Introduction
Pollen analysis was initially developed 100 years ago for inferring
past changes in vegetation and climate (von Post, 1916). Since the 1970s,
quantitative reconstructions from biological proxies have made a revolutionary
change to studies of the past climate (Imbrie and Kipp, 1971; Juggins,
2013). Numerical methods, such as the modern analogue technique (MAT;
Overpeck et al., 1985), weighted averaging partial least squares (WA-PLS;
ter Braak and Juggins, 1993), and others (Birks et al., 2010; Juggins and
Birks, 2012), are widely used in pollen-based quantitative reconstructions
(Guiot, 1990; Markgraf et al., 2002; Seppä et al., 2004; St. Jacques et
al., 2008; Tarasov et al., 2011; Xu et al., 2010b). Pollen-based
palaeoclimate reconstructions relay on modern pollen–climate relationship
studies (Li et al., 2009; Markgraf et al., 2002; Seppä et al., 2004;
Shen et al., 2006) and pollen–climate data compilations (Bartlein et
al., 2010; Prentice et al., 2000; Tarasov et al., 2005; Whitmore et al.,
2005; Yu et al., 1998; Zheng et al., 2014). Thousands of modern pollen
samples from bioclimatic regions all over the world have been collected and
analysed, for example, in the framework of the BIOME6000 Project (Prentice
et al., 2000). These pollen data have been used for testing the biome
reconstruction method and regional sensitivity (e.g. Tarasov et al., 1998;
Yu et al., 1998) and for quantitative climate reconstructions using
statistical approaches. A methodological assumption that the ecological
response of species does not change during the Quaternary based on Lyell's
uniformitarianism (Scott, 1963) is implicit in these studies, and they
require modern organism–environment relationships as calibration models
(Birks et al., 2010; Juggins and Birks, 2012).
There are several types of uncertainties in reconstructing palaeoclimate from
pollen data using calibration models (Guiot et al., 2009; Marquer et al.,
2014; Parnell et al., 2016; Xu et al., 2016b). In China and other regions
with long-term human occupation, biomes can be strongly modified rather than
natural (Ren and Beug, 2002; Zhang et al., 2010). The question of how well
modern samples reflect the natural vegetation thus needs to be addressed
(Xu et al., 2010a). It is most likely that modern pollen–climate
relationships in such regions after a long-lasting human impact are different
from what they were in the past. For example, comparing the performance of
pre-disturbance (1895–1924) and modern (1961–1990) pollen–climate
calibration sets from Minnesota, St. Jacques et al. (2008) found that the
pre-settlement model performs better than the modern one in reconstructing
past climate. The human impact on the terrestrial vegetation over the past
150 years in the American Mid-west is thus apparent in the modern
calibration set. Such a distortion in the modern pollen dataset can generate
bias in the climate reconstruction for those regions (Li et al., 2014; St.
Jacques et al., 2015, 2008; Tian et al., 2017). Palynologists therefore have
to face this challenge in vegetation and climate reconstructions when using
pollen data from densely populated regions (Juggins and Birks, 2012;
Seppä et al., 2004; Tarasov et al., 1999; Xu et al., 2010a).
In China, rich archaeological evidence suggests that crop domestication may
have taken place in the early Holocene or even earlier (Bestel et al., 2014;
Lu et al., 2009; Zhao and Piperno, 2000), and enhanced farming practices have
been
reported since 8000 years ago (Liu et al., 2015; Lu et al., 2009; Zhao,
2011). Early agriculture was usually accompanied by slash-and-burn clearance
of forest patches (Ruddiman, 2003), and pollen-inferred anthropogenic impacts
on natural vegetation are noted from as early as 6000 years ago in eastern
China (Ren and Beug, 2002; Wang et al., 2010). Due to growing demand for
land, construction materials, and fuel, disturbances to the natural vegetation
over the last 2 millennia occurred widely and are commonly detected in the
pollen records (Cao et al., 2010; Ni et al., 2014; Xu et al., 2016a; Zhao et
al., 2010, 2009). Consequently, human impact in both the modern reference
datasets and the fossil pollen records needs to be considered when
reconstructing past climate from pollen. In China, in contrast to North
America, it is not possible to establish a calibration set consisting of
pre-settlement pollen and climate data. However, it is still important to
estimate what kind of bias may appear in pollen-based quantitative climate
reconstructions using Chinese pollen data.
In the past 2 decades, a number of modern pollen studies have been
conducted in China to investigate regional pollen–vegetation–climate
relationships (Herzschuh et al., 2010; Li et al., 2009; Lu et al., 2011; Luo
et al., 2009; Shen et al., 2006; Xu et al., 2007; Zhang et al., 2012; Zheng
et al., 2008) and human impact on vegetation (Ding et al., 2011; Liu et al.,
2006; Pang et al., 2011; Wang et al., 2009; Yang et al., 2012; Zhang et al.,
2014, 2010). At the same time, representative modern reference
datasets (Cao et al., 2014; Xu et al., 2010a; Yu et al., 2000; Zheng et al.,
2008, 2014) and fossil pollen datasets (Cao et al., 2013; Ren
and Beug, 2002; Sun et al., 1999) have been assembled, which make it
possible to reconstruct the vegetation and climate for individual sites,
regions, or the of whole China (Chen et al., 2015; Ni et al., 2014; Tian et al.,
2016; Wang et al., 2014; Xu et al., 2010b). Despite the aim of these studies
to use modern surface samples from natural (i.e. likely undisturbed)
vegetation communities for establishing their calibration datasets and to
exclude samples representing human-disturbed vegetation communities (Zheng
et al., 2014), the presence in the datasets of some samples from eastern
China referred to as “troublesome” due to intense human impact surrounding
the vegetation patches (Xu et al., 2010a) suggests that the problem was not
completely resolved. Li et al. (2014) assessed the reference pollen data in
the currently available datasets from central-eastern China using a human
influence index (HII) and concluded that surface samples are biased due to
significant human impact on the natural vegetation. Pollen-based climate
reconstructions for the late Holocene in this region would thus also be
biased.
Obtaining reliable climate reconstructions in regions with a long history of
human activities is indeed a big challenge but also an important scientific
task. For example, reconstructing rainfall in temperate eastern China (TEC)
is not only necessary for understanding the East Asian summer monsoon (EASM)
variations (Chen et al., 2015; Guiot et al., 2008; Wen et al., 2013; Xu et
al., 2010b), but also very important when studying human adaptation to
climate change and the origins of agriculture and cultural evolution
(Lu et al., 2009; Mu et al., 2015; Tarasov et al., 2006). In this paper, we
(1) compiled a human-induced training set with 791 surface pollen spectra
and a corresponding natural training set with 806 spectra from TEC,
(2) compared the pollen–climate model performances of the two calibration
sets, (3) investigated the deviations in the reconstructed results for a
fossil record based on each calibration set, and (4) discuss the mechanism of
bias caused by human impact.
Regional setting
Temperate eastern China (TEC; 30–53∘ N, 100–135∘ E) was
chosen as the study area for its ecological sensitivity to climate change
(e.g. forest-steppe boundary shifts with EASM variations) and long-term human
impact on vegetation (Guiot et al., 2008; Liu et al., 2014; Ren, 2000; Xiao
et al., 2004). The region extends from the eastern margin of the Tibetan
Plateau (TP) to the Yellow Sea coastline and from the northern catchment of
the Yangtze River to the Heilong (Amur) River, covering about one-third of
China (Fig. 1a). Topographically, it encompasses three distinct levels,
showing a decrease in elevation from the TP margin (2000–4000 m) to the
Inner Mongolian plateau, Loess Plateau, Qin Mountains, Taihang Mountains, and
Greater Khingan Mountains (1000–2000 m) and to the eastern hilly areas and
flood plains (< 200–500 m). The south-eastern part of the study area is
dominated by EASM, while the north-western part is influenced by the
westerlies (Fu et al., 2008). From the coast inland, mean annual
precipitation (Pann) varies from 1400 to 35 mm, covering the
conventional humid, subhumid, semi-arid, and arid areas, and mean annual
temperature (Tann) decreases from 18 to -6 ∘C from
south to north (Domrös and Peng, 1988).
Due to the large climatic and topographic gradients, several large-scale
natural vegetation regions have been described for the study area (Fig. 1):
(I) cold temperate needleleaf deciduous forest region, (II) temperate mixed
needleleaf and deciduous broadleaf forest region, (III) warm temperate
deciduous broadleaf forest region, (VI) temperate steppe region,
(IVAi) northern subtropical broadleaf evergreen-deciduous forest zone,
(VIIBi) temperate semi-shrub and shrub desert zone, and (VIIIAi) subalpine
scrub and alpine meadow zone (Editorial Committee of Vegetation Map of China,
2007; Wu et al., 2013). However, many natural ecosystems have been intensely
modified by settlements and agricultural land use. For example, in 2015,
forest coverage in the supposedly densely forested north-east of China was about
41 % and only 26 % in the warm temperate forest region, according to
data from the China National Bureau of Statistics
(http://data.stats.gov.cn).
Maps of the study region showing the distributions
of (a) vegetation regions and types, the 1208 meteorological
stations, and (b) HII values with surface pollen sampling sites for
the
N-set (circles) and H-set (crosses). Selected large cities (red pentagram and
dots), mountains (Mts.), rivers (R.), isohyets of 200, 400, and 800 mm (for
annual precipitation), and the location of Lake Gonghai are marked.
![]()
Materials and methods
Surface pollen data
We use pollen data from a number of studies attempting to detect
human-induced changes in surface pollen assemblages, including 43 spectra
from the Anyang area in the central China Plain (Wang et al., 2009), 12 from the eastern
Hexi Corridor (Ma et al., 2009), 78 from warm temperate hilly areas (Ding et
al., 2011), 88 from the Hebei Plain and adjacent mountain area (Pang et al.,
2011), 13 from south-east China (Yang et al., 2012), and 105 spectra from
north-east China (Li et al., 2012, 2015). Additionally, 70 unpublished
spectra from the coastal plain between the Yellow River and the Yangtze River
were generated for the purpose of this study. The samples were mostly
collected from croplands, abandoned croplands, economic gardens and forests,
pasturelands, and roadside scrub and woodlands. The field sampling
strategies, laboratory procedures, analytical methods, pollen taxa, and other
detailed information are described in the aforementioned studies.
Additionally, we make use of some reference pollen datasets partly or
entirely covering the study area (Wen et al., 2013; Xu et al., 2010a, 2007;
Zheng et al., 2008), which were used to represent natural vegetation
communities.
Samples from the different natural vegetation communities were integrated
into a natural dataset (N-set), while samples from human-induced
vegetation or vegetation patches obviously disturbed by human activities were
integrated into a human-induced dataset (H-set; Fig. 1b). We used
reference samples from an approximate extent of 31–51∘ N and
102–130∘ E with a sufficiently large Pann gradient of
150–1100 mm and a Tann gradient of -3 to
16 ∘C in order to cover the greatest possible
climate range likely to be encountered in the fossil pollen record (see
Sect. 3.3). Pollen percentages were recalculated based on the sum of
terrestrial taxa. It should be clearly noted that pollen from cereal-type
Poaceae and other distinct cultivated taxa (e.g. Brassica,
Gossypium, Sesamum, and Linum) identified in the
H-set (Ding et al., 2011; Li et al., 2015) were excluded to reduce
anthropogenic noise. This strategy is similar to the one excluding aquatic
pollen and spores in order to better catch the climatic signal. Finally, 806
spectra and 151 taxa form the N-set, and 791 spectra and 147 taxa form the
H-set. The N-set includes 11 samples from vegetation region I, 61 from
region II, 300 from region III, 351 from region VI, 25 from region IVAi, 51
from region VIIBi, and 7 from region VIIIAi, while 14 samples from region I,
11 from region II, 433 from region III, 292 from region VI, and 41 from
region IVAi appear in the H-set.
Simplified pollen percentage diagram of core GH09B from Lake
Gonghai. The local pollen zone (and subzone) boundaries are based on the
results of a constrained cluster analysis with the CONISS programme in Tilia software (Grimm,
1987, 2011) used to make the pollen diagram. The detailed information on
vegetation succession was presented in Xu et al. (2016a).
![]()
Modern climate and human influence index data
Mean monthly climate averages were derived from the latest available
observation data (1981–2010) from 1208 well-distributed meteorological
stations across the study area (Fig. 1a). The original data can be accessed
from the China National Meteorological Information Center
(http://data.cma.cn). Mean values of annual precipitation
(Pann) and temperature (Tann) and the mean temperature
of the coldest (Mtco) and warmest month (Mtwa) were
selected as the transfer function variables. These four climate parameters
were estimated for each pollen site with the Polation 1.1 software
(http://polsystems.rits-palaeo.com). A vertical lapse rate of
0.6 ∘C/100 m, as suggested for China (Domrös and Peng, 1988),
was applied and leave-one-out cross validation was used to assess the
interpolation accuracy. Correlation coefficients (R) between estimated and
observed climatic values of 0.97–0.99 suggest that the results are robust.
Sanderson et al. (2002) developed a human influence index (HII) dataset for
mapping the areas with and without a human footprint. The HII dataset
quantifies human influence on terrestrial ecosystems based on four proxies
(nine datasets), including human population pressure (population density),
land transformation (land use/cover, roads and railways, built-up centres,
settlements), accessibility (roads and railways, coastlines, navigable
rivers), and electrical power infrastructure (night-time lights). Each of the
nine datasets assigns a score from 0 to 10 according to a rating (or
alternatively in a single score and 0) to assess human influence on 1 km2
of land surface. Sum scores from nine datasets were standardized as HII
values,
which range from 0 to 64 (WCS/CIESIN, 2005), with higher values indicating
a higher degree of human impact (Fig. 1b). HII reflects only modern human
impact and does not consider the human impact during the past. The HII has
been recently employed to assess human influence on pollen assemblages in
China (Li et al., 2014, 2015). In this study, HII values at 1 km2 grids
are assigned to modern pollen reference sites using ArcGIS.
Fossil pollen record from Lake Gonghai
Lake Gonghai (38∘54′ N, 112∘14′ E,
1860 m above mean sea level) is a small
(0.18 km2) hydrologically closed alpine lake with a water supply mainly
from summer precipitation (Chen et al., 2015). It is located on the
north-east margin of the Loess Plateau (Fig. 1). The lake lies close to the
modern EASM border in the forest-steppe ecotone and experiences subhumid to
semi-arid transitional moisture conditions. The upper 9.42 m of a core,
GH09B, from Lake Gonghai was subsampled at 1 cm intervals for pollen
analysis. Twenty-five (including seven from parallel core GH09C) accelerator
mass spectrometry (AMS) 14C dates of terrestrial plant macrofossils and
35 210Pb/137Cs dates of the uppermost 0.35 m of the lake sediment
were used to establish a robust age–depth model (Chen et al., 2015; Xu et
al., 2016a).
Vegetation succession in the area around Lake Gonghai has experienced five
major stages during the last 14.7 kyr (Fig. 2). Open forests and upland
meadows dominated during the last deglaciation (14.7–11.1 ka), and abrupt
strengthening of Artemisia-dominated mountain steppe association
occurred during 13.1–12.0 ka BP. In the early Holocene (11.1–9.6 ka),
Betula, Carpinus, Ostryopsis, and Ulmus
as pioneer tree species spread into the landscape. Later (9.6–7.3 ka) mixed
forest dominated by Picea and Betula started to play a
greater role. Temperate deciduous trees (e.g. Quercus) widely
expanded and the mixed broadleaf-conifer forest grew around the lake during
the Holocene climatic optimum (7.3–5.0 ka BP). The break-up of the climax
community started from 5.0 ka BP with the increasing of Pinus and
the decreasing of Quercus and Betula percentages. A major
increase in herbaceous pollen percentages occurred since 2.8 ka BP,
especially during the last 1.6 kyr when Humulus (including
Urtica), Fagopyrum, and cereal Poaceae pollen types related
to human activities became most prominent in the diagram (Fig. 2; Xu et al.,
2016a). The vegetation dynamic inferred from the GH09B pollen sequence is a valuable
source of environmental information for the EASM margin (Chen et al., 2015).
In the current study, we selected this well-dated and high-resolution fossil
record to reconstruct past climate using both the N-set and H-set of modern
pollen as calibration sets.
Ordination results of redundancy analysis (RDA) for 15 major pollen
taxa with climate variables (Pann, Tann,
Mtco, and Mtwa) and human influence index (HII) for
(a) the natural set and (b) the human-induced
set.
![]()
Numerical analyses
Relationships between surface pollen spectra and climate variables are
assessed by ordination techniques. To stabilise the variance and optimise the
signal-to-noise ratio in the data, pollen taxa which occur in at least three
samples and contribute ≥ 3 % in at least one sample were selected
and square-root transformed for further analyses (Prentice, 1980). The length
of the first axis in detrended correspondence analysis (DCA; Hill and Gauch,
1980) was used to determine whether redundancy analysis (RDA) or canonical
correspondence analysis (CCA) should be chosen for the constrained ordination
(ter Braak and Prentice, 1988). The ratio of the constrained eigenvalue to
the first unconstrained eigenvalue (λ1/λ2) for a climate
variable is used to assess its potential to be reconstructed (ter Braak,
1987). A value of λ1/λ2 greater than 1 suggests that
the variable is the main determinant in the dataset; otherwise, the
reconstruction of the variable should be conducted with caution (Juggins,
2013).
HII was also analysed in the same way to evaluate the human impact on the
pollen data and the quality of the training set. As HII is an environmental
variable with certain stochastic features in locations and intensity,
there is no robust ecological basis to estimate the optima and tolerance for
a pollen taxon using any pollen–HII calibration set. Therefore, we used an
indirect method to assess the potential bias induced from the training set due
to human impact on surface samples. At first, we found five closest modern
analogues for each fossil sample using MAT (Simpson, 2007) and then used
the mean HII value at the analogue sites to examine the human influence on the
analogue samples and to evaluate the bias in climate reconstruction for that
given fossil sample.
The WA-PLS approach combines the virtues of the WA method to model ecological
optima of species and the PLS method to select linear components from biological
assemblages (ter Braak and Juggins, 1993). It has been tested along with the WA,
MAT, and pollen response surface method (PRS) for eastern China data and
demonstrated to give better results (Cao et al., 2014; Xu et al., 2010b) due
to its generally good performance under non-analogue situations and its ability
to cope with spatial autocorrelation (Cao et al., 2014; Juggins and Birks,
2012). The optimal number of WA-PLS components was selected using a
randomization t test (van der Voet, 1994). Low root mean squared error of
prediction (RMSEP), low average and maximum biases, a high coefficient of
determination (R2) between the predicted and observed climate values,
and a rule-of-thumb threshold of 5 % (reduction in RMSEP for adding a
component) were all considered when selecting a model (Birks, 1998; Birks et
al., 2010; Juggins and Birks, 2012).
The significance of the obtained reconstructions was also tested. The
proportion of variance in the fossil sequence explained by 999 transfer
functions trained with random data was calculated from a constrained
ordination (Telford and Birks, 2011). To help understand the bias mechanism
of the human impact on pollen assemblages, we estimated the WA optima and
tolerances (Birks et al., 1990; ter Braak and Looman, 1986) of selected
climate variables for major taxa. All numerical analyses were performed using
vegan version 2.3-5 (Oksanen et al., 2016), analogue
version 0.17-0 (Simpson, 2007), rioja version 0.9-5 (Juggins, 2015),
and palaeoSig version 1.1-3 (Telford, 2015) in the R 3.2.4 environment
(R Core Team, 2016).
Scatter plots of pollen-based predicted annual precipitation
(Pann) and observed Pann using two-component weighted
averaging partial least squares (WA-PLS) models for (a) the natural
set and (b) the human-induced set.
![]()
Summary statistics for redundancy analysis (RDA) with pollen species
and climate variables (annual precipitation Pann, mean annual
temperature Tann, mean temperature of the coldest month
Mtco, mean temperature of the warmest month Mtwa) and
the human influence index (HII). Sole expl (%) is the pollen variance
explained by variables as a sole predictor; marg expl (%) is the marginal
contribution of this variable in the model with all other variables. All
P values are 0.001 (based on 999 unrestricted Monte Carlo
permutations).
Variables
Natural dataset
Human-induced dataset
λ1/λ2
Sole expl (%)
Marg expl (%)
λ1/λ2
Sole expl (%)
Marg expl (%)
Pann
1.28
20.14
10.56
0.34
6.31
5.85
Tann
0.08
2.83
1.20
0.26
5.57
1.55
Mtco
0.11
3.49
1.23
0.18
3.90
1.43
Mtwa
0.21
6.35
1.37
0.31
6.62
1.44
HII
0.03
1.12
0.76
0.10
2.29
0.55
Results
Relationship between modern pollen and climate
Ordinations are based on square-root-transformed pollen data of 99 taxa in
the N-set and 93 taxa in the H-set after noise reduction. DCA showed that the
length of the first axis is 2.65 SD (standard deviation units) in the N-set
and 2.36 SD in the H-set, suggesting that linear ordination techniques (e.g.
RDA) are appropriate to present the distribution of pollen taxa along the
climate gradients in our datasets. When using each of the climatic variables
as a sole predictor, Pann explains 20.56 % (highest) of the
pollen assemblage variance in the N-set, while the thermal variables have
much lower explanatory power (Tann: 2.83 %, Mtco:
3.49 %, Mtwa: 6.35 %). For the H-set, Pann
explains 6.31 %, which is slightly less than Mtwa
(6.62 %). If we assess the marginal contribution of a variable after
partialling out the interaction effect of other variables in an RDA,
Pann explains the highest amount of variance in both the N-set
(10.56 %) and the H-set (5.85 %). HII explains more variance in the
H-set (2.29 %) than in the N-set (1.12 %) and has a marginal
contribution of 0.55 % in the H-set and 0.76 % in the N-set.
Pann has the highest λ1/λ2 ratio in both the
N-set (1.28) and the H-set (0.34); the λ1/λ2 ratios for
all thermal variables and HII are much lower than 1 (Table 1). To better
illustrate the modern pollen–climate relationships and their difference
between natural and human-impact scenarios, 15 major pollen taxa, which are also
identified in the GH09B fossil sequence (Fig. 2), were selected to reveal the
relationship between modern pollen and climate (Fig. 3). It seems that
the general pattern of tree and shrub–herb group separation is maintained, but
the relationship of some pollen taxa (e.g. Picea, Betula,
Poaceae, and Chenopodiaceae) with climatic variables (e.g. Pann)
is altered by human influence. The greater ordination difference in Poaceae and
Chenopodiaceae in the two sets indicates that these two taxa are more sensitive
to human impact. Our ordination results suggest that Pann is the main
determinant of pollen distribution in TEC, and Pann in the N-set
is used to establish a standard calibration set. We then use the H-set to
establish a contrasting pollen–Pann calibration set to compare
the deviation in the reconstructions and to see the extent of the potential
bias induced from human impact on the modern pollen assemblages.
Summary performance statistics of the first three components of the
weighted averaging partial least squares regression (WA-PLS) for annual
precipitation (Pann) based on leave-one-out cross validation for
the natural set (N-set) and the human-induced set (H-set). Coefficient of determination
between predicted and observed Pann (R2), root mean square
error of prediction (RMSEP; mm), average bias (ave. bias) and maximum bias
(max. bias), RMSEP change in percentage (%Change), and p value are given.
The selected models are shown in bold.
Datasets
Model
R2
RMSEP
Ave. bias
Max. bias
%Change
p value
N-set
WA-PLS Component 1
0.83
96.02
-0.34
82.10
–
–
WA-PLS Component 2
0.86
89.24
1.10
72.98
-7.06
0.001
WA-PLS Component 3
0.86
87.92
-0.25
64.53
-1.48
0.098
H-set
WA-PLS Component 1
0.58
108.32
-1.39
218.50
–
–
WA-PLS Component 2
0.64
100.48
1.02
183.06
-7.23
0.001
WA-PLS Component 3
0.64
100.84
0.99
177.97
0.36
0.614
Test of the WA-PLS models
A two-component WA-PLS model performed best with the lowest RMSEP and highest
R2 for the H-set, and the best performer was a three-component model for the N-set (Table 2).
However, the improvement (1.48 % reduction in RMSEP) over the two-component
model was less than the threshold of 5 % and not significantly different
(P=0.098), and therefore we selected a two-component WA-PLS model for both
datasets. The R2 between predicted Pann values and observed
values in the N-set is 0.86 and the RMSEP is 89 mm. Both are better than
those for the H-set (R2=0.64; RMSEP = 100 mm; Fig. 4). The
percentage of RMSEP to the sampled Pann gradient (940 mm) for
the N-set is 9.47 % and for the H-set (927 mm) 10.79 %.
Overestimates at the low end of the Pann gradient (i.e. for sites
from arid areas) and underestimates at the high end (sites from humid areas),
which is an inevitable systematic bias in all WA-based models referred to
as an “edge effect”, are larger in the H-set than in the N-set.
Proportion of variance (solid lines) in Gonghai Lake fossil pollen
data explained by annual precipitation (Pann) transfer functions
with (a) the natural set (N-set) and (b) the human-induced
set (H-set). The thick black dotted lines indicate the proportion of variance
explained by the first axis of a principal components analysis (PCA), and the
fine red dotted lines indicate the 0.05 significance level. Histograms show
the amount of variance explained by 999 transfer functions with random data.
![]()
Fossil pollen record of Lake Gonghai: (a) reconstructed
annual precipitation using a natural set (N-set; blue) with upper-side standard
error of prediction (SEP) and a human-induced set (H-set; red) with lower-side
SEP. The modern instrumental value is marked on the scale axis for comparison
(purple arrow). (b) Deviations in the reconstructed precipitation
(grey) between the N-set- and H-set-based transfer functions with a 5-point
moving average smoother (black). (c) The proportion of tree pollen
taxa (%) and (d) the mean HII value of the five best analogues
in the N-set (blue, with lower-side standard deviation) and the H-set (red,
with higher-side standard deviation) for fossil samples. Six time windows
(TWs) delineated according to the deviation pattern between the two
reconstructions are separated by grey dashed lines.
![]()
Annual precipitation (<inline-formula><mml:math id="M92" display="inline"><mml:mrow><mml:msub><mml:mi>P</mml:mi><mml:mi mathvariant="normal">ann</mml:mi></mml:msub><mml:mo>)</mml:mo></mml:mrow></mml:math></inline-formula> reconstructions for Lake
Gonghai
We applied the pollen–Pann WA-PLS models to the Lake Gonghai
fossil record (Chen et al., 2015; Xu et al., 2016a). The proportion of the
variance in the fossil data explained by the first PCA axis is 47.79 %,
and the significance tests suggest that 41.57 % of the variance in the N-set can
be explained by Pann (P=0.001) and 23.32 % for the H-set
(P=0.033; Fig. 5). The two calibration sets produced very similar
reconstructed Pann patterns (including major trends and
change points) but with distinct deviations in their values (range:
-105 to 95 mm and SD = 46 mm) most of
the time (Fig. 6a). The deviation is calculated as the reconstructed N-set
Pann value minus the corresponding H-set Pann value
(Fig. 6b). From the deviation pattern, six zones or time windows (TWs) are
demarcated. From 14.7 to 13.1 ka BP (TW1), the deviation decreases
gradually from -70 to 20 mm. During 13.1–12 ka BP (TW2), the deviation
fluctuates between -30 and 30 mm and the mean value for this period is
only 3 mm. The deviation varies from -20 to 100 mm and generally appears
to increase in the 12.0–7.3 ka BP interval (TW3). A downward trend can be
observed starting from 7.3 ka BP (TW4), and the deviation decreases from a
relative stable value of 75 mm to around zero (±35 mm) during
2.8–1.6 ka BP (TW5). In the most recent period (TW6), the deviation is
mostly negative with a mean value of -40 mm. The deviation curve over the
last 14.7 kyr generally correlates (R=0.81) with the tree pollen
percentage curve (Fig. 6c).
WA optima estimates and analogue measures
The WA optima and tolerances of 15 major pollen taxa in both the N- and
H-sets were estimated (Fig. 7). The optima for Pann for most tree
taxa (Picea, Pinus, Betula, Quercus,
Carpinus, Juglans) and for Corylus (representing
shrubs and small trees) in the H-set are shifted to drier conditions compared
to those in the N-set. For most herbaceous taxa (Chenopodiaceae,
Polygonum, Poaceae, Artemisia, Ranunculaceae) and the
drought-enduring shrub Nitraria, the optima are shifted towards
wetter conditions. Ulmus (a commonly cultivated tree) and Cyperaceae
(species-rich herbaceous taxon) are exceptions in the arboreal and
non-arboreal groups, respectively. The estimated range of tolerance in the
H-set is compressed for most taxa in comparison to the N-set, especially for
tree taxa, which shrink by about 18–55 % (Fig. 7). The mean HII values of
the five best modern analogues in the H-set are generally higher than those
in the N-set, except for 13.1–12 ka BP and the last 1.6 kyr period when
they are relatively close; both are low for the mid-Holocene fossil samples
(Fig. 6d).
Caterpillar plot of weighted average (WA) optima and tolerances for
15 major pollen taxa in response to annual precipitation (Pann).
The taxa are arranged by optima values and taxa groups. Human impact
generally shifts the inferred optima of woody taxa and herb taxa in opposite
directions and compresses the tolerances for most taxa in the study area.
![]()
Discussion
Climatic signals in pollen assemblages obscured by human impact
The pollen record is a complex and non-linear function of vegetation, which in
turn is a function of climate based on some key assumptions (Birks et al.,
2010). The big challenge for pollen-based climate reconstructions is that
this indirect pollen–climate relationship can be affected by several other
(non-climatic) factors, for example, by human activities (Birks and
Seppä, 2004; Ren, 2000; Xu et al., 2010b). RDA results show that the
ability of Pann to explain pollen variance declines a lot in the
H-set in comparison to the N-set (Table 1). The statistical performance of
WA-PLS for the H-set is poorer (Table 2), suggesting that climatic signals in the
H-set have been partly obscured by human impact. Due to agricultural land use
and human-induced deforestation of the plains and hilly areas (e.g. terraced
fields) in the humid and subhumid regions, tree pollen percentages decrease
and herb percentages increase substantially in surface samples, even after
excluding distinctly cultivated taxa. It is easy to imagine that herbaceous
taxa, such as Artemisia, Chenopodiaceae, Humulus, and
weed-type Poaceae would expand after forest clearance and this will
change the regional vegetation composition and relative pollen abundances
(Ding et al., 2011; Li et al., 2015). It will also alter the pollen–climate
relationships for many pollen taxa in the response models (St. Jacques et
al., 2008). This alteration can be seen in the comparisons of RDA ordination
(Fig. 3) and estimated WA optima and tolerances (Fig. 7) for 15 major taxa
between the two datasets.
Selected pollen taxa can be separated into two tree and herb groups by
contrasting their WA optima in the N- and H-sets (Fig. 7). The inferred
optima of most woody taxa in the H-set are shifted towards drier conditions
and their tolerances compressed. This means that a fossil pollen assemblage
with a high proportion of woody taxa would be assigned a lower
Pann value when the H-set is employed in the transfer function.
Conversely, Pann values will be overestimated when herbaceous
taxa dominate in a fossil sample. This is clearly seen in the
Pann curves for Lake Gonghai (Fig. 6a). The reconstructed
Pann deviations between the N- and H-sets and the tree pollen
percentage curve demonstrate similar trends (Fig. 6) and are statistically
correlated (R=0.81). However, the Pann deviation is not
simply determined by the proportion of tree and herb taxa. For example, the
deviations in (time window) TW2 and the later TW5 are both around zero, but
the tree pollen comprises 20–30 and 40–50 %, respectively, rather than
being equal with the percentage of herbs. This is a consequence of the
vegetation composition and species characteristics.
In short, human impact obscures the climatic signals in pollen spectra by
distorting the response relationship between pollen abundance and climate
(Birks et al., 2010; Seppä et al., 2004), thus influencing the assumed
climatic optima and tolerances of pollen taxa in the model (Fig. 7). When
such a human-influenced calibration set is applied to a fossil record, which
represents mostly natural vegetation, a more or less serious bias in the
reconstructed past climate should be expected (St. Jacques et al., 2008; Xu
et al., 2010a). Using the H-set in this study, significantly lower
(P < 0.001) Pann values relative to the N-set are
reconstructed for Lake Gonghai during the early and middle Holocene (TW3 and
TW4) when tree pollen contributed more than 40 % to the total pollen
sum. This suggests that drier biases may have also been induced from surface
samples using the N-set for this period. Conversely, we note a bias towards a
wetter climate reconstruction for the late glacial (TW1) and the last 1600 years
(TW6). The sites comprising the two sets, N and H, are not perfectly even
distributed, which may also influence the optima estimates (Fig. 1b). However,
the model-inferred group-optima change pattern is statistically and
ecologically reliable (Fig. 7). We consider it likely that a similar effect
will occur in the pollen-based climate reconstructions for the whole TEC
region where the vegetation pattern has been largely shaped by EASM-induced
rainfall.
Human influence index as an assessment tool
Modern HIIs capture basic characteristics of human influence on ecosystems
and allow for a quantitative evaluation of the human impact on the land surface
(Sanderson et al., 2002). Li et al. (2014) innovatively employed an HII to
establish a calibration set with pollen data and applied it to a 6200-year
fossil record from Lake Tianchi in central China (Zhao et al., 2010). The
pollen–HII calibration model (R2=0.47) was based on 185 modern samples
from central-eastern China (a warm temperate forest region), and
the variance in the training set explained by HII (6.79 %) is comparable
to Pann (7.78 %) and Tann (6.00 %; Li et
al., 2014). A further investigation based on 189 surface pollen samples from
northern China (involving both human-induced and natural samples from
vegetation regions II, III, and VI) provided a higher correlation (R2=0.69) between pollen and HII values in a WA-PLS model (Li et al., 2015).
This good statistical performance gives us more confidence in assessing human
influence on vegetation using the HII, although we are aware of some
difficulties in applying a quantitative pollen–HII calibration model to the
fossil data.
A good correlation of the HII data with cereal-type Poaceae pollen in
northern China (Li et al., 2015) suggests that the HII can be seen as a
surrogate of indicator pollen taxa for human activities. However, cereal-type
Poaceae pollen generally has a very low abundance in a fossil sequence. For
example, the cereal-type Poaceae in a natural profile close to the
archaeological sites from Anyang, the centre of agricultural and societal
development during late Shang Dynasty, comprises only around 2 % of the
total pollen during the last 3400 years (Cao et al., 2010). In the sequence
from Lake Gonghai used in the current study, it contributes about 2–4 %
during the last 2 millennia (Fig. 2). Therefore, HII explains only
1.12 % of the variance in the N-set and 2.29 % in the H-set after
removing the cereal-type Poaceae and other distinct cultivars (Table 1). In
addition to cereal-type Poaceae, taxa such as Artemisia,
Chenopodiaceae, and weed Poaceae, which could be dominant in both steppe
areas and subhumid areas after forest clearance (Ding et al., 2011; Li et
al., 2008; Liu et al., 2006), may challenge the interpretations. Although
the human impact can be detected using additional information, including charcoal
and archaeological data (Zhao et al., 2010), compositional changes in these
taxa during the late Holocene due to human activities are hard to distinguish
from those caused by a progressively drier EASM climate, especially in fossil
pollen records from the forest-steppe ecotone in TEC.
Reconstructing human influence quantitatively from fossil pollen data with a
direct pollen–HII calibration set might not be an easy task in most cases (Li
et al., 2014), but we can still use HII as an assessment tool in a
broad spectrum way. The reconstructed climate of a certain fossil sample is
mostly determined by its closest modern analogues even though different
approaches may have been used for the reconstruction (e.g. WA-PLS; Birks et
al., 1990). By examining the mean HII values at sites of the best modern
analogues, we can evaluate the bias in the climate reconstruction of the
corresponding fossil sample. A high analogue HII value indicates greater
potential bias in the reference samples. As shown in Fig. 6d, analogue HII
values in the H-set are usually higher than in the N-set, suggesting a higher
bias in the Pann reconstruction. The mean analogue HII values in
the N-set fluctuate at around 15 during the mid-Holocene, indicating that the
climate reconstruction for this interval has the lowest human-induced bias as
an HII value of 15 is the mean background value for modern natural
vegetation patches in TEC (Li et al., 2015; Sanderson et al., 2002). Similar
to the HII trend in Lake Tianchi (Li et al., 2014), analogue HII values for
Lake Gonghai start to rise around 2.8–2.9 ka BP, which conforms to the
scenario of agricultural advancements and population growth in Bronze Age
China during the Western Zhou period (1045–771 BC; Li, 2006). Relatively
higher analogue HII values during the late glacial and early Holocene suggest
that modern analogues for this period in the current reference dataset have
experienced more human influence. Together with the common problem of no
analogues for this period (Jackson and Williams, 2004), climate
reconstructions for this interval in TEC should also be considered more
carefully.
Implications for Holocene climate reconstructions
Agriculture became the dominant subsistence strategy in today's (potentially)
warm temperate forest region (III, including the central China Plains) and
northern subtropical mixed forest zone (IVAi, including the Yangtze Plains)
from about 6.5–5 ka BP (Crawford, 2011; Zhao, 2010). Potential human
disturbance to the vegetation in eastern China since 6 ka BP has been
inferred from many pollen studies (Ren and Beug, 2002; Wang et al., 2010),
not to mention historical times (Cao et al., 2010; Zhao et al., 2010). Our
analogue HII assessments indicate that the bias in the climate reconstruction
induced from human impact via reference pollen samples or via changes in the
fossil pollen assemblages (particularly during historical times; Li et al.,
2014) is real. This raises the question of whether the Holocene climate
could be quantitatively reconstructed using pollen data from eastern China.
The answer is not a simple yes or no (Ren and Beug, 2002), although
we keep an optimistic view based on the comparison results presented in this
study.
Reconstructed Pann for the uppermost 17 samples from the Gonghai
record, representing AD 1950–2008, is 428 mm (SD = 26 mm) with the N-set
and 439 mm (SD = 25 mm) with the H-set. The modern mean
Pann around the Lake Gonghai area is 445 mm according to
instrumental data between 1959 and 2011 from Ningwu station, located 11 km
north of the lake (Chen et al., 2015). Both calibration sets provide reliable
(more or less similar) reconstructions for recent decades when compared with
historical measurements, although deviations consistently exist since
the last deglaciation (Fig. 6). By examining the deviation range between the
two datasets, and bearing in mind that surface reference samples in the H-set
are more or less strongly affected by human activities, we can assume that
the Pann reconstruction based on the N-set should be closer to
reality. Holocene pollen–climate relationships in China are relatively stable
(Tian et al., 2017). If we select appropriate sites, such as Tianchi and
Gonghai lakes, both of which are small closed alpine lakes with very limited human
impact before 3 ka BP (Institute of Archaeology CASS, 2004, 2003), we can
still hope to get a reliable Pann reconstruction prior to the
late Holocene. Our Holocene Pann reconstruction using the N-set
can further the discussion on cultural evolution and the origin of dry-land
agriculture in the study region. For example, a remarkable Pann
increase from 480 to 570 mm along the present-day EASM margin during
8.6–7.8 ka BP could have promoted the development of millet agriculture
(Liu et al., 2012; Zhao, 2011). It also supports the hypothesis that at early
Neolithic sites broomcorn millet (low Pann requirement:
350–450 mm) is more abundant than foxtail millet (optima: 450–550 mm) as
broomcorn millet is better adapted to drought conditions (Lu et al., 2009).
Regarding the potential human-induced bias in fossil records, the challenge
for pollen-based quantitative climate reconstructions is more from the lack
of natural surface samples in regions with intensive agricultural
activities. In eastern China, a calibration set only including preferenced
pollen samples from lake surface sediments with low human disturbance is
still not available (Liu et al., 2013), and surface samples are mostly
collected from mountains and steppe areas (Xu et al., 2010b). This means that
our current modern pollen datasets (Cao et al., 2014; Zheng et al., 2014)
still contain relatively few samples from central-eastern China. Collecting
extra samples from natural vegetation in mountain areas, such as
Luzhong, Qin, Dabie, and Qian, would help to improve the
pollen-based climate reconstructions for the region. The Qin Mountains, for
example, have a large and well-forested range (ca. 57 000 km2) but
are represented by relatively few pollen samples (Fig. 1b). In addition,
there are still hundreds of small forests (patches) in the hilly areas,
around the lakes, and within natural parks in eastern China which deserve
the attention of palynologists.