analyzing the linear relationships between annual burned area fraction and summertime climatic factors, prior studies identified atmospheric and soil moisture deficits as the major drivers of burned area in Siberian forests11,15. However, they did not consider the key role of precipitation in regulating moisture deficits and its indirect effects on wildfires. We estimate that the decline in fire-season precipitation may account for ~42% of the increase in land area burned by wildfires, followed by the increases in VPD (about 36%) and daily maximum temperature (about 14%). Our results align with a recent study36 that identifies hydrological drought (characterized by decreasing rainfall) as the primary driver of wildfires in Siberia.

We show that recent changes in Eastern Siberian fire-season precipitation are linked to internal climate variability that has overcome the externally forced wetting trend, and identify the SPNA as a potentially key region. Decadal cooling in the SPNA could trigger a Rossby wave train across high-latitude Eurasia, which may have partly contributed to the fire-season drying trend in Eastern Siberia over 2001–2021. A previous study identified a negative correlation between northern North Atlantic SST and Northern Scandinavian fire activity37, which aligns with our findings.

Future work should examine more drivers of wildfires in East Siberia, including vegetation, fire management, and the variability of the IPO. A growing body of evidence indicates that wildfires, terrestrial vegetation, and local meteorology exhibit complex feedback mechanisms through land-atmosphere, fire-vegetation, and aerosol-weather interactions38. For instance, large-scale tree mortality triggered by extreme climate anomalies and extensive vegetation changes induced by wildfires can pose certain effects on regional or local moisture and thermal conditions. However, such effects are not considered in our analysis. Given the prevalence of positive feedback loops in climate-fire-vegetation interactions, omitting these mechanisms may artificially underestimate the impacts of individual climatic factors on burned area—a critical knowledge gap that requires further quantification via process-based modeling. Climatic factors are interrelated and can collectively influence wildfires, posing challenges for an accurate quantification of causal relations between these factors and wildfires. Non-climatic factors (such as forest fire management and human-induced ignitions) pose additional challenges for assessing their impacts on burned areas due to a lack of relevant data. While the IPO exhibits no significant trend over our study period (2001–2021), its potential influence on Eastern Siberian precipitation merits attention. The warm phase of the IPO has been linked to reduced precipitation in the region, supported by a modest historical correlation (r = −0.33, p = 0.1; Supplementary Fig. 10). Though not a dominant driver of recent trends, IPO variability may become increasingly relevant for future decadal-scale precipitation changes. Previous studies align with our findings that SPNA variability influences regional hydroclimates asymmetrically37,39. The SPNA warm phase enhances moisture transport to Northern Scandinavia (Fig. 5c), consistent with the observed wetting and reduced fire activity there37. Conversely, it weakens northward moisture transport into Western Siberia39, a pattern mirroring the precipitation decline we find in Eastern Siberia.

Though internal climate variability in fire-season precipitation of SST origin was the primary driver of increased burned areas in the recent two decades, daily maximum temperature also exerts a significant impact on Eastern Siberian wildfires that is not confounded by precipitation. As the climate continues to warm, the effect of daily maximum temperature on wildfires may become increasingly critical. Therefore, our findings underscore the necessity of reducing carbon emissions and curbing global warming to mitigate future wildfire risk in this region. Furthermore, our study highlights the need to distinguish the roles of anthropogenic climate change and internal climate variability in shaping the wildfire risk of other carbon-rich ecosystems. Such efforts would benefit the formulation of effective strategies to alleviate wildfire impacts on carbon balance, air quality, human health, and economy40,41.

Methods

Observations and reanalysis

We depict wildfire activity in Eastern Siberia (95°E–125°E, 55°N–67.5°N) using monthly gridded burned area data from the Terra and Aqua combined satellite MCD64A1 v6.1 product42, which has a 500-m spatial resolution. This dataset was generated using 500-m Moderate Resolution Imaging Spectroradiometer (MODIS) atmospherically corrected daily surface reflectance imagery and 1-km daily active fire observations42. Specifically, MODIS surface reflectance was used to construct a burn-sensitive vegetation index and probabilistic threshold, enabling dynamic identification of burned and unburned grid cells42. The burn date for each 500-m grid cell within individual MODIS tile was labeled in ordinal day format, with zero assigned to unburned land grid cells and additional particular values for missing and water grid cells.

To cross-validate the spatiotemporal patterns of burned area in Eastern Siberia derived from MCD64A1 v6.1, we also use the FireCCI51 global burned area dataset43. This dataset was produced based on high-quality Terra MODIS satellite data and near-infrared reflectance under the Fire_cci project of the European Space Agency. These two datasets show strong consistency in the interannual variations and regional distribution of Eastern Siberian burned area (Supplementary Fig. 11). For conciseness, only results from MCD64A1 v6.1 are presented in the main text.

Additionally, we use monthly Normalized Difference Vegetation Index (NDVI) data from the MODIS product (MOD13A3) with a horizontal spacing of 1 km to characterize the seasonal cycle of vegetation greenness in Eastern Siberia. To facilitate analysis, both burned area and NDVI data are upscaled to a 2.5° × 2.5° grid by calculating the fraction of each grid cell’s area undergoing combustion and the grid-averaged value of NDVI.

We use monthly precipitation, daily maximum, mean and minimum temperatures, actual vapor pressure (AVP), and frost day frequency data from the Climate Research Unit gridded Time Series (CRU TS v4.07) dataset44, which covers land areas at a 0.5° × 0.5° resolution. Additionally, we use near-surface wind speed data from TerraClimate45, a high-resolution global dataset of monthly climate variables. VPD is calculated as follows:

$${VPD}={SVP}-{AVP}$$
(1)
$${SVP}=610.78*{ex}{p}^{\left(17.2694 \;*\; T/\left(T+237.3\right)\right)}$$
(2)

where SVP denotes saturation vapor pressure, computed from daily mean temperature T in °C. Daily mean temperature and AVP are obtained directly from the CRU TS v4.07 dataset.

We use monthly surface-layer (0–10 cm) soil moisture data from the Global Land Evaporation Amsterdam Model (GLEAM) version 4.1a dataset46, which has a spatial resolution of 0.1°. To compute monthly SST anomalies and long-term trends from 1960 to 2021, we employ the National Oceanic and Atmospheric Administration Extended Reconstruction Sea Surface Temperatures version 5 (ERSSTv5)47 at a horizontal resolution of 2°. To examine the effects of atmospheric circulation and moisture distribution on precipitation, we use monthly data of geopotential height, zonal and meridional winds, and specific humidity from the National Centers for Environmental Prediction (NCEP)/National Center for Atmospheric Research Reanalysis-Ⅰ48, with a spatial resolution of 2.5°. To verify the robustness of our results, we use additional datasets: precipitation from the GPCC precipitation database49, SST from the Hadley Centre Sea Ice and Sea Surface Temperature dataset50, and atmospheric circulation and humidity data from the NCEP-Department of Energy Atmospheric Model Intercomparison Project Reanalysis-ⅠⅠ51. Precipitation data are remapped to a 2° × 2° grid via local areal averaging, while all other raw data are bilinearly interpolated to the same 2° × 2° grid. Our analysis focuses on the period 1960–2021. This choice is motivated by two factors: scarce precipitation observations in Eastern Siberia prior to 1960, and the fact that atmospheric reanalysis data have been well constrained by global regular radiosonde observations since the late 1950s.

Model simulations

We use a suite of model simulations from initial-condition large ensembles28,29 to assess the roles of external forcings and internal climate variability in recent Eastern Siberian fire-season precipitation changes. These ensembles, produced by seven climate models, were run under historical and future (RCP8.5 and SSP5-8.5) radiative forcing scenarios52(Supplementary Table 2). All the ensemble members were driven by identical external forcings but initialized with slightly varied initial conditions. To robustly estimate externally forced signals and sample internal climate variability, we select models with at least 40 ensemble members in both historical and future-scenario experiments.

Correlation analysis, causal inference, and path analysis

We first conduct correlation analysis to assess the relations between the Eastern Siberian burned area fraction and climatic factors during the fire season (from April to September). Given that pre-fire-season climatic conditions (from January to March) may have a lagged effect on subsequent fire-season wildfires, we also analyze correlations between fire-season burned area and pre-fire-season climatic factors.

To understand the causal links between fire-season burned area and key climatic factors, we construct Pearlian-style structural causal models19 based on prior knowledge and correlation analysis results, and validate the robustness of the causal diagram according to the d-separation criterion20. The correlation analysis shows statistically significant associations, such as a strong positive correlation between fire-season precipitation and burned area fraction, which inform the structural pathways of the model. We set the significance threshold at p = 0.1. This choice reflects the goal of causal discovery: not to prove the existence of significant causal effects but to propose a reasonable background model as a reference for path analysis. Therefore, omitting edges that truly exist is more detrimental than including edges that do not exist. Specifically, we connect all significantly correlated variable pairs in the diagram, considering their causal effects reflected by the significance of correlations. The direction of causal effects (i.e., the paths) is determined based on prior knowledge.

After establishing the preliminary causal graph structure, we validate the causal relations using observational data. This involves testing whether all variable pairs not directly connected in the graph exhibit a non-significant correlation when conditioning on the d-separating set implied by the graph. Additionally, we verify the significant correlations of precipitation and VPD with burned area fraction when conditioning on each other. Detailed results of these tests are provided in Supplementary Note 1.

To assess the direct and indirect effects of climatic factors on Eastern Siberian fire-season burned area fraction, we conduct a path analysis17,53, a form of multivariate regression method that requires prior knowledge and causal inference to establish the networks of links and functional pathways among variables. Supplementary Note 2 outlines the basic information on path analysis and its differences from standard multiple linear regression. In this path analysis, fire-season precipitation and VPD are considered as direct regulators of burned area (Supplementary Notes 1 and 3). Precipitation, along with daily maximum and minimum temperatures, may affect VPD, indirectly modulating burned area. A detailed rationale for the model structure is provided in Supplementary Notes 1 and 3. We use IBM SPSS AMOS version 28.0, a widely used software for structural equation modeling, to conduct the path analysis. The maximum likelihood estimation method is employed to estimate model parameters. This method is selected for its robustness and efficiency in data handling. To evaluate the goodness of fit of the path analysis model, we use six global fit metrics: chi-square degree of freedom ratio (CDFR < 5), goodness-of-fit index (GFI ≥ 0.90), root mean square error of approximation (RMSEA < 0.1), standardized root mean square residual (RMR < 0.05), comparative fit index (CFI ≥ 0.80), and incremental fit index (IFI ≥ 0.90). A model is considered reliable if at least three criteria are satisfied. We confirm the reliability of our path analysis model as all six metrics meet their respective criteria (Supplementary Table 3). Statistical tests validating the assumptions underlying the path analysis model are provided in Supplementary Note 3.

Subpolar North Atlantic decadal variability and the Interdecadal Pacific Oscillation index

SPNA decadal variability is defined by applying a 9-year running average to the time series of area-weighted means of residual SST anomalies over a sector of the northern North Atlantic (50°W–10°W, 45°N–65°N). Specifically, residual SST anomalies are calculated by regressing grid-level SST anomalies (relative to 1960–2021) onto global mean SST anomalies, thereby removing the local component associated with global mean sea surface warming. The IPO is defined as the 9-year running mean of the differences between the area-weighted mean SST anomalies over the central equatorial Pacific (170°E–90°W, 10°S–10°N) and the averages of area-weighted mean SST anomalies over the Northwest Pacific (140°E–145°W, 25°N–45°N) and Southwest Pacific (150°E–160°W, 50°S–15°S)54.

Given that 9-year running mean smoothing can introduce inherent autocorrelation in the signal, we compute the adjusted degrees of freedom (Nadj) using the formula for autocorrelation-corrected sample size55 when assessing the statistical significance of the correlation coefficient (r) related to the SPNA decadal variability or IPO:

$${{Nadj}}=N/\left(1+2*{\sum }_{k=1}^{N-1}\frac{\left(N-k\right)}{N}*{\rho }_{{XX},k}*{\rho }_{{YY},k}\right)$$
(3)

where N denotes the original degrees of freedom, \(\rho\) is the autocorrelation coefficient, and k is lag.

Then, we use the t-test statistics and the t-distribution with the adjusted degrees of freedom to evaluate the significance of r as follows:

$${{t}}={{r}}*\sqrt{\frac{{Nadj}-2}{1-{r}^{2}}}$$
(4)

Wave activity flux

To investigate the role of SPNA decadal variability in the propagation of stationary Rossby waves, we compute the WAF based on the formula developed by Takaya and Nakamura56. Its zonal (\({{{\bf{F}}}}_{x}\)) and meridional (\({{{\bf{F}}}}_{y}\)) components are calculated as follows:

$${{{\bf{F}}}}_{x}=\frac{p}{2\left|{{\bf{U}}}\right|{r}^{2}}\left\{\frac{{{\bf{U}}}}{\cos \phi }\left[{\left(\frac{\partial {\varPsi }^{{\prime} }}{\partial \lambda }\right)}^{2}-{\varPsi }^{{\prime} }\frac{{\partial }^{2}{\varPsi }^{{\prime} }}{{\partial \lambda }^{2}}\right]+{{\bf{V}}}\left[\frac{\partial {\varPsi }^{{\prime} }}{\partial \lambda }\frac{\partial {\varPsi }^{{\prime} }}{\partial \phi }-{\varPsi }^{{\prime} }\frac{{\partial }^{2}{\varPsi }^{{\prime} }}{\partial \lambda \partial \phi }\right]\right\}$$
(5)
$${{{\bf{F}}}}_{y}=\frac{p}{2\left|{{\bf{U}}}\right|{r}^{2}}\left\{{{\bf{U}}}\left[\frac{\partial {\varPsi }^{{\prime} }}{\partial \lambda }\frac{\partial {\varPsi }^{{\prime} }}{\partial \phi }-{\varPsi }^{{\prime} }\frac{{\partial }^{2}{\varPsi }^{{\prime} }}{\partial \lambda \partial \phi }\right]+{{\bf{V}}}\cos \phi \left[{\left(\frac{\partial {\varPsi }^{{\prime} }}{\partial \phi }\right)}^{2}-{\varPsi }^{{\prime} }\frac{{\partial }^{2}{\varPsi }^{{\prime} }}{{\partial \phi }^{2}}\right]\right\}$$
(6)

where p is the ratio of air pressure between a specific level and 1000 hPa. r, ϕ, and λ represent the Earth’s radius, latitude, and longitude, respectively. U and V denote geostrophic zonal and meridional winds, and |U| represents basic-state wind velocity. \({\varPsi }^{{\prime} }\) denotes the geostrophic stream function anomaly associated with SPNA decadal variability.

Linear baroclinic model simulation

We use the linear baroclinic model57 to examine atmospheric responses to SPNA decadal variability. The linear baroclinic model is set with a horizontal resolution of spectrum triangular truncation at wavenumber 42 and a sigma coordinate of 20 vertical levels. We integrate the model with ellipse-shaped positive vorticity anomalies over the SPNA and climatological mean flows during the fire season for 50 days and attain an equilibrium solution by ~20 days. Therefore, we use the average of the final 30 days to represent the linear steady atmospheric circulation responses to the vorticity perturbation associated with positive SPNA SST anomalies.