Fire Weather Monitoring and Predictability in the Southeast Corey Davis October 9, 2014 Photo: Pains Bay fire in 2011 (courtesy Donnie Harris, NCFWS)
Outline Fire risk monitoring Fire risk climatology Fire risk predictability Image courtesy NCDENR DAQ
Motivation Wildfires are an increasing concern Tens of thousands of acres burned Smoke affects air quality, public health Response requires substantial resources
Fire Risk Monitoring Discussions with NC Forest Service began about 10 years ago Need: A real-time monitoring tool In fall 2011, work began on the Fire Weather Intelligence Portal Portal completed in summer 2013
Fire Risk Monitoring Past, current, and forecast information Weather data Point-based and gridded datasets Fire risk data National Fire Danger Rating System (NFDRS) parameters output from WIMS database
Fire Weather Intelligence Portal
Fire Weather Intelligence Portal
Fire Weather Intelligence Portal http://nc-climate.ncsu.edu/fwip
Fire Risk Climatology Goal: Examine NFDRS parameters and their variations in VA, NC, SC, GA 18 long-term weather stations 5 subregions
Why Use Fire Risk Parameters? Shown to be closely related to fire activity Represent the ability of the environment to support fires Used operationally by NCFS, USFS, etc. in risk management plans How to measure fire activity?
Fire Risk Parameters Keetch-Byram Drought Index (KBDI) Time-lagged dead fuel moisture 1-hour, 10-hour, 100-hour, 1000-hour Fire behavior parameters Burning Index, Spread Component, Energy Release Component (ERC) KBDI, ERC most useful
Keetch-Byram Drought Index (KBDI) Developed in 1960s by 2 fire scientists at USFS Represents depth of dryness in soil Varies from 0 (no dryness) to 800 Don t need to measure soil moisture Uses daily maximum temperature, daily total precipitation, annual average precipitation
Keetch-Byram Drought Index (KBDI) Sinusoidal annual trend (from Keetch & Byram, 1968)
KBDI Keetch-Byram Drought Index (KBDI) 800 700 600 500 400 300 200 100 Mountains N Piedmont C Piedmont SW Piedmont Coast 0 1/1 3/2 5/2 7/1 8/31 10/30 12/30 Date
KBDI Keetch-Byram Drought Index (KBDI) 800 700 600 500 400 300 200 100 Mountains N Piedmont C Piedmont SW Piedmont Coast 0 1/1 3/2 5/2 7/1 8/31 10/30 12/30 Date
KBDI Departure from Normal (DFN) Compares daily values against a long-term average Can t fit a sine wave for all stations Used a 31-day moving average of historical daily KBDI data Consistent with NOAA methodology for climate normal (Arguez et al., 2012)
Moving Avg. Example (Greensboro) Daily avg. KBDI 31-day moving avg.
Energy Release Component (ERC) Considered a fire behavior parameter ERC accounts for the fuel load that would burn Based on fuel moisture parameters; greatest weighting on 1000-hr fuel moisture Also includes previous 7-day ERC average
ERC Energy Release Component (ERC) 35 30 25 20 15 Mountains N Piedmont C Piedmont SW Piedmont Coast 10 1/1 3/2 5/2 7/1 8/31 10/30 12/30 Date
ERC Classes Describe fire risk with Adjective Rating classes: Low Medium High Very High Extreme Breakpoints set for each station
Percentage of ERC values ERC Classes 50% 40% VHE ERC: associated with fire activity ~80% of the time (Andrews et al., 2003) 30% 20% 10% 0% Low Moderate High Very High Extreme Adjective Rating class
Climatology Summary Among fire risk parameters, KBDI and ERC are the most useful Most commonly used, relate to physical phenomena Fire risk typically highest in the spring and summer Warmer temperatures, more moisture loss Comparisons and interpretation can be difficult KBDI DFN and ERC classes developed
Fire Risk Predictability Goal: Identify any potential climatic indicators of periods of heightened fire risk Does predictability change during the year? Motivated by need for better fire risk guidance 1+ months in advance Outside range of short-term forecasts Regional focus?
Climatic Indicators: Drought Indices Palmer indices: hydrologic accounting PDSI, PMDI, PHDI, Z-Index Standardized Precipitation Index (SPI): statistical calculation of precip. frequency Durations from 1 to 24 months Obtained from NOAA s monthly climate division dataset
Climatic Indicators: Global Climate Pattern Indices El Niño/Southern Oscillation (ENSO) Nino3.4 and ONI North Atlantic Oscillation (NAO) Pacific/North American pattern (PNA) All have known connections to North Carolina s climate
Why Use These Indices? Temperature/precipitation relationships well-studied in past research Climate pattern and drought indices are good summary parameters Easier to translate climatic indicators to fire risk guidance Redmond (2009): some diagnostic information is also (highly) prognostic
Methodology Assess historical strength of relationships on a monthly timescale CLIMATIC INDICATORS: Drought indices Climate pattern indices FIRE RISK PARAMETERS: KBDI, KBDI DFN ERC, ERC class Create statistical models using the best predictors
Initial Statistical Tests Very weak correlations between climate pattern indices and fire risk parameters PDSI, Z-Index, SPI1 & SPI2 had the strongest correlations with KBDI and ERC Summer months with PDSI and SPI1 meant the following month saw a ~25% increase in the likelihood of having abovenormal KBDI
Predictive Models Predict the next month s fire risk using one month s climate pattern or drought indices Three models: Linear Predictive Model Multiple Linear Regression Logistic Regression
Linear Predictive Model Goal: Use one month s PDSI or SPI1 values to predict next month s KBDI DFN or ERC values MODEL MAY 2000 JUNE 2000 Observed PDSI or SPI1 values Predicted fire risk Baseline: Compare performance against the persistence KBDI DFN and ERC values
Next month KBDI DFN KBDI DFN Linear Predictive Model 600 R² = 0.492 400 200 0-200 -400-600 -600-400 -200 0 200 400 600 Previous month KBDI DFN Current month KBDI DFN
Next month ERC ERC Linear Predictive Model 60 R² = 0.340 50 40 30 20 10 0 0 10 20 30 40 50 60 Previous Current month ERC
Linear Predictive Model R² values for all months:
Multiple Linear Regression Goal: Use combinations of climate pattern and drought indices to predict the next month s KBDI DFN or ERC values MODEL MAY 2000 Observed combinations of climatic indicators JUNE 2000 Predicted fire risk MLR assumes stationarity Separate analyses for spring, summer 24 models run for each season
MLR for KBDI DFN in the Summer R² values All Stations Mountains N. Piedmont C. Piedmont SW Piedmont Coast Persistence 0.391 0.629 0.342 0.426 0.384 0.255 MLR clim. patt. 1 (KBDI DFN, NAO, PNA, ONI) 0.387 0.568 0.368 0.442 0.383 0.267 MLR clim. patt. 2 (KBDI DFN, NAO, PNA, MEI) 0.390 0.569 0.373 0.445 0.383 0.267 MLR drought 1 (KBDI DFN, PDSI, SPI1) 0.457 0.613 0.410 0.539 0.464 0.337 MLR drought 2 (KBDI DFN, PDSI, SPI2) 0.430 0.592 0.396 0.515 0.431 0.299 MLR drought 3 (KBDI DFN, PDSI, SPI3) 0.429 0.594 0.396 0.511 0.432 0.295 MLR drought 4 (KBDI DFN, PDSI, SPI6) 0.438 0.605 0.405 0.523 0.443 0.297 MLR drought 5 (KBDI DFN, PDSI, SPI12) 0.444 0.587 0.405 0.532 0.454 0.315 MLR drought 6 (KBDI DFN, Z-Index, SPI1) 0.442 0.611 0.396 0.506 0.453 0.329 MLR drought 7 (KBDI DFN, Z-Index, SPI2) 0.444 0.620 0.392 0.502 0.472 0.320 MLR drought 8 (KBDI DFN, Z-Index, SPI3) 0.438 0.611 0.388 0.501 0.459 0.320 MLR drought 9 (KBDI DFN, Z-Index, SPI6) 0.438 0.610 0.389 0.498 0.458 0.320 MLR drought 10 (KBDI DFN, Z-Index, SPI12) 0.437 0.613 0.386 0.497 0.455 0.320
MLR for ERC in the Spring R² values All Stations Mountains N. Piedmont C. Piedmont SW Piedmont Coast Persistence 0.336 0.390 0.278 0.327 0.338 0.363 MLR clim. patt. 1 (ERC, NAO, PNA, ONI) 0.334 0.336 0.283 0.335 0.347 0.368 MLR clim. patt. 2 (ERC, NAO, PNA, MEI) 0.333 0.341 0.285 0.333 0.345 0.367 MLR drought 1 (ERC, PDSI, SPI1) 0.337 0.334 0.299 0.336 0.353 0.365 MLR drought 2 (ERC, PDSI, SPI2) 0.342 0.339 0.301 0.337 0.364 0.369 MLR drought 3 (ERC, PDSI, SPI3) 0.341 0.331 0.300 0.338 0.357 0.374 MLR drought 4 (ERC, PDSI, SPI6) 0.337 0.336 0.299 0.335 0.352 0.364 MLR drought 5 (ERC, PDSI, SPI12) 0.343 0.334 0.296 0.350 0.360 0.372 MLR drought 6 (ERC, Z- Index, SPI1) 0.327 0.330 0.277 0.327 0.339 0.362 MLR drought 7 (ERC, Z- Index, SPI2) 0.328 0.330 0.275 0.326 0.339 0.372 MLR drought 8 (ERC, Z- Index, SPI3) 0.329 0.331 0.276 0.326 0.339 0.373 MLR drought 9 (ERC, Z- Index, SPI6) 0.330 0.332 0.279 0.329 0.347 0.365 MLR drought 10 (ERC, Z- Index, SPI12) 0.328 0.329 0.281 0.327 0.338 0.367
Logistic Regression Goal: Use monthly climate pattern and drought indices to predict whether the next month will have at least one VHE ERC day MODEL MAY 2000 Observed index JUNE 2000 Predicted occurrence of VHE ERC days (0 or 1) Why? Highest 10% of ERC values are associated with fire activity 80% of the time (Andrews et al., 2003)
Logistic Regression with PDSI
Logistic Regression with SPI1
Logistic Regression with Z-Index
Logistic Regression Threshold values Threshold PDSI value for 75% probability of VHE ERC Threshold Z-Index value for 75% probability of VHE ERC Threshold SPI1 value for 75% probability of VHE ERC All Stations -3.35-2.89-1.63 Mountains -2.73-2.44-1.28 N. Piedmont -2.48-2.52-1.50 C. Piedmont -3.80-2.91-1.62 SW Piedmont -3.59-3.21-1.84 Coast -4.04-3.02-1.69
Predictability Summary PDSI, short-duration SPI are most closely associated with high fire risk Climate patterns offer limited guidance Persistence fire risk is a reasonable estimate Threshold PDSI, SPI1, Z-Index values could signal months with high ERC levels
Applications Fire risk guidance should be accessible, understandable, accurate, and meaningful Possible options: Threshold-based alerts? Probabilistic forecasts? Standard training?
Remaining Questions Are any non-nfdrs parameters good indicators of fire risk? What is a good indicator of fire risk in organic soils? Can we get high-resolution calculations/climatology of KBDI or ERC? Could we also predict periods with low fire risk?
Questions or Comments? Fire Weather Intelligence Portal: http://nc-climate.ncsu.edu/fwip Corey Davis cndavis@ncsu.edu (919) 515-3056