Wildfires in Florida David T. Butry US Forest Service Marcia L. Gumpertz North Carolina State University Marc G. Genton Texas A&M University Funding provided by USDA Forest Service
Study Site:SJRWMD • From 1996-2001: • 7249 small wildfires • Arson—25% • Accidents—43% • Lightning—32% • 53 large wildfires • Arson—17% • Accidents—28% • Lightning—55%
Large Wildfires Small Wildfires
Wildfire Models • Wildfire modeled as a function of: • Fuel type, fire cause, time, location • Climate and weather at time of ignition • Prescribed burning, time to respond to fire • Landscape attributes, previous fire • Three models estimated: • Area burned. Fires 1000 acres or less • Area burned. Fires >1000 acres • Probability that a fire becomes larger than 1000 acres.
Landscape Characteristics (XS): • Population density • Income • Percent of population who have attended college • Amount of road in section • Distance to nearest fire department • Landscape landcover/landuse • Latitude & longitude • Fire district • Previous wildfire • Includes current and lagged values, as well as neighboring current and lagged values • GIS “hole” • Fire Characteristics (XF): • Start time • Start year • Cause • Climate/Weather (XC): • Niño3 • KBDI • Spread Index • Buildup Index • Humidity • Wind speed • Wind direction • Management (XM): • Response time • Limited action fires • Prescribed fire • Includes current and lagged values, as well as neighboring current and lagged values
Spatial and Temporal Scales of Data • Individual fire • Location: centroid of section • Time of ignition Spatial and Temporal Variables in Models • Latitude and longitude (Albers projection) • Year • Time of day (morning, afternoon, evening, overnight)
Lagged Variables • Previous wildfire • Section: same or neighboring • Year: earlier in same year; or in previous 12 years • Prescribed burning prior to the fire • For hazard reduction: • Section: same or neighboring • Year: earlier in same year; or in previous 3 years • For silvicultural or other purposes • Section: same or neighboring • Any time in this year or previous 3 years
Model Specification Area burned: Ln(Wildfire Size) = X + , X may include lagged variables Probability of Large Fire:
Results—Small Wildfires, n=7249, R2=.19 Factors related to greater area burned Response time up to 16 hours Limited-action fires (let burn) Spread index Palmetto-Gallberry, Grass, and Pine fuel types Arson ignitions (as opposed to lightning) Afternoon ignitions (as opposed to overnight) Area burned by wildfire in neighboring sections in the previous 1-12 years Previous non-hazard mitigating prescribed fire lagged up to 3 years Factors related to smaller area burned La Niña Humidity Hazard-mitigating prescribed burning in the current year Population density, up to about 70 persons per square kilometer (which is the median) Fire districts 10, 14, & 16 Smaller fires in 1999, 2000, 2001 % of water and wetlands in section % of grasslands and upland forests in neighboring sections Interactions Several with KBDI
Results—Large Wildfires , n=53, R2=.73 Factors Related to Greater Area Burned Spread index Wildfire in previous 12 years Percent of wetland in the section Population density Income 1998 Factors Related to Smaller Area Burned Fuel type grass Arson and accidental ignitions, as opposed to lightning Fire District 12 Interactions In GIS hole areas, percent of water in the section is positively related to large fire size
Some Factors Associated with Area of Fire, Fires > 1000 Acres
Results—Probability that Area Burned will Exceed 1000 Acresn=7302, max rescaled R2=.32 Positive Relation La Niña Income Build-up Limited-action fires (let burn) Wind speed 1998 Neighboring upland forest Negative Relation El Niño Upland forest Neighboring Urban areas Interactions Log odds increases with KBDI in upland forest sections but not in other types of land cover. Log odds decreases with hazard-reducing prescribed burning if response time is short, but not if response time is longer. If the response time is long, the log odds decreases as hazard-reducing prescribed burning increases in neighboring sections.
Summary • Lagged variables in space and time • capture much of the spatial/temporal • information, but not quite all of it. • Models don’t fit or predict nearly as well • as I’d like. There’s more to be done. • Some encouraging results regarding • ability to detect effects of management • practices in section and neighbors. • Continuing work on estimating magnitude of • effect of prescribed burning -- using “propensity scores”.