Review of Palaeobotany and Palynology 140 (2006) 61 – 77
www.elsevier.com/locate/revpalbo
Quantitative relationships between modern pollen rain
and climate in the Tibetan Plateau
Caiming Shen a,b,⁎, Kam-biu Liu a , Lingyu Tang b , Jonathan T. Overpeck c
b
a
Department of Geography and Anthropology, Louisiana State University, Baton Rouge, LA 70803, USA
Nanjing Institute of Geology and Paleontology, Academia Sinica, 39 East Beijing Road, Nanjing, 210008, People's Republic of China
c
Institute for the Study of Planet Earth and Department of Geosciences, University of Arizona, Tucson, AZ 88721, USA
Received 22 July 2004; received in revised form 2 March 2006; accepted 2 March 2006
Available online 19 April 2006
Abstract
Quantitative relationships between modern pollen rain and climate are poorly studied in China, partly due to the extensive
human impact on the modern vegetation. A dataset consisting of 227 modern pollen samples from forests, shrublands, meadows,
steppes, and deserts in the Tibetan Plateau, the least anthropologically-disturbed region in China, provides a unique opportunity to
study the quantitative relationships between modern pollen rain and climate. Pollen percentage data were calculated on a sum of 20
pollen taxa. Climatic data for each site, including mean annual precipitation (MAP), mean annual temperature (MAT), July
temperature (Tjuly), and January temperature (Tjan), were derived from 214 meteorological stations in the Tibetan Plateau and
adjacent areas using natural neighbor interpolation and linear interpolation methods.
Canonical correspondence analysis (CCA) was used to reveal the climatic parameters that best reflect the main patterns of
variation in the modern pollen rain, and to detect anomalous observations. Results of CCA indicate that MAP and Tjuly are two
climatic parameters controlling the variation of modern pollen rain in the Tibetan Plateau. Pollen–climate transfer functions for
MAP and Tjuly were then developed using the inverse linear regression and weighted-averaging partial least squares regression
models. The functions derived from these two models are statistically significant at the 0.0000 level. A case study, in which these
functions were applied to a fossil pollen record from an alpine lake in the eastern Tibetan Plateau, was conducted to show the
feasibility of these functions in paleoclimate reconstruction. The results demonstrated the applicability of these pollen–climate
transfer functions to fossil pollen data.
© 2006 Elsevier B.V. All rights reserved.
Keywords: modern pollen data; numerical analysis; paleoclimatic reconstruction; transfer function; palynology; Tibetan Plateau
1. Introduction
⁎ Corresponding author. Current address: Atmospheric Sciences
Research Center, State University of New York, Albany, NY 12203,
USA. Tel.: +1 518 437 8644; fax: +1 518 372 8325.
E-mail addresses: [email protected] (C. Shen),
[email protected] (K. Liu).
0034-6667/$ - see front matter © 2006 Elsevier B.V. All rights reserved.
doi:10.1016/j.revpalbo.2006.03.001
Climate modeling provides a key to predicting future
climates and to understanding the mechanisms of past
climate changes. The use of climate models requires that
the models be thoroughly tested in order to build confidence in their results and to identify the areas for
improvement (Webb et al., 1998). Paleoclimate data,
especially quantitative data, are vital for checking the
62
C. Shen et al. / Review of Palaeobotany and Palynology 140 (2006) 61–77
model results. Pollen data can be calibrated in climate
terms, and quantitatively reconstructed paleoclimate
data are thus needed to test model results (Huntley and
Prentice, 1988, 1993; Webb et al., 1993a,b, 1998). In
this procedure, understanding the quantitative relationships between modern pollen rain and contemporary
climate is vital for estimating paleoclimatic conditions
based on fossil pollen data (Webb and Bryson, 1972;
Webb and Clark, 1977; Overpeck et al., 1985; Birks and
Gordon, 1985; Bartlein et al., 1986; Guiot, 1987; Birks,
1995). Quantitative relationships between modern pollen rain and climate are poorly studied in China (Shen
and Tang, 1992; Song et al., 1997; Wang et al., 1997),
partly due to the extensive human impact on the modern
vegetation (Liu, 1988; Liu and Qiu, 1994). Modern
pollen samples from the Tibetan Plateau, the least anthropologically-disturbed region in China, provide a
unique opportunity to study the quantitative relationships between modern pollen rain and climate.
The proven and well-established methods involved
in the quantitative reconstructions of climate include
transfer functions (Webb and Bryson, 1972; Birks,
1995), response surfaces (Bartlein et al., 1986; Webb et
al., 1993a,b; Markgraf et al., 2002), and the best modern
analogues (Overpeck et al., 1985; Guiot, 1987, 1990).
Among these approaches, response surfaces and the best
modern analogues require very large and comprehensive
training sets (generally more than 500 samples) from a
wide environmental range to provide reliable reconstructions (Birks, 1995, 2003). Having considered the
size of our modern pollen data set, the transfer function
approach was adapted in our study. In this paper, we first
used canonical correspondence analysis (CCA) to
analyze pollen data and climate data to identify the
climate variables that typify the climatic gradients
among modern pollen sampling sites and determine
the modern pollen–climate relationships. Next, transfer
functions were built by using the inverse linear (IL)
regression and weighted averaging partial least squares
(WA-PLS) regression methods. Lastly, a case study was
conducted to quantitatively reconstruct the past climate
based on a fossil pollen record from a small alpine lake
in the eastern Tibetan Plateau to illustrate the applicability of transfer functions to paleoclimatic
reconstruction.
2. Data and methods
2.1. Pollen and climate data
The pollen dataset used here consists of 227 modern
pollen spectra (Fig. 1). (For a complete listing of the
Fig. 1. Map of the Tibetan Plateau showing regional vegetation, location of surface samples and Yidun Lake.
C. Shen et al. / Review of Palaeobotany and Palynology 140 (2006) 61–77
Fig. 2. Summary pollen percentage diagram for 227 surface samples, and sample groups classified by cluster analysis. Pollen percentages are calculated based on a sum of 20 major pollen
taxa.
63
64
C. Shen et al. / Review of Palaeobotany and Palynology 140 (2006) 61–77
modern pollen dataset, see Shen, 2003) For data
standardization and consistency, 20 pollen types were
selected to represent the major and minor pollen types in
the modern pollen spectra. New pollen percentages were
recalculated based on a sum of these 20 pollen types
(Fig. 2). Among these 20 pollen types, major tree pollen
types are Abies, Picea, Pinus, Quercus, and Betula;
shrub pollen types are Rhododendron, Rosaceae, and
Salix; and herb pollen types are Gramineae, Compositae, Artemisia, Chenopodiaceae, and Cyperaceae. Minor
pollen types include tree pollen Tsuga and Corylus, and
herb taxa Ranunculaceae, Thalictrum, Caryophyllaceae,
Polygonum, and Leguminosae. Cupressaceae pollen was
excluded from the major pollen types because it is poorly
preserved in fossil pollen spectra (Shen, 2003).
Climatic data for the sampling sites were derived
from 214 meteorological stations in the Tibetan Plateau
and its adjacent areas (Fig. 3). The stations are unevenly
distributed, with more stations in the eastern part than
the western part. Most of the climate data are derived by
averaging values over a 30–40 year period, except for
some stations in remote areas or high mountains where
the lengths of the records are only several years. The
linear interpolation method was used to map the spatial
patterns of temperature parameters including mean January temperature (Tjan), mean July temperature (Tjuly),
and mean annual temperature (MAT). To take into account the regional differences in data network density
and topographic variation, the Plateau was divided into
three parts — southern Plateau (28–32°N), central Plateau (32–37°N), and northern Plateau (37–42°N). The
regression functions derived from each part of the Plateau were used to calibrate the climatic parameters of
temperature. Due to great regional variations in mean
annual precipitation (MAP) and the topographic effects
on rainfall, natural neighbor interpolation was used to
calculate the mean annual precipitation (Fig. 3). This is a
better interpolation method than others for scatter point
data (Sibson, 1981).
2.2. Methods
CCA is the constrained or canonical version of correspondence analysis (ter Braak, 1986, 1987). This
technique performs a constrained ordination of pollen
data in response to two or more climatic variables. The
CCA ordination axes are linear combinations of the
climatic variables that maximize the dispersion of the
pollen taxon scores. Unlike other multidimensional
scaling techniques, the CCA ordination diagram simultaneously displays the main patterns of variation in the
pollen data as well as the major patterns in the weighted
averages of these pollen taxa in relation to the climatic
variables (Birks, 1995). CCA was used here to reveal the
climatic parameters that best reflect the main patterns
of variation in the modern pollen rain. These climatic
parameters were then used in the transfer functions.
CANOCO 4.0 was used to implement CCA.
Fig. 3. Isohyet (mm) map for the Tibetan Plateau derived from natural neighbor interpolation of observations at 214 stations (topographic image
source: www.esri.com).
C. Shen et al. / Review of Palaeobotany and Palynology 140 (2006) 61–77
Table 1
Canonical correspondence analysis: summary statistics of pollen
variables and climatic variables, eigenvalues, pollen variable scores,
biplot scores and inflation factors of climatic variables
Pollen variable
Mean Standard
deviation
Axis 1 Axis 2
Abies
Picea
Pinus
Tsuga
Quercus
Betula
Corylus
Salix
Rosaceae
Rhododendron
Gramineae
Compositae
Artemisia
Chenopodiaceae
Ranunculaceae
Thalictrum
Caryophyllaceae
Polygonum
Leguminosae
Cyperaceae
1.98
3.46
6.86
0.09
6.46
4.67
0.43
1.96
2.84
3.41
8.70
4.02
12.43
7.73
1.94
1.71
1.59
1.25
1.43
27.03
− 1.06
− 0.78
− 0.60
− 1.12
− 0.51
− 0.58
− 0.52
− 0.37
− 0.15
− 0.35
0.36
0.24
0.17
1.06
0.05
0.16
−0.04
0.24
0.11
0.37
Variable scores
4.10
9.77
11.98
0.28
12.16
8.01
1.55
5.98
3.95
11.88
12.96
6.65
16.32
20.02
2.74
3.18
3.85
3.65
2.22
27.44
Environmental
variable
Tjuly
10.51
2.25
− 9.69
4.19
Tjan
MAT
0.95
3.03
MAP
517.54 167.90
Eigenvalue
Percentage variance
of species data
Cumulative
percentage variance
of species data
of species–
environmental
relation
Species–
environmental
correlations
Inflation
factor
7.33
26.11
41.28
2.09
0.06
0.12
0.16
−0.13
0.36
0.17
0.21
−0.04
−0.09
−0.36
−0.12
−0.07
0.16
0.80
−0.09
−0.30
−0.28
−0.21
0.12
−0.40
Biplot scores
− 0.33
− 0.75
− 0.63
− 0.98
0.24
18.70
0.93
0.48
0.66
0.02
0.09
7.21
18.70
65.85
25.91
91.25
0.91
0.74
Birks (1995) provided a detailed description of transfer function methodology and a critical discussion of
general theory, assumptions, and techniques used for
developing transfer functions. Modern pollen–climate
transfer functions in this study were developed using IL
regression (Webb and Clark, 1977; Andrews et al.,
1980; Heusser and Streeter, 1980; Bernabo, 1981; Howe
and Webb, 1983; Swain et al., 1983) and WA-PLS
regression (ter Braak and Juggins, 1993; ter Braak et al.,
1993; Birks, 1995; Seppa and Birks, 2001). IL regression involves fitting a regression equation that expresses
65
the values of a particular climate variable as a function
of the abundances of several pollen types assuming a
basically linear response model for pollen types and
their environment. WA-PLS regression is a unimodalbased technique, which has shown better performances
than other techniques (Birks, 1995, 1998; Seppa and
Bennett, 2003; Seppa et al., 2004). These two techniques were chosen not only because they are the most
commonly used linear-based and unimodal-based methods, but also for a comparison to evaluate the reliability
of derived transfer functions. Calibration procedures
of Howe and Webb (1983) for IL regression, and of
Birks (1995) for WA-PLS regression were used in this
study. Statistical procedures and computations were
made using Statistica 4.5, SPSS 10.0, and WA-PLS
program.
3. Results
3.1. Modern pollen–climate relationships
First, the dataset was analyzed several times using
CCA for automatically detecting anomalous observations or extreme values through the CANOCO software
(Jongman et al., 1987). The possible causes for these
extreme values were then examined. These extreme
values are generally caused by pollen over-representation of some locally-occurring plants or by the occurrence of azonal soil or landscape (e.g. samples from a
hot valley desert within a forest region). After deleting
some samples with anomalously extreme values, a dataset of 200 samples was finally obtained for CCA and
transfer functions.
The results of CCA are shown in Tables 1–2 and
Fig. 4. The first two axes explain 18.7% and 7.2% of the
total variation in the pollen dataset. Cumulative percentage variance of species–environment relation expresses the amount of variations explained by the axes
as a fraction of the total explainable variations, and the
two axes taken together display 91.25% of variations
Table 2
Canonical correspondence analysis: canonical coefficients, interset
correlations, and intraset correlations
Environmental Canonical
variable
coefficients
Interset
correlations
Intraset
correlations
Axis 1 Axis 2 Axis 1 Axis 2 Axis 1 Axis 2
Tjuly
Tjan
MAT
MAP
0.22
− 0.43
0.02
− 0.78
1.25
−0.56
0.17
−0.10
− 0.30
− 0.69
− 0.58
− 0.90
0.70
0.36
0.49
0.01
−0.33
−0.75
−0.63
−0.98
0.93
0.48
0.66
0.02
66
C. Shen et al. / Review of Palaeobotany and Palynology 140 (2006) 61–77
s
Fig. 4. CCA of the surface samples: ordination diagram with climatic variables represented by arrows and pollen taxa by dark dots.
Fig. 5. Scatter diagrams of mean annual precipitation versus pollen percentages of trees in the Tibetan Plateau.
C. Shen et al. / Review of Palaeobotany and Palynology 140 (2006) 61–77
Fig. 6. Scatter diagrams of mean annual precipitation versus pollen percentages of herbs in the Tibetan Plateau.
Fig. 7. Scatter diagrams of July temperature versus pollen percentages of trees in the Tibetan Plateau.
67
68
C. Shen et al. / Review of Palaeobotany and Palynology 140 (2006) 61–77
that can be explained by the variables. The first eigenvalue is fairly high, implying that the first axis represents
a fairly strong gradient. CCA axis 1 accounts for 65.85%
of the species–environment relation. The species–environment correlations indicate how much of the variation
in the pollen data on one CCA axis is explained by the
environmental variables. The large figure of 0.91 suggests that the climatic variables can account for most of
the variation in the pollen data on CCA axis 1. In Fig. 4,
the lengths and positions of the arrows depend on the
eigenvalues and on the intraset correlations (Table 2).
They provide information about the relationship between
the climatic variables and the derived axes (Jongman
et al., 1987). Arrows that are parallel to an axis indicate a correlation. The length of the arrow reflects the
strength of that correlation. Climatic variables with
long arrows are more strongly correlated with the ordination axes than those with short arrows. Thus, MAP is
most strongly related to axis 1 and least to axis 2,
whereas Tjuly is inversely related to MAP. Tjan and MAT
are not highly related to either axis 1 or axis 2. On the
other hand, inflation factors also indicate that Tjan and
MAT are not as important as MAP and Tjuly. A large
inflation factor implies that the variable is redundant with
other variables in the dataset. It is not surprising since
MAT is highly correlated with Tjuly and all temperatures
have a similar trend in variation along gradients of
altitude and latitude. Moreover, Tjuly represents the temperature of the growing season. It is evident that axis 1
and axis 2 represent the gradients of MAP and Tjuly
respectively as indicated by the scores of pollen variables
on these two axes. Pollen types of forest (Abies, Tsuga)
are located on one end and those of desert (Chenopodiaceae) on the other end of axis 1, whereas pollen types
from meadow and steppe lie in the middle. On axis 2,
pollen types from desert (Chenopodiaceae) with high
Tjuly have high scores whereas those from meadow
(Cyperaceae) with relatively low Tjuly have low scores.
Thus, we can conclude from the above that there are two
dominant factors controlling variations of pollen data in
the Tibetan Plateau: MAP and Tjuly.
As shown by the results of CCA, the gradients of
MAP and Tjuly are strongly related to the variations of
pollen from different vegetation types. The strong relationships between MAP, Tjuly and the selected pollen
types are also evident on the scatter diagrams (Figs. 5–
8). The scatter diagrams reveal pronounced linearities
for some pollen types and non-linearities for others in
the relationships. The linear relationships significant at
0.05 levels generally exist between MAP and arboreal
pollen types as well as some non-arboreal pollen types
such as Chenopodiaceae and Gramineae. Strong
Fig. 8. Scatter diagrams of July temperature versus pollen percentages of herbs in the Tibetan Plateau.
C. Shen et al. / Review of Palaeobotany and Palynology 140 (2006) 61–77
positive correlations are found between Tjuly and tree
pollen types such as Pinus, Quercus, Betula, and weak
positive relationships between Tjuly, Chenopodiaceae
and Artemisia. Markedly negative correlations occur
between Tjuly and Cyperaceae.
3.2. Transfer functions
The forward stepwise method was used in the IL
regression analysis. The performance of the IL regression is reported in Table 3 and Fig. 9. The MAP equa-
69
tion, which accounts for 90% of the total variance and
has a standard error of 55 mm, includes 14 terms. An F
test shows that the equation is significant at the 0.0000
level. Most of the correlation coefficients are significant
at the 0.000 level as indicated by the t test. The positive
coefficients for arboreal pollen types and negative coefficients for most of the non-arboreal pollen types
reflect the relationships shown in the scatter diagrams
(Figs. 5–8). The Tjuly equation, significant at the 0.0000
level, accounts for 72% of the total variance and has a
standard error of 1.3 °C. This equation has 12 terms with
Table 3
Regression summary of dataset for MAP and Tjuly
MAP
n (number of observations): 200
m (number of predictors): 14
r = 0.95
F(14, 185) = 124.52
Intercept
Abies
Chenopodiaceae
Pinus
Gramineae
Picea
Betula
Tsuga
Quercus
Corylus
Salix
Rhododendron
Rosaceae
Leguminosae
Ranunculaceae
r2 = 0.90
p b 0.0000
Adjusted r2 = 0.897
Std. error of estimate: 54.991
B
St. err. of B
t(185)
p-level
426.65
10.54
− 3.09
4.47
− 1.91
3.43
3.35
70.06
1.42
5.94
1.39
2.10
0.47
2.91
− 1.75
9.67
1.23
0.22
0.40
0.35
0.44
0.59
16.65
0.36
2.87
0.67
1.05
0.34
1.94
1.52
44.12
8.60
− 14.09
11.09
− 5.41
7.86
5.73
4.21
3.93
2.07
2.07
1.99
1.37
1.50
− 1.15
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.000
0.040
0.040
0.048
0.171
0.137
0.252
r2 = 0.72
p b 0.0000
Adjusted r2 = 0.70
Std. error of estimate: 1.28
B
St. err. of B
t(187)
p-level
8.27
− 0.01
0.09
0.06
0.06
0.05
0.14
0.04
0.02
− 0.04
− 0.05
0.10
0.02
0.38
0.01
0.01
0.01
0.01
0.01
0.04
0.01
0.01
0.03
0.03
0.07
0.02
Tjuly
n (number of observations): 200
m (number of predictors): 12
r = 0.85
F(12, 187) = 40.05
Intercept
Cyperaceae
Quercus
Chenopodiaceae
Pinus
Artemisia
Leguminosae
Betula
Picea
Caryophyllaceae
Thalictrum
Corylus
Salix
21.58
− 1.67
9.78
9.48
6.82
6.78
3.24
2.68
2.09
− 1.61
− 1.67
1.57
1.32
0
0.096
0.000
0.000
0.000
0.000
0.001
0.008
0.038
0.109
0.096
0.117
0.190
70
C. Shen et al. / Review of Palaeobotany and Palynology 140 (2006) 61–77
Fig. 9. Scatter plot of predicted values versus observed values and residuals versus observed values for inverse linear regression models.
negative coefficients for Cyperaceae, Caryophyllaceae,
and Thalictrum and positive coefficients for arboreal
pollen types and two desert/steppe components, Chenopodiaceae and Artemisia. Plots of the predicted MAP
and Tjuly versus observed MAP and Tjuly, and of the
residuals (predicted minus observed values) versus observed MAP and Tjuly (Fig. 9) show that there are no
strong biases in the MAP and Tjuly models, except for
the tendency to underestimate Tjuly at sites with Tjuly
lower than 8 °C and overestimate it at most sites with
Tjuly higher than 11 °C.
The performance of WA-PLS regression models are
shown in Table 4, where the root mean square error of
prediction (RMSEP), the coefficient of determination
(r2) between observed and predicted values, and the
maximum bias are based on the jack-knifing (Birks,
1995). According to Birks (1998), the models with low
RMSEP, low maximum bias and the smallest number of
‘useful’ components are the ideal candidates. Thus 3and 2-component WA-PLS models were selected respectively for MAP and Tjuly. There is no strong bias in
the MAP. The Tjuly model tends to overestimate at sites
with low temperature (Fig. 10), probably because of the
uphill transportation of pollen from forests to alpine
meadow areas above the treeline. It tends to underestimate at some sites with high temperature, presumably
due to the fact that some forest samples contain relatively
high percentages of herbaceous pollen such as Cyperaceae pollen as shown in Fig. 2.
3.3. A case study
In order to test the usefulness of the transfer functions for paleoclimatic reconstruction in the Tibetan
Plateau, we applied both the IL regression and WA-PLS
regression models to a fossil pollen stratigraphy from
an alpine lake. Yidun Lake (30°17.9′N, 99°33.1′E) is a
small lake situated in the subalpine conifer forest region
Table 4
Performance statistics for WA-PLS regression models
WA-PLS component
RMSEP
r2
Maximum bias
200 samples × 20 taxa
MAP (range = 66–910 mm, mean = 494.2 mm, standard deviation =
171.2 mm)
1
71.307
0.826
142.31
2
63.220
0.863
100.99
3⁎
61.225
0.872
81.47
4
60.706
0.874
78.14
5
60.058
0.877
65.85
6
59.857
0.877
73.61
Tjuly (range = 5.1–16.5 °C, mean = 10.5 °C, standard deviation
= 2.34 °C)
1
1.407
0.638
2.505
2*
1.369
0.658
2.186
3
1.359
0.663
2.237
4
1.357
0.665
2.271
5
1.354
0.666
2.269
6
1.364
0.662
2.173
* Selected models.
C. Shen et al. / Review of Palaeobotany and Palynology 140 (2006) 61–77
71
Fig. 10. Scatter plot of predicted values versus observed values and residuals versus observed values for WA-PLS regression models.
in the southeastern Tibetan Plateau (Fig. 1). At 4470 m
above sea level, the lake is situated only 300 m above
the alpine treeline. Vegetation around the lake is alpine
shrub meadow dominated by Cyperaceae and Rhododendron with some Salix. Dense forests consisting of
Picea, Abies, Pinus, Betula, Quercus, and Juniperus
occur at lower elevations down the valley. The Yidun
Lake area has a mean July temperature around 12 °C
and mean precipitation of 620 mm per year.
The core used for the transfer function study was
4.7 m long and retrieved over 2.1 m of water. The
sediment consists of 2.8 m of clayey gyttja with 10–
20% organic matter content, overlying 1.9 m of basal
clay. Dating control was provided by three AMS (accelerator mass spectrometry) radiocarbon dates derived
from the bulk sediment samples from the clayey gyttja
(Fig. 11). A date of 2070 ± 60 14C yr BP from the core
top, corroborated by another AMS date of 2040 ±
50 14C yr BP obtained from a modern sample of
aquatic sedge taken from the lake edge, suggesting that
the radiocarbon-determined ages are too old due to the
hard-water effect. By assuming that the reservoir of old
carbon remains constant through time, all three 14C
dates, including the 5710 ± 70 and 11330 ± 140 14C yr
BP dates obtained from a depth of 142 and 275 cm
respectively, were corrected by subtracting 2070 years.
The age of each pollen sample level was then estimated by extrapolation and interpolation of corrected
dates. The dating problem, which is fairly common in
Tibetan lake sediments (Fontes et al., 1993, 1996),
introduces some degree of uncertainty in the age
model.
The pollen diagram was divided into six zones based
on changes in pollen percentages and concentration
values. Pollen zones YGL6 and 5 (470–325 cm) are
characterized by low pollen concentrations and high
percentages of herbaceous pollen types such as Cyperaceae, Artemisia, Gramineae, and Caryophyllaceae.
Based on the results of discriminant analysis (Shen,
2003), the pollen data indicate that steppe and meadow
prevailed around the site during the late glacial (17.3–
11.5 ka BP). Both percentages of tree pollen and total
pollen concentrations increase in pollen zone YGL4,
suggesting that the vegetation changed to forest from
11.5 to 9.2 ka BP. After 9.2 ka BP, the pollen
assemblages are dominated first by Betula (zone
YGL3), then by Pinus (YGL2), and finally by Quercus
and Pinus (YGL1), reflecting the succession of forest
communities under different climatic conditions during
the Holocene.
The MAP and Tjuly reconstructed from transfer
functions developed by IL regression and WA-PLS
regression models show parallel and consistent trends
between these two techniques (Fig. 12). Fig. 12 also
shows the 95% confidence interval of the estimates,
which seem to be narrower for those samples at the upper
72
C. Shen et al. / Review of Palaeobotany and Palynology 140 (2006) 61–77
Fig. 11. Pollen diagram from Yidun Lake in southeastern Tibetan Plateau.
C. Shen et al. / Review of Palaeobotany and Palynology 140 (2006) 61–77
73
Fig. 12. Mean annual precipitation and July temperature (solid lines) and their 95% confidence intervals (dotted lines) for Yidun Lake reconstructed
from pollen–climate transfer functions developed by inverse linear regression and weighted-average partial least squares regression models. The
triangles indicate modern values of mean annual precipitation and July temperature.
part of the record, probably suggesting that the vegetation and climate became more modern in character. The
actual modern values of MAP and Tjuly are located within
the 95% confidence intervals of their estimates for the
top sample of the core, thus leading confidence to the
validity of the reconstructions. Due to the long-distance
transport of arboreal pollen, as indicated by low total
pollen concentrations and anomalous composition of
fossil pollen spectra, the MAP and Tjuly for the two basal
samples of Zone YGL6 and the basal sample of Zone
YGL5 are probably overestimated. The longer-term
trends in climate variations reconstructed quantitatively
by the pollen–climate transfer functions (Fig. 12) divide
the Yidun Lake record over the last 17,300 years into the
following intervals.
3.3.1. Late Glacial (17.3–11.5 kaBP; pollen zones YGL
5 and 6)
At the initial stage of lake formation, the vegetation
was dominated by steppe consisting of Artemisia and
Gramineae. The reconstructed MAP and Tjuly imply a
cold and dry climate. The MAP was about 60% of the
modern value of 620 mm, and Tjuly was 4 °C colder than
the present. This period was followed by a short cold
and wet period at 16.7–16.3 ka BP, when vegetation
changed from steppe to meadow dominated by Cyperaceae. Tjuly was 1 °C colder than that in the preceding
period, and MAP increased by about 50 mm. From 16.3
to 13.5 ka BP, the Yidun area was occupied by subalpine
meadow consisting of Cyperaceae, Artemisia, Gramineae, and Caryophyllaceae. However, forests probably
began to appear after 14.5 ka BP at lower elevations
down the valleys. MAP increased gradually to about
500 mm at ca. 13.8 ka, and then decreased gradually.
The reconstructed Tjuly rose steadily and gradually to
N 9.5 °C from 16.3 to 12.6 ka BP, which is followed by a
minor reversal around 12 ka BP.
3.3.2. Late Glacial–Early Holocene (11.5–9.2 kaBP;
pollen zone YGL4)
This interval includes the transition from meadow to
forest. The transition begins with an increase in arboreal
pollen. Birch forests invaded the Yidun area and finally
replaced the meadow at 10 ka BP. Both MAP and Tjuly show
a trend of nonlinear increase during this interval. They
reached the level close to today's at the end of this interval.
74
C. Shen et al. / Review of Palaeobotany and Palynology 140 (2006) 61–77
3.3.3. Early Holocene–Middle Holocene (9.2–6.8 ka
BP; pollen zone YGL3)
Pollen assemblage in this interval is dominated by
Betula and Pinus pollen. Abies, Picea, and Quercus are
trees frequently present. Pollen concentration also
reached the highest values of the whole sequence. Forest, probably a coniferous and deciduous broadleaved
mixed forest, occurred in this area. The climate reconstructions suggest that the Holocene MAP maximum
was achieved from 9.0 to 7.5 ka BP. The maximum
MAP is about 100–120 mm higher than the present.
However, the Tjuly still shows an increasing trend and
fluctuates around the modern value.
centages and gradual drop of Betula percentages.
Birch, the dominant forest tree in the preceding period,
was replaced by pine, and more oak invaded the forest.
MAP fluctuated around 700 mm, slightly lower than
the preceding interval but still higher than the present
except for a major fluctuation at the beginning of the
interval. Reconstructed Tjuly was higher than that of
today, and the Holocene Tjuly maximum also occurs in
this interval, which was 1–1.2 °C higher than the
present.
3.3.5. Late Holocene (2.5 ka BP to present; pollen zone
YGL1)
In this interval, Pinus pollen decreases, and Quercus
pollen replaces Pinus to become the most abundant pollen, indicating an expansion of oak in the forest surrounding the lake. Tjuly steadily fell towards its modern
value. MAP decreased dramatically to below the present
3.3.4. Middle–Late Holocene (6.8–2.5 ka BP; pollen
zone YGL2)
Pollen assemblages are characterized by high percentages of Pinus, continuous rise of Quercus perTable 5
Regression summary of data subset for MAP and Tjuly
MAP
n (number of observations): 80
m (number of predictors): 8
r = 0.94
F(8, 71) = 62.44
r2 = 0.88
p b 0.0000
B
St. err. of B
t(71)
p-level
Intercept
Chenopodiaceae
Gramineae
Artemisia
Cyperaceae
Compositae
Caryophyllaceae
Leguminosae
Thalictrum
700.82
− 6.01
− 4.43
− 3.85
− 2.65
− 3.47
− 2.55
− 3.40
− 1.73
37.13
0.40
0.47
0.49
0.43
0.89
0.89
2.23
1.41
18.88
− 14.98
− 9.40
− 7.83
− 6.23
− 3.91
− 2.86
− 1.53
− 1.22
0.000
0.000
0.000
0.000
0.000
0.000
0.006
0.131
0.225
r2 = 0.81
p b 0.0000
Adjusted r2 = 0.79
Std. error of estimate: 0.9
B
St. err. of B
t(69)
p-level
12.39
− 0.06
0.01
0.24
− 0.01
− 0.07
− 0.04
− 0.17
0.05
− 0.05
− 0.03
1.17
0.01
0.01
0.06
0.01
0.02
0.01
0.07
0.04
0.04
0.02
10.54
− 4.73
0.68
4.34
− 0.57
− 3.08
− 2.95
− 2.56
1.10
− 1.51
− 1.38
0.000
0.000
0.500
0.000
0.568
0.003
0.004
0.013
0.274
0.137
0.173
Adjusted r2 = 0.86
Std. error of estimate: 39 mm
Tjuly
n (number of observations): 80
m (number of predictors): 10
r = 0.90
F(10, 69) = 30.076
Intercept
Cyperaceae
Chenopodiaceae
Leguminosae
Artemisia
Caryophyllaceae
Gramineae
Polygonum
Ranunculaceae
Thalictrum
Compositae
C. Shen et al. / Review of Palaeobotany and Palynology 140 (2006) 61–77
value at first, then increased sharply from 1.6 to 1.3 ka
BP. It then decreased and increased again to its present
level.
4. Discussion and conclusions
Deriving transfer functions is the first step of quantitative interpretation of pollen data in climatic terms
(Bartlein and Whitlock, 1993; Seppa and Bennett,
2003). In this procedure, the most important issue is to
decide whether to use linear- or unimodal-based methods for a particular dataset. It is a general rule of nature
that the quantitative relationships between taxa and environmental variables are non-linear, and the abundance
of a taxon is often a unimodal function of the environmental variables (Gaussian model) (Birks, 1995). However, a unimodal curve will appear monotonic and
approximately linear if a limited range of the environmental variables is covered by samples. The IL regression model requires the assumption that a linear
relationship exists between climatic variables and pollen
type (Howe and Webb III, 1983; Bartlein et al., 1985;
Birks, 1995). To avoid violating this assumption, selection of an adequate geographical region is important
in reducing model specification errors (Bartlein et al.,
1985). However, to define an adequate geographical
region in the Tibetan Plateau as in North America or
Europe is difficult due to its landscape complexity, especially in the southern Plateau. In our study, the scatter
diagrams (Figs. 5–8) show that significant linear relationships exist between MAP, Tjuly, and most of the
pollen types. Thus, the assumption of linearity is satisfied in our data. For further testing, we separated a
subset of data consisting of samples from meadow,
steppe, and desert to develop transfer functions (Table 5).
Comparison of regression results shows no marked
difference between them except the data subset with
smaller standard errors of estimates for both MAP and
Tjuly, and slightly higher r2 value for Tjuly. However, it is
evident that the transfer functions developed by the data
subset could reduce the estimate errors in climatic reconstruction. WA-PLS model is a non-linear unimodal
(Gaussian)-based technique. It is thus not limited by any
geographic region. Comparison of results from linear
models and non-linear models seems to show that WAPLS models have larger estimate errors than linear
models. One explanation for this discrepancy is that
linear models almost certainly underestimate the true
uncertainty (Bartlein and Whitlock, 1993; Birks, 1995).
Ecologically, the derived transfer functions are realistic, as discussed above. Statistically, our analysis also
indicates that these transfer functions are reliable, as
75
assessed by two main lines of evidence. First, CCA
demonstrated that pollen–climate relationships in the
Tibetan Plateau are significant. Second, performance
statistics (r2 and the standard error for IL regression
model; r2 and RMSEP for WA-PLS regression model)
show that these transfer functions significantly predict
observed MAP and Tjuly. r2 is a common measure
of goodness of fit in the transfer function approach
(Bartlein and Whitlock, 1993; Seppa and Bennett,
2003). The values of r2 for our transfer functions
(0.90 and 0.87 for MAP in IL regression model and WAPLS model respectively; 0.72 and 0.66 for Tjuly) are
quite high, and comparable to those of pollen–climate
transfer functions developed in North America and Europe (e.g. Bartlein et al., 1985; Bartlein and Whitlock,
1993; Seppa and Birks, 2001; Seppa et al., 2004). The
standard error is commonly quoted as a measure of the
predictive ability of the training set (Birks, 1995). The
values of standard error for MAP are 6.5% and 7%
(11.2% and 12% for Tjuly) as percentage of the gradient
length of MAP (Tjuly) in IL and WA-PLS regression
models respectively. By comparison with the other
published models for pollen or aquatic organisms based
on similar methods (e.g. Bigler and Hall, 2002; Seppa et
al., 2004; Rull, 2006), they can be considered low. This
shows a superior performance of our derived transfer
functions in prediction power.
There are several means of evaluating how reliable
the paleoclimatic estimates are as derived from transfer
functions. The most powerful evaluation procedure is
to validate the reconstruction against known historical
records or other independent paleoenvironmental records (Birks, 1995). The estimates of MAP and Tjuly
for the top fossil sample are in agreement with the
observed values as mentioned above. It seems to indicate that derived transfer functions can provide reliable paleoclimatic estimates. Comparison with other
independent paleoenvironmental records is difficult for
the Tibetan Plateau because no quantitative climate
reconstruction has been published, although there are
some existing proxy records of temperature and precipitation derived from ice core (e.g. Thompson et al.,
1989) and lake sediments (e.g. Van Campo and Gasse,
1993). A useful but informal evaluation procedure is to
estimate the same climatic parameter by several numerical methods (Birks, 1995). In our case study, the
climates reconstructed by transfer functions show an
excellent consistency in estimate values between two
different techniques, although the standard errors are
larger in WA-PLS models than IL models due to the
reason mentioned above. Additionally, the main trend
of climate reconstructions is also consistent with
76
C. Shen et al. / Review of Palaeobotany and Palynology 140 (2006) 61–77
vegetation succession inferred from pollen data. However, the occurrence of long-distance pollen probably
leads to a bias towards overestimates of MAP and Tjuly
for a few samples such as two basal samples of Zone
YGL6 and the basal sample of Zone YGL5. In summary, our study of quantitative relationships between
modern pollen rain and climate in the Tibetan Plateau
shows that:
(1) MAP and Tjuly are two dominant factors controlling the variations in the modern pollen rain, i.e.
spatial distribution of modern vegetation.
(2) The transfer functions developed in terms of MAP
and Tjuly using IL and WA-PLS regression models
are reliable in reconstructing paleoclimate quantitatively based on the fossil pollen data.
Acknowledgment
This research was supported by grants from the U.S.
National Science Foundation (NSF grants ATM9410491, ATM-0081941), The Chinese National Science Foundation (grants No. 49371068 and 49871078),
and dissertation research grants from the Geological
Society of America (GSA), Association of American
Geographers (AAG), and the Robert C. West Field
Research Award (LSU Department of Geography and
Anthropology). We thank Dr. H.J.B. Birks and Dr.
Stephen Juggins for providing the WA-PLS program.
We also thank Dr. A. Peter Kershaw for comments on
the manuscript.
References
Andrews, J.T., Mode, W.N., Davis, P.T., 1980. Holocene climate based
on pollen transfer functions, eastern Canadian Arctic. Arctic and
Alpine Research 12, 41–64.
Bartlein, P.J., Whitlock, C., 1993. Paleoclimatic interpretation of the
Elk Lake pollen record. Geological Society of America Special
Paper 276, 275–293.
Bartlein, P.J., Prentice, I.C., Webb III, T., 1985. Mean July temperature at
6000 yr BP in eastern North America: regression equations for
estimates from fossil-pollen data. In: Harington, C.R. (Ed.), Climatic
Change in Canada 5: Critical Periods in the Quaternary Climatic
History of Northern North America Syllogeus, vol. 55, pp. 301–342.
Bartlein, P.J., Prentice, I.C., Webb III, T., 1986. Climatic response
surfaces from pollen data for some eastern North American taxa.
Journal of Biogeography 13, 35–57.
Bernabo, J.C., 1981. Quantitative estimates of temperature changes
over the last 2700 years in Michigan based on pollen data.
Quaternary Research 15, 142–159.
Bigler, C., Hall, R., 2002. Diatoms as indicators of climatic and
limnological change in Swedish Lapland: a 100-lake calibration set
and its validation for paleoecological reconstructions. Journal of
Paleolimnology 27, 97–115.
Birks, H.J.B., 1995. Quantitative palaeoenvironmental reconstructions. In: Maddy, D., Brew, J.S. (Eds.), Statistical Modelling of
Quaternary Scicene Data. Technical Guide, vol. 5. Quaternary
Research Association, Cambridge, pp. 161–254.
Birks, H.J.B., 1998. Numerical tools in quantitative paleolimnology—
progress, potentialities, and problems. Journal of Paleolimnology
20, 301–332.
Birks, H.J.B., 2003. Quantitative palaeoenvironmental reconstructions
from Holocene biological data. In: Mackay, A., Battarbee, R.W.,
Birks, H.J.B., Oldfield, F. (Eds.), Global Change in the Holocene.
Arnold, London, pp. 107–123.
Birks, H.J.B., Gordon, A.D., 1985. Numerical Methods in Quaternary
Pollen Analysis. Academic Press, London. 317 pp.
Fontes, J.-C., Melieres, F., Gibert, E., Liu, Q., Gasse, F., 1993. Stable
Isotope and radiocarbon balances of two Tibetan Lakes (Sumxi Co,
Longmu Co) from 13 000 B.P. Quaternary Sciences Reviews 12,
875–887.
Fontes, J.-C., Gasse, F., Gibert, E., 1996. Holocene environmental
changes in Bangong Co Basin (Western Tibet). Part 1: Chronology
and stable isotopes of carbonates of a Holocene lacustrine core.
Palaeogeography, Palaeoclimatology, Palaeoecology 120, 25–47.
Guiot, J., 1987. Late Quaternary climatic changes in France estimated
from multivariate pollen time series. Quaternary Research 28,
100–118.
Guiot, J., 1990. Methodology of the last climatic cycle reconstruction
in France from pollen data. Palaeogeography, Palaeoclimatology,
Palaeoecology 80, 49–69.
Heusser, C.J., Streeter, S.S., 1980. A temperature and precipitation
record of the past 16 000 years in southern Chile. Science 210,
1345–1347.
Howe, S., Webb III, T., 1983. Calibrating pollen data in climatic terms:
improving the methods. Quaternary Science Reviews 2, 17–51.
Huntley, B., Prentice, I.C., 1988. July temperatures in Europe from
pollen data, 6000 years before present. Science 241, 687–690.
Huntley, B., Prentice, I.C., 1993. Holocene vegetation and climate of
Europe. In: Wright Jr., H.R., Kutzbach, J.E., Webb III, T.,
Ruddiman, W.F., Street-Perrott, F.A., Bartlein, P.J. (Eds.), Global
Climates Since the Last Glacial Maximum. University of
Minnesota Press, Minneapolis, pp. 136–168.
Jongman, R.H.G., ter Braak, C.J.F., van Tongeren, O.F.R., 1987. Data
Analysis in Community and Landscape Ecology. Pudoc,
Wageningen.
Liu, K.-b., 1988. Quaternary history of the temperate forests of China.
Quaternary Science Reviews 7, 1–20.
Liu, K.-b., Qiu, H.-l., 1994. Late-Holocene pollen records of
vegetational changes in China: climate or human disturbance.
Terrestrial, Atmospheric and Oceanic Sciences 5, 393–410.
Markgraf, V., Webb, R.S., Anderson, K.H., Anderson, L., 2002.
Modern pollen/climate calibration for southern South America.
Palaeogeography, Palaeoclimatology, Palaeoecology 181,
375–397.
Overpeck, J.T., Webb III, T., Prentice, I.C., 1985. Quantitative
interpretation of fossil pollen spectra: dissimilarity coefficients
and the method of modern analogs. Quaternary Research 23,
87–108.
Rull, V., 2006. A high mountain pollen–altitude calibration set for palaeoclimatic use in the tropical Andes. The Holocene 16, 105–117.
Seppa, H., Bennett, K.D., 2003. Quaternary pollen analysis: recent
progress in palaeoecology and palaeoclimatology. Progress in
Physical Geography 27, 548–579.
Seppa, H., Birks, H.J.B., 2001. July mean temperature and annual
precipitation trends during the Holocene in the Fennoscandian
C. Shen et al. / Review of Palaeobotany and Palynology 140 (2006) 61–77
tree-line area: pollen-based climate reconstructions. The Holocene
11, 527–539.
Seppa, H., Birks, H.J.B., Odland, A., Posta, A., Veski, S., 2004. A
modern pollen–climate calibration set from northern Europe:
developing and testing a tool for palaeoclimatological reconstructions. Journal of Biogeography 31, 251–267.
Shen, C., 2003. Millennial-scale variations and centennial-scale events
in the southwest Asian monsoon: pollen evidence from Tibet. PhD
Dissertation, Louisina State University, Baton Rouge. 286 pp.
Shen, C., Tang, L., 1992. Holocene climate based on pollen transfer
function in Changbaishan mountain and Xiaoxinanling Ranges.
In: Shi, Y., Kong, Z. (Eds.), Climate and Environment during
Holocene Megathermal in China. China Ocean Press, Beijing, pp.
34–39.
Sibson, R., 1981. A brief description of natural neighbour interpolation. In: Barnett, V. (Ed.), Interpreting Multivariate Data. John
Wiley, Chichester, pp. 21–36.
Song, C., Lu, H., Sun, X., 1997. Establishment and application of
pollen–climate transfer functions in North China. Chinese Science
Bulletin 42, 2182–2186.
Swain, A.M., Kutzbach, J.E., Hastenrath, S., 1983. Estimates of
Holocene precipitation for Rajasthan, India based on pollen and
lake-level data. Quaternary Research 19, 1–17.
ter Braak, C.J.F., 1986. Canonical correspondence analysis: a new
eigenvector technique for multivariate direct gradient analysis.
Ecology 67, 1167–1179.
ter Braak, C.J.F., 1987. The analysis of vegetation–environment
relationships by canonical correspondence analysis. Vegetatio 69,
69–77.
ter Braak, C.J.F., Juggins, S., 1993. Weighted averaging partial least
squares regression (WA-PLS): an improved method for reconstructing environmental variables from species assemblages.
Hydrobiologia 269/270, 485–502.
ter Braak, C.J.F., Juggins, S., Birks, H.J.B., van der Voet, H., 1993.
Weighted averaging partial least squares regression (WA-PLS):
definition and comparison with other methods for species–
77
environmental calibration. In: Patil, G.P., Rao, C.R9 (Eds.),
Multivariate Environmental Statistics. Elsevier Science Publishers,
Amsterdam, pp. 525–560.
Thompson, L.G., Mosley-Thompson, E., Davis, M.E., Bolzan, J.F.,
Dai, J., Yao, T., Gundestrup, N., Wu, X., Klein, Z., Xie, Z., 1989.
Holocene–Late Pleistocene Climatic ice core records from
Qinghai–Tibetan Plateau. Science 246, 474–477.
Van Campo, E., Gasse, F., 1993. Pollen- and diatom-inferred climatic
and hydrological changes in Sumxi Co Basin (Western Tibet) since
13,000 yr B.P. Quaternary Research 39, 300–313.
Wang, F., Song, C., Sun, X., Cheng, Q., 1997. Climatic response
surface from pollen data for four arboreal taxa in North China.
Acta Botanica Sinica 39, 272–281.
Webb III, T., Bryson, R.A., 1972. Late- and post-glacial climatic
change in the northern Midwest, USA: quantitative estimates
derived from fossil pollen spectra by multivariate statistical
analysis. Quaternary Research 2, 70–115.
Webb III, T., Clark, D.R., 1977. Calibrating micropaleontological data
in climatic terms: a critical review. Annals of the New York
Academy of Science 288, 93–118.
Webb, R.S., Anderson, K.H., Webb, T., 1993a. Pollen responsesurface estimates of late-Quaternary changes in the moisture
balance of northeastern United States. Quaternary Research 40,
213–227.
Webb III, T., Ruddiman, W.F., Street-Perrott, F.A., Markgraf, V.,
Kutzbach, J.E., Bartlein, P.J., Wright Jr., H.R., Prell, W.L., 1993b.
Climatic changes during the past 18,000 years: regional syntheses,
mechanisms, and causes. In: Wright Jr., H.R., Kutzbach, J.E.,
Webb III, T., Ruddiman, W.F., Street-Perrott, F.A., Bartlein, P.J.
(Eds.), Global Climates Since the Last Glacial Maximum.
University of Minnesota Press, Minneapolis, pp. 514–535.
Webb III, T., Anderson, K.H., Bartlein, P.J., Webb, R.S., 1998. Late
Quaternary climate change in eastern North America: a comparison of pollen-derived estimates with climate model results.
Quaternary Science Reviews 17, 587–606.