www.publish.csiro.au/journals/ijwf International Journal of Wildland Fire, 2003, 12, 309–322
Using simulation to map fire regimes: an evaluation of approaches,
strategies, and limitations*
Robert E. Keane
, Geoffrey J. Cary
and Russell Parsons
USDA Forest Service, Rocky Mountain Research Station, Fire Sciences Laboratory,
PO Box 8089, Missoula, MT 59807, USA.
School of Resources, Environment and Society, Australian National University, Canberra, ACT 0020, Australia.
Corresponding author. Telephone: +1 406 329 4846; fax: +1 406 329 4877; email: firstname.lastname@example.org
This paper is derived from a presentation at the conference ‘Fire and savanna landscapes in northern Australia:
regional lessons and global challenges’, Darwin, Australia, 8–9 July 2002
Abstract. Spatial depictions of fire regimes are indispensable to fire management because they portray important
characteristics of wildland fire, such as severity, intensity, and pattern, across a landscape that serves as important
reference for future treatment activities. However, spatially explicit fire regime maps are difficult and costly to create
requiring extensive expertise in fire history sampling, multivariate statistics, remotely sensed image classification,
fire behaviour and effects, fuel dynamics, landscape ecology, simulation modelling, and geographical informa-
tion systems (GIS). This paper first compares three common strategies for predicting fire regimes (classification,
empirical, and simulation) using a 51 000 ha landscape in the Selway-Bitterroot Wilderness Area of Montana, USA.
Simulation modelling is identified as the best overall strategy with respect to developing temporally deep spatial
fire patterns, but it has limitations. To illustrate these problems, we performed three simulation experiments using
the LANDSUM spatial model to determine the relative importance of (1) simulation time span; (2) fire frequency
parameters; and (3) fire size parameters on the simulation of landscape fire return interval. The model used to sim-
ulate fire regimes is also very important, so we compared two spatially explicit landscape fire succession models
(LANDSUM and FIRESCAPE) to demonstrate differences between model predictions and limitations of each on
a neutral landscape. FIRESCAPE was developed for simulating fire regimes in eucalypt forests of south-eastern
Australia. Finally, challenges for future simulation and fire regime research are presented including field data, scale,
fire regime variability, map obsolescence, and classification resolution.
Additional keywords: mapping; GIS; LANDSUM; FIRESCAPE; simulation modelling; landscape modelling.
Successful wildland fire management is partly dependent on
accurate and consistent predictions of fire regimes at multiple
spatial and temporal scales (Hardy et al. 2001). Fire regime
maps can portray historical burning characteristics, such as
severity, frequency, and pattern, of natural and human-caused
fire (Morgan et al. 2001). Using fire regime maps, landscape
fire treatments can be prioritized, designed, and scheduled
from fire frequency and severity descriptions (Heinselman
1985; Barrett and Arno 1992; Agee 1995; Brown 1995). Fire
regime maps can also be used to quantify input parame-
ters for landscape models that simulate effects of alternative
fire management strategies on landscape dynamics (Keane
*This paper was written and prepared by U.S. Government employees on official time, and therefore is in the public domain and not subject to copyright.
The use of trade or firm names in this paper is for reader information and does not imply endorsement by the U.S. Department of Agriculture of any product
et al. 2002b). Fire regime maps also provide a context for
interpreting and understanding landscape and fire ecological
interactions (Barrett and Arno 1992;Turner et al. 1997). And
last, fire regime maps can be used to stratify fire monitoring
and landscape inventory sampling (Lutes et al., in press) (see
Development of fire regime maps is a difficult and costly
task, requiring extensive expertise in fire ecology, fire his-
tory sampling, and statistical analysis (Morgan et al. 2001).
Moreover, since fire regimes are the expression of the inter-
actions between climate, fire, vegetation, and topography,
mapping them would require extensive knowledge of fire
dynamics, fuels, landscape ecology, simulation modelling,
310 R. E. Keane et al.
remote sensing, and Geographic Information Systems (GIS)
(Keane et al. 2003). This wide variety of skills is especially
important because there are several approaches, strategies
and techniques that are usually fused to create accurate and
realistic fire regime maps. The key to successful fire regime
prediction is to recognize the strengths and limitations of all
methods, data, and models and then merge the best parts into
a comprehensive prediction vehicle.
In this paper, wefocusonthe comparativelynew method of
using simulation modelling for developing fire regime maps.
In doing so we will first compare three common strategies for
predicting and mapping fire regimes: (1) classification; (2)
statistical modelling; and (3) simulation modelling to explore
the performance of simulation modelling against these other
more traditional methods. Second, we present the advantages,
disadvantages, and limitations of simulation modelling to
spatially describe fire regimes using three simulation experi-
ments where we vary the simulation time span, fire frequency
parameters, and fire size parameters. Then, we will demon-
strate how differences in design between two simulation
models can affect fire regime predictions by comparing pre-
dicted fire regimes generated from the FIRESCAPE (Cary
and Banks 1999; McCarthy and Cary 2002) and LANDSUM
(Keane et al. 1997b, 2002b) landscape simulation models
on a neutral landscape. Future research, including compila-
tion of comprehensive field databases, scale issues, inherent
variability, and improved fire regime classifications, are
Fire regimes are general descriptions of wildland fire charac-
teristics across discrete time and space bounds. Common fire
characteristics used to define fire regimes include frequency,
size, pattern, intensity, severity, type of fuel burned, and sea-
son of burn (Gill 1975, 1998; Heinselman 1981; Agee 1993).
Fire frequency is best defined at the scale of application.
Point measures, such as fire return interval and fire probabil-
ity, describe the number of fire events experienced over time
at one point on the landscape. Landscape measures of fire
rotation and fire cycle estimate the number of years it takes
to burn an area the size of the relevant landscape (Agee 1995;
Lertzman et al. 1998).The distribution of burn sizes on a land-
scape or region depends primarily on the number of large fire
events; typically only a few fires burn the majority of the area
(Yarie 1981; Strauss et al. 1989; Bessie and Johnson 1995).
Vegetation, topography, antecedent weather, and fuels will
often dictate the mosaic of burned patches within and across
fires on the landscape (Skinner and Chang 1996; Kushla and
Ripple 1997). Fire intensity describes the physical heat out-
put from a fire, whereas fire severity describes the subsequent
fire-caused damage to the biota and soils (DeBano et al.
In this paper, we use only frequency and severity to
describe fire regimes because they are most important to fire
effects and they are used in the majority of studies. The point-
based average fire return interval (years) is used to describe
frequency. Fire severity is described by three categories com-
monly adopted for northern hemisphere coniferous forests.
Non-lethal surface fires burn surface fuels at low intensities
but do not kill many overstory trees. Stand-replacement burns
kill that majority of the dominant vegetation, often trees and
shrubs (greater than 90% mortality) (Brown 1995). These
fires include both lethal surface fires and active crown fires
(Agee 1993; Brown 1995). Mixed severity burns contain ele-
ments of both non-lethal surface and stand-replacement fires
mixed in time and space and include passive crown fires,
patchy stand-replacement burns, and mixed severity under-
burns (Brown 1973, 1995; Shinneman and Baker 1997; Arno
et al. 2000). Other types of severity classes exist, such as
ground fires (i.e. smouldering fire burning extensive duff
layers), but for brevity they are not presented here.
It is the cumulative interaction of fire, vegetation, climate,
topography, and humans over time that ultimately creates a
fire regime (Crutzen and Goldammer 1993). These interac-
tions are spatially and temporally correlated; future burns
are influenced in space by the adjacency to burnable stands
and fire-resistant topographic features (e.g. lake shores, rock
outcrops) and in time by the occurrence and severity of past
climate (e.g. El Niño, drought) and disturbance events (e.g.
previous burns, insect epidemics). A change in any of these
factors will ultimately cause a change in the fire regime and,
since all four factors are constantly changing across many
scales, fire regimes are inherently dynamic. For example, cli-
mate change can affect fire regimes by modifying the weather
(Flannigan and Wagner 1991; Cary and Banks 1999), alter-
ing fire ignition patterns (i.e. lightning) (Price and Rind 1994;
Stocks et al. 1998), and increasing fuels and smoke (Keane
et al. 1997a, 1999). Exotic plants, such as cheatgrass and
spotted knapweed, have modified fire regimes as they invaded
into many arid ecosystems (Whisenant 1990). Humans have
influenced past and present fire regimes throughout the world
(Barrett and Arno 1982; Pyne 1982; Russell 1983). Native
Americans started many burns for a wide variety of reasons
including land clearing, wildlife habitat improvement, culti-
vation, defence, communication, and hunting (Gruell 1985;
Lewis 1985; Bahre 1991; Kay 1995). In parts of Australia,
Aboriginal burning was common for at least 60 000 years
(see Bradstock et al. 2002). Because of this dynamic nature
of fire, fire regimes should not be viewed as attributes or
characteristics of ecosystems or vegetation types. Fires are
landscape-level disturbances that do not follow discrete map-
ping units and are influenced by many factors besides fuels
and vegetation (Agee 1993). Attempts to predict fire regimes
solely from fuels (Olsen 1981), vegetation (Frost 1998),
or topography (Barrett and Arno 1992) have only partially
succeeded because these studies have not recognized the per-
vasiveness of fire on the landscape and the interactions of all
factors that control fire dynamics across multiple scales.
Using simulation to map fire regimes 311
Many techniques and methods have been used to pre-
dict fire regimes for both stands and landscapes (Morgan
et al. 2001; Keane et al. 2003). We have identified three
broad strategies for mapping fire regimes: (1) Classifica-
tion; (2) Statistical Modelling, and (3) Simulation Modelling.
The classification strategy involves assigning a fire regime
category to one or more categories in related classification
schemes often based on vegetation, biophysical settings, or
climate. The popular statistical analysis strategy can use the
entire suite of multivariate statistical techniques, such as
regression, ordination, general additive models, and logistic
regression, to create deterministic or stochastic fire regime
predictive models. Last, the simulation strategy uses stand
or landscape models to simulate fire events and vegetation
development (i.e. succession) over time to generate some
spatial expression of fire regime. This strategy is somewhat
new because recent advancements in computer technology
have allowed an independent spatial simulation of fire spread
coupled with weather and topography (Finney 1998).
Each of these strategies can be implemented using three
approaches: (1) Stochastic; (2) Empirical; and (3) Physical
(see Gardner et al. 1999; Turner et al. 2001 for review).
The stochastic approach uses probabilities and stochastic
functions to quantify or describe fire regime. An empirical
approach uses field data to derive deterministic relationships
to represent characteristics of a fire regime. Examples include
regression models, discriminant functions, and other multi-
variate statistical modelling. The classification strategy may
use empirical approaches such as regression trees and neu-
ral networks. The physical approach uses formulations of
the physical processes driving ecosystems and landscapes
to create fire regime descriptions. These approaches are not
mutually exclusive and, in fact, the best fire regime predictive
models are often created from a melding of approaches.
Since this paper emphasizes simulation modelling, it is
important to understand the design and components of the
landscape fire succession models that can spatially predict
fire regimes. There are usually at least four elements in a
landscape fire succession model: vegetation succession, fire
ignition, fire spread, and fire effects (see McCarthy and
Cary 2002; Keane and Finney 2003 for review). Succes-
sion is simulated using a variety of approaches such as a
stochastic Markov transition model (Acevedo et al. 1995);
a species-based vital attributes scheme (Roberts and Betz
1999); an empirical frame-based multiple pathway model
(Chew 1997; Keane et al. 1999, 2002b); a deterministic age-
since-disturbance function (Baker 1994; Li et al. 1997); a
fuel accumulation function (Cary and Banks 1999); an indi-
vidual plant gap model design (Miller and Urban 1999); or
a physical biogeochemical model (Keane et al. 1996b). Fire
ignition is usually modelled with stochastic functions based
onWeibull probability distributions (He and Mladenoff 1999;
Keane et al. 1996b) that can be linked to indices of fire
weather (Gardner et al. 1996; Li et al. 2000). Fire spread
is often simulated using cell automata, percolation, or vector
propagation based on simple topographic rules to physically
based fire behaviour models (see Gardner et al. 1999; Turner
et al. 2001 for summaries). The effects of fires are often mod-
elled using a rule-based approach, probabilistic functions, or
explicit simulations of fire damage (see Keane and Finney
2003 for summary).
Mapping strategy comparison
Fire regime maps of fire frequency and severity were created
using the three broad mapping strategies presented in the
previous section: classification, statistical analysis, and sim-
ulation modelling (Keane et al. 2003). A portion of the Lower
Selway watershed (51 761 ha), located on the western edge of
the Selway-Bitterroot Wilderness in the mountains of central
Idaho, was used as the analysis landscape. Field data used
in this comparison were taken from 64 plots located within
this landscape and collected by Keane et al. (2002a) in 1995
for an intensive ecological inventory of the area. Fire fre-
quency is described by three categories of fire return intervals
(0–40 years, 40–100 years, and 100+ years), and fire sever-
ity is defined by three general fire type categories: non-lethal
surface fire, mixed-severity fire, and stand-replacement fire.
Maps created using the classification strategy employed
an empirical approach where the rule-based terrain model of
Barrett and Arno (1992) for the greater Selway-Bitterroot
Wilderness Area was coded into a GIS to create the fre-
quency and severity maps. Discriminant analysis was used
for the statistical analysis strategy to create fire regime
maps with an empirical approach. The extensive ecologi-
cal gradient-based field dataset (Keane et al. 2002a) used
in this analysis contained over 200 variables to predict fire
regime including topography, weather, ecosystem processes
(i.e. evapotranspiration, net primary productivity simulated
from the Biome-BGC models; Thornton 1998), satellite
imagery, and soils information. Only discriminant analysis
was employed in this statistical approach because Keane
et al. (2002a) found that other more complex approaches (e.g.
general additive models, logistic regression) only marginally
increased overall map accuracy over discriminant analysis.
Last, the probabilistic-deterministic LANDSUM landscape
fire succession simulation model (Keane et al. 1997a, 2002b)
was used to generate fire regimes for the lower Selway land-
scape to demonstrate a simulation approach. Fire frequency
estimates were averaged across all pixels over all years in a
1000-year run and fire severity was computed from the modal
value simulated for each pixel.
Simulation sensitivity analysis
In the Keane et al. (2003) study, it became apparent that simu-
lation modelling was one of the best strategies for generating
fire regime maps. Yet, they found many limitations to this
312 R. E. Keane et al.
complex and demanding strategy. To address these limita-
tions, we conducted three simulation experiments to assess
the importance of various modelling parameters for generat-
ing fire regime maps using the same Lower Selway watershed.
First,we multiplied the input point-based fire frequency prob-
abilities (inverse of fire return interval) in the LANDSUM
model by 0.5, 1.5, 2.0, and 2.5 to assess the sensitivity of
fire ignition parameters in generating accurate fire regimes.
Next, to illustrate the importance of temporal scale in fire
regime descriptions, we created fire regime maps from 100,
250, 500, 1000, and 1500-year simulation runs. We attempted
to include a 10 000-year simulation in this exercise but lacked
sufficientcomputing resources and time.The average fire size
parameter in LANDSUM (see Keane et al. 1997b, 2002b for
details) was then assigned five values (50, 100, 500, 1000,
and 5000 ha) to ascertain the importance of the fire size dis-
tribution function on landscape fire regimes (the observed
value for the study area was 50 ha). For the two fire parame-
ter simulation experiments, we averaged and computed modal
values of fire occurrence and severity respectively over 500-
year simulation runs. We used the observed fire parameter
values for the simulation time span experiments (1.0 fire
size multiplier and 50 ha fire size parameters). Even though
LANDSUM has stochastic elements, we performed only one
run for each simulation experimentbecauseprevious analyses
showed inherent fire ignition stochasticity had a minor effect
on landscape-level LANDSUM results (Keane et al. 2002b).
And, we only report fire frequency characteristics in the form
of landscape fire return interval (average fire return interval
for all pixels on the landscape) for simplicity and brevity.
Simulation model comparison
Differences in simulation models also have a great effect
on resultant fire regime prediction. Each simulation model
is developed with specific purposes, ecosystems, and land-
scapes in mind, and this limits the application of models to
Fig. 1. Fuels input layers used in the comparison of FIRESCAPE (Cary and Banks 1999) and LANDSUM (Keane et al. 2002b) for predicting
fire regimes. (a) Elevation (dark areas are low); (b) random fuel assignment; (c) clustered using ellipses of varying size. The homogeneous fuel
layer is not shown because it would be a square of one color (i.e. one fuel type).
other areas, ecosystems, or situations (Gardner et al. 1999).
We compared the complex mechanistic model FIRESCAPE
(Cary and Banks 1999; Cary 2002) to the simple pathway
model LANDSUM (Keane et al. 2002b) to explore the impor-
tance of topography, fuels pattern, and climate in fire regime
generation across models. We applied these models to neu-
tral landscapes of 1000 ×1000 pixels of 50 m width where
topography, fuels, and climate were varied in a factorial
design. Simulation parameters (i.e. succession, climate, fire
frequency, and fire size) for each model were taken from the
native landscapes for which the model was built to eliminate
geographical bias in model development.
Topography was modelled as flat, moderate, and moun-
tainous using a 2-dimensional sine function with a periodicity
of 16.7 km or 333 pixels (i.e. elevation relief for flat was
zero, moderate was 1250 m and mountainous was 2500 m)
(Fig. 1a). Spatial fuel distributions were created using three
patterns: homogeneous, random, and clumped (Fig. 1b, c).
Ten replicates of the clumped fuel pattern were generated
using an unpublished algorithm (personal communication,
R.H. Gardner). It involved invoking randomly orientated and
located elliptical disturbance patterns of varying size to set
back the ‘age’ of fuel or community by 0.1 from a max-
imum of 1 indicating the highest fuel loadings. The 90th
percentile size of the ellipses was 100 ha and the aspect ratio
of the axes was set at 0.8. The set back of ‘age’ in over-
lapping ellipses was additive and new ellipses were added
until a landscape average ‘age’ setback of 0.25 was achieved.
For FIRESCAPE, these setback age values were translated
into fuel amounts via a negative exponential fuel accumu-
lation curve (Olson 1963) using a steady-state litter load of
1.637 kg m
, which is the average steady-state fine litter
loading observed for high elevation (>1500 m) sites in the
Australian Capital Territory (ACT) region, and a decomposi-
tion constant of 0.3 (Cary 1998). For LANDSUM, fuel ‘age’
was classified into eight classes of sequential stages of veg-
etation succession where the youngest fuel (<0.12) and the
Using simulation to map fire regimes 313
climax community allocated to the oldest fuel (>0.88). In
this fashion, the 10 replicates of clumped fuel ‘age’ were
transformed into input data appropriate for each model.
For the observed weather, 10 yearlong sequences of daily
weather were chosen from 42 years of daily records at
Glacier National Park, USA (LANDSUM) and 42 years
of simulated weather from a weather generation algorithm
(Richardson 1981) implemented for the Australian Capital
Territory Region (FIRESCAPE). The generation algorithm
produces sequences of weather with similar statistical qual-
ities as that of observed data from the region (Cary and
Gallant 1997) and is the primary weather component of the
FIRESCAPE model. Weather years were chosen so that they
best matched the variation in average daily maximum tem-
C) and average daily precipitation (mm) across
all years available at each location. These weather streams
defined a Current scenario that was then modified to create
two additional climate change weather scenarios by adding
C to daily temperatures (Cubasch et al. 2001) and mul-
tiplying daily rainfall by 0.8 and 1.2 to create the warm, dry
scenario and warm, moist scenario, respectively.
A total of 2700 1-year-long simulations without succes-
sion were run for each model given 27 unique combina-
tions of elevation (mountainous, moderate, flat), fuel pattern
(clumped, random, homogeneous), and climate (observed,
warm-dry, warm-moist) and 10 replicate maps of each fuel
pattern and 10 replicate sequences for each weather scenario.
Since LANDSUM has several stochastic functions, each sim-
ulation was repeated 10 times. The number of fire ignitions
and the total area burnt per year were recorded for each
1-year simulation. Fire ignitions were defined as ignitions
that successfully spread to at least one pixel adjacent to that
where the fire was initially ignited. The area burned and num-
ber of ignitions was written to computer files that were later
analysed using a fully factorial ANOVA design with the SAS
statistical package (SAS Institute 1990). This model com-
parison method was the prototype for a more robust model
comparison using several other spatially explicit landscape
fire succession models by the authors.
Results and discussion
Mapping strategy comparison
Fire regime maps created from the three strategies were
quite different in both frequency and severity for a variety
of reasons (Fig. 2a–c). Rule sets and parameters used in
the classification strategies are syntheses of empirical data,
expert knowledge, and observed experience for the study area
(Barrett and Arno 1992). As such, these rules represent a
coarser spatial, temporal, and category resolution than that of
statistical and simulation strategy and this resolution is man-
ifest in the maps because Barrett and Arno (1992) believed a
large portion of the landscape was in frequent, mixed sever-
ity fire regimes (Tables 1 and 2). Neither the classification
nor statistical strategy incorporated spatial relationships into
predictive models; that is, adjacent stands or surrounding
topography did not influence fire dynamics. Statistical strate-
gies ultimately rely on comprehensive and accurate field data,
and the field data from this study were reasonably repre-
sentative of the biophysical environment, but were limited
by sample size (only 64 plots) and temporal depth (approxi-
mately 200–300 years of fire history). As a result, very little
of the landscape was mapped to frequent, non-lethal surface
fire regimes because there were few plots that represented
this regime (Tables 1 and 2). Simulation modelling provided
greater temporal depth (1000 years) but results were further
removedfrom reality because they incorrectly assume that the
model accurately simulated fire ignition, growth, and inten-
sity. Most of the landscape is in the moderate fire return
interval class because fire frequencies were averaged over
only 1000 years, and most fires were long fire return interval
and stand replacement, so many pixels did not have a rich
history of all three severity types (Keane et al. 2003).
Statistical strategies are the most accurate because they
best represent the data used to construct and validate the
models (Tables 1 and 2). But, the Kappa statistic is low
because of the uneven distribution of field plots across fre-
quency and severity types due to the absence of fire history
records on the study landscape for long fire return interval
ecosystems. A surprising result is the value of physically
based variables in statistical analysis. Ecosystem process
variables, summarized from Biome-BGC simulations, not
only increased accuracies by 10% for both maps, they also
tended to portray spatial relationships at the scale of fire
regime dynamics better than other indirect variables. This
is because the Biome-BGC model integrated coarse scale
weather (2 km resolution) with mid-scale soils data (100–
500 m resolution) and fine scale vegetation (30 m resolution)
to quantify ecosystem process such as net primary produc-
tivity, heterotrophic respiration, and evapotranspiration, that
directly influence fire and fuel dynamics, at the appropriate
predictive scale (Keane et al. 2002a). Statistical analysis with
onlytopographical variables tended to exploit inconsistencies
in the DEM (Digital Elevation Model) because there is only
one scale represented—30 m (Keane et al. 2003).
Both classification and statistical strategies are used exten-
sively because they are somewhat simple, data-driven, and
easy to implement (Morgan et al. 2001; Keane et al. 2003).
Using classification strategies, land managers can quickly
and easily create predictive fire regime maps with little field
data using expert opinion, but these maps often lack a mea-
sure of error or statistical variability so they may contain
considerable errors. Fire regime maps created by statistical
strategies provide these measures but they are limited by the
scope of the data (e.g. geographic region, temporal depth,
ecosystem, and biophysical setting) and it is important that
the independent or predictor variables be mapped across the
entire analysis landscape, which is rarely the case for many
314 R. E. Keane et al.
Frequent (interval 40 years)
Infrequent (interval 100 years)
Moderate (interval 40 –100 years)
Fig. 2. Resultant fire regime maps of fire frequency created using the three strategies of classification,
statistical analysis, and simulation modelling. (a) Fire frequency using classification strategy (Barrett and
Arno 1992); (b) fire frequency using statistical analysis strategy (discriminant analysis); (c) fire frequency
using simulation modelling (LANDSUM).
lands. The complexity of classification and statistical strate-
gies is much less than a simulation strategy where extensive
expertise in computer programming, landscape ecology, fire
dynamics, and mapping is needed to create landscape fire suc-
cession models. Moreover, simulation models are notoriously
difficult to parameterize, initialize, and execute.
Nevertheless, we found simulation modelling superior
to classification and statistical strategies for mapping fire
regimes with respect to several desirable aspects. First, the
temporal depth of fire history field data needed to develop
extensive and comprehensive rule-based or statistical models
is often limited. Fire history studies in many forested eco-
systems have only a 200–500-year fire record taken from a
spatially and temporally discontinuous record (i.e. fire scars).
Many fire dominated ecosystems of the world, such as in
Australia and central Africa, lack extensive fire scar records
because there are few plants that record a fire scar and a
discrete annual growth ring record. Simulation models can
be executed for long periods so resultant fire regimes are
summarized across ecosystem-appropriate time spans. Many
simulation models integrate important landscape processes,
such as fire, weather, and succession, across multiple scales
Using simulation to map fire regimes 315
Table 1. A comparison of the fire interval classes generated from the three fire regime
Numbers inside the table represent the percentage of the Lower Selway landscape that was
predicted for each fire return interval class. The last row shows the percentage of plots in each fire
return interval category (n =64)
Strategy Fire return interval Map agreement with plot data
Frequent Moderate Infrequent Overall % Kappa statistic
(<40 years) (40–100 years) (>100 years) correct
Classification 46.2 36.7 13.4 33.00 0.15
Statistical 17.8 41.5 40.7 64.58 0.40
Simulation 14.4 82.3 3.3 31.25 0.00
Percentage of 12.5 50.0 37.5 100.0
Table 2. A comparison of the fire severity classes generated from the three fire regime mapping strategies
Numbers inside the table represent the percentage of the Lower Selway landscape that was predicted for each fire
severity class. The last row shows the percentage of plots in each fire severity category (n =64)
Strategy Fire severity class Map agreement with plot data
Non-lethal surface Mixed severity Stand-replacing fire Overall % Kappa statistic
fire (% landscape) fire (% landscape) (% landscape) correct
Classification 14.9 66.3 15.1 50.0 0.19
Statistical 5.7 52.3 41.9 72.9 0.51
Simulation 7.1 22.4 70.5 39.6 0.00
Percentage of 8.3 56.3 35.4 100.0
into one comprehensive application. Landscape fire succes-
sion models also have the ability of integrating spatially
discontinuous point data for fire history, biophysical set-
tings, and vegetation composition (used to quantify input
parameters) into one cohesive spatial application. Simulation
models can be modified to include explicit representations of
the factors that influence fire regimes, such as climate and
humans. For example, Cary and Banks (1999) simulated fire
regimes under different climate scenarios for a landscape in
the Brindabella Range,ACT (Cary 2002), and Wimberlyet al.
(2000) simulated fire regimes under current fire exclusion
policies for a large landscape in western Oregon, USA. And
last, simulation models that predict fire regimes can be used
for many other purposes, such as predicting wildlife habitat,
watershed erosion, and fuel loadings for various management
Simulation sensitivity analysis
As expected, the simulation parameters greatly influenced
simulated spatial fire regimes (Fig. 3; for brevity only fire fre-
quency is shown). Landscape fire return intervals increased
8–15% when site-level fire ignition probabilities were mul-
tiplied by 0.5, and return intervals decreased 8–15% when
probabilities were multiplied by 1.5. However, when these
probabilities were multiplied by 2.0 and greater, it appears
the landscape fire return interval stabilized at ∼25% below
that observed for the landscape. This indicates that the pat-
tern of recently burned communities and the landscape size
greatly influence fire return interval as fire ignition frequen-
cies increase. The same phenomenon is found in the fire size
parameter experiment (Fig. 3c). Landscape fire return inter-
vals tend to stabilize around 28 years as the average fire size
parameter increases compared to the 60–80 years observed
for the Lower Selway landscape. Again, the small landscape
extent(51 000 ha) and the increased presence of seral commu-
nities that have low ignition probabilities heavily influence
landscape fire regime characteristics. Between the two fire
parameters, average fire size appears to have the greatest
effect on simulated fire regimes and, therefore, should be
estimated with greater accuracy (Fig. 4). An error of 20%
in estimating fire probabilities might only result in a 5–10%
error in landscape fire return interval, but the same error in
average fire size estimation might result in an error of greater
than 50% in landscape fire return interval (see the example
in Fig. 4).
Simulation length is very important in the computation
of landscape fire return interval (Fig. 3b). Short simulation
periods (<300 years) were not long enough to adequately
represent fire frequency and severity, especially in long fire
return interval ecosystems such as subalpine forests. The fire
return interval started to converge to observed values after
∼300 years of simulation indicating that at least 500 yearsand
preferably 1000 years should be used to compute fire severity
316 R. E. Keane et al.
Landscape fire interval
0 0.5 1 1.5 2 2.5 3
Ignition probability multiplier
Landscape fire interval
Length of simulation (annual time steps)
0 200 400 600 800 1000 1200 1400 1600
Landscape fire interval
1 10 100 1000 10000
Fig. 3. Response of landscape fire interval to simulation parameters
in LANDSUM. (a) Ignition probability multiplier (using 500 year sim-
ulation period, 50 ha fire size parameter); (b) length of simulation (1.0
ignition probability multiplier, 50 ha fire size parameter); and (c) fire
size parameter (500 year simulation, 1.0 ignition probability multiplier).
Triangle data points show landscape fire interval calculated by includ-
ing areas that do not burn (i.e. rock/barren); diamond data points show
landscape fire interval calculated without these unburnable areas.
and frequency maps for the Lower Selway landscape. Cary
(1998) found a similar result for inter-fire interval, intensity
and season of occurrence in the Brindabella Range study.
The diversity of fire history increases as simulation length
becomes longer, resulting in a greater resolution in fire return
interval calculations (Fig. 5). It appears that the first 200–400
yearsshould not be included in the computation (Fig. 3b).The
simulation time span and the initial span of years to exclude
from the analysis is landscape specific, depending on topog-
raphy, fire return intervals, and climate, so it is probably best
to run the model long enough so there are at least 5 fires per
map unit (i.e. pixel or polygon). In short interval fire systems,
such as the tropical savannas of Australia, this time would be
significantly shorter, especially when burn sizes are often
larger than those of the USA Rocky Mountains.
An unexpected result is the importance of unburnable
landscape elements (e.g. rock, water, snowfields) to the
computation of landscape fire regime characteristics (trian-
gle v. diamond data points in Fig. 3). The average landscape
fire return interval increased ∼15% when these areas were
included in the calculation for this study area, regardless of
input fire frequency or fire size parameterization (Fig. 3a, c).
Landscape fire return interval monotonically increased as
simulation length increased because the fire return interval
for these areas is assigned the simulation time length in our
algorithm (Fig. 3b). This is illustrated in Fig. 5 by the number
of pixels with a return interval equal to simulation length (bar
furthest right on histograms). When these areas are excluded,
the landscape fire return interval eventually will converge to
a somewhat stable value; it did not in our experiment because
we were unable to generate a simulation length long enough
for the study area due to computational limitations.
Simulation model comparison
There were major differences between the two landscape fire
succession models FIRESCAPE and LANDSUM (Fig. 6).
The three factors of terrain, fuels, and climate explained
only 25% of variation for LANDSUM simulations but
over 60% for FIRESCAPE simulations. This indicates that
FIRESCAPE has a more complete integration of the pro-
cesses influencing fire regimes, mainly climate and topog-
raphy. FIRESCAPE contains a climate driver linked to fire
ignitions coupled with a comprehensive fire spread model.
However, LANDSUM appears to have a greater sensitivity
to the patterns of fuels on the landscape, probably because
the model simulates a greater differentiation in fuels (Fig. 6).
These results are entirely explained by the inherent design
of each simulation model. LANDSUM was developed to
simulate fire, vegetation, and landscape dynamics with a
minimal set of simple input parameters (Keane et al. 1996a,
1997b, 2002b). As a result, LANDSUM does not include
daily weather, direct estimations of fuel load, and a highly
mechanistic fire spread model, thereby explaining its lowsen-
sitivity (low r
) to climate and topography factors (Fig. 6).
FIRESCAPE contains a daily climate driver, a direct simula-
tion of fuel loads, and a comprehensive fire spread component
(Cary and Banks 1999; McCarthy and Cary 2002). LAND-
SUM was built for those landscapes where fire history
evidence can be collected and summarized into appropriate
model parameters (e.g. fire frequencies from fire scars). The
ecosystems that FIRESCAPE was developed for rarely con-
tain fire history evidence so a more complex integration of
fire processes with climate was warranted (Cary and Banks
Four factors ultimately dictate landscape fire succession
model design: (1) diversity of ecosystem processes affecting
fire regimes; (2) availability of field data; (3) planned
application; and (4) computing resources. LANDSUM was
developed for northern Rocky Mountain ecosystems in the
Using simulation to map fire regimes 317
0–20 20–50 50–150 150–300 300
Fire return interval (years)
Fig. 4. Fire frequency maps created over 500 year simulations using the observed fire ignition prob-
abilities but changing the fire size parameter. The same random number sequence was used for each
simulation to eliminate stochastic variability betweenruns. Fire size parameters used in these simulations
were (a) 100 ha, (b) 500 ha, (c) 1000 ha, and (d) 5000 ha.
western United States while FIRESCAPE was developed for
eucalyptforests of south-easternAustralia. Differences in fire
regime characteristics within these two areas, along with the
available field data and planned application, dictated model
design and development. The selection of the most appropri-
ate model for application to other landscapes requires the user
to evaluate several factors. First, the output must be pertinent
to the user’s application. Second, the data to parameterize and
initialize the model must be available and of sufficient qual-
ity and quantity. Next, the model must contain an explicit
simulation of the processes that control fire regimes for the
landscape in question. Last, there must be sufficient comput-
ing resources and expertise available to execute the model
and interpret its results.
Challenges and opportunities
There are six primary challenges in predicting fire regimes
across a landscape. The first is matching or rectifying the
spatial and temporal scales that govern fire and landscape
dynamics (Simard 1991). Fire is a complex disturbance pro-
cess manifest at many time and space scales, yet many
fire regime studies describe fire dynamics at the stand-level
318 R. E. Keane et al.
Fig. 5. Changes in resolution in the computation of fire intervals with increasing simulation length. Shown is the frequency distribution of pixels
by fire return interval (years) for the following simulation lengths: (a) 100 years (5 distinct interval values); (b) 250 years (9 distinct interval values);
(c) 500 years (16 distinct interval values); and (d) 1000 years (32 distinct interval values).
Climate Fuel Terrain Model Residual
Variance explained (r
Fig. 6. A comparison of the FIRESCAPE and LANDSUM landscape
fire succession models using anANOVA analysis on factorial simulation
experimental design of climate, fuels, and terrain for simulated output
of burned area and number of fires. Shown is the amount of variance
explained by each factor and their interactions using the correlation
coefficient or r
. Residual is the variation in model output not explained
by the three factors.
across relatively short time spans. For example, lightning
dynamics is an important and complex process that is rarely
integrated into predictive fire regime models because of its
large scale requirements (Knight 1987; Agee 1991), although
Cary (1998) includes a lightning location model in the
Brindabella Range Study.
The second challenge is including all the elements that
characterize fire regimes (frequency, severity, pattern, fuel
type, seasonality) into a comprehensive predictive model.
Frequency can be easily quantified from point- or stand-level
data, and fire pattern (i.e. size and shape of burned patches)
is evaluated from fire atlases or simulated from models; how-
ever, fire seasonality requires long-term climate records and
an assessment of the phenological stages of affected veg-
etation. Fire severity not only depends on stand-level fuel
characteristics (Ryan and Noste 1985), but also the location
of that stand in the landscape matrix (Camp et al. 1997) and
the dynamics of coarse-scale wind patterns for that stand
(Swanson et al. 1997).
The third challenge is accounting for the inherent spatial,
temporal, and process (severity or intensity) variabilitywithin
Using simulation to map fire regimes 319
a fire regime, which is ultimately responsible for landscape
structure and composition (Heinselman 1981; Agee 1993;
Gill and McCarthy 1998; McKenzie 1998). For example,
the variability in fire return interval is essential for assess-
ing the degree of departure from historical conditions along
with designing fire treatments, assessments and schedules for
landscape management (Landres et al. 1999). Fire pattern
variability dictates the size and shape of future treatments
that attempt to emulate natural fire (Hunter 1993). And, the
range of fire severities experienced within a fire regime will
guide treatment design and implementation (Keane 2000).
Yet, most fire regime models rarely characterize the inher-
ent variability of the predicted variables of fire frequency or
The fourth challenge is avoiding the possibility that devel-
oped fire regime prediction models will become obsolete
once predicted climate change, exotic invasions, and changes
in land management become reality (Shinn 1980; Weber
and Flannigan 1997). Future changes in climate will render
most maps inaccurate unless the models used to describe fire
regime contain a link to climatic processes or other change
The fifth challenge is designing useful fire regime classi-
fication categories for use across diverse local, regional, and
national applications. The range of fire return intervals that
define a frequency category may not be optimal across all
landscapes in a region or across all space scales. For exam-
ple, Hardy et al. (2001) defined their frequent fire category
as lands having mean fire return intervals between 0 and 35
years, yet 35 years might be especially long for some grass-
lands and very short for some forests. Such a broad range
may not provide sufficient distinction between important fire
regimes on dry, fire-prone landscapes. And, the interpreta-
tion of this somewhat arbitrary range might be misleading
for many management applications; managers might use the
mid-point of this class as a target fire return interval without
first evaluating evidence collected from their local landscapes
and quantifying return interval and variability.
The last challenge is the collection and analysis of field
data, which are critical for a myriad of fire regime prediction
tasks because they provide the only truth for understand-
ing, predicting, and interpreting fire dynamics (Morgan et al.
2001). Field data are needed to (1) design and describe fire
regime characteristics and resultant classifications; (2) create
predictive algorithms using statistical techniques; (3) param-
eterize and initialize simulation models; and (4) assess and
validate predictive models and their results. However, many
fire history sampling techniques have problems: (1) diffi-
cult to determine the spatial extent or pattern of fire events;
(2) expensive to collect and analyse the data; (3) shal-
low temporal depth in some ecosystems; (4) insufficient
records on the landscape; and (5) difficult to employ standard
analytical and collection techniques because of ecosystem
The three major strategies for mapping fire regimes
have unique advantages and limitations that dictate their
• The classification strategy is relatively quick, easy, and
simple and works best when there is very little fire history
data. However, it does not account for spatial relationships
and tends to be inaccurate and portray fire regimes at a
• The statistical strategy is also relatively easy, straightfor-
ward, and popular, and it is the most accurate, but this
data-driven approach requires copious field data that are
expensive and difficult to collect, and this strategy requires
extensive expertise in statistical modelling.
• Simulation models are complex computer programs that
are often difficult to parameterize, initialize, and exe-
cute. However, the simulated fire regimes have the deepest
temporal depth; integrate complex multiscale spatial inter-
actions; and can be used to explore alternative fire regimes
under changing climate and vegetation.
The simulation strategy appears to generate the most
robust and realistic fire regimes because of the deep tem-
poral record, but quantification of several parameters is very
important in the simulation:
• Estimation errors for fire ignition probability parame-
ters may result in minor differences in simulation results,
whereas minor errors in the average fire size parameter
could have major influences on subsequent fire regime
simulations and mapping.
• The simulation time span should be long enough for the
majority of the map units (pixelsorpolygons) to experience
at least 3–5 fires.
• Results from these simulation experiments are landscape
specific and may be significantly different for other land-
scapes, ecosystems, or models.
We thank the Global Change Terrestrial Ecosystems (GCTE)
Task 2.2.2 Landscape Fires working group for guidance and
assistance in designing the model comparison study, espe-
cially Bob Gardner, University of Maryland, Sandra Lavorel,
Mike Flannigan, Canadian Forestry Service, and the six other
members. We acknowledge Matt Rollins, James Menakis,
and Wendel Hann, USDA Forest Service, and John Ludwig
and Dick Williams, CSIRO Tropical Ecosystems Research
Centre, for their technical reviews; the National Center
for Ecosystem Analysis and Synthesis for funding GCTE
Acevedo MF, Urban DL, Ablan M (1995) Transition and gap models of
forest dynamics. Ecological Applications 5, 1040–1055.
320 R. E. Keane et al.
Agee JK (Ed.) (1991) Fire history of Douglas-fir forests in the Pacific
Northwest. USDA Forest Service, Pacific Northwest Research
Station General Technical Report PNW-GTR-285.
Agee JK (1993) ‘Fire ecology of Pacific Northwest forests.’ (Island
Press: Washington, D.C.)
Agee JK (1995) Fire regimes and approaches for determining fire his-
tory. In ‘The use of fire in forest restoration’. (Eds CC Hardy and
SF Arno) pp. 12–13. USDA Forest Service, Intermountain Research
Station General Technical Report INT-GTR-341.
Arno SF, Parsons DJ, Keane RE (2000) Mixed-severity fire regimes in
the northern Rocky Mountains: consequences of fire exclusion and
options for the future. In ‘Wilderness science in a time of change
conference. Volume 5: Wilderness ecosystems, threat, and manage-
ment’. pp. 225–232. Missoula, MT, 23–27 May 1999. (USDA Forest
Service Rocky Mountain Research Station: Fort Collins CO)
Bahre CJW (1991) ‘A legacy of change: Historic human impact on
vegetation in the Arizona borderlands.’ (The University of Arizona
Baker WL (1994) Restoration of landscape structure altered by fire
suppression. Conservation Biology 8, 763–769.
Barrett SW, Arno SF (1982) Indian fires as an ecological influence in
the northern Rockies. Journal of Forestry 80, 647–651.
Barrett SW, Arno SF (1992) Classifying fire regimes and defining
their topographic controls in the Selway-Bitterroot Wilderness. In
‘Proceedings of the 11th Conference on Fire and Forest Meteo-
rology, Missoula, Montana, USA’. (Eds P Andrews and DF Potts)
pp. 299–307. (Society of American Foresters: Bethesda, MD)
Bessie WC, Johnson EA (1995) The relative importance of fuels and
weather on fire behavior in subalpine forests. Ecology 76, 747–762.
Bradstock RA, Williams JE, Gill AM (Eds) (2002) ‘Flammable
Australia: The fire regimes and biodiversity of a continent.’
(Cambridge University Press: Cambridge, UK)
Brown JK (1973) Fire cycles and community dynamics of lodgepole
pine forests. In ‘Management of lodgepole pine ecosystems’. Vol.
I. (Ed. DB Baumgartner) pp. 23–55. (Washington State University
Brown JK (1995) Fire regimes and their relevance to ecosystem man-
agement. In ‘Proceedings of the Society of American Foresters 1994
Annual Meeting’. pp. 171–178. (Society of American Foresters:
Camp A, Oliver C, Hessburg P, Everett R (1997) Predicting late-
successional fire refugia pre-dating European settlement in the
Wenatchee Mountains. Forest Ecology and Management 95, 63–77.
Cary GJ (1998) ‘Predicting fire regimes and their ecological effects
in spatially complex landscapes.’ PhD Thesis, Australian National
Cary GJ (2002) Importance of a changing climate for fire regimes in
Australia. In ‘Flammable Australia: The fire regimes and biodiver-
sity of a continent’. (Eds RA Bradstock, JE Williams and AM Gill)
pp. 26–49. (Cambridge University Press: Cambridge, UK)
Cary GJ, Banks JCG (1999) Fire regime sensitivity to global cli-
mate change: An Australian perspective. In ‘Advances in global
change research’. (Ed. JL Innes) pp. 233–246. (Kluwer Academic
Publishers: Dordrecht and Boston)
Cary GJ, Gallant JC (1997) Application of a stochastic climate gen-
erator for fire danger modelling. In ‘Bushfire ’97’. pp. 123–134.
Australasian Bushfire Conference. (CSIRO: Darwin)
Chew JD (1997) Simulating vegetation patterns and processes at land-
scape scales. In ‘Integrating spatial information technologies for
tomorrow: GIS ’97 conference proceedings, 17–20 February 1997’.
pp. 287–290. (GIS World: Fort Collins, CO)
Crutzen PJ, Goldammer JG (1993) ‘Fire in the environment: The eco-
logical, atmospheric, and climatic importance of vegetation fires.’
(John Wiley and Sons: Chichester)
Cubasch U, MeehlGA, Boer GJ, StoufferRJ, Dix M,NodaA, SeniorCA,
Raper S,Yap KS (2001) Projections of future climate change. In ‘Cli-
mate change 2001: The scientific basis’. (Eds JT Houghton, Y Ding,
DJ Griggs, M Noguer, P van der Linden, X Dai, K Maskell and
CI Johnson) Contribution of Working Group I to the Third Assess-
ment Report of the Intergovernmental Panel on Climate Change,
pp. 525–582. (Cambridge University Press: Cambridge, UK)
DeBano LF, Neary DG, Ffolliott PF (1998) ‘Fire’s effect on ecosystems.’
(John Wiley and Sons: New York)
Finney MA (1998) FARSITE: Fire Area Simulator—model devel-
opment and evaluation. USDA Forest Service, Rocky Mountain
Research Station Research Paper RMRS-RP-4. Fort Collins, CO.
Flannigan MD, Wagner CEV (1991) Climate change and wildfire in
Canada. Canadian Journal of Forest Research 21, 66–72.
Frost CC (1998) Presettlement fire frequency regimes of the United
States: a first approximation. Tall Timbers Fire Ecology Conference
Gardner RH, Hargrove WW, Turner MG, Romme WH (1996)
Climate change, disturbances and landscape dynamics. In ‘Global
change andterrestrial ecosystems’. (EdsBH Walker andWL Steffen)
pp. 149–172. (Cambridge University Press: Cambridge, UK)
Gardner RH, Romme WH, Turner MG (1999) Predicting forest fire
effects at landscape scales. In ‘Spatial modeling of forest land-
scape change: Approaches and applications’. (Eds DJ Mladenoff
and W Baker) pp. 163–185. (Cambridge University Press:
Gill AM (1975) Fire and the Australian flora: a review. Australian
Forestry 38, 4–25.
Gill AM (1998) An hierarchy of fire effects: impact of fire regimes on
landscapes. In ‘III International Conference on Forest Fire Research,
14th Conference on Fire and Forest Meteorology, Luso. Portugal’.
Gill AM, McCarthy MA (1998) Intervals between prescribed fires
in Australia: what intrinsic variation should apply? Biological
Conservation 85, 161–169.
Gruell GE (1985) Indian fires in the interior west: a widespread
influence. USDA Forest Service General Technical Report
Hardy CC, Schmidt KM, Menakis JP, Sampson NR (2001) Spatial
data for national fire planning and fuel management. International
Journal of Wildland Fire 10, 353–372.
He HS, Mladenoff DJ (1999) Spatially explicit and stochastic simula-
tion of forest-landscape fire disturbance and succession. Ecology 80,
Heinselman ML (1981) Fire and succession in the conifer forests of
North America. In ‘Forest succession: concepts and applications’.
(Eds DC West, HH Shugart and DB Botkin) pp. 375–405 (Springer-
Verlag: New York)
Heinselman ML (1985) Fire regimes and management options in eco-
systems with large high-intensity fires. In ‘Symposium and work-
shop on wilderness fire’. (Eds JE Lotan, BM Kilgore, WC Fischer
and RW Mutch) pp. 101–109. USDA Forest Service, Intermoun-
tain Forest and Range Experiment Station General Technical Report
INT-1982. Missoula, MT.
Hunter ML, Jr (1993) Natural fire regimes as spatial models for
managing boreal forests. Biological Conservation 65, 115–120.
Kay CE (1995) Aboriginal overkill and native burning: Implications
for modern ecosystem management. Western Journal of Applied
Forestry 10, 121–126.
Keane RE (2000) Landscape fire succession modeling: Linking ecosys-
tem simulations for comprehensive applications. In ‘Landscape fire
modeling—challenges and opportunities’. (Eds BC Hawkes and
MD Flannigan) pp. 5–8. Northern Forestry Centre, Victoria, British
Columbia, Information Report NOR-X-371.
Using simulation to map fire regimes 321
Keane RE, McNicoll C, Rollins MG (2002a) Integrating ecosys-
tem sampling, gradient modeling, remote sensing, and ecosystem
simulation to create spatially explicit landscape inventories. USDA
Forest Service General Technical Report RMRS-GTR-92.
Keane RE, Hardy CC, Ryan KC, Finney MA (1997a) Simulating
effects of fire on gaseous emissions from future landscape of
Glacier National Park, Montana, USA. World Resources Review 9,
Keane RE, Long DG, Menakis JP, Hann WJ, Bevins C (1996a) Sim-
ulating coarse scale vegetation dynamics with the Columbia River
Basin succession model CRBSUM. USDA Forest Service General
Technical Report INT-GTR-340.
Keane RE, Finney MA (2003). The simulation of landscape fire,
climate, and ecosystem dynamics. In ‘Fire and global change in tem-
perate ecosystems of the Western Americas’. (Eds TT Swetnam, WL
Baker, G Montenegro and T Veblen) (Springer-Verlag: New York)
Keane RE, Long D, Basford D, Levesque BA (1997b) Simulating
vegetation dynamics acrossmultiple scales toassess alternativeman-
agement strategies. In ‘Integrating spatial information technologies
for tomorrow: GIS ’97 conference proceedings, 17–20 February
1997’. pp. 310–315. (GIS World: Vancouver)
Keane RE, Morgan P, White JD (1999) Temporal pattern of ecosys-
tem processes on simulated landscapes of Glacier National Park,
Montana, USA. Landscape Ecology 14, 311–329.
Keane RE, Parsons R, Hessburg P (2002b) Estimating historical range
and variation of landscape patch dynamics: limitations of the
simulation approach. Ecological Modelling 151, 29–49.
Keane RE, Parsons R, Rollins MG (2003) Predicting fire regimes across
multiple scales. In ‘Emulating natural disturbances: concepts and
techniques’. (Ed. L Buse) (Cambridge University Press: Cambridge,
UK) (In press)
Keane RE, Ryan KC, Running SW (1996b) Simulating effects of fire
on northern Rocky Mountain landscapes with the ecological process
model FIRE-BCG. Tree Physiology 16, 319–331.
Knight DH (1987) Parasites, lightning, and the vegetation mosaic in
wilderness landscapes. In ‘Landscape heterogeneity and distur-
bance’. (Ed. MG Turner) pp. 59–83. (Springer-Verlag: New York)
Kushla JD, Ripple WJ (1997) The role of terrain in a fire mosaic of a
temperate coniferous forest. Forest Ecology and Management 95,
Landres PB, Morgan P, Swanson FJ (1999) Overview and use of natu-
ral variability concepts in managing ecological systems. Ecological
Applications 9, 1179–1188.
Lertzman K, Fall J, Brigitte D (1998) Three kinds of heterogeneity in
fire regimes: at the crossroads of fire history and landscape ecology.
Northwest Science 72, 4–23.
Lewis HT (1985) Why Indians burned: specific verses general rea-
sons. USDA Forest Service, Intermountain ResearchStation General
Technical Report INT-182. Ogden, UT.
Li C, Ter-Mikaelian M, Perera A (1997) Temporal fire disturbance
patterns on a forest landscape. Ecological Modelling 99, 137–150.
Li CF, Flannigan MD, Corns IGW (2000) Influence of potential cli-
mate change on forest landscape dynamics of west-central Alberta.
Canadian Journal of Forest Research 30, 1905–1912.
Lutes D, Keane RE, Caratti J, Gangi L, Key CH (In press) FIREMON:
A fire monitoring and inventory sampling and analysis system.
USDA Forest Service General Technical Report RMRS-GTR.
McCarthy MA, Cary GJ (2002) Fire regimes in landscapes: models and
realities. In ‘Flammable Australia: the fire regimes and biodiver-
sity of a continent’. (Eds RA Bradstock, JE Williams and AM Gill)
pp. 77–94. (Cambridge University Press: Cambridge, UK)
McKenzie D (1998) Fire, vegetation, and scale: toward optimal models
for the Pacific Northwest. Northwest Science 72, 49–65.
Miller C, Urban DL (1999) A model of surface fire, climate, and forest
pattern in the Sierra Nevada, California. Ecological Modelling 114,
Morgan P, Hardy CC, Swetnam TW, Rollins MG, Long DG (2001)
Mapping fire regimes across time and space: Understanding coarse
and fine-scale fire patterns. International Journal of Wildland Fire
Olsen JS (1981) Carbon balance in relation to fire regimes. In ‘Pro-
ceedings of the conference Fire Regimes and Ecosystem Prop-
erties’. (Technical Coordinators HA Mooney, TM Bonnicksen,
NL Christensen, JE Lotan and WA Reiners) pp. 327–378. USDA
Forest Service General Technical Report WO-26.
Olson JS (1963) Energy storage and the balance of producers and
decomposers in ecological systems. Ecology 44, 322–331.
Price C, Rind D (1994) The impact of a 2×CO
climate on lightning-
caused fires. Journal of Climate 7, 1484–1494
Pyne SJ (1982) ‘Fire in America—a cultural history of wildland and
rural fire.’ (Princeton University Press: Princeton, NJ)
Richardson CW (1981) Stochastic simulation of daily precipitation,
temperature, and solar radiation. Water Resources Research 17,
Roberts DW, Betz DW (1999) Simulating landscape vegetation dynam-
ics of Bryce Canyon National Park with the vital attributes/fuzzy
systems model VAFS.LANDSIM. In ‘Spatial modeling of forest
landscape change: approaches and applications’. (Eds DJ Mladenoff
and WL Baker) pp. 99–123. (Cambridge University Press:
Russell EWB (1983) Indian-set fires in the forests of the northeastern
United States. Ecology 64, 78–88.
Ryan KC, Noste NV (1985) Evaluating prescribed fires. In ‘Wilderness
Fire Symposium’. (Eds JE Lotan, BM Kilgore, WC Fischer and
RW Mutch) pp. 230–237. USDA Forest Service, Intermountain
Research Station General Technical Report INT-182. Ogden, UT.
SAS Institute (1990) ‘SAS Procedures Guide Version 6, 3rd edn.’(SAS
Institute Inc.: Cary, NC)
Shinn DA (1980) Historical perspectives on range burning in the Inland
Pacific Northwest. Journal of Range Management 33, 415–423.
Shinneman DJ, Baker WL (1997) Nonequilibrium dynamics between
catastrophic disturbances and old-growth forests in ponderosa pine
landscapes of the Black Hills. Conservation Biology 11, 1276–1288.
Simard AJ (1991) Fire severity, changing scales, and how things hang
together. International Journal of Wildland Fire 1, 23–34.
Skinner CN, Chang C-R (1996) Fire regimes, past and present. Sierra
Nevada Ecosystem Project: Final report to Congress, Volume II.
Wildland Resources Center Report Number 37, University of
California at Davis, Davis, California, USA.
Stocks BJ, Fosberg MA, Lynaham TJ, Mearns L, Wotton BM, et al.
(1998) Climate change and forest fire potential in Russian and
Canadian boreal forests. Climatic Change 38, 1–13.
Strauss D, Bednar L, Mees R (1989) Do one percent of forest fires cause
ninety-nine percent of the damage? Forest Science 35, 319–328.
Swanson FJ, Franklin JF, Sedell JR (1997) Landscape patterns, distur-
bance, and management in the Pacific Northwest, USA. In ‘Chang-
ing landscapes: an ecological perspective’. (Eds IS Zonnneveld and
RTT Forman) pp. 191–213. (Springer-Verlag: New York)
Thornton PE (1998) Regional ecosystem simulation: Combining
surface- and satellite-based observations to study linkages between
terrestrial energy and mass budgets. PhD Dissertation, University of
Turner MG, Gardner RH, O’Neill RV (2001) ‘Landscape ecology in
theory and practice.’ (Springer-Verlag: New York)
Turner MG, Romme WH, Gardner RH, Hargrove WW (1997) Effect
of fire size and pattern on early succession in Yellowstone National
Park. Ecological Monographs 67, 411–433.
322 R. E. Keane et al.
Weber MG, Flannigan MD (1997) Canadian boreal forest ecosystem
structure and function in a changing climate: impact on fire regimes.
Environmental Review 5, 123–145.
Whisenant SG (1990) Changing fire frequencies on Idaho’s Snake River
Plains: Ecological and managementimplications. In ‘Proceedings on
cheatgrass invasion, shrub dieoff, and other aspects of shrub biol-
ogy’. (Ed. ED McArther) pp. 4–10. USDA Forest Service General
Technical Report INT-276. Ugden, UT.
Wimberly MC, Spies TA, Long CJ, Whitlock C (2000) Simulating his-
torical variability in the amount of old forest in the Oregon Coast
Range. Conservation Biology 14, 167–180.
Yarie J (1981) Forest fire cycles and life tables: a case study from interior
Alaska. Canadian Journal of Forest Research 11, 554–562.