Predicting Fire Season Severity in the Pacific Northwest Paul Werth1 Abstract—Projections of fire season severity that integrate historical weather and fire information can be used by fire managers when making decisions about allocating and prioritizing firefighting resources. They enable fire managers to anticipate fire activity and pre-position resources to maximize public and firefighter safety, reduce environmental impacts, and lower firefighting costs. This research determines the potential severity of fire seasons in the Pacific Northwest by using statistical techniques that correlate weather data and annual-acreage-burned figures for five fire management agencies in Washington and Oregon (U.S. Forest Service, Bureau of Land Management, Bureau of Indian Affairs, Oregon Department of Forestry, and Washington Department of Natural Resources). Weather and fire trends for the 1970 to 2004 time period were calculated, and thresholds for above average, average, or below average fire seasons were determined based upon annual acres burned. Eight weather parameters were then correlated using scatter diagrams, contingency tables, and multivariate regression equations to predict above average, average, or below average fire seasons based upon projected acres burned. Results show considerable variance in predictors by fire agency with accuracy rates of 60 to 85% for predictions of above average fire seasons and 85 to 90% for average and below average fire seasons. Introduction Several considerations affect fi re managers’ decisions regarding allocation of firefighting resources including: (1) public and firefighter safety (2) the potential effect of fires on local environments, and (3) the increasing impact of fi refighting costs on agency budgets. Over the past several years, the Northwest Interagency Coordination Center has demonstrated that pre-positioning resources throughout Washington and Oregon in advance of fi re outbreaks, improves their effectiveness in achieving all three of the above-listed goals. The obvious question arises, “How do fi re managers determine the most effective placement of resources prior to the fi re season?” One tool they use is a pre-season assessment of historical weather and fi re information that produces projections of expected fi re season severity for any given area in the Pacific Northwest. This research takes that assessment to the next level by applying statistical techniques to weather and annual acres-burned data for five, fi re management agencies in Oregon and Washington, including the U.S. Forest Service, Bureau of Land Management, Bureau of Indian Affairs, Oregon Department of Forestry, and Washington Department of Natural Resources. USDA Forest Service Proceedings RMRS-P-41. 2006. In: Andrews, Patricia L.; Butler, Bret W., comps. 2006. Fuels Management—How to Measure Success: Conference Proceedings. 28-30 March 2006; Portland, OR. Proceedings RMRS-P-41. Fort Collins, CO: U.S. Department of Agriculture, Forest Service, Rocky Mountain Research Station. 1 Weather Research and Consulting S er v ice s , L L C , B at t le G rou nd , Washington. pwerth@prodigy.net 757 Werth Predicting Fire Season Severity in the Pacific Northwest Data The assumption that wildland fi re severity is primarily driven by low fuel moisture has historically directed research towards drought (Westerling and others 2002; Hall and Brown 2003) as a pre-season, predictor of fi re season severity. This research also uses drought, but expands the list of potential predictors to include: seasonal precipitation, mountain snowpack, snowmelt date, and the sea surface temperature of the Pacific Ocean. Monthly precipitation figures for seven weather stations in Washington and Oregon were used in this analysis. The seven stations used were Medford, Portland, Redmond, Burns and Pendleton in Oregon, and Yakima and Spokane in Washington (fig. 1) They were selected based on their location near fi re-prone areas and completeness of record since 1970. Monthly precipitation data was divided into four groups: (1) winter (November-March), (2) spring (April-May), (3) June, and (4) summer (July-August). June is a group by itself because precipitation during the month of June can significantly impact the duration of significant fi re danger. Snow pack water equivalency (SWE) data for the Columbia River Basin of Washington, Oregon, Idaho and portions of British Columbia, Montana, and Wyoming was also used in this analysis. The April 1 SWE is of particular importance because the snowpack typically peaks around April 1st. SWE figures for May 1st were used to determine the rate of spring snowmelt in the mountains. SWE data was used to track the annual snowmelt date at 39 Natural Resources Conservation Service (NRCS) SNOTEL sites in Washington and Oregon from 1986 to 2005 (fig. 2). These sites represent every major river basin and different elevations within Washington and Oregon. Historic Palmer Drought Severity Index (PDSI) values for climate zones in Washington and Oregon were collected from the National Climatic Data Center (NCDC) database. Average March values for each state along with the number of climate zones classified in moderate drought were used in this research. Monitoring sea surface temperature anomalies in the central Pacific Ocean is essential in determining the phases of the El Niño / Southern Oscillation (ENSO). The warm phase, commonly called El Niño, is characterized by abnormally warm sea surface temperatures in the central and eastern equatorial Pacific Ocean. The cool phase of this natural cycle is called La Niña. El Niño often results in warm, dry winters and below normal snow packs in the Pacific Northwest. La Niña has the opposite effect, producing cool, wet winters and above average snow packs. Both phases appear to have minimal effect on summer weather in the Pacific Northwest. The Multivariate ENSO Index (MEI) combines six variables (sea-level pressure, zonal and meridianal components of the surface wind, sea surface temperature, surface air temperature, and total cloudiness fraction of the sky) to monitor ENSO. Negative values of the MEI represent the La Niña phase while positive values indicate El Niño. Bi-monthly values of MEI were retrieved from the NOAA-CIRES Climate Diagnostics Center in Boulder, Colorado. The Eastern North Pacific (ENP) (fig. 3) sea surface temperature index is a component of the Pacific (P) index (Castro, McKee, and Pielke 2001) that combines tropical and North Pacific SSTs into one index. The P index has been correlated with upper-level atmospheric circulation patterns over the North Pacific Ocean and the Western and Central United States. It has also been correlated to the onset of the Southwest Monsoon and precipitation anomalies in the Great Plains states. Data to compute the ENP was downloaded from the Comprehensive Ocean Atmospheric Dataset (COADS). 758 USDA Forest Service Proceedings RMRS-P-41. 2006. Predicting Fire Season Severity in the Pacific Northwest Figure 1—Seasonal Precipitation Stations. Werth Figure 2—WA and OR SNOTEL Stations. Figure 3—ENP and NINO 3.4 Pacific SST Regions. USDA Forest Service Proceedings RMRS-P-41. 2006. 759 Werth Predicting Fire Season Severity in the Pacific Northwest Area-burned figures (in acres) for federal- and state-protected land in Washington and Oregon was obtained from the Oregon Department of Forestry (ODF), Washington Department of Natural Resources (DNR), and the Northwest Interagency Coordination Center annual summaries dating back to 1970. The acres burned statistics include both lightning and human-caused fi res. Weather and Area Burned Trends The fi rst step in determining the significance of seasonal precipitation on fi re season severity in the Pacific Northwest is to determine whether there are long-term trends in both weather and fi re data. This was accomplished by constructing time lines for each dataset and then performing a regression analysis to determine whether there are identifiable trends in the data. Linear regression equations were developed for each data set in the form of: y = mx + b. The equation algebraically describes a straight line for a set of data with (x) the independent variable, (y) the dependent variable, (m) the slope of the line, and (b) the y-intercept. The sign (+ or –) and magnitude of m signify whether the independent variable is increasing or decreasing and at what rate. Regression analysis indicates decreasing winter rainfall (November-March) and Columbia River Basin April 1 SWE since 1970 (figs. 4 and 5). The decrease is more apparent in SWE, indicating warmer winter temperatures are also a contributor in addition to decreased precipitation. However, the trend in spring rainfall (April and May) is for wetter conditions (fig. 6). Rainfall amounts for July and August also show a trend toward drier weather during the summer in the Pacific Northwest (fig. 7). Similar regression techniques were used to establish trends in acres burned for federal and state land management agencies in Washington and Oregon. All agencies trend toward more acres burned per year, especially since the mid-1980s. This is most evident in the U.S. Forest Service data (fig. 8), which shows the largest trend in acres burned of all the agencies. Defining Fire Season Severity Defi ning fi re season severity is a difficult question, one that may have many answers. Some base it on the total number of fi res or the number of days in high to extreme fi re danger; others use the number of large fi res during the year. In order to predict fi re season severity, one must fi rst defi ne it. The standard used in this research is the annual acres burned by fi re agency. The dataset includes thirty-five years of annual acres burned by agency from 1970 to 2004. Data was sorted by agency and by year from the highest to the least number of acres burned. Data was then divided into thirds, or terciles. Years in the top tercile, (i.e., those with the largest number of acres burned,) were classified as “Above Average” fi re seasons. Years in the middle tercile were classified as “Average” fi re seasons, and years in the bottom third were classified as “Below Average” fi re seasons. This classification was performed for each of the five federal and state fi re agencies. Threshold acres were identified for each category as displayed in this graph for the Bureau of Land Management (BLM) (fig. 9). 760 USDA Forest Service Proceedings RMRS-P-41. 2006. Predicting Fire Season Severity in the Pacific Northwest Werth Figure 4—Winter Precipitation Trend. Figure 5—April 1 Columbia Basin Snowpack Trend. USDA Forest Service Proceedings RMRS-P-41. 2006. 761 Werth Predicting Fire Season Severity in the Pacific Northwest Figure 6—Spring Rainfall Trend. Figure 7—Summer Rainfall Trend. 762 USDA Forest Service Proceedings RMRS-P-41. 2006. Predicting Fire Season Severity in the Pacific Northwest Werth Figure 8—USFS Acres Burned Trend. Figure 9—Sorted BLM Acres Burned and Severity Thresholds. USDA Forest Service Proceedings RMRS-P-41. 2006. 763 Werth Predicting Fire Season Severity in the Pacific Northwest Analysis Methods Various statistical techniques were used to determine which variables would be the best predictors of fi re season severity. Polynomial regression analysis was used to create multivariate forecast equations. Graphical regression was also used in conjunction with contingency tables. All analysis was performed using Microsoft Excel. Multivariate Equations The fi rst step in this process was to identify which variables (seasonal precipitation, snow pack SWE, spring snowmelt date, March PDSI, and Pacific SSTs) were the best predictors of acres burned for each agency. Each variable was ranked from best to worst based on its correlation (R-squared value) with acres burned. Table 1 displays the rankings of each variable by agency. Overall, summer rainfall (July/August) was the best predictor, with March PDSI, April 1 SWE, and May 1 SWEs a close second. There were considerable differences in the predictor rankings by agency. However, even the best predictors did not do a good job of forecasting acres burned alone. Much better results were achieved when all the variables were used. This was accomplished by creating multivariate (multiple regression) equations unique to each fi re agency using all the variables. Each variable was “weighted” according to its correlation factor. The equation forecasting acres burned took the form y=a1(m1x12 +n1x1)+...+an(mnx n2 +nnx n)+b, where (y) is the dependent variable (acres burned), (x1) through (xn) the independent variables, (a1) through (an) are variable weighting factors, (m1,n1) through (mn,nn) are coefficients of each independent variable, and (b) a constant. The resulting equation predicts acres burned by fi re agency using either observed or forecasted values as input for each independent variable. Scatter Diagrams and Contingency Tables A second method of predicting acres burned is the utilization of scatter diagrams and contingency tables. This technique plots one variable against the other (i.e., April 1 SWE versus Spring Precipitation) on an x-y scatter diagram, and then labels the intersection of those two variables as either an “Above Average” fi re season or not. In this manner, threshold values for each variable can be constructed, dividing the diagram into “YES - high probability” or “NO - low probability” risk areas of fi re season severity (fig. 10). The results from multiple scatter diagrams, correlating a selection of variables, are then input into a 2-way YES / NO contingency table (fig. 11) that predicts the probability of an “Above Average” fi re season and the range of acres burned in similar years dating back to 1970. Table 1—Correlation Factor Ratings by Agency. Parameters Winter Rain Apri1 1 Snowpack Spring Rain June Rain Summer Rain March PDSI April ENP Snowmelt Date May 1 Snowpack 764 USFS BLM BIA ODF WDNR Ave Rank 9 5 4 3 1 6 2 6 6 7 3 5 7 1 6 7 3 2 8 5 2 4 1 6 9 2 6 3 4 5 8 5 1 7 9 2 3 5 7 8 1 2 9 4 5 6.00 4.40 4.60 6.00 1.80 4.20 6.80 4.80 4.20 USDA Forest Service Proceedings RMRS-P-41. 2006. Predicting Fire Season Severity in the Pacific Northwest Werth Figure 10—WA DNR May SWE vs June Rain. Figure 11—USFS YES/NO Contingency Table. USDA Forest Service Proceedings RMRS-P-41. 2006. 765 Werth Predicting Fire Season Severity in the Pacific Northwest Results The combination of scatter diagrams, contingency tables, and multivariate equations produces the following outputs used to predict the severity of fi re seasons in the Pacific Northwest: • a defi nition of fi re season severity (above average, average, below average) based on acres burned by fi re agency, • the projected acres burned for the coming fi re season, and • the probability of an “Above Average” fi re season. The program is based on thirty-five years of weather and fi re data (1970 to 2004). The relatively small number of data points is near the minimum needed to draw confidence in the statistical analysis. However, significant changes in fi refighting strategy, resource availability, and wildland fuel regimes over the years produce additional uncertainty if data from years prior to 1970 is included. Thus, the current evaluation of how well the program performs is based upon “dependent” rather than ‘independent” data. Statistics in future years will be able to provide more relevant verification. Accuracy rates indicate the program will produce correct forecasts of fire season severity in Washington and Oregon in 70 to 85% of the years on which the data was based. A forecast of an “Above Average” fire season should verify correctly 60 to 85% of the time, and a forecast of “average” or “below average” 85 to 90% of the time. There appear to be better accuracy rates in predicting acres burned in timber fuels compared to grass / brush fuels, which isn’t surprising when considering the sensitivity of fire spread rates in grass fuels. 2006 Northwest Fire Season Early projections of 2006 fi re season severity in Washington and Oregon are based on correlations with past fi re seasons and the following factors: weak La Niña conditions, a wet winter, lack of drought, an above normal snowpack, and projected late spring snowmelt dates. Additional assumptions are that spring and summer will experience “near normal” or “typical” precipitation patterns (i.e., periodic rains through June, followed by dry weather during July and August) and there will be an average amount of lightning. Considering the above factors, it is highly unlikely that Washington and Oregon will experience a severe fi re season in 2006. However, the threat of large fi res will vary considerably by fuel type. Forest fuels in the mid and higher elevations of the Cascade and Blue Mountains will have the lowest probability of sustaining large fi re growth. The threat of large fi res will be the highest in grass fuels, primarily in the “High Desert” of central and southeastern Oregon. Other locations that may experience a greater chance of large fi res are the pine forests along the lower eastern slopes of the Cascades and the lower slopes of the Blue Mountains, where grass is the primary carrier of fi re. Table 2 displays the severity forecast for each of the five federal and state agencies, as well as the projected acres burned. In general, western Washington and western Oregon, including the crest of the Cascades, will likely see a Below Average fi re season. Eastern Washington can expect an Average fi re season. Eastern Oregon may also see a Below Average fi re season in the Klamath Basin and most of the Blue Mountains. Central and southeastern Oregon are projected to experience an Average to Above Average fi re season (fig. 11). 766 USDA Forest Service Proceedings RMRS-P-41. 2006. Predicting Fire Season Severity in the Pacific Northwest Werth Table 2—Projected 2006 Acres Burned by Agency. Threshold acres burned for an above average fire season 2006 Fire season Probability of an above average fire season Projected 2006 acres burned USFS Average to below average 10% 25,000 to 50,000 acres 120,000 acres BLM Average to above average 40% 50,000 to 90,000 acres 90,000 acres BIA Average to above average 30% 10,000 to 20,000 acres 20,000 acres ODF Average to below average 10% 5,000 to 9,000 acres 14,000 acres WADNR Average to below average 10% 4,000 to 9,000 acres 10,500 acres Agency Figure 12—Northwest Fire Season Severity. USDA Forest Service Proceedings RMRS-P-41. 2006. 767 Werth Predicting Fire Season Severity in the Pacific Northwest Summary and Conclusions Potential fi re season severity in the Pacific Northwest is projected using statistical techniques correlating weather data and annual-acreage-burned figures for five fi re management agencies in Washington and Oregon. Weather and fi re trends for the period 1970 to 2004 are calculated. Thresholds for above average, average, or below average fi re seasons were determined based on annual acres burned. Eight weather parameters were correlated using scatter diagrams, contingency tables, and multivariate regression equations to predict above average, average, or below average fi re seasons based on projected acres burned. Future modifications to this research may include replacing existing variables with new and better variables, and the development of equations that predict fi refighting costs and resource needs. Although this research is specific to the Pacific Northwest, the concept of using multiple predictors to forecast fi re season severity is adaptable to other areas, nationally and internationally. Acknowledgments The author thanks Brian Potter, AirFIRE Team Seattle WA, USDA Forest Service; Tony Westerling, University of California, San Diego; and John Werth, National Weather Service for their helpful comments and suggestions in the review of this paper. References Castro, C. L.; McKee, T. B.; Pielke Sr., R. A. 2001. The Relationship of the North American Monsoon to Tropical and North Pacific Sea Surface Temperatures as Revealed by Observational Analysis. J. Climate. 14: 4449-4473. Hall, B. L.; Brown, T. J. 2003. A Comparison of Precipitation and Drought Indices Related to Fire activity and Potential in the U.S. AMS Fifth Symposium on Fire and Forest Meteorology. J11.3. Westerling, A.L.; Gershunov, A.; Cayan, D. R.; Barnett, T. 2002. Long Lead Statistical Forecasts of Area Burned in Western U.S. Wildfi res by Ecosystem Province. International Journal of Wildland Fire. 11: 257-266. 768 USDA Forest Service Proceedings RMRS-P-41. 2006.