Frontiers of Biogeography Annual precipitation drives fire occurrence across sub-humid and semi-arid ecological gradients

Fire is an integral part of semiarid to moderately humid ecosystem dynamics in North America. The biogeographical settings in which fires readily occur are affected by global processes like climate change, as well as local and regional characteristics such as terrain, proximity to human infrastructure, and vegetation structure. Increasing numbers and severity of fires today requires high-resolution and accurate predictions of fire probability. Species distribution models (SDM) allow researchers to identify environmental predictors of fire and depict the probability of fire occurrence. We applied a Maximum Entropy (Maxent) SDM to identify fire predictors and fire risk across a broad biogeographic humid to semi-arid climate gradient within the state of Texas. We used 15 years (2001-2016) of remotely sensed fire occurrence data, along with 13 biophysical variables representing climate, terrain, human activity, and landcover to generate multiple models. Annual precipitation was the primary predictor of fire occurrence, followed by elevation and landcover. After projecting fire probability onto three climate scenarios, we found moderate change in fire distribution. Humid and sub-humid areas had higher probabilities of fire occurrence while arid regions had lower probabilities under those scenarios. Overall, the linkage between fire occurrence and annual precipitation suggests that climate-driven fire probabilities will be variable under projected future climates.


Introduction
Global climate change is predicted to change fire activity across all terrestrial biomes, with fires generally increasing in number and severity (Dennison et al. 2014). In particular, fire activity in sub-humid and semi-arid drylands has increased alongside climate change and human activity over the past 50 years (Ortega et al. 2012). These trends in fire activity vary based on historical climate, vegetation type, and topography, each influencing fire-environment interactions (Parisien et al. 2012). Given projected changes in fire activity due to changes in the fire environment, there is an accompanying need to study how these changes may affect current and future fire occurrences (Morgan et al. 2001).
Sub-humid and semi-arid drylands comprise 41% of all terrestrial land area, and 34% of the world's population resides within these biomes (Maestre et al. 2012, SCBD 2013. Vegetation productivity in these regions is frequently water-limited and subjected to regular drought and fire activity (Vallejo et al. 2012). Temperature and precipitation seasonality determine effective length of the fire season, and temperature extremes hasten or slow the drying of fuel (Dale et al. 2001). Although climate warming is expected to alter climate controls on fire occurrence globally (Flannigan et al. 2009), a regional understanding of climate controls and their future shifts is necessary to prepare for alternative fire-regime futures.
Landscape characteristics and human activity are additional factors that modify the distribution of fire occurrence (Cary et al. 2006, Syphard et al. 2007, Mann et al. 2016. Landscape characteristics such as terrain (e.g., elevation, ruggedness) and landcover influence the distribution of fire because they create fine-scale mosaics of heterogeneous fire environments (Kane et al. 2015). Human activity is also well-known to influence fire distribution directly through fire ignitions and suppression and indirectly through climate and fuel modification (Archibald 2016). Modeling fire occurrence using climate and landscape characteristics will help identify important predictors of fire distribution and their relative importance to accurately modeling fire distribution across the sub-humid to dryland gradient.
One method of parsing out interactions within the fire environment is using species distribution models (SDMs) to estimate the probability of fire occurrence and quantify the relative importance of local fire environments under present and future climate scenarios Moritz 2009, Parisien et al. 2014). Using recorded fire occurrences and descriptive environmental variables, these models allow for flexible applications across different scales and regions (Parisien et al. 2012). In previous work, SDMs and similar models have been successfully used to model fire occurrence across heterogeneous landscapes with varying spatial extents (Parisien et al. 2012, Young et al. 2017. However, there is a knowledge gap regarding the drivers of fire across ecoregion transitions such as the sub-humid to semi-arid gradient. Harnessing the spatial distribution of recorded fires, we sought to assess whether the landscape and climatic controls that mediate wildfire across the sub-humid to dryland transition can be predicted.
We applied SDMs to determine the extent to which climate and landscape characteristics are capable of predicting fire occurrences across a broad biogeographical gradient that represents a sub-humid to dryland transition within the state of Texas. We ask how well climate alone predicts fire occurrence and how including landscape characteristics modify the prediction of fire occurrence. We hypothesize that climate is a major driving factor of the distribution of fire occurrence, but that including landscape characteristics will improve model performance, especially because humans now cause most fire ignitions and are responsible for fire suppression (Syphard et al. 2007, Parisien et al 2016. We project the probability of fire occurrence on multiple future climate scenarios to determine where we should anticipate changes in the distribution of fire occurrence.

Materials and Methods
We used a Maximum Entropy (MaxEnt) SDM with fifteen years of remotely sensed fire occurrence data to explore the influence of environmental characteristics in predicting fire occurrences in the south-central United States under contemporary and predicted future climate scenarios. We accomplished this by modeling the probability of fire occurrence using a suite of climate, terrain, human, and landcover descriptors at a one-km resolution across the sub-humid and semi-arid climate gradient in the U.S. state of Texas, an area encompassing nearly 696,000 km 2 . We generated fire occurrence models for each of four descriptive variable categories and compared them to a fifth comprehensive model using all variables. With these five models, we sought to identify geographic areas with different fire prediction levels and the specific variables contributing to these fire occurrences. Finally, we applied our comprehensive fire prediction model to future climate scenarios to identify how they may potentially change fire probability.

Study Area
Within the south central United States, the state of Texas encompasses a transition from humid to semi-arid environments. Precipitation ranges from more than 1,340 mm in the east to less than 360 mm annually in the west (Hijmans et al. 2005). Elevation across the transition ranges from sea level to 2667 m (USGS 2008). Texas is home to three of the ten most populous cities in the U.S. and has population densities ranging from 1007 people per sq. km in Harris County to 0.03 people per sq. km in Loving County (Census Bureau, 2010). Vegetation across the state changes along a longitudinal precipitation gradient, with pine and mixed pine-hardwood forests dominating the humid eastern portions of the state, oak-juniper savannas prevailing in the central part of the state, and mixed shrub-grasslands covering the western part of the state (McMahan et al. 1984). The cartographic boundary for the state was defined based on the 2010 Census boundary and served as the clipping and processing extent for all predictor variables.

Remotely sensed fire occurrence
Fire occurrence data were collated from the USDA Active Fire Mapping Program (https://data.fs.usda.gov/geodata/maps/active-fire.php) between the years 2001 and 2016, and includes 135,798 fires. These fire occurrences are generated from the MODIS active fire product, which detects fire in one km pixels using middle-infrared and thermal infrared brightness (Giglio et al. 2009). This platform provides daily detections across the United States. Prior to input into our model, data were clipped to retain only the points within the study area, and data from multiple years were merged into a single comprehensive dataset for the 15 year time period. It is worth noting that due to the temporal resolution of the detection platform, data do not include fires that started and were suppressed prior to the sensor flyover. Additionally, the spatial resolution of the detection algorithm inhibits our ability to identify small fires that were suppressed before they reached a measurable extent (~1km). Thus, our data do not include short or small fires. We used the kernel density function in ArcMap to visualize the density of fire occurrences as the number of fires per 1km 2 raster cell within our study area.

Environmental variables
Environmental variables were selected from a suite of climate, terrain, landcover, and human variables chosen for their hypothesized influence on fire occurrence and spread (Table 1). To remove redundancy, improve model performance, and improve interpretation of results, we conducted a correlation analysis of 21 environmental variables, reviewed those variables that were strongly correlated (r > 0.7), and retained the variables most strongly connected to fire occurrence or variables that were previously used in SDM studies of fire occurrences, so as to make direct comparisons among studies (Merow et al. 2013). Removing highly correlated variables improves interpretation of MaxEnt models and projection of models into future climate scenarios (Braunisch et al. 2013). Seven variables were removed. The remaining 13 variables were grouped into four variable sets based on their representation of climate, terrain, human, and landcover influence (categories are marked in Table 1). All environmental variables were downloaded at, or resampled to, a 1 km 2 spatial resolution and projected to NAD1983 projected coordinate system. The climate dataset was sourced from a suite of 19 biologically important climate variables provided in the WorldClim database of global weather and climate data (Hijmans et al. 2005). The climate variables are interpolated at approximately one km spatial resolution from 50 year averages of climate station data across the globe. All the climate data were downloaded at 30 arc-second resolutions and resampled to a one km spatial resolution. These data are commonly used for SDMs and are particularly well-suited for occurrence-based distribution modeling (Booth et al. 2014). After conducting the correlation analyses, we selected 5 variables that are particularly relevant to fire occurrence in our study area. Given the seasonality of drought and fire-conducive weather, measuring mean annual temperature, and the maximum temperature of the warmest month provide context for the hot conditions that regularly occur and vary across the area. Additionally, the selection of annual precipitation and the mean temperatures of the wettest and driest months provide insight into the role that precipitation and the interactions between temperature and precipitation have in predicting fire occurrence.
To determine potential effects of terrain on fire occurrence, we used a suite of terrain variables generated from a one km digital elevation model (DEM) from the US National Elevation Dataset (USGS 2008). These calculated and derived rasters included elevation, topographic roughness index, and aspect. Topographic roughness, which represents elevation difference between adjacent pixels (Riley 1999), is the squared root of the sum of squared differences in elevation between a center cell and its eight neighbors. Roughness values ranged from near zero, representing relatively smooth topography, to 1227, indicating large abrupt variations in adjacent elevation. Aspect was calculated using the DEM to determine the direction of a slope as a range between 0 and 360 degrees. Solar radiation is a measure of the insolation of a landscape in watt hours per square meter per year. This measure reflects direct sun exposure and, therefore, indirectly captures evapotranspiration and drying potential for fine fuels.
We incorporated the influence of human activity and infrastructure on fire occurrence by calculating population density (population per km 2 ), distance to the nearest city boundary (km), and the distance to the nearest interstate highway (km) These data were downloaded January 2017 as shapefile vectors and converted to rasters matching the 1km 2 spatial resolution as the rest of the data using ArcMap. Population density was extracted from the 2010 census block data (Census of Population and Housing, 2010), locations of municipal boundaries were also obtained from the USGS Municipal Boundaries shapefile, which includes all incorporated and unincorporated local communities within the state. The distances to the closest interstate and nearest community were calculated using the Euclidian Distance tool in ArcMap to identify isolated areas with large distances between municipalities or interstate highways, which may be more removed from firefighting capabilities and road access.
The influence of landcover type on fire occurrence is tied to the fuels present and the contiguity of fuel for fire spread (Littell and Gwozdz 2011). We used the 2011 National Landcover Database (Homer et al. 2015) to define the primary form of landcover present. This database contains 20 classes of landcover types including shrublands, coniferous forests, urban, and water, among others. We resampled this layer to the 1km 2 spatial resolution and clipped it to the extent of the study area. Using categorical data for landcover type allows the MaxEnt model to test the predictivity of the various discrete classes within the dataset.

Modeling the probability of fire occurrence
We modeled the probability of fire occurrence using the Maxent model (Version 3.3.3k) (Phillips et al. 2006). Maxent is an SDM, which uses values of environmental variables at occurrence points to build algorithms to predict probability of occurrence (Elith et al. 2011). Maxent uses randomly selected background points within the extent of the study area as pseudo absences to calibrate entropy algorithms. To validate a Maxent model, the occurrence data are split into training and testing datasets. The model generates response curves that are used to map the relative probability of occurrence within the area of study and outputs a list of the relative contributions of each input variable (Phillips et al. 2006).
We ran a series of models to estimate the probability of fire occurrence and projected models onto three future climate scenarios. The series included a comprehensive model with all variables and one sub-model for each variable suite: climate, terrain, human impact, and landcover. By partitioning these variable suites, we were able to identify regions of dissimilar estimation and compare the influence of each variable suite with a model that included all of the variable suites. For each model, response curves and jackknife calculations were performed to identify variable predictive ranges and to calculate variable importance. Variable response curves detail predictivity across the range of values for each environmental variable. Jackknife calculations generate independent Maxent models using only one variable and compare the predictivity of that single variable model to the overall model using all of the environmental variables. Additionally we defined 135,000 background points, which provide the Maxent model a random sample of values from within the study extent to compare the environmental data of the occurrence points to the background points for fitting model algorithms (Elith et al. 2011). We withheld a random sample of 20% of the occurrence points for model validation (i.e. for the testing dataset).

Assessing model performance
We assessed model performance using the area under the receiving operator curve (AUC) (Elith et al. 2011). This curve represents the plot of sensitivity of true positives over the specificity of false positives. AUC values range from 0.5, where the model prediction is no better than any random selection of test points, to one, which represents perfect model prediction accuracy (Phillips et al. 2006). Previous work in using SDMs for modeling fire occurrence found that models with an AUC greater than 0.6 are considered informative (Parisien and Moritz 2009). Additionally, the percent contributions of each variable are calculated to provide insight in the relative predictivity of each input into the model. To identify areas with differing model results, we calculated a series of anomaly maps -the prediction from the comprehensive model minus the prediction from the model using each variable category. The resulting raster values are zero where there is no difference between predictions, negative where the comprehensive model had a lower probability of occurrence than the other models, and positive where the comprehensive model had a higher probability of occurrence than the other models.

Modeling future climate predictions
In order to predict changing probabilities of fire occurrence given future climates, we projected our models into various future climate scenarios. We used the NOAA Geophysical Fluid Dynamics Laboratory Climate Model 3 (GFDL-CM3) (Griffies et al. 2011). This climate model has been used for previous studies in Texas, and is well suited for analyzing precipitation and temperature at the state and regional level (Rainwater 2013). We selected three climate scenarios generated from the climate model to represent the best case (RCP26), middle case (RCP45), and worst case (RCP85) emission scenarios for 2050. Each of these scenarios was clipped to the boundary of the study area and added as a projection in our Maxent model. We calculated anomaly maps between modern and future projections of probability of fire occurrence to evaluate change in probability of fire occurrence.

Density of fire occurrences
Fifteen years of remotely-sensed fire occurrences in our study area showed the distribution of fire across the semi-arid to sub-humid gradient (Figure 1). There was a high density of fire occurrences in the pine and mixed pine-hardwood forests in the sub-humid to humid east and there were relatively few remotely-sensed fire occurrences in the mixed shrubgrasslands in the semiarid west, except in a few places like the Davis Mountains hosting a cooler wetter landscape than the surrounding arid lowland desert. In the oak-juniper savannas in the central region, there were heterogeneous densities of fires ranging from regions of extremely high fire occurrence to many regions with extremely low fire occurrence. There were also a few high density fire occurrences in the thorn shrub and subtropical woodlands of the south.

Modeling fire occurrences
Modeling fire occurrences with SDMs showed that the comprehensive model had the highest AUC, and therefore predictability, followed by the climate, terrain, landcover, and human impacts models, respectively ( Table 2). The human impacts model did not reach the AUC model threshold of 0.6.
The comprehensive model predicted higher probability of fire occurrence in the central, eastern, and coastal portions of the state, as well as a particularly high area in the Davis Mountains surrounded by very low probability in the arid west (Figure 2). This is similar to the climate and terrain sub-models ( Figure 2B,C), which also predicted higher probabilities of fire occurrence across the eastern, central, and coastal regions of the state and in the Davis Mountains region in the arid west. The human impacts model showed a surprisingly high probability across most of the state and moderate levels near major roadways ( Figure 2D). The landcover model primarily predicted fire occurrence in the eastern portions of the state dominated by pine forests ( Figure 2D) and failed to identify regions of higher fire density in the central and southern portions of the state indicated by the distribution of fire density, as well as in the comprehensive, climate and terrain models.

Predictors of Fire Occurrence
Fire occurrence was best predicted in the comprehensive model by annual precipitation, accounting for 53.4% of the variation in that model (Table 2). In the climate model, annual precipitation was again the primary predictor of fire occurrence (76.1% model contribution) with the other climate variables contributing less than 23.7% to the model. The second and third highest contributing predictors of the comprehensive model, elevation and landcover respectively, were also top predictors of the terrain and landcover models. The human impact variables did not contribute meaningfully to the comprehensive model. As landcover was the only variable of the landcover model, it accounted for all variation within that model. The environmental variables with higher overall contributions to the five models indicated the range of conditions that are associated with fire occurrence. The most predictive variable, annual precipitation, indicated that fires were more likely to occur in the central semiarid to sub-humid transition zones and eastern sub-humid regions which receive over 450 mm of precipitation each year. While other climate variables did not contribute as much to the overall model prediction, fire was more probable in areas that had annual mean temperatures over 17°C and maximum temperatures of the warmest quarter under 37°C. Fire probability was also higher in dry semi-humid areas dominated by evergreen forests, deciduous forests, and shrublands ( Figure 2).

Projected changes in future climate and fire occurrence
Future climate scenarios show an increase in mean annual temperature across our study area, averaging 0.24 C, 0.28 C, and 0.33 C for the RCP 25, RCP45, and RCP85 scenarios, respectively. Annual precipitation is projected to increase by an average of 37 mm, 20 mm, and 10 mm, respectively. Climate trends show the area growing warmer and slightly wetter over time, even in the most optimistic of carbon emissions scenarios. These climate trends drive changes in predicted fire occurrence across this sub-humid to semi-arid gradient (Figure 3). Across climate scenarios, there was a consistent decline in fire probability across the central and northern part of the state, whereas the rolling plains and the high plains of the Panhandle and the gulf prairies and marshes had an increase in fire probability (Figure 3).

Discussion
In a 15-year period with 135,798 remotely-sensed fires, annual precipitation was the strongest predictor of fire occurrence across the sub-humid to semi-arid transition in our study area. This corresponds with the previous work such as that of Parisien & Moritz (2009) which found similar relationships in fire-vegetation-climate analysis. The strong influence of annual precipitation on predicted fire occurrence likely reflects a linkage between increased precipitation and increased plant growth and biomass accumulation, which dries out in the droughty summer months and becomes susceptible to ignition. The rain-drought cycle has been identified in other, similar, semi-arid and sub-humid regions (Turner et al 2008, Wang et al. 2016, and plays a critical role in where and when fires may occur. However, as our comprehensive and landcover models show, precipitation is more explanatory than vegetation type for the patterns we observed where landcover had a weaker predictive relationship with fire occurrence. This disconnect may illustrate that fuels and precipitation are both critical for fire to occur. Even when fuels accumulate, if they are not dry enough to ignite, then a fire is unlikely to occur and propagate. Under several future climate scenarios, our models showed that the increase in temperature and precipitation drive heterogeneous changes in the probability of fire occurrence. While our fire occurrence projections consistently included widespread areas of reduced fire probability and pockets of increased fire probability, they do not inform on the expected severity of future fires. Fire severity is likely to increase over time, regardless of the numbers of fires, due to the higher overall temperatures and increased drought periods (Barbero et al. 2015). Future modeling efforts should include the impacts of these scenarios on fire severity in the study region. Overall, we found that annual precipitation was a strong predictor of fire occurrence, both directly and through secondary linkages to ecological characteristics like vegetation type in subhumid and semi-arid lands. The patterns we found support similar fire-vegetation-climate cycles identified in analogous ecological regions across the globe. Projection on future climate scenarios showed a widespread decrease in the probability of fire occurrence across great parts of the grasslands, juniper and oak woodlands, and live oak and mesquite savannah, as well as in the cross timbers and prairies in the central parts of the state most closely related to the transition zone between sub-humid to semi-arid environments. Interestingly, given the uncertainties that future climate conditions bring, we found little changes to the projections of the distribution of probability of fire occurrence among the three projected climate scenarios.
In light of the climate-driven drought cycles, land management practices provide the most direct modifications to the annual precipitation-fire occurrence relationship by inducing or suppressing fire and changing the vegetative structure. This is particularly relevant in sub-humid and semi-arid rangelands that are being converted into woodlands, cropland, and other ecosystems (Twidwell et al. 2013, Bestelmeyer et al 2015, Leis et al 2017. As an example, the conversion of grass-dominated rangelands to afforested woodlands introduces a shift in fire occurrence that may become more detectable to remote observation platforms such as the MODIS active fire product (Roy et al 2008). Additionally, these changes in land-use and landcover across the Great Plains are a major driver in changes in carbon storage (Bouchard et al 2011) and decreasing soil moisture (Zou et al 2018). These vegetative shifts may therefore be managed to reduce the impact of climate shifts on fire occurrence and fire detection.
Rural landownership in Texas has been slowly trending away from traditional agricultural use and towards lifestyle and multiple-use management which generally shifts grassland ecosystems towards afforested juniper woodlands (Sorice et al 2012). As these historical rangelands are encroached upon by woody plants, thus reducing the livestock carrying capacity while incurring significant land clearing costs to the landowner (Teague 2001). Our models of future projected increase in fire probability in the south-central plains combined with this continued landowner change suggests that we should continue to focus fire management efforts and vegetation restoration in the south-central plains.
Limitations to interpreting models projected on future climate scenarios include uncertainty surrounding the controls on fire occurrence with SDM, uncertainty in future climate scenarios, no-analog future climate conditions, and not incorporating future projections of vegetation cover or human impact. Projected models of future probability of fire occurrence are best viewed as representing the expectation of fire occurrence given various future suites of climate conditions. Future modeling efforts could include more in-depth projections for climate, landcover, and human impact variables over time and how such integrated future projections will change fire probability in the region.
SDMs like Maxent provide one method of exploring patterns of disturbance in modern and future climates, and are readily applicable to a variety of regional and global problems. Using tools like Maxent to explore fire occurrence as it relates to vegetation density, fire size, and comparing predictions based on annual climate, fire season, and fire size would all make for interesting future investigations. Another important aspect for future investigation includes evaluating different types of precipitation variables. Here, we chose to focus on annual precipitation because it was highly correlated to other precipitation variables. However, one study in the humid southeastern United States found precipitation seasonality or variability is more important than annual precipitation in that high moisture environment (Lafon and Quiring 2012). Thus, future research should investigate this in the sub-humid to semi-gradient across Texas. The global significance of sub-humid and semi-arid lands makes them critical for understanding fire disturbance patterns in the modern environment and into the future. Understanding how environmental characteristics influence fire occurrence in these areas is a key component to facilitating their roles as natural ecosystem components when managing lands and to mitigating negative effects on these regions.

Data Accessibility
Data and code for this study are available via FigShare at: https://figshare.com/projects/Fire_occurrence_across_Texas_-_An_analysis_of_fire_occurrence_predictors_across_ecological_gradients/98438

Supplementary Material
The following materials are available as part of the online article at https://escholarship.org/uc/fb Table S1. Table of all climate variables considered and tested for correlation prior to running the MaxEnt model. Figure S1. Correlation matrix between bioclim variables used to select climate model inputs.