Projected increases in western US forest fire despite growing fuel constraints

Data sets

Monthly climate data of maximum and minimum temperature, dewpoint temperature, and precipitation at a 1/24th degree horizontal resolution from 1950 to 2020 was acquired from the Parameterized Regression on Independent Slopes Model (PRISM)⁴⁴. Monthly surface downward shortwave radiation and 10-m wind speeds at a 0.25-degree horizontal resolution were acquired from ERA-5⁴⁵ for the same period and bilinearly interpolated to the PRISM grid. Monthly data for the same variables from a single ensemble member from each of 30 climate models participating in the Sixth Coupled Model Intercomparison Project (CMIP6) were acquired from the historical climate experiment for 1950–2014 and from the SSP2-45 experiment for 2015–2050 and interpolated to a common 1.0-degree horizontal resolution grid (Supplementary Table 4).

Following Abatzoglou and Williams, we calculated three proxies of aridity using monthly climate data: mean vapor pressure deficit (VPD), Penman-Monteith reference evapotranspiration (ETo), and climatic water deficit (CWD⁴⁶, defined as ETo minus actual evapotranspiration³). We modified ETo to account for potential reduced stomatal conductance due to increasing atmospheric carbon dioxide, which reduces surface resistance to evapotranspiration. We made this modification following the method of Yang et al.⁴⁷. Importantly, the effect of CO₂ on surface resistance at the scale of the western US is highly uncertain and this method derives the strength of this effect from earth system models. Each index was calculated as follows. At each grid cell, we calculated mean Mar–Sep VPD, the sum of Mar–Sep ETo, and Jan-Dec CWD; each of these time series was standardized to the 1991–2020 baseline using z-score transformations to create a fuel aridity index f for each grid cell. The regionally averaged fuel aridity index F was calculated by first taking the average of f over grid cells that have a majority of land classified as forest or woodland in the LANDFIRE environmental site potential product⁴⁸. We then re-standardized F relative to the 1991–2020 reference period and applied equidistant quantile mapping⁴⁹ to each model. The latter ensures that the distributions of modeled Z match those of observed Z for the 1991–2020 period while preserving changes in Z from this reference period. Herein we used CWD for F because it presents a more balanced view of precipitation and atmospheric demand than VPD or ETo alone, exhibits strong links to the forest-fire area over the observational record, and has more conservative increases in fire under future climate (Supplementary Fig. 2). The variance explained in forest-fire area when defining F as VPD, ETo, and detrended CWD is presented in Supplementary Table 1. We note that our approach does not explicitly incorporate daily meteorology such as the number of dry days or critical fire-weather patterns¹⁰ beyond that already included in F.

Burned area data from wildland fires were acquired from Monitoring Trends in Burn Severity (MTBS) during 1984–2018⁵⁰ and from the version 6 MODIS burned area dataset during 2001–2020⁵¹. The forested burned area was aggregated by lands classified as forest or woodland⁴⁸. MTBS includes primarily fires ≥404 ha that comprises>95% of burned area in the region⁵². We further excluded areas in the unburned-to-low burn severity class⁵³ as well as fires classified as prescribed burns in MTBS. Further, we did not include forested area treated by prescribed fire as a contemporary area for prescribed fire is more than an order of magnitude less than that of forest-fire area⁴¹. Forest-fire area estimates for 2019–2020 were obtained using adjusted burned areas from MODIS based on a linear model that relates MODIS and to the MTBS forest-fire area time series during the overlapping 2001–2018 period²⁶.

Experimental design

We focus on macroscale climate–fire models operating at the scale of the entire western US forested area. While there is value in spatially refined models, efforts to parameterize empirical relationships at localized scales can be limited by the stochastic nature of ignitions and fire weather—particularly in locations with long fire return intervals with zero-inflated distributions of annual burned area. Strong interannual relationships between fuel aridity and strain on national fire suppression resources shared across the region highlight the implicit value in considering larger spatial scales⁵⁴. The macroscale approach is further justified because the leading mode of variability in fuel aridity across forested land is a commonly signed regionwide pattern that is strongly correlated (r² = 0.79) to the logarithm of forest-fire area (Supplementary Fig. 3).

Static model

Following previous empirical models of annual forest-fire area^3,25, we first consider a static model of western US annual forest-fire area (FFA) based on F (fuel aridity) of the form:

where t is the year, α_s and β_s, are regression coefficients, and ε represents an error term. We use annual CWD for F as it accounts for precipitation and atmospheric demand, exhibits strong interannual relationships with FFA, and provide more conservative estimates of projected changes in aridity and thus area burned than other aridity metrics such as VPD^3,7,12. The error term ε is drawn from the population of the log-residual of observed minus modeled FFA. This error term represents variability not captured in the FFA–F relationship (e.g., extreme fire-weather conditions, human ignitions) that is important for the full distribution of FFA.

Dynamic models

The contemporary climate–fire relationship in Eq. 1 should persist with increased F until increased burned area and severity cause fuel limitations¹⁵. Fire-fuel feedbacks that alter the climate–fire relationship primarily occur through temporary reduction of fine fuels; such feedbacks can reduce the burning potential for approximately three decades post-fire^38,55. Further, longer-lived reductions in the forest-fire area can occur when forests do not recover from fire and instead transition to non-forest vegetation that can still carry fire. However, constraints on the area burned imposed by fire-fuel feedbacks are weakened by concurrent drought, which allows the fire to propagate across sparser fuels, and can markedly shorten the window of reduced burning¹⁸.

We incorporate these effects through a term L, which represents the fraction of contemporary forested land that is incapable of carrying fire in a predominately forested environment in a given year, in a dynamic model of the form:

where the response of log(FFA) to fuel aridity reduces as a function of L. We present various potential forms and strengths of fire-fuel feedbacks in L that are guided by the ecological literature and account for post-fire tree regeneration failure, fuel limitations imposed by recent fire history, and waning of fuel limitations during drought^18,22,23,24. L is influenced by semi-permanent limitations due to failure of post-fire forest regeneration (L_rf), and temporary limitations due to recent fire history (L_f):

Importantly, L is poorly constrained and likely varies in geographically and temporally complex ways^18,34. For example, L can differ for a fixed fraction of recently burned forest. A relatively small L implies weak feedbacks allowing forests to more easily reburn. A relatively large L implies strong feedbacks, for example, where heterogeneous fire effects create patch mosaics that constrain fire spread even though there is ample fuel. Finally, the age threshold for L may decrease with continued climate change, with some indications that recent fires burned through forests <10 years post-fire.

Post-fire tree regeneration failure L
_rf

Summer soil moisture deficits in the years post-fire may reduce tree regeneration^19,38. Absent reforestation, the failure of post-fire forest regeneration represents a semi-permanent reduction in forested land available to burn. Post-fire regeneration is strongly dependent on site-specific thresholds in growing-season moisture requirements^23,56, burn severity and distance to the seed source, as well as species composition. We estimate the fraction (ρ) of forest that is permanently lost due to the failure of post-fire regeneration using the mean fuel aridity over the 3-year post-fire period (F_3y) because protracted drought stress in the years immediately following fire has been shown to limit the establishment of some tree species⁵⁷. This is done using a simple linear transform:

where μ is set at 0.1 (Eq. 4 is plotted in Supplementary Fig. 4a). Hence, the fraction of forested land that is semi-permanently ineligible to carry forest fire because previously burned forest did not regenerate as forest (L_rf) is the cumulative sum of the product of annual FFA and ρ since 1984:

where T refers to the contemporary area of forested land⁴⁸. Note that Eq. 4 and μ can be modified to account for the diversity of species-specific responses at local-to-regional scales given the acknowledgement that some species are more resilient than others and local plant water stress alters regeneration probabilities^58,59. Overall, L_rf as parameterized here resulted in values approaching L_rf ~0.01 by 2050, suggesting that the inability of trees to regenerate post-fire is a minor contributor to fire-fuel feedbacks through mid-century. Modifications to the parameters in Eq. 4 resulted in only minor differences in projected FFA (Supplementary Table 3).

Temporary fire-fuel feedbacks L
_f

Most studies in forested environments show strong fire-fuel feedbacks in the first 5–10 years post-fire^55,60. This temporary fire-fuel feedback, which we refer to here as L_f, tends to wane after 10 years⁶⁰, with the longevity τ of the fire-fuel feedbacks varying geographically, from as short as ~15 years in warmer sites in the southwest to over ~30 years in cold mesic systems in the northern Rockies¹⁸. Herein, we use a baseline τ = 30 years, which results in a conservative estimate of future area burned.

We consider two forms for how L_f incorporates information on annual fire histories over the previous τ years: a constant feedback and a fading feedback. These forms of L_f are defined below in Eqs. 6 and 7 and plotted in Supplementary Fig. 4c.

In the case of the constant feedback, the effect of burned area on L_f remains constant over the τ years following fire. At the scale of the whole western US forested area, the constant form, therefore, assumes that the transient limitation is simply proportional to the total FFA over the preceding τ years:

In Eq. 6, parameter γ represents the strength of the feedback, described in more depth below.

The fading feedback form of L_f more heavily weights the contribution from recent FFA compared to older FFA. At the scale of the whole western US forested area, this form applies constant weight to FFA in the five most recent years given strong fire-fuel feedbacks of recent fires, and increasingly reduces the contributions from prior years based on a sinusoid function:

Given the uncertainty in the efficacy of the fire-fuel feedback, we present results using both the constant and fading formulations for the temporary fire-fuel feedbacks.

We additionally considered three different fuel-limitation strengths γ in Eqs. 6 and 7 to account for direct and indirect potential effects of past fires: γ = 0.5, referred to as weak; γ = 1, referred to as moderate; and γ = 1.5, referred to as strong. For the weak (γ = 0.5) fuel-limitation case using the constant feedback model, the fractional forested area ineligible to burn is only half of the total area burned in the past 30 years, indicating that half of recent burned areas can reburn. For the strong-constant fuel-limitation case, the forested area ineligible to burn post-fire exceeds the total recent burned area by 50%. An example of a strong fuel limitation is a burn mosaic with reduced connectivity that constrains the ability of subsequent fire spread into the adjacent forest that did not burn in the previous τ years. We considered higher values of γ, but these yielded degraded cross-validation skills when modeling the historical period (Supplementary Table 2).

Longevity of fire-fuel feedbacks during drought

Finally, some temporary fuel limitations can be overcome during extreme fire-weather conditions and during periods of drought. For example, while reduced fuel loads in a post-fire landscape serve as an effective barrier for fire propagation under moderate fuel aridity, the fire spread probability increases with increasing F³⁴. Studies have found that the longevity of fire-fuel feedbacks was a third shorter during periods of extreme drought than in periods without drought stress^18,34. For example, there is evidence of short-interval (<20 years) stand-replacing fires in systems with 100–300 year mean fire return intervals—suggesting that such systems can carry fire under warmer and drier fire-weather conditions^20,33,61. We incorporate this effect by making the longevity parameter τ a function of contemporaneous F(t), where τ reduces linearly from 30 years toward 20 years as fuel aridity increases from 1991 to 2020 mean (F = 0) to two standard deviations above the mean (F ≥ 2):

The form of τ is plotted in Supplementary Fig. 4b. The resultant weighting of L_f in the fading model using the end-members of τ is displayed in Supplementary Fig. 4d.

Running the model

Dynamic models were run over the twentieth century (1916–1983) using proximal estimates of FFA²⁵ calibrated to the observational record⁶². Nominally, LANDFIRE’s gridded mean fire return interval (MFRI), an estimate of the average number of years between fires, was used to approximate a baseline FFA prior to 1916. This was used to initialize the model and serve as a reference point for projected FFA. We calculated three estimates based on the reported range for each categorical MFRI across the landscape (e.g., 11–15 years): a midpoint MFRI (e.g., 13 years), as well as the lower (e.g., 11 years) and upper bounds (e.g., 15 years) of MFRI estimates. We recognize that MFRI estimates are prone to uncertainty⁶³, and that such estimates reflect conditions prior to European colonization which likely differ from those in the late 1800s and early 1900s^28,64. However, the LANDFIRE MFRI represents the only wall-to-wall dataset of estimated fire regime parameters relevant to our study here and provide a range of likely estimates.

FFA estimates prior to the satellite era are prone to uncertainties^25,65. However, initial conditions do not influence model sensitivity over the 1984–2020 period. Rather, FFA estimates serve to highlight that our formulation of L captures the increase in fuel extent in the second half of the twentieth century due to the fire deficit associated with forest management practices^5,28. Finally, no significant difference in climate–fire correlations over the observational record were seen using dynamic models compared to the static models, suggesting that the influence of changes in L in recent decades has had little impact on recent fire-climate relationships—either because L has not changed significantly or that L ≪ 1 over the observational period (1984–2020). Our dynamic models suggest that suppressed L during the observational period has heightened FFA sensitivity to F.

Static and dynamic models were applied to the climate output of each CMIP6 model for the period 1950–2020 using L calculated from observations prior to 1950 and model-derived thereafter. For the 2021–2050 period, we used the observed FFA record through 2020 in calculating L and used model-derived FFA thereafter. Both the static and dynamic models implicitly include human and management effects in the regression parameters, and we assume these factors are time-invariant for future projections. Further, for each climate model run, we used a Monte-Carlo resampling procedure (n = 1000) that randomizes ε from the observed residual population therein creating 1000 replicates of modeled FFA for each climate model. Herein, we report statistics for the median of the Monte-Carlo simulations for each climate model.

In both static and dynamic models, we limit the area that can burn in a single year. The maximum FFA in the static model is defined by the estimated 100-yr return period FFA for the observational period using extreme value analysis. This works out to be 7.5% of the total forested land T—nearly four times that burned in 2020. For the dynamic models, the maximum forest that is allowed to burn in any year is 7.5% of the forested area that is eligible to burn: i.e., 7.5% of (1 − L)*T. We acknowledge that this limit is not well quantified physically but is implemented to constrain the exponential FFA–F relationships used herein. In practice, such limits occur in ~1% of model years for 2021–2050. Similar quantitative results were seen by capping F ≤ 3, which implies a saturation effect of F on FFA. Finally, while we ran the model at the aggregated scale of western US forests, the model may conceptually be run at subregional scales. We note that the strength of fire-feedbacks, climate–fire relationships, as well as model parameterizations will likely vary on subregional scales.

Statistical information

Model cross-validation was performed to assess the skill of the models. Models using Eqs. 1 and 2 were built using training data during 1984–2000 and validated for 2001–2020. We evaluated skill by three metrics: model bias as the ratio of total modeled FFA to total observed FFA; the coefficient of efficiency (CE) as a measure of accuracy relative to a null model; and the correlation coefficient between modeled and observed log(FFA). Note that measures of bias and CE were calculated using non-log transformed data.

We assessed the statistical significance of cross-validation skill by resampling the residual term ε in Eqs. 1 and 2. This involved developing and testing 1000 iterations of the model subject to a random sampling of ε. We deem a model to have significant skill when >95% of the iterations had bias CE > 0, >95% of the iterations had r > 0, and the inner 95% of the simulations included a bias of 0.

Supplementary Table 2 shows that the static model and many of the dynamic models have significant cross-validated skills. However, skill decreased in the dynamic models as the feedback strength increases. While the weak dynamic feedback models had similar cross-validation skill as the static model, dynamic models with very strong feedbacks (γ ≥ 2) had sizeable underpredictions in FFA by up to 46% for the validation period. Hence, we excluded such parameters from the further analysis given that such results were incongruent with the observational record.

Three statistical metrics of annual variability of FFA were calculated for both static and dynamic models. First, we used generalized extreme value theory to estimate recurrence intervals for FFA greater than equal to that of the 2020 fire season. Second, we calculated the interquartile range (IQR) in modeled FFA to examine changing interannual variability. Lastly, we examined the percent of years with modeled FFA below the 1991–2020 observed median as a measure of quiescent fire years. Calculations were performed separately for each climate model for 1991–2020 and 2021–2050.

Source: Ecology - nature.com