SENSITIVITY ANALYSIS OF THE 2003 SIMI WILDFIRE EVENT
1* 1 1 1 1 2 1
S.H. Peterson , N.C. Goldstein , M.L. Clark , K.Q. Halligan , P. Schneider , P.E. Dennison , D.A. Roberts
Department of Geography, University of California at Santa Barbara, Santa Barbara, CA 93106-4060
*Telephone: (805) 893-4434; Fax: (805) 893-7782; Email: firstname.lastname@example.org
Department of Geography, University of Utah, Salt Lake City, UT, 84112-9155
PhD student, UCSB Department of Geography, 2002-Present; MA in Geography, SDSU Dept. Of Geography,
2000; Research interests in remote sensing and fire fuel mapping; published in International J. of Remote
In recent years , several extreme wildfires in the Western United States and Southeastern Australia have burned
within the mosaic of natural and residential land uses known an the wildland-urban interface. One such
destructive fire was the Simi Fire, part of the 2003 Southern California Fire Complex. The Simi Fire burned from
October 25 to November 5, 2003, consumed over 108,000 acres , destroyed 94 structures, and cost approximately
$10 million to suppress.
Advances in geospatial technology are making it easier for spatial modeling tools to be incorporated into the
wildfire fighting and mitigation process. These include GPS-enabled products, remote weather stations,
integrated local and regional GIS databases, and remotely sensed imagery. The focus of this work is the use of
state-of-the art geospatial products to simulate the Simi Wildfire, including testing the sensitivity of the wildfire
simulation package, FARSITE (Finney 1998), to data and model inputs. This work illustrates that when using a
spatial-dynamic model, sensitivity testing is necessary for understanding the variability of model output.
THE FARSITE WILDFIRE MODEL
FARSITE is a deterministic, equation-driven fire spread model that utilizes the semiempirically-based Rothermel
equations to predict fire behaviour (Rothermel 1972, 1983). FARSITE models fire propagation as a wave front
spreading from multiple vertices along a fire front, based on Huygens’ Principle (Richards 1990). The fire
direction and velocity at each vertex are dependent upon wind and fuel conditions and slope and aspect; the fire
front at time t+1 is generated by connecting the individual fire fronts generated at each vertex (Finney 1998).
FARSITE includes a well-developed user interface, and inputs/outputs are GIS-based.
FARSITE is currently mandated for use by the US Forest Service (USFS). It uses a variety of inputs, some of
which, like elevation, slope and aspect, are standardized and readily available. Weather-related inputs
(temperature, wind direction and speed, and dead fuel moisture) commonly come from a Remote Automated
Weather S tation (RAWS), in which case they are spatially invariant. Alternatively, forecast models , such as
MM5 (Grell et al. 1994) produce spatially varying output that can be forecast. In our simulations we used a 60
hour forecast at a grid spacing of 4km. Thus, future fire behaviour can also be predicted, operationally.
Vegetation information is supplied in 2 layers: Fuel models (Anderson, 1982) and canopy cover. Fuel models
es sentially are simplifications of the amount and distribution of surface fuels , used for a fire model. In addition to
the geospatial inputs needed for ru nning a FARSITE simulation, user-defined parameters such as the perimeter
resolution of the fire front are required. There is limited documentation concerning the ideal perimeter resolution
for a FARSITE run, and there is no discussion in the literature.
This work is a sensitivity analysis of FARSITE with respect to fuels data (including spatial resolution), perimeter
resolution and weather inputs. The data for examination were from the 2003 Simi Wildfire. Model validation was
accomplished by comparing spatial overlap between simulated fires and fire perimeters derived from the Active
Fire Product of the MODIS sensor, which is produced 4 times a day, at 1km pixel resolution; fire polygons from
MODIS were generated using a convex hull technique. The primary metric of comparison was the Lee-Sallee
metric (Lee and Sallee 1970), which measures agreement between two shapes. A Lee-Sallee of 1 indicates that the
two shapes are congruent, and anything less than 1 is a measure of their deviation from congruence. For this
test, we ran the FARSITE model for the first 58 hours of the Simi Fire (the period of maximum growth), from Oct
25, 1300 PST to Oct 27, 2300.
To examine the effect of fuel heterogeneity, we used a mode filter on the wildfire fuels. We filtered the input fuels
data at a generalization of 3,5,7,9,11,13,15, and 21 pixel boxes. We tested the different mode filters at a variety of
perimeter resolutions, at 60m and then 50m increments from 100m to 450m. We found that a mode-9 filter,
corresponding to a spatial resolution of 270m pixel cells, provided the best and least variable fit of the fire
perimeter at 58 hours (Figure 1).
Effect of mode filtering on mean Lee-Salee
0 2 4 6 8 10 12 14 16 18 20
Using the Mode-9 filter, we examined different perimeter resolutions while running the FARSITE model. As
shown in figure 2, a shallow gradient exists from highest (59m resolution) to lowest (450m) accuracy. However,
s ince the longer the perimeter length, the faster the model run, on the order of hours /days , using small perimeter
lengths operationally may not be necessary .
Effect of perimeter resolution on mode 9 Lee-Salee
0 10 20 30 40 50 60
Elapsed time, hours
Finally, we tested the sensitivity of FARSITE to weather inputs. Here, we tested using the RAWS data and
output from MM5, at three different perimeter lengths. The modelled weather underperformed at all of the time
steps except the final two.
This work illustrates the importance of sensitivity analysis when evaluating model performance. Every
geospatial model, especially those to be used in such important societal efforts as disaster modelling, should be
tested with respect to the effect of varying model parameters and inputs on model output . Spatially filtering
(coarsening) the fuels map caused the simulated fire to speed up, increasing agreement with actual perimeters.
Varying perimeter resolution had a relatively small effect on accuracy but affected model run time by orders of
magnitude. We hope that these results can inform decision makers in assessing optimal FARSITE inputs and
parameters. These results may have implications for mapping and management of chaparral in the wildland
interface of southern California.
Anderson, H.E. (1982) Aids to determining fuel models for estimating fire behavior, USDA Forest Service GTR-
INT-122, Intermountain Forest and Range Experiment Station.
Finney, M. A. (1998) FARSITE: Fire Area Simulator - Model development and evaluation, USDA Forest Service
RMRS-RP-4, Rocky Mountain Forest and Range Experiment Station.
Grell, G. A., Dudhia , J. and Stauffer, D.R. (1994), A description of the fifth-generation Penn S tate/NCAR
mesoscale model (MM5), NCAR Technical Note, NCAR/TN-398+STR.
Lee, D. and Sallee G. (1970) A method of measuring shape, Geographical Review 60, 555-563.
Linn, R., Reisner, J., Colman, J.J. and Witerkamp, J. (2002) Studying wildfire behavior usi g FIRETEC,
International Journal of Wildland Fire 11, 233-246.
Richards, G.D. (1972) An elliptical growth model of forest fire fronts and its numerical solution, International
Journal for Numerical Methods in Engineering 30, 1163-1179.
Rothermel, R. C. (1972) A mathematical model for predicting fire spread in wildland fuels, USDA Forest Service
RP-INT-115, Intermountain Forest and Range Experiment Station.
Rothermel, R.C. (1983) How to predict the spread and intensity of forest and range fires, USDA Forest Service
GTR-INT-143, Intermountain Forest and Range Experiment Station.