Contrasting the role of human- and lightning-caused wildfires on future fire regimes on a Central Oregon landscape

The study area covered 3.3 million ha along the eastern slope of the Cascade Mountains in central Oregon, USA (figure 1), and was the focus of several studies integrated in the 'Forests–People–Fire' project (Spies et al 2018) centred around the social-ecological system of a fire-prone forest landscape. The study area extends along a west-to-east moisture gradient that supports diverse vegetation types ranging from alpine forest to semi-arid juniper woodlands desert in an intermix of public, private and tribal land. The west side of the study area is dominated by moist forest managed as designated wilderness areas where mechanical treatments are limited and wildfires are rare but typically high-severity. Moving east, vegetation types transition from dry mixed forest to semi-arid juniper woodlands along the eastern boundary of the study area (Spies et al 2017). Overall, forested lands ranging from subalpine white fir to dry ponderosa pine cover 70% of the study area, and semi-arid juniper woodlands account for 20% of the study area. Federal forest lands managed by the U.S. Forest Service dominate the study area (48%) and include the Deschutes and the Fremont-Winema National Forests. The northern end of the study area includes the Confederated Tribes of the Warm Springs. Private industrial (12%), state forest (1%) and small forest ownerships (8%) occur on the drier portions of the study area and the wildland urban interface (12%) is concentrated around cities and major transportation corridors (Kline et al 2017). The study area has experienced an increase in population in response to the amenities associated with the region, e.g. environmental quality, recreation opportunities and access to public lands (Olsen et al 2017).

Zoom In Zoom Out Reset image size

2.2.1. Fire occurrence data

We obtained historical ignition data for 1992–2015 (figure 2) for the entire study area from the spatial wildfire database of the U.S. (Short 2017). There were 15 006 recorded ignitions (figure 2(A)) in the study area, but only 400 generated fires larger than 10 ha (figure 2(B)). Fires >10 ha represented 2.6% of all ignitions and 99% of the total burned area. Each fire record was associated with the x, y coordinate for ignition location, cause, ignition date and final fire size.

Zoom In Zoom Out Reset image size

The maximum area burned in a year was 63 000 ha (1996), and there were eight years with >20 000 ha burned. The largest single fire was 47 846 ha (Simnasho fire, 1996) followed by 37 626 (Barry Point fire, 2012) and 32 267 (B&B Complex fire, 2003). Lightning-caused ignitions correspond to 43% of the fires >10 ha, and 55% of the area burned, whereas human-caused ignitions leading to fires >10 ha were frequent (57%) but associated with a smaller share of total area burned (45%).

Temporal and spatial ignition patterns were distinct for human- and lightning-caused ignitions. Human ignitions were concentrated near populated areas and roads, whereas lightning-caused ignitions were distributed evenly across the study area (figure 2(A)). In terms of seasonality, lightning ignitions peak in mid-summer, whereas human ignitions tend to occur more broadly and extend towards spring and fall (figure 2(C)).

2.2.2. Land cover data

2.2.3. Historical ERC data

We developed a suite of spatially explicit statistical models that estimate wildfire ignition location (hereafter referred to as occurrence) and expected fire size as a function of ERC, location, and day-of-year for human and lightning-caused fires following the methods of Ager et al (2018). The ignition prediction model included separate models for human- and lightning-caused ignitions due to their distinct seasonalities (figure 2(C)) and spatial occurrence (figure 2(B)). The modelling workflow starts by estimating daily wildfire occurrence for the study area. Specifically we used the following formula to estimate the probability of any ignition per km² per day:

$$\begin{equation}{P_{{\text{ig}}}} = {P_{{\text{light}}}}x\left( {1 - {P_{{\text{human}}}}} \right) + {P_{{\text{human}}}}x\left( {1 - {P_{{\text{light}}}}} \right)\end{equation} \tag{ 1 }$$

where P_ig , P_light , P_human are the probabilities of any ignition, lightning-caused ignition and human-caused ignition, respectively.

We used logistic regression to estimate the probabilities of lightning- and human-caused ignitions. Specifically,

$$\begin{equation}{\text{Pr}}[Y = 1|X] = {\left[ {1 + {\text{exp}}\left( { - \theta } \right)} \right]^{ - 1}}\end{equation} \tag{ 2 }$$

and

$$\begin{align}\theta & = {\beta _0} + s\left( {{\text{ERC}}} \right) + s\left( {{\text{day}}{\text{.of}}{\text{.year}}} \right)\nonumber\\ & \quad + s\left( {{\text{longitude}}, {\text{latitude}}} \right){\text{ + land}}\end{align} \tag{ 3 }$$

where, Y = a binary response variable (lightning-caused ignition, human-caused ignition or occurrence of a fire >10 ha given ignition), X = (X₁, ..., X_n ) = a matrix of explanatory variables (ERC, location coordinates, day-of-year, land cover type) and s(X) = a smooth function of the three explanatory variables estimated from the data via spline functions. We used a periodic spline to estimate the effect of day.of.year and a two-dimensional spline to estimate a surface of locational effect. The location term was used as a surrogate to account for spatially-explicit variables (e.g. population density, roads, etc) that might have an effect on the probability of ignition or large fire occurrence and that were not included explicitly in the models.

For the estimation of fire size given a fire >10 ha, we used a log-normal regression with the final model used given by:

$$\begin{align}{\text{log}}\left( {{\text{fire}} {\text{size}}} \right) & = {\beta _0} + s\left( {{\text{ERC}}} \right) + s\left( {{\text{day}}{\text{.of}}{\text{.year}}} \right)\nonumber\\& \quad + s\left( {{\text{longitude}}, {\text{latitude}}} \right) + {\text{ land }}\nonumber\\& \quad + {\text{error}}\end{align} \tag{ 4 }$$

where, error was a Gaussian random error. All models were created utilizing the open-source R statistical program (R Core Team 2016).

2.3.1. Model application to future ERC

We then used projected ERC streams from 18 different GCMs in the spatiotemporal ignition prediction model to quantify changes in the number of fires per season, fire size, and frequency of EWEs by cause. We defined EWE as the maximum fire size by climate model for the first ten years of the simulation record (2006–2015) and tested how often the established maximum size record is expected to be exceeded over time.

To incorporate climate change into the spatiotemporal model of fire ignition and size, complementary grids of daily ERC were produced from statistically downscaled GCM data using the Multivariate Adaptative Constructed Analogs approach (Abatzoglou and Brown 2012; supplementary material). We downscaled daily meteorological fields from 18 different GCMs from CMIP5 (Rupp et al 2013) for both historical climate forcing (1950–2005) and future forcing (2006–2060) experiments, with the latter using RCP 4.5. The RCP 4.5 is a posited future pathway with continued climate mitigation whereby the globe warms approximately 2 °C above pre-industrial levels by 2060. These data were downscaled using the gridMET data to foster compatibility with the observed record. However, due to slight biases in covariability among GCM variables, we applied a trend preserving quantile delta mapping of resultant ERC such that the distribution of daily data for each model during 1979–2012 adhered to the distribution of observational data for the same period.

The output of the downscaling process was a series of daily ERC grids for each of the 18 GCMs. Each grid series was used to create additional daily synthetic ERC grid series by randomly shuffling years in a moving five-year window, creating four additional replicate grids (figure 3). All daily ERC grids share the same overall trends of the original ERC stream but allow for variability in daily ERC. This variability in ERC streams allows us to sample a larger range of possible daily ERC values for a given grid cell location and day-of-year. The spatiotemporal ignition prediction model randomly sampled from empirically-derived relationships between ERC, seasonality, location, and fire ignition and size to generate a list of fire ignitions expected to occur in a given fire year (henceforth, firelist). Each ignition listed in a firelist is associated with the following attributes: year, day-of-year, ERC, location, cause, land cover type and expected fire size. Figure 4 describes the time periods for all model data and analyses.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

For the model ensemble, we calculated the change between modelled contemporary (2006–2015) and projected future conditions (2031–2060) for the following variables: number of fires >10 ha, fire size, and occurrence of EWEs for human- and lightning-caused ignitions. We present results in terms of changes between contemporary and future conditions for the model ensemble average. We used the Kruskal–Wallis non-parametric test to compare the medians of the contemporary and future fire size, number of ignitions, and total area burned.

We defined the benchmark EWE as the maximum fire size by climate model for the first ten years of the simulation record (2006–2015) and tested (a) how often the established maximum size record is expected to be exceeded over the following simulated 45 years, and (b) how often the EWE benchmark was reset during that period, i.e. we tracked every year the maximum simulated fire size record was broken throughout the simulation period. We tested whether the probability of the maximum size record being broken differed from a theoretical model under stationary conditions. This was done by comparing modelled probabilities with a theoretical probability curve following decay of 1/n where n is the number of years from the start of the simulation record (Meehl et al 2009). This analysis accounts for the fact that as time passes (and records accumulate), it is increasingly unlikely for a stationary process to break a record.

Our analysis projected that number of fires >10 ha per season remained approximately the same between the contemporary and future time periods. This uniform fire occurrence is likely a response to the shape of the fire ignition-ERC distribution for the region (figure 2(B)), emphasizing the importance of regional studies that capture local patterns of fire occurrence. In the contemporary record, lightning ignitions were more likely on days with moderate ERC, and this likelihood diminished with higher ERC. One possible reason for the reduction in lightning ignitions under higher ERC is that in the region, lightning-caused fires occur after the passage of light precipitation, including that from cold fronts that can lead to a drop in ERC. Numerous fires from widespread cloud-to-ground lightning can flare for days to weeks after the lightning event due to extreme fire weather, reduced fuel moisture, and temperature increases (Schroeder and Buck 1970). It is also important to note that the ignition date that was used to assign a daily ERC value corresponds to the day of detection which may not coincide with day of ignition. Regardless of what explains the observed relationship between number of lightning ignitions and ERC this relationship may not necessarily carry on into the future. For example, Romps et al (2014) projected that by 2088 the rate of lightning flashes over the contiguous United States is likely to increase by half relative to the 1996–2005 reference period. However, the frequency of lightning strikes alone does not necessarily correlate well with fire starts, particularly in the Pacific Northwest (Rorig and Ferguson 1999) where lightning events can be associated with precipitation (wet lightning) or occur outside of the area of precipitation or with precipitation that evaporates before reaching the ground (dry lightning).

Changes in the number and seasonality of lightning-caused ignitions have implications for restoration efforts that plan on lightning ignitions under low-risk conditions to increase the pace and scale of forest restoration. While the number of lightning-caused ignitions was slightly reduced (−14%), this reduction occurred during summer months when fire managers might be less willing to manage lightning ignitions to accomplish ecological objectives due to greater risk of escape, concurrent fires, and potentially fewer resources available. On the other hand, we projected no changes in spring and autumn lightning ignitions when the weather might facilitate management of lightning- ignited fires for ecological objectives (Barros et al 2018).

The likelihood of human-caused fires was similar for ERC values between 40 and 70 and then increased slightly. This suggests the frequency of human-caused fires is more likely to increase in response to increases in daily fuel dryness (ERC) than lightning-caused fires, a problem that is compounded by evidence of the growing frequency and impact of human-caused fires (Abatzoglou et al 2018, Nagy et al 2018, Mietkiewicz et al 2020).

Projected fire size increased for human-caused ignitions—with fires on average 30% larger as a response to an increase in dryness (ERC) only. We did not account for concurrent high wind events or changes in fuels management, landcover, or suppression capacity that may affect future fire size-ERC relationships. We held all these factors constant and looked at the effect of future ERC on the occurrence of EWEs. Our results suggest that future fire regimes in this study landscape may be characterized by an increase in record-breaking fire sizes.

The increase (albeit small) in the number of human-caused fires combined with an increase in fire size can have implications for operating plans that focus on staffing, preparedness, and prevention (Cattau et al 2020). Decision points (e.g. daily ERC, number of concurrent ignitions) based on past conditions may be inadequate as most projections show historical records of area burned and fire size will continue to be broken over time. Adaptation strategies may include setting new ERC thresholds at which restrictions and emergency closures to public lands are enforced during periods when human ignitions are expected to be high (e.g. Fourth of July) and re-dimensioning of daily resource availability and initial attack response during pre-identified times.

Parsing out the relative contribution of human- and lightning-ignited fires in future fire regimes is also critical to inform stakeholder engagement in diverse settings (Bendix and Hartnett 2018). Despite overwhelming evidence in support of increased wildfire frequency and extent in response to a changing climate, at local and regional scales, perceptions of wildfire risk and risk drivers are less unequivocal. A survey of residents in rural communities of the Blue Mountains of Oregon showed that most participants perceive the rising frequency of large wildfires, but the causes and drivers of that increase were perceived differently and predicted by their sociopolitical identity (Hartter et al 2020). In some communities, climate change and its impact on fires today and in the future remain a polarizing theme. This will likely impact how decision-makers and land managers engage with stakeholders and how they craft effective risk mitigation and communication programs—particularly in cases where the risk associated with human-caused ignitions is expected to be compounded by future climate.

We did not account for changes in population density or human behaviour that may affect contemporary patterns of human ignitions and fuels. Similar to other rural areas in the Pacific Northwest (Christensen et al 2000), the study area has seen a shift from timber to a recreation economy with amenity-based property buyers acquiring small tracts of land as seasonal or secondary homes. This has been accompanied by a shift in the way land is perceived and managed (Hartter et al 2020). The many amenities rural spaces provide—environmental quality, access to public recreational and cultural resources—will continue to drive the migration from urban to rural spaces. Presumably, shifts in work practices and the increased ability to telework will compound these effects, however anticipating the role of human impacts on future fire regimes is challenging.

Our projections suggest that in the future individual fire events will be larger (and more frequent) than observed in the available historical record (1992–2015) (figure 9). These projections are in agreement with other regional studies that used alternative future climates to project expected changes in size and frequency of extreme fires albeit with different modelling approaches and definitions of extreme events (Stavros et al 2014b, McEvoy et al 2020). Whether the projected effects on fire occurrence and size will materialize depends on fire and fuel feedbacks not accounted for in our models. For example, in real landscapes, the occurrence of a future EWE is linked to whether and when EWEs occurred in the same area. The effect of past fires on future fire occurrence is typically associated with negative feedbacks (Halofsky et al 2020). In our model, ignitions simulated for any given year assume the same fuel availability every year and have 'no memory' of past fires. Previous fires reduce the fuel available to burn (Donato et al 2013), and as such, can reduce the likelihood of ignition (Parks et al 2016), the likelihood of an ignition becoming a large fire, and expected fire size—the components of our statistical model. However, while past fires can limit the occurrence and spread of subsequent fires, particularly in less productive systems, this effect dampens as time passes and vegetation recovers (Parks et al 2015) and is less pronounced under extreme fire weather characterized by strong winds and warm, dry conditions (Parks et al 2015).

We defined EWE based on size, but it is important to note that not all EWEs are necessarily socio-ecological disasters, i.e. events that caused fatalities, consumed primary homes, and/or were defined as so by a national government (Bowman et al 2017, Joseph et al 2019), and large fires can have an important role in restoring ecological conditions in fire excluded landscapes. However, large fires also have the potential to disproportionally impact vulnerable communities and magnify pre-existing vulnerabilities in ways that play out over years and decades. Long-term recovery and adaptation towards socio-ecological resilience, e.g. ability to afford insurance, rebuild structures, create and maintain defensible spaces, and live with smoke, may be disproportionally hindered in vulnerable communities (Davies et al 2018, Palaiologou et al 2019). In the conterminous U.S., 29 million people coexist with potential for extreme wildfires, but 12 million are socially vulnerable people, meaning that a wildfire event could be disastrous (Davies et al 2018). Furthermore, census tracts that are majority Black, Hispanic or Native American are associated with greater vulnerability to wildfire than other census tracts (Davies et al 2018). Under the status quo, projected climate change is expected to exacerbate social inequalities unless ecosystems and communities adapt to a changing climate and more wildfire (Schoennagel et al 2017) in an equitable way. Adaptation plans must integrate climatic, ecological, and socio-economic data at a scale relevant to identify where social risks are the greatest.

Large fire events can also impact public attitudes and trigger new policy in response to particularly extreme wildfire years (Fischer and Jasny 2017, Mockrin et al 2018, 2020) and can incentivize the development of fire-adapted communities (Schumann et al 2020). For example, the 210 ha Pine Forest community in Winthrop, WA, was exposed to fire during the 2015 Twisp River fire that burned within a quarter-mile of the community. Following the fire, the community organized to update their forest stewardship plan and has completed fuel reduction treatments in 75% of individual lots and 70% of common spaces (WADNR 2020).

Large fire events can also have long-lasting ecological impacts on ecosystems (Foster et al 1998), and those impacts can be compounded by subsequent reburns (Halofsky et al 2020), which our model indicates may be more frequent and extensive in the future. Fire size alone is not always indicative of the ecological outcomes (negative or positive) and should be accompanied by measures of severity per vegetation or land cover type. Future wildfire projections need to broaden the scope of fire regime descriptors to include cause and EWEs in addition to other fire regime metrics that relate to fire effects such as severity.

On the other hand, in many fire-adapted systems where fire reintroduction is a key management goal, more area burned can help accomplish important restoration work. Under certain conditions, fire managers can take advantage of opportunities created by large fires to achieve important forest health work by managing fires for multiple ecological benefits. Increasing and maintaining the capacity to manage very large events across multiple state, local and federal agencies with fire responsibilities will be key to adapt to a future where EWEs will be more frequent and larger.

Our models projected increases in area burned and the potential of increasingly large fires in this central Oregon study area by mid-century. These results quantify the potential effect of future ERC distribution on selected fire regime attributes. We highlight three major assumptions in this study: (a) our model assumes that contemporary ERC-ignition and ERC-fire size relationships will hold under future conditions; (b) the model did not explicitly account for changes in fire weather events, such as the east downslope winds that impacted the western Oregon Cascades in 2020; and (c) our model does not account for interactions between fire and other disturbances that may become more prevalent in a warming climate (Halofsky et al 2020). For example, on hotter and drier sites, the co-occurrence of drought, insect outbreaks, failed regeneration and non-native invasions (Kerns et al 2020) in response to disturbance have the potential to change ERC-occurrence and ERC-size relationships modelled in this study.