Relationships of climate, human activity, and fire history to spatiotemporal variation in annual fire probability across California: Source Code and Core Data

In the face of recent wildfires across the Western United States, it is
essential that we understand both the dynamics that drive the spatial
distribution of wildfire, and the major obstacles to modeling the
probability of wildfire over space and time.  However, it is well
documented that the precise relationships of local vegetation, climate,
and ignitions, and how they influence fire dynamics, may vary over space
and among local climate, vegetation, and land use regimes.  This
raises questions not only as to the nature of the potentially nonlinear
relationships between local conditions and the fire, but also the
possibility that the scale at which such models are developed may be
critical to their predictive power and to the apparent relationship of
local conditions to wildfire.  In this study we demonstrate that
both local climate – through limitations posed by fuel dryness (CWD) and
availability (AET) – and human activity – through housing density, roads,
electrical infrastructure, and agriculture, play important roles in
determining the annual probabilities of fire throughout
California.  We also document the importance of previous burn
events as potential barriers to fire in some environments, until enough
time has passed for vegetation to regenerate sufficiently to sustain
subsequent wildfires. We also demonstrate that long-term and short-term
climate variations exhibit different effects on annual fire probability,
with short-term climate variations primarily impacting fire probability
during periods of extreme climate anomaly.  Further, we show
that, when using nonlinear modeling techniques, broad-scale fire
probability models can outperform localized models at predicting annual
fire probability.  Finally, this study represents a powerful tool
for mapping local fire probability across the state of California under a
variety of historical climate regimes, which is essential to avoided
emissions modelling, carbon accounting, and hazard severity mapping for
the application of fire-resistant building codes across the state of
California.

Climate data used in this study was drawn from the California Basin
Characterization Model v8, and consists of monthly estimates of cumulative
water deficit (CWD) and actual evapotranspiration (AET) from 1951 –
2016.  This dataset represents a 270-m grid-based model of water
balance calculations that incorporates climate inputs through PRISM data
in addition to solar radiation, topographic shading, cloudiness, and soil
properties to estimate evapotranspiration.  Using these monthly
values, we calculated the 1980 – 2009 mean CWD and AET normals, as well as
mean deviations from those normals over a three-year period preceding each
year of interest. Cultivated and agricultural areas were identified using
the 2016 National Land Cover Database data, which estimated dominant land
cover throughout North America at 30-m resolution.  The
proportion of cultivated area and of water features that covered each 1-km
pixel were then calculated by resampling to 1-km scale. Mean housing
density data was drawn from the Integrated Climate and Land-Use Scenarios
(ICLUS) dataset, which provides decadal estimates of housing density
throughout the United states from 1970 - 2020.  As precise
continuous estimates of housing density were not available, housing
density within each pixel was set to the mean of its class. 
Annual values were estimated from decadal data using linear
interpolation.  Ecoregions within California (hereafter referred
to as “regions”) were delineated using CalVeg ecosystem provinces data.
Road data were drawn from 2018 TIGER layer data, and consisted of all
primary and secondary roads across California. Electrical infrastructure
data was drawn from 2020 transmission lines data.  In both cases,
the distance of nearest roads or transmission lines to each pixel were
then calculated.  Pixels which contained roads or electrical
infrastructure were assigned distances of 0 km. Fire history data was
drawn from FRAP fire perimeter data, which incorporates perimeters of all
known timber fires >10 acres (>0.04 km2), brush fires
>30 acres (>0.12 km2), and grass fires >300 acres
(>1.21 km2) from 1878 – 2017. Using this data, the presence of fire
in each 1-km pixel was classified in a binary fashion (e.g. 1 for burned,
0 for unburned) for each year of interest.  Due to computational
limits and the quantity of data involved in this study, we did not
calculate the burned area within each pixel, or distinguish pixels in
which a single fire occurred in a given year from those in which multiple
fires occurred.   This data was also used to calculate
the number of years since the most recent fire within any pixel, prior to
each year in which fire probability was projected.  Thus,
locations in which no fire was observed throughout the fire record were
treated as having gone a maximum of 100 years without a fire event for the
purposes of model construction. These pixels comprised 29% - 33% of data
annually (depending on year), and included both locations in which fire
would not be expected (such as highly xeric regions) as well as locations
in fire-prone areas in which no fire had been documented within the FRAP
fire perimeter data used in this study. 

Please refer to Readme.txt file.