Forest fire threatens global carbon sinks and population centres under rising atmospheric water demand

Recent years have seen a succession of mega-fires in forests and woodlands globally, occurring in regions and seasons not traditionally considered fire-prone. Forests possess abundant live and dead biomass that becomes highly flammable as fuels dry, enabling fires to traverse natural barriers and spread into fire-sensitive communities, including rainforests, during major drying events. Quantifying how forest fire activity relates to fuel moisture variation clarifies environmental drivers of fire risk, particularly when paired with spatially explicit fuel moisture predictions. Vapour pressure deficit (VPD), derived from air temperature and humidity, directly measures atmospheric water demand, predicts dead fuel moisture, and influences live fuel moisture and mortality. Prior work has linked VPD to fire variability and identified VPD-based fuel moisture thresholds associated with increased burned area in several regions. The research question is whether global forest ecosystems exhibit VPD thresholds marking a switch from humid, non-flammable conditions to dry, flammable states, and how future climate change will alter the frequency of days exceeding such thresholds, with implications for carbon stocks and human health from wildfire smoke.

Multiple studies have demonstrated VPD as a reliable predictor of dead fuel moisture and a key driver of plant stress and fuel flammability. VPD-based thresholds have indicated critical increases in burned area in southeast Australia and Mediterranean Europe, and VPD has been strongly associated with fire variability in temperate and tropical forests. There is accumulating evidence of increasing atmospheric dryness (e.g., Mediterranean basin, western US, tropical South America) and rising forest carbon emissions linked to fuel moisture changes. Recent analyses using hourly data have similarly identified VPD thresholds tied to fire activity in the Americas. However, many prior studies focused on seasonal aggregates (e.g., total burned area), potentially conflating climate signals with suppression or regional differences in burn extent; fewer have modeled daily ignition probability as a function of climate at global biome scales.

Study area: The analysis covered all global forest biomes selected from a global terrestrial ecosystem classification, grouped into subtropical and tropical, mediterranean, and temperate and boreal categories, and masked using a 1 km global forest cover product.

Fire data: Fire extent was represented with the MODIS active fire/burned area products from February 2000–2020, using only highest QA observations at ~500 m daily resolution. To capture regional variation, 12 predefined sub-continental windows (excluding small island windows) were used. Prescribed/cultural burns likely contribute minimally due to smaller size/intensity and proximity to detection biases.

Climate data: Daily maximum VPD was computed from ERA5 reanalysis (0.25°, hourly) using daily maximum air temperature and dew point temperature (at the time of Tmax) for the same period as fire data. For future projections, three CMIP5 GCMs were selected for skill, independence, and scenario span: ACCESS1.0, CNRM-CM5, and GFDL-CM3, under RCP4.5 and RCP8.5. Daily maximum VPD was derived from 3-hourly GCM fields for 2026–2045 and 2081–2100; early-period data (1951–2000) were used for quantile-mapping bias correction. Climate deltas were calculated by subtracting modeled present (1981–2000) from modeled future values. Analyses used native GCM grid resolution (results apply to forest within each grid cell).

Modeling and threshold detection: For each biome-by-window combination (n = 70), generalized linear models (binomial error, logit link) estimated the probability of fire occurrence (grid cell recorded as burned) as a function of daily maximum VPD. Presence data were VPD values matched to burned pixels and dates; absences were randomly sampled unburned grid cells/dates within the study area, with equal numbers of presences and absences per year. Grid cells that burned in the previous five years were excluded; extreme wildfire years were removed in a subset of cases. Model performance was assessed with ROC metrics. The VPD threshold (VPD50) was defined as the daily VPD at which modeled fire probability equals 50%.

Frequency and exposure metrics: For each region, scenario, and epoch, the annual frequency of days exceeding VPD50 was computed using ERA5 (2003–2020) for current climate and bias-corrected GCMs for futures. Relationships between exceedance frequency and burned area were examined in supplementary analyses. To estimate exposure, gridded population projections (1 km) were multiplied by the change in days per year above VPD50 to obtain person-days of exposure to critical fire activity conditions. Forest carbon exposure was estimated by multiplying ESA Biomass CCI aboveground biomass (2010; 100 m, resampled to analysis grid) by the change in days per year above VPD50 to yield tonne-days of exposure.

• Fire activity across all global forest biomes responds strongly and predictably to daily maximum VPD, with clear separation of VPD distributions on fire versus non-fire days. Models achieved a median true positive rate of 0.73 across 70 biome-window combinations, indicating 73% probability of correctly predicting fire on fire days.
• VPD thresholds (VPD50) vary systematically by biome, broadly following latitude: highest in subtropical and tropical forests (median ~2.7 kPa), intermediate in mediterranean biomes (median ~2.3 kPa), and lowest in temperate and boreal forests (median ~1.3 kPa).
• The mean annual frequency of days exceeding VPD50 (potential fire days) during 2003–2020 was highest in East Asia, southeast Australia, western Europe, and the eastern United States. Some temperate, boreal, subtropical, and tropical regions averaged fewer than 30 exceedance days per year, whereas mediterranean forests commonly exceeded VPD50 on fewer than 66 days per year but against an already fire-prone background.
• Under RCP8.5, by 2026–2045 all models project at least 45 additional days per year above VPD50 in parts of tropical South America; two of three models project similarly large increases in North America, East Africa, and large parts of Europe. By 2081–2100, all continents’ forest biomes are projected to see ≥45 additional exceedance days per year, with tropical South America projected to see increases ≥150 days per year across models.
• Under RCP4.5, widespread increases in exceedance frequency also occur, though generally of smaller magnitude. Increases are largest and most widespread in ACCESS1.0 and GFDL-CM3 and more moderate in CNRM-CM5.
• Regions globally significant for carbon storage, notably the Amazon, and densely populated regions in parts of South Asia and East Africa face substantial increases in exposure (tonne-days and person-days, respectively) to conditions conducive to forest fire.
• Findings at daily resolution reinforce fuel moisture’s central role in governing ignition probability and forest fire activity globally.

The study demonstrates that daily maximum VPD provides a robust, biome-specific threshold predictor for the switch from non-flammable to flammable conditions in forests. Threshold values scale with climatic gradients, reflecting higher atmospheric water demand in lower latitudes. By focusing on daily ignition probability rather than seasonal aggregates, the models more directly capture the fuel moisture constraint on fire initiation. Projections indicate pervasive increases in the frequency of days exceeding VPD50 under climate change, implying heightened wildfire potential nearly everywhere, with especially large increases in tropical forests (Amazon), then northern temperate and boreal forests, and continued high risk in mediterranean regions. These changes carry major implications for the global carbon cycle—potentially pushing the Amazon toward a tipping element—and for public health via smoke exposure in heavily populated regions. While human factors (ignition sources, suppression capacity), fire weather components (e.g., wind), and long-term drying trends also influence realized fire activity and may dampen or amplify the VPD–fire link locally, the consistent global relationship underscores VPD’s utility for risk assessment and planning. Model performance could be enhanced with additional predictors (e.g., evaporation, soil moisture, wind) and improved data, but the primary conclusion—climate-driven increases in days with high ignition probability—is robust across models and scenarios.

This work identifies and maps VPD-based ignition probability thresholds across all global forest biomes using daily fire observations and hourly reanalysis, and applies them with skill-selected GCMs to project changes in the frequency of high-risk days. The study finds systematic biome differences in thresholds and widespread future increases in exceedance frequency, with particularly large increases in tropical South America, threatening carbon stocks and elevating human exposure to smoke in populous regions. These results provide a practical, physically grounded basis for forecasting forest fire risk under climate change and for prioritizing mitigation and adaptation measures. Future research should investigate seasonal and interannual variability of atmospheric water demand, integrate additional biophysical and human predictors, refine spatial resolution and bias corrections, quantify links to burned area and severity, and more fully assess health and economic impacts of smoke and fire under different mitigation pathways.

• Human influences (ignition patterns, detection, suppression capacity) and other fire weather components (e.g., wind) were not explicitly modeled and can weaken or modulate the VPD–fire relationship regionally.
• Seasonal and interannual variability in atmospheric water demand and trends were not explored in depth.
• The presence–absence sampling and exclusion rules, as well as removal of extreme years in some cases, may influence threshold estimation.
• Spatial resolution mismatches (500 m fire vs. 0.25° climate; coarse native GCM grids) and reporting at coarse grid scales limit local applicability and may obscure fine-scale heterogeneity; results apply to forests within each grid cell.
• Bias correction and the choice of three CMIP5 models introduce uncertainty; model-dependent differences in projected temperature–humidity changes affect VPD projections.
• VPD exceedance frequency does not directly translate to magnitude of burned area or severity; realized fire depends on ignition, fuels, and suppression.
• ERA5 and satellite fire products carry known uncertainties (e.g., humidity representation, burned area detection biases).