Pollen-based temperature and precipitation inferences for the montane forest of Mt . Kilimanjaro during the last Glacial and the Holocene

ing multivariate statistical methods. Canonical correspondence analysis revealed that the mean annual temperature, mean annual precipitation and minimum temperature each account for significant fractions of the variation in pollen taxa. A training set of 107 modern pollen taxa was used to derive temperature and precipitation transfer functions based on pollen subsets using weighted-averaging-partial-least-squares 10


Introduction
Many environmental issues, like global warming or altered precipitation patterns, have increased the interest in fossil species assemblages as indicator of the palaeoenvironment.In palaeoecology the direct measurement of environmental variables is impossible; consequently there is the need to resort to indirect methods.Fossil taxa assemblages can provide a record of the palaeoenvironment since each biological species requires particular environmental conditions for regeneration, establishment and growth.Today, this principle idea is primarily used in biomonitoring (Spellerberg, 1991).Nevertheless, it is also possible to infer the past environmental conditions at a site from the species composition that occurred there.Fossil taxa records are valuable for obtaining a historical perspective of current environmental problems such as acid rain (Battarbee, 1984) and global warming (Fritz et al., 1991;Walker et al., 1991).
As mentioned above, it requires a representative training set consisting of data on species assemblages and associated environmental measurements.Relationships between species and environmental variables are generally nonlinear.Because of Shelford's law of tolerance (Odum, 1971) and niche-space partitioning (Whittaker et al., 1973), species abundance or probability of occurrence is often a unimodal function of the environmental variables.
Palaeoecological transfer functions share the several characteristics (Sachs et al., 1977).They produce calibrated quantitative estimates of environmental parameters of a past environment.They make use explicit algebraic methods to formulate these estimates.The applied algorithms rely on multivariate techniques to analyse multicomponent fossil data.The transfer functions are calibrated from a training data set, an adequate sample of modern distributional data.The calibrated function is then applied to older samples to estimate environmental parameters for past times.
In palaeo-ecology, the weighted averaging (WA) method gained popularity because it combines niche-space partitioning with empirical predictive power (Birks et al., 1990;Fritz et al., 1991;ter Braak and Juggins, 1993).Ter Braak and Juggins (1993) could show that the addition of a partial least square regression (PLS) to the weighted averaging (WA) method could even further improve the predictive power.They demonstrated that the application of weighted averaging partial least square regression (WA-PLS) could reduce the prediction error by up to 70 %.The main advancement of the WA-PLS procedure is that it is unimodal-based (since it is based on the niche-space partitioning concept and ecological optima of species) unlike the PLS which is linearbased (ter Braak and Juggins, 1993;Birks, 1993).Further strength of this multivariate calibration-function approach is its relative robustness to spatial autocorrelation and its applicability in extrapolation of parameters (Birks et al., 2010;Bartlein et al., 2011).These recent advances in quantitative environmental reconstruction procedures now allow palaeoecologicts to explore many critical ecological questions, answers to which require the unique long-term temporal perspective provided by palaeoecology.What are the rates of change in ecological processes?Which processes result in given ecological thresholds being crossed, or resilience?Did novel systems, non-analogue systems, occur in the past and which combination of factors did lead to such systems (Willis et al., 2010a)?Answers to these questions (Froyd and Willis, 2008;Willis et al., 2010b;Macdonald et al., 2008) highlight the contributions that palaeoecological reconstruction can make in understanding ecological and evolutionary processes responsible for biodiversity patterns (Birks et al., 2010).
The central assumption in plant ecology is that climate has the dominant control on the distribution of the major vegetation types of the world.The climate space required to describe taxa distributions requires at least two dimensions related to temperature and water availability because of the existence of bioclimatic limits on the distribution of taxa (Woodward and Williams, 1987;Harrison et al., 2010).At the global scale responses to extreme minimum temperatures and to the hydrological budget predict the distribution limits of the major vegetation (Woodward and Williams, 1987).Further, frost stress plays a decisive role in plant distribution since it serves as a very selective filter and operates over very long time scales (Larcher and Bauer, 1981;Sakai and Larcher, 1987).Hence, the composition of natural vegetation in an area is also adjusted to local low temperature extremes (Körner, 2003).
Fossil pollen records provide evidence of changing plant compositions through time (Pickett et al., 2004;Marchant et al., 2009;Prentice et al., 2000).Pollen can often be identified to genus and family level, but only rarely to species level.Higher taxa levels tend to have distributions as coherent as lower taxa or species in climate space (Wake et al., 2009;Huntley et al., 1989).Bioclimatic limits determine the fundamental niche of taxa, but niches of taxa, especially at higher level, commonly overlap.They also vary in abundance; either with maximum values near physiological optima, or displaced by competition.Therefore more-or-less continuous variation in taxa abundance along environmental gradients in space are found, and we can establish relationships between taxa composition and climate parameter space (Jackson and Williams, 2004).
This present study aims at the development of an ecologically plausible and significant transfer functions through the calibration of the relationships between modern pollen assemblages, vegetation belts and several ecologically determinant climate parameters such as mean annual temperature (MAT), mean annual precipitation (MAP) and minimum temperature (Tmin) in the Kilimanjaro area.The transfer functions are then applied to the fossil WeruWeru 26 pollen record from 2600 m a.s.l. on the SE slope of Mt.Kilimanjaro to reconstruct the past vegetation and climate dynamics of equatorial East Africa of the past about 50 kyr BP.

Study site
Mt. Kilimanjaro is a relict of an ancient volcano, which was formed as part of the East African Rift Zone.It is located at the border to Kenya 300 km south of the equator in Tanzania and about 300 km west of the Indian Ocean (Fig. 1a).It rises from savanna plains at 700 m up to its glaciated summit at 5895 m a.s.l.

Vegetation
Mt. Kilimanjaro has several different bioclimatic zones (Fig. 1b and c): the dry and hot colline savanna zone surrounds the mountain base between 700 and 1100 m a.s.l.Most of this area is used for crop production (maize, beans and sunflowers) or as meadows.Remnants of the former savanna woodlands (Acacia, Terminalia, Grewia, and Combretum) are encountered mainly around Lake Chala in the eastern foothills and on the northwestern side of the mountain.The submontane and lower montane forest belts between 1000 and 1800 m have been converted to coffee-banana fields, a special type of agroforestry called "Chagga home gardens".Remnants of the former forests of this belt (Newtonia, Strombosia and Entandrophragma) are mainly restricted to deep valleys and gorges.Montane tropical rain and cloud forests on Mt Kilimanjaro cover an area of about 1000 km 2 .
The WeruWeru 26 (WW26) study site (03 • 07 18.1 S, 37 • 16 17.7 E) is located at 2650 m a.s.l. on the southern slope of Mt.Kilimanjaro in the WeruWeru river area (Fig. 1b and c).In the lower montane zone (1800-2000 m a.s.l.) of the WeruWeru area Macaranga kilimandscharica, Agauria salicifolia and Ocotea umsabarensis are the dominant tree species.This part of Kilimanjaro's forests is influenced by human activities such as fodder and wood collection and grazing by cattle.WW26 is located in the mid montane zone, which starts in the WeruWeru area at 2000 m and end at 2700 m elevation.In this zone, humidity reaches its maximum, indicated by the wealth of epiphytes and ferns, in particular filmy ferns and tree ferns.Evergreen forest species such as Ocotea umsabarensis, Xymalos monospora, Ilex mitis and Schefflera volkensii build up the tree layer.More detailed description of these vegetation types are published by Hemp (2001Hemp ( , 2006a, b), b).

Climate
Due to its equatorial location, the Mt.Kilimanjaro area is characterized by a typical tropical daytime climate.The precipitation pattern follows the position of the intertropical convergence zone (ITCZ) and is strongly influenced by the activity of the Indian monsoon system.According to Köppen and Troll/Pfaffen (in Müller, 1989), Mt.Kilimanjaro can be assigned to a seasonal dry tropical climate with two pronounced wet seasons: a short one in November/December and a longer one from March to May.Annual precipitation is modified by elevation and exposure to prevailing winds blowing inland from the Indian Ocean and varies on the wet southern slope from 500 mm at the mountain foothills to about 3000 mm at 2200 m a.s.l.(Fig. 1b).At higher altitudes of 2700 m, 3000 m and 4000 m a.s.l., rainfall amounts to roughly 70, 50 and 20 %, respectively of this annual maximum (Hemp, 2001(Hemp, , 2006b)).The northern slopes receive less annual rainfall than the southern slopes.The mean annual temperature decreases from 23.4 • C at the foothills in Moshi, 813 m a.s.l.(Walter et al., 1975)  at the summit (Kibo, 5895 m a.s.l.) (Thompson et al., 2002).Temperature lapse rates vary from 0.51 to 0.56 • C per 100 m elevation (Hemp, 2006b).Frost occurs from 2700 m upwards (Hemp, 2006a).
3 Material and method

The pollen record
In total 76 samples were taken every 2 cm from a 165 cm long WeruWeru 26 (WW26) sediment core.The core was obtained from a soil pit by hammering five 50 cm long zinc metal cases consecutively into the pit wall.Due to the high content of un-decomposed material it was only possible to extract sediment from 15 cm below the soil surface downwards.The material was wrapped into plastic foil and kept under cool and dark conditions until transported back to the University of Göttingen.For pollen analysis a sample volume of 0.5 cm 3 was prepared in the lab applying the standard method (Faegri and Iversen, 1989).The extracted pollen samples were mounted in glycerine jelly for pollen identification and counting.
The chronology is based on 13 AMS dated samples (see Supplement); 11 samples were dated at the Physical Institute of the University Erlangen-Nuremberg/Germany and two samples by Beta Analytic in Miami, Florida/USA.According to the results, samples from 17 to 52 cm core depth can be assigned to the Holocene.Samples between 66 and 165 cm core depth belong to the Late Pleistocene.There was no pollen in the sediment below 123 cm depth.

Modern pollen-rain data
The pollen and climate data sets used for the training data set included the 14 plots of the WeruWeru transect on the southern slope of Kilimanjaro.The modern pollen-rain was captured in pollen traps installed on 14 montane forest plots every 100 m elevation between 1900 and 3200 m a.s.l.On each plot the rainfall was recorded using dipping bucket rain gauges and in parallel funnel-gauges.Data loggers were used for temperature.The instruments were accurate within ±1 mm and ±0.1 • C, respectively.Mean annual precipitation (MAP), mean annual temperature (MAT) and absolute minimum temperature (Tmin) were calculated based on data recorded from the years 2005 to 2013.For the climate parameter reconstruction the developed transfer functions were applied to the data set of the WW26 sediment core containing in total 76 samples.Chemical preparation of past and modern pollen samples followed standard methods (Faegri and Iversen, 1989).The extracted pollen samples were mounted in glycerine jelly for pollen identification and counting.At least 300 pollen grains were counted for each sample.Pollen determination was carried out using light microscopy with 400× and 1000× magnification.Pollen identification referred to collections at Göttingen University, the African Pollen Database (Lézine, 2001) and Assemien et al. (1974).The pollen sum was used for the calculation of the pollen taxa abundance (in %).The fern spore abundance (%) was computed based on the pollen sum but it is not included in it.

Numerical analyses and subset selection
An ordination derives quantitative information on the relationship among species, environmental variables and sites (Legendre and Legendre, 2012) In quantitative environmental reconstructions the first step is to examine whether linear or unimodal methods are appropriate for the available training set in relation to the environmental variables of interest (Birks, 1995;Lotter et al., 1997).For this purpose the modern pollen-rain data set was analysed by Detrended Correspondence Analysis (DCA) to estimate the lengths of environmental gradient inherit in the data.To visualize the response of each pollen taxa to the climate variables, the taxa percentage at each site/elevation was plotted again the climate parameters measured at these sites.A local regression function (alpha: 0.45) was fitted to visualize the response trend.Further, in order to estimate the explanatory power of MAP, MAT and Tmin, three Canonical Correspondence Analyses (CCA) were performed using each climate parameter as sole constraining variable.This ordination method is preferably applied to data of species abundance and relative frequency that have a large number of zero values (Legendre and Legendre, 2012).DCAs and CCAs were carried out using CANOCO 4.54 (ter Braak and Šmilauer, 1997;2002) and summarized in Table 2.
Forward selection was used to select a subset of the explanatory pollen taxa for each climate variable individually so that each pollen taxa present in the subset specifically contributes to the explanatory power of the subset.Pollen taxa with a contribution of ≥ 20 %, representing the percentage contribution of the particular taxa to the explanatory power of the whole subset of taxa, were selected (Table 3).

WAPLS climate reconstruction and assessment of model performance
For the climate variables MAP, MAT and Tmin, pollen-based inference models were developed using weighted-averaging-partial-least-squares (WAPLS) models (ter Braak and Juggins, 1993).The WA-PLS procedure and the reconstruction was computed using the program R and the "rioja" package (Juggins, 2012).In applications like the present one, large test sets are not available; instead the prediction errors are simulated by bootstrap cross-validation (number of boot cycles: 1000).The goal of bootstrapping is to assign measures of accuracy to sample estimates.Since it is not easily possible to quantitatively estimate to what degree pollen taxa abundance represents rather a function of local origin and/or pollen under-or overrepresentation, three WA-PLS calibration functions were developed, each on nontransformed and transformed species subsets (as selected by forward selection for each climate variable) for a comparison.Finally, the models were used to reconstruct MAP, MAT and Tmin based on the WW26 pollen data set.The root-mean-square error of prediction (RMSEP) is a frequently used measure of the differences between values predicted by a model and the values observed.Further, checking the predictive worth of an environmental model includes a goodness of fit metric (R 2 ) to quantify the degree of matching to recorded data, thereby giving a measure of model performance.Standard goodness-of-fit statistics as R 2 are an appropriate measure for model evaluation  (Guiot and Vernal, 2007).A randomized t test was used to test the significance level of the WAPLS components.Subsequently, the significant WAPLS components were used for the transfer-functions.Error bars indicate the standard error of bootstrap climate parameter estimates.
A principle component analysis (PCA) was performed on the fossil pollen data set of WW26 to derive main trends in vegetation change over time.The data set was square root transformed prior to the ordination to reduce the effects of over-represented taxa in the pollen record.

MAP, MAT and Tmin along the WeruWeru transect
Figure 2 shows the trend of MAP, MAT and Tmin along the elevational gradient between 1900 and 3200 m a.s.l. of the southern slope of Mt.Kilimanjaro.Starting off with 2750 mm (at 1900 m a.s.l.) the MAP peaks at 2300 m elevation with about 3450 mm rainfall.It then gradually decreases with increasing elevation to 1600 mm at 3200 m a.s.l.MAT (Fig. 2b) shows a linear decrease in temperature uphill.While the MAT reaches 14.2 • C at 1900 m a.s.l., in drops to 7.06 • C at 3200 m a.s.l.Tmin of the recorded time period shows a linearly decreasing trend with increasing elevation (Fig. 2c).At 1900 m a.s.l., the lowest temperature recorded was 6.01 • C.
A Tmin of −1.2 • C was recorded at 3200 m a.s.l.The frost line is crossed at about 2840 m a.s.l. as suggested by interpolation of the local regression function.However, temperatures below zero were also recorded at 2700 m a.s.l.

Pollen taxa response to MAP, MAT and Tmin gradients
In Araliaceae pollen grains are most abundant at site with 2090 mm and at 2220 mm MAP (23 % of the recorded pollen sum and 18 % respectively) (Fig. 3a).Lowest abundance values are recorded at sites with annual rainfall of 1610 to 1830 mm and with MAP > 2900 mm.Despite the strong variation of abundance, Araliaceae pollen shows a unimodal response to MAP with an optimum abundance at sites with 1800 to 2400 mm of rainfall.Olea pollen has the highest abundance at 1930 mm MAP.At lower and higher precipitation, the abundance decreases gradually.Macaranga pollen grains show an abundance of about 5 to 15 % at every site (MAP range: 1600 to 3450 mm) except for three: at locations with 3670, 2920 and 2750 mm MAP this pollen taxon is very abundant (70 to 75 %).Podocarpus pollen grains show a unimodal response to MAP with highest abundance of 43 % at 2210 mm of rainfall.The abundance decreases gradually with higher and lower MAP.
The MAT plot (Fig. 3b) shows that Araliaceae pollen grains are abundantly recorded (18 to 24 %) at sites ranging from 8.5 to 11.0 • C MAT.At sites with higher or lower MAT the abundance of Araliaceae pollen is < 2.5 %.Olea pollen grains are most abundant

Multivariate analyses of the modern pollen-rain and subset selection
The Detrended Correspondence Analysis (DCA) of the untransformed modern pollenrain data revealed a length of environmental gradient of 2.49 inherit in the data.This result and the predominantly unimodal response of single pollen taxa suggested to The CCA triplots (Fig. 4a and b) represent simultaneously the ordination of samples (sites labelled with elevation in m a.s.l.), pollen taxa (30 % best fit) and their relationship to climate variables.In the ordination diagrams, the climate variables (MAP, MAT, Tmin) are represented by vectors, the point of direction indicates increasing values of the respective parameter.
The proximity of pollen taxa and sites to any climate parameter can be interpreted as a relative measure for the correlation between those data point.Sites that a closer positioned to climate variable vector respectively, experience higher temperature or precipitation.Taxa that show a positive correlation with either of the climate parameters are most abundant at sites with high values of this parameter.
Hagenia and Erica pollen grains are negatively correlated with all three climate parameters whereas Macaranga shows a positive correlation.Ilex is more strongly correlated with MAP than with MAT, the opposite is the case for Podocarpus pollen grains, which are slightly more correlated with MAT.Taxa that are positioned close to the centre of the ordination diagram do either not show a clear linear relationship with any climate variable, show a unimodal response and/or do not significantly contribute to the variance of the whole data set (hence, the dispersion of taxa within the ordination space).
The summary of the CCA (Table 1) shows that the first two CCA axes with MAP as single constraining variable accounted for 33.46 % of the cumulative variance in the pollen data.For the MAT the first two CCA axes account for 34.52 %, for Tmin this is 34.92 %.Pseudo-canonical correlation between pollen taxa and the climate variable is 0.943 for MAP, 0.9616 for MAT and 0.9665 for Tmin.The Monte Carlo permutation test results reveal high significance levels for all three climate parameters (p = 0.002).
The forward selection revealed 24 pollen taxa with an explanatory contribution ≥ 20 % for MAP, 25 taxa for MAT and 15 taxa for Tmin (Table 2).

Model performance and climate reconstruction of the WW26 pollen record using Weighted-Averaging Least Square Regression (WA-PLS)
The model performance is summarized by several measures (Table 3).Due to sediment loss (hiatus), the palaeosoil sequence between 56 and 66 cm the Last Glacial Maximum (LGM) is missing.According to the AMS chronology the sequence above the hiatus (< 56 cm depth) is made up of mid to late Holocene sediment (ca.6 to 0 kyr BP) and the sediment below the hiatus (> 66 cm depth) belongs to the mid Last Glacial (ca.46 to 36 kyr BP).
Earlier than 44 kyr BP our model reconstructs a MAP is between 1530 and 1620 mm.It then reaches a temporary high around 42 kyr BP (around 90 cm depth, above 1680 mm) and then drops back to 1550 mm soon after (at 85 cm depth).The remaining time prior to 36 kyr BP MAP fluctuates between 1670 and 1550 mm.In the mid Holocene, just before 6 kyr BP, our model reconstructs similar MAP values as for the late Last Glacial (ca.1650 mm).Between 6.1 and 2.6 kyr BP MAP increases rapidly up to 1940 mm.Maximum MAP (1950 mm) is reached in the late Holocene at about 1.9 kyr BP (29 cm depth).2030 mm are reconstructed for the modern pollen rain sample at the study site at 2600 m a.s.l., which very close to the observed precipitation from the site (2077 mm).The lowest MAP of 1530 mm is reconstructed for about 45 kyr BP (at 117 cm depth).
The reconstructed MAT ranges between 7.9 and 8.2 • C in the lowest part of the record between 95 and 123 cm depth.It shows an increase to 8.6 • C around 42 kyr BP (90 cm depth).It then drops back to 8.1 to 8.4 • C until 36 kyr BP.The lowest MAT (7.9 • C) is recorded at about 45 kyr BP (113 cm depth).Temperature around 8.7 • C are recorded for the mid Holocene.After that, MAT increases rapidly to 8.9 • C at about 3 kyr BP (45 cm).Throughout the late Holocene, the temperatures fluctuate between 8.3 and 9.2 • C. The MAT reconstructed for the study sites is 9.6 • C, which is close to today's temperature from the site of 10.1 • C.
The reconstructed Tmin values vary between −1.1 and −0.4 • C throughout the pre- LGM.At no point Tmin exceeds the frostline.In the mid and late Holocene fluctuates strongly between −1.5 • C and + 0.5 • C The lowest Tmin is recorded in the late Holocene at 2.6 kyr BP (−1.2 • C) and at about 1 ka (−1.5 • C).In the mid and late Holocene the Tmin values frequently exceed the frostline.The Tmin reconstructed for the study site today is 1.1 • C which is very close to the Tmin of 1.3 • C observed at 2600 m a.s.l.today.
The PCA based on the WeruWeru pollen record revealed a first principle component (PC 1) reflecting 46.6 % of the variation inherent in pollen record and a second principle component reflecting 25.4 %.PC 1 is plotted next to the quantitative climate reconstruction (Fig. 5) to facilitate the detection and interpretation of past forest community and climate patterns.While during the pre-LGM the pollen taxa composition of the site seems to have been rather stable, the Holocene is characterized by major taxa composition changes and fast turnover rates indicated by a rapid change in standard deviation units of the first principle component.

Interpretation and discussion
The results of the climate record along the WeruWeru transect on the southern slope of Mt.Kilimanjaro between 1900 and 3200 m elevation have shown that the MAP and MAT follow the typical climate gradient observed from mountains located in dry tropical region (Lauer, 1976).While MAT and Tmin show a linear trend along the elevational gradient, the MAP shows a normal distribution with a maximum in mid-altitudes at Kilimanjaro (Hemp, 2006a).Since the investigated elevational gradient starts at an elevation of 1900 m a.s.l., this unimodal trend is not very strong reflected in our results.

Pollen taxa responses to climate
The response of the pollen taxa to the climate variables differs between taxa.Despite the fact that the majority of pollen taxa seems to follow a normal distribution along the gradient of climate variables, there are pollen taxa that show a rather directional trend such as Macaranga vs. MAT and Tmin.However, it is known that species abundance is usually a unimodal function of the environmental variable (Odum, 1971;Whittaker et al., 1973).Despite the relative abundance of taxa within the pollen record, the ecological unimodal response to climate parameters can still be observed.In the case of Macaranga and partly Olea, the full elevational abundance gradient, hence ecological optimum curve, was not covered by our modern pollen-rain study.The results of the modern pollen-rain suggest that the abundance of Macaranga as well as Olea extents to lower elevations below 1900 m a.s.l.accompanied by higher temperatures and lower precipitations (Hemp, personel observation, 2006).Despite potential differences in pollen dispersal patterns between species (as suggested in Schüler, 2013), this unimodal distribution would also hold true for the pollen record.Hence, it can be assumed that directional trends in the taxa abundance along the elevation gradient recorded in this study are rather caused by the incomplete inventory of the climate gradient.This is supported by the fact that the precipitation gradient is not fully captured as can be seen from the recorded MAP (Fig. 2) and when compared to Hemp (2006a).The evaluation of unpublished data of A. Hemp (600 forest relevées from Kilimanjaro) reveals that all these taxa have a unimodal distribution regarding abundance and biomass along the climatic gradient.Studies on pollen representivity have shown that despite pollen production and dispersal patterns the underlying vegetation composition can still be identified (Schüler, 2013).
Taxa such as Podocarpus and Araliaceae seem to have a rather narrow MAT optimum area whereas Olea and Macaranga are rather tolerant and display a broader temperature optimum.This unimodal response of taxa was further supported by the long environmental gradient revealed by the DCA.The CCA revealed that all the climate variables are important environmental factors influencing the distribution of the pollen taxa.In all ordinations, the explained variance in the pollen data set is high.The correlation between pollen taxa and the climate variables indicates a strong relationship between species and environment, which is supported by the permutation test results.Based on these outcomes, the quantitative inference models for these variables from the pollen assemblages were developed.

Model performance
The assessment of the pollen training subset selection and the WA-PLS model performance shows that the type of data transformation used plays an important role for the reconstruction.Skewed data and data with many rare or absent taxa are often transformed by taking the square-root (Legendre and Legendre, 2012).Since in our study not only the most abundant pollen taxa (e.g.Ericaceae and Macaranga) show a good correlation with MAP, MAT and Tmin but also taxa that have a low occurrence (e.g.Olea), square-root transformations amplify the weight of these taxa.The normalizing effect of the square-root transformation allows the selection of rare taxa into the subset and relativizes the effect of dominant taxa on the transfer function.Our results show that this is very important to consider when climate parameters are to be reconstructed from pollen taxa compositions instead of single taxa only.Data transformation facilitates rare taxa or the presents and absents of taxa, which can still act as discriminant 211 taxa in the plant ecological context, to act as delimiters or determinants in the climate parameter reconstruction.Hence, the information inherent in rare taxa must not be neglected.Instead, these taxa should receive special attention and need to be reinforced prior to model application.This need is further supported by findings from the montane vegetation of Mt.Kilimanjaro, which demonstrate, that several pollen taxa are over or underrepresented due to differences in plant pollen productivity and possibly wind-caused pollen drift (Schüler, 2013).

Pre-LGM climate reconstruction
The WA-PLS reconstructed climate parameters based on the WeruWeru pollen record give a very nice estimation of the climatic conditions in the Kilimanjaro area in the pre-LGM and the Holocene.Due to the hiatus in the pollen record, the climate reconstruction is not possible for the LGM.
During late pre-LGM (ca.44 to 36 kyr BP) MAP is reconstructed to be between 1540 and 1680 mm.This equals a maximum reduction in MAP by 480 mm, which is 24 % than today.Before this period (> 43 kyr BP), our model suggests a MAP between 1530 and 1620 (20 and 25 % respectively, less than today).This suggests that the early pre-LGM was slightly drier than the late pre-LGM.During the early pre-LGM, the reconstructed MAT of 7.9 to 8.2 • C means that it was on average 1.4 to 1.7 • C cooler than today.After this period until 36 kyr BP, our model suggests marginally warmer conditions with temperature of 8.1 to 8.6 • C. Such cool and humid conditions during the period directly preceding the LGM is also proposed by the pollen and d13C record from Mt. Kenya (Olago et al., 1999).
Reconstructed minimum temperatures (Tmin) persistently remain between −0.4 and −1.1 • C throughout the pre-LGM.The climate fluctuations are lowest in the early pre-LGM (> 44 kyr BP) and largest during the Holocene.Humid but stable climate conditions during the early pre-LGM were also observed at Maundi Crater (2780 m a.s.l.) on Mt.Kilimanjaro (Schüler et al., 2012).The relatively warm climatic conditions similar to today are supported by vegetation reconstructions from the Rukiga highlands in southwest Uganda (Taylor, 1990).In addition, contemporaneous records from Sacred Lake, Mt.Kenya (Coetzee, 1967;Olago, 2001) and Kamiranzovu Swamp (Hamilton, 1982) are in accordance with these results.The increase in shrubs and trees indicate a shift of vegetation zones to higher elevations suggest humid and comparatively warm climate conditions at Maundi Crater (Schüler et al., 2012).Based on Mt.Kilimanjaro temperature lapse rates varying from 0.51 to 0.56 • C per 100 m rise (Hemp, 2006b) the reconstructed temperatures at our study site at 2600 m of 7.9 to 8.2 • C translate into a vegetation shift of about 300 m during the late pre-LGM.This is more that Mumbi et al. (2008) recontructed, who observe an altitudinal shift of montane forest from 1700 to 1800 m a.s.l. to 1800-1900 m a.s.l. in the Eastern Arc Mountains of Tanzania.However, this difference could be caused by the shadowing effect of the surrounding mountains in the Eastern Arc, which leads to less extreme climate conditions within in the mountains range compared to freestanding mountains as the Kilimanjaro.Hence, our MAT reconstruction reflects the decreasing temperature trend approaching the LGM.Lake Chala temperature reconstruction based on distributions of branched GDGTs infer LGM temperatures 7-10 • C lower than today (Sinninghe Damsté et al., 2012) which suggests that a rapid temperature decrease in the Kilimanjaro area occurred after 36 ka.Interesting is, that despite the MAT being slightly higher in the late pre-LGM compared to the early pre- LGM, the lowest Tmin is reconstructed for the late pre-LGM.This suggests that in spite of the relatively moderate temperatures on average, the frequency and severity of low temperature events increased towards the LGM.
The pollen-based temperature and precipitation reconstruction from Kashiru Swamp at 2104 m a.s.l. in the Burundi Highlands, estimate the LGM climate being by 3-4 • C cooler and drier (ca.30 % less rainfall) than today (Bonnefille and Chalié, 2000;Bonnefille et al., 1992).Such LGM temperatures would have caused a descent of vegetation zone on Mt.Kilimanjaro by another 300-400 m after the late pre-LGM.Recent vegetation reconstruction on Mt.Killimanjaro suggest a vegetation shift of about 1000 m for the LGM (Schüler et al., 2012;Zech, 2006) LGM temperature decrease of 5.1 • C suggested for East African Lakes (Loomis et al., 2012).The lowering of temperatures during the LGM in East Africa decreased to 5-6 • C on average, which also matches the reduction of the SST's in the western Indian Ocean between 0 • N and 20 • S and west of 50 • E (Anhuf, 2000).Results from other east African high mountains show that the overall lowering of the vegetation belts during the LGM amounted to ±700 m in the dry high mountains, whereas in the humid high mountains it amounted to ±1000 m (Coetzee, 1967;Hamilton, 1982;Perrott and Street-Perrott, 1982;Maitima, 1991;Olago et al., 2000).
Despite the assumed vegetation zone shifts during the pre-LGM the taxa composition within the prevailing vegetation remained rather stable as proposed by the ordination of the pollen taxa.The variations in MAT, Tmin and MAP had only a minor impact on the taxonomic structure of the montane plant communities during that time.

Holocene climate reconstruction
Precipitation and temperature fluctuate strongly throughout the Holocene.Mid and late Holocene MAT is between 0.4 and 1.4 • C lower than today.MAP is between 4 and 20 % lower than today and the Tmin exceed the frostline just after 6 ka for the first time in the record.This suggests increased unstable climate conditions throughout the Holocene, which is a common phenomenon in the tropics (Mayewski et al., 2004).
Our pollen-based reconstruction proposes very relatively warm and wet mid Holocene around 6 kyr BP).Heavy convective precipitation for around 6 kyr BP also is suggested by δ 18 O diatom from Lake Challa (Barker et al., 2011).The Kilimanjaro ice core records reflect this substantial cooling between 6.5 to 5.2 kyr BP (Thompson et al., 2002) which cannot be derived from our results.Neither does can we observe the dry conditions assumed from pollen record of Maundi Crater, Kilimanjaro (Schüler et al., 2012) and hydrology reconstructions (BIT index) of Lake Challa at the foot of Mt.Kilimanjaro (Verschuren et al., 2009;Moernaut et al., 2010).
A second abrupt dry event picked up in our record around 2.5 kyr BP is also recorded from the Kilimanjaro ice cores as enhanced dust input and lowered δ 18  O ice (Thompson  , 2002).While δ 18 O diatom of Lake Challa also indicates greatest aridity around 2.5 ka, the BIT index suggests increased surface runoff relative to the mid Holocene.Considering the slightly cooler late Holocene climate suggested by the Tanganyika TEX86 record (Tierney et al., 2008), our results support the suggestion by Barker (Barker et al., 2011) that a weakened southeastern monsoon resulted in a reduced long rainy season and protracted main dry season during that period.Several other East African Lakes record major drought in the late Holocene: at Lake Victoria, the diatom-inferred lake level reached a minimum at 2.7-2.4 kyr (Stager and Mayewski, 1997) and Lake Naivasha, south Kenya, stood dry around 3 ka (Richardson and Dussinger, 1986).After the increase in temperature around 6 kyr BP, MAT shows strong variation throughout the remaining Holocene but does not experience any substantial long-term change.
The mid Holocene to present cooling by 4 • C as suggested by temperature reconstruction from Sacred Lake, Kenya (Loomis et al., 2012) cannot be observed in our record.Synchronously with the rise of annual temperature and increased precipitation in the Holocene, Tmin exceeds the frostline for longer periods.Since freezing stress is known to operate on long time scales and has a strong impact on plant communities (Körner, 2003;Holdridge, 1967) this microclimatic change to non-freezing minimum temperature caused profound change in the montane forest community at the WeruWeru site, which is indicated by the large increase in PC 1.The consistent vegetation change continues for at least 1000 yr after the first occurrence of phases without frost in the mid Holocene.This shows that despite more or less stable climate conditions, the only slightly higher Tmin triggered the reorganization of montane forest composition.This trigger effect would also explain the rather stable vegetation composition during the pre-LGM where minimum temperatures predominantly remain around below −0.6 to −1.0 • C.
It seems that despite repeated frost events throughout the Holocene, the generally more favourable, warmer and wetter climate conditions enabled the afro-montane forest to develop and expand on the slopes of Kilimanjaro.However, since our results also suggest, that the annual amount of precipitation is a crucial climate factor for the In prospective studies, the quantitative climate reconstruction can be improved by additional modern pollen rain data, especially from lower elevations with submontane dry forests and colline savanna vegetation.It is unlikely that the temperature and precipitation patterns revealed by our record would be strongly modified since potential Late Glacial climate conditions are well covered by our modern pollen rain data set.However, warmer and especially drier phases would probably be refined since the climate gradient covered here does not span elevations < 1900 m a.s.l.(with higher temperatures and decreasing precipitation).Further, outlier samples, such as samples with unusual pollen assemblages or unusual combination of environmental variables, or a pollen assemblage with poor relationship to MAP and/or MAT can strongly affect the predictive power of transfer functions (Birks et al., 1990;Hall and Smol, 1992) or eventually the respective reconstruction.Hutson (1976) points out that failure of transfer functions due to environmental conditions that have no modern analogue may be cryptic, but that chances for discovery are increased if more than one transfer function algorithm is applied to a given set of samples.Moreover, it needs to be explored in more detail, which taxa are to be included, and how they are weighted in the calibration, and eventually influence the reconstruction.
Hence, there is an urgent need of modern pollen-rain studies in tropical East Africa in order to establish pollen-climate relationships that show a strong and stable correlation.This will strongly improve the model performances and lead to much better predictions of climate parameter variations based on fossil pollen distribution.

Conclusions
Studies of the Quaternary period, especially the period from the LGM onwards, have two great advantages: a wealth of accurately dated information, and the fact that very little macroevolution or natural extinction has taken place on this time scale so that observed biotic transitions can be interpreted based on knowledge of the present-day biota.It is possible to compute pollen transfer functions relating the distribution of fossil pollen to the distribution of the climatic parameter of interest.However, this implies the existence of an adequate set of fossil samples and a representative training data set, hence, meaning good data on the distribution of environmental parameters under modern conditions.This relationship can then be used to estimate the climate values under past conditions.This computation is crucially dependent on a strong and stable correlation between climate parameter variations and fossil distribution.If these constraints are satisfied, the transfer functions derived through WA-PLS yield comparable results.The response of the pollen taxa to the climate variables differs between taxa.In our study, the majority of pollen taxa seems to follow a normal distribution along the gradient of climate variables.Directional trends in the taxa abundance along the elevation gradient recorded in this study are probably caused by the incomplete inventory of the climate gradient.The assessment of the WA-PLS model performance shows that the type of data transformation used plays an important role for the reconstruction.Our best-fit transfer function is the one based on square root -transformed modern pollen-rain data, which is subset individually for each reconstructed climate variable by forward selection.We can show that it is important to include the information inherent in rare taxa.These taxa should receive special attention and need to be reinforced prior to model application.The reconstructed climate patterns for our fossil pollen record from the afro-montane forest on Mt.Kilimanjaro are consistent with the interpretation of pollen data in terms of vegetation history of montane forest and with general climate trends during the past 45 kyr BP in tropical East Africa.The pre-LGM is characterized by a climate about 1.7 • C cooler and by about 25 % drier than today.
Towards the LGM, the annual minimum temperatures decreased and showed more fluctuations despite mostly unchanging mean annual temperature and only slightly increased annual precipitation.Due to a hiatus, the LGM is missing in the record.During the mid and late Holocene, our reconstructions reveal warmer and particularly more

197
Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper |

199
Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | linearly to −7.1 • C 201 Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper |

Discussion
Paper | Discussion Paper | Discussion Paper | Discussion Paper | 203 Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Fig. 3 the response functions of several representative pollen taxa (Araliaceae, Olea, Macaranga and Podocarpus) to MAP, MAT and Tmin are depicted.205 Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | (0.3 to 1.4 %) between 7 and 10.5 • C.Then, with increasing MAT (> 11 • C), Olea decreases and is not recorded at sites with MAT > 12.6 • C. Macaranga pollen grains show an abundance of about 5 to 15 % at every sites with MAT < 12 • C. Between 12 and 14.3 • C this pollen taxon is very abundant (70 to 75 %).Podocarpus pollen grains are not very common at sites with MAT < 9 • C (< 9 %).They are most abundant between 9 and 11 • C and then the record of Podocarpus pollen decreases with increasing MAT.The response function of the pollen taxa to Tmin is very much alike the one observed from MAT.
Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | continue with a Canonical Correspondence Analysis (CCA, a unimodal based constrained ordination).

207
Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper |

209
Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper |

Discussion
Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | , which corresponds well with 213 Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper |

Discussion
Paper | Discussion Paper | Discussion Paper | Discussion Paper | et al.

215
Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | existence of montane forests, even a minor reduction in precipitation over the past decades as observed by Hemp (2005) could already have a major effect on the vegetation.
Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper |

217
Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | humid climate conditions.For the first time in this record, the annual minimum temperature exceeds 0 • C, which triggered profound changes in the vegetation and lead to the development and expansion of afro-montane forest on Mt.Kilimanjaro as it is found at 2600 m a.s.l.today.Supplementary material related to this article is available online at http://www.clim-past-discuss.net/10/195/2014/cpd-10-195-2014-supplement.pdf.