Dry Live Fuels Increase the Likelihood of Lightning‐Caused Fires

Live fuel moisture content (LFMC) is a key determinant of landscape ignition potential, but quantitative estimates of its effects on wildfire are lacking. We present a causal inference framework to isolate the effect of LFMC from other drivers like fuel type, fuel amount, and meteorology. We show that in California when LFMC is below a critical flammability threshold, the likelihood of fires is 1.8 times as high statewide (2.25% vs. 1.27%) and 2.5 times as high in shrubs, compared to when LFMC is greater than the threshold. This risk ratio is >2 times when LFMC is 10% less than the threshold. Between 2016 and 2021, the risk ratio was highest in 2020 (2.3 times), potentially contributing to the record‐breaking wildfire activity in 2020. Our estimates can inform several wildfire prediction and management applications, including wildfire suppression, prescribed burn planning, and public safety power shutoff implementation.


10.1029/2022GL100975
2 of 10 at high spatio-temporal resolution. As a result, many fire danger models at most only partially account for the effects of soil and plant traits on LFMC (Bradshaw et al., 1984;Fujioka et al., 2008;Hardy & Hardy, 2007;Yebra et al., 2013). It is thus crucial to understand whether these modeling limitations significantly affect wildfire predictions.
Recent developments in remote sensing-based LFMC mapping (Quan et al., 2021;Rao et al., 2020) now enable detailed studies of LFMC's role in wildfire occurrence. Several studies show a strong threshold-like correlation between low LFMC and wildfire occurrence (Chuvieco et al., 2009;Jurdao et al., 2012;Luo et al., 2019) or burned area (Dennison et al., 2008;Nolan et al., 2016). However, correlative studies such as these may inflate the effect of LFMC on wildfire occurrence due to cross-correlation between LFMC and other wildfire drivers like atmospheric aridity and winds. Additionally, previous studies of LFMC's influence on fire occurrence have generally lacked a reliable counterfactual. For example, if a region didn't burn, was it due to LFMC being too wet, a lack of ignition sources, or other, non-LFMC-related differences with regions that did burn?
Existing studies use two methods to approximate this alternative. Some use a random sampling of all unburned areas as a comparison (Chang et al., 2013;Chuvieco et al., 2009), thereby implicitly assuming that the unburned areas also experienced an ignition, an assumption which may not always hold (Abdollahi et al., 2019;Villarreal et al., 2022). Other studies use the areas just outside the burn scar as a comparison, assuming that LFMC was too wet outside the burn scar for the burn scar to extend further (Jurdao et al., 2012;Yebra et al., 2018). Although this method has the benefit of relying on areas near known ignitions (which may have similar fuel distribution as adjacent burned areas), factors such as human containment and meteorology may also affect the shape of the burned scar, undermining the resulting inferred LFMC effects. Our solution is to focus on lightning-caused fires, taking advantage of the fact that lightning occurrence is both an observable and pseudo-random source of potential ignitions enabling robust causal inference.

Methods
We design a randomized control experiment to estimate the effect of LFMC on lightning ignition probability holding all other geographic, topographic, meteorological, and fuel factors constant. We create lightning pairs within land cover types with one having dry LFMC and another having wet LFMC while minimizing differences in other covariates ( Figure 1). We apply our approach to forests, shrubs, and grasses in California between 2016 and 2021.

Creating Matched Pairs of Lightning Events
We classify each lightning event as occurring over either dry or wet pre-fire LFMC based on land cover-specific thresholds, with dry-LFMC lightning events defined as those associated with a pre-fire LFMC below 55% for Figure 1. Approach to select lightning events used for causal inference. Brown (cyan) circles indicate lightning events at places where live fuel moisture content is less (greater) than a pre-determined threshold. To find matched pairs of lightning events with similar covariates, we compute the Mahalanobis distance (calculated over all covariates) across all possible dry-LFMC and wet-LFMC pairs and matched pairs were chosen to minimize the sum of the between-pair Mahalanobis distance across all matched pairs. Note that while 12 features are used for matching, only two feature axes are shown here for visualization. grasslands, below 106% for shrublands, or below 81% for forests. We choose these thresholds because previous work in Argentina showed that fire size varies across these breakpoints (Argañaraz et al., 2018), which are similar to thresholds established in Californian chaparral (Dennison & Moritz, 2009). We classify LFMC to enable causal estimation of LFMC's effect on fire likelihood based on the potential outcomes framework (see Text S1 in Supporting Information S1). However, the relevance of such thresholds may be context-specific (Pimont et al., 2019). We therefore also test the sensitivity of our results to varying degrees of LFMC difference from this threshold.
Each lightning pair consists of a dry-LFMC strike and a wet-LFMC strike, with land cover being identical and all other wildfire-related properties being as similar as possible. As covariates, we consider latitude and longitude, above-ground biomass (AGB) and dead biomass to control for fuel availability (Hessilt, 2022;Safford et al., 2022), wind speed on the lightning day (Goss et al., 2020), cumulative precipitation 12 months before the lightning day, average vapor pressure deficit (VPD) over the 4 months before the lightning day to control for soil moisture (Dadap et al., 2019), and litter moisture effects (Dorph et al., 2022) (Table S1 in Supporting Information S1). We use lagged aggregates for precipitation and VPD rather than instantaneous values because soil moisture and litter moisture have memory (McColl et al., 2017).
We control for dead fuels by replicating the Safford et al. (2022) approach to estimate dead biomass. We use annual tree mortality maps from the aerial detection and monitoring program (USFS, 2017) from 2012 to 2016, when drought-induced tree mortality was high across California (Lund et al., 2018). We filter out background mortality (shapefiles with <1 dead tree per acre) to isolate areas with recent mortality. We then create a cumulative map by summing the remaining dead trees across 2012-2016, resulting in a map with the number of dead trees at the end of 2016, the starting year of our analysis. We also use mean canopy height, its spatial variance, and the spatial variance of normalized difference vegetation index (NDVI) in the lightning pixel to control for vegetation structure and its sub-grid scale heterogeneity (Fernández-Guisuraga et al., 2021), and elevation and slope to control for topographic influences (Holden & Jolly, 2011).
To find the lightning strike pairs with the most comparable covariates, we perform a global optimization to minimize the sum of Mahalanobis distances between each pair's covariates (Text S1 in Supporting Information S1). We restrict our analysis to the peak fire period (June to November, Westerling et al., 2006) and to the period when LFMC data are available (2016-2021).

Estimating the Average Effect of Dry LFMC
The resultant data set has pairs of lightning events, each with a dry-LFMC strike and a wet-LFMC strike from the same land cover, and other similar covariates. For each pair, we analyze whether lightning ignited a wildfire in either member. We do so by checking whether a burned fire pixel appeared within 2 days of the lightning date in the lightning grid cell, which allows for errors in the detection date of a fire. This way, we label each lightning event as "fire" or "no fire." As an independent check on the accuracy of our labels, we compare our labeled fires to fires from the CAL FIRE's Fire and Resource Assessment Program (FRAP) (CAL FIRE, 2023). The FRAP investigates the causes behind select fires. Only 3.1% of FRAP fires caused by something other than lightning (covering 6% by area) are considered lightning-caused fires in our data set, adding confidence to our method. We calculate the likelihood of a fire for the dry-LFMC and wet-LFMC lightning events separately and divide the two to obtain the risk ratio (Equations 1 and 2). The risk ratio quantifies the effect of dry LFMC on the likelihood of fire relative to that of wet LFMC.
To further characterize the effects of LFMC, we re-compute the risk ratio separately for lightning events: (a) in each land cover type, (b) belonging to four equal-sized bins of LFMC difference between the two lightning events of each pair, and (c) in each year. Finally, we quantify the uncertainty associated with each estimate of the risk ratio by running a conditional maximum likelihood estimation (Fisher) test to obtain 95% confidence intervals.

Data Sets
We use daily lightning data from the National Lightning Detection Network (Orville et al., 2011), which has a spatial accuracy of 0.1° (∼10 × 10 km). We obtain fire occurrence data from the Moderate Resolution Imaging Spectroradiometer (MODIS) burned area product MCD64A1 (Boschetti et al., 2013) at 500 m resolution. To classify lightning into dry-LFMC lightning and wet-LFMC lightning, we use LFMC mapped across the western United States at 250 m resolution and a 15-day timescale (Rao et al., 2020, Figure S1 in Supporting Information S1). These maps were created using a recurrent neural network based on synthetic aperture radar data (along with optical reflectance, topographic, and soil characteristics) and trained on ground measurements from the National Fuel Moisture Database (United States Forest Service, 2018). These maps have low uncertainty across the diverse vegetation and topography of California (R 2 = 0.63) and are further described in Text S2, Figures S1 and S2 in Supporting Information S1. We use the most recent pre-lightning LFMC to classify the lightning events.
We use 2016 and 2019 land cover data from the National Land Cover Database (Homer & Fry, 2012), precipitation and VPD data from the Parameter-elevation Regression on Independent Slopes Model (PRISM Climate Group Oregon State University, 2004), elevation data from USGS (2018), canopy height from Simard et al. (2011), NDVI from MODIS (Huete et al., 2010), wind speed from GRIDMET (Abatzoglou, 2011) and the 2010 harmonized AGB map from Spawn et al. (2020). Time-variable and high-resolution AGB datasets do not exist for California; however, we assume spatial variations in AGB dominate temporal variations, and neglect the latter. Because the lightning pairs are also matched on more recent land cover data, and major changes in AGB would be expected to be associated with land cover change, we expect the impact of this assumption to be small.
Because the lightning data has the coarsest spatial resolution of the datasets, we linearly average all other data to that resolution (10 × 10 km), except land cover and fire. For land cover, we assign the most common land cover type to each 10 km grid and filter out the 3.2% of grid cells where it occupies less than 40% of the grid area. For fire, if any 500 m pixel within the lightning footprint burned, we associate that lightning event with an ignition. The data sources and their aggregation are summarized in Table S1 in Supporting Information S1.

Change in Covariate Imbalance After Matching Lightning Strikes
Compared to wet-LFMC lightning strikes, dry-LFMC strikes are biased toward low precipitation, low AGB and dead biomass, shorter vegetation, and high VPD (square points in Figure 2). Multiple covariates (altitude and past VPD) differ by more than one standard deviation between dry-LFMC and wet-LFMC lightning occurrences due to strong cross-correlation with LFMC ( Figure S3 in Supporting Information S1), motivating the need to match dry-LFMC and wet-LFMC lightning events to robustly determine the influence of LFMC.
After matching, the standardized difference between the covariates is less than 0.10 for all covariates except VPD (0.30), terrain slope (0.15) and altitude (0.32) (Figure 2). Due to the high bias in dry-LFMC lightning occurrences, we can match only 13,358 out of 38,581 lightning events (Text S1 in Supporting Information S1).
Of the resulting 13,358 lightning events, 235 (or 1.8%) caused an ignition (Figure 3d). Ignition likelihood was strongly correlated to LFMC (correlation = −0.87; Figure S4 in Supporting Information S1). Most of the lightning-caused ignitions are concentrated in northern California, and hence show a different spatial pattern than the strikes without ignitions (Figure 3a). However, we observe no related spatial patterns for the LFMC associated with lightning (Figure 3b), or across land cover types (Figure 3c).

Figure 2. Effects of matching lightning strikes on covariates.
Standardized differences in covariates before (squares) and after (circles) matching of dry-Live fuel moisture content (LFMC) and wet-LFMC lightning strikes. Lower absolute standardized differences indicate lower differences between covariates, limiting their ability to confound our inference. The symbol represents spatial variance within the lightning pixel (10 × 10 km).

Effects of LFMC on Fire Likelihood
Across the 13,358 matched lightning events in our data set, 85 fires ignited in wet-LFMC areas, while 150 fires ignited in dry-LFMC areas. The likelihood of wet-LFMC lightning igniting a fire is thus ∼1-in-78, whereas the likelihood of dry-LFMC lightning igniting a fire is ∼1-in-43. This represents a risk ratio for dry-LFMC lightning of 1.8 (95% confidence interval: 1.4-2.3). That is, when LFMC is less than the land cover-specific threshold-and under constant vegetation, fuel availability and prior meteorology-the likelihood of a fire is 80% higher (Figure 4a). The risk ratio is 2.5 (1.3-4.7) in shrubs, 1.6 (1.2-2.2) in forests, and 1.7 (0.4-6.9) in grasslands. The high uncertainty in grasslands is likely because only 8 out of 235 fires occurred in grasslands.
Segregating the matched pairs of lightning strikes based on their difference in LFMC (LFMC dry-LFMC lightning − LFMC wet-LFMC lightning ) reveals that greater differences in LFMC are associated with a higher likelihood of fire ( Figure 4b). The risk ratio is relatively low (1.2; 0.7-1.9) when the difference in LFMC is less than 10% between the matched pairs but exceeds 2.0 when the LFMC differs by more (reaching 2.2 in the highest bin). The risk ratio for all Δ LFMC categories is significant at the 95% confidence level, except when the LFMC difference is less than 10%.

Annual Variation in Risk Ratio at Lightning Sites
When compared within years, 2020 had the highest risk ratio (2.3; 1.3-4.1), followed by 2017 (2.0; 1.2-3.1, Figure 4c). All other years' risk ratios are not significant (p > 0.05) which may be due to a combination of factors such as where lightning occurs each year, the interannual variability of LFMC, lower sample sizes, etc. (Table S2 and Figure S1 in Supporting Information S1). The risk ratio in 2020 was high even though neither the fraction of dry-LFMC regions (0.7), nor the average difference of LFMC between each matched pair of lightning strikes (−19%) showed significant deviations from other years. Since risk ratios of shrubs and forests differ significantly, we checked for differences in land cover of lightning locations but found that 2020 is not an outlier. However, the lightning strikes leading to ignition in 2020 were overwhelmingly concentrated in August, unlike in other years when the lightning strikes with ignitions were more distributed throughout the warm season ( Figure S7 in Supporting Information S1). Note that 2017 also had high risk ratio relative to other years, but did not have an anomalous lightning ignition pattern, nor other clear explanatory patterns in LFMC or land cover to explain the high risk ratio (Table S2 in Supporting Information S1). Further research is needed to explain the high risk ratio in 2017 and more generally, the annual variation of risk ratios.

Discussion and Conclusions
Wildfires are complex phenomena that can occur under many conditions, making it difficult to determine the influence of different drivers using correlative studies alone. We provide the first causal evidence that LFMC increases the likelihood of a lightning-caused wildfire after controlling for fuel availability, precipitation, VPD, winds, and other location-specific differences in fuel flammability (Figures 3 and 4). This result is robust in the presence of unobserved confounders (Text S3 and Figure S5 in Supporting Information S1). Thus, under the same fire-weather conditions, LFMC may vary enough to influence live fuel flammability.
The effect of LFMC on ignition risk varies by land cover. The risk ratio is significantly higher for shrubs (2.5) than forests (1.6, Figure 4a), despite the higher LFMC threshold in shrubs. Although few studies have directly measured the lightning ignition efficiency of individual vegetation classes, the available research demonstrates that vegetation composition plays a significant role in lightning fire occurrence. Even outside California, specific conifer species are associated with increased lightning ignition efficiency, leading to high fire occurrences in forests relative to other types of land cover (e.g., Oregon:  In all plots, points indicate the mean and whiskers extend to 95% confidence intervals. The dashed line indicates a risk ratio of 1 (no effect).
ignition probability due to an abundance of fine fuels. In contrast, in Tasmania, the highest ignition efficiencies were observed in button grass (Nampak et al., 2021). These results further demonstrate that lightning ignition efficiency may vary geographically, as a function of broader climatological and physiological controls.
California's shrubs and forests are located in areas with contrasting climatology and fire-climate regimes (Keeley & Syphard, 2017). Among areas with frequent lightning, shrubs are predominantly located in the high-elevation desert east of the Sierra Nevada, while occupying a higher average elevation than California's forested areas (Figure 3c). As a result, lightning strikes in shrubs have a significantly higher average elevation ( Figure S6 in Supporting Information S1) than those in forests, which may influence solar radiation and its effects on live and dead fuel flammability. Further, our LFMC data are derived from a C-band synthetic aperture radar, which does not fully penetrate forest canopies. Because the retrieved LFMC is mostly sensitive to the crown-rather than surface-fuels, LFMC may less accurately reflect the actual live fuels burning in forests (Krawchuk et al., 2006), lowering the apparent influence of LFMC.
Our analysis also sheds light on the role of several factors in driving the record-setting 2020 California fire season, when ∼4.3 million acres burned (Safford et al., 2022). 2020 was associated with the highest risk ratio (i.e., high effect of LFMC on ignition probability; Figure 4c), but the risk ratio is not explained by several expected controls. Thus, the higher apparent effect of LFMC on lightning-ignited wildfire must have been promoted by other factors besides those in our framework. Tropical storm Fausto reached California between 15 and 17 August 2020, producing 859 lightning strikes-about 15 times more than the average for any three days between June and November from 2016 to 2021 according to the National Lightning Data Network. Given the unusual concentration of 2020 lightning ( Figure S7 in Supporting Information S1), which coincided with an extreme heatwave over California, suppression activities may have been particularly challenged while firefighters attempted to contain large areas simultaneously (Safford et al., 2022). In addition, the COVID-19 pandemic may have challenged suppression efforts in 2020. Although more research is needed to better understand this unusual fire year, our results suggest that human fire management characteristics might modulate how different landscape and climate factors affect natural ignition risk.
Although we report an approximate doubling of the likelihood of a wildfire under dry-LFMC conditions (Figure 4b), fire probability remains low. When LFMC is dry, the probability of a fire given a lightning strike is 2.25%, compared to 1.27% when LFMC is wet (Figure 4d). These low probabilities of lightning ignitions are consistent with previous research (Abatzoglou et al., 2016;Dorph et al., 2022;Nampak et al., 2021).
The improved understanding of LFMC's effect on fuel flammability can be used to inform several fire management practices. LFMC can vary significantly across landscapes and even between nearby species or individuals depending on soil and plant hydraulic traits (Nolan et al., 2022;Rao et al., 2022;Scarff et al., 2021), but these variations are not fully accounted for in fire danger models. Our robust, causal demonstration of the significant role of LFMC in regulating ignition probability motivates further efforts to account for LFMC in wildfire danger estimation, either through using remote sensing-based maps or through ecohydrological models that can forecast spatially variable LFMC responses to weather conditions. Because California uses a wide network of LFMC measurements (United States Forest Service, 2018) to determine staffing levels of local and regional wildfire response units, a more accurate estimation of LFMC's effect on flammability may improve the allocation of firefighting resources. Similarly, a more accurate quantification of ignition probability can help to target public safety measures, such as power shutoffs to reduce the potential for accidental ignitions. Moreover, there is a severe backlog of prescribed burns in California (Miller et al., 2020). Since the safety of prescribed burns depends partially on LFMC (Biswell, 1999), improved understanding of flammability under different land cover and LFMC conditions may improve prescribed burn planning. As the climate continues to warm and dry, prescribed burn windows may shrink (Dong et al., 2022;Ellis et al., 2021), further fueling the need for an accurate estimate of LFMC's effect on flammability. Finally, as a changing climate alters wildfire regimes in California and globally (Ellis et al., 2021;Keeley & Syphard, 2016) our causal framework may be used to attribute and update the effects of other drivers on wildfire occurrence.

Data Availability Statement
The scripts supporting the analysis can be found at https://doi.org/10.5281/zenodo.8128765. Source of all data used is available in Table S1 in Supporting Information S1.