Deriving temperature estimates from Southern Hemisphere leaves

Palaeogeography, Palaeoclimatology, Palaeoecology 412 (2014) 80–90

Contents lists available at ScienceDirect

Palaeogeography, Palaeoclimatology, Palaeoecology journal homepage: www.elsevier.com/locate/palaeo

Deriving temperature estimates from Southern Hemisphere leaves Elizabeth M. Kennedy a,⁎, Nan Crystal Arens b, Tammo Reichgelt c, Robert A. Spicer d,e, Teresa E.V. Spicer e, Lena Stranks f, Jian Yang e a

GNS Science, P.O. Box 30-368, Lower Hutt, New Zealand Department of Geoscience, Hobart & William Smith Colleges, Geneva, NY 14456, USA Department of Geology, University of Otago, P.O. Box 56, Dunedin, New Zealand d The Open University, Milton Keynes MK7 6AA, United Kingdom e Institute of Botany, Chinese Academy of Sciences, Beijing, China f 1 The Long House, Main Street, Willersey, Broadway WR12 7PJ, United Kingdom b c

a r t i c l e

i n f o

Article history: Received 7 March 2014 Received in revised form 8 July 2014 Accepted 11 July 2014 Available online 23 July 2014 Keywords: Leaf Margin Analysis CLAMP New Zealand Southern Hemisphere Palaeoclimate analysis

a b s t r a c t The percentage of woody dicots with entire-margined leaves in a flora is known to be positively correlated with mean annual temperature (Leaf Margin Analysis — LMA) but this relationship is not globally uniform. In particular the floras of Australia and New Zealand have been regarded as displaying a different physiognomic relationship to climate than floras seen in the Northern Hemisphere. This difference is more marked in New Zealand where the LMA relationship appears entirely absent. Here we amass data for both Northern and Southern hemispheres using standard protocols and show that regional variations in the leaf margin–mean annual temperature relationship are real but become less significant when other characters are included. Even New Zealand falls into line and most of the mean annual temperature signal in New Zealand floras is encoded in non-margin features. We introduce a new CLAMP (Climate Leaf Analysis Multivariate Program) calibration dataset for the Southern Hemisphere, comprising leaf physiognomic data from Argentina, Bolivia, South Africa, Australia, New Zealand and other Pacific Islands that offers comparable precision for climate prediction to similar datasets derived from the Northern Hemisphere. © 2014 Elsevier B.V. All rights reserved.

1. Introduction With the development of global palaeoclimate models, the value of quantitative palaeoclimate data for testing model reliability has stimulated new ways of producing quantitative climate reconstruction data from the fossil record. There are few quantitative palaeoclimate proxies applicable to terrestrial sediments. The use of leaf morphology is one of the most powerful, particularly because leaves respond directly to conditions in the atmosphere and fossil leaves can therefore be used as proxies for an array of climate parameters. Unfortunately, deriving a globally applicable foliar physiognomic climate reconstruction technique that provides accurate and precise results over geological timescales is no simple task. This is particularly so for the Southern Hemisphere where leaf form is often regarded as being poorly and/or differently correlated with climate compared to the Northern Hemisphere (Upchurch and Wolfe, 1987; Greenwood, 1992; Jordan, 1997; Stranks and England, 1997; Kennedy, 1998; Greenwood et al., 2004). The most commonly used measure of climate for model validation is mean annual temperature (MAT), even though this is probably not a critical limiting measure for plant growth and distribution. Measures ⁎ Corresponding author. Tel.: +64 4 5704838. E-mail address: [email protected] (E.M. Kennedy).

http://dx.doi.org/10.1016/j.palaeo.2014.07.015 0031-0182/© 2014 Elsevier B.V. All rights reserved.

related to freezing such as the cold month mean temperature (CMMT), or potential heat stress such as the warm month mean temperature (WMMT) are likely to be climatic parameters that limit plant distribution. Nevertheless MAT can readily be compared both to climate model output and to measures provided by independent geochemical proxies such as those based on isotopes. MAT is also a key measure in determining the global mean surface temperature and latitudinal temperature gradients, a primary driver of the climate system. For this reason it is important to evaluate the capacity of leaf morphological proxies to reconstruct MAT both accurately and precisely. The use of foliar physiognomic analysis as a tool for climate reconstruction assumes that leaf form is optimised through natural selection for maximising primary productivity and minimising structural investment, while managing water relations and radiation balance. Because this is such a basic evolutionary strategy the physiognomic approach is likely to be time-stable and has been successfully applied as far back as the early radiation of the angiosperms approximately 100 million years ago (e.g. Spicer and Herman, 2010, and references therein). The term ‘successfully’ here needs qualification. Any proxy method can generate climate retrodictions but the critical issue is how well they reflect what the actual values were in the past (accuracy) and what the uncertainties are in those retrodictions (precision). The only way to evaluate accuracy for past climates is through consilience with

E.M. Kennedy et al. / Palaeogeography, Palaeoclimatology, Palaeoecology 412 (2014) 80–90

other proxies (preferably based on independent methodologies) and for measures of precision we need to define carefully a standardised methodological framework so that calibration is consistent throughout the methodology and the statistics generated are therefore valid. In this paper we discuss temperature estimation from two foliar physiognomic methods of analysis — the Climate Leaf Analysis Multivariate Program (CLAMP) and Leaf Margin Analysis (LMA). We investigate hemispheric disparities using a multivariate approach to foliar physiognomy at 90 locations throughout the Southern Hemisphere including newly collected data from New Zealand and Australia.

2. The universality of the foliar physiognomy/climate relationship In the context of using standardised methodologies for collecting both physiognomic and climate data here we focus on two discussion points that are often associated with foliar physiognomic palaeoclimate proxies: 1. Is the relationship between leaf physiognomy and temperature the same in Northern and Southern hemispheres? 2. Is the relationship between leaf margin and temperature always the most dominant leaf character/climate relationship in these methodologies?

2.1. Is the relationship between leaf physiognomy and temperature the same in Northern and Southern hemispheres? One of the most critical questions regarding the application of quantitative leaf morphology-based methods such as LMA and CLAMP concerns their global validity. For the most part, leaf physiognomic proxies have been developed in the Northern Hemisphere with Northern Hemisphere modern analogue datasets. The validity of their application to the Southern Hemisphere is questionable (e.g. Greenwood et al., 2004). It is clear that various LMA studies have shown important differences in regression statistics but because of the lack of standardisation regarding both the physiognomic and climate data the sources and magnitude of these differences have been unclear. To explore the applicability of LMA and CLAMP to the Southern Hemisphere we assembled datasets following the standardised field collecting protocols used in CLAMP, accompanied by high-resolution gridded climate data assembled from global station data for the 1961–1990 interval. Global, hemispheric and regional variations in leaf form/climate relationships can thus be investigated within a common leaf sampling and climatic framework. These datasets are composed of previously published sites, combined with newly collected material in the case of the Southern Hemisphere dataset.

2.2. Is the relationship between leaf margin and temperature always the most dominant leaf character/climate relationship in these methodologies? There has been considerable debate as to the comparative usefulness of the univariate versus multivariate foliar physiognomic methods (Wolfe, 1979; Wing and Greenwood, 1993; Wolfe, 1993; Wilf, 1997; Wilf et al., 1998; Gregory-Wodzicki, 2000; Spicer et al., 2005; Spicer and Yang, 2010; Steart et al., 2010; Spicer et al., 2011). Univariate methods are attractive because of their simplicity. They are relatively straight-forward to score and to calculate, but are they always the most accurate and precise? Wilf (1997) stated that the CLAMP method does not improve temperature estimates produced using LMA because the temperature signal is dominated by the leaf margin character suite in the CLAMP dataset, masking any useful influence from other characters. We will test this assertion here.

81

3. Materials and methods 3.1. Leaf margin analysis (LMA) 3.1.1. Introduction: LMA The first attempt to relate leaf form to temperature was that of Bailey and Sinnott (1915, 1916) who recognised that the leaf margin type in woody dicots is correlated with MAT, although they used no quantitative temperature observations in their work, using instead qualitative climate classifications such as ‘tropical’ and ‘warm temperate’. Wolfe (1979) made use of the leaf margin/temperature relationship and established it as a quantitative palaeoclimate proxy. This quantitative approach is now commonly termed Leaf Margin Analysis (LMA). LMA utilises the positive correlation between observed temperature and the proportion of woody dicotyledonous species in a modern vegetation assemblage that have leaves with entire margins (E), and remains widely used. Regression equations produced using modern datasets of leaf margin and temperature information are then used to calculate mean annual temperature from the E value of the fossil assemblage. To-date the most widely applied palaeotemperature regression equation (MAT (°C) = (E × 0.306) + 1.141) for this correlation is based on a Southeast Asian dataset (Wolfe, 1979; Wing and Greenwood, 1993). Several other LMA calibrations based on regional datasets and CLAMP datasets have also been applied (e.g. Wilf, 1997; Greenwood et al., 2003; Kowalski and Dilcher, 2003; Greenwood et al., 2004; Hinojosa and Villagrán, 2005; Miller et al., 2006). Although still widely used, LMA suffers from a number of limitations. Wolfe (1979) emphasised that the technique did not perform well in dry or cold climates where water is limiting to growth. In these situations leaf size is small and water loss through marginal teeth would be disadvantageous (Bailey and Sinnott, 1915; Wolfe, 1993). Wolfe (1979) also noted that spinose margins adapted to deter browsing should be regarded as entire (untoothed). The complication most detrimental to the routine use of LMA as a climate reconstruction tool is that there is no single globally applicable LMA regression and differences exist between Northern and Southern hemispheres (Wolfe, 1979; Upchurch and Wolfe, 1987). Wolfe's (1979) Northern Hemisphere LMA gradient, based on the monsoonaffected Southeast Asian vegetation, showed the approximate relationship between MAT (°C) and % entire margins (E) to be an increase in E of 3% for every 1 °C increase in MAT, with 60% E at the 20 °C isotherm. Wolfe employed only a few Southern Hemisphere floras but suggested that they indicated a margin/MAT relationship that was closer to 4% entire/1 °C for Southern Hemisphere floras. In addition to hemispheric differences several workers have noted regional variations in LMA regressions (e.g. Greenwood, 1992; Greenwood et al., 2004; Steart et al., 2010) and these differences are often most strongly displayed in floras with high endemism. In the most extreme cases there appears to be no correlation at all between leaf margin (%E) and temperature, such as in Costa Rica (Dolph and Dilcher, 1980) and New Zealand (Stranks, 1996; Kennedy, 1998). While many of these regional variations in LMA have been interpreted in terms of biogeographic history (Greenwood et al., 2004) and phylogeny (Little et al., 2010) part of the reason for the differences in LMA regressions can be attributed to different leaf sampling strategies, and to climate datasets collected over different time intervals and from stations with differing relationships (altitude, aspect, distance) to the vegetation sampled. This ‘calibration noise’ remains unquantified but can be broken down into that due to variation in collection methodology and that due to uncertainties in climate data. 3.1.2. Leaf data as a source of uncertainty in LMA Flawed sampling strategies were embedded in foliar physiognomic research from the beginning. Bailey and Sinnott (1915) took their data from regional floras documented for taxonomic purposes and defined by political rather than phytogeographical boundaries, and which

82

E.M. Kennedy et al. / Palaeogeography, Palaeoclimatology, Palaeoecology 412 (2014) 80–90

represented sample areas of vastly differing sizes, and did not consider the effects of elevation on climate. Thus leaf margin form reported in a flora from ‘Brazil’ is compared to the material derived from ‘the Los Angeles area’. The use of floral lists derived from vastly differing and poorly defined areas persisted in the foliar physiognomic work of Wolfe (1979), Wilf (1997), Greenwood et al. (2004) and numerous other workers. Moreover, in some instances not all woody taxa were included in the analyses. For example in Wilf (1997) some sites included only trees with stems/trunks ≥10 cm diameter and thus excluded vines, small bushes and juvenile trees. This lack of consistency in sampling methodology can influence regressions, reduce precision and contribute to differences in outcomes between studies. 3.2. Multivariate foliar physiognomy and CLAMP 3.2.1. Analytical method Recognising that aspects of leaf morphology other than teeth could also provide climate information and that LMA had limitations, Wolfe (1993) developed the multivariate method CLAMP (Climate Leaf Analysis Multivariate Program) that utilises an array of leaf features to estimate both temperature and precipitation variables. Rather than using herbarium or floral list data Wolfe chose to collect his calibration material directly from the field and from localised areas. Initially Wolfe used Correspondence Analysis (CA) (Benzecri, 1973; Hill, 1973, 1979) to arrange vegetation sites in multidimensional space such that sites with similar leaf physiognomies plotted close to each other and those with different leaf form spectra plotted apart. CA was employed because it is a technique that is robust to missing data such as is common with fossil leaves, moreover CA does not assume independence of the variables nor normality in their measurements. Most importantly CA, unlike multiple regression analysis (e.g. Wing and Greenwood, 1993; Gregory and McIntosh, 1996; Wiemann et al., 1998), tolerates and displays correlations among variables. This is relevant in CLAMP because all leaf character variables are correlated through both common ancestry and selection for efficiency (optimisation), and all climate variables are correlated with one another through the physics of the atmosphere (Spicer and Yang, 2010). Multiple regression analysis assumes that the variables are independent, there is an absence of multicollinearity, and data describing them are normally distributed (Kent and Coker, 1992). In the case of both leaf and climate variables these assumptions cannot be met and the use of multiple regression analysis imposes false constraints on the analyses and increases the likelihood of spurious outcomes. In Wolfe's (1993) first incarnation of CLAMP, climate trends through this physiognomic space were determined by eye and so introduced the potential for significant operator error, but subsequently Canonical Correspondence Analysis (ter Braak, 1986) was used to explicitly align climate vectors (Kovach and Spicer, 1995). Thus CLAMP currently comprises two calibration datasets, one of leaf morphology data for each site and one of corresponding meteorological data for each site (http://clamp.ibcas.ac.cn). 3.2.2. The foliar physiognomic datasets Inevitably during the evolution of CLAMP the number and type of sites used for calibration have varied and grown. The leaf morphology dataset currently consists of percentage scores for 31 different leaf character states across, ideally, a minimum of 20 woody dicot taxa for each vegetation site, including margin, size, base and apex, and general shape. The use of 20 taxa as a minimum reduces uncertainties to an acceptable level while maximising the number of fossil assemblages to which the technique can be applied (Povey et al., 1994). In its original form the standard CLAMP dataset included 106 leaf assemblages, all of which were from the Northern Hemisphere, predominantly from North America. A revised 103 site dataset was in use for some time until the current standard CLAMP PHYSG3BR dataset composed of 144 sites, predominantly from North America and Japan was introduced,

together with a larger (173 sites), but less precise, calibration set known as PHYSG3AR that includes sites where significant cold is experienced. The PHYSG3BR dataset offers the greatest precision for many mid and high latitude Northern Hemisphere fossil assemblages (e.g. Uhl et al., 2007; Spicer and Herman, 2010; Yang et al., 2011) but often fossil sites, particularly from low palaeolatitudes, plot outside the cloud of modern sites that make up the ‘physiognomic space’ of the calibration set. In this situation, as with LMA when all the leaves are either toothed or entire, the uncertainties become infinite. To try to overcome this, the PHYSG3BR sites were supplemented with sites influenced by the East Asia Monsoon (PHYSGASIA1) (Jacques et al., 2011), the South Asia Monsoon (PHYSGINDIA1) (Srivastava et al., 2012) and from sites in the Southern Hemisphere discussed here. Unlike the previous LMA datasets all CLAMP samples comprising the calibration datasets have been collected directly from the field, from areas of natural or naturalised vegetation (Spicer et al., 2004) less than 1 km in diameter, from precisely located sites at known altitudes, following publicly defined protocols (http://clamp.ibcas.ac.cn). These protocols include sampling all visible leaf physiognomic variation within a taxon. This standardised approach reduces spurious patterns in physiognomic space as far as possible and reduces uncertainty to a minimum, but includes leaf form variability arising from spatial heterogeneity that is a functional component in ecosystems and not merely noise generated from random processes (Legendre, 1993). 3.3. Climate data as a source of uncertainty in LMA and CLAMP The principle of standardisation also applies to the climate data used in the calibration process. An appreciation of the uncertainties in the climate data used to calibrate a palaeoclimate proxy is essential to evaluate its reliability. When climate data are derived from different, poorly constrained sources, estimates of the precision and accuracy of that proxy become meaningless. Several issues arise in the derivation of climate data associated with modern sample sites. Climate stations local to the vegetation sampling sites often have incomplete records, records of different length or records spanning different time intervals. These issues, as well as differences in recording instrumentation, all affect the calculated averages and are well understood by climatologists compiling global meteorological data (e.g. New et al., 1999, 2002). More important is the spatial relationship between the climate station and the vegetation being sampled. Distance, altitude and aspect can introduce large differences between climate recorded at a meteorological station and that experienced by the sampled vegetation, again altering regressions and reducing precision and accuracy. In the case of LMA it becomes difficult to compare regional differences in LMA regressions or to understand the nature of uncertainty when applying the technique to the fossil record. This variability in climate data precision affects not only LMA and CLAMP but all terrestrial climate proxies. Wolfe (1993) used climate stations close to, and at the same altitude as, the vegetation sites in his CLAMP calibration. However as sampling for CLAMP was extended across Asia it became apparent that most climate stations were in agricultural land or urban centres far from the natural vegetation. The concept of ‘virtual climate stations’ was therefore introduced, where globally gridded climate data, initially at a 0.5° spatial resolution (New et al., 1999) but now at 0.16° (New et al., 2002), were interpolated to the exact location of the vegetation sample and adjusted for altitude following a clearly defined methodology (Spicer et al., 2009). By using a globally gridded dataset, regional differences in data recording, and spatial variation in the density of records could be reduced to a single uniform dataset using standard protocols and statistically determined uncertainties. Moreover the climate records underpinning the grid were all collected for the same time interval (1961–1990). Apart from providing a robust climate calibration for CLAMP this approach also allows regional LMA regressions to be compared within a standardised climatic framework for the first time.

E.M. Kennedy et al. / Palaeogeography, Palaeoclimatology, Palaeoecology 412 (2014) 80–90

Unfortunately the gridding process is not without limitations. Topographic heterogeneity, particularly on a spatial scale smaller than grid cell size, cannot capture aspect-related climate variation even when the gridded data are underpinned by digital elevation models. Interpolation schemes are oblivious to such ‘hidden’ topography. Temperature differences between north and south facing slopes can be in the order of several degrees Celsius, will vary between open and closed canopy sites, and depend on soil moisture content, particularly when dry air masses predominate (Fritts, 1961). To illustrate the magnitude of differences in temperature measures offered by three gridded datasets available for New Zealand Fig. 1 displays the maximum minus the minimum values for MAT, WMMT and CMMT for 49 sites shown in Fig. 2, encompassing the CLAMP sample sites used in our analyses. Although most of the sites exhibit only small differences in estimates, in some instances estimates of MAT differ by as much as 4.1 °C, the WMMT by 7.5 °C and the CMMT by 4.6 °C. The mean difference for MAT is 0.9 °C, the WMMT is 1.3 °C and for the CMMT it is 1.5 °C with standard deviations of 0.9, 1.3 and 1.1 °C respectively.

3.4. Datasets for this study Our starting point in this study is a pool of 239 sites derived from both hemispheres and encompassing an altitudinal range from 0.5 m a.m.s.l. to over 3500 m a.m.s.l. It is sourced from the CLAMP PHYSG3AR dataset with 12 additional sites from Bolivia (GregoryWodzicki, 2000), 2 sites from Argentina, 14 sites from South Africa (Steart et al., 2010), 21 from New Zealand (Stranks and England, 1997) and recently collected samples from Australia (21 sites) and New Zealand (14 sites). This dataset pool was subdivided to make two calibration sets; one for each hemisphere. The Northern Hemisphere calibration set (NH149) consists of 149 sites from the CLAMP PHYSG3AR dataset that are free from the influence of the Asian monsoon system and excludes sites where the MAT b = 4 °C. This site

83

selection ensured that the NH149 dataset was comparable to that of the Southern Hemisphere (SH90), which lacks sites with strong monsoonal climates and cold temperatures. The SH90 dataset consists of 90 sites distributed across Bolivia, Argentina, South Africa, Australia, New Caledonia, Fiji and New Zealand. The same high-resolution gridded climate data based on New et al. (2002), spatially interpolated and adjusted for height following procedures outlined in Spicer et al. (2009), was used in all analyses. The dataset combinations are listed in the online Supplementary information. 3.5. CLAMP analyses applied in this study The 239 site dataset was analysed in its entirety using Canonical Correspondence Analysis (ter Braak, 1986) before separate analyses of the NH149 and SH90 dataset combinations were made. To test the claim that the effectiveness of CLAMP depends upon the inclusion of leaf margin characters we excluded all six margin characters from additional CLAMP analyses of both the NH149 and SH90 datasets (Table 1). To explore the predictive capabilities of the two hemispheric calibration datasets, an ‘extract and replace’ or ‘jacknife’ approach was employed on NH149 and SH90, where each site was successively removed from the dataset, subjected to CLAMP analysis using the remaining sites as the calibration dataset, and then replaced in the dataset for the successive analysis. 4. Results 4.1. LMA (univariate) regressions — Northern and Southern hemisphere data combined The relationships between leaf margin form (percentage of entire margins; %E) and MAT across the entire dataset considered here are shown in Fig. 3. The regression equations and goodness of fit (R2) are

Fig. 1. Graph showing the range of estimates of temperature (Tmax–Tmin) exhibited by three gridded climate datasets across 49 sites in New Zealand. The sites are arranged in descending order of mean annual temperature (MAT) differences but warm month mean (WMMT) and cold month mean temperature (CMMT) differences are also given. The three meteorological datasets are a low resolution gridded dataset (at 0.5° spatial resolution) and a high resolution gridded dataset (at 0.16° spatial resolution) that can be generated through the CLAMP website (New et al., 1999, 2002; Spicer et al., 2009; http://clamp.ibcas.ac.cn), and a dataset generated using a New Zealand climate surface model on a 1 km grid (Leathwick and Stephens, 1998). Leaf morphology samples were collected using CLAMP protocols from all 49 of these sites, but some of these sites had low dicot angiosperm diversity and were not included in the SH90 dataset (see site list in the online Supplementary information).

84

E.M. Kennedy et al. / Palaeogeography, Palaeoclimatology, Palaeoecology 412 (2014) 80–90

Fig. 2. Map of New Zealand showing the locations of the sites featured in Fig. 1.

given in Table 2. It is immediately clear that overall the correlation between leaf margin form and MAT across the complete dataset is poor and displays a large scatter. Some regions, such as New Zealand, display a weak or non-existent correlation between %E and MAT, while other regions show markedly different regression line slopes. The North American and Japanese regression slopes are similar but with different intercepts and the slope of the global regression is shallower. Overall the island floras, and those of Australia, display

shallower regression slopes than the Northern Hemisphere and global regressions. When the data are divided into Northern and Southern Hemisphere cohorts the regression slopes are almost identical, but are separated by an intercept difference of 4.75 °C (Fig. 4; Table 2). For any given MAT, Southern Hemisphere leaves have a higher proportion of entire margins than those in the Northern Hemisphere. Such a relationship has been previously noted for Australia (Greenwood et al., 2004).

E.M. Kennedy et al. / Palaeogeography, Palaeoclimatology, Palaeoecology 412 (2014) 80–90

85

Table 1 The 31 leaf characters measured for CLAMP analysis. Definitions of all the characters and details of how to score the characters can be found on the CLAMP website (http://clamp.ibcas.ac.cn/). Lamina dissection

Margin characters

Size characters

Apex characters

Base characters

Length to width characters

Shape characters

Lobed

No teeth Teeth regular Teeth close Teeth round Teeth acute Teeth compound

Nanophyll Leptophyll I Leptophyll II Microphyll I Microphyll II Microphyll III Mesophyll I Mesophyll II Mesophyll III

Emarginate Round Acute Attenuate

Cordate Round Acute

L:W b1:1 L:W 1–2:1 L:W 2–3:1 L:W 3–4:1 L:W N4:1

Obovate Elliptic Ovate

4.2. CLAMP (multivariate) regressions — Northern and Southern Hemisphere data combined

4.3. Northern and Southern Hemisphere CLAMP predictive capabilities compared

Fig. 5 shows the equivalent plot for CLAMP and associated regression statistics are in Table 2. Here the horizontal axis records the position of the modern vegetation sites along the MAT vector (the MAT vector score) and the vertical axis displays the observed MAT for those sites. The vector score position of a site is determined by its position in multidimensional space based on scores for all the leaf character states including tooth characters. The inclusion of characters additional to those relating to the leaf margin (leaf size, shape, apex and base shape; Table 1) now centres the New Zealand sites on the global regression line; a similar effect is seen with other regions that with a single variable alone (percentage of entire margins) plot off the global regression. Consequently the overall scatter about the linear regression line is much reduced over that displayed by LMA and improves the predictive capability by approximately 25%.

4.3.1. Mean annual temperature Fig. 6 shows the CLAMP MAT regression characteristics of the Northern and Southern Hemisphere datasets individually using a simple linear regression. For any given vector score the SH90 dataset yields a slightly higher predicted MAT than the NH149 dataset but the slopes of the two regressions are almost identical. This shows that the same relationship between multiple leaf form characters and MAT exists in both hemispheres despite demonstrably different taxonomic compositions and biogeographic histories. The predominantly disconnected landmass ecosystems in the Southern Hemisphere display a similar convergent relationship with MAT to those of the more connected land surfaces of the Northern Hemisphere. The use of the 2nd order polynomial regressions (Fig. 7) raised the R2 values to 0.88 for the NH149 dataset and 0.82 for the SH90 dataset but reveals a tendency for the Southern Hemisphere data to predict a slightly higher MAT at low and high vector scores associated with larger scatter. This may be a function of the range extremities being dominated

Table 2 Regression statistics for Figs. 3 to 6.

Fig. 3. Leaf margin analysis (LMA) plot of the percentage of entire margined woody dicot leaves against mean annual temperature for the 239 Northern and Southern Hemisphere sites used in this analysis. Local regressions show regional variations in the relationship while the regression for all sites is shown by a grey line. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

No. of sites

Regression equation

R2

SD °C

Fig. 3 (LMA) All sites Australia Japan New Zealand North America Pacific Islands Puerto Rico South America Southern Africa

239 21 35 35 104 6 10 14 14

y y y y y y y y y

0.5733 0.0559 0.7396 0.0174 0.7919 0.3491 0.5904 0.9256 0.3708

3.9715 2.3389 1.8701 1.6893 2.6302 1.3303 0.7117 1.7282 2.3009

Fig. 4 (LMA) SH linear regression NH linear regression

90 149

y = 0.257x − 1.8412 y = 0.2475x + 2.9066

0.40406 0.81236

4.8258 2.5521

Fig. 5 (CLAMP) All sites Australia Japan New Zealand North America Pacific Islands Puerto Rico South America Southern Africa

239 21 35 35 104 6 10 14 14

Y Y Y Y Y Y Y Y Y

0.7961 0.368 0.3295 0.1514 0.8724 0.0591 0.0017 0.9161 0.2704

2.7454 1.9137 3.0008 1.5699 2.0593 1.5994 1.1111 1.8352 2.4777

Fig. 6 (CLAMP) SH linear regression NH linear regression

90 149

y = 4.3848x + 14.3563 y = 5.3029x + 14.0566

0.7564 0.8279

3.0852 2.4439

= = = = = = = = =

= = = = = = = = =

0.2039x 0.0571x 0.2338x 0.0239x 0.2471x 0.0862x 0.1446x 0.4396x 0.1094x

+ + + + + + + − +

3.6562 16.4785 4.6321 10.561 2.3305 17.6547 12.8523 8.3637 9.5173

4.8796x + 14.1031 1.9164x + 18.4872 2.9836x + 13.3254 1.5051x + 10.2956 5.607x + 14.0095 1.7087x + 19.8836 −0.1569x + 24.6882 5.7524x + 18.5936 1.8984x + 16.0432

86

E.M. Kennedy et al. / Palaeogeography, Palaeoclimatology, Palaeoecology 412 (2014) 80–90

Fig. 4. LMA plot showing the same data as in Fig. 3 but divided into Northern and Southern Hemisphere cohorts. Although the slopes are similar there is considerably more scatter in the Southern Hemisphere data.

by New Zealand at the cool end and northeastern Australia at the warm end, both displaying a degree of variation between sites. The standard deviation of the NH149 MAT regression is 2.1 °C, while that of SH90 is 2.7 °C. The predictive power of the regression models is demonstrated in Fig. 8. This displays the relationship between the CLAMP-predicted MAT and temperatures observed in terms of the high-resolution gridded data. Perfect prediction would equate to the observed and result in the linear regression in Fig. 8 having a slope of 1, an intercept of zero and an R2 of 1. The NH149 dataset has an R2 of 0.87 while that of SH90 is 0.79.

Fig. 5. The observed mean annual temperature (MAT) versus the CLAMP MAT vector score for the combined Northern and Southern Hemisphere datasets identified by region. This plot is comparable to Fig. 3 but shows much less dispersion about the overall regression line and New Zealand now plots in line with the other sites.

Fig. 6. The observed mean annual temperature (MAT) versus the CLAMP MAT vector score for the Northern and Southern Hemisphere data cohorts identified separately and with linear regressions.

4.3.2. CLAMP hemispheric comparison for MAT with leaf margin characters removed In order to investigate the importance of leaf margin characters to the ability of CLAMP to predict MATs accurately and precisely (see Section 2.2) we removed all six leaf margin characters from both the Northern and Southern Hemisphere physiognomic datasets and repeated the analysis described in Section 4.3.1.

Fig. 7. The observed mean annual temperature (MAT) versus the CLAMP MAT vector score for the Northern and Southern Hemisphere data cohorts identified separately with 2nd order polynomial regressions.

E.M. Kennedy et al. / Palaeogeography, Palaeoclimatology, Palaeoecology 412 (2014) 80–90

87

Fig. 10. The observed warm month mean temperature (WMMT) versus the CLAMP WMMT vector score for the Northern and Southern Hemisphere data cohorts identified separately with 2nd order polynomial regressions.

Fig. 8. The relationship between the CLAMP-predicted MAT and that which is observed in terms of the high resolution gridded data.

Fig. 9 shows that the effect of removing all tooth characters is minimal. The shapes of the regression models are almost identical to those of Fig. 7. The removal of tooth characters from the SH90 dataset results in a smaller change (0.028) in the R2 value than when tooth characters are removed from the NH149 dataset (0.08), but the effect on both datasets is small.

Fig. 9. The observed mean annual temperature (MAT) versus the CLAMP MAT vector score for the Northern and Southern Hemisphere data cohorts identified separately with 2nd order polynomial regressions when all tooth-related physiognomic character states were removed from the analysis. Note the similarity to Fig. 7 that shows the results for the complete dataset.

4.3.3. Warm month mean temperature The CLAMP WMMT regression models for both calibration datasets are shown in Fig. 10. Overall the R2 values are not as good as for MAT, although that for SH90 (0.8) is better than that for NH149 (0.68). The CLAMP predictive capabilities of the two calibrations for the WMMT are shown in Fig. 11. Here the Northern and Southern Hemisphere slopes are quite similar with the Southern Hemisphere being slightly steeper and with a higher R2.

Fig. 11. The relationship between the CLAMP-predicted WMMT and that which is observed in terms of the high-resolution gridded climate data. Note the higher R2 value and slightly steeper slope of the Southern Hemisphere cohort compared to that representing the Northern Hemisphere.

88

E.M. Kennedy et al. / Palaeogeography, Palaeoclimatology, Palaeoecology 412 (2014) 80–90

datasets produced good R2 values, 0.77 for SH90 and 0.80 for NH149. A prediction of 0 °C provided by the NH149 dataset equates to an observed CMMT of 1.0 °C, whereas a prediction of 0 °C from the SH90 dataset equates to an observed CMMT of 2.3 °C. 5. Discussion 5.1. Universality, margin characters, and multi-character coding for climate

Fig. 12. The observed cold month mean temperature (CMMT) versus the CLAMP CMMT vector score for the Northern and Southern Hemisphere data cohorts identified separately with 2nd order polynomial regressions.

4.3.4. Cold month mean temperature The CLAMP regression models for both hemispheres regarding the cold month mean are shown in Fig. 12. The R2 values are similar and reasonably good, both being above 0.81. The results of the tests of predictive power are shown in Fig. 13. The predictive power of both SH90 and NH149 is similar for CMMT. Both

It is clear from our results that hemispheric differences in the relationship between leaf margin type and temperature largely disappear when foliar characters such as leaf size and shape are included in the analysis. Moreover the removal of all tooth characters shows that the remaining features code for MAT almost as well as the full suite of characters. This confirms that leaf physiognomic relationships with climate involve numerous characters and that climate data are encoded in several characters simultaneously. Our results also suggest a high degree of redundancy in leaf form as far as coding for climate, which adds resilience to the CLAMP palaeoclimate proxy in the face of taphonomic information loss (Spicer et al., 2005). Biogeographic factors may result in individual leaf characters, such as margin form, having a different relationship with individual climate parameters to that seen elsewhere. In the case of New Zealand the percentage of entire margined woody dicot leaves shows no relationship with MAT but when other characters are included modern New Zealand vegetation falls into line with other floras that show a clear relationship between leaf form and MAT. A single character, such as that used in LMA, displays pronounced regional variations in its relationship with climate. However, when a suite of leaf characters is examined together, the optimised interplay among characters falls within the global range of possible solutions for adaptation to a particular climate. The use of a multivariate approach to foliar physiognomy therefore also provides greater resilience to evolutionary change and biogeographic history than when single characters are used. 5.2. Climate data quality The predictive uncertainties and inaccuracies associated with these datasets have to be viewed in the light of the quality of the climate data used to calibrate the method. As shown by the climate data for New Zealand (Fig. 1) there is a mean difference of 0.9 °C for MAT and 1.5 °C for the CMMT between three gridded datasets describing the modern climate of that region. The standard deviations (SD) as a measure of spread about these means are around 1 °C, so the usual measure of uncertainty of ±2 SD encompasses a spread of approximately 4 °C. By comparison with many parts of the world, the climate station data for New Zealand is tolerably good. Nevertheless if we take the New Zealand data to be typical of much of the developed world, expecting any climate reconstruction technique based on this quality of climate calibration data to have a precision better than ~ 2 °C is unrealistic. This level of uncertainty is not a function of plant sensitivity to climate but purely the ‘noise’ in our measurement of modern climate, and the gridding process. 5.3. Application of a Southern Hemisphere calibration

Fig. 13. The relationship between the CLAMP-predicted CMMT and that which is observed in terms of the high-resolution gridded climate data. Here the Southern Hemisphere cohort (SH) has a similar R2 value to the Northern Hemisphere (NH) but both cohorts display errors at the intercept with the SH cohort underpredicting temperature by 2.3 °C and the NH cohort underpredicting by 1.0 °C.

Analyses of previous CLAMP calibration datasets (Jacques et al., 2011; Yang et al., 2011; Srivastava et al., 2012) have suggested that the larger the calibration set, the lower the precision. The reason for this most likely lies in the complexity of the physiognomy/climate relationships. As more climate types are included, the range of optimisation compromises present in leaf form increases. As an example, consider the contrast in adaptations between a maritime climate such as that of New Zealand, where severe seasonal drought is rare compared to a monsoon climate, such as in India or southern China, where severe drought alternates with extremely high seasonal rainfall. Evergreen

E.M. Kennedy et al. / Palaeogeography, Palaeoclimatology, Palaeoecology 412 (2014) 80–90

species in monsoon climates have to be adapted to both dry and wet conditions, while dry deciduous leaves only have to be adapted to the wet season. Unsurprisingly, leaf physiognomic spectra are distinctive in monsoon vegetation (Jacques et al., 2011). Simple vectors are a poor way to summarise these complexities in physiognomic space and as a result there is greater scatter in climate parameter regression models as more sites and climate regimes are added. However, at the same time restricting the number of sample sites and climate regimes in the calibration compromises the ability of the technique to encompass climate regimes of the past. For example, the northward passage over time of the New Zealand landmass combined with secular climate change has inevitably resulted in changes in New Zealand vegetation and climate through time. It would therefore be inappropriate to calibrate CLAMP for New Zealand fossil assemblages using only modern New Zealand sites. Previous work on several New Zealand fossil floras (e.g. Kennedy et al., 2002; Kennedy, 2003; Pancost et al., 2013; Reichgelt et al., 2013a,b) has successfully applied the standard Northern Hemisphere-dominated CLAMP calibration datasets to leaf assemblages from the Late Cretaceous, Paleocene, Early Eocene and Miocene. Again, we qualify the term ‘successful’ here to be that the climate estimates produced from CLAMP in these studies have for the most part been consistent with other available proxy data. Here we introduce a standardised calibration dataset for the Southern Hemisphere as a whole instead of one for smaller regionally specific floras and climate such as those of Australia or New Zealand. This Southern Hemisphere dataset can be used as an alternative, or in addition, to the Northern Hemisphere-dominated CLAMP calibration datasets. Application of both the Southern and the Northern Hemisphere calibration datasets, as well as validation from independent climate proxy data, may improve the robustness of interpretations of palaeoclimate from Southern Hemisphere leaf fossil assemblages. In the case of New Zealand, the Southern Hemisphere calibration may be more appropriate for geologically young New Zealand floras (post-Miocene) than the existing Northern Hemisphere-dominated datasets, but this is yet to be tested. 6. Conclusions Is the relationship between leaf margin form and temperature the same in Northern and Southern hemispheres? The relationship between leaf margin and temperature (the Leaf Margin Analysis calibration) is variable, both between hemispheres and regionally. When other leaf characters are included in a multivariate analysis the hemispheric differences largely disappear. Is the relationship between leaf margin and temperature always the most significant leaf character/climate relationship in these methodologies? No, not always. In some instances no clear relationship exists at all. In a multivariate context non-margin leaf characters code for temperature. A CLAMP calibration dataset of 90 Southern Hemisphere leaf assemblages and associated gridded climate data gives temperature estimates with a precision similar to those of Northern Hemisphere datasets. Acknowledgements EMK was funded by the GNS Science Global Change through Time Programme (GCT). RAS was supported by a William Evans Fellowship from Otago University, TR was funded by GCT and Otago University. LS, EMK, RAS, TEVS and TR gratefully acknowledge the assistance of David Norton, Nick Akerman, Richard Jongens, Nick Moore, Graeme, Win and Jon Kennedy, and Jane Chewing in the collection of New Zealand samples during a number of trips. Robyn Burnham provided advice on collection techniques. John Leathwick facilitated access to the Landcare Research, New Zealand climate data and advised on its application. Field work by NCA in Australia was partially supported by the Center for Global Education, Hobart & William Smith Colleges and

89

David Kendrick provided essential support in the field. We thank Ian Raine, Joe Prebble and two anonymous reviewers for constructive comments on the manuscript. Appendix A. Supplementary data Supplementary data to this article can be found online at http://dx. doi.org/10.1016/j.palaeo.2014.07.015. References Bailey, I., Sinnott, E., 1915. A botanical index of Cretaceous and Tertiary climates. Science 41, 831–834. Bailey, I.,Sinnott, E., 1916. The climatic distribution of certain types of angiosperm leaves. Am. J. Bot. 3, 24–39. Benzecri, J.P., 1973. L'analyse des données: L'analyse des correspondences. Dunod, Paris. Dolph, G.E., Dilcher, D.L., 1980. Variation in leaf size with respect to climate in Costa Rica. Biotropica 12, 91–99. Fritts, H.C., 1961. An analysis of maximum summer temperatures inside and outside a forest. Ecology 42, 436–440. Greenwood, D.R., 1992. Taphonomic constraints on foliar physiognomic interpretations of Late Cretaceous and Tertiary palaeoclimates. Rev. Palaeobot. Palynol. 71, 149–190. Greenwood, D.R., Moss, P., Rowett, A., Vadala, A.J., Keefe, R., 2003. Plant communities and climate change in southeastern Australia during the Early Paleogene. In: Wing, S., Gingerich, P., Schmitz, B., Thomas, E. (Eds.), Causes and Consequences of Globally Warm Climates in the Early Paleogene. Geological Society of America Special Paper, Boulder, Colorado, 369, pp. 365–390. Greenwood, D.R., Wilf, P., Wing, S.L., Christophel, D.C., 2004. Paleotemperature estimation using leaf-margin analysis: is Australia different? Palaios 19, 129–142. Gregory, K.M., McIntosh, W.C., 1996. Paleoclimate and paleoelevation of the Oligocene Pitch-Pinnacle flora, Sawatch Range, Colorado. Geol. Soc. Am. Bull. 108, 545–561. Gregory-Wodzicki, K.M., 2000. Relationships between leaf morphology and climate, Bolivia: implications for estimating paleoclimate from fossil floras. Paleobiology 26, 668–688. Hill, M.O., 1973. Reciprocal averaging — an eigenvector method of ordination. J. Ecol. 61, 237–249. Hill, M.O., 1979. Correspondence analysis — a neglected multivariate method. Appl. Stat. 23, 340–354. Hinojosa, L.F., Villagrán, C., 2005. Did South American mixed paleofloras evolve under thermal equability or in the absence of an effective Andean barrier during the Cenozoic? Palaeogeogr. Palaeoclimatol. Palaeoecol. 217, 1–23. Jacques, F.M.B., Tao, S., Spicer, R.A., Xing, Y., Huang, Y., Wang, W., Zhou, Z., 2011. Leaf physiognomy and climate: are monsoon systems different? Glob. Planet. Chang. 76, 56–62. Jordan, G.J., 1997. Uncertainty in palaeoclimatic reconstructions based on leaf physiognomy. Aust. J. Bot. 45, 527–547. Kennedy, E.M., 1998. Cretaceous and Tertiary megafloras from New Zealand and their climate signals(Ph.D. Thesis) Earth SciencesThe Open University, Milton Keynes, United Kingdom. Kennedy, E.M., 2003. Late Cretaceous and Paleocene terrestrial climates of New Zealand: leaf fossil evidence from South Island assemblages. N. Z. J. Geol. Geophys. 46, 295–306. Kennedy, E.M.,Spicer, R.A.,Rees, P.M., 2002. Quantitative paleoclimate estimates from Late Cretaceous and Paleocene leaf floras in the northwest of the South Island, New Zealand. Palaeogeogr. Palaeoclimatol. Palaeoecol. 184, 321–345. Kent, M., Coker, P., 1992. Vegetation Description and Analysis: A Practical Approach. Belhaven Press, London. Kovach, W.L.,Spicer, R.A., 1995. Canonical correspondence analysis of leaf physiognomy: a contribution to the development of a new palaeoclimatological tool. Palaeoclimates 1, 125–173. Kowalski, E.A., Dilcher, D.L., 2003. Warmer paleotemperatures for terrestrial ecosystems. Proc. Natl. Acad. Sci. U. S. A. 100 (1), 167–170. Leathwick, J.R.,Stephens, R.T.T., 1998. Climate surfaces for New Zealand. Landcare Research Contract Report LC9798/126, June 1998. Landcare Research, Hamilton. Legendre, P., 1993. Spatial autocorrelation: trouble or new paradigm? Ecology 74, 1659–1673. Little, S.A., Kembel, S.W., Wilf, P., 2010. Paleotemperature proxies from leaf fossils reinterpreted in light of evolutionary history. PLoS ONE e15161. Miller, I.M., Brandon, M.T., Hickey, L.J., 2006. Using Leaf Margin Analysis to estimate the mid-Cretaceous (Albian) paleolatitude of the Baja BC block. Earth Planet. Sci. Lett. 245, 95–114. New, M., Hulme, M., Jones, P., 1999. Representing twentieth-century space–time climate variability. Part I: development of a 1961–90 mean monthly terrestrial climatology. J. Clim. 12, 829–856. New, M.,Lister, D.,Hulme, M.,Makin, I., 2002. A high-resolution data set of surface climate over global land areas. Clim. Res. 21, 1–15. Pancost, R.D., Taylor, K., Inglis, G.N., Kennedy, E.M., Handley, L., Hollis, C.J., Crouch, E.M., Pross, J., Huber, M., Schouten, S., Pearson, P.N., Morgans, H.E.G., Raine, J.I., 2013. Early Paleogene evolution of terrestrial climate in the SW Pacific, Southern New Zealand. Geochem. Geophys. Geosyst. 14. http://dx.doi.org/10.1002/2013GC004935. Povey, D.A.R., Spicer, R.A., England, P.C., 1994. Palaeobotanical investigation of Early Tertiary palaeoelevations in northeastern Nevada: initial results. Rev. Palaeobot. Palynol. 81, 1–10.

90

E.M. Kennedy et al. / Palaeogeography, Palaeoclimatology, Palaeoecology 412 (2014) 80–90

Reichgelt, T.,Kennedy, E.M.,Mildenhall, D.C.,Conran, J.G.,Greenwood, D.R.,Lee, D.E., 2013a. Quantitative palaeoclimate estimates for Early Miocene southern New Zealand: evidence from Foulden Maar. Palaeogeogr. Palaeoclimatol. Palaeoecol. 378, 36–44. Reichgelt, T., Kennedy, E.M., Conran, J.G., Lee, D.E., 2013b. Using floral proxies to reconstruct Late Eocene to Late Miocene terrestrial paleoclimates in southern New Zealand. In: Reid, C.M., Wandres, A. (Eds.), Abstracts, Geosciences 2013 Conference, Christchurch, New Zealand. Geoscience Society of New Zealand Miscellaneous Publication, 136A, pp. 76–77. Spicer, R.A.,Herman, A.B., 2010. The Late Cretaceous environment of the Arctic: a quantitative reassessment using plant fossils. Palaeogeogr. Palaeoclimatol. Palaeoecol. 295, 423–442. Spicer, R.A.,Yang, J., 2010. Quantification of uncertainties in fossil leaf paleoaltimetry — does leaf size matter? Tectonics 29, TC6001. http://dx.doi.org/10.1029/2010TC002741. Spicer, R.A.,Herman, A.B.,Kennedy, E.M., 2004. The foliar physiognomic record of climatic conditions during dormancy: CLAMP and the cold month mean temperature. J. Geol. 112, 685–702. Spicer, R.A., Herman, A.B., Kennedy, E.M., 2005. The sensitivity of CLAMP to taphonomic loss of foliar physiognomic characters. Palaeogeography, Palaeoclimatology, Palaeoecology 184, 321–345. Spicer, R.A., Valdes, P.J., Spicer, T.E.V., Craggs, H.J., Srivastava, G., Mehrotra, R.C., Yang, J., 2009. New developments in CLAMP: calibration using global gridded meteorological data. Palaeogeogr. Palaeoclimatol. Palaeoecol. 283, 91–98. Spicer, R.A., Bera, S., De Bera, S., Spicer, T.E.V., Srivastava, G., Mehrotra, R.C., Mehrotra, N.C., Yang, J., 2011. Why do foliar physiognomic climate estimates sometimes differ from those observed? Insights from taphonomic information loss and a CLAMP case study from the Ganges Delta. Palaeogeogr. Palaeoclimatol. Palaeoecol. 299, 39–48. Srivastava, G., Spicer, R.A., Spicer, T.E.V., Yang, J., Kumar, M., Mehrotra, R.C., Mehrotra, N.C., 2012. Megaflora and palaeoclimate of a Late Oligocene tropical delta, Makum Coalfield, Assam: evidence for the early development of the South Asia Monsoon. Palaeogeogr. Palaeoclimatol. Palaeoecol. 343, 130–142. Steart, D.C.,Spicer, R.A.,Bamford, M.K., 2010. Is Southern Africa different? An investigation of the relationship between leaf physiognomy and climate in southern African mesic vegetation. Rev. Palaeobot. Palynol. 162, 607–620.

Stranks, L., 1996. Physiognomic and taphonomic studies in New Zealand and Australia; implications for the use of palaeobotany as a tool for palaeoclimate estimation(Ph.D. Thesis) Earth SciencesOxford University. Stranks, L., England, P., 1997. The use of a resemblance function in the measurement of climatic parameters from the physiognomy of woody dicotyledons. Palaeogeogr. Palaeoclimatol. Palaeoecol. 131, 15–28. ter Braak, C.J.F., 1986. Canonical correspondence analysis: a new eigenvector technique for multivariate direct gradient analysis. Ecology 67, 1167–1179. Uhl, D., Traiser, C.,Griesser, U.,Denk, T., 2007. Fossil leaves as palaeoclimate proxies in the Palaeogene of Spitsbergen (Svalbard). Acta Palaeobot. 47, 89–107. Upchurch, G.R.,Wolfe, J.A., 1987. Mid-Cretaceous to Early Tertiary vegetation and climate: evidence from fossil leaves and woods. In: Friis, E.M., et al. (Eds.), The Origin of Angiosperms and Their Biological Consequences, pp. 75–105. Wiemann, M.C.,Manchester, S.R.,Dilcher, D.D.,Hinojosa, L.F.,Wheeler, E.A., 1998. Estimation of temperature and precipitation from morphological characters of dicotyledonous leaves. Am. J. Bot. 85, 1796–1802. Wilf, P., 1997. When are leaves good thermometers? A new case for Leaf Margin Analysis. Palaeobiology 23, 373–390. Wilf, P., Wing, S.L., Greenwood, D.R., Greenwood, C.L., 1998. Using fossil leaves as paleoprecipitation indicators: an Eocene example. Geology 26, 203–206. Wing, S.L., Greenwood, D.R., 1993. Fossils and fossil climate: the case for equable continental interiors in the Eocene. Philos. Trans. R. Soc. Lond. B Biol. Sci. 341, 243–252. Wolfe, J.A., 1979. Temperature parameters of humid to mesic forests of eastern Asia and relation to forests of other regions of the Northern Hemisphere and Australasia. Geol. Surv. Prof. Pap. 1106, 1–37. Wolfe, J.A., 1993. A method of obtaining climatic parameters from leaf assemblages. US Geol. Surv. Bull. 2040, 1–73. Yang, J., Spicer, R.A., Spicer, T.E.V., Li, C.-S., 2011. ‘CLAMP Online’: a new web-based palaeoclimate tool and its application to the terrestrial Paleogene and Neogene of North America. Palaeobiodivers. Palaeoenviron. 91, 163–183.