Forecasting live fuel moisture of Adenostema fasciculatum and its relationship to regional wildfire dynamics across southern California shrublands

 In seasonally dry environments, the amount of water held in
living plant tissue—live fuel moisture (LFM)—is central to vegetation
flammability. LFM-driven changes in wildfire size and frequency are
particularly important throughout southern California shrublands, which
typically produce intense, rapidly spreading wildfires. However, the
relationship between spatiotemporal variation in LFM and resulting
long-term regional patterns in wildfire size and frequency within these
shrublands is less understood. In this study, we demonstrated a novel
method for forecasting the LFM of a critical fuel component throughout
southern California chaparral, Adenostema fasciculatum (chamise) using
gridded climate data. We then leveraged these forecasts to evaluate the
historical relationships of LFM to wildfire size and frequency across
chamise-dominant California shrublands. We determined that chamise LFM is
strongly associated with fire extent, size, and frequency throughout
southern California shrublands, and that LFM–wildfire relationships
exhibit different thresholds across three distinct LFM domains.
Additionally, the cumulative burned area and number of fires increased
dramatically when LFM fell below 62%. These results demonstrate that LFM
mediates multiple aspects of regional wildfire dynamics, and can be
predicted with sufficient accuracy to capture these dynamics. Furthermore,
we identified three distinct LFM ‘domains’ that were characterized by
different frequencies of ignition and spread. These domains are broadly
consistent with the management thresholds currently used in identifying
periods of fire danger.

All live fuel moisture data used in the model calibration were drawn from
the National Fuel Moisture Database (NFMD,
https://www.wfas.net/nfmd/public/index.php (accessed on 9/1/2020)), and
these consisted of 19,639 individual observations of chamise fuel moisture
across 61 sites throughout California, spanning the years 1977 through
2017. Climate data used in this study were drawn from the California Basin
Characterization Model v8 [28], and consisted of monthly estimates of
cumulative water deficit (CWD) and actual evapotranspiration (AET)
measurements through the years 1951–2016. This dataset represents a 270 m
grid-based model of water balance calculations that incorporates not only
climate inputs (through PRISM climate data [29]) but also solar radiation,
topographic shading, and cloudiness, along with soil properties to
estimate evapotranspiration [30]. Using these monthly values, we
calculated the mean maximum temperature (TMX), mean actual
evapotranspiration (AET), mean climatic water deficit (CWD), mean
precipitation (PPT), and mean soil moisture storage (STR) at 1, 6, and 12
month periods with lags of 1, 2, 3, 4, 5, and 6 months. Fire history data
were drawn from FRAP fire perimeter data [31], which incorporate the
perimeters of all known fires from 1878 through 2017. Vegetation data used
to identify chamise vegetation in this study were drawn from both CALFIRE
FRAP FVEG data [32] and the LANDFIRE 2016 Existing Vegetation Type (EVT)
dataset [33]. Data Preparation Observations of LFM were merged with data
recording the latitude and longitude of each site and then filtered to
exclude those observations not pertaining to chamise. As an LFM below 50
can represent dead material on the sampled shrubs, observed in situ
estimates of LFM below 50% (which were exceedingly rare) were also
excluded. Because LFM within each site was often recorded at inconsistent
intervals that did not align with the monthly climate data used in this
study, and many sites incorporated observations from multiple individual
plants (the number of which also varied over time), we then calculated a
single mean LFM within each month and site in which observations were
present. In order to reduce the computational load to a manageable scale,
all climate data were rescaled to 1 km pixels through spatial averaging,
conducted using Rasterio in Python v3.7 [34]. Six-month and twelve-month
mean TMX and total PPT, AET, CWD, and STR were then extracted at monthly
timesteps using python v3.7. Predicting Live Fuel Moisture across
California The most relevant climate parameters, lags, and window
durations were identified by regressing each LFM observation against the
corresponding monthly climate parameters (including TMX, PPT, AET, CWD,
and STR) with lags of 1 to 6 months, as well as against the six-month
means of each parameter over the six months preceding each observation.
Overall relationships between chamise LFM and local climate at monthly
timescales were then modeled using a generalized additive model (GAM)
framework. To minimize computational time while allowing for nonlinear
relationships between local climate and LFM, a maximum of five smoothing
terms was allowed for each climate parameter. In order to determine the
ability of this modeling technique to predict LFM in both (a) novel
locations and (b) months not present in the training data, model
performance was assessed using multidimensional k-fold cross-validation.
All data were divided by month and year into to one of five randomly
assigned temporal groups of equal size, and all were similarly divided
into five randomly assigned spatial groups. GAM models were then
constructed iteratively, while holding out one temporal and one spatial
group as a testing data set within each iteration. The ability of these
models to successfully predict LFM at monthly timescales was evaluated by
calculating the mean Pearson correlation coefficient between the predicted
LFM at training sites and months not used in model development, and the
observed mean monthly LFM recorded at those sites and months across all
model iterations. In order to avoid unnecessary complexity within these
models and to limit the computational requirements, only parameters of
which the inclusion increased the mean Pearson correlation coefficient by
0.02 or more were excluded from the selected model. In order to
incorporate as long a wildfire series as possible, LFM was predicted
monthly from 1952 through 2017. Identifying Fires of Interest First, we
identified those fires in which chamise was likely to represent a major
component of the overall fuel by eliminating those fires in which
<50% of the burned area was predicted to consist of either Southern
California coastal scrub or dry mesic chaparral according to the FVEG land
cover dataset produced by CALFIRE-FRAP. Similarly, we eliminated all fires
in which <50% of the burned area was predicted to consist of either
mixed chaparral, chamise-redshanks chaparral, or coastal scrub according
to EVT vegetation maps. Because of concerns surrounding mismatches among
vegetation types between FVEG land cover data and EVT vegetation maps,
only those fire scars which met both of these sets of criteria were
selected for further analysis. It should be noted that these vegetation
maps were static over time and did not attempt to incorporate variation in
vegetation cover that may have occurrred across the study period or
immediately after disturbance events. However, annual assessments of
vegetation cover throughout the study period were not available. Thus,
although land cover may have fluctuated somewhat throughout the study
period and immediately after fires or other disturbance events, these data
nevertheless represented the best available data pertaining to the spatial
distribution of chamise-dominated vegetation across California. To
evaluate the relationship of chamise LFM to the mean fire size, frequency,
and cumulative area burned across southern Californian forests, it was
first necessary to measure the predicted (and observed) LFM within the
area burned during each fire. In order to summarize the predicted LFM
within each fire at the time of ignition based on the gridded LFM
estimates produced in this study, the mean predicted LFM in the month and
year in which the initial ignition occurred was calculated across the
entirety of each fire scar. The resulting data included 1818 individual
fires from the year 1952 through 2017. Identifying Critical Thresholds in
LFM and Relationship to Burned Area To evaluate the relationship between
LFM and fire, and to identify critical LFM thresholds associated with
shifts in fire behavior, we first calculated the cumulative area burned
with decreasing (simulated) LFM for all selected fire scars. As previous
studies have shown that observed thresholds in LFM–wildfire relationships
may be biased due to differences in the freequency with which different
values of LFM occur over space and time [27], we converted these LFM
values into percentile ranks based on the distribution of simulated LFM
across the duration and spatial extent of this study. By carrying out this
step, we corrected for any differences in the spatial or temporal
frequency of LFM across the study area, which might otherwise bias the
apparent relationships to cumulative area burnt. Using these percentile
LFM values, we then conducted piecewise or ‘broken stick’ regression [35]
in order to identify transition points in LFM that were associated with an
increasing burned area. After identifying thresholds in LFM–wildfire
relationships using LFM percentiles, these percentile ranks could then be
converted back into actual LFM values in order to identify the transition
points in LFM–cumulative burned area relationships. Identifying Critical
Thresholds in LFM and Relationship to Mean Fire Size In order to determine
whether the mean size of wildfires varied significantly with LFM, we
similarly conducted piecewise analyses of the relationship between LFM and
the mean size of all wildfires in which the predicted LFM (based on the
mean LFM value across the burned area of each wildfire event) fell within
a 5 percentile span (e.g., all fires in which LFM fell within the 5th to
the 9.99th percentile). By evaluating mean fire size within set percentile
ranges of LFM, this analysis eliminated any effects of differential fire
frequency across the range of LFM, and enabled us to evaluate only the
relationship of LFM to wildfire size. As with analyses of cumulative
burned area, the identified percentile thresholds in LFM–wildfire
relationships could then be converted back into actual LFM values in order
to identify the actual transition points in LFM–mean-fire-size
relationships. Identifying Critical Thresholds in LFM and Relationship to
Cumulative Number of Fires Finally, in order to determine the degree to
which low LFM was associated with a higher number of fires, and to
identify critical thresholds of LFM below which fires occurred more
frequently, we similarly conducted piecewise analyses of the relationship
between LFM percentile ranks and the cumulative number of fire events that
had occurred. As with our analyses of the cumulative area burned, these
analyses were conducted using LFM percentiles rather than raw LFM in order
to compensate for potential differences in the spatial and temporal
frequency of different ranges of LFM across the study area, and then
converted post hoc data into actual LFM data in order to identify the
actual threshold values. As percentile ranks of LFM inherently compensate
for variable frequencies of different ranges of LFM values over space and
time, the rates at which fires accumulate may be considered to be a
measure of mean fire frequency within each range of LFM.  

Software presented here was developed using python v3.7, and requires
installation the following packages: Rasterio Pandas Numpy Geopandas
Jupyter Copy Matplotlib