Frequency of cool summers in interior North America over the past three centuries



[1] An innovative technique utilizes a tree-ring marker to investigate long-term changes in the frequency of cool summers in Interior North America (INA), a region that currently suffers important gaps in knowledge concerning annual to secular temperature changes. Using multivariate adaptive regression splines (MARS), we established a threshold for the formation of climatically-induced light rings recorded in Pinus banksiana trees from INA. Then, we used the MARS model to reconstruct negative departures in summer maximum temperatures (April–September) from 1717 to 2007. The estimates explain 45% of the variance in instrumental temperature data. The reconstruction indicates the presence of significant multidecadal changes in the frequency of cool summers, with maximums in 1780, 1900 and 1960 and minimums in 1740, 1860, 1920 and 2000. No evidence of secular changes was found.

1. Introduction

[2] Across Interior North America (INA), 2004 had the coldest May to August daytime temperatures recorded since 1948 (Figure 1a). The city of Winnipeg (Manitoba, Canada) recorded its coldest summer since regular observations began in 1872 (Environment Canada, The top ten Canadian weather stories for 2004,, 2004), with May to August seasonal means of daily maximum temperatures (hereafter: equation imagemax) at 3.5°C below the 1971−2000 average (Figure 1b). The summer of 2008 also was a cool one, at 1.6°C below the normal. The impact of these cool summers on society and nature were not very strong and primarily affected recreational activities (e.g., decrease in the number of park visitors). Nevertheless, a fair question may be addressed: what is the change, if any, in extreme summer cooling events in INA associated with the widely accepted human-caused global warming [Jansen et al., 2007; Trenberth et al., 2007; Mann et al., 2008]?

Figure 1.

(a) Average May to August temperatures (°C) for the year 2004 expressed as departures from the climatology mean of 1961−1990 [Fan and van den Dool, 2008]. Location of Duck Mountain is shown (DMPF; 51°40′N; 100°55′W). Crosses indicate the spatial distribution of proxy databases (tree rings and non-tree-rings) used in the latest-reconstruction of surface temperature at hemispheric and global scales [Mann et al., 2008]. A box/dotted line delineates the region identified as ‘Interior North America’ (INA). (b) May to August seasonal means of maximum daily temperatures (equation imagemax) for the city of Winnipeg, Manitoba, Canada (1873−2008 [Vincent and Gullett, 1999; Environment Canada, 2008]). The city is located 300 km southeast of DMPF. The red line is a second-order least squares polynomial fitting across a moving 10-year window. The 2004 and 2008 summers are circled. The map was created using the KNMI Climate Explorer.

[3] Knowledge gaps of temporal changes in temperature means and extremes across INA are caused by the limited time coverage of meteorological observations and low availability of temperature-sensitive proxies (Figure 1a). Temperature reconstruction from tree rings, which is the largest source of information for hemispheric- and global-scale temperature reconstructions, has mainly been developed at altitudinal and latitudinal treelines where temperature is a strong limiting factor of growth [Jansen et al., 2007; Mann et al., 2008]. These temperature limiting conditions are less severe in the continental climate setting of INA, where moisture and temperature are both strong limiting factors of growth [e.g., Girardin and Tardif, 2005; Cook et al., 2007; St. George et al., 2009]. In this study, we use a novel technique that utilizes a tree-ring marker to develop an annually-resolved regional reconstruction of summer equation imagemax in INA spanning 1717−2007. This temperature reconstruction was inferred from changes in the proportion of climatically-induced light rings (hereafter: LR) recorded in jack pine (Pinus banksiana Lamb.) trees growing in a mixed boreal forest of central Canada (Duck Mountain Provincial Forest, DMPF; Figure 1a). A threshold for the formation of LR is established and changes in the rate of occurrence (OccRate) of summer cooling events in INA are detected with the help of bootstrap confidence bands.

2. Background

[4] The use of LR in dendrochronology has a long history [Gill, 1930]. LR have frequently been used to crossdate tree-ring series. Their presence in numerous trees within a region also forms the basis for their use as bioindicators of environmental events during the growing season [Liang and Eckstein, 2006]. LR are defined as tree-rings in which the latewood zone is made up of a few layers of latewood cells [Filion et al., 1986] that are light-colored due to incomplete cell wall development. LR are easily identifiable under a dissecting microscope (see auxiliary materials for LR images). In trees at northern and subalpine treelines, climatically-induced LR have been associated with below-average temperatures during the overall/end of growing season and/or reduced growing season length [Filion et al., 1986; Yamaguchi et al., 1993; Szeicz, 1996; Gindl, 1999; Wang et al., 2000; Hantemirov et al., 2000]. There have been no previous studies focusing on climatically-induced LR in central North America.

3. Data and Methods

[5] Sampling took place in the DMPF, which covers approximately 376,000 ha (Figure 1a). The DMPF is part of the Boreal Plains ecozone, a transition zone between the boreal forest to the north and the aspen parkland and prairie to the south. It makes up part of the Manitoba Escarpment, which is characterized by its higher elevation compared with the plains to the east. Baldy Mountain is the highest point, at 825 m above sea level. During the summers of 2000 and 2001 the DMPF (Figure 1a) was surveyed with the objective of reconstructing fire history. Jack pine cores (2 radii/tree) and stem cross-sections from living and dead trees were collected from 59 sites. Each of the cores and sections were dried, sanded, and cross-dated using the pointer-year method [Yamaguchi, 1991]. LR were identified based on visual comparison with adjacent rings [Liang and Eckstein, 2006; Filion et al., 1986] and all instances were recorded as part of the cross-dating procedure. The LR chronology was not made from jack pine trees evenly distributed through time, as many of the trees originated from post-fire recruitment, as for example in the 1880−1890s (Figure 2c). To ensure accurate dating of LR, annual growth increments were measured and cross-dating and measurements were statistically validated [Holmes, 1983]. The proportion of trees recording a LR in the sampled tree population for any given year was computed and a LR chronology covering the period extending from AD 1717 to AD 2001 was created. For the purpose of this work, the LR chronology was updated to 2007 after revisiting a random subsample of trees.

Figure 2.

Reconstruction of April to September seasonal means of daily maximum temperatures (equation imagemax; April–September) in DMPF. (a) Sub-calibration and verification of the MARS model; R2 refers to the calibration model R-square over 1936−2006, whereas r2 is the square of the Spearman rank correlation on data withheld from the calibration (1900−1935). The model exhibits significant predictive skills as r2 is greater than 0.11 (P < 0.05). (b) Reconstruction of negative deviations in equation imagemax (cool summers) after calibration of all available data (1900−2006). (c) Number of trees used through time. The position of the horizontal dashed line approximates the minimum number of samples needed to estimate a proportion of LR in 2.0% of the samples with 95% confidence that the estimate is within ±5% of the true population percentage. (d) Occurrence rate (OccRate) of years during which deviation of equation imagemax was estimated to be below 17°C, with 90% confidence bands.

[6] Monthly means of equation imagemax for the DMPF were obtained for the 1900−2006 period for the location 51.45°N; 100.55°W (740 m asl) using BioSIM [Régnière and Bolstad, 1994]. In this procedure, daily data were obtained by interpolating data from the four closest weather stations and adjusting for differences in latitude, longitude and elevation between the data sources and the location. Seasonal means were obtained from the average of the daily data. It should be noted that while the number of meteorological stations used to generate the weather data was kept constant through time (N = 4), the mean distance between the meteorological stations and the location increased with elapsed time because of the lower density of meteorological stations in earlier times compared with today. Greater uncertainty is thus expected in the earlier part of the record.

[7] Calibration of the equation imagemax reconstruction was carried out using multivariate adaptive regression splines (MARS) [Friedman, 1991]. MARS is a technique in which non-linear relationships between a predictand (i.e., variable to predict) and a predictor are described by a series of linear segments of differing slopes, each of which is fitted using a basis function. Breaks between segments were defined by an inflection point in a model that initially over-fitted the data, and which was then simplified using a backward/forward stepwise cross-validation procedure to identify terms to be retained. At each step, the model selected the inflection point and its corresponding pair of basis functions that gave the greatest decrease in the residual sum of squares. Selection was proceeded until some maximum model size was reached, after which a backward-pruning procedure was applied in which those basis functions that contributed least to model fit were progressively removed. The sequence of models generated from this process was then evaluated using generalized cross-validation, and the model with the best predictive fit was selected. For the present study, equation imagemax was calibrated against the proportion of trees (equation image) recording a LR in the sampled tree population over the 1900−2006 period. The program MARS [Salford Systems, 2001] was used for calibration.

[8] Changes in the occurrence rate of cool summers (i.e., summers during which average daytime temperatures were unusually low) were analyzed over the 1717−2007 period using kernel functions [Mudelsee et al., 2004]. Kernel estimation allows detailed inspection of time-dependent event occurrence rates and assessment of significant changes with the help of confidence bands. We used a Gaussian kernel, K, to weigh observed LR event dates, T(i), i,…, N (total number of events), and calculated the occurrence rate, λ, at time t. Selection of the band width (h = 12 years) was guided by cross-validation. Confidence bands (90%) around λ(t) were determined using the bootstrap technique and a percentile-t confidence band was calculated. We used the program XTREND [Mudelsee, 2002] for occurrence rate estimation.

4. Results

[9] The number of trees used to develop the LR chronology ranged from a minimum of three in AD 1717 to a maximum of 406 in AD 1927 (Figure 2c). LR were observed in 86 years (29.6% of all years), and in 18 of those years, LR appeared in more than 5% of the samples. We determined that the sample size needed [Cochrane, 1977] to estimate the proportion of appearance of LR in 2% of the samples was 30 trees (with 95% confidence that the estimate is within ±5% of the true population percentage), and 73 trees for a proportion of 5%. Thus, sample size was sufficient to estimate LR proportions over much of the chronology with a minimum bias (auxiliary material). LR were found to form in young (10−50 yrs) and mature (> 50 yrs) jack pine trees (auxiliary material).

[10] LR in jack pine growing in the DMPF generally developed during years with a late growing season onset (cool April and May) and a cool end of the growing season (August–September) (Figure 3). Overall, April–September seasonal means of equation imagemax during years of abundant LR formation were ∼15.3°C (a difference of ∼2°C with non-LR years), with the increasing LR proportion corresponding to cooler growing season temperatures (auxiliary material). The statistical significance of this finding was confirmed using an analysis of variance between equation imagemax and LR proportion classes (0%, < 5% and ≥ 5%) and Scheffe post-hoc test for pairwise comparisons (auxiliary material). Compared with minimum and mean temperatures, cloud cover and drought indices, equation imagemax from April to September explained more variance, and this predominance may reflect the stronger association between low equation imagemax in early spring and years of high LR proportion (Figure 3). Temperature prior to the growing season and precipitation had little influence on LR formation. By means of contemporaneous correlations we also examined major climate indices for potential connections with LR formation (e.g., Cold Tongue Index, Pacific Decadal Oscillation, Pacific North American pattern, and Northern Annular Mode). Our results (not shown) were inconclusive and suggested that climate conditions leading to LR were primarily induced by internal atmospheric dynamics rather than large-scale atmosphere-ocean couplings. Enhanced cyclonic activity southwest of the Hudson Bay advecting moist polar air masses over the DMPF promoted LR formation.

Figure 3.

Daily maximum temperatures from April to September for years with no LR (proportion of 0%; average of n = 58 years), few LR (proportion < 5%; n = 41) and abundant LR (proportion ≥ 5%; n = 8). Shaded area indicates days during which temperature was significantly different (P < 0.05) between years of abundant LR and years without light rings based on two-sample t-tests.

[11] We calibrated the LR chronology against seasonal means of April–September equation imagemax using MARS. The MARS model explained 45% of the deviation between the predictand and the predictor (P < 0.0001) and took on the following form:

equation image

In comparison, a linear regression of the same data (equation imagemax = 17.731 − 1.824 · log (100equation image + 1)) explained 40% of the variance (P < 0.0001). In the MARS model, the inflection point takes on a value of 0.018, and BF1 takes a value of equation image minus 0.018 when equation image is > 0.018, but otherwise takes a value of zero. The values of BF2 have a value of 0.018 when equation image is zero, declining to zero as equation image approaches 0.018. They remain fixed at zero at values of equation image > 0.018. Coefficients applied to each of the basis functions define the slopes of the non-zero sections. Dividing the calibration interval did not lead to significant changes in MARS regression parameters. Furthermore, verification of estimates over a period withheld from calibration (Figure 2a) suggested that the MARS model had sufficient predictive skills to be applied backward in time. Estimates and observations over an independent period (1900−1935) were highly correlated (r2 = 0.39) and a sign test result indicated good fidelity in the direction of year-to-year changes between the observed and estimated data with 72% success (P = 0.02). That being said, LR in jack pine will not automatically be observed each time the mean April–September equation imagemax reaches the threshold necessary for LR formation (Figure 2a). This commonality observed in most LR studies [Filion et al., 1986; Yamaguchi et al., 1993; Szeicz, 1996; Gindl, 1999] brings a certain level of error to our reconstruction.

[12] We applied the MARS model to the LR chronology as far back as possible to produce a series of equation imagemax estimates. Owing to the non-linear interaction between LR occurrences and equation imagemax, only those negative deviations in equation imagemax from a threshold of 17.8°C were estimated with a standard error of regression of ±0.89°C. Years recording LR in more than 5% of the samples yielded the lowest equation imagemax estimates (<16.0°C) (Figure 2b): 1725, 1744, 1756, 1758, 1765, 1787, 1795, 1815, 1817, 1837, 1907, 1935, 1945, 1954, 1965, 1974–1975 and 2004. No summer cooling events were observed from 1838 to 1880. This absence occurred despite this interval being represented by 81 jack pine trees from 26 sampling sites, and with 65% of those trees having registered LR for a total of 125 LR events. Therefore, the absence of LR at the end of the 19th century does not seem to relate with the post-fire renewal of the forest. The regional nature of this phenomenon will eventually be addressed with expansion of the sampling to surrounding mountains.

[13] The OccRate of summer cooling events in INA was analyzed using kernel functions and mathematical bootstrap simulations (Figure 2d). During the 1717–2007 interval, OccRate of cool summers (<17°C) was not constant over time, but featured highs (0.34 event yr−1 around 1780, 0.30 yr−1 around 1920, and 0.41 yr−1 around 1960) and lows (0.09 yr−1 around 1740, 0.10 yr−1 around 1860, and 0.11 yr−1 around 1920). Estimates of OccRate for the late 20th century, 0.15 event yr−1 (i.e., every 6th year was a year with cool April–September equation imagemax), appear within the range of the historical period. Although the statistical uncertainty of OccRate is substantial, the amount of sample years led to the conclusion that multi-decadal changes were significant. Results on other temperature thresholds followed similar trends, albeit with higher uncertainties around OccRate owing to reduced sample size (auxiliary material).

5. Concluding Remarks

[14] In this study we used a novel technique that utilizes a non-linear modeling approach (MARS) to the calibration of a tree-ring marker, and investigated long-term changes in the frequency of cool summers in INA using kernel functions and mathematical bootstrap simulations. Our results suggest that climatically-induced LR in jack pine growing in the DMPF were the results of processes similar to those observed in studies carried out near the northern and altitudinal treelines, i.e., LR were primarily induced by below-average temperatures during the overall growing season. OccRate of cool summers since 1717 did not record long-term secular changes and rather exhibited a pattern of multidecadal variations.

[15] A possible explanation for the absence of secular changes in cool summers in a region that experienced significant warming [Trenberth et al., 2007] is that our inferences only apply to seasonal means of equation imagemax. Many regions across North America have shown important differences in the trend between minimum (nighttime) and maximum (daytime) temperatures over the last few decades, with a greater absolute rate of increase in minimum temperatures. In the DMPF, data show that minimum temperatures have increased at a rate of 0.01°C per year over 1900−2006 (P < 0.001), whereas the long-term mean of equation imagemax has remained constant at 17.3°C, but with some important multidecadal variations (auxiliary material). The period of maximum LR occurrence, i.e., 1940–1970 (Figure 2a), was synchronized with a period of relatively cool equation imagemax, and no such concordance existed with minimum temperatures. Therefore, LR chronologies in INA might be unsuitable for the detection of past temperature changes, other than those occurring in the lower tail of the distribution of daytime temperatures. Nevertheless, the significance of our results should form the basis for the development of other tree-ring proxies in INA (e.g., density and isotope measurements). Once included in the MARS model, they could provide insights into other aspects of temperature changes, namely on warm spells, average and variance.


[16] This research was funded by the Canada Research Chair program, University of Winnipeg, NSERC, and SFMN. We thank two anonymous reviewers for helpful comments on an earlier draft of this manuscript.