Fires, mostly natural but could also have anthropogenic origins, are a recurring phenomenon in various ecosystems over Southeast Asia, particularly during the boreal spring seasons1. These fires have significant impacts on vegetation type, composition, landscape structure, and ecological processes. These fires also have substantial impacts on human health and well-being, primarily through exposure to heat flux, emissions, and altered ecosystem functioning2,3. Tropical ecosystems in Asia, characterized by dry deciduous, thorn, and mixed deciduous forests, are particularly susceptible to fire disturbances because of the increasing reliance of local communities on these forests4. The severity of the 1997/1998 wildfire season serves as an example of this vulnerability, with an estimated emission of about 1 Gt of carbon into the atmosphere5,6. As global warming leads to the intensification of heatwaves and droughts, it is anticipated that more frequent wildfires will worsen and last longer in Southeast Asia, with far-reaching effects7,8.

Fire occurrences in Southeast Asia exhibit notable variations over interannual to decadal timescales and are closely linked to climate fluctuations9. Notably, El Niño-Southern Oscillation (ENSO)-induced exacerbated dry conditions have a major influence on these fires. During 1982/1983, 1997/1998, and 2015, there were three significant forest fires in Sumatra and Borneo, each of which coincided with the warm phase of ENSO10,11. Additionally, the joint impact of ENSO and the Indian Ocean Dipole (IOD) on forest fires in Southeast Asia has been identified, with the warm phase of IOD exacerbating fire activity when combined with El Niño events12. This condition might be more complicated when considering significant correlations identified between ENSO and the IOD13. However, previous research has primarily focused on the role of atmospheric and oceanic variability in the tropics and low latitudes, whereas the linkage between Southeast Asian fires and high-latitude climate variability is not well recognized.

There has been a notable decline in Arctic sea ice since the late 1970s, as a result of significant Arctic warming, or the Arctic amplification14,15. This change can have far-reaching effects beyond the high-latitude region16. For example, previous studies have suggested that the Arctic sea ice reduction could impose a substantial impact on the regional climate in Southeast Asia, with the reduction of sea ice over the Barents Sea during the autumn leading to intensified droughts in Southeast Asia during winter by changes of jet stream characteristics and modulation of East Asia Winter Monsoon (EAWM) strength and its associated northerly wind anomalies17,18. The Arctic sea ice also has a profound impact on other low-latitude areas by anomalous meridional circulation and Rossby wave propagation19. Nevertheless, a thorough and quantitative evaluation of the impact of the Arctic on the fire activities over Southeast Asia, as well as on the population exposure to fires is still lacking. Investigating possible connections between Arctic sea ice and fire occurrences in Southeast Asia is critical given the mounting evidence of connections between high-latitude climate change and mid-to-low-latitude weather and environment in recent decades20,21. Furthermore, considering the dense population in Southeast Asia and the significant socioeconomic and environmental impact of fires, it is crucial to comprehensively consider the population exposure to potential fire weather changes associated with high-latitude variability22.

The objective of this research is to establish a connection between fire weather and its population exposure in Southeast Asia and Arctic variability, especially the change of sea ice. To investigate the potential mechanisms underlying this linkage, we will perform model simulations and analyze observational and reanalysis datasets. We further quantify the population under fire weather exposure in decreased sea ice concentration (SIC) considering its social-economic impacts. Our results will shed light on the forecasting of Southeast Asia fires and emphasize the vital role of Arctic climate variability in driving mid-to-low-latitude climate and environmental changes.

Results

Teleconnection between Southeast Asia fire potential and Beaufort SIC

The domain of the Southeast Asian region in this study includes the latitudes of 70°E to 110°E and 10°N to 30°N (Fig. S1). Note that this domain contains mainland Southeast Asia and South Asia, which slightly differs from the conventional definition of Southeast Asia. The reason for selecting this region is that the peak seasons of fires in mainland Southeast Asia and equatorial Southeast Asia are different but those for mainland Southeast Asia and South Asia are similar11. Based on Fig. S2 and an earlier study1, we consider late winter/early spring (February, March, and April, FMA) as the peak fire season and focus our analysis on fire activities and meteorological variables in these three months. Note that in all of the following analyses, the time series have all been detrended. Therefore, the relationship established between Beaufort SIC and the regional fire weather index (FWI, the index measuring the weather conditions for fire, see “Methods” section) should not be influenced by the long-term trend of global warming. Moreover, in order to obtain a more robust conclusion, we estimated all p-values by considering the autocorrelation using the method of the previous study23.

A lagged correlation analysis between the FMA FWI of Southeast Asia and Arctic SIC (120°W-160°W, 75°N-85°N, outlined in Fig. S3a) in the August, September, and October (ASO) of the preceding year reveals that there is the strongest (negative) correlation over the Beaufort region (Fig. S3a). Figure S3a also indicates that there is no such strong correlation between FMA FWI over Southeast Asia and the other regions of the Arctic Ocean, suggesting that Beaufort SIC is the most likely the main remote source of oceanic forcing influencing Southeast Asian climate in FMA, more likely to have far-reaching consequences. We thus focus on Beaufort SIC in the following discussion.

The time lag could be explained by the cumulative effect of ice-albedo feedback in the polar region24. This time lag can also be attributed to the time delays inherent in the transfer of oceanic heat to the atmosphere and the subsequent adjustment of atmospheric circulation patterns in response to the modified surface conditions25,26,27. The Beaufort has the lowest SIC as well as the highest interannual variability in ASO, as illustrated in Fig. S3b. Moreover, there is also a strong correlation (R = −0.68) between ASO Beaufort SIC and FMA Beaufort T2M, although weaker than a one-month lag as illustrated in the previous study, this correlation is still relatively strong (Fig. S3c)24. As such, we will primarily focus on ASO Beaufort SIC in our following analysis.

Analysis using various fire proxies, including smoke records, fire radiative power (FRP), burned area from satellite observation (satellite BA), global fire emissions database version 4 (GEFDv4 BA), burned area fraction from global fire emissions database version 4.1 (GEFDv4.1 BAF), and fire count (FC), along with Singular Value Decomposition (SVD) analysis between Southeast Asia’s FWI and Beaufort SIC (Fig. 1 and Fig. S4), consistently reveals a strong negative relationship between FWI and SIC. We applied SVD instead of Empirical Orthogonal Function (EOF) because EOF typically analyzes the spatiotemporal patterns of a field, while SVD can assess the joint variability and interaction between two fields (i.e., Southeast Asia’s FWI and Beaufort SIC). The first SVD, which accounts for more than 85% of the total covariance, demonstrates strong negative signals over the Beaufort Sea and positive signals across Southeast Asia. The principal components (PCs) of these two variables show a moderate negative correlation of −0.54. Notably, in 2014, the FWI PC experienced a significant surge, while the SIC PC saw a dramatic decrease, as evidenced by the negative trend in the SIC time series in Fig. S5c for that year. We also note that the negative relationship between the PCs of these two variables has been enhanced in recent decades, primarily due to the higher variability of both SIC and FWI time series. Moreover, over the past decade, the regionally averaged FWI exhibits a marked increase (Fig. S5a), while the SIC shows a notable decline (Fig. S5b), aligning well with the trends observed in the PCs. All these linear trends have passed the significance test with p ≤ 0.05 with the Mann–Kendall method.

Fig. 1: Correlation between various fire proxies in Southeast Asia and Beaufort SIC.
figure 1

a The correlation coefficient distribution between FMA Southeast Asia FWI and ASO Beaufort SIC from 1980 to 2019. b Like (a), but for FMA burned area from the satellite observations (2001–2019). c Similar to (a), but for FMA burned area from GEFDv4 (1996–2016). d Similar to (a), but for FMA burned area fraction from GEFDv4.1 (1997–2015). e Similar to (a), but for the FMA Fire Count (2003–2019). f Similar to (a), but for the smoke records at weather stations (1980–2019, only stations with p ≤ 0.1 are displayed). The Student’s t-test indicated that the region containing white dots passed the significance test with p ≤ 0.1. As a result of the Beaufort SIC reduction, the correlation coefficients are interpreted negatively.

Full size image

Given that the FWI is the only metric that can be derived directly from reanalysis data and model outputs, and that all other fire proxies exhibit similar correlation patterns, we will focus on FWI as our primary indicator of fire activity in subsequent discussions. To deeper understand the connection between Southeast Asia’s FWI and Beaufort SIC, it is essential to probe into the meteorological variables in Southeast Asia that are associated with variations in SIC. We address this by analyzing the results from the Atmospheric Model Intercomparison Project (AMIP, see “Methods” section) and the Ocean Basin experiment (OBE, see “Methods” section) simulations, in addition to reanalysis data.

Since FWI is explicitly calculated by local meteorological variables such as surface temperature, total precipitation, and surface wind speed. In this work, we investigate the response of these parameters to remote forcing that comes from various ocean basins. We primarily use surface meteorological variables from the interpolated ERA5 reanalysis datasets28. The variables chosen are T2M (2-meter temperature), TP (total precipitation), SLP (sea level pressure), WND10 (10-meter wind speed calculated its meridional and zonal component), relative humidity at 850 hPa (RH850), volumetric soil moisture (VSM), and vapor pressure deficit (VPD).

The negative correlation coefficients between Beaufort SIC and Southeast Asian meteorological variables, corresponding to the SIC decline, are calculated using multiple datasets and experiments and are shown in Fig. 2 (also Fig. S6). According to analysis using ERA5 data, reduced Beaufort SIC appears to be correlated with higher 2 m temperature (T2M), lower total precipitation (TP), stronger surface wind (WND10), and higher sea level pressure (SLP) over Southeast Asia (Fig. 2a–c), particularly in the northeastern part. These meteorological changes all encourage the ignition and spread of fires, which thus lead to significant increases in FWI (Fig. 2d). It is noted that FWI in Fig. 1a is obtained by raw Copernicus Emergency Management Service FWI calculated using weather forecast from historical simulations provided by ERA5 reanalysis, while that in Fig. 2d is calculated by our random forest regression (RFR) model using ERA5 meteorological data as the input. Therefore, the values shown in these two figures are similarly opposite, but not absolutely. Analysis using different SIC (HadISST, COBE-SST2, and NOAA OI SST V2, Fig. S5a) and reanalysis datasets (ERA5, NCEP-NCAR, and MERRA-2, Fig. S6a–h) indicates similar correlation patterns and thus confirms these results. To further validate our conclusion, we also consider more meteorological variables related to fire conditions including relative humidity at 850 hPa (RH850), volumetric soil moisture (VSM), and vapor pressure deficit (VPD). Reduced Beaufort SIC also corresponds to decreased RH850 and VSM with increased VPD (Fig. S7). All these meteorological conditions are favorable to the ignition and spread of fires in Southeast Asia (Fig. S6 and S7). As indicated by a previous study, Beaufort SIC has experienced an unprecedented rapid decline in recent decades, with the most remarkable SIC shrinkage since 2005 (Fig. S5a)29. This declining trend has been attributed to enhanced sea ice transport to the north of Canada30.

Fig. 2: Correlation between meteorological variables in Southeast Asia and Beaufort SIC in reanalysis datasets (left column) and model simulations (right column).
figure 2

The distribution of the negative correlation coefficients between FMA meteorological variables and ASO Beaufort SIC in (a–d) the ERA5 reanalysis (1980–2019), and (e–h) is similar to (a–d), but for the OBE simulation. a, e T2M, b, f TP, c, g SLP (shaded) +U10 + V10 (arrows), and d, h FWI calculated by the Random Forest Regression (RFR) models. The correlation coefficient is taken as a negative sign to correspond to the Beaufort SIC reduction. Student’s t-test indicated that the region containing white dots passed the significance test with p ≤ 0.1. As a result of the Beaufort SIC reduction, the correlation coefficients are interpreted negatively.

Full size image

Reanalysis datasets involve a wide range of physical processes, so we further analyze the climate responses over Southeast Asia to Beaufort SIC reduction in the AMIP experiments. AMIP shows consistent results with reanalysis datasets, i.e., reduced Beaufort SIC is associated with higher temperatures, less precipitation, higher SLP, and northward wind anomalies in Southeast Asia (Fig. S6i–k), all of which are contributing factors to the significant increases in FWI that have been observed in this region (Fig. S6l).

The AMIP experiment is driven by all ocean variability. To better isolate the role of the Arctic Ocean and confirm the causality relationship, we are ongoingly analyzing how key meteorological variables respond to Beaufort SIC change in climate model simulations (i.e., the OBEs). The OBE results reveal that meteorological variables in Southeast Asia show similar response patterns to Beaufort SIC reduction to those in reanalysis and AMIP experiments. Specifically, decreased Beaufort SIC causes higher T2M, lower TP, and higher SLP over Southeast Asia (Fig. 2e–g). These meteorological changes lead to strong significant increases in FWI estimated by the Random Forest Regression (RFR) model (Fig. 2h). It has been observed that the responses of meteorological variables and FWI to the Beaufort SIC are not strictly consistent. This discrepancy can be attributed to the presence of biases in the reanalysis data and inaccuracies in the model parameterization. Moreover, reanalysis data involve many different forcing mechanisms and physical processes, whereas model results represent forcing from the Arctic Sea Ice only. It is thus possible that some differences are observed. Nonetheless, the overall spatial response pattern remains notably consistent, characterized by decreased Beaufort SIC corresponding to increased FWI in southeast Asia (Fig. 2d, h).

In sum, observations, reanalysis, AMIP and OBE results jointly reveal a robust relationship between Beaufort SIC decline and Southeast Asian fire increases in the recent decades. OBE results further confirm the causality of this relationship.

Possible physical processes underlying the Beaufort SIC – Southeast Asian fire teleconnection

In order to clarify the physical mechanisms by which Beaufort SIC variations affect the likelihood of fire in Southeast Asia, we examine the responses of geopotential height (GPH200) at 200 hPa, meridional wind (U200), and zonal wind (V200) over Southeast Asia to Beaufort SIC forcing in ERA5 reanalysis datasets and the OBE. We also investigate the corresponding zonal cross-sections of the variables averaged from 70°E to 110°E. Changes in stream functions in both reanalysis datasets and OBE are also investigated to reveal the possible circulation patterns. The T-N wave activity flux (WAF) is also calculated as an indicator of the wave energy propagating direction31.

The ice-albedo feedback mechanism causes a warm anomaly over the Beaufort Sea region associated with SIC decline, which extends to the higher troposphere and serves as the source of further wave propagation. The warm anomaly might also be associated with increased downward longwave radiation revealed by other studies32,33,34. This results in a localized low-pressure center forming in the upper troposphere (Fig. 3a, b). This atmospheric pressure anomaly acts as a source for the generation of the Rossby wave, which propagates from the west to the east in a distinct pattern characterized by alternating high and low-pressure centers (Fig. 3c, d). Notably, this wave train configuration culminates in a low-pressure anomaly over North Asia and a high-pressure anomaly over Southeast Asia (Fig. 3a, b). The changes in the stream function across the upper levels over South Asia in response to a decrease in SIC in the Beaufort Sea align with the development of a high-pressure system accompanied by descending airflows in that region. Furthermore, the T-N WAF also reveals the eastward and southward directions of Rossby wave propagation (Fig. 3c, d), offering solid evidence for the suggested mechanism. The existence of this high-pressure system hinders the upward movement of moisture in the atmosphere, reducing cloud formation and the chances of precipitation. This lack of rainfall, coupled with the heating effect, accelerates the evaporation of moisture in the air, exacerbating drought-like climate conditions.

Fig. 3: Physical processes involved in the teleconnection.
figure 3

a, b The distribution of the negative correlation coefficients between FMA 200 hPa GPH (shaded) and U + V (arrows) and ASO Beaufort SIC in the a ERA5 reanalysis (1980–2019) and b OBE. c, d Similar to (a) and (b), but for the 200 hPa stream function (SF, shaded) and T-N WAF (black arrows). e, f Similar to (a) and (b), but for the cross-section of the zonal mean (70°E-110°E) GPH (shaded) and V + W (arrows). g, h Similar to (a) and (b), but for the cross-section of the zonal mean (70°E-110°E) meridional circulation stream function (SF, shaded). Student’s t-test indicated that the region containing white dots passed the significance test with p ≤ 0.1. The red arrows represent the propagation trajectory of the Rossby waves. As a result of the Beaufort SIC reduction, the correlation coefficients are interpreted negatively.

Full size image

Moreover, a high-pressure system is typically associated with a high-pressure ridge, which is an elongated vertical area of high pressure in the atmosphere. The high-pressure ridge disrupts the flow of the atmosphere, preventing moist airflow from entering the region where the high-pressure system is located. This impediment to moisture supply typically results in prolonged droughts and heatwaves. It is also noteworthy that this pressure configuration opposes the typical mode of the Asian winter monsoon system, which is characterized by continental high pressure and oceanic low pressure. Consequently, the East Asian winter monsoon and South Asian winter monsoon are both weakened by these pressure anomalies. The weakening of the winter monsoon contributes to the faster transition to the summer monsoon and creates a favorable circulation background for the rapid increase in temperature in Southeast Asia. In addition, the weakened monsoon induces the divergence of water vapor from equatorial advection, leading to increased soil dryness, which has also been pointed out by Dong et al.17. The weakening of the monsoon also corresponds to the prevalence of southerly winds in the region, further leading to stronger surface wind speeds (Fig. 3a, b). This dryness, coupled with reduced moisture transport to Southeast Asia, further reduces precipitation in the region.

Furthermore, vertical cross-sections of reanalysis and OBE consistently reveal the enhancement of the northern hemispheric meridional circulation cell induced by the declined sea ice (Fig. 3e, f). The source of the meridional cell might be the warming of high-latitude regions caused by the reduction of sea ice in the Beaufort Sea, resulting in the ascending airflow. This anomalous meridional cell ascends in the Arctic region and descends in the low-latitude region (near Southeast Asia, clockwise), while the other cell ascends in equatorial areas and descends in low-latitude areas (near Southeast Asia, anticlockwise) (Fig. 3e, f). These anomalous circulation patterns are corroborated with the responses of stream function in both reanalysis datasets and OBE (Fig. 3g, h), which further contribute to the intensification of sinking airflow in Southeast Asia that leads to elevated surface temperatures and dryness, thereby enhancing the FWI. However, it is noted that the reduction of SIC and propagation of Rossby waves are coupled together, rather than a one-way causal relationship. Additionally, we have to acknowledge that there exist some disparities between the outputs of the ERA5 and OBE reanalysis systems within Fig. 3, primarily attributed to the coarser spatial resolution and the less comprehensive parameterization employed within the OBE simulation. Moreover, the reanalysis dataset contains many forcing signals, which are not all independent. In contrast, the data in the OBE only includes the signal of SIC without interferences from other forcing signals. Nonetheless, our primary emphasis lies in the discernment of the overarching patterns and the systemic responses of the atmospheric circulation as represented by these models.

A schematic diagram summarizing these processes is presented in Fig. 4. Briefly, a decrease in Beaufort SIC results in increased absorption of solar radiation by the sea surface, leading to the heating of the air above the sea surface. The development of a warm anomaly leads to the formation of a local low-pressure area, which in turn initiates the eastward movement of Rossby waves. This wave train generates a high-pressure anomaly over the North Pacific and South Asia with a low-pressure anomaly over East Asia, creating a pressure pattern that is contrary to the typical pressure pattern of East Asian and South Asian winter monsoons. The weakening of the winter monsoon enhances dry and hot weather in Southeast Asia during the FMA season by leading to the divergence of water vapor near Southeast Asia with reduced precipitation, increased temperature, stronger surface wind, and elevated soil aridity. Moreover, two meridional cells from the equatorial areas and polar regions respectively intensify the sinking airflow in Southeast Asia, contributing to the persistent dry and hot conditions. The above mechanisms combined lead to prolonged and intense fire seasons in Southeast Asia.

Fig. 4: A schematic diagram summarizing the teleconnection between Beaufort SIC and Southeast Asia fire.
figure 4

The propagation path of Rossby waves is represented by white solid lines and arrows. “A” represents an anticyclone with a high-pressure center, while “C” represents a cyclone with a low-pressure center. The meridional circulation is represented by thick solid dark arrows. EAWM East Asia Winter Monsson, SAWM South Asia Winter Monsoon.

Full size image

Enhanced population exposure to FWI induced by decreasing sea ice

The increase of FWI over Southeast Asia associated with Beaufort SIC decrease tends to deteriorate local air quality and detriment public health, especially considering that the study area is one of the most densely populated areas in the world. In order to further evaluate this health effect, we attempt to quantify the change in population exposure to fire weather by calculating the population-weighted FWI over Southeast Asia under two scenarios, namely decreasing Beaufort SIC by one (-std) and two standard deviations (-2std) respectively. The -std corresponds to the decadal SIC trend of shared socioeconomic pathways (SSP) 126, while -2std corresponds to the decadal SIC trend of SSP585 in CMIP635.

For FWI change with 1 std SIC decrease, as shown in Fig. 5a, when Beaufort SIC is one standard deviation below the climatology value, FWI will increase by 10~25% in Southeast Asia. Consistent with Fig. 2h, the most significant FWI change is found in the northeast of Mainland Southeastern Asia (i.e., Myanmar and eastern Thailand), where growth has exceeded 20%. Under this scenario, considering the population distribution in this region (Fig. 5b), the population exposure to FWI will also significantly increase, with the most remarkable changes found in northern India (over 20%), followed by southern India and central Myanmar (Fig. 5c). The spatial displacement of FWI change and that of FWI exposure is mainly due to the dense population in northern India, making this area especially vulnerable to fire-related air pollution.

Fig. 5: Population exposure to FWI with reduced Beaufort SIC.
figure 5

a The contribution of the decrease of the standard deviation in Beaufort SIC to FWI increasing percentage in Southeast Asia. b Population distribution in Southeast Asia. c Population-weighted FWI change percentage with 1 std SIC decrease. d Population exposure to FWI under different SIC reduction scenarios in Southeast Asia. The blue bar corresponds to the present scenario, the orange bar corresponds to a standard deviation reduction in Beaufort SIC, and the red bar corresponds to a double standard deviation reduction in Beaufort SIC. FWI values are divided into five categories: lower than 20 (low), 20–38 (moderate), 38–48 (high), 48–58 (very high), and higher than 58 (extreme)57.

Full size image

The change in population exposure is further evaluated under different FWI categories (Fig. 5d). Overall, the reduction of SIC, even in -std scenarios, will lead to a sharp increase in the number of people exposed to high and extreme FWI (with an increase of over 100%, Fig. 5d). Specifically, in the present scenario, most population are exposed to high FWI with few in extreme FWI (fewer than 200 thousand people). While in the moderate reduction scenario, most population (around 600 million people) is exposed to very high FWI, and more than 200 million people are exposed to extreme FWI (approximately 1000 times more than the present scenario). Furthermore, in the -2std scenario, nearly 1 billion people are exposed to very high FWI (five times more than the present scenario). We also analyze the scenario of the 0.5 std reduction of SIC and find that even in this moderate reduction scenario, more than two times people will be faced with very high FWI compared to the present scenario (Fig. S8).

The above results indicate that as SIC decreases, over 500 million people will be exposed to very high levels of FWI, and those exposed to extreme FWI are 1000 times more exposed than in the present scenario in this region. This further emphasizes the urgency and importance of adopting appropriate measures to delay and adapt to the reduction of SIC.

Conclusions and discussion

In this work, we reveal a teleconnection from the Arctic Ocean to the likelihood of fires (mainly represented by FWI) in Southeast Asia and evaluate its impact on public health. The decrease of SIC in the Arctic’s Beaufort Sea leads to increased springtime FWI over Southeast Asia, mainly through two pathways. First, the reduction in Beaufort SIC initiates the propagation of Rossby waves, which serves to weaken the East and South Asian winter monsoon by generating a pressure pattern opposite to the climatological monsoon pressure anomaly pattern. Second, this unique pressure pattern gives rise to the occurrence of heatwaves, convergence of water vapor, and the enhancement of surface wind in Southeast Asia. Moreover, two meridional cells originating from equatorial and polar regions intensify the sinking airflow in Southeast Asia, contributing to persistent dry and hot conditions. Consequently, these conditions exacerbate the already hot and dry environment, thereby creating a favorable setting for the ignition and propagation of fires. The interplay of reduced Beaufort SIC, Rossby waves, pressure patterns, and subsequent heatwaves and water vapor convergence, collectively contribute to the increased risk of prolonged and intense fire seasons. By quantifying the population under FWI exposure in decreased SIC, it is indicated that as SIC equals one standard deviation below the climatological mean (similar to the SIC decadal trend in the SSP126 scenario), over 500 million people will be exposed to very high levels of FWI, and those exposed to extreme FWI are 1000 times more than those in the present scenario.

While prior research has connected fire activity in Southeast Asia to a range of atmospheric and oceanic variability, including ENSO and IOD10,12, our findings highlight the unneglectable role of variability in the high latitudes. Given that the Arctic SIC is expected to continue shrinking from 19% (SSP126) to 62% (SSP585) by 2100 s and Beaufort seas surface temperature will also rise faster than the global average, increasing by +4 °C by 2040 compared with the 1981–2010 mean, the vulnerable Southeast Asian region may face more severe fire management challenges36,37. These worsened conditions might greatly increase the exposure level of the population to extreme fire weather considering its increasing population38,39. Firstly, global warming causes the vegetation to flourish, serving as the “fuel” for the wildfire. Moreover, this warming further increases FWI in Southeast Asia by teleconnection between reduced Beaufort SIC and increased fire potential. Additionally, such teleconnection might be enhanced in the future warming scenario24. Based on the increasing risk of wildfires and more population in this region, there may be more people in this area exposed to extreme fire weather conditions with higher risks, thus presenting great challenges to future attempts to mitigate climate change40.

Methods

Fire proxy data

The present study employs monthly fire activity proxies, including the Fire weather index (FWI) quantifying meteorological conditions favorable for fire, burned area (BA) indicating the total brunt spatial extent, fire count (FC) revealing the number of fire hotspots, fire radiative power (FRP) suggesting the intensity, and station smoke observation data as an indirect indicator of fire.

FWI is one of the most widely used numerical indexes measuring the weather conditions for fire ignition, spread, and sustainability41. The FWI data used in this study are obtained from the fire danger indices historical data calculated by the European Center for Medium Weather Forecasting (ECMWF) Reanalysis v5 (ERA5) reanalysis. After that, monthly values are created by averaging the daily FWI data for additional analysis.

The monthly BA product derived from satellite observations (BA satellite) from 2001 to 2019 is applied in this study41. When it comes to the pixel product, the BA is identified by the date that the burned signal was first detected, and when it comes to the grid product, it is identified by the total BA per grid cell. BA from global fire emissions database version 4 (GEFDv4) and BAF from global fire emissions database version 4.1 (GEFDv4.1) were also applied in this study for further analysis42. The Moderate Resolution Imaging Spectroradiometer (MODIS) fire products in Collection 6 are the source of the FC and FRP datasets used in this study43. These datasets were selected because, in comparison to more recent sensors, they span a longer time period. FRP stands for fire radiative intensity, a measure of biomass combustion and fire energy. It has improved detection capabilities in boreal regions. The monthly FC and FRP data in this study are resampled to a 0.25° × 0.25° grid for analysis after being interpolated from 2003 to 2019. The FC and FRP data processing techniques are similar to those applied in an earlier investigation44.

The hourly observational meteorological surface dataset is the source of the monthly station smoke observation data45. This dataset includes surface weather observations from all over the world ranging from 1981 to 2019, providing hourly records of station weather observations around the world. To determine fire days, we use the manually reported weather (MW) for every hour. We use smoke detection as a stand-in for fire because MW does not directly provide information on fire detection. A day is considered smoky when, at any time of day, MW equals 04 (smoke). Our approach to processing station smoke observation data is consistent with earlier research46, and daily observations are used to compute the annual station smoke frequency in the spring.

In light of the considerable influence of ENSO on the fire proxies in Southeast Asia, we also remove ENSO’s signal from all further analyses. This involves regressing the Niño 3.4 index, which is a key indicator of ENSO, onto the fire proxies in Southeast Asia. The remaining data, or residuals, after this regression will be the basis for our final calculations. This approach is also extended to other weather variables within the reanalysis data. By doing so, we aim to reduce the influence of ENSO on fire-related and meteorological variables, thereby allowing us to isolate the effects of Beaufort Sea ice on the meteorological conditions47.

Meteorological data

The meteorological variables chosen in this study span the years 1981 to 2019 and have a 0.25° × 0.25° spatial resolution, in line with Lawson and Armitage’s methodology48. To assess the sensitivity of our results to alternative reanalysis datasets, we also incorporate corresponding variables from the NCEP-NCAR Reanalysis 1 datasets (2.5° × 2.5°)49 and the MERRA-2 datasets (0.625° × 0.625°)50. The analysis of stream function and T-N wave activity flux (WAF) is also used for further investigation31.

Furthermore, to minimize the impact on the results by the selection of one specific dataset, we include monthly SIC datasets from multiple sources, including the National Oceanic and Atmospheric Administration (NOAA) Optimum Interpolation Sea Ice Concentration (1° × 1°) obtained from the Hadley Center (HadISST), as well as Sea Surface Temperature and Ice datasets from Centennial in situ Observation-Based Estimated (COBE-SST2, 1° × 1°) and NOAA Optimum Interpolation (OI) SST version2 (V2, 1° × 1°) datasets for the same period51,52. The correlation coefficients between meteorological variables and SIC are regarded as the response of meteorological variables to SIC variability. ENSO and IOD signals may also impact the Arctic sea ice, but neither the correlation between Beaufort SIC and boreal winter Niño 3.4 in the preceding year (ENSO index) nor that between Beaufort SIC and boreal autumn dipole mode index (IOD index) is statistically significant. Therefore, we do not consider the impacts of ENSO and IOD on Beaufort SIC in this study.

Ocean basin experiments

We conducted a range of Arctic forcing experiments with the Community Earth System Model-Community Atmosphere Model version 5 (CESM-CAM5)53. These experiments were designed to investigate how the Arctic SIC affects Southeast Asia, leveraging the improved capabilities of CESM-CAM5 in accurately modeling climate variability and high-latitude dynamical processes. Given the exploration of large-scale teleconnections between Southeast Asia FWI and Beaufort SIC in this study, employing global model simulations is crucial for comprehensive analysis. The trend in Arctic SIC from 1979 to 2019, derived from actual time series data, was included in the model. Meanwhile, the SIC in all other ocean basins was kept as seasonally changing climatological data. This setup allowed for an examination of how meteorological variables in Southeast Asia responded to the distant influence of the Arctic SIC changes. The atmospheric reaction to the SIC forcing in the specific ocean area was ascertained by determining the ensemble mean based on an ensemble of eight simulations. Note that each ensemble member has a different starting condition. The first simulation for eight model years was powered by climatological forcing. The start files for each year were used as the starting point for each of the eight ensemble members. These experiments are similar to simulations in earlier research46.

AMIP experiments

The Atmospheric Model Intercomparison Project (AMIP) experiments provide monthly meteorological variables simulated using multiple atmospheric models that support the reanalysis data. The atmospheric models used in these simulations are mandated by SST and SIC observations54. Detailed information about the 13 models used here is given in Table S1. We use data spanning from 1980 to 2014 to investigate how Southeast Asian meteorological variables have responded to Arctic SIC since the 1980s. For further analyses, every dataset mentioned above has been re-gridded as 1° × 1°in the AMIP.

The random forest regression model

While daily meteorological variables like T2M, TP, and WND10 can be explicitly used to calculate the FWI40, the scarcity of outputs from OBE and AMIP results makes it challenging to calculate the FWI using the standard formulas. Thus, following the previous work, we build a statistical model utilizing the Random Forest Regression (RFR) method to simulate FWI in these experiments with four predictors, namely T2M, SLP, TP, and WND1024. The assessment of relative importance for all input variables is conducted through the application of Gini importance, a metric characterized by the aggregate reduction in node impurity, averaged across the entirety of the ensemble tree collection. Finally, four predictors, namely T2M, SLP, TP, and WND10 were selected, and the output is FWI as a direct indicator of fire weather potential.

A 200-tree RFR model is trained 100 times in order to minimize the error in the prediction result associated with varying initial values and weights. The prediction performance was examined and the model parameters were tuned using a five-fold cross-validation method. The average correlation coefficient between the observation and prediction is 0.88, which passes a significance level of p ≤ 0.01 according to a Student’s t-test.

Population exposure to FWI

To evaluate the socioeconomic effects of FWI changes, we used Landscan’s 2020 global population density data55. The original spatial resolution of this dataset is 0.05 ° × 0.05 °. In order to match with the OBE results, we interpolated it to 2.5 ° × 1.875 °. Following previous studies concerning population exposure to fire air pollution, the population-weighted FWI is calculated according to the following formula56.

$${{FWI}}_{{weighted},i,j}=\frac{{{FWI}}_{i,j}\times {{Population}}_{i,j}}{\sum _{i,j}{{Population}}_{i,j}}$$

where \({{FWI}}_{{weighted},{i},j}\) is the population-weighted FWI in the ith and jth grid, indicating the population impact of an increase in FWI, \({{FWI}}_{i,j}\) is the fire weather index in the ith and jth grid, and the \({{Population}}_{i,j}\) is the population is the total population in the ith and jth grid. The fire weather is further classified into five grades according to its FWI value, namely low (<20), moderate (20–38), high (38–48), 48–58 (very high), and extreme (>58), using a similar criterion as the previous work57.

Furthermore, we also consider SIC reduction varies in various future warming scenarios with various shared socioeconomic pathways (SSPs)58,59. For instance, SSP126 represents a scenario that integrates the socioeconomic aspects based on SSP1 and the energy-emissions-land use aspects based on the Representative Concentration Pathway (RCP) 2.6. Specifically, the declining trend of Beaufort Sea SIC for SSP126 is −2%~−10%/decade, while the estimated downward trend for Beaufort Sea SIC in the SSP585 scenario is −6%~−14%/decade, and the standard deviation for SIC is 4.4%–7.1%35. Therefore, we analyze FWI exposure under the above three scenarios: climatology (present), decreasing one standard deviation (-std), and decreasing two standard deviations (-2std). The present scenario refers to the SIC remains in the climatology. The -std scenario corresponds to SIC equals to the climatology minus one standard deviation (approximately equal to the decadal trend in SSP126), while the -2std corresponds to SIC equals to the climatology minus two standard deviations (approximately equal to the decadal trend in SSP585).