Feb 28, 2012 - William's-corrected G statistics are shown above each plot (critical value of .... Kitzberger, T., P. M. Brown, E. K. Heyerdahl, T. W. Swetnam, and T. T. ... Mantua, N. J., and S. R. Hare (2002), The Pacific Decadal Oscillation, J.
GEOPHYSICAL RESEARCH LETTERS, VOL. 39, L04703, doi:10.1029/2011GL050645, 2012
Does proxy uncertainty affect the relations inferred between the Pacific Decadal Oscillation and wildfire activity in the western United States? Kurt F. Kipfmueller,1 Evan R. Larson,2 and Scott St. George1 Received 16 December 2011; revised 20 January 2012; accepted 20 January 2012; published 28 February 2012.
[1] We examined a set of five proxy reconstructions of the Pacific Decadal Oscillation (PDO) to test whether the choice of reconstruction affected the association between the PDO and widespread forest fires in the western United States. Exact binomial tests suggest the PDO has little direct impact on wildfires, with a statisticallysignificant association between the phase of the PDO and regional fire activity observed with only one reconstruction. Region-wide fires were not consistently associated with specific phase combinations of ENSO and the PDO. Any conclusion that extensive wildfires are more or less common when the PDO is in one phase or the other depends entirely on the choice of PDO reconstruction. Without a better understanding of low-frequency behavior in the north Pacific prior to 1900, efforts using proxy data to determine whether or not the PDO affects wildfire activity in the western United States will produce ambiguous results. Citation: Kipfmueller, K. F., E. R. Larson, and S. St. George (2012), Does proxy uncertainty affect the relations inferred between the Pacific Decadal Oscillation and wildfire activity in the western United States?, Geophys. Res. Lett., 39, L04703, doi:10.1029/2011GL050645.
1. Introduction [2] Because of the prevalence of fire suppression in the western United States during the last 100 years, many insights into the associations between extensive wildfires and large-scale patterns in the climate system are based on evidence from proxy records. In a classic study conducted in the American Southwest, Swetnam and Betancourt [1990] showed that the El Niño-Southern Oscillation exerts a major control on wildfire activity in coniferous forests of the southwestern United States. Combining fire-scar records and drought estimates from tree rings demonstrated that wet winters related to El Niño conditions increased the growth of grasses and other fine fuels, which then feed wildfires during subsequent dry periods associated with La Niña. Paleofire records from the Pacific Coast of South America exhibit a similar association with ENSO [Kitzberger et al., 2001; Veblen et al., 1999]. [3] More recently, several studies have argued that lowfrequency behavior in the climate system acts in a similar fashion to influence wildfire activity on timescales of one to 1
Department of Geography, University of Minnesota, Twin Cities, Minneapolis, Minnesota, USA. 2 Department of Social Sciences, University of Wisconsin-Platteville, Platteville, Wisconsin, USA. Copyright 2012 by the American Geophysical Union. 0094-8276/12/2011GL050645
several decades. Hessl et al. [2004] suggested that fire activity in the Pacific Northwest of the United States exhibits greater spatial synchrony during the warm phase of the Pacific Decadal Oscillation (PDO). Paleofire records from scarred trees have also been used to show that the PDO, in conjunction with ENSO and other climate modes, affects the occurrence of wildfires in other parts of the western United States [Schoennagel et al., 2005; Sibold and Veblen, 2006; Kitzberger et al., 2007; Heyerdahl et al., 2008; Sherriff and Veblen, 2008; Taylor et al., 2008]. In all of these studies, the inferred association between fire activity and the state of the north Pacific Ocean is determined by comparing paleofire records with proxy reconstructions of the PDO index. [4] There are now several independent proxy estimates of the PDO index during the past several centuries, each developed from different sets of tree-ring data and a variety of reconstruction approaches [Biondi et al., 2001; D’Arrigo and Wilson, 2006; D’Arrigo et al., 2001; Gedalof and Smith, 2001; MacDonald and Case, 2005]. Although these reconstructions are able to reproduce the major shifts in the PDO index observed during the instrumental period, they are only weakly similar to each other prior to the 20th century [Mantua and Hare, 2002; Cook, 2009] (Table S1 of the auxiliary material).1 Because the available set of PDO reconstructions exhibit some degree of dissimilarity, it is possible that the strength or perhaps even the direction of the inferred relations between regional wildfires and the PDO could be affected by the choice of reconstruction. [5] In this study, we test whether differences between PDO reconstructions are substantive enough to affect the main conclusions drawn from comparing major climate modes against a network of paleofire records from the western United States. First, we use intervention analysis to identify persistent ‘regime shifts’ in each PDO reconstruction and determine whether low-frequency behavior is similar across the set of reconstructions. Next, we develop composite records of fire activity for three regions in the western United States and test if widespread fires are more or less common during certain phases (or phase combinations) of ENSO and the PDO. If the associations between these modes and wildfires are the same regardless of the choice of PDO reconstruction, our results would indicate that paleoclimatic records of the north Pacific Ocean are adequate to determine whether or not that sector influences regional wildfire dynamics. On the other hand, if the observed relationship between climate modes and fire activity does depend on reconstruction choice, it may be necessary for the fire-climate community to reconsider the 1 Auxiliary materials are available in the HTML. doi:10.1029/ 2011GL050645.
L04703
1 of 5
L04703
KIPFMUELLER ET AL.: PROXY UNCERTAINTY AND WILDFIRE
L04703
Figure 1. (a) Tree ring-based fire records from the International Multiproxy Paleofire database, after nearby sites were combined. (b) Annual time series of percent sites recording fires in each region. Triangles mark region-wide fire years, which are defined as years where the number of sites burned exceeded the 90th percentile. (c) Proxy reconstructions of El Niño-Southern Oscillation and the Pacific Decadal Oscillation used in this analysis. Solid lines represent reconstructed values. Dotted lines represent instrumental data. Triangles represent fire years in different analysis regions as in Figure 1b. (d) Warm (white) and cool (black) phases of the El Niño-Southern Oscillation and the Pacific Decadal Oscillation as estimated from proxy reconstructions. Gray shading indicates the reconstruction does not span that interval. role that the North Pacific is believed to play in modulating fire activity in the western United States.
2. Data and Methods 2.1. Fire History Records and Regional-Fire Years [6] We obtained all tree-ring based fire records from the western United States housed by the International Multiproxy Paleofire database (IMPD). Records from the IMPD are annually-resolved and are based on fire scars observed in multiple trees at an individual site. Because many paleofire records from the database are either located at the same geographic coordinates or are very close together, we combined sites within 10 km of each other to minimize biases towards small sub-regions that have been studied more intensively. This step reduced the number of fire records in our network from 322 to 128, with 39 sites in the Pacific northwest, 33 in the central Rockies and 56 sites in the Southwest (Figure 1a). At each site, the fire history described by each individual tree was converted to a binary variable (fire/no fire) for every year of record and then all tree-level records were combined to produce summary site chronologies. Each year with one tree or more containing a fire scar was classified as a fire year for that site. Regionwide fire years were defined as years when the number of sites that recorded fires was greater than the 90th percentile
of the number of sites burned within that region over the 1700–1900 interval (Figure 1b). This approach identified 21 region-wide fire years in the Pacific Northwest, 22 in the Central Rockies, and 27 in the Southwest (Table S2). These counts are roughly equivalent to those obtained by other studies investigating geographically-extensive wildfires in the western United States [Trouet et al., 2010; Heyerdahl et al., 2008; Schoennagel et al., 2005; Sibold et al., 2006]. 2.2. Proxy Reconstructions of ENSO and the PDO [7] We obtained one proxy reconstruction of ENSO [D’Arrigo et al., 2005] and five reconstructions of the PDO [Biondi et al., 2001; D’Arrigo and Wilson, 2006; D’Arrigo et al., 2001; Gedalof and Smith, 2001; MacDonald and Case, 2005] (Table S3). Comparison between these reconstructions and tree-ring reconstructed drought [Cook et al., 2004] reveal some differences in the spatial pattern of drought related to the different reconstructions (Figure S1). All reconstructions are based on tree ring-width data, which matches the temporal resolution and dating accuracy of the tree-ring paleofire records. We classified yearly values from the reconstructed ENSO series as either positive (El Niñolike) or negative (La Niña-like) depending on whether or not the reconstructed value was above or below zero. Because the PDO index exhibits high variance at decadal or
2 of 5
L04703
KIPFMUELLER ET AL.: PROXY UNCERTAINTY AND WILDFIRE
L04703
Figure 2. Expected (black bars) and observed (white bars) number of fire years tested using an exact binomial test for five reconstructions of the Pacific Decadal Oscillation and one reconstruction of the El Niño-Southern Oscillation. Bold values indicate associations that are significant at p < 0.05. multidecadal timescales [Mantua and Hare, 2002], we used intervention analysis [Box and Tiao, 1975; Rodionov, 2004] to identify persistent warm or cool phases in each PDO reconstruction following the approach described by D’Arrigo and Wilson [2006] and Trouet and Taylor [2010]. Because the total number of fire events in each region is relatively low (n ranges between 21 and 27 fires), we sorted reconstructed values for the ENSO and PDO series into only two classes (positive or negative) to ensure that the results of our contingency tests are stable (see next section). 2.3. Fire-Climate Analyses [8] For each region, we used an exact binomial test and log-likelihood G-tests [Gotelli and Ellison, 2004; Sokal and Rohlf, 1994] to examine whether region-wide fires were more or less common during particular phases or phase combinations of ENSO and PDO. Compared to the number of classes (k = 4), the number of fire years in our analysis was relatively small and several phase combinations had fewer than five expected fires. Because using a low number of events can produce inaccurate results in these types of contingency tests [Sokal and Rohlf, 1994], we employed a Monte Carlo procedure to test the significance of our statistical results [McDonald, 2009] (modified from a spreadsheet available at http://udel.edu/mcdonald/statintro.html). The Monte Carlo simulation used 1000 random draws to determine the probability of a single iteration producing a Gtest statistic that was greater than the observed p-value. Expected values for both the exact binomial and G-tests were determined using an extrinsic hypothesis based on the proportion of years from each reconstruction falling in a given class (sensu Sokal and Rohlf [1994]). We restricted our analysis to the period 1700–1900 because this interval is common to all of the PDO reconstructions and excludes the modern era of fire suppression. The G-statistic was also corrected to approximate a chi-square distribution more
closely and to account for the small sample size [Williams, 1976].
3. Results and Discussion [9] The timing of phase shifts associated with each PDO reconstruction demonstrates that the set of proxies exhibits important differences in low-frequency behavior prior to 1900 (Figure 1c and Table S1). Over the entire 200-year period, there are only two brief intervals when all five reconstructions report as the same phase (around 1770 and again circa 1840). Comparing pairs of reconstructions shows similar disagreements. For example, DAR01 and DAW06 are out of phase with each other for several decades (particularly from 1750 to 1820 and 1850 to 1890), with the DAW06 reconstruction tending toward more frequent warm phase conditions than the DAR01 reconstruction. [10] The exact binomial test confirmed that ENSO is associated with fire activity in the central Rockies and Southwest. Across both regions, widespread fires occurred more frequently during La Niña-like conditions, while El Niño-like conditions were associated with fewer fires (Figure 2). This pattern is consistent with prior studies of fire history in these regions [Kitzberger et al., 2001; Sherriff and Veblen, 2008; Sherriff et al., 2001; Swetnam and Betancourt, 1990, 1998]. In contrast, the PDO appears to have had little direct impact on wildfires. A statisticallysignificant association between the phase of the PDO and regional fire activity existed in only one case, suggesting that fires in the central Rockies occurred more often during the negative phases of the PDO as estimated by the BIO01 reconstruction. Although the association is not statistically significant, the DAW06 and G&S01 reconstructions showed exactly the opposite relationship, with fire activity in the central Rockies being more common during the positive phase of the PDO. Regardless of the PDO reconstruction under consideration, the phase of the PDO is not
3 of 5
L04703
KIPFMUELLER ET AL.: PROXY UNCERTAINTY AND WILDFIRE
L04703
Figure 3. Mosaic plots of contingent relationships between widespread fire years and proxy estimates of the PDO and NINO3 for three regions across the western United States. The numbers in each box represent the observed number of fire events falling within each contingent class. William’s-corrected G statistics are shown above each plot (critical value of 2 X [3,p
