Model overview
iLand is a landscape-scale forest model that simulates ecological
processes at multiple spatial and temporal resolutions in a
hierarchical framework (online documentation:
http://iland.boku.ac.at). iLand simulates tree growth and mortality of
individual trees and the interactions among them in spatially explicit
landscapes as a function of canopy light interception, radiation,
temperature, soil water, and nutrients. The model also explicitly
simulates tree regeneration processes, such as seed production
(including serotiny), dispersal, and environmental controls (such as
drought) on seedling establishment and sapling growth (Hansen et al.
2018). Both sexual reproduction (i.e., by seed) and re-sprouting are
simulated for aspen. Thus, following fire, tree regeneration is
influenced by the age of the trees that burned (determining the size
of the canopy seed bank for serotinous species), distance to the
nearest unburned seed source, soil moisture conditions in subsequent
growing seasons, and plant reproductive traits. The model has been
well tested in Greater Yellowstone.
Climate and soil are assumed spatially homogeneous within a 1-ha grid
cell, but within-cell variation in light and tree regeneration is
simulated at 2×2 m resolution based on forest structure. iLand is
forced with daily temperature, precipitation, vapor pressure deficit,
and radiation. For this application, we used gridded climate data sets
that were statistically downscaled (4-km resolution) with the
Multivariate Adaptive Constructed Analogs approach (Online:
http://www.climatologylab.org/maca.html). These included two general
circulation models (GCMs), CNRM-CM5 and GFDL-ESM2M that represent
20th-century climate well in Greater Yellowstone. For each GCM, the
first of the five runs from the Inter-Governmental Panel on Climate
Change AR5 experiment were downscaled. We also included two
representative concentration pathways (RCPs) 8.5 and 4.5, which assume
continued increases in radiative forcing to 8.5 W m2 by 2100 and
stabilization of radiative forcing at 4.5 W m2 by 2100, respectively.
Both GCMs show similar temperature trends with about 5degC of summer
warming by 2099 under RCP 8.5.
iLand dynamically simulates wildfire at 20 m x 20 m resolution in a
modeling framework designed initially for the Northern Rocky Mountains
and western Oregon. Briefly, fire is simulated based on statistical
distributions of fire occurrence and size, fuel load (including
surface litter and downed coarse wood pools, excluding live fuels and
dead canopy fuels), and drought (using the Keetch Byram drought index,
KBDI). KBDI is a cumulative daily metric of water balance for the fuel
layer that accounts for effects of both precipitation and temperature.
In the fire module, daily KBDI is summed for each simulation year, and
compared to a reference KBDI (1980 to 2016) to compute a KBDI anomaly.
Fire ignition in any given 20 m x 20 m cell that has sufficient
available fuels (≥ 0.05 kg m2 or 500 kg ha-1) is modeled based on the
20th-century fire return interval and adjusted by the KBDI anomaly so
that ignition is more likely when conditions are hot and fuels are dry
and less likely when conditions are cool and fuels are wet. Fire size
is modeled by first drawing a maximum potential fire size from a
negative exponential distribution, fit to 20th-century fires, and then
dynamically spreading the fire across the landscape using a cellular
automaton approach. Because fire size in subalpine forests is strongly
driven by aridity, we modified the distributions that determine
maximum fire size so that only fires greater than 400 ha are chosen
when KBDI anomaly is greater than 1.7 (hot-dry conditions), and only
fires less than 10 ha are chosen when KBDI anomaly is less than 1
(cool-wet conditions). A KBDI anomaly cutoff of 1.7 delineated the 5 %
driest years in the contemporary climate record (1980-2016) for the
study landscape, and the minimum fire size selected when KBDI anomaly
exceeded this threshold was determined based on historical fire
records for the region from 1970-2016. However, many other factors
(e.g., winds, flash droughts) can cause fires to become large when our
metric of aridity (KBDI anomaly) are at intermediate values. So, when
the KBDI anomaly values are between 1 and 1.7, simulated fire size is
drawn from a negative exponential distribution with a mean size of 75
ha and a maximum size of 20,000 ha. Thus, large fires can occur at
intermediate KBDI anomaly values (between 1 and 1.7), but large fires
have a lower probability of occurrence.
Once a maximum fire size is selected, fire spreads dynamically through
the landscape, with probability of spread to adjacent cells contingent
on fuel load, wind, and slope. Fuel constraints were set so that a
fire can only spread if ≥ 0.05 kg m2 (500 kg ha-1) of fuel is present
in neighboring cells, the same threshold for fire ignition. Wind is
simulated with a given speed and direction per fire event (both
randomly selected for each event from user-defined ranges). Spread
rates differ if fires burn upslope or downslope and vary with slope
angle. In the burned cells, fire severity is modeled as percent crown
kill based on fuels, KBDI anomaly, tree size, and bark thickness. To
ensure iLand could re-create 20th-century fire activity with
reasonable skill, we parameterized the model and compared the
simulated fire regime to historical fire records (1970-2016) from the
study area.
Simulation experiment
To ensure the initial simulation landscape corresponded with the
tree-species composition, forest structure, and stand age
distributions in the actual landscape, while also creating conditions
consistent with the internal model logic, we conducted a 300-yr
spin-up under historical climate (CNRM-CM5 period: 1950-2005, climate
years randomly chosen with replacement) and fire. This procedure
generated a simulation landscape similar to the actual landscape in
1989. Using the dynamic fire module, we then simulated the resulting
landscape from 1989 to 2098 while varying climate conditions (two GCMs
and RCPs, as described above) and fire management (two scenarios).
The two fire-management scenarios were designed to be generally
consistent with how fire management operates in western subalpine
forests, but they were not meant to precisely replicate past
management actions in GRTE, nor predict what will occur in this
landscape, as that is not feasible. The two management scenarios were:
(1) a managed wildfire use scenario where all fires were allowed to
burn naturally, and (2) a fire suppression scenario in which fires
that ignited when drought was moderate (KBDI anomaly less than = 1.7)
were suppressed and never grew larger than 0.04 ha, but fires that
ignited when drought was extreme (KBDI anomaly greater than 1.7)
burned unhindered. Thus, we represented effective suppression of fires
when conditions are cool and wet, as was typical of many years in the
historical record, and the inability to suppress fires when conditions
are hot and dry. We assumed that drought was the dominant factor
influencing fire suppression for the purpose of this analysis; other
variables (topography, wind speed, proximity to roads, distance to
structures) that might influence suppression effectiveness were not
represented. Because climate projections do not vary within a given
GCM and RCP but fire is stochastic in iLand (probability and location
of fire ignitions and the fire sizes and severities that result), we
simulated 20 replicates for each combination of GCM x RCP x
suppression scenario (n = 160).
Model outputs
To determine the relative importance of fire management vs. climate
change in the simulations, we analyzed the spatial and temporal
patterns of fire (number of fires, area burned, area-weighted mean
fire size, and proportion of stand-replacing fire (greater than 90% of
mature trees killed in 1-ha grid cells) and forests (forested area,
fuel loads, stand age, and dominant tree species in 1-ha grid cells)
in the different scenarios. Annual number of fires was tallied within
four size classes (less than 10 ha, 10-100 ha, 100-225 ha, and greater
than 225ha). Fire severity was calculated as the proportion of area in
stand-replacing fire (greater than 90% crown kill) within each fire
perimeter. Forested area was defined as areas with ≥ 50 trees ha-1. We
also calculated the median fuel load (coarse and fine downed wood) in
forested areas. Stand age was tallied within four classes (less than
40 yrs, 40-150 yrs, 151-250 yrs, and greater than 250 yrs). Species
dominance in each 1-ha cell was quantified by using species importance
values (IV). Importance values sum the relative abundance (number of
individuals of a species divided by number of individuals of all
species in a grid cell) and relative basal area (basal area attributed
to that species divided by total basal area in a grid cell) for each
species on a plot. Thus, species IV ranges from zero (species is not
present) to two (pure stand of the species). We then tallied the
forested cells in the simulated landscape that were dominated (i.e.,
IV greater than 1) by lodgepole pine, Douglas-fir, spruce-fir, or
aspen.
Analysis
We quantified differences in response variables among scenarios from
1989 to2017 and 2018 to 2098. We evaluated differences among scenarios
by comparing means and bootstrapped 95% confidence intervals (CIs).
However, in interpreting model results, we emphasize ecologically
meaningful differences rather than statistical ones.
model.zip file content
Model_output folder: From each output, results were condensed using R
statistical software to create a txt file of each output results.
Files of condensed iLand simulation output were processed in each
subfolder folder.
Model_excutable folder: All computation was conducted on the
University of Wisconsin high throughput cluster. Operating systems and
sofware: OS: HTCondor , iLand version: version 1.06, QT library:
5.8.0, R version: 3.4.0 Files: At the highest file level is the iLand
model executable folder which contains the model source code and the
necessary QT library to compile iLand on the cluster.
Model_run folder: Twenty replicate simulations were run with each
project folder in the model runs folder. Within the runs folder there
are 8 project folders and one shared folder. These contain materials
necessary for running iLand simulations under different GCM climate
scenarios and with and without fire suppression. In each of the
project folders, there is a gis folder which contains an environment
file that tells iLand information about each resource unit and a stand
grid that tells iLand where each resource unit is located. There is a
lip_nr folder that tells iLand how to conduct light calculations for
different types of trees. There is a scripts file which contains the
java scripts necessary to implement the stand replacing fires. There
is also shared folder that contains four folders, one for each GCM,
containing a climate database and a species parameter database. There
is also an initialization database that contains a spinup snapshot of
vegetation on the landscape which is used to initialize the
simulations.
