Winter snow as a skillful indicator for spring wildfire in Northeast Asia

Abstract

Wildfire in Northeastern Asia presents significant challenges to climate, ecosystems, and socio-economic systems. Here, we investigate the influence of preceding winter snow on spring wildfires in Northeastern Asia. Our results reveal soil moisture serves as a key mediator linking winter snow to spring fires. Anomalously low (high) snowfall during winter leads to drier (wetter) land surfaces in spring, thereby increasing (decreasing) spring wildfire risk. A multiple linear regression model based on this mechanistic association achieves a robust four-month lead forecast of spring burned area, with a prediction-observation Spearman correlation of 0.65 (p < 0.01), and explains 41% of the variability in the spring burned area across Northeast Asia. Cross-validation further confirms the significance and reliability of this lead-time relationship. Our findings highlight the pivotal role of winter snow in modulating spring wildfire activity in Northeastern Asia and demonstrate the potential to leverage snow conditions to improve seasonal wildfire forecasts.

Introduction

Wildfire constitutes a major disturbance in the global ecosystem^1,2, altering surface vegetation and releasing vast quantities of greenhouse gases, aerosols, and trace gases into the atmosphere. It also exacerbates soil erosion and air pollution, thereby exerting a crucial impact on environmental and climate security^3,4,5,6. Boreal forests, which cover approximately 35% of the world forested area⁷, contribute up to 600 MW of radiative heat annually and exhibit some of the highest fire radiative power on Earth⁸. Under climate warming, the wildfire risks in boreal forests have generally been on rising^9,10,11. During the time period of 2000–2005, the area of the boreal forests decreased by 4%, with wildfires accounting for 60% of this reduction¹².

Meteorological conditions, such as temperature, precipitation, drought, heatwaves, and wind, play crucial roles in affecting wildfire behaviors^13,14,15. For example, rising temperatures coupled with declining precipitation substantially elevate ignition risks¹⁶. Enhanced atmospheric aridity further amplifies fire activity. In North America, the increase in extreme wildfire events and the acceleration of fire spread in North America were closely linked to a decline in relative humidity, which explained approximately 75% of the variability in these aspects of wildfire behavior^17,18.

Besides, snow dynamics are also key modulators of wildfire activity. The extent and duration of snow cover, as well as the timing of melt, influence fire risk through multiple pathways^19,20. Early snowmelt, for instance, often results in reduced spring soil moisture²¹. It also triggers an earlier onset of vegetation greening²², implying vegetation emerges earlier and is then subject to a longer drying season. These all can increase fire susceptibility^23,24. Moreover, smoke particles from wildfires, particularly dark-brown carbon, can accelerate snowmelt by enhancing shortwave radiation absorption over snow-covered surfaces^25,26,27. Evidently, a negative feedback relationship exists between wildfires and snow cover. Nonetheless, most existing studies have focused on their concurrent associations, while lead–lag relationships that could underpin seasonal wildfire predictability remain underexplored.

Northeast Asia is highly susceptible to wildfires, but a detailed understanding of fire environmental drivers, particularly in the context of prediction, remains limited. In the mid-to-high latitudes of the Northern Hemisphere, snow typically persists through winter and melts as temperatures rise in spring, offering a potential early season indicator of wildfire risk. Given the strong climate memory and broad environmental influence of snow, we propose exploring the association between the December–March snow and April–May wildfires in Northeastern Asia. This paper is organized as follows: we first explore the relationship between winter snow and spring wildfires in Northeast Asia, and a potential mechanism is proposed. Subsequently, based on the uncovered physical connections, a statistical model is constructed to forecast the spring burned area in Northeastern Asia. Finally, we provide the discussion part.

Results

The spatial-temporal pattern of wildfire in Northeastern Asia

Severe wildfires are observed commonly across the Northeast of Asia, as indicated by the spatial distribution of the 20-year (1997–2016/2001–2020) averaged burned area in spring (Fig. 1a, b). Over the past two decades, 18% of grid within the study domain (45°–57°N and 120°–130°E) have experienced a burned area fraction exceeding 0.04%. Similarly, 20% of the region exhibits burned areas larger than 32 km². Seasonally, wildfire activity peaks in April and May (Fig. 1c), with mean burned fractions reaching 0.03% and 0.02%, and corresponding burned areas averaging 23.3 km² and 26.5 km², respectively. Northeastern Asia is characterized by a cold temperate continental monsoon climate. From 2001 to 2020, the mean spring precipitation was approximately 36 mm, with only 23 mm occurring in April. During this period, temperatures rise sharply from subzero winter levels to over 8 °C in early spring. This rapid warming, combined with low precipitation, results in reduced atmospheric humidity, mean relative humidity values stand at 49.3% in April and 50.7% in May. Such climatic conditions converge to induce spring wildfire risk, making Northeast Asia particularly susceptible to fire outbreaks during this season.

Fig. 1: Spatial and temporal pattern of wildfire and associated environment conditions.

Spatial distribution of the mean burned area fraction (a) and burned area (b) in April and May over Northeast Asia. Long-term seasonal variations in the burned area fraction, burned area, air temperature, precipitation and relative humidity for each month in Northeastern Asia (c), with units of %, km², °C, mm and %, respectively. Burned area fraction and burned area are sourced from GFED 4.1 s and FireCCI51, with data periods of 1997–2016 and 2001–2020, respectively.

Significant association between snow and Northeast Asian wildfire

To assess the interannual relationship between antecedent snow conditions and subsequent spring wildfires in Northeast Asia, we employed Spearman rank correlation to quantify their associations. Given that snow cover typically persists from autumn through early spring in this region, a 6-month lag was introduced to capture potential delayed effects of snowpack on fire activity. Significant negative correlations were observed between December and March snow water equivalent (SWE) and both burned area fraction (from GFED4.1 s) and burned area (from FireCCI51) during April–May (Fig. 2a). For GFED4.1 s, the coefficients are –0.57 (p < 0.01), –0.48 (p < 0.05), –0.58 (p < 0.01), and –0.64 (p < 0.01), respectively. Corresponding values for FireCCI51 are –0.48 (p < 0.05), –0.45 (p < 0.1), –0.45 (p < 0.1), and –0.49 (p < 0.05). When averaged over the December–March period, SWE is also significantly correlated with both burned area fraction (r = –0.57, p < 0.01) and burned area (r = –0.48, p < 0.05).

Fig. 2: Association between the burned area and snow cover.

a Correlation coefficients between the spring burned area fraction, burned area and snow water equivalent from October to May over Northeast Asia. b is the same but correlated with the snow depth derived from the ERA5 datasets. *, ** and *** denotes the significant correlation at 0.1, 0.05, and 0.01 level, respectively.

We employed snow depth datasets to further assess the robustness of such association (Fig. 2b). Similar negative associations were found. For burned area fraction, the correlation coefficients with snow depth from December to March are –0.47 (p < 0.05), –0.44 (p < 0.1), –0.48 (p < 0.05), and –0.40 (p < 0.1), respectively. For burned area, the coefficients are –0.49 (p < 0.05), –0.44 (p < 0.01), –0.46 (p < 0.05), and –0.50 (p < 0.05), respectively. Averaged across the four-month period, the correlation with snow depth remains significant for both burned area fraction (r = –0.45, p < 0.05) and burned area (r = –0.49, p < 0.05). Together, these results reveal a statistically robust and temporally consistent inverse relationship between winter snow cover and spring fire activity over the past two decades. The above highlights snow as a valuable precursor of spring wildfire risk in Northeast Asia, with a potential predictive lead time of up to four months.

Possible physical mechanisms

Building on the observed strong correlation between spring wildfires and antecedent snow cover from December to March, we next elucidate such links from the perspective of physical mechanisms and hypothesize that the preceding winter snow affects spring wildfire by modulating simultaneous climate factors. We explore the correlation coefficients between the burned area fraction and the meteorological variables, including precipitation, soil moisture, air temperature, relative humidity, VPD and wind, across different datasets (Fig. 3a). A significant negative correlation exists between the burned area fraction and precipitation, soil moisture, as well as relative humidity. Conversely, a significant positive correlation is observed between the burned area fraction and the VPD. For precipitation derived from CRU, CMAP, GPCP and ERA5 datasets, the respective coefficients are –0.46 (p < 0.05), –0.39 (p < 0.1), –0.42 (p < 0.1) and –0.38 (p < 0.1). Regarding soil moisture sourced from GLDAS, ERA5, CPC and GLEAM, the corresponding coefficients are –0.76 (p < 0.01), –0.61 (p < 0.01), –0.69 (p < 0.01) and –0.43 (p < 0.1), respectively. Moreover, for relative humidity and VPD, the corresponding coefficients are –0.65 (p < 0.01) and 0.67 (p < 0.01), respectively. Wind speed exhibits a negligible impact on spring fires variability. Consistent results are observed when analyzing burned area datasets (Fig. 3b). These findings indicate that soil moisture is likely a more critical factor in governing the variability of spring wildfires in Northeastern Asia.

Fig. 3: Climate factors related to the burned area in spring.

Correlation coefficients between the spring burned area fraction (a), burned area (b) and simultaneous regional climate factors derived from different datasets. The dot–dash denotes a significant correlation coefficient at 0.05 and 0.01 level, respectively.

Snow serves as a vital water reservoir for terrestrial ecosystems, exerting a key hydrological influence on spring soil moisture recharge. When winter snowfall is below average, reduced snowmelt may result in inadequate soil water replenishment in spring²⁸. This effect is particularly acute in Northeast Asia, where low soil moisture, coupled with limited spring precipitation, exacerbates drought risk. Moreover, reduced snow accumulation lowers surface albedo, amplifying solar absorption and accelerating snowmelt. The resulting early melt not only reduces water retention but also enhances evaporative losses, further drying the land surface. These compounded effects intensify surface aridity and, in turn, elevate wildfire risk. We therefore propose that winter snow cover influences spring wildfire activity primarily through its modulation of soil moisture dynamics in Northeast Asia.

To test our hypothesis, we examined the relationship between winter snow anomalies (December–March) and spring (April–May) in situ soil moisture across Northeast Asia. We employed the soil moisture datasets from the CPC and GLDAS to evaluate the reliability of this linkage. Results indicate that a one-unit-standardized increase in snow water equivalent during winter is associated with a spring soil moisture increase of approximately 15 mm in the CPC data (Fig. 4a, shading) and 2 kg/m² in the GLDAS data (Fig. 4b, shading), with spatial patterns and magnitudes showing strong consistency across datasets. Temporal analyses (Fig. 4c, d) further highlight a clear lagged response, whereby winter snow anomalies significantly influence subsequent soil moisture conditions in spring. Notably, the correlation between winter snow and spring soil moisture is substantial, and the coefficients reach 0.57 (p < 0.01) for snow water equivalent and 0.66 (p < 0.01) for snow depth.

Fig. 4: Relationship between the December–March snow cover and April–May soil moisture.

Regression coefficients of the normalized December–March snow water equivalent and snow depth upon the April–May soil moisture (SM) derived from the CPC (a) and GLDAS (b) datasets, respectively. (c) and (d) represents the temporal correlation between the soil moisture and snow water equivalent (SWE) and snow depth (SD), respectively. In (a) and (b), contours represent the soil water content anomalies associated with the snow depth and excluding the spring precipitation (PRCP). Only values significant at the 0.05 level are shown.

Note that spring soil moisture is partially contributed by the concurrent precipitation, which may obscure the true statistical relationship between the winter snow and spring soil moisture. To mitigate this, we removed the spring precipitation signal from the in situ soil moisture time series across Northeast Asia. Following this adjustment, in association with one-unit-standardized changes of December–March snow depth, the April–May residual anomalous pattern and maximum center remain largely consistent (contours in Fig. 4a, b). In terms of temporal variation, the residual soil moisture time series also exhibits significant correlations with snow water equivalent (Fig. 4c) and snow depth (Fig. 4d), with coefficients of 0.64 (p < 0.01) and 0.65 (p < 0.01), respectively. These results underscore that winter snow anomalies exert a pronounced influence on spring soil moisture variability.

Predicting model

In Northeast Asia, a significant and physically robust relationship existed between preceding winter snow cover and spring wildfires. Given that our previous study²⁹ demonstrated that the winter North Atlantic sea surface temperature (SST) tripole pattern could be a precursor for spring wildfires in Northeastern Asia. That is, the preceding winter SST is characterized by negative SST anomalies in the midlatitude and tropical North Atlantic and positive SST anomalies in the subtropical North Atlantic, and such a structure can persist to spring, resulting in a significant high-pressure anomaly over southeastern Siberia, which corresponds to decreasing precipitation and enhances the wildfire risk. Motivated by these findings, we develop a multivariable linear regression model to predict spring wildfire activity in Northeast Asia, using winter snow cover and the SST tripole index as predictors, and burned area fraction (BAF) as the target variable. The tripole pattern index of the winter North Atlantic SST is defined as the difference between the mean SST of a positive center (between 30°–40°N and 280°–310°W) minus an averaged SST in the two negative centers (approximately 10°–20°N and 290°–330°W, and 45°–65°N and 300°–315°W, respectively).

We first built separate regression models using either the SST or SWE as predictors. The SST-only model explained 20% of the interannual variance in spring burned area, with a significant correlation of 0.47 (p < 0.05) between the predicted and observed burned area. By comparison, the SWE-only model accounted for 36% of the variance, yielding a higher correlation of 0.57 (p < 0.01). When both predictors were combined, model performance improved substantially, explaining 41% of the variance and reaching a correlation of 0.65 (p < 0.01). This prediction improvement was particularly evident in specific years (e.g., 2004, 2007, 2009, 2010 and 2016; Fig. 2 in Supporting Information S1). Consequently, the combined SST and SWE model was ultimately selected for subsequent analysis (Table 2 in Supporting Information S1). These findings indicate spring wildfire activity in Northeast Asia can be skillfully predicted up to four months in advance, based on physically interpretable predictors including snow water equivalent and North Atlantic SST.

To assess the predictive performance of the regression model and minimize the risk of overfitting associated with using all available samples, we conducted a leave-k-out cross-validation to hindcast spring burned area fraction (BAF) in Northeast Asia. For k values ranging from 1 to 4 (Table 1), the cross-validated correlation between predicted and observed BAF remains at 0.56 ~ 0.57 (p < 0.05), with R² values stable at 0.41. This indicates that the model is robust. Notably, the final leave-k-out predicted values represent the mean across all iterations, thereby reducing variability in the hindcasts. Overall, these findings show that the preceding winter snow and North Atlantic SST have indictive significance for wildfires in Northeastern Asia, and such association is significant and robust.

Table 1 Spearman correlation coefficient between the observed BAF and leave-k-out validation BAF from 1997 to 2016

Discussions

Our findings demonstrate that anomalous winter snow exerts a significant influence on spring wildfire activity in Northeast Asia, primarily through its regulation of soil moisture. The statistical linkages among December–March snow cover, North Atlantic sea surface temperatures (SST), and April–May burned area are consistent with established physical mechanisms. Previous studies highlighted that tropical climate factors can also impact mid-high latitude regional climates. For instance, Wu et al.³⁰ demonstrated that El Niño-Southern Oscillation (ENSO) modifies midlatitude circulation via South Asian heating, thereby influencing temperatures in northeastern China prior to the late 1970s. To evaluate whether ENSO contributes to the predictability of spring wildfires in Northeast Asia, we removed its signal from both the wildfire and snow water equivalent time series. The model performance remained largely unchanged, with R² values stable at 0.42, indicating that ENSO has a negligible impact on the model’s predictive skill (Table 1 in Supporting Information S1).

Additionally, the model’s capability to estimate extreme values of the burned area in Northeast Asia needs to be evaluated. Extreme events were defined as those with standardized burned area fraction (|BAF | ) exceeding 0.7. Based on this criterion, peak fire years were identified as 2003, 2006, 2008, and 2015, whereas 2004, 2005, 2007, 2010, and 2013 were classified as low-fire years. The model skillfully identifies the timing of these extreme events, but it systematically underestimates the magnitude of high-fire years and overestimates that of low-fire years (Fig. 1 in Supporting Information S1). These biases suggest that, while the model captures the occurrence of extremes, its performance in reproducing their intensity remains limited.

Over the past two decades, Northeast Asia has experienced increasingly severe surface drought conditions. Both snow water equivalent and snow depth exhibit a widespread declining trend across the region, with maximum rates of decrease reaching –2 mm yr⁻¹ and –1.6 mm yr⁻¹ (Fig. 5a, b), respectively. Their temporal evolution further supports this downward trajectory (Fig. 5c, d), reflecting a persistent reduction in seasonal snow accumulation and associated surface moisture availability. These findings are consistent with previous studies³¹.

Fig. 5: Trend of the snowfall.

(a, c) denote interannual trend of December–March snow water equivalent during 1997–2016. (b, d) is the same but for snow depth from 2001 to 2020. The cross symbols represent variation trend values are significant at 0.1 level.

Projections from the Coupled Model Intercomparison Project Phase 6 (CMIP6) suggest that this trend is likely to continue, with pronounced losses in winter and early spring snow cover over northern China and eastern Siberia^32,33. The diminishing seasonal snow is expected to curtail snowmelt-driven infiltration, thereby reducing spring soil moisture recharge. In parallel, climate projections indicate a strengthening of anticyclonic anomalies over Northeast Asia under continued global warming³⁴, which will likely exacerbate regional warming and drying. These changes are expected to intensify surface aridity, elevate fuel flammability, and amplify the risk of spring wildfire activity in a warming future.

However, short-term weather events, such as abrupt warming, drought episodes, or cold spells, can also exert immediate and sometimes overriding influences on wildfire activity^35,36. While our findings emphasize the critical role of winter snow in shaping seasonal fire dynamics, future research should seek to integrate high-resolution meteorological and fire datasets to more explicitly quantify how transient weather variability interacts with snow-related legacies in modulating spring wildfire behavior in Northeast Asia.

The Northeast Asia ecosystem experiences frequent wildfires that exert substantial ecological and climatic impacts³⁷. Following the catastrophic 1987 fires in the Great Xing’an Mountains, China has implemented aggressive fire suppression policies, resulting in a marked decline in wildfire activity over recent decades³⁸. Integrating snow-based seasonal outlooks can enhance anticipatory fire management. For example, controlled burns conducted in alignment with these snow-driven predictions can reduce fuel accumulation prior to high-risk conditions, thereby lowering the probability of large-scale wildfires. Our findings provide insights for adaptive wildfire management in snow-dominated boreal regions.

Methods

Burned area data

To characterize wildfire activity, we made use of two monthly burned area datasets. The first, from the Global Fire Emissions Database version 4.1 s (GFED4.1 s), which incorporates small fire detections at a spatial resolution of 0.25°, covering the time span from 1997 to 2016³⁹. The second dataset, covering 2001 to 2020, is sourced from the European Space Agency Climate Change Initiative version 51 (CCI51), also with a spatial resolution of 0.25°, based on the Moderate Resolution Imaging Spectroradiometer (MODIS) observations aboard the Terra satellite⁴⁰. To guarantee reliability, the confidence level layer is applied to the selected burned area regions.

Climate data

To explore the impact of regional climate factors on wildfire occurrence in Northeast Asia, we examined the relationships among the precipitation, soil moisture, air temperature, relative humidity, vapor pressure deficit (VPD), wind and burned area, respectively. Four monthly precipitation datasets are employed. The first is obtained from the Climatic Research Unit Time-series dataset (CRU-TS version 4.04), featuring a spatial resolution of 0.5°. The second, the Climate Prediction Center Merged Analysis of Precipitation (CMAP), spans 1979–2020 with a resolution of 2.5°⁴¹. The third precipitation dataset, the Global Precipitation Climatology Project (GPCP version 2.3), offers gridded precipitation estimates at 2.5° resolution, based on in-situ observations and satellite precipitation data⁴². The fourth dataset is the total precipitation derived from the fifth generation European Centre for Medium-Range Weather Forecasts Re-Analysis data (ERA5), downloaded at a spatial resolution of 0.25°⁴³.

Four soil moisture datasets were used. The first is sourced from the Climate Prediction Center (CPC) soil moisture data, which provides soil moisture data at a spatial resolution of 0.5° and represents the water content of the entire soil column⁴⁴. The second dataset is from the Global Land Data Assimilation System version 2 (GLDAS v2), offering 0.25° resolution and representing surface-layer soil moisture derived from model–observation fusion⁴⁵. The Global Land Evaporation Amsterdam Model surface soil water dataset (GLEAM) is also applied, with a time period of 1980–2021. It integrates reanalysis radiation and air temperature, as well as a combination of gauge-based, reanalysis and satellite-based precipitation^46,47. Lastly, the ERA5 volumetric soil water layer 1 with a spatial resolution of 0.25° was used⁴³. This dataset incorporates information regarding soil classification, soil depth, and groundwater levels.

We used two monthly air temperature data in this research. The first, CRU-TS dataset, is derived from station observations and offers a spatial resolution of 0.5°. The second dataset comprises 2-meter air temperature from the ERA5 reanalysis, downloaded with a spatial resolution of 1.5°⁴³.

Relative humidity and wind datasets are also derived from the ERA5 reanalysis data. Additionally, we also utilize the VPD to depict atmospheric humidity variations during the fire season in Northeastern Asia. The VPD describes the difference between the saturation and the actual water vapor pressure. As a composite indicator, it can reflect the combined effects of temperature and humidity on the force driving water loss from plants^18,48. The VPD values are computed using air temperature and relative humidity data from the ERA5 dataset⁴⁹.

Weekly snow water equivalent (SWE) data for the Northern Hemisphere are obtained from the National Snow and Ice Data Center. The SWE data originates from the European Space Agency GlobSnow-1 dataset, which combines time-series measurements from two space-borne passive microwave sensors. The original SWE data are projected onto the Equal-Area Scalable Earth Grid, and we re-grid the dataset to a 0.25°spatial resolution. Moreover, the snow depth (SD) dataset obtained from the ERA5 reanalysis is also applied to monitor snow variations in the Northeast of Asia, representing the instantaneous grid-box average of snow thickness on the ground.

Analysis

Spearman rank correlation analysis is adopted to investigate the connections between wildfire and key meteorological elements, with statistical significance reported using p-values. Additionally, a multiple linear regression model was developed to predict interannual variations in wildfire activity across Northeast Asia, based on winter snow fall and North Atlantic sea surface temperatures. The model takes the following form:

$${{\rm{BA}}}_{{\rm{Predicted}}}={a}_{1}{\rm{Snow}}+{a}_{2}{\rm{SST}}+{a}_{0}$$

Here, a₀, a₁ and a₂ denote the constant term and regression coefficients, respectively. R-square (R²) serves as the metric to assess the model performance. And we use Spearman correlation coefficient to examine the relationship between the observed and the predicated burned area values.

Furthermore, the leave-k-out cross-validation method is applied to evaluate the predictive skill of the model. This method ensures each data point contributes fully to model validation and maximizes data utilization, which is particularly important for temporally limited datasets. Specifically, k samples from n samples are deleted as the validation data. Subsequently, the remaining n-k samples are used to derive a forecasting model, and the removed samples are employed to test the built forecasting model. This process is repeated n/k times until forecasting values are obtained for all the removed samples. To balance over-fitting risk and optimize data usage, the k values are typically set within the range of n/10~n/5⁵⁰. In this study, the n value is 20.

We utilize the Theil-Sen analysis to investigate the snow trend over the past two decades. This robust non-parametric statistical method does not assume a specific probability distribution for the data, making it well suited for environmental time series⁵¹. To evaluate the statistical significance of the trends, the Mann–Kendall test is employed. Tt does not necessitate that the variables conform to a normal distribution, and is resilient to missing values and outliers⁵².

To isolate the influence of winter snow on spring soil moisture, independent of concurrent precipitation effects, we follow the approach of Gong et al.⁵³. We use a linear regression technique to remove the signals of spring precipitation on the interested variables. Initially, the spring precipitation is regressed on the soil moisture in Northeastern Asia. Then the estimated spring precipitation-related components are subtracted from the original soil moisture time series, and the residuals represent spring precipitation-free parts and are yielded to the subsequent analysis. All spring data are calculated as the average of April and May values, while winter data are determined as the mean of December, January, February, and March values. We only focus on the interannual variations of the wildfire and climate factors, and all datasets are detrended prior to conducting the analysis.