de B uin, Sophie e al.
A icle — Published Ve sion
Scena io p ojec ions o Sou h Asian mig a ion pa e ns
amids en i onmen al and socioeconomic change
Global En i onmen al Change
P o ided in Coope a ion wi h:
Leibniz Ins i u e o Ag icul u al De elopmen in T ansi ion Economies (IAMO), Halle (Saale)
Sugges ed Ci a ion: de B uin, Sophie e al. (2024) : Scena io p ojec ions o Sou h Asian mig a ion
pa e ns amids en i onmen al and socioeconomic change, Global En i onmen al Change, ISSN
1872-9495, Else ie , Ams e dam, Vol. 88, pp. 1-12,
h ps://doi.o g/10.1016/j.gloen cha.2024.102920 ,
h ps://www.sciencedi ec .com/science/a icle/pii/S0959378024001249
This Ve sion is a ailable a :
h ps://hdl.handle.ne /10419/302109
S anda d-Nu zungsbedingungen:
Die Dokumen e au EconS o dü en zu eigenen wissenscha lichen
Zwecken und zum P i a geb auch gespeiche und kopie we den.
Sie dü en die Dokumen e nich ü ö en liche ode komme zielle
Zwecke e iel äl igen, ö en lich auss ellen, ö en lich zugänglich
machen, e eiben ode ande wei ig nu zen.
So e n die Ve asse die Dokumen e un e Open-Con en -Lizenzen
(insbesonde e CC-Lizenzen) zu Ve ügung ges ell haben soll en,
gel en abweichend on diesen Nu zungsbedingungen die in de do
genann en Lizenz gewäh en Nu zungs ech e.
Te ms o use:
Documen s in EconS o may be sa ed and copied o you pe sonal
and schola ly pu poses.
You a e no o copy documen s o public o comme cial pu poses, o
exhibi he documen s publicly, o make hem publicly a ailable on he
in e ne , o o dis ibu e o o he wise use he documen s in public.
I he documen s ha e been made a ailable unde an Open Con en
Licence (especially C ea i e Commons Licences), you may exe cise
u he usage igh s as speci ied in he indica ed licence.
h p://c ea i ecommons.o g/licenses/by/4.0/
Scena io p ojec ions o Sou h Asian mig a ion pa e ns amids
en i onmen al and socioeconomic change
Sophie de B uin
a,*
, Jannis Hoch
b
, Jens de B uijn
a,c
, Ka hleen He mans
d
, Amina Maha jan
e
,
Ma i Kummu
, Jaspe an Vlie
a
a
Ins i u e o En i onmen al S udies, VU Uni e si y, De Boelelaan 1087, 1081HV Ams e dam, he Ne he lands
b
Fa hom, B is ol, UK
c
In e na ional Ins i u e o Applied Sys ems Analysis (IIASA), Laxenbu g, Aus ia
d
Leibniz Ins i u e o Ag icul u al De elopmen in T ansi ion Economies (IAMO), Theodo -Liese -S . 2, 06120 Halle (Saale), Ge many
e
In e na ional Cen e o In eg a ed Moun ain De elopmen (ICIMOD), Ka hmandu, Nepal
Wa e and De elopmen Resea chakno G oup, Aal o Uni e si y, Finland
ARTICLE INFO
Keywo ds:
En i onmen al change
Clima e change
Mig a ion
Sou h Asia
Machine-lea ning
Scena io p ojec ions
ABSTRACT
P ojec ing mig a ion is challenging, due o he con ex -speci ic and discon inuous ela ions be ween mig a ion
and he socioeconomic and en i onmen al condi ions ha d i e his p ocess. He e, we in es iga e he use ulness
o Machine Lea ning (ML) Random Fo es (RF) models o de elop h ee ne mig a ion scena ios in Sou h Asia by
2050 based on his o ical pa e ns (2001–2019). The model o he di ec ion o ne mig a ion eaches an accu acy
o 75%, while he model o he magni ude o mig a ion in pe cen age eaches an R
2
alue o 0.44. The a iable
impo ance is simila o bo h models: empe a u e and buil -up land a e o p ima y impo ance o explaining
ne mig a ion, aligning wi h p e ious esea ch. In all scena ios we ind ho spo s o in-mig a ion No h-wes e n
India and ho spo s o ou -mig a ion in eas e n and no he n India, pa s o Nepal and S i Lanka, bu wi h dis-
pa i ies ac oss scena ios in o he a eas. These dispa i ies unde sco e he challenge o ob aining consis en esul s
om di e en app oaches, which complica es d awing i m conclusions abou u u e mig a ion ajec o ies. We
a gue ha he applica ion o mul i-model app oaches is a use ul a enue o p ojec u u e mig a ion dynamics,
and o gain insigh s in o he unce ain y and ange o plausible ou comes o hese p ocesses.
1. In oduc ion
Human mig a ion is in insically ela ed o socie al change and
de elopmen . People lea e hei place o o igin o mul iple easons,
including be e economic o educa ional oppo uni ies elsewhe e,
amily ma e s o escaping con lic o pe secu ion (IOM, 2022).
Inc easingly, en i onmen al change is unde s ood o shape mig a ion
pa e ns ia impac s on ag icul u e and he habi abili y o egions
(Adge e al., 2015; Ho on e al., 2021). Ye , he iden i ica ion o causal
linkages be ween en i onmen al change and mig a ion is challenging o
e en impossible, due o he in e sec ing and con ex -speci ic na u e o
ac o s a ec ing mig a ion decisions (Boas e al., 2019; Ca aneo e al.,
2019; He mans and McLeman, 2021). Ye , unde s anding mig a ion
dynamics in an e a o clima e change is bo h scien i ically pe inen and
socie ally ele an , especially in a wo ld whe e iews and oices in
na ional and global mig a ion deba es a e o en ideologically d i en
(Boas e al., 2019; de Haas, 2010a).
The consequences o en i onmen al change, especially clima e
a iabili y impac s, on mobili y and immobili y ha e been in es iga ed
ex ensi ely in he pas wo decades (McLeman e al., 2021; Pigue ,
2022). Speci ically, he e is a g owing ecogni ion ha clima e- ela ed
mig a ion is mul i-causal and con ex -speci ic (Adge e al., 2024; Ca -
aneo e al., 2019; Hun e and Simon, 2023). Empi ical s udies based on
in e iew, su ey, o census da a ha e p o ided an o e iew o common
d i e s (A i i, 2011; He mans-Neumann e al., 2017; an de Gees ,
2011). E idence s ongly sugges s ha en i onmen al condi ions ha e a
g ea e impac on mig a ion wi hin coun ies as compa ed o in e na-
ional mig a ion (Ciss´
e e al., 2022; Cundill e al., 2021). Howe e , while
some quan i a i e esea ch on he impo ance o en i onmen al condi-
ions o in e na ional mig a ion dynamics is a ailable, (Abel e al., 2019;
Beine and Pa sons, 2015; Falco e al., 2019; G eceque e al., 2017), his
ype o esea ch wi hin coun ies is la gely absen .
* Co esponding au ho .
E-mail add ess: [email p o ec ed] (S. de B uin).
Con en s lis s a ailable a ScienceDi ec
Global En i onmen al Change
jou nal homepage: www.else ie .com/loca e/gloen cha
h ps://doi.o g/10.1016/j.gloen cha.2024.102920
Recei ed 23 Feb ua y 2024; Recei ed in e ised o m 14 Augus 2024; Accep ed 27 Augus 2024
Global En i onmen al Change 88 (2024) 102920
A ailable online 2 Sep embe 2024
0959-3780/© 2024 The Au ho (s). Published by Else ie L d. This is an open access a icle unde he CC BY license ( h p://c ea i ecommons.o g/licenses/by/4.0/ ).
The lack o subna ional longi udinal mig a ion da a has limi ed
p og ess on la ge-scale quan i a i e subna ional mig a ion esea ch
(Pigue , 2022). In line wi h his, li le esea ch has been published on
de eloping quan i a i e mul i-coun y mig a ion scena io p ojec ions
on a subna ional le el ha encompass he po en ial impac o en i on-
men al change (Beye e al., 2023; McLeman, 2013; Oakes e al., 2023).
De eloping hese scena ios is complica ed by se e al ac o s. The e a e
mul iple di ec ions o in luence o clima e change impac s on mig a ion.
Impac s such as p olonged d ough s and ising empe a u es can bo h
educe and inc ease mig a ion, depending on he speci ic socioeconomic
condi ions o he egion and he expe iences and esou ces o hose
being a ec ed (Dallmann and Millock, 2017; Muelle e al., 2020;
Muelle e al., 2014). Addi ionally, he a ailabili y o subna ional so-
cioeconomic scena io p ojec ion da a a e limi ed, especially wi hin he
amewo k o clima e change (Buhaug and Ves by, 2019). Some agen -
based and in eg a ed modelling s udies exis o u u e human mig a-
ion which ocus on indi idual coun ies (Thobe e al., 2018). These
s udies use na ional demog aphic su ey da a, educing hei compa-
abili y ac oss coun ies. Fu he mo e, he wo G oundswell epo s
(Clemen e al., 2021; Rigaud e al., 2018) and he A ican Shi s epo
(Amak ane, 2023) de elop a ious scena ios by deploying a g a i y
model wi h and wi hou hyd oclima ic and ag icul u al a iables o
Sub-Saha an A ica, Sou h Asia, and La in Ame ica. By showing he
di e ence be ween he scena io wi h and wi hou he hyd oclima ic and
ag icul u al a iables, he s udies ob ain he numbe o in e nal mi-
g an s ha could be a ibu ed o hyd oclima ic condi ions. G a i y
models ace c i icism when used o de eloping scena io p ojec ions o
mig a ion, because hey a e unable o adequa ely accoun o he
changes in-mig a ion pa e ns o e ime (Beye e al., 2022). Fu he -
mo e, g a i y models canno handle he discon inues impac s o d i e s
(Robinson and Dilkina, 2018).
The main objec i e o his a icle is o be e unde s and he use-
ulness o a new app oach o de eloping mig a ion scena ios in he
con ex o clima e change: Machine Lea ning (ML) Random Fo es (RF)
models. RF app oaches can combine mul iple inpu da a and include
discon inuous ela ions (Robinson and Dilkina, 2018), making hem
po en ially well-sui ed o unde s anding and modelling mig a ion.
Howe e , RF app oaches ha e no been used be o e o de elop mig a-
ion scena ios. To u he explo e he po en ial o RF app oaches, we
employ wo di e en models: a RF classi ica ion model and a RF
eg ession model. The o me p ojec s he di ec ion o ne mig a ion pe
egion, while he la e p ojec s he magni ude o he ne mig a ion. We
es hese RF app oaches in o Sou h Asia. To do so, we i s ain bo h
models o explain his o ical ne mig a ion pa e ns based on known
d i e s o mig a ion, using a no el high- esolu ion la ge-scale mig a-
ion da ase (Ni a e al., 2023a). We in e p e he esul o his aining
as he capabili y o ou models o explain ne mig a ion pa e ns, which
p o ides an indica ion o hei capaci y o also explain u u e pa e ns.
We p ojec ne mig a ion in he yea 2050 unde di e en socioeco-
nomic and en i onmen al change scena ios using bo h ained models.
Resul s a e analysed mainly in he con ex o he usabili y o he wo RF
app oaches o en isioning mig a ion scena ios. By doing so, we can
disce n he insigh s hey o e , acili a ing a mo e obus e alua ion and
in e p e a ion o app oaches o de eloping mig a ion scena ios.
2. Da a & me hods
2.1. Case s udy egion
We ocus on Sou h Asia, including Bangladesh, Bhu an, India, Nepal,
Pakis an, and S i Lanka, ollowing he delinea ion o his egion by he
Wo ld Bank. Wo ld Bank egions ep esen ela i ely homogenous so-
cioeconomic egions, making hem app op ia e o assessing mig a ion
dynamics and in e p e ing model esul s wi h he same model. Sou h
Asia was selec ed o wo easons. The egion was selec ed o wo
easons. Fi s , he egion is cha ac e ised by a high dependency on
clima e-sensi i e li elihoods, mos ly in ag icul u e (Tucke e al., 2015).
Second, he dea h and bi h a es a e well documen ed on he subna-
ional le el, compa ed o la ge pa s o he Middle Eas egion and he
A ican con inen (Ni a e al., 2023a). Sou h Asian ne mig a ion da a
can he e o e be ega ded as mo e accu a e han o he egions wi h a
high dependency on clima e-sensi i e li elihoods. We excluded
A ghanis an ex-an e om he analysis. In his coun y, 3.9 million e -
ugees we e epa ia ed o A ghanis an be ween 2002 and 2015 (UNHCR,
2022). O e all, 5.3 million people e u ned o A ghanis an since 2002
un il he Taliban ook o e con ol in 2021 (UNHCR, 2023). These
excep ional condi ions make he coun y unsui able o assessing gen-
e al mig a ion dynamics d i en by socioeconomic and en i onmen al
changes.
As spa ial uni s we employed he Global Adminis a i e A eas
(GADM) le el 2 (Global Adminis a i e A eas, 2022), which ep esen s
sub-di isions o p o inces o dis ic s. Le el 2 a eas which we e smalle
han 100 km
2
o inhabi ed by less han 500 people in he yea 2001 we e
me ged wi h he adjacen a ea in he same p o ince wi h which i sha ed
he longes bo de . This a oids a po en ially la ge in luence o small
popula ion numbe s on he model. Table 1 p o ides he demog aphic
cha ac e is ics o hese egions pe coun y.
2.2. Da a desc ip ion
2.2.1. Ta ge da a
We used he global ne mig a ion da ase de eloped by Ni a e al.
(2023a) as a ge da a, which is based on annual ha monised subna-
ional da a o bi hs and dea hs o he 2001–2019 pe iod. This sub-
na ional da a was downscaled o a 5 a c-minu e esolu ion globally and
hen combined wi h obse ed changes in popula ion numbe s
(Wo ldPop, 2021). Ne mig a ion is de ined as he di e ence be ween
he o al popula ion change and he na u al popula ion change (bi hs
minus dea hs) pe yea , e lec ing he ne numbe o people mo ing in o
o ou o a gi en a ea. Simula ed da a was alida ed agains epo ed
de ailed ne mig a ion da a o a ious coun ies, indica ing a good
pe o mance o he downscaling me hod (Ni a e al., 2023a). The inal
da ase p o ides in o ma ion on he ne numbe o people mo ing in o
ou o a g id cell pe yea . Following his app oach, mig a ion en ails
bo h na ional and in e na ional mig a ion, and he wo canno be
sepa a ed. Consequen ly, he case s udy egion o his s udy is no a
closed sys em, and he sum o all mig a ion can be highe o lowe han
0.
We p e-p ocessed he a ge mig a ion da a o acili a e he se up o
ou RF models. Fo he Random o es eg ession (RFR), he ne pe -
cen age o he popula ion pe adminis a i e a ea ha mig a es was
calcula ed by di iding he o al numbe o ne mig an s by he o al
popula ion o ha a ea. We used he Wo ldPop popula ion da a
(Wo ldPop, 2021) as used by Ni a e al. (2023a). Fo he Random o es
classi ica ion (RFC) model, an a ea was classi ied as 0 when he a ea
aced ne ou -mig a ion in a pa icula yea , and as 1 when he a ea
aced ne in-mig a ion in a pa icula yea . In se en a eas we obse ed
Table 1
Demog aphic cha ac e is ics o he coun ies in he s udy egion.
Coun y Popula ion
2015
X 1 000
Numbe
o admin
le el 2
a eas
included
A e age popula ion*/admin le el 2
a ea
X 1 000
Bhu an 740 133 6
Bangladesh 152 513 65 2 346
India 1 284 631 664 1 935
Nepal 26 813 14 1 915
Pakis an 193 696 33 5 870
S i Lanka 19 607 227 86
* Popula ion numbe s as used by Ni a e al. (2023b) omWo ldPop (2021).
S. de B uin e al.
Global En i onmen al Change 88 (2024) 102920
2
ne mig a ion alues exceeding 20 %, which a e likely e o s ollowing
he skewed pa e n (e.g., one yea acing ne ou -mig a ion o 20–40 %
ollowed by ne in-mig a ion in a ollowing yea o 20–40 %). The e o e,
we decided o emo e hese egions om ou analysis. Fig. 1A p esen s
he a e age yea ly ne mig a ion a io (%) o he e e ence pe iod. 1B
p esen s he a e age o ne in (1) and ne ou (0) mig a ion pe yea o
he en i e e e ence pe iod 2001–2019.
2.2.2. Independen a iables
We used a combina ion o socioeconomic and en i onmen al a i-
ables as he independen a iables which a e u he de ailed in Table 2.
The a iables we e selec ed based on h ee c i e ia. Fi s , we selec ed
a iables ha a e po en ially ele an o explaining mig a ion acco d-
ing o p e ious esea ch. Second, we selec ed hose a iables o which
consis en his o ical and u u e da a o he whole s udy egion is
a ailable. Thi d, we analysed mul i-collinea i y, and emo ed a iables
ha a e oo s ongly co ela ed wi h each o he (Chan e al., 2022).
Al hough RF-models can handle (mul i)collinea i y in e ms o p edic-
ion, ha ing highly co ela ed p edic o s can a ec he in e p e abili y
o he a iable impo ance. I migh assign ela i ely lowe impo ance
o one o he co ela ed a iables compa ed o wha i would i he
a iables we e no highly co ela ed (Chan e al., 2022). Fo mo e de ails
on (mul i)collinea i y be ween he selec ed a iables and cu -o alues,
see Fig. S1 and Table S1 on (mul i)collinea i y in he Supplemen a y
Ma e ial.
The socioeconomic a iables included a e buil -up land, con lic ,
educa ion, G oss Na ional Income (GNI) pe capi a and economic
inequali y (GINI). Buil -up land was included as a p oxy o he le el o
u banici y and he access o se ices, condi ions gene ally a ac ing
mig an s (Selod and Shilpi, 2021). Con lic was included since i can
lead o o ced displacemen (B ai hwai e e al., 2019), al hough i is no
clea o wha ex en con lic e en s a e a ac o o impo ance o
mig a ion o e la ge egions and long- ime spans. Educa ion le els
a ec mig a ion ollowing he oppo uni ies ha come wi h being
Fig. 1. (A) The a e age mig a ion a io o he e e ence pe iod 2001–2019. (B)
The a e age o ne in (1) and ne ou (0) mig a ion o he e e ence
pe iod 2001–2019.
Table 2
Cha ac e is ics o he socioeconomic and en i onmen al annual a iables.
Va iable B ie
desc ip ion
Uni Sou ce O iginal
esolu ion
Buil -up a ea The a e age
buil -up a ea.
Km
2
/cell Wol e al.
(2018)
9.25 ×9.25 km
Con lic A med
con lic
e en s wi h
o e 10
dea hs.
Numbe o
e en s
UCDP
Geo e e enced
E en Da ase
V23.1 (Da ies
e al., 2023;
Sundbe g and
Melande ,
2013)
Geo e e enced
e en s
Educa ion A e age o al
yea s o
schooling pe
pe son.
Yea s o
schooling
Smi s and
Pe manye
(2019),
ha monised
and as e ised
using me hods
by Kummu e
al (2018).
Admin 1 le el,
excep o S i
Lanka, which is
coun y-based.
GNI pe cap G oss
domes ic
income pe
capi a.
US$2017 Smi s and
Pe manye
(2019)
ha monised
and as e ised
using me hods
by Kummu e
al (2018) and
u he
downscaled o
admin 2 le el.
5 a cminu es
Inequali y
(GINI)
Income
inequali y
be ween
people,
ollowing he
Gini index.
0–1 index Sol (2020)
downscaled
using
subna ional
GINI da a
Admin 1 le el,
excep o S i
Lanka, which is
coun y-based.
D y spell Numbe o
days in he
longes
pe iod
wi hou
signi ican
p ecipi a ion
o a leas 1
mm.
Numbe o
days
Wo ld Bank
G oup (2023)
0.5x0.5
deg ees
Flu ial lood
Volume
Volume o
he la ges
looding
e en in an
a ea.
M
3
Su anudjaja
e al. (2018)
5 a cminu es
P ecipi a ion To al
p ecipi a ion.
Millime es Wo ld Bank
G oup (2023)
0.5x0.5
deg ees
Tempe a u e A e age
empe a u e.
Deg ee C◦IMAGE model (
Doelman e al.,
2018; S eh es
e al., 2014),
2023 upda e.
0.5x0.5
deg ees
Ac ual c op
yield
The ac ual
simula ed
o al c op
yield.
Ton/km2 IMAGE model (
Doelman e al.,
2018; S eh es
e al., 2014),
2023 upda e.
5 a cminu es
S. de B uin e al.
Global En i onmen al Change 88 (2024) 102920
3
educa ed (de Haas, 2010b; Neumann and He mans, 2017). Income
le els, in his s udy GNI pe capi a, a e an impo an equi emen o
mig a ion (Neumann and He mans, 2017; Ni a e al., 2021). The ole o
economic inequali y in-mig a ion dynamics is equi ocal. Howe e ,
ea lie esea ch has showed ha inequali y is o en high in as -g owing
ci ies wha would imply ha i could be a ac o explaining u al–u ban
mig a ion (Øs by, 2016).
The en i onmen al a iables included in his s udy a e he leng h o
d y spells, lood olume, p ecipi a ion, empe a u e, and ac ual c op
yields. D y spells ha e been included since hese ha e been associa ed
wi h ou -mig a ion (Ca ico and Dona o, 2019). Flood olume was
included because loods can a ec ag icul u al yields, as well as habi -
abili y o an a ea a ec ing mig a ion dynamics (Ho on e al., 2021;
Muelle e al., 2014). P ecipi a ion and empe a u e ha e been ound o
a ec mig a ion mainly ia ag icul u al p oduc i i y, al hough o en o a
limi ed ex en (Boh a-Mish a e al., 2014; Ca aneo and Pe i, 2016;
Muelle e al., 2014). Finally, we included ac ual c op yields o accoun
o he la ge-scale ends o p oduc i i y o an a ea. P oduc i e a eas
migh indica e ha hese a eas a e a ac i e o u al- u al mig an s
(Ha hie e al., 2015). A he same ime, people li ing in p oduc i e a eas
migh ha e he inancial means equi ed o mig a e (G o h e al., 2020).
F om mos a iables, we calcula ed he yea ly a ying a e age alue
pe adminis a i e a ea. Fo con lic , he o al numbe o con lic e en s
pe yea pe a ea was used. Fo looding, he maximum annual lu ial
lood olume pe a ea was aken, ep esen ing he la ges lood e en o
ha yea ollowing exceeding i e discha ge ollowing excess ain all
a e accoun ing o soil in il a ion and e apo anspi a ion. Fo all uns,
a one-yea ime lag was applied o all a iables excep o con lic
e en s and lood olume, assuming ha hose igge mig a ion di ec ly.
Fo mo e con ex ega ding he socioeconomic and en i onmen al a -
iables ega ding he minimal, maximal, a e age alues and s anda d
de ia ion, see Table S2 in he Supplemen a y Ma e ial.
2.2.3. P ojec ion da a
Th ee SSP-RCP combina ions we e employed o e lec a ange o
socioeconomic and clima e de elopmen s: SSP1 wi h RCP2.6, SSP2 wi h
RCP4.5, and SSP3 wi h RCP7.0. See Table 3 o a b ie desc ip ion o
hese scena ios. The SSP-RCP scena io p ojec ion da a a e consis en
wi h he his o ical da a. Fo he en i onmen al da a, CMIP6-based ISI-
MIP-3 p o ocol p ojec ions we e used o RCP2.6, RCP4.5 and RCP7.0,
excep o lood olume since his da a was no a ailable in CMIP6.
The e o e, we use CMIP5-based ISIMIP-3 p o ocol da a o RCP2.6,
RCP4.5 and RCP6.0 o lood olume. Al hough CMIP5 and CMIP6 a e
di e en in e ms o a iables, esolu ion, and scena ios, he phases a e
consis en in e ms o simula ion p o ocols and model e alua ion s an-
da ds (Tebaldi e al., 2021). I is he e o e jus i ied o use da a om bo h
phases in one analysis. Fo mo e scena io de ails pe a iable see
Table S3 in he Supplemen a y Ma e ial. Con lic e en s we e no
included in any scena io because p ojec ions o con lic isk a e no
a ailable o he egion. While p ojec ions o con lic a e lacking, we
inco po a e his o ical con lic e en s none heless, o gain insigh s in o
he impo ance o hese e en s in shaping he mig a ion pa e ns.
2.3. Se -up o he Random Fo es models
To e alua e he impo ance o he socioeconomic and en i onmen al
a iables o he magni ude and di ec ion o mig a ion, we employed an
RFR model and an RFC model. Fo ou analysis we used he exis ing
CoP o amewo k, which is designed o apply machine-lea ning ap-
p oaches o p ojec ions, used be o e o de elop con lic isk p ojec ions
(modi ied om Hoch e al., 2021a). Wi hin he CoP o amewo k, he
Sciki -lea n lib a y is used o implemen he RFR and he RFC algo i hms
(Ped egosa e al., 2011). Bo h algo i hms a e ained wi h 20 yea s o
da a (2001–2019) o quan i y he his o ical ela ion be ween he inde-
penden a iables and ne mig a ion, he a ge da a.
To explo e he explana o y powe o he wo app oaches, we i s
ained he RFR and RFC o each coun y sepa a ely. Fo ou inal
p ojec ions in he en i e egion, we included only hose coun ies whe e
his o ical mig a ion could be explained by he model. Mo eo e , by i s
assessing he indi idual coun ies, we could also examine o wha ex en
he a iable impo ance was compa able be ween he indi idual coun-
ies o he wo app oaches.
Fo p ojec ing u u e mig a ion pa e ns, we agg ega ed hose
coun ies whe e mig a ion could be (pa ly) explained by he RFR and
RFC app oach. This agg ega ion excludes Bhu an and Pakis an (see e-
sul s and discussion). The RFR and he RFC we e subsequen ly ained
o e his combined s udy egion o de elop he ou -o -sample scena io
p ojec ions.
Fo each yea in he e e ence pe iod, alues we e ex ac ed om he
independen a iables o each adminis a i e a ea. In o al, his yields
da a poin s equal o he numbe o adminis a i e a eas imes he
numbe o yea s. Subsequen ly, o bo h he eg ession and he classi-
ica ion un, 100 RF ees we e ini ialised o cap u e he a iance in he
da a wi hou o e i ing. Fo each ee, 70 % o he da a poin s we e
andomly d awn o ain he model, and he emaining 30 % we e used
o alida ion using se e al e alua ion me ics (Skici -lea n, 2023; Ting,
2011):
- Accu acy: e lec s he ac ion o co ec classi ica ions [0–1] −
highe ep esen s mo e co ec classi ica ions.
- Recall: p esen s he o al numbe o ue posi i es ou o he o al
numbe o posi i es [0–1] – highe ep esen s mo e co ec
classi ica ions.
- P ecision: he numbe o ue posi i es ou o he o al numbe o
posi i es p edic ed, including he alse posi i es [0–1] – highe
ep esen s mo e co ec classi ica ions.
- The ROC AUC (Recei e Ope a ing Cha ac e is ic −A ea Unde he
Cu e): his alue quan i ies he o e all abili y o a bina y classi i-
ca ion model o dis inguish be ween he posi i e and nega i e classes
[0 −1] – highe ep esen s mo e co ec classi ica ions.
- R
2
: Fo he eg essions uns he R2 is p esen ed. This is a s a is ic ha
measu es he p opo ion o he a iance in he a ge a iable
explained by he independen a iables in he model. I p o ides a
Table 3
Summa y o he scena io na a i e o he SSP-RCP combina ions used in his
s udy based on O’Neill e al. (2016) and K iegle e al. (2012).
SSP-RCP Scena io desc ip ion
SSP1 – RCP2.6 Sus ainable
De elopmen ¡Low
Emissions
This scena io en isions a u u e whe e he
wo ld ollows a sus ainable de elopmen
pa hway wi h low g eenhouse gas emissions. I
assumes ha socie y places a s ong emphasis
on en i onmen al sus ainabili y, ene gy
e iciency, and he educ ion o ca bon
emissions. Challenges o mi iga ion and
adap a ion a e low. Unde RCP 2.6, o al
adia i e o cing inc eases o 3.0 W m −2 un il
mid-cen u y be o e a decline begins. The goal
is o limi global wa ming o well below 2
deg ees Celsius abo e p e-indus ial le els.
SSP2 – RCP4.5 Middle o he
Road
This scena io ep esen s a middle-o - he- oad
de elopmen pa hway, wi h mode a e
g eenhouse gas emissions wi hou
undamen al b eak h oughs. I an icipa es a
wo ld whe e e o s o mi iga e clima e change
a e mode a e, wi h a ocus on balancing
economic g ow h and en i onmen al
conce ns.
SSP3 – RCP7.0 Regional Ri al y
¡High Emissions
In his scena io, a u u e is en isioned whe e
he e is a lack o global coope a ion, leading o
egional i al ies and agmen a ion o e o s.
G eenhouse gas emissions a e ela i ely high,
esul ing in subs an ial global wa ming. This
pa hway highligh s he po en ial consequences
o limi ed in e na ional collabo a ion.
S. de B uin e al.
Global En i onmen al Change 88 (2024) 102920
4
alue be ween 0 and 1, a highe R2 indica es a be e i o he model
o he da a.
We calcula ed he a iable impo ance based on he same aining
uns o bo h models. Va iable impo ance e e s o he ela i e impo -
ance o each inpu a iable in explaining mig a ion, he a ge a iable.
In o al, he a iable impo ance o he combined independen a iables
is 1. Quan i ying he impo ance o he sepa a e a iables o he model
p edic ions enhances ou unde s anding o he impo ance o he so-
cioeconomic and en i onmen al condi ions o ne mig a ion. Al hough
a iable impo ance p o ides he magni ude o in luence i does no
p o ide he di ec ion o in luence, since he di ec ion can go bo h ways
wi hin he same model, depending on he o he a iables. This p ope y
is bo h a s eng h and a weakness o RF models since i makes hem
lexible bu also ha de o in e p e .
F om he end o he e e ence pe iod (2019) un il 2050, we made
annual ou -o -sample p ojec ions. The p ojec ions a e made o he
combined egion o coun ies whe e a sha e o he his o ical mig a ion
could be explained in bo h he RFR and he RFC. To main ain he in-
e nal consis ency o each p ojec ion pa hway, his was done o each
selec ed SSP-RCP combina ion. Fo he RFR, he p ojec ions ep esen
he a e age ne in o ou -mig a ion a io in 2050 o he 100 ees in he
model. Fo he RFC, he p ojec ions ep esen he a e age p obabili y o
ne in-mig a ion in 2050 o he 100 ees in he model. Fo bo h he RFR
and RFC we compa ed he 2050 alues wi h he a e age ne mig a ion
alue pe egion o he e e ence pe iod o indica e whe e u u e
mig a ion migh de ia e om his o ical pa e ns. The las s ep in ol ed
a compa ison o he scena io p ojec ions based on he RFR and he RFC
app oach o check o consis ency. This was done by compa ing he
p ojec ed ne in- and ne ou -mig a ion. Fo he RFC app oach, ne in-
mig a ion was de ined as a p obabili y o ne in-mig a ion o o e 0.5,
while o he RFR ne in-mig a ion was de ined as a posi i e mig a ion
pe cen age.
3. Resul s
3.1. Model alida ion
Fo he RFR un, which aims o p edic he annual ne mig a ion as a
pe cen age o he popula ion, he R
2
a ies ma kedly be ween he
coun ies. Fo India, he explained a iance is low wi h 0.17. The R
2
o
Bangladesh, Nepal and S i Lanka a e ela i ely high, espec i ely 0.59,
0.91 and 0.53. Fo Bhu an and Pakis an, mig a ion could no be
explained, as indica ed by he nega i e R
2
alues in Table 4. These
coun ies a e he e o e excluded om he uns o make scena io p o-
jec ions o he combined egion (see Discussion o u he in e p e a-
ion o hese esul s). The explained a iance o he combined egion,
Bangladesh, India, Nepal and S i Lanka is 0.44.
Fo he RFC un, which aims o p edic bina y in (1) o ou (0)
mig a ion, he o e all model pe o mance is good, as indica ed by ROC-
AUC sco es o abo e 0.8, excep o Bhu an (Table 4). Fo Bhu an,
his o ical mig a ion canno be explained wi h he RFC app oach. O e all
accu acy o he o he coun ies— he ac ion o co ec classi-
ica ions—is easonable o good. Mean p ecision − he abili y o he RFC
no o label an obse a ion as in-mig a ion ha is ou -mig a ion – is also
easonable o good. The ecall sco es a e lowe o all a eas, indica ing a
es ic ed abili y o he classi ie o ind all posi i e obse a ions.
Mig a ion in he combined egion o Bangladesh, India, Nepal and S i
Lanka can be explained ela i ely well, as indica ed by an o e all ac-
cu acy o 0.75.
3.2. P edic o s o ne mig a ion
The a iable impo ance in he RFR and RFC un pe coun y a e
a he compa able (Fig. 2). The a e age annual empe a u e is he mos
impo an a iable in bo h analyses. Buil -up land, yields and annual
p ecipi a ion a es also ha e a ela i ely high impo ance ac oss he
coun ies and o he combined egion in bo h uns. Con lic is no
impo an o explain ne mig a ion in he e e ence pe iod. The impo -
ance o he a iables in he wo models is ai ly compa able be ween
coun ies o many, hough no all, a iables. Fo he a iable impo -
ance pe coun y see Fig. S2 in he Supplemen a y Ma e ial. Especially
in Nepal he e a e majo di e ences in he impo ance o he same
a iable be ween uns. The impo ance o some a iables, including
inequali y and empe a u e, di e conside ably among coun ies.
3.3. Scena io p ojec ions
Based on he his o ical ela ions lea ned in he RFC and RFR un,
p ojec ions wi h da a om h ee SSP-RCP combina ions a e made o he
combined Sou h Asian egions o Bangladesh, India, Nepal and S i
Lanka.
The RFR p ojec ions, which yield he magni ude o ne in- o ou -
mig a ion in 2050, show a sca e ed image ac oss he egion (Fig. 3a,
c, e). Some egions a e p ojec ed o ace ne ou -mig a ion in all SSP-RCP
combina ions, including wes e n Nepal, no h-eas e n India, and la ge
pa s o no he n and sou he n S i Lanka and coas al zones o
Bangladesh. A he same ime, ne in-mig a ion is p ojec ed in all sce-
na ios o he coas al a eas o wes e n India, Sou h-Eas India and la ge
pa s o inland Bangladesh, including he capi al egion o Dhaka. The
magni udes o ne in- o ou -mig a ion a e in gene al mo e mode a e o
SSP1-RCP26, as compa ed o he o he wo scena ios. The SSP2-RCP45
scena io p ojec ion is mos ou spoken in e ms o di e ences in ne in-
and ou -mig a ion. Mos a eas a e p ojec ed o ace ne in-mig a ion in
he SSP3-RCP70 scena io, 474 ou o he 969 a eas. This numbe is 286
o he SSP1-RCP70 and 273 o he SSP2-RCP45 scena io. Howe e , he
a e age ne mig a ion o e all adminis a i e a eas is mo e simila
among he scena ios, wi h a weigh ed a e age o 0.1 % in SSP3-RCP70,
while his is −0.1 % o he SSP1-RCP70 and −0.5 % o he SSP2-RCP45
scena io. The e a e some no able di e ences be ween he scena io
p ojec ions and he e e ence pe iod. Especially in no he n S i Lanka,
no he n Nepal, no h-eas India and pa s o cen al-wes e n India, he
di e ences be ween he scena io p ojec ions and he e e ence pe iod
a e subs an ial (Fig. 3b, d, ).
The RFC p ojec ions show a a he uni o m ne mig a ion pa e n
ac oss he egion o all h ee scena io p ojec ions (Fig. 4a, c, e). Fo he
majo i y o he egions, he p ojec ed p obabili y o ne in-mig a ion is
below 0.5, indica ing ha hese egions a e mos ly cha ac e ised by
socioeconomic and en i onmen al condi ions his o ically associa ed
wi h ou -mig a ion. In SSP1-RCP26 64 ou o 969 egions a e p ojec ed
o ace ne in-mig a ion, while his is 69 o SSP2-RCP4.5 and 289 o
SSP3-RCP70. In all scena ios, he No h-Wes o India is p ojec ed o
ha e he highes p obabili y o ne in-mig a ion. The di e ences be-
ween he SSP-RCP uns a e no able, wi h SSP3-RCP70 acing he lowes
le el o ou mig a ion. Compa ing he scena io p ojec ions o he
a e age o he e e ence pe iod, he pic u e is sca e ed hough. Changes
in ne mig a ion a e conside able. Majo pa s o Nepal, he no h and
Table 4
Mean e alua ion me ics o RFC and RFR un.
Coun y RF Classi ie RF
Reg ession
ROC
AUC
Accu acy P ecision Recall R
2
Bangladesh 0.88 0.88 0.80 0.53 0.59
Bhu an 0.51 0.50 0.51 0.53 −0.19
India 0.82 0.75 0.74 0.66 0.17
Nepal 0.98 0.93 0.95 0.86 0.91
Pakis an 0.90 0.83 0.80 0.72 −0.25
S i Lanka 0.80 0.73 0.67 0.56 0.53
Sou h Asia (excl.
Bhu an and
Pakis an)
0.82 0.75 0.72 0.61 0.44
S. de B uin e al.
Global En i onmen al Change 88 (2024) 102920
5
cen al pa s o S i Lanka, mos o inland Bangladesh, cen al and
sou he n India a e p ojec ed o ha e a p obabili y o ne in-mig a ion
below 0.5, indica ing a highe chance on ou -mig a ion hen in-
mig a ion. Since he a e age sha e o ne ou -mig a ion in he e e -
ence mig a ion was lowe han he p ojec ed p obabili y o ne in-
mig a ion by 2050, hese egions ace a posi i e di e ence when
compa ing he e e ence pe iod wi h he p ojec ion pe iod (Fig. 4b, d, ).
The ag eemen be ween he RFR and RFC app oaches is mode a e
(Fig. 5). A eas a e ma ked g een i he RFR a io is posi i e and he RFC
p obabili y is abo e 0.5, o i he RFR a io is nega i e and he RFR
p obabili y is below 0.5. Fo SSP1-RCP26, 72 % o he a eas o e lap in
e ms o di ec ion, 73 % o SSP2-RCP45 and 61 % o SSP3-RCP70. The
co ela ion coe icien be ween he p ojec ions is 0.2, 0.22 and 0.27(p <
0.01) o espec i ely SSP1-RCP26, SSP2-RCP45 and SSP3-RCP70. The
app oaches a e hus weakly in ag eemen .
4. Discussion
4.1. In e p e a ion o he esul s
We use wo di e en models o explain his o ical mig a ion pa e ns
in Sou h Asia and ind la gely simila pa e ns in a iable impo ance o
bo h models. The e alua ion esul s o bo h models show ha he deg ee
o which ne mig a ion can his o ically be explained di e s be ween
coun ies and be ween he wo app oaches. An ROC o 0.83 o he RFC
model and he R
2
o 0.44 o he RFR model indica e ha bo h a e able o
explain a conside able sha e o his o ical mig a ion. A he same ime,
hese e alua ion me ics also indica e ha a signi ican pa o his
p ocess is ei he explained by a iables ha a e no included o ha he
p ocess is, o some ex en , in insically unce ain. Fo example, socie al
and poli ical sen imen owa ds mig an s o mig a ion and e ugee
policies (Ha am, 2019), could no be included due o he absence o
sui able subna ional da a. Also, pe cep ions o isks and oppo uni ies in
he place o o igin o des ina ion can a ec mig a ion decisions, which
migh no always e lec he objec i e eali y (Fische e al., 1997).
Fu he mo e, ac o s such as place a achmen , people’s aspi a ions, and
cul u al p e e ences con ibu e o he complexi y o unde s anding
mig a ion, ende ing i o en an i a ional and unp edic able decision
(de Haas, 2021).
The impo ance o di e en a iables la gely con i ms he exis ing
unde s anding o p ocesses d i ing mig a ion. In bo h app oaches, he
annual a e age empe a u e is he mos impo an a iable, ollowed by
buil -up land. This implies ha ising empe a u es and a ising p o-
po ion o buil -up land could change mig a ion pa e ns conside ably.
Exis ing li e a u e has also showed ha empe a u e can a ec mig a-
ion in a ious ways. Fo example Ca aneo and Pe i (2016) show ha
highe empe a u es can inc ease ou -mig a ion le els in middle-income
coun ies by lowe ing ag icul u al p oduc i i y. Muelle e al. (2014)
conclude ha hea s ess consis en ly inc eases he long- e m mig a ion
o men in Pakis an, d i en by a nega i e e ec on a m and non- a m
income. Ano he explana ion o his s udy could be ha in he e e -
ence pe iod, people ended o mo e om he colde , mo e inaccessible
moun ainous a eas in Nepal and India o wa me , mo e e ile and u ban
a eas (Biella e al., 2022; Maha jan e al., 2020). This could explain why
ne in-mig a ion is p ojec ed in mos a eas in he RFR p ojec ion o SSP3-
RCP70, since empe a u es ise quicke and buil -up land g ows as e
compa ed o SSP1-RCP26 and SSP2-RCP45. This possibly makes a eas
mo e a ac i e in he p ojec ions based on he ob ained his o ical e-
la ions, al hough i is ques ionable o wha ex en his his o ical ela ion
will be a p edic o o u u e dynamics. In he RFR SSP3-RCP70 p o-
jec ions, leas ou -mig a ion is p ojec ed in he colde clima es o he
Moun ainous No h o India and Nepal, compa ed o he o he wo
scena ios. This dynamic is no cap u ed in he RFC scena ios, since he e
ou -mig a ion is mo e p obable in he colde moun ainous egions in all
scena ios.
Con a y o exis ing insigh s (Neumann and He mans, 2017; Ni a
e al., 2021), we did no ind income o be o p ima y impo ance o
explaining mig a ion. This obse a ion migh be explained by he ac
ha we assess ne mig a ion, a he han absolu e lows o in- and ou -
mig a ion While high income egions a e po en ially a ac i e o mi-
g an s ollowing he pe cei ed economic oppo uni ies while a he same
ime ou -mig a ion is also ound o inc ease wi h highe incomes due o
associa ed mig a ion cos s (Clemens, 2020; de Haas, 2021; G o h e al.,
2020). Ye , his migh no be isible in he ne mig a ion numbe s. The
Fig. 2. Dis ibu ion o he a iable impo ance based on he ini ial ees o he RFR and he RFC un o he sepa a e coun y uns. The boxes ep esen he
in e qua ile ange, he line in he box is he a e age alue, he whiske s maximum is 1.5 imes he in e qua ile ange, and he ed do s a e da a ou lie s. (Fo
in e p e a ion o he e e ences o colou in his igu e legend, he eade is e e ed o he web e sion o his a icle.)
S. de B uin e al.
Global En i onmen al Change 88 (2024) 102920
6
Fig. 3. Le maps: RFR scena io p ojec ions o he magni ude o ne mig a ion in pe cen age. Righ maps: The absolu e di e ence be ween he p ojec ed 2050 ne
mig a ion pe cen age and he a e age ne mig a ion pe cen age o e he e e ence pe iod 2001–2019.
S. de B uin e al.
Global En i onmen al Change 88 (2024) 102920
7
Fig. 4. Le maps: RFC scena io p ojec ions o he p obabili y o ne in-mig a ion. Righ maps: The absolu e di e ence be ween he p ojec ed 2050 p obabili y o ne
in- (1) o ou - (0) mig a ion and he a e age ne in- (1) o ou - (0) mig a ion o e he e e ence pe iod.
S. de B uin e al.
Global En i onmen al Change 88 (2024) 102920
8
