The purpose of this experimental forecast is to provide managers with a forecast of the area burned in Interior Alaska for the upcoming fire season.
Make predictions about what time of the season fires will occur
Predictive capacity for Alaska fire falls behind what is available in the lower 48 states. Increases in wildfire frequency, severity, duration, and total area burned are among the most significant expected ecological effects of climate warming. Two of the three most extensive wildfire seasons in Alaska’s 50-year record occurred in 2004 and 2005 and 60% of the largest fire years have occurred since 1990 (Kasischke et al. 2006).
In 2004, the largest fire season on record in Alaska, over 2.5 million hectares burned, costing state and federal fire agencies nearly $150 million. A Fairbanks neighborhood was evacuated multiple times and air quality in Fairbanks was classified as hazardous or unhealthy for nearly one quarter of the fire season. Population growth, road-building and resource development are increasing the need for fire suppression by expanding the area of wildland-urban interface. Furthermore, increased fire activity in Alaska increases nation-wide competition for limited and shared fire fighting resources.
Designed in close collaboration with fire managers from a range of state and federal agencies participating in the Alaska Wildland Fire Coordination Group, this project takes advantage of the strong weather/fire link in Alaska to produce estimates for the severity of the 2009 and 2010 fire seasons. The regression model developed by Duffy et al. (2005) estimates the logarithm of annual area burned as a function of monthly weather and teleconnection indices with an R-squared of greater than 75%. We extend this modeling framework through the application of gradient boosting models (GBM). Preliminary results show significant improvement over the already high R-squared from the regression model. The uncertainty associated with the forecasts will be quantified resulting in a set of possible values for area burned in Alaska and confidence intervals for the forecast.
These results will provide a web-based decision-support tool that will help Alaska fire mangers adapt to a changing climate in their suppression and natural resource planning.
Methods
The general approach can be simply stated as "sequentially fit a predictive model for annual area burned in Alaska after the data for each month (March-July) become available". As with any modeling process there are a number of different decisions that must be made regarding model complexity, selection of explanatory variables, spatio-temporal resolution of interest, and others. A brief outline of the specifics of the approach is presented below, with an example for the data available at the end of April 2008.
One of the key results of Duffy et al. (2005) was the identification of the approximate spatial and temporal resolution that displays the strongest linkage between climate and fire. This work keeps the focus on linkages at an annual timescale across interior Alaska for the time period of 1950-2008. Explanatory variables used in this analysis are monthly teleconnection indices and monthly temperature/precipitation. Predictions are made monthly from March through July. The collection of potential explanatory variables includes monthly values for the following: the arctic oscillation; the east Pacific/North Pacific teleconnection; the Polar teleconnection; the West Pacific teleconnection; and average temperature and total precipitation. The teleconnection data are available from two NOAA sites (Climate Prediction Center (CPC) Climate and Weather Linkages website and the CPC Teleconnection website). The temperature and precipitation data are assembled following the methods of Duffy et al. (2005) and are available from the Western Region Climate Center.
Details regarding the specifics of the software used to implement this approach can be found at The R Project. A detailed description of the generalized boosting model approach is beyond the scope of this document. Conceptually, the goal of the GBM approach is to find the 'combination' of explanatory variables that best characterizes the annual area burned. In a regression model this is done by using the data to find the best parameters in a linear model. In the GBM approach, regression trees are constructed through binary recursive partitioning.
For more information on how the explanitory variables were selected and a discussion on uncertainty, see the Experimental Forecast of Area Burned for Interior Alaska website.
References
Duffy, P. A., Walsh, J.E., Graham, J.M., Mann, D.H., Rupp, T.S.(2005) Impacts of large-scale atmospheric-ocean variability on Alaskan fire season severity. Ecological Applications 15(4):1317-1330.
Friedman, J.H. (2002) Stochastic Gradient Boosting. Computational Statistics and Data Analysis 38:367-378.
Friedman, J.H. (2001) Greedy Function Approximation: A gradient Boosting Machine. The Annals of Statistics 29:5
1189-1232.
