2.1.3 Fuel and topographic properties
We used the fractional cover maps of major fuel components across
Alaskan tundra (He et al., 2019) to represent fuel type distribution.
Here we considered three fuel components, namely woody, herbaceous and
nonvascular fuels. Four vegetation indices that are directly related to
leaf water content were adopted as estimates of fuel moisture state for
large-scale monitoring (Yebra et al., 2008), including two Normalized
Difference Infrared Indices using MODIS bands 6 and 7
(NDII6 and NDII7; Hardisky et al.,
1983), Normalized Difference Water Index (NDWI; Gao, 1996), and Global
Vegetation Moisture Index (GVMI; Ceccato et al., 2002). We computed
these indices using the MODIS 8-day surface reflectance data (MOD09A1;
Vermote et al., 2015) for our study area (Table S1). The 5m Digital
Elevation Model (DEM) data developed with airborne Interferometric
Synthetic Aperture Radar (IfSAR) data for Alaska was then used to
extract topographical features, including elevation, slope, aspect, and
roughness.
2.2 Tundra fire occurrence modeling
Five groups of influencing factors were used as independent variables
for modeling tundra fire occurrence: fuel type, fuel moisture state,
fire weather, topography, and ignition source (Figure S2). Fire weather,
ignition source (CG lightning probability), and fuel moisture state are
weather-related conditions and can change rapidly on a daily basis
throughout a short period. Although vegetation shifts and fuel type
transitions can occur from years to decades under disturbances or
climatic variability and change, the vegetation compositions and fuel
type distributions are relatively stable without substantial seasonal or
diurnal changes.
To fully understand how these dynamic weather-related variables affect
the probability of tundra fire occurrence, we developed two types of
models, referred to as “Current-day model” and “Previous-day model”.
Here we categorized the ignition source, fire weather and fuel moisture
state as “dynamic” variables considering their temporal variabilities
during fire seasons. While topographic properties and fuel type
distributions were considered as “static” variables. The two types of
models selected the dynamic variables on different dates as independent
variables. The “Current-day model” adopts the dynamic variables
simulated on the exact dates of fire occurrence, while the
“Previous-day model” uses those extracted from the dates before the
detected fire occurrence. Fire occurrence points detected in Section
2.1.1 were used to represent the presence of “Fire” events. We
randomly sampled points across the tundra regions on the same fire
ignition dates to represent “No Fire” events.
Empirical models were then developed with both the RF classification and
logistic regression algorithms to identify the key factors driving
tundra fire occurrence and quantify their impacts. Although RF
algorithms can provide relative rankings of variable importance in
predicting the dependent variable, they are limited in showing the
quantitative relationships between each independent variable and fire
occurrence probability. We therefore developed logistic regression
models as well, to quantify the impacts of environmental factors. Before
modeling, we tested the correlations of variables among the five groups
of environmental factors using Pearson’s r correlation and removed the
highly correlated ones. For both RF classification and logistic
regression models, 70% of the records were randomly selected for model
training, and the rest 30% were reserved for validation. Welch’s t-test
was also conducted to assess the differences of environmental factors
between “Fire” and “No fire” events across the study area.
3 Results
3.1 Wildfire occurrences in Arctic tundra of Alaska
Individual fire events were first identified using the MCD14ML data
between 2001 and 2019 (Figure 1). The occurrences of wildfire events
vary across space in Arctic tundra of Alaska. The majority of the fires
occurred in Southwest Alaska (~39.62%), followed by the
North Slope (~36.92%) and the Seward Peninsula
(~23.46%). A slightly increasing trend of tundra fire
occurrences was found during the study period (Figure 1 b). Temporal
variability also exists regarding fire season severity, as indicated by
the number of annual fire events. During 2001 and 2019, thirteen years
have relatively low fire events (< 20 fires per year), and
four years have a moderate fire season with 20 ~ 30 fire
events per year. An exceptionally severe fire season was detected in
2015, with 49 fire events in total. To cover a variety of fire season
severities, we sampled five seasons (2002, 2006, 2008, 2013, 2017) with
light severity, two years with moderate severity (2007, 2010), and the
year of 2015 as severe with very high fire activity for model
development (Table S2).