Seasonal Fire Prediction for Alaska

1
Seasonal Fire Prediction for Alaska
Alaska Center for Climate Assessment and Policy

• Dan White, University of Alaska, Fairbanks
• Paul A. Duffy, Neptune Inc.
• Sarah Trainor, University of Alaska, Fairbanks

2
Targeted Collaborating Stakeholders

• State of Alaska, Department of Natural Resources
• Tanana Chiefs Conference
• U.S. Fish and Wildlife Service
• Bureau of Land Management, Alaska Fire Service
• Bureau of Indian Affairs
• National Park Service
• U.S. Forest Service
• National Interagency Fire Center

3
Why is Fire Important in Alaska?

• Dominates
• the
• disturbance
• regime

• Succession
• modifies
• forest
• structure

4
(No Transcript)
5
(No Transcript)
6
(No Transcript)
7
Why is Fire Important in Alaska?

• Interior Alaska contains 60 million burnable
  hectares (approx. equal to MT and ID combined)
• Average annual area burned is 340,000 ha median
  is 135,000 ha
• Largest year burned 2.6 million ha, which is
  about 14 of the area in Oklahoma

8
Fire perimeters from 1950-2005
9
CLIMATE
1
2
What are the relevant spatial and temporal scales?
VEGETATION
FIRE
3
10
CLIMATE
Obvious link between climate/weather and
fire Spatial and temporal scales of
interest. not so obvious
2
FIRE
11
Annual Area Burned in Alaska from 1950-2003
12
Statistical Model Development

• Response log(Annual Area Burned)
• 7 Explanatory Variables
• Monthly temperatures (April, May, June, July) and
  precipitation (June)
• Teleconnection indices PDO and East Pacific
• R-squared for the model is 0.79

13
Data Sources

• Area Burned Bureau of Land Management
• Monthly temperatures and precipitation Western
  Region Climate Center
• Pacific Decadal Oscillation Joint Institute for
  the Study of the Atmosphere and Ocean
• Teleconnection indices National Oceanic and
  Atmospheric Administration -Climate Prediction
  Center

14
Historical Fire perimeters are from the LFDB
15
Observed
Estimated
Duffy et al (2005)
16
Fire and Climate

• Monthly weather/teleconnection indices drive
  annual variability in statewide area burned
• Can this model be used to forecast?

17
Forecasting in 2004

• May 2004 estimate 135,000 ha
• 70th percentile May, June, July temperature
• 50th percentile June precipitation

18
Forecasting in 2004

• May 2004 estimate 135,000 ha
• 70th percentile May, June, July temperature
• 50th percentile June precipitation
• Actual area burned 2004 2,300,000 ha

19
Forecasting in 2004

• May 2004 estimate 135,000 ha
• 70th percentile May, June, July temperature
• 50th percentile June precipitation
• Actual area burned 2004 2,300,000 ha
• Fall 2004 estimate 2,100,000 ha
• Actual data (2nd hottest June, etc.)

20
Several Issues with Prediction

• Assumptions of the linear model framework can be
  restrictive
• Need a model that does not rely on in-season
  data

21
Observed vs. GBM estimated Area Burned in Alaska
Observed
Estimated
Gradient Boosting Model (GBM) Friedman (2001)
22
Partial Dependence Plots for GBM model
August Precipitation
June Temperature
Precipitation (mm)
Temperature (C)
Vertical axis shows expected hectares as a
function of the explanatory variable
23
Building Predictive Models

• Next step is to apply GBM approach using
  pre-season variables
• We know teleconnection indices influence area
  burned
• Construct GBM model with information from several
  different teleconnection indices

24
Building Predictive Models

• Use a stepwise procedure to select the
  teleconnections to be used for explanatory
  variables
• Currently, this process is performed monthly for
  March through June
• Data are available at the end of each month

25
Building Predictive Models

• Use a stepwise procedure to select the
  teleconnections to be used for explanatory
  variables
• East Pacific/North Pacific (Jan, April avg)
• Polar (Jan, Feb avg)
• April Temperature
• January Precipitation

End of April Model
26
80 Uncertainty Intervals of Cross-Validated
Predictions
Cross-Validation performed by re-fitting the
model 5000 times, each time eliminating 26 years
of data
27
Prediction(s) for the 2008 Season (April data)
28
Error Table for Predictions Based on April Data
lt 500,000 ha
gt 1,000,000 ha
29
Prototype Interactive Web-Tool
For the Base layer, choose between a Google
Physical Map or a Hybrid Satellite Image Map with
roads and place names
Choose the resolution and position on the map
30
Prototype Interactive Web-Tool
Choose to add monthly Fire Forecasts and other
layers that are currently in development,
including Fire History and Management Options
31
Conclusions

• Annual area burned in Alaska is strongly driven
  by climatic factors
• This link can be used to generate forecasts
• Experimental Forecasts will be available monthly
  starting the first week of April at
  www.uaf.edu/accap

32
Acknowledgements

• Alaska Wildland Fire Coordination Group
• NOAA CPO and CPC
• Scenarios Network for Alaska Planning
• CLIMAS and WWA
• For more information contact Sarah Trainor,
  sarah.trainor_at_alaska.edu 907-474-7878