US wildfire potential: a historical view and future projection using high-resolution climate data

The KBDI is calculated using daily maximum temperature, daily precipitation, and annual accumulated precipitation (Keetch and Byram 1968). As a substitute for fuel availability in an area, the KBDI uses cumulative annual precipitation and assumes that higher annual rainfall corresponds to more vegetation, and therefore more fuel available to burn. The KBDI formula approximates the amount of evapotranspiration as a proxy to account for the dryness of the upper soil layers and the flammability of organic matter. The calculated KBDIs fall within a range from 0 to 800, which is equivalent to the amount of water necessary (in hundredths of an inch) to bring the soil up to an assumed complete saturation of 8 inches. For example, a KBDI value of 0 indicates that the soil is completely saturated, and the potential of a wildfire is low; whereas a KBDI value of 800 indicates extreme drought and unpredictable wildfire behavior. In practice, the KBDI is commonly calculated on a daily basis, and the index either increases or decreases by

$$\begin{align}& \frac{{\left( {800 - Q} \right)\left( {0.968{e^{0.0875{T_{max}} + 1.552}} - 8.30} \right) \times {{10}^3}}}{{1 + 10.88{e^{ - 0.0174R}}}}\nonumber\\ &\qquad = dQ/dt\end{align} \tag{ 1 }$$

where T_max is the daily maximum temperature (°C); R is the annual accumulated precipitation (cm); dt is the time increment, 1 d here; and Q is the KBDI for the previous day. When T_max falls below 6.78 °C, dQ is 0. In addition, the value of Q is reduced by the amount of daily precipitation in excess of 0.51 cm (0.20 inch). This equation allows for input in SI units but gives the KBDI in units of hundredths of an inch. See Keetch and Byram (1968) for more details.

Ideally, according to Keetch and Byram (1968), KBDI could be set to zero after a period of 15.24 − 20.32 cm (6–8 inches) of precipitation or more accumulated in 1 week. However, this requirement may not be met often in dry regions such as the southwestern CONUS. We developed an initialization procedure as follows. First, we find the week of maximum precipitation for the first year of data and set the KBDI to zero for that week. Using it as an initial point, KBDI is calculated on the basis of the KBDI value from the previous day. During the initialization (or spin-up) stage of KBDI, the 1st year data is used in a loop to repeat the daily KBDI calculations. When the differences between the daily KBDIs calculated in loop (n) and loop (n-1) are smaller than a threshold, i.e. 0.3 inch, it is determined that the calculated daily KBDIs are stabilized as being the closest to the 'real' KBDIs of the first year. At this point, the initialization of KBDI is completed. The KBDI on the last day of the first year, which is calculated in the loop (n), is then used to estimate the KBDI on the first day of next year, and so forth so on for the subsequent time periods. For relatively wet regions, the KBDI does not need a long spin-up time (1–2 loops of the 1st year data are sufficient); however, in dry regions, the KBDI needs longer time (at least 4 loops) to reach equilibrium (see figure S1 (available online at stacks.iop.org/ERL/16/034060/mmedia) in supporting information). This method accounts for the specific conditions of the first year and can therefore achieve a more effective initialization than previous studies (Liu et al 2013), while the error due to inaccurate representation of the initialization year eventually settles out (Fujioka 1991).

Daily precipitation and maximum temperature were obtained from the Parameter elevation Regression on Independent Slopes Model (PRISM; Daly et al 1994, 1997, 2008) at a spatial resolution of 4 km. PRISM utilizes station measurements and a weighted regression scheme to determine precipitation and temperature values in complex terrain by considering factors such as location, elevation and orographic effectiveness. Data after 2002 also incorporate radar measurements to supplement station data (Daly et al 2008). The KBDI was calculated using PRISM data from 1981 through 2017, with 1981 used for initialization.

Future projections of KBDI were calculated from the model simulations of a regional climate model—Weather Research and Forecast model (WRF, Skamarock et al 2008)—at a spatial resolution of 12 km (Wang and Kotamarthi 2015). Most of the results presented in this study are derived from the WRF simulations driven by Community Climate System Model version 4 (CCSM4, Gent et al 2011) with bias-corrected vertical profiles of temperature, wind, and geopotential height. The KBDI was calculated for one historical (1995–2004), and two future periods (2045–2054 and 2085–2094) under representative concentration pathway (RCP) 8.5, a pathway that assumes very high levels of greenhouse gas emissions by 2100 with an effective radiative forcing increase of 8.5 W m⁻² due to large populations and little technology improvement (Riahi et al 2011). To understand the KBDI's sensitivity to different climate projections, we also calculated KBDI in the late 21st century under RCP8.5 using another two sets of WRF outputs driven by Geophysical Fluid Dynamics Laboratory Earth System Model with generalized ocean layer dynamics component (GFDL-ESM2G, Donner et al 2011), and the Hadley Centre Global Environment model, version 2-earth system (HadGEM2-ES, Jones et al 2011). Please see Zobel et al (2018) for detail evaluation of these WRF regional climate model simulations. The three WRF simulations driven by different global model boundary conditions represent high (HadGEM2-ES), median (CCSM4), and low sensitivity (GFDL-ESM2G), respectively, of global average air temperature to the doubling of CO₂ (Sherwood et al 2014).

In addition to calculating the KBDI across the entire CONUS and Alaska, we also calculated the KBDI over forested lands—deciduous broadleaf (located in southeastern CONUS) and evergreen needleleaf (located in western and southern CONUS)—determined by Moderate Resolution Imaging Spectroradiometer (MODIS) as well as US Geological Survey land use data. These land use data were also utilized as static input for the WRF post-processing.

Monitoring Trends in Burn Severity (MTBS) dataset spanning 1984–2017 is employed here to explore the relationship between KBDI and fire activities. MTBS is developed utilizing satellite and local-report on a basis of 30 m spatial resolution (Eidenshink et al 2007). It provides burned area for each fire event and for different burned severity for fires greater than 1000 acres in the western US, and 500 acres or greater in the eastern US. Total burned area, fire count, averaged burned area per fire count are derived in this study at annual scale to calculate correlations with all KBDI and high KBDI values over entire CONUS for forest lands and all land types.

To compare the observationally (PRISM) and WRF-derived KBDI, the CONUS is divided into seven regions (figure 1) based on the national climate assessment (Wuebbles et al 2017): northeast (NE), southeast (SE), midwest (MW), southwest (SW), northwest (NW), Northern Great Plains (NGP), and Southern Great Plains (SGP). Four KBDI categories were defined following Weatherly and Rosenbaum (2017) with KBDI values 0–300, 300–600, 600–750, and 750–800. KBDI values larger than 600 indicate that fire starts easily, burns quickly, and is difficult or impossible (if KBDI >750) to control. This is referred to as 'high' KBDI to quantify high fire risk potential, as by previous studies (Janis et al 2002; Hamadeh et al 2015; Weatherly and Rosenbaum 2017) for various areas in North America. We also explore the relationship between KBDI (both annual mean and occurrence of high KBDI) and burned area using MTBS data in the observational time period.

Zoom In Zoom Out Reset image size

In order to investigate the relative impacts of changing precipitation and temperature on the changes of KBDI, three additional sets of KBDI were computed as follows. Let T_p (or T_w ) and P_p (or P_w ) represent the daily maximum temperature and daily precipitation from PRISM (or WRF output), respectively; and ΔT_w,j and ΔP_w,j represent the change of maximum temperature and precipitation in season j (spring, summer, fall, or winter), projected by WRF for the late 21st century under RCP8.5. Then, three new sets of KBDI are calculated using (1) T_p +ΔT_w,j and P_p ; (2) T_p and P_p +ΔP_w,j ; and (3) T_p +ΔT_w,j and P_p +ΔP_w,j , respectively, to investigate KBDI changes due to (1) only temperature changes, (2) only precipitation changes, and (3) both temperature and precipitation changes.

In order to investigate the statistical significance of the results (including long-term trend, historical versus future changes, and uncertainty due to multiple model results) presented here, we conduct student-t test as well as calculation of signal-to-noise ratio (SNR). SNR is a measure of a desired signal relative to the background noise, with SNR > 2 implying statistical significant signals. For example, studies in climate science have been using SNR to present the significance of the signal of climate change (Heo et al 2014, Scaife and Smith 2018).

Figure 1 shows the seasonal mean PRISM-derived KBDI from 1982 to 2017, divided into seven sub-regions. All seven regions have the highest KBDI in fall and the lowest KBDI in winter or spring, except within the SE region, Florida has the highest KBDI in winter and spring. Although most of the regions show KBDI peaks in the fall, the maximum KBDI occurs later over NGP, NW, and SW than over SGP, NE, MW, and SE (figure S2), suggesting the wildfire season generally begins in the southern and eastern CONUS and then moves to the western and northern CONUS.

Over the long term (figure 2), most of the regions show little variation in the estimated annual mean KBDI in the past 36 years (1982–2017), except for the SW and NW regions, which increase by 1.8 and 1.4 per year, respectively, over all land types. In particular, the KBDI increased the most over forested lands in the SW and NW; the annual mean (and annual 95th percentile) increases by +5.3 (4.0) and +2.9 (3.3) per year over the forested lands in SW and NW, respectively. The increasing trend of KBDI suggests a persistent long-term drying condition occurred over the West US, which may have played a role in mediating the increased fire activities over this region. Furthermore, the number of days with high wildfire potential (using the 95th percentile of KBDI in 1982–2011 as a baseline) increases by 16 and 25 d, respectively, over the forested SW and NW. The SW and NW also had more areas experiencing high KBDI (>600), with +0.85% (1 156 km²) and +2.4% (529 km²) increases per year (figure 2). In addition, the NW has changed more rapidly than the SW in magnitude, frequency, and space of calculated KBDI. These results are consistent with a previous study by Abatzoglou and Williams (2016), which focuses only over forested regions in the western CONUS.

Zoom In Zoom Out Reset image size

In order to use the WRF model outputs to project future wildfire potential, we evaluate the performance of WRF-derived KBDI for 1996–2004. The WRF run used here is driven by CCSM4 with bias correction (WRF-CCSM4 hereafter). Figure S3 shows the probability density functions (PDFs), as well as scatter plots using daily and monthly averaged KBDI. Results are for entire subregions, including all land types. Compared to PRISM-derived KBDI, the WRF-derived KBDI has a higher density of 0 < KBDI ≤ 100, and a lower density of KBDI > 100 over all the subregions. This indicates that, in general, WRF underestimates the frequency of moderate to high KBDIs seen in summer and fall and overestimates the frequency of low KBDIs seen in the spring. However, both WRF- and PRISM-derived KBDIs show the SW has the highest density of KBDI greater than 400 followed by the SGP and SE. The WRF-derived KBDI also aligns with PRISM during the high-KBDI seasons (figure S2), which represent the most critical times for fire management and planning over the fire-prone regions such as NW, SW, and SGP, as well as the SE. In particular, the WRF-derived KBDIs correlate strongly with PRISM-derived KBDIs over the SW and NW, with correlation coefficients of 0.76 and 0.86, respectively; the relative root mean square error (rRMSE) is also the smallest over the SW with a value of 35.5% (figure S3), indicating that the WRF model captures the daily variation of KBDI data with relatively small error over SW and NW, compared to the other subregions with lower correlations and higher rRMSEs. The discrepancy between KBDIs derived from PRISM and those derived from WRF-CCSM4 can be explained mainly by the model bias in precipitation and temperature (Wang and Kotamarthi 2015, Zobel et al 2018). Overall, the WRF model underestimates the KBDI, but captures the seasonal cycle, peak values, and PDF distributions reasonably well, especially over regions with high wildfire potential such as the SW and NW.

Figure 3 shows the WRF-derived daily KBDI in historical (1982–2017), mid-21st century, and late-21st century periods averaged for eight regions, including Alaska. Results are for entire subregions, including all land types. Although the peak of daily KBDI occurs at the same season in historical and future periods (except for MW), the KBDI values in the future are projected to be much higher than that in the historical period. The SGP, SE, and SW see the largest increases in KBDI magnitude with increases of 145, 132, and 116 during peak fire season in the fall by the late 21st century (table S1). The regional mean KBDI in the future is greater than the annual maximum KBDI in historic periods for more than 90 d over all seven subregions and 65 d in Alaska. This phenomena reflects that, for those regions that already have high fire danger, the fire season will be longer; for those regions that used to have low wildfire potential, such as the MW, SE, and NGP, there is a large increase in periods with high KBDI, indicating that in the future these regions could experience high wildfire danger (table S2). Spatially, the change in seasonal mean KBDI is also highly variable (figure S4). Over the MW, NGP, and NW, the changes (mostly increases) in summer and fall are large and significant, while those in the winter and spring are small and not significant. This is also true for northern and central Florida, as well as all of Nevada. The seasonal mean change over parts of the NW and SW are smaller than other regions, with some areas showing no change in seasonal mean daily KBDI.

Zoom In Zoom Out Reset image size

The seasonal mean changes in daily KBDI are important, and have been studied previously (Liu et al 2010, 2013). However, in order to understand the impact of climate change on extreme wildfire risk, the focus should be on high KBDI (Weatherly and Rosenbaum 2017), which is related to fires that are difficult to control and cause severe damage to society (e.g. critical infrastructure; public health) and economics. We investigated the change in number of days with KBDI > 600. As shown in figure 4, spatial features of changes in number of days with high KBDI are different from those in seasonal mean KBDIs (figure S4). For example, seasonal mean KBDI values are higher in the MW, NGP, NE, and SE in all four seasons, but high-KBDI events show very minor or no change over these regions. This is because these regions are mostly dominated by low- to moderate-KBDI events, and high-KBDI events are rare, especially in historic periods. However, over the SW, although the seasonal mean KBDI shows no significant change, the frequency of high-KBDI events increases significantly (by 40–60 d) in summer and fall, when the KBDI reaches its peak (figure 4). This indicates that in both history and the future, the SW is dominated by high-KBDI events, which will occur more frequently in the future over the SW and SGP. The NW behaves similarly to the SW; it experiences a significant increase in the frequency of high-KBDI events in the fall and summer (figure 4) but has areas of small changes in seasonal mean KBDI (figure S4). These findings are consistent with the long-term trend of number of days with high KBDI over the past 36 years, based on PRISM-derived KBDI.

Zoom In Zoom Out Reset image size

In addition, we examined the change of area with high KBDI (>600) from the historic period to the late 21st century. Alaska and the SE, SW, and NW have more grid cells (or area increase) during the summer and fall with KBDI > 600, with increased area of 1 392, 58 416, 46 560, and 27 780 km², respectively (table S3). This emphasizes that wildfire danger (which is already high) over these regions will be not only more frequent but also broader in area. Moreover, there are more grid cells with KBDI > 600 over the MW and NGP, where the wildfire potential used to be low in historic but will have longer fire season (figure 3 and table S1) and broader fire areas. The area increases are largest in summer and/or fall over all the subregions except the SGP, where the area increase is the largest in winter (40 812 km²). This indicates that the high KBDI over the SGP will occur not only in warm seasons but also in cold seasons, due to future temperature increase.

To examine the uncertainty of the modeled KBDI changes, we calculated the seasonal mean changes of KBDI in the late 21st century over all eight regions using three WRF model outputs from WRF-CCSM4, WRF-HadGEM, and WRF-GFDL (figure S5). In general, all of the three models project KBDI increases consistently; variations are relatively small compared to the mean changes projected by the three different models. In the fall, when most regions have the peak KBDI, all regions except Alaska have SNRs > 1.5, indicating that the WRF-CCSM4-derived KBDI is representative for all three models at a confidence level of 90%. In summer, SNRs are larger than 2 over the SGP and SE, which means all the models agree with each other about the projections in these regions. The SNRs in spring and winter are low, mostly because the KBDI and the predicted changes from historic to future periods in these seasons are low and not robust.

The future changes of KBDI depend on temperature and precipitation changes. Assuming temperature remains unchanged, lower KBDIs are generally predicted in response to precipitation increases projected over most of the regions such as the MW, SE, and NE (figure S6). One exception is over the SW in spring, where precipitation is projected to decrease as many studies have also reported robust dryness in future based on both regional climate models and GCMs (Cayan et al 2010, Gao et al 2014), thus increases of KBDI are predicted as well as over the MW in winter. In contrast, future KBDIs are predicted to increase predominantly due to the projection of warmer temperatures assuming no changes in precipitation (figure S7). The magnitude and area of the KBDI increases due to temperature changes only are much larger than those of the KBDI reductions due to precipitation changes only (figure S6), especially in the summer and fall. As a result, increases of KBDI for higher fire potential are prevalent in future projections when considering both temperature and precipitation changes (figure S4), and the amplified intensity, duration, and frequency of high KBDIs are mainly attributed to the temperature warming.