全 文 :武汉植物学研究 2003, 21 (2) : 137~ 142
J ourna l of W uhan B otan ica l Resea rch
紫茎泽兰 E up a tor ium adenop horum Spreng.
在中国入侵分布预测
M on ica Papes, A. Tow n send Peterson
(N atu ra l H istory M useum and B iod iversity R esearch Cen ter, T he U n iversity of K ansas, L aw rence, Kansas 66045, U SA )
摘 要: 原产于墨西哥的紫茎泽兰 E up atorium ad enop horum Sp reng. 作为一个有害的外来物种在印度、新西兰和
澳大利亚生长已有很长时间。在中国, 尤其是在南方和西南地区其蔓延速度之快, 带来了不可忽视的经济和社会后
果。我们采用了生态位模拟新方法来预测紫茎泽兰可能入侵的范围。据此, 预测的潜在分布区包括该植物在中国境
内已分布的省份及未来华中、华东易受入侵的区域。
关键词: 生态位模拟; GA R P; 紫茎泽兰; 入侵物种
中图分类号: Q 949. 783. 5 文献标识码: A 文章编号: 10002470X (2003) 0220137206
Pred icting the Poten tia l Invasive D istr ibution for
E up a tor ium adenop horum Spreng. in Ch ina
M on ica Papes, A. Tow n send Peterson
(N atu ra l H istory M useum and B iod iversity R esearch Cen ter, T he U n iversity of K ansas, L aw rence, Kansas 66045, U SA )
Abstract: O rig inat ing in M ex ico , E up a torium ad enop horum Sp reng. , is a nox iou s w eed long p re2
sen t as a non2nat ive species in India, N ew Zealand, and A u stra lia. In Ch ina, it is sp reading
rap id ly, part icu larly in the sou thern and sou thw estern parts of the coun try, w ith seriou s econom 2
ic and socia l con sequences. A new m ethod, eco logica l n iche modeling, w as app lied fo r p redict ing
its po ten t ia l geograph ic range of invasion. P redicted po ten t ia l d ist ribu t ional areas included the
Ch inese p rovinces w here the p lan t is know n to occu r, as w ell as addit ional areas in cen tra l and
eastern Ch ina that appear su scep t ib le to fu rther sp read of th is species.
Key words: Eco logica l n iche modeling; GA R P; E up a torium ad enop horum ; Invasive species
C rofton w eed ( E up a torium ad enop horum
Sp reng. ) is nat ive to M ex ico , w here it is a com 2
mon w eedy sh rub. Its in t roduct ion in India, N ew
Zealand, and A u stra lia had seriou s econom ic con2
sequences fo r local agricu ltu re [1 ]. It w as in troduced
in Yunnan P rovince, Ch ina, around 1940. It has
sp read rap id ly th roughou t sou thw estern Ch ina,
and p resen t ly can be found in Gu izhou, X izang (T i2
bet) , Guangx i, Yunnan, and Sichuan p rovinces[2 ].
Becau se it invades m ain ly grasslands, econom ic im 2 p lica t ion s are sign if ican t. T h is p lan t m ay cau sech ron ic pu lmonary disease in ho rses, o r even deathin livestock [1 ]. Such con sequences led u s to invest i2gate fu rther the po ten t ia l d ist ribu t ion of th isspecies in Ch ina.Peterson and V iegla is[3 ] p resen ted a m ethodo2logy fo r app lica t ion of eco logica l n iche modelingtechn iques in p redict ion of the geograph ic po ten t ia lof species’ invasion s. T he essence of the app roachis developm en t of an eco logica l n iche model basedΞ Received date: 2002206205, A ccep ted date: 2002212209.B iograph ies:M onica Papes(1976- ) , fem ale, Graduate Research A ssistan t, B iodiversity research; A ndrew Tow nsend Peterson
(1964- ) , m ale, A ssociate P rofesso r and Curato r, B iogeography and system atics.
on the eco logica l characterist ics of know n occu r2
rences on the nat ive dist ribu t ion of a species. T h is
techn ique, o rig inally developed fo r p redict ing
species’ d ist ribu t ion s on nat ive geograph ic dist ribu2
t ional areas[4, 5 ] , w as modif ied [3 ] to include p ro jec2
t ion of the n iche model to po ten t ia lly invaded re2
gion s to p redict geograph ic po ten t ia l on the invaded
dist ribu t ion.
W e app lied th is m ethodo logy to recon struct
the nat ive dist ribu t ion of E up a torium in M ex ico
and to p redict the invasive range of th is species in
Ch ina. W e then compared ou r resu lts w ith the
availab le info rm at ion concern ing the nat ive and in2
vasive dist ribu t ion of th is p lan t.
1 M ethods
Eco logica l n iche models w ere developed based
on geo referenced occu rrence po in ts taken from di2
verse sou rces, including herbarium specim en
reco rds, scien t if ic litera tu re including flo ras and
system at ic t rea tm en ts, etc. In a ll, 45 occu rrence
po in ts w ere ob ta ined w ith in the nat ive range of the
species. T he seem ingly low num bers of occu rrence
po in ts ob ta ined after mon th s of search ing litera tu re
and con tact ing m u seum cu rato rs emphasizes the
crit ica l need fo r compu teriza t ion of natu ra l h isto ry
m u seum ho ld ings, and in tegra t ion of such data
sets via the In ternet to p rovide eff icien t access to
b iodiversity data.
Eco logica l n iches w ere modeled u sing the
Genet ic A lgo rithm fo r R u le2set P redict ion
(GA R P) [4 6 ]. In genera l, the p rocedu re focu ses on
modeling eco logica l n iches ( the con junct ion of eco2
log ica l condit ion s w ith in w h ich a species is ab le to
m ain ta in popu la t ion s w ithou t imm igra t ion ) [7 ].
Specif ica lly, GA R P rela tes the eco logica l charac2
terist ics of know n occu rrence po in ts to tho se of
po in ts random ly samp led from the rest of the study
region, seek ing to develop a series of decision ru les
that best summ arize tho se facto rs associa ted w ith
the species’ p resence.
O ccu rrence po in ts are d ivided tw ice even ly in2
to t ra in ing and test data sets—— tha t is, an in it ia l
50% of the data po in ts are set aside fo r a comp lete2
ly independen t test of model quality (ex trin sic test
data ) ; of the rem ain ing po in ts, 50% are u sed fo r
develop ing models ( t ra in ing data ) and 50% are
u sed fo r tests of model quality in ternal to GA R P
( in trin sic test data). GA R P w o rk s in an itera t ive
p rocess of ru le select ion, evaluat ion, test ing, and
inco rpo ra t ion o r reject ion: a m ethod is cho sen from
a set of po ssib ilit ies ( e. g. , log ist ic regression,
b ioclim at ic ru les ) , app lied to the tra in ing data,
and a ru le is developed o r evo lved. P redict ive accu2
racy is then evaluated based on 1 250 po in ts resam 2
p led from the in trin sic test data and 1 250 po in ts
samp led random ly from the study region as a
w ho le. R u les m ay evo lve by a num ber of m ean s
that m im ic DNA evo lu t ion: po in t m u ta t ion s, dele2
t ion s, cro ssing over, etc. T he change in p redict ive
accu racy from one itera t ion to the nex t, m easu red
via the in trin sic data, is u sed to evaluate w hether a
part icu lar ru le shou ld be inco rpo ra ted in to the
model, and the algo rithm run s either 1 000 itera2
t ion s o r un t il convergence. Ξ
A ll modeling in th is study w as carried ou t on a
desk top imp lem en ta t ion of GA R P now availab le
fo r pub lic, free dow n load 1). T h is imp lem en ta t ion
offers m uch2imp roved flex ib ility in cho ice of p re2
dict ive environm en ta löeco logica l G IS (Geograph ic
Info rm at ion System s) data coverages: in th is case,
in it ia lly, w e u sed 12 data layers summ arizing ele2
vat ion; slope; aspect (a ll from the U. S. Geo logi2
cal Su rvey’s 2) H ydro21K data set) ; aspects of cli2
m ate including diu rnal tempera tu re range; fro st
days; m ean annual p recip ita t ion; so lar rad ia t ion;
m ax im um , m in im um , and m ean annual tempera2
tu res; vapo r p ressu re; and w et days (annual m ean s
196121990; from the In tergovernm en ta l Panel on
C lim ate Change 3). T he area of analysis included
M ex ico , w here E up a torium is nat ive, and su r2
831 武 汉 植 物 学 研 究 第 21 卷
Ξ 1) h ttp: ööbeta. lifem apper. o rgödesk topgarpö.2) h ttp: ööedcdaac. usgs. govögtopo30öhydroö.3) h ttpööipcc. chö.
rounding areas. GA R P’s p redict ive ab ilit ies have
been tested under d iverse circum stances[3, 8 15 ].
To reduce environm en ta l coverage sets to ju st
tho se coverages that p rovide h ighest p redict ive ac2
cu racy, w e u sed a variety of jackkn ife m an ipu la2
t ion s. In genera l, w e ran m u lt ip le itera t ion s ( 12
20) of models om it t ing each coverage, o r su ites of
coverages, system at ica lly. W e then exam ined the
co rrela t ion betw een inclu sion o r exclu sion of each
coverage (coded as 1s and 0s) and om ission erro r
(percen t of ex trin sic test data no t p redicted as p re2
sen t). Po sit ive co rrela t ion s w ere con sidered indica2
t ive of detrim en ta l con tribu t ion of a part icu lar cov2
erage to model quality, and tho se coverages w ere
removed from fu rther analysis. It is impo rtan t to
no te that the jackkn ife m an ipu la t ion s w ere done
so lely on the nat ive dist ribu t ion, and so do no t de2
t ract from the independen t natu re of the invaded2
range p redict ion s p resen ted herein.
To op t im ize model perfo rm ance, w e developed
100 rep lica te models of species’ eco logica l n iche
based on random 50- 50 sp lits of availab le occu r2
rence po in ts. U n like p reviou s app lica t ion s, w h ich
either u sed single models to p redict species’ d ist ri2
bu t ion s[12, 14 ] o r summ ed m u lt ip le models to inco r2
po rate model2to2model varia t ion [3 ] , w e u sed a new
p rocedu re (A nderson et a l. subm it ted ) fo r choo2
sing best sub sets of models. T he p rocedu re is
based on the ob servat ion s that (1) models vary in
quality, ( 2) varia t ion among models invo lves an
inverse rela t ion sh ip betw een erro rs of om ission
( leaving ou t t rue dist ribu t ional area) and comm is2
sion ( including areas no t actually inhab ited) , and
(3) best models (as judged by experts b lind to er2
ro r sta t ist ics in the o rig inal derivat ion of the
m ethodo logy) are clu stered in a region of m in im um
om ission of independen t test po in ts and modera te
area p redicted (an ax is rela ted direct ly to comm is2
sion erro r). T he rela t ive po sit ion of the cloud of
po in ts rela t ive to the tw o erro r axes p rovides an as2
sessm en t of the rela t ive accu racy of each model.
H ence, to choo se best sub sets of models, w e (1)
elim inated all models that had non2zero om ission
erro r based on independen t test po in ts, (2) ca lcu2
la ted the average area p redicted p resen t among
these zero2om ission po in ts, and ( 3 ) iden t if ied
models that w ere w ith in 1% of the overa ll avera2
ge. O u t of 342 models w ith no om ission erro r w e
summ ed the 14 models w h ich confined to the 1%
th resho ld. M odel quality w as tested via the ex trin2
sic test data: a ς 2 test w as u sed to compare ob2
served success in p redict ing the dist ribu t ion of test
po in ts w ith that expected under a random model
(p ropo rt ional area p redicted p resen t p rovides an
est im ate of occu rrence po in ts co rrect ly p redicted
w ere the p redict ion to be random w ith respect to
the dist ribu t ion of the test po in ts). P ro ject ion of
the ru le2sets fo r these models on to m ap s of sou2
thern and eastern A sia p rovided p redict ion s of po2
ten t ia lly invaded dist ribu t ion s.
2 Results
T he jackkn ife m an ipu la t ion s iden t if ied cover2
ages that detracted from the p redict ive ab ilit ies of
the algo rithm. T he first round of jackkn ife m an i2
pu la t ion s detected 6 coverages po sit ively co rrela ted
w ith om ission erro r (ground fro st frequency, r =
0102; annual m ean m in im um temperatu re, r =
0105; annual m ean temperatu re, r= 0111; annual
m ean m ax im um temperatu re, r= 0103; vapo r p res2
su re, r= 0101; and aspect, r= 0102). T hese cove2
rages w ere elim inated from fu rther analysis. T he
rem ain ing six layers (d iu rnal tempera tu re range,
elevat ion, slope, p recip ita t ion, so lar rad ia t ion,
and w et day frequency) w ere exam ined in jackkn ife
m an ipu la t ion s in w h ich all po ssib le com b inat ion s of
the 6 coverages w ere tested. T he best p redict ion of
the nat ive dist ribu t ion of E up a torium ( i. e. , zero
om ission ) w as ob ta ined u sing fou r layers: d iu rnal
tempera tu re range, w et days, elevat ion, and
slope.
A ll sub sets models w ere h igh ly sta t ist ica lly
sign if ican t. T he ς 2 tests based on the independen t
ex trin sic test data sets, w h ich included 22 po in ts of
know n occu rrence, a ll ind ica ted p redict ive ab ility
far bet ter than random models (a ll P ≤ 2. 14 ×
10- 20). H ence, a ll best sub sets models w ere h igh ly
p redict ive fo r the nat ive dist ribu t ion, and, fo r that
931 第 2 期 M onica Papes 等: 紫茎泽兰 E up atorium ad enop horum Sp reng. 在中国入侵分布预测 (英)
reason, w e p roceeded to exp lo re their p redict ion s
fo r invaded dist ribu t ional areas in Ch ina.
Fou r of the five p rovinces w here E up a torium
is know n to occu r in Ch ina w ere p redicted as h igh ly
su itab le in ou r p redict ion of the invasive dist ribu2
t ion (F ig. 1). O ne addit ional p rovince from w h ich
the species is know n , X izang ( T ibet ) , w as
F ig11 M odeled p redict ions fo r E up atorium ad enop horum on its native geograph ic distribu tion (top ) and
p ro jection to po ten tia lly invaded range in eastern A sia (bo ttom ). Increasingly dark shades of gray indicate
greater confidence in p redict ions of p resence (w h ite2zero models p redicted p resen t, and b lack214 models p redicted
p resen t; w h ite circles2po in ts of native occurrence; w h ite triangle2M exico C ity; balck triangle2Beijing)
041 武 汉 植 物 学 研 究 第 21 卷
m arginally included in ou r p redict ion s. How ever,
in each case, the lack of mo re p recise d ist ribu t ional
data from w ith in Ch ina lim its ou r ab ilit ies to assess
the p redict ive natu re of ou r models. GA R P models
indica ted addit ional po ten t ia l areas of invasion,
m ain ly in cen tra l and eastern Ch ina.
3 D iscussion
O ne of the mo st impo rtan t th rea ts to nat ive
comm un it ies is rep resen ted by the invasive exo t ic
species[16 ]. T hey often disp lace nat ive dom inan ts,
a ltering comm un ity funct ion and compo sit ion [17 ] ,
and can have seriou s econom ic con sequences. P re2
dict ing the geograph ic po ten t ia l species’ invasion s
u sing eco logica l n iche modeling techn iques seem s
to be the mo st su itab le app roach to th is p rob2
lem [3 ]. It can be app lied no m atter w here the geo2
graph ica l reg ion of in terest is, as long as data on
the nat ive dist ribu t ion of the species are availab le.
O u r p redict ion concern ing the po ssib le areas of
invasion of E up a torium in Ch ina p rovinces w here
th is w eed is know n to be included the fou r estab2
lished [2 ]. N at ive to M ex ico and Cen tra l Am erica,
th is p lan t species is clearly no t yet in equ ilib rum
dist ribu t ionally, and fo r that reason does no t yet
inhab it it s en t ire po ten t ia l d ist ribu t ional area in
A sia. How ever, lack of p recise occu rrence data fo r
the p lan t in Ch ina lim ited ou r ab ility to test the ac2
cu racy of ou r models. Indeed, such w as the mo ti2
vat ion fo r pub lica t ion of th is paper: w e p resen t ou r
model p redict ion s in the hope that they st im u la te
the p resen ta t ion of p recise d ist ribu t ional info rm a2
t ion in the fu tu re.
T h is study detected o ther po ten t ia l areas of in2
vasion in Ch ina. T he eastern part of the coun try
w as in genera l p redicted as mo re su itab le fo r inva2
sion than the w estern part. M o re o r less con t inu2
ou s w ith the areas cu rren t ly occup ied, to the
no rth, are six p rovinces that appear to be direct ly
vu lnerab le: Gan su, N ingx ia, Shaanx i, Shanx i,
H enan, and H ubei. O u r model a lso detected tw o
po ten t ia l d ist ribu t ional areas that are d isjunct from
the p resen t d ist ribu t ion, in the no rtheast and
sou theast, in w h ich th is species cou ld likely estab2
lish if in t roduced. T hese po ten t ia l d ist ribu t ional
areas include L iaon ing and H eilongjiang p rovinces
in the no rtheast and Fu jian and Zhejiang p rovinces
in the sou theast.
Acknowledgmen ts: D ata w ere k indly p rovided by Jam es
C. So lomon from M issouri Bo tan ical Garden, M arjo rie
Know les from the Sm ith son ian Inst itu te, and by the R ed
M undial de Info rm ación sobre B iodiversidad (R EM IB 4) ) fo r
data from the H erbario del Inst itu to de Eco logía, A. C.
Guo jun Chen k indly assisted w ith transla t ion of crit ical b ib2
liograph ic references and review of the Ch inese abstract.
T h is study w as funded by the U. S. N ational Science Foun2
dation and the U. S. Environm ental P ro tection A gency. Ξ
References:
[ 1 ] O elrich s P B, Calanasan C A , M acL eod J K, et a l.
Iso la t ion of a compound from E up atorium ad enop ho2
rum (Sp reng. ) (A g era tina ad enop hora (Sp reng ) )
causing hepato tox icity in m ice [J ]. N atu ra l2T ox ins,
1995, 3: 350 354.
[ 2 ] Q iang S. T he h isto ry and sta tus of the study on
crofton w eed (E up atorium ad enop horum Sp reng. ) ; a
w o rst w o rldw ide w eed[J ]. J W uhan B ot R es, 1998,
16: 366 373.
[ 3 ] Peterson A T , V ieglais D A. P redict ing species inva2
sions using eco logical n iche modeling [ J ]. B io2
S cience, 2001, 51: 363 371.
[ 4 ] Stockw ell D R B , N oble I R. Induction of sets of
ru les from anim al distribu tion data: A robust and in2
fo rm ative m ethod of analysis [ J ]. M ath Com p u t
S im u l, 1992, 33: 385 390.
[ 5 ] Stockw ell D R B , Peters D P. T he GA R P modelling
system : P rob lem s and so lu tions to au tom ated spatia l
p redict ion [J ]. In t J Geog Inf orm S y st, 1999, 13:
143 158.
[ 6 ] Stockw ell D R B. Genetic algo rithm s Ê [A ]. In:
F ielding A H ed. M ach ine L earn ing M ethods fo r
Eco logical A pp licat ions[M ]. Bo ston: K luw er A cade2
m ic Publishers, 1999. 123 144.
[ 7 ] Grinnell J. F ield tests of theo ries concern ing distri2
butional con tro l [J ]. A m er N at, 1917, 51: 115
128.
[ 8 ] Peterson A T , Cohoon K C. Sensit ivity of distribu2
t ional p redict ion algo rithm s to geograph ic data com 2
141 第 2 期 M onica Papes 等: 紫茎泽兰 E up atorium ad enop horum Sp reng. 在中国入侵分布预测 (英)
Ξ 4) h ttp: ööwww. conab io. gob. m xörem ibödocto sörem ib2esp. h tm l.
p leteness[J ]. E col M od el, 1999, 117: 159 164.
[ 9 ] Peterson A T , Sóberon J , Sánchez2Co rdero V. Con2
servatism of eco logical n iches in evo lu tionary tim e
[J ]. S cience, 1999, 285: 1 265 1 267.
[ 10 ] Godow n M E, Peterson A T. P relim inary distribu2
t ional analysis of U. S. endangered b ird species [J ].
B iod ivers Conserv , 2000, 9: 1 313 1 322.
[ 11 ] Peterson A T , Egbert S L , Sánchez2Co rdero V , et
a l. Geograph ic analysis of conservation p rio rit ies
using distribu tional modelling and comp lem entarity:
Endem ic b irds and m amm als in V eracruz, M exico
[J ]. B iolog Conserv , 2000, 93: 85 94.
[ 12 ] Peterson A T. P redict ing species’ geograph ic distri2
butions based on eco logical n iche modeling [ J ].
Cond or, 2001, 103: 599 605.
[ 13 ] Peterson A. T , Sánchez2Co rdero V , Sóberon J , et a l. Effects of global clim ate change on geograph icdistribu tions of M exican C racidae [J ]. E colog M o2d el, 2001a.[ 14 ] Peterson A T , Ball L G, Cohoon K C. P redict ingdistribu tions of trop ical b irds [J ]. Ibis, 2002, 144:e27 e32.[ 15 ] Stockw ell D R B , Peterson A T. Effects of samp lesize on accuracy of species distribu tion models [J ].E colog M od el, 2002, 148: 1 13.[ 16 ] M ooney H A , D rake J A. B io logical invasions ofN o rth Am erica and H aw aii [ M ]. N ew Yo rk:Sp ringer2V erlag, 1986.[ 17 ] V itousek P M , W alker L R , W h itaker L D , et a l.B io logical invasion by M y rica f ay a alters eco systemdevelopm ent in H aw aii [ J ]. S cience, 1987, 238:802 804.241 武 汉 植 物 学 研 究 第 21 卷