Haiganoush K. Preisler PSW, Albany A collaborative project with: Alan Ager – Threats Assessment Center, PNW, Oregon Steve Taylor - Pacific Forestry Centre, Victoria, Canada Barbara Bentz - RMRS, Logan UT Jeff Hicke - University of Idaho Anthony Westerling - UC Merced PSW Workshop on Climate Change IFG, Placerville, CA April 30 – May 1 2008 Motivation • Greenhouse gas balance of forests appear to be strongly affected by naturally occurring fires and by cyclical insect outbreaks1. • Fires cause direct emissions. • Insect outbreaks reduce growth rates; kill trees, resulting in transfer of carbon from biomass to dead organic matter. • Methods are needed for forecasting these stochastic disturbances in the presence of changing climate and unpredictable weather systems. 1 ‘Mountain pine beetle and forest carbon feedback to climate change’ Kurz et al. Nature April 2008 Biologically based statistical models are needed to 1. Quantify relationships and uncertainties. 2. Assessing goodness-of-fit. 3. Appraise skill of forecasting model. 4. Forecast fire and pest outbreak intensities with known accuracies and confidence bounds. 5. Appraise skill of process-based simulation models. Estimates are needed for two probabilities 1. Probability{ Ai,j(t) =1 | Hj(t-1), Xj(t-1), Vj } 2. Probability{ Bi,j(t) =1 | Hj (t-1), Xj(t-1), Vj } where Ai,j(t) = fire of size class i at location j and time t. Bi,j(t) = insect infestation of size i at location j and time t. Hj(t) = history of fire and insect at location j. Xj(t) = history of climate and weather at location j. Vj = topography and vegetation at location j. Historic Data (1980-present) • Weather stations ÎPrism Îgridded weather • Gridded products ÎVIC1 Îgridded hydrological data (e.g. moisture deficit) • Fire, insect, vegetation, topography • Weather data ÎBioSIM2 (adaptive seasonality) Îsuitable habitats for insect outbreaks • LANSAT fire data + Arial insect survey maps • Ozone data 1 Script Institute of Oceanorgaphy 2 Service Canadien des Forêts Proposed Plan of Work Historic fire and insect data Statistical Model VIC Forecasts/scenarios for weather, climate Estimated probabilities Forecasts of fire and insect damage BioSIM Assess fit Weather, climate, vegetation, topography, ozone and nitrogen data Insect life cycle and development info Forecasts of Greenhouse Gas emissions In the following slides we present work from two pilot studies demonstrating some of the proposed techniques Study 1: Data on three insect species from Oregon and Washington are used to estimate effects of climate and other variables on probabilities of infestation. Study 2: Data on large federal fires in California are used to forecast one-month-ahead fire probabilities with the help of some fire weather indices. Western Pine Beetle - Forest Service lands Study 1: Data Yearly sum (sqrt-scale) 30 25 Number of 1km2 pixels with more than 10 trees killed by the beetle 20 15 10 5 1985 1990 1995 Year Spatial distribution of infestation estimated by historic (1981 -2005) probability of a pixel having > 10 trees killed by beetle. 2000 2005 Moutain Pine Beetle - Forest Service lands Study 1: Data Yearly sum (sqrt-scale) 80 70 Number of 1km2 pixels with more than 10 trees killed by the beetle 60 50 40 30 1985 1990 1995 Year Spatial distribution of infestation estimated by historic (1981 -2005) probability of a pixel having > 10 trees killed by beetle. 2000 2005 Spruce Beetle - Forest Service lands Study 1: Data Yearly sum (sqrt-scale) 20 Number of 1km2 pixels with more than 10 trees killed by the beetle 15 10 5 1985 1990 1995 Year Spatial distribution of infestation estimated by historic (1981 -2005) probability of a pixel having > 10 trees killed by beetle. 2000 2005 Two climate variables examples Average tem perature July 1972 Study 1: Data M oisture Deficit July 1972 Climate Research Division Scripps Institution of Oceanography La Jolla, California Study 1: Data Western North American Large Fire History (fires greater than 200 hectares) Study 1: Model Statistical Model Yij = 1 if more than 10 trees killed by insects at location i (grid size 1x1km) year j = 0 otherwise Yij ~ Binomial{1, π(θij )} θ ij = linear predictor = α + ∑ g m ( X mij ) + g ( year j ) + h (lon ij , lat ij ) m Xmij = value of mth explanatory variable at location i year j g( ) and h( ) ~ smoothing functions e.g. basis splines and thin plate splines Study 1: Results Preliminary Results Beetle MPB WPB Spruce B Infestation (last year) 2.3 – 2.8 times Increase in odds 1.9 – 2.6 times Increase in odds 2.9 – 4.9 times Increase in odds Drought history Previous 2 years moisture deficit Previous 2 years moisture deficit Previous year moisture deficit Temperature Winter temperature variability Not significant Average temperature Fire history Not significant Not significant 2.0 – 4.5 times increase in odds if fire 2 years ago Explanatory Study 2 • A statistical model that estimates relationships between observed climate; simulated hydrology and observed large wildfire incidence. • the hydrology is simulated with the VIC hydrologic model using observed climate. Study 2: Data • Location and size of fires in California Federal lands • Moisture Deficit* observed prior to the date the forecast is made • Surface Temperature observed prior to the date the forecast is made • Percent Forest Cover in a grid cell • Location and Month of year *Potential - Actual Evapotranspiration simulated with the VIC hydrologic model Study 2: Model Statistical Model Yij = k if fire of size class k at location i year j Yij ~ Multinomial {1, πk(θij )} k=1, 2, 3, 4, 5 with π1 + … + π5 = 1 π k = exp(θ k )/ ∑ exp(θ k ) k θ ijk = linear predictor = α k + ∑ g m ( X mij ) + g ( yearj ) + h(lonij , latij ) m Xmij = value of mth explanatory variable at location i year j g( ) and h( ) ~ smoothing functions e.g. basis splines and thin plate splines Study 2: Model Skill One month ahead forecasts 5 Observed Forecasted 4 Precentage Large fires (200ha) on California Federal lands 1985-2003 ~95% CL 3 2 1 0 -8 -7 -6 -5 Linear predictor -4 -3 Study 2: Model Skill In the following slides • We map the forecasted probability of a large fire (>200ha) on a 1/8 degree grid over California. • Superimposed on this we first show simulated fires: we use the forecasted probabilities to generate random fires. We show 10 simulations for selected dates/examples. • For each date we show a final map of forecasted probabilities and the observed large fire locations • The selected examples are August 1994 (high fire year), August 1995 (low fire year) and August 1987 (high fire year with clustered lightning ignitions) August 1994 forecasts 8.0 7.0 6.0 5.0 4.0 3.0 2.0 1.0 0.0 Map of Probabilities (%) of > 200ha burns August 1994 forecasts 8.0 7.0 6.0 5.0 4.0 3.0 2.0 1.0 0.0 Map of Probabilities (%) of > 200ha burns Red dots=model forecasted pixels with >200ha burn area August 1994 forecasts 8.0 7.0 6.0 5.0 4.0 3.0 2.0 1.0 0.0 Map of Probabilities (%) of > 200ha burns Red dots=model forecasted pixels with >200ha burn area August 1994 forecasts 8.0 7.0 6.0 5.0 4.0 3.0 2.0 1.0 0.0 Map of Probabilities (%) of > 200ha burns Red dots=model forecasted pixels with >200ha burn area August 1994 forecasts 8.0 7.0 6.0 5.0 4.0 3.0 2.0 1.0 0.0 Map of Probabilities (%) of > 200ha burns Red dots=model forecasted pixels with >200ha burn area August 1994 forecasts 8.0 7.0 6.0 5.0 4.0 3.0 2.0 1.0 0.0 Map of Probabilities (%) of > 200ha burns Red dots=model forecasted pixels with >200ha burn area August 1994 forecasts 8.0 7.0 6.0 5.0 4.0 3.0 2.0 1.0 0.0 Map of Probabilities (%) of > 200ha burns Red dots=model forecasted pixels with >200ha burn area August 1994 forecasts 8.0 7.0 6.0 5.0 4.0 3.0 2.0 1.0 0.0 Map of Probabilities (%) of > 200ha burns Red dots=model forecasted pixels with >200ha burn area August 1994 observed 8.0 7.0 6.0 5.0 4.0 3.0 2.0 1.0 0.0 Map of Probabilities (%) of > 200ha burn areas Red dots are observed pixels with >200ha burns August 1995 forecasts 8.0 7.0 6.0 5.0 4.0 3.0 2.0 1.0 0.0 Map of Probabilities (%) of > 200ha burns Red dots=model forecasted pixels with >200ha burn area August 1995 forecasts 8.0 7.0 6.0 5.0 4.0 3.0 2.0 1.0 0.0 Map of Probabilities (%) of > 200ha burns Red dots=model forecasted pixels with >200ha burn area August 1995 forecasts 8.0 7.0 6.0 5.0 4.0 3.0 2.0 1.0 0.0 Map of Probabilities (%) of > 200ha burns Red dots=model forecasted pixels with >200ha burn area August 1995 forecasts 8.0 7.0 6.0 5.0 4.0 3.0 2.0 1.0 0.0 Map of Probabilities (%) of > 200ha burns Red dots=model forecasted pixels with >200ha burn area August 1995 observed 8.0 7.0 6.0 5.0 4.0 3.0 2.0 1.0 0.0 Map of Probabilities (%) of > 200ha burn areas Red dots are observed pixels with >200ha burns August 1987 forecasts 8.0 7.0 6.0 5.0 4.0 3.0 2.0 1.0 0.0 Map of Probabilities (%) of > 200ha burns Red dots=model forecasted pixels with >200ha burn area August 1987 forecasts 8.0 7.0 6.0 5.0 4.0 3.0 2.0 1.0 0.0 Map of Probabilities (%) of > 200ha burns Red dots=model forecasted pixels with >200ha burn area August 1987 forecasts 8.0 7.0 6.0 5.0 4.0 3.0 2.0 1.0 0.0 Map of Probabilities (%) of > 200ha burns Red dots=model forecasted pixels with >200ha burn area August 1987 forecasts 8.0 7.0 6.0 5.0 4.0 3.0 2.0 1.0 0.0 Map of Probabilities (%) of > 200ha burns Red dots=model forecasted pixels with >200ha burn area August 1987 observed 8.0 7.0 6.0 5.0 4.0 3.0 2.0 1.0 0.0 Map of Probabilities (%) of > 200ha burn areas Red dots are observed pixels with >200ha burns Study 2: Conclusions • The model with the climate variables appears to have useful forecast skill • The pattern of large wildfire forecasts as shown in the simulations appear similar to the observed pattern of historic wildfires • This particular model cannot predict ‘clusters’ of ignitions like those observed in 1987 • We showed forecasts for one fire size class, but we are modeling probabilities for multiple size classes • Future work will revise this model to generate scenarios based on historical patterns of lighting ignitions Some relevant references • • • • • • • • • Bentz, B.J., Logan, J.A., and Amman, G.D. 1991. Temperature dependent development of mountain pine beetle and simulation of its phenology. Canadian Entomologist 123: 1083-1094. Hicke, J.A., Logan, J.A., Powell, J., and Ojima, D.S. 2006. Changing temperatures influence suitability for modeled mountain pine beetle (Dendroctonus ponderosae) outbreaks in the western United States. Journal of Geophysical Research-Biogeosciences 111: G02019, doi:02010.01029/02005JG000101. Kruz, W.A., Dymond, C.C., Stinson, G., Rampley, G.J., Neilson, E.T., Carroll, A.L., Ebata, T. and Safranyik,L. 2008. ‘Mountain pine beetle and forest carbon feedback to climate change’ Nature, Vol 452124, April 2008. doi:10.1038/nature06777. Preisler,H.K., Westerling, A.L. 2007. "Statistical model for forecasting monthly large wildfire events in western United States". Journal of Applied Meteorology and Climatology 46, 1020-1030. Preisler,H.K., Ager, A.A. and Hayes, J.L. (In press). "Probabilistic risk models for multiple disturbances: An example of forest insects and wildfires". Encyclopedia of Forest Threats. Preisler,H.K.,Chen, S.C. Fujioka, F., Benoit, J.W. and Westerling, A.L. (In press). "Wildland fire probabilities estimated from weather model-deduced monthly mean fire danger indices". International Journal of Wildland Fire. Preisler, H.K., D.R. Brillinger, R.E. Burgan, and J.W. Benoit. (2004) Probability based models for estimation of wildfire risk. Journal of Wildland Fire, 13, 133-142. PDF. Régnière, J., Cooke, B., and Bergeron, V. 1996. BioSIM: a computer-based decision support tool for seasonal planning of pest management activities; User's Manual. Canadian Forest Service. Information Report LAU-X-155. Westerling, A.L., H.G. Hidalgo, D.R. Cayan, T.W. Swetnam 2006: "Warming and Earlier Spring Increases Western U.S. Forest Wildfire Activity" Science, 313: 940-943. DOI:10.1126/science.1128834