Note to data users: Please carefully review the metadata provided with each layer. We request that users consult with the ForestERA project in advance of using these data in publications and/or presentations to ensure that the strengths and limitations of the data are considered.

Description

Fire “risk” is defined as the probability of a fire occurring in a given area (Sampson et al. 2000).  Thus, fire risk layers are usually based on either ignition probabilities or probabilities that a fire will spread to a given area.  This layer provides an estimate of the probability that the ignition for a large fire (>50 acres or 20 ha) will occur in any given square kilometer across the landscape over a 16 year period.  The spatial and temporal aspects of the probability are based on the fire ignition data, which have a spatial resolution of 1km² and come from the period 1986 - 2001.  A second layer, which provides a measure of the uncertainty associated with the probability of occurrence, is also provided.  Both of these layers were created at 1km² resolution and resampled to 90m to match the resolution of other ForestERA data.  This was done using the mean of all 90m pixel values in an 11x11 area (11x11 focalmean filter).  The 11x11 pixel area approximates the 1km² resolution of the original data.  Thus the value for any 90m pixel is representative of the 1km² area around that 90m pixel rather than the 90m pixel alone.

Purpose

These layers were created as part of the ForestERA project to support landscape-scale forest restoration planning efforts by a broad group of stakeholders including federal and state agencies, academic institutions, and non-governmental entities.  These data are intended for regional analyses over spatial extents on the order of tens to hundreds of thousands of acres, and were not developed for use at finer spatial scales, although they may be useful for some applications at finer scales.

Development

We compiled a digital database of federal fire ignition occurrence data obtained directly from the US Departments of Agriculture and Interior, and from the National Fire Occurrence Database maintained by the USDA Forest Service Fire Sciences Laboratory.  These data included point-of-origin records of lightning- and human-caused ignitions occurring on lands managed by the Forest Service, Bureau of land Management, National Park Service, Fish and Wildlife Service, and Bureau of Indian Affairs.  We refined this database by including only those ignitions that occurred between 1 April and 30 September, the season that captures the driest months in the region and the monsoonal storm patterns, during which lightning strikes are most common (Swetnam & Betancourt 1998).  Additionally, within this period, we included only those occurrence records of ignitions resulting in large fires ≥20 ha in extent.

To quantify fire risk across the landscape, we used a weights of evidence (WOE) modeling approach.  WOE is a Bayesian method of event prediction using known locations of event occurrence.  The approach was originally developed for medical diagnoses (Spiegelhalter and Knill-Jones 1984), but has more recently been extended to the prediction and spatial analysis of mineral deposits (Agterberg 1989, 1990; Raines and Mihalasky 2002), archaeological sites (Hansen 2000), fossilized packrat middens (Mensing et al. 2000), and plant migration (Lyford 2003).  Weights of evidence models use the spatial location of known occurrence points to determine coefficients for a set of categorical input maps (Bonham-Carter et al. 1989).  For each analysis unit, or unit cell, these coefficients represent the probability of the input map characteristic or pattern being: a) present with a known occurrence, b) present without an occurrence, c) absent with an occurrence, or d) absent without an occurrence.  The WOE model takes a log-linear form, and the final product is a posterior probability map showing the cumulative conditional probability for presence of an occurrence at each unit cell.

Following the procedure described by Bonham-Carter et al. (1989), steps in our WOE analysis included: 1) estimation of a prior probability given only the occurrence data; 2) use of conditional probability ratios to calculate positive (W⁺) and negative (W^-) weights for each input map; 3) a pairwise test of conditional independence for each of the input maps, with respect to the known occurrences; 4) calculation of the posterior probability and estimate of uncertainty for the higher contrast input maps; and 5) testing for conditional independence in the input layers using a goodness-of-fit statistic (see Dickson et al. (in review).  In our analysis we found significant relationships between the spatial distribution of ignitions resulting in large fires and four predictor variables; vegetation type, topographic roughness, aspect, and road density.  These variables were used in the final WOE analysis, from which the fire risk layer was developed.  We used the Arc-SDM (Kemp et al. 2001) spatial data modeler extension to ArcView 3.3 (ESRI Corporation) to conduct all weights of evidence calculations and analyses. 

Accuracy Assessment

No accuracy assessment is possible for this layer until a sufficient number of new large fires occur across the analysis area.  A measure of uncertainty based on the variance of the weights for any given location on the landscape (see Bonham-Carter et al. 1989; Dickson et al., in review) is provided in the weights of evidence analysis.  The level of uncertainty is considered significant, and the results of the analysis are suspect for a given location, if the probability of ignition divided by the variance (uncertainty) is < 1.96 at that location (Bonham-Carter et al. 1989).  This analysis suggested that significant levels of uncertainty did not occur on any portion of our landscape.

Sources of errors

This fire risk assessment is a retrospective analysis of large fire ignitions across the landscape.  It is always possible that the factors associated with fire ignitions could change.  For example, changes in weather patterns could alter patterns of ignition caused by lightning, while increasing human development could alter patterns of ignition caused by campfires.  Such changes could alter the pattern of future large fires.  In addition, this analysis combined the effects of human and natural fires.  We did not have quite enough large fires started by human causes to analyze the human and natural fire ignition patterns separately.  This could have some effects on the overall pattern, as large fires from human and natural causes sometimes show different relationships to a single input variable.  For example, large natural fires are more likely to occur in areas with low road densities, while large fires caused by humans are more likely to start in areas of high road density.

Use of this layer

We recommend this layer be used to determine general fire risk in the 1 km² area around each pixel.  Use of this layer to make inferences at spatial extents smaller than 1 km² is not advised. 

There are many challenges associated with determination of fire risk.  Many experts would argue that the ignitions that cause large fires are random.  However, we believe that while we cannot predict which individual ignition will become a large fire, the ignition patterns themselves are nonrandom (e.g. more lighting-associated ignitions occur in areas of rough terrain).  Samples selected randomly from a nonrandom pattern should approximate the nonrandom pattern if sample size is high.  Thus by looking at the landscape-level pattern of large fire ignitions we believe we can predict fire risk across large landscapes, even though we cannot predict exactly which ignitions will lead to a large fire.

We believe this fire risk layer should be used in combination with layers representing fire hazard and wind direction for purposes of analysis, as both ignitions (risks) and fuels (hazards) are important aspect of overall fire threat.  Thus, areas with both high fire risk and high fire hazard might be higher priority than area where one of the other is low.   Likewise, treatments might best be placed in areas downwind of those identified as high fire risk, especially if those areas of high risk are in rough terrain or roadless areas.  Use of the fire risk layer alone to determine priorities or treatments may be less effective than using it in combination with other layers.

Literature Cited

Agterberg, F. P., G. F. Bonham-Carter, Q. Cheng, and D. F. Wright.  1993.  Weights of Evidence Modeling and Weighted Logistic Regression for Mineral Potential Mapping: Pp.13-32 in Computers in Geology -- 25 Years of Progress,(J. C. Davis, and U. C. Herzfeld, eds.).  Oxford University Press, New York.

Bonham-Carter, G. F., F. P. Agterberg, and D. F. Wright.  1988.  Integration of geological datasets for gold exploration in Nova Scotia.  Photogrammetric Engineering and Remote Sensing 54: 1585-1592.

Bonham-Carter, G.F., Agterberg, F.P. and D. F. Wright.  1989.  Weights of evidence modelling: a new approach to mapping mineral potential: Pp. 171-183 in Statistical Applications in the Earth Sciences (F. P. Agterberg and G. F. Bonham-Carter, eds.)  Geological Survey of Canada Paper 89-9.

Dickson, B. G., Y. Xu, J. W. Prather, E. N. Aumack, H. M. Hampton, and T. D. Sisk.  (in review).  Mapping the probability of large fire ignition occurrence in northern Arizona.  Landscape Ecology.

Kemp, L.D., G. F. Bonham-Carter, G. L. Raines, and C. G. Looney.  2001.  Arc-SDM: ArcView extension for spatial data modelling using weights of evidence, logistic regression, fuzzy logic and neural network analysis.

Lyford, M. E., S. T. Jackson, J. L. Betancourt, and S. T. Gray.  2003.  Influence of landscape structure and climate variability on a Late Holocene plant migration. Ecological Monographs 73: 567-583.

Mensing, S. A., R. G. Elston, Jr., G. L. Raines, R. J. Tausch, and C. L. Nowak.  2000.  A GIS model to predict the location of fossil packrat (Neotoma) middens in central Nevada.  Western North American Naturalist 60: 111–120.

Raines, G. L., and M. J. Mihalasky.  2002.  A reconnaissance method for delineation of tracts for regional-scale mineral-resource assessment based on geologic-map data. Natural Resources Research 11: 241-248.

Sampson, R. N., R. D. Atkinson, and J. W. Lewis.  2000.  Mapping Wildfire Hazards and Risks.  The Haworth Press, New York.

Speigelhalter, D. J. and R. P. Knill-Jones.  1984.  Statistical and knowledge-based approaches to clinical decision-support systems, with an application in gastroenterology.  Journal of the Royal Statistical Society A 147: 35-77.

Swetnam, T. W., and J. L. Betancourt.  1998.  Mesoscale disturbance and ecological response to decadal climatic variability in the American Southwest.  Journal of Climate 11: 3128-3147.

Page last updated February 23, 2005