Landscape Ecol (2007) 22:1143–1154

DOI 10.1007/s10980-007-9096-4

RESEARCH ARTICLE

Predicting the spatial distribution of an invasive plant

species (Eupatorium adenophorum) in China
Li Zhu Æ Osbert J. Sun Æ Weiguo Sang Æ
Zhenyu Li Æ Keping Ma

Received: 24 January 2006 / Accepted: 22 March 2007 / Published online: 12 April 2007
Ó Springer Science+Business Media B.V. 2007

Abstract Invasive alien species can pose a severe air temperature strongly influenced the predictions.
threat to biodiversity and stability of the ecosystems Our results indicate that crofton weed may break out
they invade. Predicting distribution patterns of inva- in Yungui Plateau, Sichuan Basin, southeastern
sive species in regions outside their native range is a Coastlands, Hainan Island, and Taiwan although
fundamental component of early warning systems. currently it is either absent or has only recently been
Crofton weed (Eupatorium adenophorum Spreng.) recorded in these regions. Redundancy analysis
was first discovered in the Yunnan Province of China (RDA) ordination results demonstrated that temper-
around the 1940s. The well-documented invasion ature and precipitation play an important role in
history of this plant species provided the opportunity confining the spread of crofton weed. Over the past
for us to examine the spatiotemporal patterns of 20 years, crofton weed has spread from subtropical
biological invasion by crofton weed. Using the areas with higher annual mean temperature and lower
datasets documenting 441 known localities invaded climatic fluctuations to much cooler and dryer areas
by crofton weed in China over the past 50 years and at higher altitudes. The distribution of crofton weed
23 environmental variables generated by the genetic was restricted mainly to regions with mean annual air
algorithm for rule-set production (GARP) model, we temperature ranging from 10 to 228C and annual
tested the predictability of crofton weed distribution precipitation from 800 to 2000 mm. Our results could
with a high degree of accuracy. Both the Kappa help in developing and implementing early detection
statistics and the receiver–operator characteristic measures to minimize the ecological impacts of
(ROC) analysis indicated that it is possible to predict crofton weed invasion in China.
the geographical spread of crofton weed in China.
Precipitation in the coldest quarter of the year, Keywords Geographic potential
Ecological niche
extremely low air temperature, and maximum annual modeling
Eupatorium adenophorum Spreng.

Spatiotemporal pattern
Biological invasions
China

L. Zhu
O. J. Sun
W. Sang (&)
Z. Li
K. Ma

Laboratory of Quantitative Vegetation Ecology, Institute
of Botany, The Chinese Academy of Sciences, Beijing Introduction
100093, China
e-mail: [email protected] Biological invasion by alien plants has been identified
as one of the most important contemporary ecological
L. Zhu
Graduate University of the Chinese Academy of Sciences, problems because of the potential threat it poses to
Beijing 100049, China biodiversity and stability of ecosystems (Wilcove

123
1144 Landscape Ecol (2007) 22:1143–1154

et al. 1998; Reichard and White 2003) and its adverse regression, in an iterative, artificial-intelligence-based
impacts on agricultural and forest productivity (Tur- system and uses raster-based environmental and
pie et al. 2003; Knowler and Barbier 2005). Increasing biological data to produce a niche-based model of
global economic trade and trans-boundary activities the environmental requirements of the given species
make it almost impossible to avoid the spread of alien (Stockwell and Noble 1992; Stockwell and Peters
species across continents (McNeely 1999; Normile 1999). These models can then be projected onto other
2004) or the movement of species across different landscapes to predict the geographic distribution of
habitats within a region. Being able to understand the species (Lim et al. 2002; Parra et al. 2004). By
factors that regulate the spread of invasive species is extrapolating the model to a region free from the
an important goal of landscape ecology (With 2002). species in question, the potential of that species to
The ability to predict distribution patterns of invasive invade the region can be predicted (Peterson and
plants in regions outside their native range is funda- Vieglais 2001). The predictive power of this
mental to developing early detection systems and approach has been well tested in applications such
minimizing the ecological impacts of biological as biogeography and conservation (Ortega-Huerta
invasion by alien plants. and Peterson 2004; Araújo et al. 2005) and in
In China, a total of 108 plant species have been assessing the effects of global climate change (Pet-
identified as alien weeds, of which 15 are distributed erson et al. 2002; Parra-Olea et al. 2005).
throughout most regions or the whole country (Qiang Incomplete sampling and smaller populations are
and Cao 2000). One of the most notable and noxious common problems in compiling or evaluating data on
invasive alien plants is crofton weed (Eupatorium the distribution of alien species, particularly in the
adenophorum Spreng.), a perennial herb native to case of recently established species that are not yet to
Central America. It naturally spread into southern reach their limits of successful growth and reproduc-
Yunnan province of China from Myanmar around the tion imposed by climate. Stockwell and Peterson
1940s (Xie et al. 2001). The invasion of large areas of (2002) explored the sample size needed for accurate
grassland and forest land by crofton weed has modeling via re-sampling of data for well-sampled
contributed to the decline in the numbers of indige- species. But random re-sampling methods ignore the
nous animals and plants and decrease in biodiversity temporal correlations that may exist within real
(Sun et al. 2004). Crofton weed is also a noxious plant distribution data.
known to cause acute asthma, diarrhea, depilation, We used a 50-year chronological data set of crofton
and even death of livestock (Wu et al. 2004). The weed covering a large area of China, to build a more
weed has invaded more than 30 countries including realistic scenario of typical early stages of crofton
China, India, Thailand, New Zealand, Australia and weed invasion, incorporated time series re-sampling
the United States, resulting in extinction of many as a historical simulation approach, modeled the
plants and severe economic losses (Fuller 1981; Wang fundamental niche of this species using the GARP
et al. 1997). The ecological and socio-economic approach, and analyzed the environmental parameters
implications of biological invasion have prompted the of known localities and within niche areas predicted
urgent call for better methods to predict the areas as likely to be invaded by the weed. Our objectives
where introduced species may become a threat were to (1) describe the spatiotemporal patterns of
(Rejmánek 2000; Welk et al. 2002; Richardson 2004). crofton weed invasion in China and (2) identify the
Many different modeling approaches have been regions vulnerable to invasion by crofton weed.
developed and used for predicting the likely distri-
bution of plant species based on climatic and edaphic
constraints (Segurado and Araújo 2004; Guisan and Methods
Thuiller 2005). One of the approaches that have
proven to be especially useful is the genetic algorithm Datasets
for rule-set production (GARP) modeling system
(Peterson and Cohoon 1999). The GARP modeling Ecological niche models are based on non-random
approach employs four inferential sub-routines, correlations between known occurrence of species
namely atomic, BIOCLIM, range rules, and logistic and environmental datasets that describe parameters

123
Landscape Ecol (2007) 22:1143–1154 1145

related to the dimensions of the ecological niche of edcdaac.usgs.gov/gtopo30/hydro/); soil data (i.e. soil
the species. To build models, environmental param- pH, soil moisture, and soil carbon) from the IGBP-
eters of localities of known occurrence are deter- DIS dataset (http://www.sage.wisc.edu/atlas/) and
mined and then used to identify geographic regions soil units from the Global Ecosystems Database
that have similar environmental parameters. Time (http://www.ngdc.noaa.gov/seg/ecosys/ecosys.shtml/
series data are usually obtained by periodically re- ); and land cover classification based on AVHRR
sampling the target objects. satellite imagery for 1992–1993 from University of
In our study, the occurrence of crofton weed was Maryland Global Land Cover Facility (http://
established for 441 localities using herbarium records glcf.umiacs.umd.edu/index.shtml/). The coverage
in local and national museums throughout China and was limited to 2–548N latitude and 73–1368E
from literature and field investigations. The localities longitude, which was re-sampled to 1-km grid
of known occurrence were then georeferenced using resolution. We used a jackknife manipulation and
the Electronic Sinomap (2004). The dataset for analysis and inspection of omission statistics to
predicting the distribution of the species consisted reduce the original sets of environmental layers to
of dated crofton weed specimens from 1940 to 2003. 17 ecological dimensions (Peterson and Cohoon
Because there had been a marked increase in 1999).
regionally clustered records as a result of systemat-
ically increased sampling efforts since 1990, we Niche modeling
removed 20 records for the period 1990–2003 to
compensate for any bias due to the increased The GARP method (http://www.lifemapper.org/
sampling. A total of 390 historical records were desktopgarp/) works as an iterative process of rule
selected and stratified, respectively in random and selection, evaluation, testing, and incorporation or
chronological order, into 39 sets of 10 records each as rejection to produce a heterogeneous rule-set
intrinsic training data in the modeling process. The describing the species’ ecological niche. The general
prediction analysis began with the first set of 10 approach is described in greater detail by Feria and
records, and successive sets of 10 records were added Peterson (2002). For each of the time-series dataset
for each of the consecutive runs. The sequential model runs, rules were generated for 1000 iterations
addition of chronologically stratified data sets, which and repeated 100 times using a differential data ratio
is a form of quasi-retrospective sampling, constitutes of 80:20 for training and rule-set validation. We
the temporal dimension of the dataset compared with selected the 10 best-subset models out of the 100
randomly arrayed datasets to examine the influence replicates based on optimal combinations of error
of invasion time on the model performance. An components (Anderson et al. 2003). The intersection
additional set of 51 records was used for evaluating of all the 10 best-subset models generated a final
the performance of GARP models as the extrinsic test map with values ranging from 0 to 1 (1 for regions
data. where all the models predicted niche presence; 0 for
In the modeling process, 23 eco-geographic data regions of niche absence). The resulting value
layers were used for the initial assessments of (hereafter referred to as overlap index, OI) is
ecological niche dimensions, including climate data interpreted as a measure of relative likelihood of
(i.e. annual total radiation, maximum, minimum, correctly predicting the niche presence for crofton
mean, extreme high and extreme low annual air weed.
temperatures, relative humidity, mean annual pre-
cipitation, precipitation in the warmest and the Test approach
coldest quarters, and precipitation in the wettest and
the driest months) for the period 1971–2000 from To test our model predictions, the extrinsic test data
Chinese Ecosystem Research Network dataset were overlaid on the mosaic of models and evaluated
(http://www.cern.ac.cn/0index/); topographic and through Kappa analysis (Cohen 1960) and receiver–
hydrologic data (i.e. elevation, slope, aspect, topo- operator characteristic (ROC) analysis (Hanley and
graphic index, flow accumulation, and flow direc- McNeil 1982) as more reliable measurements, (Pe-
tion) from the USGS’s Hydro-1K dataset (http:// arce and Ferrier 2000; Manel et al. 2001) advocated

123
1146 Landscape Ecol (2007) 22:1143–1154

by some researchers (Cumming 2000; Mcpherson To assess the prediction performance related to
et al. 2004). The Kappa statistic is a reliability metric increasing invasion time and data completeness, we
that uses the classification probabilities of the model tested the 10 best models using the original extrinsic
and of reality to compute the likelihood of agreement testing dataset segregated from the training datasets
by chance. In other words, the Kappa value is useful with random addition and chronological addition for
for testing the null hypothesis that there is no more the known localities of the species based on the two
agreement than might occur by chance alone. types of statistical analyses mentioned above.
Monserud and Leemans (1992) have suggested the
following ranges of agreement for the Kappa statistic: Environmental envelope analysis
poor, K < 0.4; moderate, 0.4 < K < 0.55; good,
0.55 < K < 0.85; excellent, 0.85 < K < 0.99; and The zonal statistics routine (ArcMap 8.3, Spatial
perfect, 0.99 < K < 1.00. Unfortunately, the calcu- Analyst; ESRI 2003) was used to extract from the
lation of Kappa values requires that probabilistic digital maps the values of 23 environmental variables
predictions of occurrence be divided into concrete in the localities where crofton weed is known to be
predictions of absence or presence, based on a single, present and in 1,000 random locations generated with
potentially arbitrary classification threshold. a random point generator in ArcMap 8.3 (ESRI
The ‘‘area under the curve’’ (AUC) of the ROC 2003). These values were used in the Redundancy
plot, instead, is a threshold-independent measure of Analysis (RDA; implemented in the software CA-
model accuracy, which juxtaposes correct and incor- NOCO 4.5; ter Braak and Šmilauer 2002) to order the
rect predictions over a range of thresholds. It ranges invasion probabilities and spreading dynamics of
from 0 to 1, with a value larger than 0.50 indicating a crofton weed along the given environmental gradi-
better than random-event performance (Fielding and ents. The significant environmental variables
Bell 1997). Given the various advantages and disad- (P < 0.05) were selected after a forward selection
vantages to using these different measures, we chose using an unrestricted Monte Carlo permutation test
not to use any one single measure to assess model based on 9,999 random permutations.
accuracy in our analyses.

Model evaluation Results

The model was evaluated using two criteria: the Spatiotemporal trends
influence of each environmental layer on the models’
prediction accuracy and the prediction performance The chronological development and spatial autocor-
related to increasing invasion time and data com- relation for the locations of occurrence of crofton
pleteness. weed within the invasion time series illustrated a
To assess the influence of the environmental general trend despite the occurrence of numerous new
layers, 10 repeated runs of GARP modeling, each records in the areas that were first invaded (Fig. 1).
consisting of 100 iterations, were performed in Crofton weed populations in China tended to be more
which each of the 23 layers was sequentially recent towards the northern and southern limits of
eliminated from the analyses, which resulted in 10 their distribution range. In the initial phase of the
prediction maps for each of the eliminations. These invasion, the range was limited to several isolated
maps were overlaid to produce a composite map, locations centered on the Jinghong region (Yunnan).
with each pixel having a value between 0 and 10 Subsequently, the weed continued to expand and
that indicated the number of replicate model runs established clusters of invasion foci. Over the last
predicting the presence of the weed. These maps 60 years, this weed has spanned 23.48 of latitude and
were further compared pixel by pixel with compos- 9.68 of longitude. The locations along Yibin (Sich-
ite maps generated by GARP for all the 23 layers uan) and Nandan-Hechi area (Guangxi) appeared to
following the same procedures. The level of agree- spearhead a more recent movement of the species by
ment between maps was analyzed by Kappa and providing nascent foci for further invasion to the
ROC statistics. north and the east.

123
Landscape Ecol (2007) 22:1143–1154 1147

the layers using the jackknifing procedure (Fig. 2).

The ROC analysis indicated that maximum mean
annual air temperature, precipitation in the coldest
quarter, and extreme low air temperature had the
most influence on the predictions whereas the Kappa
analysis identified only the latter two layers as the
most influential factors.
As can be seen in Fig. 3, the predictive accuracy
was related to sample size and especially to the
spatial distribution of input data points. The curves of
both Kappa and AUC values showed an increasing
trend with the random and positive case addition until
leveling off at the 7th run with 70 data points
(Fig. 3a) and did not reach a steady state until the
22nd run with 220 data points in the time series
(Fig. 3b) comprising 50 years of invasion history.

Predicted geographic distribution of crofton weed

in China
Fig. 1 Spatiotemporal trends in crofton weed (Eupatorium
adenophorum) invasion in south-west China. Geographical
The GARP analysis was based on a time series
distribution of training data was plotted on the bottom x–y axis. training set of 390 location records (Fig. 4). This
Invasion history of distribution points based on recording dates analysis predicted a much wider distribution than
was plotted on the vertical axis. The black lines represent the current occurred areas for crofton weed extending to
equal invasion time and the grey arrows indicate the direction
of spread. The boundaries of administrative regions are marked
192 counties and cities across seven provinces. Each
on the top plane of the 10 best-subset models from this analysis
(v2 = 217.99–331.56; P < 0.001) was imported into
ArcMap (ESRI 2003) and displayed in geographic
Tests of model performance space as a mosaic of model intersections. Given that
ecological niche models yield highly accurate pre-
Each of the more important environmental layers can dictions of the distribution of a species, the most
be identified by the lower agreement value of the likely areas for future invasion (i.e. OI = 1) include
statistics after the layer is excluded from the rest of significant portions of Yunnan, Guizhou, Guangxi,

1.0

0.8
Kappa / AUC

0.6

0.4

0.2

0.0
pcq elt dem pre asp dir slp mat cov pH rad cbn mst acc eht top hum

Fig. 2 Identification of the influence of each environmental elevation, pre: mean annual precipitation, asp: aspect, dir: flow
layer on the models’ prediction accuracy based on Kappa and direction, slp: slope, tmx: maximum annual air temperature,
AUC (area under curve) values calculated from jackknife cov: land cover, pH: soil pH, rad: annual total radiation, cbn:
experiments. Layers excluded from model building are listed soil carbon, mst: soil moisture, acc: flow accumulation, eht:
on the horizontal axis. The white and grey bars represent extreme high annual air temperature, top: topographic index,
Kappa and AUC values respectively. pcq: precipitation in the and hum: relative humidity
coldest quarter, elt: extreme low annual air temperature, dem:

123
1148 Landscape Ecol (2007) 22:1143–1154

a) 1.00 stricted Monte Carlo permutation test showed that the

selected environmental variables significantly ex-
.95 plained the total variance (P < 0.05) as well as the
variation along the first ordination axis (P < 0.05).
.90
Kappa / AUC

The first ordination axis is related mainly to extreme

.85 low, minimum, mean, and maximum annual air
temperatures whereas much of the variance of the
.80
second ordination axis is explained by precipitation
.75 AUC in the coldest quarter and mean annual precipitation.
kappa The higher OI values were associated with increased
.70 levels of moisture-related variables, especially mean
1 5 9 13 17 21 25 29 33 37
Model-Run (10 random records added each run)
annual precipitation and precipitation in the driest
month. It can be inferred that both temperature and
b) 1.00 precipitation play a pivotal role in the spread of
.90 crofton weed to northern and eastern parts of China
.80
following the decrease in the value of OI. Figure 5b
shows the relationships among the invasion dynamics
Kappa / AUC

.70
of crofton weed and selected environmental param-
.60 eters. In early stages of the invasion, most of the
.50 likely localities are in the highlands (average altitude
.40 of about 1,300 m) except on steeply sloped terrain
.30
(>158); the occurrence of crofton weed is positively
correlated with precipitation and temperature. In the
.20
1/10 5/20 9/30 13/40 17/50 21/50 25/50 29/60 33/60 37/60 later stages of the invasion, the occurrences of crofton
Model-Run (10 time-series records added each run) / Year of Invasion weed were correlated negatively with temperature
and precipitation but positively with geographic and
Fig. 3 Quality assessments by Kappa and AUC values of the
performance of the prediction model of modeling methods for
edaphic factors such as elevation, land cover, soil
under recorded or incompletely distributed items with random type, and soil pH. Over the past 20 years, crofton
addition (a) and chronological addition (b). The larger value in weed has spread from subtropical areas with higher
this curve indicates greater reliability of the outputs of the annual mean temperature and lower climatic fluctu-
corresponding model. The minimum sample size in this instance
is just over 70 and the invasion records of at least 50 years are
ations to much steeper, cooler, and drier areas at
needed to attempt reliable prediction on a large-scale higher elevations.

Hainan, Chongqing, Hubei, eastern Sichuan, south- Visualization of ecological niches

eastern boundary of Tibet, western Hunan, southern
Shanxi; the coastal zone of Guangdong, Fujian, A graphical display of the pattern of crofton weed
Taiwan, and Zhejiang; and isolated areas in Henan, distribution influenced by concrete, tangible ecolog-
Anhui, Jiangsu and Shandong. ical parameters provides fertile ground for insights
into the pattern of distribution. The known localities
Factors determining the invasion pattern of of crofton weed were well arranged (Fig. 6a) along
crofton weed across China the margin of predicted niches defined by the
bivariate environmental space of mean annual air
Figure 5a illustrates the relationships among the temperature and annual precipitation, with mean
probabilities for crofton weed invasion and selected annual air temperature ranging from 4.48C (Taoyu-
environmental parameters. The first two ordination anlang in Taiwan) to 23.18C (Taidong in Taiwan) and
axes explain 61.6% of the variation in distribution of annual precipitation from 698 mm (Yuanmou in
the species and 97.7% of the variation in the Yunnan) to 2254 mm (Jiangcheng in Yunnan).
relationship between the distribution of crofton weed The mean annual air temperature in the known
and environment parameters respectively. An unre- localities (16.88C; SD = 2.9) was above the average of

123
Landscape Ecol (2007) 22:1143–1154 1149

Fig. 4 Geographic predictions for crofton weed (Eupatorium testing and validations. The different colors represent regions
adenophorum) in China. Blue triangles represent the 390 spots where crofton weed presence was predicted with varying
used to train the predicted models and where the weed is degrees of certainty or likelihood (i.e. overlap index 1 means a
known to occur, and white circles represent 51 extrinsic testing higher possibility in that 10 out of 10 models predicted the
data used to test the predicted geographic range. To derive a presence; overlap index 0 means a lower possibility in that
robust result, all the available data (including extrinsic testing none of the 10 models did so)
data) were used as input data in the final modeling after model

that in randomly predicted locations (15.68C; the coldest quarter of the year than in those locations
SD = 3.2) and the differences were significant where the invasion was held to be unlikely (Fig. 6b).
(Mann-Whitney U = 68,284.5, P < 0.001), which Precipitation in the coldest quarter of the year for the
indicated a trend for the spread of crofton weed predicted localities was also higher than that in the
toward locations with lower temperatures, whereas known localities of crofton weed, suggesting that
the mean annual precipitation in the known localities crofton weed may be expected to spread further
(1292 mm; SD = 252) was not significantly different toward wetter regions.
from the average of that in randomly predicted
presences (1326 mm; SD = 340; Mann–Whitney
U = 84,019, P = 0.244). When temperature and Discussion
precipitation in the sub-regions in which the presence
of crofton weed was predicted were compared with Predictive power and practical consequences of
those in which it was unlikely to spread, it became the model
apparent that a broad combination of higher temper-
ature and greater precipitation favored crofton weed. Hierarchical modeling schemes of environmental
Within the regions of predicted presence, the majority control on the spread of species, such as those
of habitats had mean annual air temperatures between discussed by Wu and David (2002), may be essential
10 and 228C and mean annual precipitation between to the improvement of spatial modeling. In this study,
800 and 2,000 mm. however, we opted for the simpler niche modeling
The predicted locations for continued invasion of approach because our primary objective was not so
crofton weed are characterized by lower values of much to eradicate the weed from localities where it is
extreme low temperatures and greater precipitation in already established as to identify regions potentially

123
1150 Landscape Ecol (2007) 22:1143–1154

a) pcq a)
3000

Mean annual precipitation (mm)

0.8 Predictied presences
Known localities
2500 Predictied absences
pre
2000
0.7 mit
pwm eht mst hum tmp 1500
0.4 0.6 0.9 mat
0.3 0.5
0.1 pwq elt
dir 1000
soil cbn
0.2
top
c
aascp
slp pdm
rad cov
0 500
pH dem
1
0
-20 -10 0 10 20 30
-1.0 1.1 o
Mean annual temperature ( C)
b)

Precipitation of coldest quarter (mm)

b) 500
dem 10
rad 400
acc

cbn pH
slp 20
300
soil cov
dir eht
asp 200
mat
tmp top
50
30 60 2 100
40
hum
mit mst
pwm pwq elt 0
-60 -50 -40 -30 -20 -10 0 10 20
o
pre pcq Extreme low annual temperature ( C)

pdm Fig. 6 Ecological niches of crofton weed (Eupatorium

-0.8 0.6 adenophorum) in environmental space, . White circles indicate
known localities; black dots and gray triangles indicate
Fig. 5 (a) Redundancy analysis (RDA) ordination diagram modeled predicted presences and absences, respectively, in
(showing first and second ordination axes) displaying the the area of analysis
relation of crofton weed (Eupatorium adenophorum) invasion
probabilities (arrows with dashed lines show OI value ranging
from 0.1 to 1) and the correlation pattern of selected
environmental characteristics (arrows with thick lines) in
native range are often used as a constraint (Peterson
China. (b) RDA diagram displaying the relation of the et al. 2003; Iguchi et al. 2004). In this study, our
temporal dynamics by crofton weed invasion (arrows with predictions used the data on occurrence of the species
dashed lines show invasion years ranging from 2 to 60) and the in China. Although crofton weed has existed in China
correlation pattern of selected environmental characteristics
(arrows with thick lines): tmn: minimum annual air tempera-
for about 60 years, it continues to expand its’ range.
ture, tmp: mean annual air temperature, pwm: precipitation in Our datasets reflected the geographical non-equilib-
the wettest month, pdm: precipitation in the driest month, pwq: rium in the distribution of the species. However,
precipitation in the warmest quarter and soil: soil category. For given the long invasion history and diverse environ-
the explanations of the other abbreviations, see Fig. 2
mental conditions across the invaded regions, crofton
weed can be considered to have occupied extremes of
vulnerable to invasion and implementing appropriate ecological niches in China. For many plant species,
control measures. Our exercise proved effective in the current environmental conditions in their native
that it has achieved excellent predictive accuracy by ranges may not be necessarily consistent with the
identifying environmental parameters that influence ecophysiological characteristics of the species (Haw-
the potential spread of the species. kins and Sweet 1989; Sun and Sweet 1996). The
In predicting new regions where a given species is distribution of a plant species reflects to a large extent
likely to spread, the environmental conditions in the the influences of a suite of factors related to evolution

123
Landscape Ecol (2007) 22:1143–1154 1151

and geological history of habitats and may not always range using climate-matching models, invasiveness
fit the most suitable environmental conditions. and invasibility are mainly determined by additional
Papes et al. (2003) predicted the distribution of processes that act on much smaller spatial scale
crofton weed in China based on environmental (Williamson 1999, 2001).
conditions in the native habitats and suggested that
the species would occur mainly in eastern Yunnan,
and central and south-eastern China. Our results, Dispersal mode and invasion pattern of crofton
based on the locations in China where it has invaded weed
and become established, seemed to fit the known
localities better, which allowed for considering the Baker (1986) noted that invasion occurs through two
possibly of changed niches regulated by different modes: the steady advance and the stratified diffusion
biotic (e.g. competitive exclusion) and abiotic factors models. Crofton weed has been expanding from the
(e.g. habitat specificities) (Davis et al. 1998) than in south-west boundaries of China to the southern and
its’ natural range. Certainly, to achieve more reliable eastern provinces as scattered satellite populations
spatial predictions over a shorter time span, distribu- from the original center of introduction followed by
tional data in the native range should be incorporated infilling of the gaps. These long-distance hops are
into such models. characteristic of ‘‘stratified diffusion’’, which is the
Spatial models of species distribution are com- most frequent pattern of plant invasions such as that
monly subject to two types of errors: errors of observed in the post-glacial dispersal of oak trees in
omission (exclusion of regions inhabited) and errors Great Britain (Hengeveld 1989). The extent to which
of commission (inclusion of regions not actually species spread by stratified diffusion may influence
inhabited) (Fielding and Bell 1997). Insufficient the choice of control strategies. For example, control
sampling and ignoring the potential of non-equilib- measures that prevent the establishment of new foci
rium invasion would tend to increase the first type of or eliminate newly established foci are far more
error, leading to predicted values being lower than the effective than those in which efforts are concentrated
actual values. Additionally, failure to take into on invasion fronts already established (Moody and
account interactions among species (i.e., mutualism, Mack 1988). As demonstrated in this study, the
competition, and predation), long distance dispersal, establishment of new foci through jump dispersal
propagule pressure (Rouget and Richardson 2003), resulting from a heterogeneous landscape is of
anthropogenic disturbance, and other restricting fac- paramount importance in the spread of crofton weed.
tors could lead to a gap between real and predicted Identification of factors that affect patterns and
distributions resulting from both types of errors. processes of the spread of species and potentially
Approaches that take these processes into account enhance or hamper species invasion is expected to
will not only result in more accurate predictions but suggest ways for effectively managing invasive
also provide better insight into specific factors species (Sakai et al. 2001; Fagan et al. 2002). The
controlling the distribution of invasive plants. RDA analyses revealed that environmental charac-
Despite the possible sources of errors, statistical teristics are the key features determining the invasion
analysis of our results indicated that our niche model patterns of crofton weed and provided justification for
based on records where crofton weed has invaded using environmental data in niche visualization. The
China and a 50-year record of crofton weed invasion occurrence of crofton weed was found to be posi-
had a high degree of predictive power. It included tively correlated with precipitation and temperature
most known localities for crofton weed, suggesting and was more likely in regions with sufficient rainfall
low omission errors. Ninety-six percent of known and higher temperatures. The predicted northern limit
localities were predicted by all models (OI = 1). The is 33.58N corresponding to 0 8C isotherm in the
high predictive power of distribution ranges from coldest month in China. This temperature may
niche modeling is widely reported (Anderson 2003; represent a particular threshold for the weed to adapt
Raxworthy et al. 2003). Although the large-scale and to cold climates. Elevation and slope, despite having
long-term prediction of establishment can be based no direct biological effects on plants (Austin 2002),
on the characteristics of a species’ native and invaded can influence crofton weed invasion through their

123
1152 Landscape Ecol (2007) 22:1143–1154

correlation with temperature and precipitation. The At present, the invasion of crofton weed is so
preference of crofton weed to highlands largely extensive that eradication would be extremely labor-
reflects their range in their native habitats (Lu and intensive, and removal projects, although possible on
Ma 2004); most of their known localities occur at local scales, will have to be accompanied by long-term
flatter, mid-elevation sites where soil moisture is management efforts. An important means of conserv-
relatively high. Those known localities may also ing native biodiversity is to be vigilant in preventing
represent staging areas where the weed could slowly crofton weed from entering the vulnerable areas.
adapt to an environment of steeper slopes and higher Under those circumstances, long-term monitoring
elevations. systems to provide early detection and strict quaran-
The spread of crofton weed in China is far from tine measures to avoid anthropogenic spread of crofton
over. Human transformation of landscapes has facil- weed would be a high priority for the management of
itated the dispersal of a species that already has crofton weed invasion. Our predictions of potential
strong reproductive capacity and long-distance dis- distribution of crofton weed can provide a strong basis
persal capability (Wang et al. 1994). In south-western for identifying areas where detection efforts would be
China, crofton weed is often found along major river most effective and beneficial.
systems such as the Jinsha, Yalong, Hong and
Lancang, with the highest densities along river Acknowledgements This research was supported by the Key
Knowledge Innovation Project of the Chinese Academy of
valleys, in riparian areas, and on riverine beaches. Sciences (grant no. KSCX1-SW-13-03), the National Natural
Hence, rivers may have been a major channel of Science Foundation of China (contract No. 30470337) and a
spread for crofton weed. River corridors consist of an fellowship reward to W. Sang from the Royal Society, London.
array of landscape elements (Ward et al. 2002) with a We would like to thank Wenting XU, Mei YU, and Shibo
FANG for their kind help with GIS analyses, and M.
high frequency of open ground for colonization Williamson and G. Jetschke for constructive comments on an
(Malanson 1993), thus forming dispersal networks earlier version of the manuscript. We much appreciated the
connecting different landscapes (Forman and Godron excellent suggestions and comments of the three referees and
1986) and serving as a dispersal vector for alien coordinating editor.
invasive organisms (Renöfält et al. 2005). Crofton
weed is reported to have reached the Yangtze River
in 2003 and may continue to spread eastward along References
the river course if not contained.
Anderson RP (2003) Real vs. artefactual absences in species
distributions: tests for Oryzomys albigularis (Rodentia:
Remarks on invasion by crofton weed Muridae) in Venezuela. J Biogeogr 30:591–605
Anderson RP, Lew D, Peterson AT (2003) Evaluating predic-
Crofton weed, a species characteristic of warmer tive models of species’ distributions: criteria for selecting
regions with a tropical climate, appears to tolerate a optimal models. Ecol Model 162:211–232
Araújo MB, Thuiller W, Williams PH, Reginster I (2005)
wide range of environmental conditions within trop- Downscaling European species atlas distributions to a fi-
ical and subtropical regions. Our study suggested that ner resolution: implications for conservation planning.
Yungui Plateau, Sichuan Basin, south-eastern Coast- Glob Ecol Biogeogr 14:17–30
lands, Hainan Island, and Taiwan might be at high Austin MP (2002) Spatial prediction of species distribution: an
interface between ecological theory and statistical mod-
risk of invasion by this weed although currently the elling. Ecol Model 157:101–118
weed is either altogether absent in these regions or Baker HG (1986) Patterns of plant invasion in North America.
recorded only recently from isolated locations. In: Mooney HA, Drake JA (eds) Ecology of biological
Despite its already widespread occurrence, crofton invasions of North America and Hawaii. Springer-Verlag,
Berlin, Germany, pp 44–57
weed may expand further into regions susceptible to Cohen JA (1960) A coefficient of agreement for nominal
its invasion. Identification of such sites would help in scales. Educ Psychol Meas 20:37–46
formulating measures to contain its spread. One area Cumming GS (2000) Using between-model comparisons to
of particular concern is south-east China given the fine-tune linear models of species ranges. J Biogeogr
27:441–455
number of open ports, which makes it highly Davis AJ, Jenkinson LS, Lawton JH, Shorrocks B, Wood S,
susceptible to the introduction of alien weeds through (1998) Making mistakes when predicting shifts in species
frequent trade. range in response to global warming. Nature 391:783–786

123
Landscape Ecol (2007) 22:1143–1154 1153

Electronic Sinomap (2004) Sinomaps Press, Beijing, China Monserud RA, Leemans R (1992) Comparing global vegetation
ESRI (2003) Arcview GIS, version 8.3. Environmental Sys- maps with the Kappa statistic. Ecol Model 62:275–293
tems Research Institute Inc., Redlands, CA, USA Moody ME, Mack RN (1988) Controlling the spread of plant
Fagan WF, Lewis MA, Neubert MG, van den Driessche P invasions: the importance of nascent foci. J Appl Ecol
(2002) Invasion theory and biological control. Ecol Lett 25:1009–1021
5:148–157 Normile D (2004) Expanding trade with China creates eco-
Forman RTT, Godron M (1986) Landscape ecology. Wiley, logical backlash. Science 306:968–969
New York, USA Ortega-Huerta MA, Peterson AT (2004) Modelling spatial
Feria ATP, Peterson AT (2002) Prediction of bird community patterns of biodiversity for conservation prioritization in
composition based on point-occurrence data and inferen- North-eastern Mexico. Divers Distribut 10:39–54
tial algorithms: a valuable tool in biodiversity assess- Papes M, Peterson AT (2003) Predicting the potential invasive
ments. Divers Distribut 8:49–56 distribution for Eupatorium adenophorum Spreng. in
Fielding AH, Bell JF (1997) A review of methods for the China. J Wuhan Bot Res 21:137–142
assessment of prediction errors in conservation presence/ Parra JL, Graham CC, Freile JF (2004) Evaluating alternative
absence models. Environ Conserv 24:38–49 data sets for ecological niche models of birds in the An-
Fuller TC (1981) Introduction and spread of Eupatorium ade- des. Ecography 27:350–360
nophorum in California. In: Vebkata RB (ed) Proceedings Parra-Olea G, Martı́nez-Meyer E, Pérez-Ponce de León G
of the 8th Asian-Pacific Weed Science Society confer- (2005) Forecasting climate change effects on salamander
ence. The Organizing Committee eighth conference of distribution in the highlands of Central Mexico. Biotro-
Asian-Pacific Weed Society, Bangalore, India, pp 277– pica 37:202–208
280 Pearce J, Ferrier S (2000) Evaluating the predictive perfor-
Guisan A, Thuiller W (2005) Predicting species distribution: mance of habitat models developed using logistic
offering more than simple habitat models. Ecol Lett regression. Ecol Model 133:225–245
8:993–1009 Peterson AT, Cohoon KP (1999) Sensitivity of distributional
Hanley JA, McNeil VJ (1982) The meaning and use of the area prediction algorithms to geographic data completeness.
under a receiver operating characteristic (ROC) curve. Ecol Model 117:159–164
Radiology 143:29–36 Peterson AT, Vieglais DA (2001) Predicting species invasions
Hawkins BJ, Sweet GB (1989) Evolutionary interpretation of using ecological niche modeling: new approaches from
high temperature growth response in five New Zealand bioinformatics attack a pressing problem. BioScience
forest tree species. New Zeal J Bot 27:101–107 51:363–371
Hengeveld R (1989) Dynamics of biological invasions. Chap- Peterson AT, Ortega-Huerta MA, Bartley J, Sánchez-Cordero
man & Hall, London, UK, 160 pp V, Soberón J, Buddemeier RH, Stockwell DRB (2002)
Iguchi K, Matsuura K, McNyset K, Peterson AT, Scachetti- Future projections for Mexican faunas under global cli-
Pereira R, Powers KA, Vieglais DA, Wiley EO, Yodo T mate change scenarios. Nature 416:626–629
(2004) Predicting invasions of North American basses in Peterson AT, Papes M, Kluza DA (2003) Predicting the po-
Japan using native range data and a genetic algorithm. tential invasive distributions of four alien plant species in
Trans Am Fish Soc 133:845–854 North America. Weed Sci 51:863–868
Knowler D, Barbier E (2005) Importing exotic plants and the Qiang S, Cao XZ (2000) Survey and analysis of exotic weeds
risk of invasion: are market-based instruments adequate? in China. J Plant Resour Environ 9:34–38
Ecol Econ 52:341–354 Raxworthy CJ, Martı́nez-Meyer E, Horning N, Nussbaum RA,
Lim BK, Peterson AT, Engstrom MD (2002) Robustness of Schneider GE, Ortega-Huerta MA, Peterson AT (2003)
ecological niche modeling algorithms for mammals in Predicting distributions of known and unknown reptile
Guyana. Biodivers Conserv 11:1237–1246 species in Madagascar. Nature 426:837–840
Lu ZJ, Ma KP (2004) The influence of topographical factors on Reichard SH, White P (2003) Invasion biology: an emerging
the invasion of the alien species, Eupatorium adenopho- field of study. Ann MO Bot Gard 90:64–66
rum. Acta Phytoecol Sin 28:761–767 Rejmánek M (2000) Invasive plants: approaches and predic-
Malanson GP (1993) Riparian landscapes. Cambridge Uni- tions. Austral Ecol 25:497–506
versity Press, Cambridge, UK Renöfält BM, Jansson R, Nilsson C (2005) Spatial patterns of
Manel S, Williams HC, Ormerod SJ (2001) Evaluating pres- plant invasiveness in a riparian corridor. Landscape Ecol
ence-absence models in ecology: the need to account for 20:165–176
prevalence. J Appl Ecol 38:921–931 Richardson DM (2004) Plant invasion ecology dispatches from
McNeely JA (1999) The great reshuffling: how alien species the front line. Divers Distribut 10:315–319
help feed the global economy. In: Sandlund OT, Schei PJ, Rouget M, Richardson DM (2003) Inferring process from
Viken A (eds) Invasive species and biodiversity man- pattern in alien plant invasions: a semi-mechanistic model
agement. Kluwer Academic Publishers, Dordrecht, the incorporating propagule pressure and environmental fac-
Netherlands, pp 11–31 tors. Am Nat 164:713–724
Mcpherson JM, Jetz W, Rogers DJ (2004) The effects of Sakai AK, Allendorf FW, Holt JS, Lodge DM, Molofsky J,
species’ range sizes on the accuracy of distribution With KA, Baughman S, Cabin RJ, Cohen JE, Ellstrand
models: ecological phenomenon or statistical artefact? J NC et al (2001) The population biology of invasive spe-
Appl Ecol 41:811–823 cies. Annu Rev Ecol Syst 32:305–332

123
1154 Landscape Ecol (2007) 22:1143–1154

Segurado P, Araújo MB (2004) An evaluation of methods for Wang HJ, He P, Ma JL (1994) An investigation and research
modelling species distributions. J Biogeogr 31:1555–1568 report on the dissemination of Ageratina adenophora on
Stockwell DRB, Noble IR (1992) Introduction of sets of rules rangeland areas in Liangshan District of Sichuan Prov-
from animal distribution data: a robust and informative ince. Grassland China 1:62–64
method of data analysis. Math Comput Simulat 33:385– Ward JV, Tockner K, Arscott DB, Claret C (2002) Riverine
390 landscape diversity. Freshwater Biol 47:517–539
Stockwell DRB, Peters D (1999) The GARP modelling system: Welk E, Schubert K, Hoffmann MH (2002) Present and po-
problems and solutions to automated spatial prediction. tential distribution of invasive garlic mustard (Alliaria
Int J Geogr Inf Sci 13:143–158 petiolata) in North America. Divers Distribut 8:219–233
Stockwell DRB, Peterson AT (2002) Effects of sample size on Wilcove DS, Rothstein D, Dubow J, Phillips A, Losos E (1998)
accuracy of species distribution models. Ecol Model Quantifying threats to imperiled species in the United
148:1–13 States: assessing the relative importance of habitat
Sun OJ, Sweet GB (1996) Genotypic variation in light and destruction, alien species, pollution, overexploitation, and
temperature response of photosynthesis in Nothofagus disease. Bioscience 48:607–615
solandri var. cliffortioides and N. menziesii. Aust J Plant Williamson M (1999) Invasions. Ecography 22:1–8
Physiol 23:421–428 Williamson M (2001) Can the impact of invasive plants be
Sun XY, Lu ZH, Sang WG (2004) Review on studies of predicted? In: Brundu G, Brock J, Camarda I, Child L,
Eupatorium adenophorum—an important invasive species Wades M (eds) Plant invasions, species ecology and
in China. J Forest Res 15:319–322 ecosystem management. Backhuys Publishers, Leiden, pp
ter Braak CJF, Šmilauer P (2002) CANOCO reference manual 11–20
and CanoDraw for windows user’s guide: software for With KA (2002) The landscape ecology of invasive spread.
canonical community ordination (version 4.5). Micro- Conserv Biol 16:1192–1203
computer Power, Ithaca, NY, USA, 500 pp Wu J, David JL (2002) A spatially explicit hierarchical ap-
Turpie JK, Heydenrych BJ, Lamberth SJ (2003) Economic proach to modeling complex ecological systems: theory
value of terrestrial and marine biodiversity in the Cape and applications. Ecol Model 153:7–26
Floristic region: implications for defining effective and Wu ZH, Qin GL, Deng TJ (2004) Invasion and establishment
socially optimal conservation strategies. Biol Conserv of Eupatorium adenphorum (Spreng.) and its risk ap-
112:233–251 praisal in some areas of Guangxi. Southwest China J
Wang F, Summerell BA, Marshall DR, Auld BA (1997) Inoc- Agric Sci 17:469–471
ulum production of Phaeoramularia, a potential mycoher- Xie Y, Li ZY, Gregg WP, Li DM (2001) Invasive species in
bicide for crofton weed. Aust Plant Pathol 26:173–178 China - an overview. Biodivers Conser 10:1317–1341

123