A Paleoclimate-Compatible Framework for Modeling Lightning-Caused Ignition Probability in Alaska

Abstract

Understanding the role of historical lightning-driven fire regimes in shaping terrestrial ecosystems and carbon cycles requires reconstructing fire from data beyond the instrumental record. Previous efforts have relied on paleo proxies, such as charcoal records, but these approaches are limited by their coarse spatial extent. Alternatively, process-based modeling offers a spatially continuous pathway for simulating lightning-caused fire regimes. However, existing lightning prediction models use upper-atmospheric variables, such as convective available potential energy (CAPE), that are not available in paleoclimate reconstructions, limiting their use beyond the instrumental period. Here, we develop a probabilistic framework for simulating lightning-caused fire ignitions that (1) relies on variables available in paleo reconstructions (near-surface climate, fuel moisture, and land cover) and (2) decomposes lightning-caused fire occurrence into two components: lightning strike rate and lightning ignition efficiency. Both components were trained on modern observational data for Alaska during 2002–2011, and then combined in a Bernoulli model to estimate daily fire probability. Near-surface climate predictors captured spatial and temporal variability in lightning activity with performance comparable to CAPE-based models, and ignition efficiency models showed strong discrimination between fire-causing and non-fire-causing strikes. Despite overestimation under high-risk conditions, the Bernoulli model demonstrated strong discriminatory skill (ROC AUC = 0.894), effectively ranking fire risk across space and time. By explicitly separating lightning occurrence from ignition efficiency and relying on variables available in paleo reconstructions, this approach provides a transferable framework for simulations of historical lightning-fire regimes.

1. Introduction

Wildfires play an important role in ecosystem disturbance regimes, vegetation dynamics, and carbon cycling. Lightning-caused wildfires account for a disproportionately high fraction of the total area burned in many fire-prone regions [1,2]. This is in part due to their association with holdover fires that can smolder for days before detection [3] with longer holdover durations in boreal forests [4,5]. Furthermore, lightning-caused fires respond to climatic and environmental drivers that differ from those of anthropogenic ignitions [6,7]. In Alaska, lightning-caused fires dominate area burned in part because fire management systems allow fires in remote regions to burn with minimal suppression. This corresponds to roughly two-thirds of interior Alaska [8]. Human-caused fires, however, tend to occur near communities and roads where aggressive suppression keeps them small [9,10]. This Alaska-specific fire management approach means that lightning- and human-caused fires behave in distinct ways, emphasizing the need for a targeted approach to modeling lightning-caused ignition. In recent decades, increasing lightning frequency in boreal North America (including Alaska) has driven a rise in fire occurrence [11], and this trend is expected to continue under future climate projections [12].

Reconstructing historical lightning-driven fire regimes offers a pathway for understanding future regime shifts by establishing long-term baselines for climate, fire, and ecosystem interactions. Previous approaches have estimated past fire regimes from charcoal and ice core records [13,14,15,16,17]. However, these proxies are limited in their spatial extent because charcoal deposition declines sharply with distance from a fire. That is, sediment records primarily capture fire activity in the immediate vicinity of a lake [18]. Process-based modeling offers a spatially continuous complement to these proxy-based approaches. By coupling paleoclimate reconstructions with lightning and fire prediction models, fire dynamics can be simulated across centuries or millennia. Before this is possible, however, two limitations must be addressed.

First, most models designed to predict climate-driven changes to lightning rely on atmospheric variables that are unavailable in paleoclimate reconstructions [19,20,21]. Examples include convective available potential energy (CAPE) [22,23,24], column saturation level, lifting condensation level [22], cloud top height, updraft intensity, cold cloud depth [25,26], and geopotential height [27]. Reliance on upper-atmospheric variables is also evident in Alaska-specific models: Bieniek et al. [28] identify convective precipitation as an important predictor and Chen et al. [29] apply CAPE. Recent work from Uden et al. [30] demonstrated that lightning strike rates can be predicted using near-surface variables with comparable accuracy to CAPE-based approaches, providing a viable pathway for paleo-compatible data.

Second, existing lightning-fire parameterizations for boreal North America rarely represent both lightning occurrence and ignition efficiency as independent climate-driven processes. Several studies have modeled fire from lightning occurrence, climate, fuel data, or landscape characteristics, but omit climate-driven changes to lightning [31,32,33]. Other work simulates lightning from climate but does not incorporate ignition efficiency [11]. Neither approach captures the joint conditions required for a fire: an ignition source (lightning) and the effectiveness of the source to result in combustion (ignition efficiency). An exception is Hessilt et al. [12], who use climate-driven simulations of lightning frequencies and ignition efficiency probabilities to estimate future lightning-caused fire counts for boreal North America. Separating these processes preserves interpretability and allows climate signals to differ across lightning occurrence and ignition.

Here, our modeling approach is motivated by a specific paleo application: driving lightning-caused fire predictions using last-millennium climate reconstruction at daily resolution [34,35], and last-Holocene estimates of vegetation cover [36,37,38]. While other paleoclimate products exist, for example, CHELSA-TraCE21k [39] provides global temperature and precipitation coverage at 100-year timestep over the last 21,000 years, their coarser temporal resolution and limited variable availability are not compatible with the framework developed here. This work therefore constitutes a foundational step in reconstructing past lightning-caused fire regimes: developing and validating a modeling framework against modern observational data for Alaska. Our approach allows for future studies to apply these models to reconstruct last millennium estimates of fire and vegetation dynamics beyond traditional paleo fire proxies [40].

With these motivations in mind, we develop a probabilistic modeling framework for lightning-caused fire designed for application in paleoclimate contexts. Specifically, we decompose lightning-caused fire probability into two interacting components: lightning strike rate and lightning ignition efficiency. Both components are trained using predictor variables available in paleoclimate datasets, including near-surface climate variables, fuel moisture indices, and land cover for Alaska, 2002–2018. We then combine these processes into a Bernoulli trial that accounts for both an ignition source (lightning strike rate) and the probability of ignition given a strike (ignition efficiency) to estimate the probability of at least one lightning-caused fire within a given gridcell and day. This approach provides a transferable, probability-based pathway for applying paleoclimate data to reconstruct historical lightning–fire regimes.

2. Materials and Methods

2.1. Data

2.1.1. Lightning

Lightning strike data, including location and time, were obtained from the Alaska Lightning Detection Network (ALDN), maintained by the Alaska Fire Service (AFS) [41,42]. We used two datasets: an earlier dataset (2002–2011) and a newer dataset (2012–2018). From 2002 to 2011, lightning was detected using an impact-based system that reported cloud-to-ground lightning primarily as flashes. During this period, changes in sensor technology, processing software, and network configuration resulted in variable detection efficiency and relatively coarse spatial accuracy, particularly in remote regions of Alaska. As a result, this dataset is characterized by lower positional accuracy. Beginning in 2012, the ALDN transitioned to a time-of-arrival (TOA)-based system. The TOA system records individual strokes and includes cloud-to-ground strikes. Due to differences in stroke versus flash reporting, expanded sensor coverage, and improved detection range, the AFS reports detecting ~2.25 times more lightning events relative to the earlier systems; there are 1,882,857 strikes in the earlier dataset and 1,784,663 strikes in the later dataset. Due to these differences in detection methods and accuracy, we train and test our models on each period separately.

While the models developed in this study are applied to the entire state of Alaska, it should be noted that sensors are heterogeneously spaced [28], making detection efficiency lower in remote areas far from sensors, particularly for the pre-2012 data. This may introduce spatial bias in lightning observations, for example, gridcells located in southeastern Alaska.

2.1.2. Wildfires

Wildfire ignition data were obtained from the Alaska–Yukon–Northwest Territories Fire Emissions Database (AKFED), version 2, archived at the Oak Ridge National Laboratory Distributed Active Archive Center [43]. AKFED provides information on wildfire occurrence across boreal North America from 2002 onward. Fire ignition locations and dates are derived using a combination of large fire perimeter records and satellite-based fire detections from MODIS. Ignition timing is estimated to within approximately one day, and ignitions are identified at spatial resolutions ranging from 500 m to 1 km, depending on whether a fire is detected from a single pixel or multiple adjacent detections. When multiple active fire detections occur simultaneously, the ignition location is defined as the centroid of the detected pixels.

2.1.3. Climate and Fuel Moisture

Climate predictors were downloaded from the Copernicus Climate Change Service. We obtained hourly ERA5 reanalysis data on a 0.25° × 0.25° grid for summer months (June–August) from 2002 to 2018 [44]. The raw variables include CAPE, 2 m air temperature, 2 m dewpoint temperature, total precipitation, 10 m wind speed, surface downward shortwave radiation, and surface pressure (Figure 1a–f). After aggregating to daily means (and daily sums for precipitation), temperature and dewpoint temperature were converted to Celsius and used to calculate relative humidity. The product of CAPE and precipitation was also calculated at the daily scale. We also computed the Canadian Forest Fire Danger Rating System indices following Van Wagner and Pickett [45] from the ERA5 data. These include daily fine fuel moisture code (FFMC), duff moisture code (DMC), and drought code (DC) (Figure 1g–i).

2.1.4. Land Cover

We used land cover data from the North American Land Change Monitoring System (NALCMS) to derive vegetation cover predictors [46]. The NALCMS products provide land cover classification across North America using a standard 19-class legend based on the Land Cover Classification System (LCCS) of the Food and Agriculture Organization of the United States. Four timesteps were used: 2005, 2010, 2015, and 2020. For the years 2005 and 2010, land cover maps are available at 250 m spatial resolution using MODIS satellite imagery. For the 2015 and 2020 datasets, newer generation NALCMS products were produced at 30 m spatial resolution using Landsat multispectral satellite imagery.

For downstream paleo applications of our models, we reclassified the data from 19 classes to three broad cover types consistent with paleo reconstructions of land cover (REVEALS; [36,37,38]). These include conifer (representing temperate or sub-polar needleleaf forests and sub-polar taiga needleleaf forest), broadleaf (tropical/sub-tropical and temperate/sub-polar broadleaf forests), and unforested/open land (shrubland, grassland, polar dwarf vegetation, wetland, barren, and snow/ice). Classes corresponding to cropland, urban land, water, and other non-natural surfaces were excluded. Mixed forest was allocated proportionally into the conifer and broadleaf categories based on the relative prevalence of pure conifer versus broadleaf forest within a gridcell. To match the land cover data to the ERA5 climate and fuel moisture grid, we assigned each land cover pixel to the nearest ERA5 gridcell. Within each ERA5 gridcell, the sum of each reclassified category (conifer, broadleaf, and unforested/open land) was divided by the number of NALCMS pixels so that each cover type is a proportion that sums to 1 (Figure 1j–l). This aligns modern land cover data with available paleo land cover reconstructions.

2.2. Model Definitions

Lightning-caused fire occurrence was modeled as a two-step process in which (1) the rate of lightning strikes and (2) the probability that an individual strike ignites a fire are estimated separately and then combined to estimate the overall probability of fire. Figure 2 provides an overview of our workflow. We first trained lightning and ignition models, which varied according to input predictors, modeling approach, and training period (2002–2011 vs. 2012–2018). Models were evaluated across multiple performance metrics. The lightning and ignition models with the best overall balance across metrics were then implemented into the fire probability Bernoulli model. Finally, fire probability predictions from this working model were evaluated against fire observations.

Lightning and fire occurrence are rare events, making observational datasets highly zero-inflated. Excessive zero values can dominate model fitting, reduce a model’s ability to detect relationships between event and predictors, and bias parameters toward predicting no occurrence. To reduce the number of zero outcomes and focus the analysis on periods when lightning and fire activity are most likely, we limit the training data to the summer fire season, June–August. This practice has been applied to other lightning prediction and fire ignition models [12,23,29,47].

2.2.1. Lightning Strike Rate, r_strike

To predict the spatial and temporal distribution of lightning activity, we retrained a model developed by Uden et al. [30] that relates near-surface climate variables to cloud-to-ground lightning strike rates. We deliberately use near-surface climate predictors rather than upper-atmospheric data, as they are widely available in paleoclimate reconstructions, and have demonstrated comparable performance. These models were re-parameterized for Alaska using the climate and lightning data described above. For each summer season (June–August) and gridcell, we calculated the mean lightning strike rate (strikes/km²/month) and associated averages of temperature, precipitation, wind, relative humidity, shortwave radiation, and surface pressure. The data were split into train (80%) and test (20%) sets and predictor variables were z-score standardized.

Seasonal lightning activity was modeled as a climate-dependent random variable representing the expected rate of lightning strikes per unit area. Let $y_i=r_{strike}$ denote the observed summer-season lightning strike rate in gridcell i and let $x_i^{\top }$ be a vector of standardized near-surface climate predictors. We tested three statistical approaches to predicting $r_{strike}$: a linear model, a Gamma generalized linear model (GLM), and a Gamma Bayesian model.

Linear Model

In the linear formulation, lightning strike rate was assumed to follow a Gaussian distribution:

$$\begin{gathered} y_i|{\mu }_i, \sigma ~Normal\left({\mu }_i,{\sigma }^2\right), \\ {\mu }_i=a+x_i^{\top }b, \end{gathered}$$

(1)

where $a$ is an intercept, $b$ is a vector of regression coefficients, and σ is the residual standard deviation. This model was fitted using the base R lm() function [48].

Gamma GLM

To account for the strictly positive and right-skewed nature of lightning strike rates, we also fit a Gamma GLM with a log link:

$$\begin{gathered} y_i| {\mu }_i,\kappa ~ Gamma(k, k/{\mu }_i), \\ {log(\mu }_i)=a+x_i^{\top }b, \end{gathered}$$

(2)

where $k$ is a shape parameter controlling dispersion and ${\mu }_i$ is the expected strike rate. $a$ and $b$ were fitted in base R using glm() with family = Gamma(link = log) [48].

Gamma Bayesian

Finally, we fit a Bayesian Gamma model where both the shape and rate parameters of the Gamma distribution were allowed to vary as functions of climate:

$$\begin{gathered} y_i| {\alpha }_i,{\beta }_i ~ Gamma({\alpha }_i,{\beta }_i), \\ log\left({\alpha }_i\right)=a_{\alpha }+x_i^{\top }{b}_{\alpha }, \\ {log(\beta }_i)=a_{\beta }+x_i^{\top }{b}_{\beta }. \end{gathered}$$

(3)

All regression coefficients were assigned Normal priors,

$$a_{\alpha },a_{\beta }, b_{\alpha }, b_{\beta } ~ Normal(0,1),$$

(4)

and posterior inference was performed using four Hamiltonian Monte Carlo chains in the rstan package version 2.21.8 [49] in R. This formulation allows both the mean and variance of lightning strike rates to respond to climate variability.

2.2.2. Lightning Ignition Efficiency, P_ignite

We model ignition efficiency using climate (precipitation, temperature, relative humidity, wind), fuel moisture variables (fine fuel moisture code, duff moisture code, and drought code), and vegetation cover (conifer, broadleaf, and unvegetated/open land). Following methods from Hessilt et al. [12], spatial buffers were applied to each fire to select lightning strikes within 2.5 km or 500 m (depending on the earlier or later dataset, respectively). We then filtered strikes that occurred within 1 day after (to account for inaccuracies in fire detection) and 120 h before the fire. One hundred and twenty hours represents the point at which true lightning-caused ignitions can no longer be reliably distinguished from chance co-occurring lightning strikes (according to methods from Hessilt et al. [12]), rather than the true upper bound on holdover fires, which can extend far beyond this period [4,5]. When multiple strikes fell within the spatial and temporal buffer of a fire, proximity index A was calculated (derived from distance to and time since the fire) to select the best match [50]. This resulted in 278 and 32 fire-causing strikes for the 2002–2011 and 2012–2018 training periods, respectively.

Fire-causing lightning strikes were assigned a value of 1. Non-fire-causing strikes were sampled without replacement from the same temporal windows as the fire-causing strikes, but not spatially near to fire ignitions. Non-fire-causing strikes were assigned a value of 0. This approach preserves the temporal structure of lightning activity and fire weather conditions while preventing spatial autocorrelation between 1 and 0 outcomes.

To mitigate zero inflation, we sample ten non-fire-causing strikes for every fire-causing strike. This strategy improves model generalization by providing sufficient contrast between ignition and non-ignition cases while avoiding the strong class imbalance characteristics of the raw lightning dataset. We also tested a ratio of five 0 outcomes to each 1 outcome and compared sampling methods using model fit (Appendix A). While the 1:5 sampling ratio provided small improvements to class discrimination, models trained on the larger dataset (sampling ratio of 1:10) were consistently better calibrated.

Each lightning strike was associated with climate, fuel, and land cover data (Figure 1c–l) based on date and proximity to the nearest ERA5 gridcell. The data were split into train (80%) and test (20%) sets and predictor variables were z-score standardized. We tested the performance of all combinations of climate, fuel moisture, and landcover predictors (i.e., additive effects of just one predictor grouping, as well as combining predictor groupings) and compared performance between the two training periods (2002–2011 and 2012–2018).

Ignition efficiency was modeled as the probability that an individual lightning strike results in a wildfire ignition. Let $z_j\in \{0,1\}$ indicate whether lightning strike j ignited a fire and $x_j^{\top }$ denote a vector of standardized daily predictors, including near-surface climate, fuel moisture indices, and vegetation cover. Ignition outcomes were modeled using logistic regression with a Bernoulli likelihood:

$$\begin{gathered} z_j|p_j ~ Bernoulli(p_j), \\ logit\left(p_j\right)={\eta }_j={\beta }_0+x_j^{\top }\beta , \\ {p}_j=P_{ignite, j}={\displaystyle \frac{1}{1+exp(-{\eta }_j)}}, \end{gathered}$$

(5)

where $p_j=P_{ignite}$ is the probability of ignition for strike j. We used two approaches to fit ${\beta }_0$ and $\beta$: a ridge-penalized logistic regression (also applied by Hessilt et al. [12]) and a Bayesian logistic regression.

Ridge-Penalized Logistic Regression

Hessilt et al. [12] use a ridge-penalized logistic regression to account for correlation between predictor variables (see Appendix C.2), which often leads to unstable logistic regression. This approach also handles extreme class imbalance and focuses on predictive performance rather than variable selection. Here, we apply the same model. ${\beta }_0$ and $\beta$ are estimated by minimizing the ridge-penalized loss function, $l_{ridge}$:

$$l_{ridge}\left({\beta }_0,\beta \right)= -\sum _{j=1}^n\left[z_{j}log\right(p_j)+(1-z_j) log(1-p_j\left)\right]+\lambda \sum _{\kappa =1}^{\kappa }{\beta }_{\kappa }^2.$$

(6)

$\lambda \ge$0 is the ridge penalty parameter, and $\kappa$ is the number of predictors. The intercept ${\beta }_0$ is not penalized. The coefficient estimates are found as

$$\begin{gathered} \beta ~Normal\left(0,{\tau }^2\right) \\ \lambda ={\displaystyle \frac{1}{2{\tau }^2}}, \end{gathered}$$

(7)

where $\tau$ is the prior standard deviation of each ${\beta }_{\kappa }$. This model was fitted using the glmnet package version 4.1.9 in R with 5-fold cross validation [51].

Bayesian Logistic Regression

The Bayesian logistic regression replaces the penalty with explicit priors centered at zero for both ${\beta }_0$ and $\beta$, and estimates the full posterior distribution of these parameters:

$$\begin{gathered} {\beta }_0 ~ Normal\left(0, {\sigma }_{{\beta }_0}^2\right), \\ \beta ~ Normal(0, {\sigma }_{\beta }^2), \end{gathered}$$

(8)

with posterior distribution

$$p({\beta }_0, \beta |z)\propto \prod _{j=1}^{n}Bernoulli\left(p_j\right)p\left({\beta }_0\right)p\left(\beta \right),$$

(9)

where

$$p_j={logit}^{-1}({\beta }_0+x_j^{\top }\beta ).$$

(10)

Posterior inference was performed using four Hamiltonian Monte Carlo chains in R’s rstan package version 2.21.8 [49].

2.2.3. Fire Probability, P_fire

After selecting the best $r_{strike}$ and $P_{ignite}$ models using performance metrics described in Section 2.3 and

Table 1. Metrics used to test the performance of $r_{strike}$, $P_{ignite}$, and $P_{fire}$.

Metric	Test	Value Interpretation
Lightning Strike Rate, rstrike
nRMSE	Magnitude of prediction error relative to observed mean	0 = perfect agreement; larger values = increasing error
Pearson Correlation	Linear co-variation between observed and predicted strike rates	1 = perfect agreement; 0 = no relationship; negative = inverse relationship
S-score	Distributional agreement; similarity of predicted and observed distributions	1 = identical distributions; 0 = no overlap
Spatial Correlation	Agreement in spatial patterns of lightning activity	1 = strong correspondence in geographic patterns; 0 = weak spatial agreement
Anomaly Correlation Coefficient	Agreement in interannual anomalies	1 = perfect, 0 = no skill, negative = inverse relationship
nRMSE of anomalies	Magnitude of error in interannual variability	values < 1 = anomaly errors smaller than observed variability; values > 1 = poor reproduction of interannual variability
Ignition Efficiency, Pignite
ROC AUC	Ranking skill; ability to discriminate ignitions from non-ignitions	1 = perfect discrimination; 0.5 = no skill
PR AUC	Rare-event skill; discrimination under class imbalance.	higher values = better identification of true ignitions; baseline = event prevalence
Brier Score	Overall probabilistic accuracy; mean squared error of probabilistic predictions	Lower values = better performance; 0 = perfect
Calibration Intercept	Systematic bias in predicted probabilities	0 = no bias; positive values = underprediction; negative values = overprediction
Calibration Slope	Calibration of predictions	1 = perfect calibration; slope < 1 = overconfidence, slope > 1 = underconfidence
Fire Probability, Pfire
ROC AUC	Discrimination	1 = perfect; 0.5 = no skill
PR AUC	Rare-event discrimination	higher values = better identification of true ignitions; baseline = event prevalence

, we combine these components to estimate the total probability of lightning-initiated fire, P_fire. Because our strike rate model outputs an expected rate, we do not simulate integer strike counts. Instead, let $r_{strike,m}$ be the modeled monthly strike rate (strikes/km²/month for June, July, and August) and $D_m$ be the number of days in that month m. We convert this to a daily strike rate:

$$r_{strike,d}={\displaystyle \frac{r_{strike,m}}{D_m}}$$

(11)

The probability of at least one ignition in a given gridcell day, $I_d$, is

$$P_{fire,d}=P(I_d\ge 1)=1-{(1-P_{ignite,d})}^{r_{strike},d}$$

(12)

As the probability of a positive outcome (a single strike results in fire ignition; $P_{ignite,d}$) increases and the number of independent events (rate of lightning strikes; $r_{strike,d}$) increase, then so does the overall probability of a positive outcome (at least one strike ignites a fire; $P_{fire,d}$). Alternatively, when climate conditions are conducive to a high $P_{ignite}$, but low $r_{strike}$ (or vice versa), the overall probability of a fire ignition remains low.

Although the Bernoulli model usually assumes discrete trials, using a continuous rate ($r_{strike,d}$) as the exponent is equivalent to a Poisson-derived formulation under rare-event approximation: $1-exp(-r_{strike,d}·P_{ignite,d})$ [52]. This approximation holds when the expected number of ignition events per gridcell-day (the product of strike rate and ignition probability) is very small. Given the observed fire prevalence of ≈1.55 × 10⁻⁵ (fire days divided by total days) during 2002–2011, this condition is satisfied.

2.2.4. Random Forest Models

Prior to model fitting, we used random forest models to assess the relative importance of candidate predictor variables to both lightning strike rate and ignition efficiency. Random forests evaluate variable importance by measuring the increase in prediction error when predictor values are permuted (%IncMSE), and the reduction in node impurity attributed to splits on each variable (IncNodePurity). These analyses were conducted in R using the random Forest package version 4.7.1.1 [53] and are described in more detail in Appendix C. Results informed predictor selection and are reported alongside model comparisons in Section 3.

2.3. Performance Metrics

Given that the target outcome differs for each model, we apply different performance metrics for $r_{strike}$, $P_{ignite}$, and $P_{fire}$. Metrics include for $r_{strike}$: normalized RMSE (nRMSE), Pearson correlation, S-score [54], spatial correlation, anomaly correlation coefficient (ACC), and nRMSE of anomalies; for $P_{ignite}$: area under the receiver operating characteristic curve (ROC AUC), area under the precision-recall curve (PR AUC), Brier score, and calibration regression intercept and slope; and for $P_{fire}$: ROC AUC and PR AUC. We also tested the model’s ability to reproduce observed fire counts by comparing ∑$P_{fire}$ with the total number of observed fires in the simulation period. Appendix B provides in-depth descriptions of these metrics. For brief metric descriptions and interpretations, refer to Table 1.

The two training periods differ substantially in both sample size (2002–2011: n = 1112, 2012–2018, n = 128) and lightning detection accuracy, which introduce a trade-off between data quantity and quality. To quantify the uncertainty associated with each model’s performance, we calculated 95% bootstrap confidence intervals for all evaluation metrics. For each model, we generated 1000 bootstrap samples by resampling the test dataset with replacement, each sample having the same size as the original test set (20%). Performance metrics were recalculated for all bootstrap samples and confidence intervals were obtained using the 2.5th and 97.5th percentiles of the resulting bootstrap distribution. This approach provides a non-parametric estimate of uncertainty that reflects differences in sample size and data accuracy across training periods.

Performance metrics for $r_{strike}$ and $P_{ignite}$ were calculated from test data. To calculate metrics for P_fire, observations of fire point locations were converted to binary events, so that each day and gridcell (0.25° × 0.25°, matching the resolution of the training data) contains either a 1 (at least one lightning-caused fire occurred) or 0 (no lightning-caused fire occurred). This allows predicted probabilities to be compared with binary observations.

We also tested out-of-period performance of $r_{strike}$ and $P_{ignite}$ models selected for implementation into the $P_{fire}$ model based on in-period performance by applying parameters estimated from one period to the alternate period (i.e., 2002–2011 parameters → 2012–2018 data and 2012–2018 parameters → 2002–2011 data).

2.4. Use of Artificial Intelligence

OpenAI’s ChatGPT (GPT-4o; March 2026) was used during development of R and Python code. All code and data output was reviewed for quality and accuracy. We take full responsibility for the content of this manuscript.

3. Results

3.1. Lightning Strike Rate, r_strike

We applied a random forest [53] to understand the relative importance of each near-surface climate variable in predicting lightning strike rate for the 2002–2011 and 2012–2018 periods (Appendix C.1). Some variables, particularly relative humidity, consistently rank high in importance, although the relative importance of individual variables differed somewhat between metrics and time periods. Because all variables contributed some predictive value, we chose to include all six as predictors for our lightning prediction models. Using the six metrics described in Table 1, we compare model families (Linear, Gamma GLMs, and Gamma Bayesian), predictors (CAPE-based vs. near-surface), and training period (2002–2011 vs. 2012–2018) in Figure 3.

3.1.1. Modeling Approach and Predictor Groups

Linear models exhibited low nRMSE (0.84–0.94), moderate correlation (0.33–0.55), relatively strong spatial correlation (0.35–0.62) and successfully captured interannual variability during the longer training period (ACC = 0.61–0.79; nRMSE of anomalies = 0.54–0.79). However, these models predict occasional negative strike rates, which are physically unrealistic, and were, therefore, excluded from subsequent fire calculations.

Gamma GLMs showed the best overall balance among models that enforce non-negative strike rates. For CAPE-based predictors, Gamma GLMs performed poorly across most metrics (nRMSE = 2.14–13.31; correlation = 0.12–0.32; spatial correlation = 0.13–0.36; nRMSE of anomalies = 1.02–4.50) but achieved high S-scores (S-scores = 0.99), indicating that they reproduce the marginal distribution of lightning despite failing to capture its magnitude, spatial pattern, or interannual variability. In contrast, near-surface Gamma GLMs showed substantially improved fit across metrics (nRMSE = 0.98–1.02; correlation = 0.35–0.49; spatial correlation = 0.42–0.54; nRMSE of anomalies = 0.54–1.20) but had lower S-scores (0.61–0.96). This suggests that the model captures the geographic pattern and distributional shape of lightning reasonably well (spatial correlation and S-score), but not the precise magnitude (correlation).

Bayesian Gamma models did not consistently outperform the Gamma GLMs. These models degraded correlation (−0.02–0.19) and spatial correlation (−0.01–0.22) and showed mixed anomaly performance (ACC = −0.82–0.80; nRMSE of anomalies = 0.56–1.88), though they captured the observed frequency distribution well (S-score = 0.95–0.99).

3.1.2. Training Period

A comparison between training periods demonstrates that models trained on the longer 2002–2011 period (circles, Figure 3) generally outperformed those fit to 2012–2018 (triangles, Figure 3), despite lower lightning detection accuracy prior to 2012. For example, the near-surface Gamma GLM achieved higher correlation (0.49 vs. 0.35), higher spatial correlation (0.54 vs. 0.42), and better anomaly magnitudes (nRMSE of anomalies = 0.54 vs. 1.20) when trained on the earlier period, while maintaining similar nRMSE (0.98 vs. 1.02). Bootstrapped confidence intervals confirmed that nRMSE differences across periods were not statistically significant, indicating that additional years of climate variability outweighed improvements in detection accuracy for model generalization.

3.1.3. Model Selection

These results demonstrate that near-surface predictors, when used in combination, are a viable substitute for CAPE-based models, consistent with findings from Uden et al. [30]. For the intended paleo application, stronger spatial agreement and interannual anomaly tracking seen in the Gamma GLMs are more relevant, as regime–scale comparisons of relative fire activity are more robust than absolute strike rate estimates at the daily or gridcell level. Based on these findings, we selected the near-surface Gamma GLM approach trained on 2002–2011 data as the $r_{strike}$ component of $P_{fire}$.

3.1.4. Cross-Validation Across Time Periods

To check for temporal transferability of the near-surface Gamma GLM (selected for application in $P_{fire}$), we applied parameters estimated from each period to data from the alternate period and compared their performance (Figure 3a–f, top facets) with in-period predictions (Figure 3a–f, bottom facets). The 2002–2011 parameters applied to 2012–2018 data perform comparably or better than the 2012–2018 in-period predictions according to most metrics (nRMSE: from 1.02 to 0.96; spatial correlation: from 0.42 to 0.68; S-score: from 0.93 to 0.94). ACC shows the most notable degradation (worsening from −0.1 to −0.15), indicating a decline in the model’s ability to track year-to-year changes in lightning (though the magnitude of anomalies is better captured according to nRMSE of anomalies).

The 2012–2018 parameters applied to 2002–2011 data show improvements to spatial correlation, ACC, and nRMSE of anomalies, though nRMSE (from 0.98 to 1.27) and S-score (from 0.96 to 0.86) degradation indicate the 2012–2018 parameters are less robust when applied outside their training period. Given these results, while neither near-surface Gamma GLM shows perfect out-of-period performance, the forward cross-period transfer confirms that the 2002–2011 parameters generalize well to new data, supporting their selection for paleo applications.

3.2. Ignition Efficiency, P_ignite

We again applied a random forest to assess variable importance, this time in predicting ignition efficiency (Appendix C.2). Fuel moisture indices, temperature, relative humidity, and precipitation hold the greatest importance, while wind and vegetation show less contribution to ignition prediction. Overall, the earlier 2002–2011 period displayed better importance values than the 2012–2018 period. We then evaluated two modeling approaches to predicting lightning ignition efficiency (glmnet ridge regression vs. a Bayesian logistic model), seven predictor groupings (climate, fuel moisture, vegetation, and their combinations), and two training periods (2002–2011 vs. 2012–2018). Model performance was assessed using five metrics described in Table 1 and compared in Figure 4.

The event prevalence is used as a point of comparison for PR AUC, representing a model with no skill in ranking ignition probability (Figure 4b). Because lightning-caused fires are extremely rare events, the baseline event prevalence before subsampling zero-outcome lightning strikes (lightning-caused fire count/total lightning count) is very low, corresponding to 1.55 × 10⁻⁴ for 2002–2011 and 1.89 × 10⁻⁵ for 2012–2018. When interpreting PR AUC, absolute values are not directly comparable to real-world prevalence. Instead, relative comparisons across models are used to assess performance.

3.2.1. Predictor Groups

Models which used only a single-predictor group (i.e., just climate, fuel indices, or land cover) showed clear differences in skill. Climate-only models performed best among single group predictors (ROC AUC = 0.66–0.85). Vegetation-only models performed worst (ROC AUC = 0.50–0.61; PR AUC = 0.28–0.72; Brier = 0.18–0.28), although all models substantially exceed the PR AUC no-skill baseline given the rarity of ignition events. Fuel moisture-only models performed moderately well (ROC AUC = 0.61–0.66), indicating some ability to rank ignition risk. However, for the glmnet fuel-only and vegetation-only models, calibration slopes could not be estimated (Figure 4e). This occurred because ridge penalization shrank all regression coefficients toward zero, yielding nearly constant predicted ignition probabilities (ROC AUC = 0.5; no skill). With all model predictions collapsing to similar values, there is no variation in predicted probabilities against which to assess whether the model is well-scaled, confirming that these models have no meaningful discriminative skill. Combining predictor groups improved performance: models using climate + vegetation or all predictors consistently produced higher discrimination (ROC AUC = 0.67–0.71 for 2002–2011 and 0.58–0.76 for 2012–2018) and lower Brier scores (=0.17–0.24), indicating that ignition efficiency depends on both climate and fuel related conditions.

3.2.2. Modeling Approach

Model families differed in calibration. Bayesian models exhibited the best-calibrated probability estimates during 2002–2011, with calibration intercepts near zero (−0.16–+0.36; 0 = no bias) and calibration slopes close to one (0.88–1.33; 1 = perfect calibration), indicating that predicted probabilities were neither systematically too high nor too low, and scaled proportionally to observed ignition rates. That is, ignition probabilities produced by the Bayesian models are well-scaled to observed fire. In contrast, glmnet models showed slopes farther from ideal (3.7–8.2) and larger intercepts (2.8–7.7) during the same period, consistent with overconfident predictions. The 2012–2018 glmnet models achieved high discrimination (ROC AUC = 0.60–0.85; PR AUC = 0.70–0.90) but displayed highly unstable calibration parameters with extremely wide bootstrap confidence intervals, reflecting small sample sizes and class imbalance. Bayesian models during 2012–2018 were more stable but showed reduced discrimination relative to the earlier period (ROC AUC = 0.39–0.74; PR AUC = 0.63–0.81).

3.2.3. Training Period

Training on the longer 2002–2011 period consistently produced ignition probabilities that more closely reflect observed ignition rates and reduced uncertainty across both model families. The shorter 2012–2018 period, despite improved lightning detection accuracy, amplified data scarcity and class imbalance, producing higher discrimination for glmnet but at the cost of unstable calibration and wider confidence intervals. In fact, for several glmnet models trained on the 2012–2018 period (e.g., climate-only and climate + vegetation), bootstrap confidence intervals for calibration slope and intercept were extremely wide (upper limits > 100), reflecting quasi-complete separation in some bootstrap resamples rather than meaningful parameter values. This indicates that, under the limited sample size and strong class imbalance of the later period, small changes in the data lead to large changes in the calibration fit, and the resulting probabilities are not numerically stable. Statistical stability in these ignition models benefits more from sample size and climate variability than from improvements in lightning detection accuracy alone.

3.2.4. Model Selection

Model choice depends on whether ignition probabilities are used for ranking or for physical interpretation. If well-calibrated ignition probabilities are required for downstream modeling (e.g., fire probability), Bayesian models trained on 2002–2011 offer the best balance of discrimination, calibration (among in-period predictions), and stability. If only relative risk ranking is needed, glmnet models from 2012–2018 achieve the highest ROC AUC values but produce unreliable probability estimates. Vegetation-only and fuel + vegetation models show limited discriminative power (results that are consistent with Hessilt et al. [12]).

However, vegetation in combination with climate and fuel does not degrade model performance and may allow spatial transferability across ecoregions. We therefore chose to retain vegetation predictors, prioritizing generalizability over maximum parsimony. Given these results, we selected the Bayesian logistic regression trained on climate, fuel moisture, and vegetation cover from the 2002–2011 period.

3.2.5. Cross-Validation Across Time Periods

We again check the transferability of parameter estimates by applying models from each period to data from the alternate period (top vs. bottom facets of Figure 4a–e), comparing the earlier and later Bayesian logistic regression trained on climate, fuel moisture, and vegetation cover. Neither model transfers perfectly: both out-of-period calibration slopes (0.33 and 0.26) are far from ideal (1), indicating overconfident predictions in both cases. A key difference is in Brier score: the 2002–2011 parameters applied to 2012–2018 data produce a Brier score of 0.20, which is an improvement over the 2012–2018 in-period Brier score of 0.29. The reverse transfer substantially worsens Brier score (2012–2018 out-of-period: 0.34 vs. in-period: 0.17; lower values indicate better prediction accuracy).

ROC AUC shows a different pattern. The 2002–2011 out-of-period predictions produce lower ROC AUC values than the in-period model, with wide confidence intervals reflecting the limited sample size of the 2012–2018 data (consistent with findings in Section 3.2.3). For the intended paleo application, we chose to prioritize accurate probability estimates (Brier score) over discrimination ranking (ROC AUC). With this in mind, we maintained our selection of the 2002–2011 parameter estimates for application in $P_{fire}$.

3.3. Fire Probability, P_fire

Based on results from Section 3.1 and Section 3.2, $P_{fire}$ was estimated using the Gamma GLM lightning model and Bayesian ignition efficiency model, both trained on the 2002–2011 period using all predictors. Figure 5a compares predicted probabilities with observed frequencies of lightning-caused fire using quantile-based bins. At very low probabilities, predicted and observed frequencies fall slightly below the 1:1 line, indicating that the model is calibrated to the background risk regime with slight overprediction. As predicted probabilities increase, however, overestimation of the likelihood of fire progressively increases. In the highest probability bin, predicted probabilities sum to ~262 fires, but only 164 fires are observed. This indicates that the model captures the relative ranking of fire risk (good discrimination) but inflates probabilities in high-risk conditions (imperfect calibration).

The ROC curve (Figure 5b) indicates strong discrimination skill (ROC AUC = 0.894, with ROC AUC = 1 indicating a perfect model); days with lightning-caused fires are consistently assigned higher probabilities than non-fire days. Because lightning-caused fires are rare events (272 fires across ~4 million gridcell-days), the baseline PR AUC (Figure 5c, dashed horizontal line) equals the event prevalence (≈0.000067). The model achieved a PR AUC of 0.00041, approximately 2.6 times higher than the baseline rate, indicating improvement over random classification. Furthermore, the PR line (solid line) lies above the baseline prevalence rate (dashed horizontal line), demonstrating that high-probability predictions occur when observed fires occur. Together, these results indicate that the model effectively ranks fire risk (when and where fire risk is elevated), though it overestimates the magnitude of this risk.

These results are reflected in Figure 6, which compares deviations of annual observations (yellow) and predictions (blue) from the 10-year climatology (dotted horizontal line). Across most years, model predictions correctly track the direction of observed deviations from the mean. However, the model fails to capture the magnitude of anomalies.

Figure 7 provides a spatial comparison of predictions and observations by mapping strike rates, lightning ignition efficiency, and fire occurrence, as well as lightning and fire probability prediction bias. Predicted lightning strike rates (Figure 7a) correspond well to observations (Figure 7d), with coastal areas experiencing strike rates close to 0 strikes/month/km² and rates increasing towards the interior of Alaska. However, the lightning model over-estimates strike rates in certain regions, for example assigning average values of up to 0.1 strikes/month/km² in east-central Alaska, where only 0.05–0.075 strikes/month/km² were observed (Figure 7a,d). Figure 7g shows the spatial distribution of lightning prediction bias. Bias appears greatest in Eastern interior Alaska. Overprediction (orange gridcells) occurs in low-lying relatively flat regions, though the higher-elevation Copper River Basin in southern Alaska also sees overprediction. Bias switches from over to underprediction (purple gridcells) where elevation starts to increase. Low bias occurs in coastal regions where observed lightning is infrequent.

The spatial distribution of ignition efficiency predictions is compared with observations by dividing lightning-caused fire counts by total lightning counts for each gridcell. Though predicted probabilities of lightning-caused ignitions differ from observations by one order of magnitude, this is due to the large class imbalance in the raw data, which contains 1,882,857 lightning strikes and just 278 fires. To mitigate zero inflation, we subsample 0-outcome strikes, allowing the model to successfully separate ignitions from non-ignitions. In Figure 7b, predictions capture the relative spatial distribution of observations seen in Figure 7e but overestimate absolute values. A per-gridcell bias metric is not computed for ignition efficiency because the observed ignition rate (fires/strikes) is not directly comparable to the model’s ignition efficiency estimates.

Despite this, mean fire probability predictions (Figure 7c) correspond well to observed rates (Figure 7f; calculated for each gridcell as the number of fire days/total days), with areas of high predicted fire risk occurring where fires were more frequently observed. Figure 7h displays locations of under (purple gridcells) and overprediction (orange gridcells) of fire probability. A comparison with elevation indicates that overprediction occurs in flat, low-lying regions of Eastern interior Alaska, while underprediction occurs in flatter regions of Western interior Alaska. Spatial patterns in fire probability bias correspond to those of the lightning model.

4. Discussion

This study aimed to develop a lightning-caused fire framework that (1) decomposes fire occurrence into lightning frequency and ignition efficiency and (2) relies exclusively on predictor variables available in paleoclimate reconstructions. Our results demonstrate that both objectives were achieved. Near-surface climate predictors reproduced spatial and temporal variability in lightning strike rates with comparable skill to CAPE-based models, and Bayesian ignition models provided well-calibrated ignition probabilities. When combined in a Bernoulli framework, these components effectively ranked fire risk across Alaska while preserving physical interpretability. All predictor variables used here (near-surface climate, fuel moisture indices, and vegetation cover) are available in paleo datasets, making these methods transferable into the past.

4.1. Process Separation and Interpretability

A key contribution of our methods is the explicit separation of lightning occurrence and ignition efficiency, which preserves interpretability by allowing for climate sensitivities to differ across processes. Figure 7 illustrates this spatially: regions of high ignition efficiency in northern Alaska do not translate to high fire probability due to low strike rates, while the interior, where high strike rates and ignition efficiencies coincide, has higher predictions of fire probability. This compound structure constrains unrealistic fire occurrence in regions lacking ignition sources. With the exception of Hessilt et al. [12], current lightning–fire parameterizations do not make this separation [11,31,32,33].

Figure 8 further demonstrates the interpretive value of this separation by displaying the conditional effects of temperature, relative humidity, precipitation, and wind (selected because they serve as predictors in both the lightning and ignition efficiency models) on each model component and their combined fire probability. Across all panels, fire probability (right column) peaks only where lightning (left column) and ignition efficiency (center column) are concurrently high. Temperature acts as a compound predictor: both lightning occurrence and ignition efficiency increase with warming above 10–15 °C, reflecting the role of temperature in driving convective instability and fuel drying simultaneously (Figure 8a–i). Relative humidity reveals a more complex relationship. Ignition efficiency requires the combined effects of low humidity and wind (Figure 8q), whereas lightning activity tends to occur under low humidity and warm temperatures (Figure 8a). Precipitation suppresses ignition efficiency across all variable combinations (Figure 8e,k,n), while lightning is less constrained by precipitation, requiring warm and relatively dry conditions. Wind-precipitation interactions show weak gradients across all three response variables (Figure 8m–o), suggesting limited discriminative power in isolation; the influence of these variables on fire probability is largely contingent on temperature and humidity conditions already favorable for lightning ignition. This ability to interpret how climate drives each component independently makes our modular framework suitable for integration into larger modeling systems.

4.2. Overprediction and Calibration

Overprediction in the highest-risk bins (Figure 5a) may arise from several factors, including the nature of our model framework, unresolved fine-scale variability in the training data, and assumptions about human suppression. Even small overestimates in ignition probability or strike rate will propagate in the expression $1-{(1-P_{ignite,d})}^{r_{strike},d}$; the Bernoulli model inflates fire probability under high-risk conditions. For example, Figure 7g shows that overestimates of lightning (orange gridcells) concentrate in the Yukon Flats, Tanana–Kuskokwim Lowlands, and Copper River Basin [56], likely driving inflated fire probabilities in the same regions, as seen in Figure 7h.

Furthermore, overprediction could result from the coarse resolution of the training data (0.25° grid). Under extreme conditions, ignition can fail at finer scales due to smaller scale variability in weather, fuel conditions, or vegetation structure that is not represented in the model. That is, while ignition predictions might be high, sub-gridcell variability may not necessarily result in an observed ignition.

Finally, because the framework we have developed intentionally excludes anthropogenic fire interactions, fire detection and suppression could reduce realized fire occurrence relative to climate and fuel conditions alone. In other words, the model may approximate a climate-limited ignition regime rather than the observed regime modified by humans. This was proposed by Krause et al. [57], who report overestimates of global lightning-caused fire predictions and suggest that exclusion of suppression was a contributing factor.

However, this mechanism may be limited for Alaska’s interior region. DeWilde and Chapin [8] found that approximately two-thirds of interior Alaska operates under a mostly natural fire regime with minimal suppression activity, and that large lightning-caused fires dominate burned area in these remote regions. Conversely, human-caused fires in this region tend to be smaller and shorter-lived than lightning-caused fires [10], and human influence on fire occurrence is concentrated within approximately 40 km of settlements, corresponding to ~31% of interior Alaska [9]. Since the highest predicted fire probabilities in our model arise in interior Alaska (Figure 7c), exclusion of anthropogenic interactions in our models may contribute to overestimation of fire probability in this ~31% of the region where human influence on fire is strongest (Figure 7h).

In the context of reconstructing paleo fire regimes, consequences to overprediction may vary according to the period of interest. Charcoal records from interior Alaska suggest that the 2002–2011 training period represents a period of high fire activity relative to Holocene reconstructions. Kelly et al. [13] found that recent burning in the Yukon Flats already exceeds upper fire regime limits of the past 10,000 years, and ice core-derived black carbon records analyzed by Sierra-Hernandez et al. [16] show a marked increase in Alaska’s fire severity since the 1980s.

For paleo simulations targeting periods of low fire activity due to, for example, the absence of flammable red spruce during the early Holocene (10,000–6000 cal BP), or the cool wet climate of the Little Ice Age (500 cal BP) [13], predicted fire probabilities will fall in the low-to-moderate range where our model is well calibrated (Figure 5a; points are close to the 1:1 line). In other words, instances of high-risk conditions where overprediction is concentrated will be rare or absent. However, we caution that during periods of high fire activity, such as the mid-Holocene (6000–3000 cal BP) when flammable black spruce expanded, or the warm dry climate of the Medieval Climate Anomaly (1000–500 cal BP) [13], instances of high-risk bins are more likely (Figure 5a; points fall below the 1:1 line). During these periods, when climate conditions approach or exceed the modern training range, the potential for overestimates of fire risk will increase.

4.3. Training Period Selection

Our comparison across training periods reveals a tradeoff between detection accuracy and sample size relevant to paleo applications. Despite the lower detection accuracy of the 2002–2011 training period, models trained on these data achieved comparable performance across most metrics (Figure 3 and Figure 4). Detection errors in the earlier period introduce noise to the outcome variable. However, despite improved detection, the shorter 2012–2018 period provided insufficient climate variability and fire events for stable parameterization, resulting in wide confidence intervals. For paleo applications, where the goal is to extrapolate from climate states outside the modern range, a training period that captures greater climate variability is more robust.

4.4. Predictor Selection

Lightning polarity has been used in past efforts to estimate ignition efficiency, with the assumption that positive strikes, which provide a longer current, have a greater potential for ignition [58]. For example, the Wildland Fire Assessment System (WFAS) is used by the U.S. Forest Service to determine ignition efficiency and includes polarity. However, other ignition efficiency models applied to Alaska [12] and the contiguous United States [3,59] do not find an association with polarity. Given these contradictions, we chose to exclude this variable from our models.

Elevation was not included as a predictor in our models. Prior studies in Alaska have documented relationships between elevation and lightning strike density, specifying that this relationship is the result of elevation’s influence on convective activity [60,61]. Figure 7g displays the relationship between elevation and lightning prediction bias. Interior Alaska shows a clear pattern: at higher elevations, where observed lightning is frequent, the model underpredicts lightning (purple gridcells) whereas in low-lying flat regions, where observed lightning is infrequent, the model overpredicts (orange gridcells).

This pattern becomes less clear towards coastal regions: bias decreases and over vs. underprediction occurs at both low and high elevation. There appear to be different convective mechanisms occurring here, potentially resulting from coastal proximity where shallower cloud tops and weaker updrafts diminish lightning activity [22,62]. Dissing and Verbyla [61] found frequent lightning occurrences in both mountainous terrain and flat river valleys and Peterson et al. [31] saw differences in fire and lightning activity between Eastern and Western Alaska. This suggests that elevation’s influence on lightning is mediated by local convective conditions rather than acting as a direct predictor.

In the context of ignition efficiency, Hessilt et al. [12] found that elevation and other landscape characteristics were outperformed by fire weather predictors according to ROC AUC values and Calef et al. [9] reported that elevation was a minor driver of lightning-caused fire probability. Still, future work could test whether including elevation improves the performance of our models. Fire probability bias shown in Figure 7h corresponds to locations of lightning prediction bias in Figure 7g. Though not surprising given the pipeline structure of our framework, addressing bias in the lightning model will likely diminish errors in fire prediction.

4.5. Model Integration

Postglacial charcoal analyses reveal that vegetation composition has played an important role in wildfire regimes of boreal Alaska [13,15,16,17]. For example, Kelly et al. [13] propose that warm, dry conditions during the Medieval Climate Anomaly (1000–500 cal BP) drove a rise in fire severity in the Yukon Flats region of interior Alaska, promoting recruitment of early successional, less flammable deciduous species. This vegetation feedback likely limited fire frequency despite climatic conditions conducive to elevated ignition probability. In contrast, the expansion of flammable black spruce during the mid-Holocene (6000–3000 cal BP) drove a sustained rise in fire frequency [13]. These contrasting feedbacks demonstrate that vegetation-fire dynamics are inconsistent across time scales and reinforce the value of process-based models.

Dynamic global vegetation models (DGVMs) are designed to simulate these climate-fire-vegetation feedbacks. However, existing DGVM fire modules such as JULES-INFERNO [63,64], SPITFIRE [65], BLAZE [66], CLM fire module [67], and CTEM fire module [68,69], link fire activity to factors like human population density, human propensity to start or suppress fires, and socioeconomic conditions (as well as lightning ignition). While appropriate for modern contexts, including those discussed in Section 4.2, these assumptions limit their application in paleo settings where human–fire interactions are poorly constrained or simply differ. The framework presented here provides a probability-based ignition mechanism that could be incorporated into DGVM fire modules to understand past climate, lightning, wildfire, and vegetation dynamics.

4.6. Limitations

The use of monthly strike rate aggregates, while necessary to reduce zero inflation common in rare-event modeling, introduces a temporal mismatch with daily ignition probabilities. With a single strike rate prediction assigned to all days within a month, within-month lightning activity is not captured, leaving ignition efficiency as the predominant driver of day-to-day variation in fire probability. For the intended paleo application, where climate reconstructions carry greater uncertainty at sub-seasonal timescales, this is an acceptable tradeoff; regime–scale fire dynamics are better constrained at annual or seasonal resolutions than at the daily time scale. In other words, paleo fire reconstructions would focus on seasonal, decadal, or centennial change in fire regimes, rather than daily forecasting. With this in mind, future work could explore retraining these models on seasonal or summer-aggregated predictors to better accommodate paleoclimate datasets with coarser temporal resolution [39]. Nonetheless, our methods provide a pathway forward for reconstructing past lightning-driven fire regimes from paleo data.

As demonstrated in Figure 5a and discussed in Section 4.2, our fire model tends to overestimate absolute fire probability under high-risk conditions. Other modeling approaches have reported a similar issue [12,57,58], suggesting that this limitation is not unique to this study. Nonetheless, absolute fire counts (${\sum P}_{fire}$), particularly those derived from high-risk conditions, should be interpreted with caution since they will likely exceed true fire counts. However, the strong discriminatory skill of the model (ROC AUC = 0.894) indicates that the relative ranking of fire risk across space and time is preserved. Once again, for paleo applications, where the goal is to reconstruct long-term trends and relative changes to fire, rather than absolute fire counts, this limitation is less consequential. We expect that the overall temporal and spatial pattern of predicted fire probability remain a reliable indicator of fire variability, even where absolute magnitudes are inflated.

5. Conclusions

We developed a probabilistic framework for modeling lightning-caused fire that (1) isolates the effects of lightning strike rate and ignition efficiency on fire probability and (2) is compatible with available paleo data. All modeling choices, including predictor selection, were guided by compatibility with paleo datasets, prioritizing long-term applicability alongside performance under modern drivers. We used near-surface climate, fuel moisture, and vegetation cover to model lightning strike rate and ignition efficiency in Alaska during 2002–2011. Near-surface climate predictors show comparable performance to CAPE in capturing lightning activity, demonstrating that paleoclimate-compatible variables adequately capture lightning activity. These models were combined in a Bernoulli trial to predict lightning-caused fire probability, effectively ranking the spatial and temporal probability (ROC AUC = 0.894), though with overestimation under high-risk conditions. This framework provides a viable pathway for simulating past lightning–fire regimes, with the potential to extend fire–climate reconstructions beyond the instrumental record and provide baselines for evaluating future fire risk.