Submi ed 22 Janua y 2018
Accep ed 6 Ma ch 2018
Published 21 Ma ch 2018
Co esponding au ho
Ad ián Regos, [email p o ec ed]
Academic edi o
Louise Willemen
Addi ional In o ma ion and
Decla a ions can be ound on
page 14
DOI 10.7717/pee j.4540
Copy igh
2018 Regos and Domínguez
Dis ibu ed unde
C ea i e Commons CC-BY 4.0
OPEN ACCESS
The con ibu ion o Ea h obse a ion
echnologies o he epo ing obliga ions
o he Habi a s Di ec i e and Na u a
2000 ne wo k in a p o ec ed we land
Ad ián Regos1,2and Jesús Domínguez1
1Depa amen o de Zooloxía, Xené ica e An opoloxía Física, Uni e sidade de San iago de Compos ela,
San iago de Compos ela, Spain
2P edic i e Ecology G oup, Cen o de In es igacão em Biodi e sidade e Recu sos Gené icos da Uni e sidade
do Po o, CIBIO/InBIO, Vai ão, Po ugal
ABSTRACT
Backg ound. We lands a e highly p oduc i e sys ems ha supply a hos o ecosys em
se ices and bene i s. None heless, we lands ha e been d ained and illed o p o ide
si es o building houses and oads and o es ablishing a mland, wi h an es ima ed
wo ldwide loss o 64–71% o we land sys ems since 1900. In Eu ope, he Na u a 2000
ne wo k is he co ne s one o cu en conse a ion s a egies. E e y six yea s, Membe
S a es mus epo on implemen a ion o he Eu opean Habi a s Di ec i e. The p esen
s udy aims o illus a e how Ea h obse a ion (EO) echnologies can con ibu e o he
epo ing obliga ions o he Habi a s Di ec i e and Na u a 2000 ne wo k in ela ion o
we land ecosys ems.
Me hods. We analysed he habi a changes ha occu ed in a p o ec ed we land (in
NW Spain), 13 yea s a e i s designa ion as Na u a 2000 si e (i.e., be ween 2003 and
2016). Fo his pu pose, we analysed op ical mul ispec al bands and wa e - ela ed and
ege a ion indices de i ed om da a acqui ed by Landsa 7 TM, ETM+and Landsa
8 OLI senso s. To quan i y he unce ain y a ising om he algo i hm used in he
classi ica ion p ocedu e and i s impac on he change analysis, we compa ed he habi a
change es ima es ob ained using 10 di e en classi ica ion algo i hms and wo ensemble
classi ica ion app oaches (majo i y and weigh ed o e).
Resul s. The habi a maps de i ed om he ensemble app oaches showed an o e all
accu acy o 94% o he 2003 da a (Kappa index o 0.93) and o 95% o he 2016 da a
(Kappa index o 0.94). The change analysis e ealed impo an empo al dynamics
be ween 2003 and 2016 o he habi a classes iden i ied in he s udy a ea. Howe e ,
hese changes depended on he classi ica ion algo i hm used. The habi a maps ob ained
om he wo ensemble classi ica ion app oaches showed a educ ion in habi a classes
domina ed by sal ma shes and meadows (24.6–26.5%), na u al and semi-na u al
g asslands (25.9–26.5%) o sand dunes (20.7–20.9%) and an inc ease in o es (31–
34%) and eed bed (60.7–67.2%) in he s udy a ea.
Discussion. This s udy illus a es how EO–based app oaches migh be pa icula ly
use ul o help (1) manage s o each decisions in ela ion o conse a ion, (2)
Membe S a es o comply wi h he equi emen s o he Eu opean Habi a s Di ec i e
(92/43/EEC), and (3) he Eu opean Commission o moni o he conse a ion s a us
o he na u al habi a ypes o communi y in e es lis ed in Annex I o he Di ec i e.
How o ci e his a icle Regos and Domínguez (2018), The con ibu ion o Ea h obse a ion echnologies o he epo ing obliga ions o
he Habi a s Di ec i e and Na u a 2000 ne wo k in a p o ec ed we land. Pee J 6:e4540; DOI 10.7717/pee j.4540
None heless, he unce ain y a ising om he la ge a ie y o classi ica ion me hods
used may p e en local manage s om basing hei decisions on EO da a. Ou
esul s shed ligh on how di e en classi ica ion algo i hms may p o ide e y di e en
quan i a i e es ima es, especially o wa e -dependen habi a s. Ou indings con i m
he need o accoun o his unce ain y by applying ensemble classi ica ion app oaches,
which imp o e he accu acy and s abili y o emo e sensing image classi ica ion.
Subjec s Conse a ion Biology, Na u al Resou ce Managemen , En i onmen al Impac s, Spa ial
and Geog aphic In o ma ion Science
Keywo ds En i onmen al moni o ing, Habi a mapping, We land conse a ion, Remo e sensing,
Supe ised classi ica ion, Landsa sa elli e image y, Wa e - ela ed indices, Conse a ion Eu opean
di ec i es, Ensemble classi ica ion app oach, P o ec ed a eas
INTRODUCTION
We lands a e highly p oduc i e sys ems ha p o ide a hos o ecosys em se ices and
bene i s, including local clima e egula ion, e osion con ol, ec ea ional ishing, lood
con ol and long- e m supply o good quali y g ound wa e , s o age o pollu an s, a e
species habi a , and cul u al he i age and educa ional alue (De G oo e al., 2006;Ho wi z
& Finlayson, 2011). None heless, we lands ha e been pe cei ed as a sou ce o ec o s o
wa e bo ne in ec ious diseases, and his o ically conside ed wo hless and an impedimen
o de elopmen . Consequen ly, we lands ha e been d ained and illed o p o ide si es o
building houses and oads o o es ablishing a mland, wi h an es ima ed wo ldwide loss
o 64–71% o we land sys ems since 1900 (Da idson, 2014).
P o ec ion o we lands can come in many o ms, anging om local p ac ices and
na ional legisla ion o in e na ional ecogni ion h ough insc ip ion on he Ramsa Lis
and/o he Wo ld He i age Lis (Tho sell, Le y & Siga y, 1997). In Eu ope, he Na u a
2000 ne wo k is he co ne s one o cu en en i onmen al conse a ion s a egies. This
ne wo k includes Special P o ec ion A eas o wild bi ds (SPAs), designa ed by he Membe
S a es unde he Bi ds Di ec i e (2009/147/EC) wi h he aim o conse ing he habi a s
o pa icula ly h ea ened species and mig a o y species. I also includes Special A eas o
Conse a ion (SACs), designa ed o o he axa and habi a s unde he Habi a s Di ec i e
(92/43/ EEC). E e y six yea s, Membe S a es mus epo on implemen a ion o he
measu es aken unde hese Eu opean Di ec i es. This epo mus include in o ma ion
on he conse a ion measu es conce ning he na u al habi a ypes lis ed in Annex I o he
Habi a s Di ec i e (A . 6), as well as e alua ion o he impac s and su eillance (A . 2) o
hose measu es in ela ion o hei conse a ion s a us, wi h pa icula ega d o p io i y
na u al habi a ypes and p io i y species.
Ea h obse a ion (EO) echnologies ha e made signi ican con ibu ions o na u e
conse a ion in he las ew decades (Muchoney, 2008;O’Conno e al., 2015 and e e ence
he ein). Inc easingly la ge amoun s o geospa ial in o ma ion a e being p o ided
by sa elli e and ae ial image p ocessing and analysis—also known as emo e sensing
(RS)—which has eno mous po en ial o conse a ion applica ions (Leyequien e al.,
2007;Alca az-Segu a e al., 2009;Pe ou, Manakos & S a haki, 2015;Skidmo e e al., 2015;
Regos and Domínguez (2018), Pee J, DOI 10.7717/pee j.4540 2/19
Adamo e al., 2016, among o he s). Access o EO da a has imp o ed g ea ly in ecen yea s,
and many ae ial and sa elli e da a a e now eely a ailable (Tu ne e al., 2015).
Despi e he abo e-men ioned p og ess, he lack o a single, uni ying habi a ea u e as
well as he highly dynamic na u e o we lands (which may lead o highly a iable spec al
signa u es) and hei s eep en i onmen al g adien s (which o en p oduce na ow eco one
a eas) may cons ain and o e whelm he capaci y o cu en emo e senso s (Gallan , 2015).
Recen ad ances in compu ing and he de elopmen o image classi ica ion echniques
ha e made RS-based land-co e mapping easie , as e and mo e widely a ailable o use
in bo h conse a ion and applied ecology (Kha ami, Moun akis & S ehman, 2016). Faced
wi h his wide ange o echniques, many esea che s ha e ocused on compa ing he image
classi ica ion pe o mance o land-co e mapping o o he applica ions (e.g., Hube -Moy
e al., 2001;C acknell & Reading, 2014;Regos e al., 2015). One e ec i e solu ion o dealing
wi h he unce ain y a ising om he use o a wide ange o echniques is o gene a e a
classi ica ion ensemble by combining some indi idual classi ie s. This is e e ed o as a
mul iple classi ica ion sys em o ensemble classi ica ion app oach ( o a e iew, see Du e
al., 2012). The ensemble classi ica ion app oach, ecen ly applied by he emo e sensing
communi y, is iewed as an e ec i e way o imp o ing he classi ica ion pe o mance o
emo ely sensed image y (B iem, Benedik sson & S einsson, 2002;Lu & Weng, 2007).
The main goal o he p esen wo k is o illus a e how EO echnologies may con ibu e
o he epo ing obliga ions o he Habi a s Di ec i e and Na u a 2000 ne wo k ega ding
we land ecosys ems. We analysed he habi a changes ha ha e aken place in a p o ec ed
we land (in NW Spain), 13 yea s a e i s designa ion as Na u a 2000 si e (2003–2016). Fo
his pu pose, we analysed op ical mul ispec al bands and wa e - ela ed and ege a ion
indices de i ed om da a cap u ed by Landsa 7 TM, ETM+and Landsa 8 OLI senso s.
To quan i y he unce ain y a ising om he algo i hm used in he classi ica ion p ocedu e
and i s impac on he change analysis, we compa ed he habi a change es ima es ob ained
using 10 di e en classi ica ion algo i hms and wo ensemble classi ica ion app oaches.
MATERIAL & METHODS
S udy si e
The s udy a ea is a coas al we land included in he Na u a 2000 ne wo k in 2003 and
designa ed as Special A ea o Conse a ion (SAC) and Special P o ec ion A ea (SPA)
o wild bi ds. The si e co e s an a ea o 984 ha, co esponding o he bounda ies o he
‘‘Dunas de Co ubedo e lagoas de Ca egal e Vixán’’ Na u al Pa k (Fig. 1). The in e na ional
impo ance o he we land was ecognised when i was designa ed a Ramsa si e, in 1993.
This we land includes one o he la ges dune sys ems in he NW Ibe ian Peninsula,
wi h ex ensi e s e ches o sand (Ladei a, Fe ei a and Vila beaches) lanked by la ge
dune and coas al lagoon ecosys ems (Lagunas de Ca egal and Vixán), oge he wi h an
adjacen dune sys em, and an emb yonic shi ing dune (1-km long, 200–250 m wide and
12–15 m high) (Vázquez-Paz & Pé ez-Albe i, 2002). The dune sys em, comp ising a sandy
ba ie , has a ou ed he c ea ion o an in e io sedimen a y a ea composed o ixed dunes
(‘g ey dunes’), ma shes, sandy and muddy in e idal zones, as well as wo coas al lagoons
Regos and Domínguez (2018), Pee J, DOI 10.7717/pee j.4540 3/19
Figu e 1 Loca ion o he s udy a ea and p o ec ed-a ea sys ems. Ramsa we land (dashed-do ed line),
Na u al pa k and SAC (black dashed line) and SPA ( illing lines).
Full-size DOI: 10.7717/pee j.4540/ ig-1
wi h e y di e en aqua ic cha ac e is ics: (1) he Ca egal lagoon co e s an i egula ,
delimi ed space be ween he ma sh and he dune sys em. The a ea adjacen o he coas line
co esponds mo phologically o an es ua ine channel co e ed by sandy deposi s whe e
looding depends on he idal cycle (Fig. 1); (2) he Vixán lagoon, loca ed in he a ea
dis al o he coas line, has a dense eed bed (Ph agmi es aus alis) ha occupies mos o
he euli o al and sup ali o al en i onmen s (Fig. 1). In he a ea adjacen o he coas ,
he eed bed is eplaced by bul ushes (Typha la ipholia) and, o a lesse ex en , by we
g asslands. The d ainage channel zigzags h ough he dune sys em un il eaching he beach
(Ramil-Rego, 2007).
P e-p ocessing EO da a
We used sa elli e emo e sensing image y o map and moni o he habi a changes ha ha e
aken place be ween 2003 and 2016. We analysed op ical mul ispec al bands (Pa h/Row:
205/30) de i ed om ou cloud- ee images acqui ed by NASA’s Landsa missions on
20 Ma ch (Landsa 7 ETM+) and 6 Oc obe 2003 (Landsa 5 TM) and on 2 May and
23 Sep embe 2016 (Landsa 8 OLI) (de ailed in o ma ion a ailable o each band is
a ailable a : h p://landsa .usgs.go /band_designa ions_landsa _sa elli es.php). Landsa
scenes cap u ed in sp ing and au umn (e.g., in May and Sep embe ) we e analysed o ake
in o accoun seasonal di e ences in ege a ion phenology (e.g., common eed g ass). The
images a e all a ailable ee o cha ge om he US Geological Su ey (USGS) Cen e o
Regos and Domínguez (2018), Pee J, DOI 10.7717/pee j.4540 4/19
Ea h Resou ces Obse a ion and Science (EROS) and we e ob ained by di ec download
om he GloVis acili y (h p://glo is.usgs.go ).
All downloaded images we e L1T (a p ocessing le el ha includes a geome ic co ec ion
pe o med wi h g ound con ol poin s and he use o a digi al ele a ion model) and
p ojec ed in he UTM coo dina e sys em (WGS 84 da um, UTM p ojec ion, Zone 29
No h). Digi al numbe s (DNs) we e con e ed o op-o -a mosphe e adiance and
physically meaning ul uni s by adiome ic calib a ion and applica ion o senso - and band-
speci ic calib a ion pa ame e s. The classi ica ion p ocess was based on he adiome ic
in o ma ion ob ained om e lec i e bands and wo mul ispec al indices o each image:
(1) he No malized Di e ence Vege a ion Index (NDVI; Rouse e al., 1974) and (2) he
No malized Di e ence Wa e Index (NDWI; Gao, 1996). This p ocedu e enhanced he
spec al sepa abili y o ege a ion associa ed wi h aqua ic and halophilic en i onmen s.
Classi ica ion p ocedu e
Supe ised classi ica ion o he emo ely-sensed da a was ca ied ou using he ollowing
10 classi ica ion algo i hms a ailable in he R-based package Ca e and implemen ed in he
RS oolbox package, e sion 0.1.5 (Kuhn, 2017;Leu ne & Ho ning, 2017): amdai (Adap i e
Mix u e Disc iminan Analysis), a NNe (Model A e aged Neu al Ne wo k), gbm
(S ochas ic G adien Boos ing), knn (k-Nea es Neighbou s), mda (Mix u e Disc iminan
Analysis), pls (Pa ial Leas Squa es), (Random Fo es ), s mPoly (Suppo Vec o
Machines wi h Polynomial Ke nel), s mLinea (Suppo Vec o Machines wi h Linea
Ke nel) and s mRadial (Suppo Vec o Machines wi h Radial Basis Func ion Ke nel). In
addi ion, wo ensemble p ocedu es we e pe o med: (1) a simple o ing sys em (‘Ens_SV’;
he so-called ‘majo i y o ing’ and ‘selec all majo i y’ sys em, sensu Baue e al., 1999),
conside ing each habi a map as an equally weigh ed o e; and (2) a weigh ed o ing
app oach (‘Ens_WV’), using o e all accu acy ob ained by indi idual classi ie s as weigh s
(Du e al., 2012).
Eigh habi a classes, de ined as a eas wi h common ecological and biophysical
cha ac e is ics and, he e o e, wi h a homogeneous spec al signa u e, we e iden i ied
in he s udy a ea. Fo hese habi a classes, we adop ed he e minology used in he Annex
I o he Habi a s Di ec i e. These eigh habi a classes co espond wi h 23 speci ic habi a s
lis ed in his Annex I in ou s udy a ea (Table 1). The s udy a ea is e y well desc ibed, and
he whole lis o habi a s is al eady de ined in p e ious epo s (see e.g., Ramil-Rego e al.,
2008). T aining and alida ion a eas o each habi a class we e es ablished by on-sc een
digi izing in QGIS so wa e, and consis ed o a se o pixels iden i ied o e well-known
homogeneous a eas in each Landsa image, hus p o iding a e e ence spec al signa u e
o each class. In pa icula , we applied a s a i ied andom design as sampling s a egy,
wi h a o al o abou 259–346 aining and alida ion a eas p opo ionally dis ibu ed
h oughou he en i e s udy a ea o each yea (Table 2;Da ase S1). Speci ically, o 2003
we used di e en Red-G een-Blue (RGB) composi es om he Landsa bands and digi al
o hopho os in na u al colou s a a scale o 1:18,000 ob ained om he Plan Nacional de
O o o og a ía Aé ea (PNOA) o 2004, while o 2016 we used digi al o hopho os om
2014.
Regos and Domínguez (2018), Pee J, DOI 10.7717/pee j.4540 5/19
Table 1 Lis o b oad habi a classes used in he change analysis and hei co espondence wi h he na u al habi a s (and codes) lis ed in he
Annex I o he Habi a s Di ec i e. As e isk indica es habi a s wi h highes p io i y o conse a ion acco ding o he Habi a s Di ec i e.
Habi a class Na u al habi a s lis ed in he Annex I o he Habi a Di ec i e
Sand dunes 1110 Sandbanks which a e sligh ly co e ed by sea wa e all he ime.
1140 Mud la s and sand la s no co e ed by sea wa e a low ide.
1210 Annual ege a ion o d i lines.
2110 Emb yonic shi ing dunes.
2120 Shi ing dunes along he sho eline wi h Ammophila a ena ia (‘whi e dunes’).
Tidal a eas 1130 Es ua ies.
1150* Coas al lagoons.
1160 La ge shallow inle s and bays.
1170 Ree s.
Fo es –
Reedbed –
Sea dunes o A lan ic coas 2130* Fixed coas al dunes wi h he baceous ege a ion (‘g ey dunes’)
2150* A lan ic decalci ied ixed dunes (Calluno-Ulice ea).
2190 Humid dune slacks.
2230 Malcolmie alia dune g asslands.
2260 Cis o-La endule alia dune scle ophyllous sc ubs.
Na u al and semi-na u al g asslands 6220* Pseudo-s eppe wi h g asses and annuals o he The o-B achypodie ea
6410 Molinia meadows on calca eous, pea y o clayey-sil -laden soils (Molinion cae uleae).
6420 Medi e anean all humid g asslands o he Molinio-Holoschoenion.
6430 Hyd ophilous all he b inge communi ies o plains and o he mon ane o alpine le els.
6510 Lowland hay meadows (Alopecu us p a ensis,Sanguiso ba o icinalis)
Sal ma shes and meadows 1310 Salico nia and o he annuals colonizing mud and sand.
1330 A lan ic sal meadows (Glauco-Puccinellie alia ma i imae).
1420 Medi e anean and he mo-A lan ic halophilous sc ubs (Sa coco ne ea u icosi).
Bu ned a eas –
Table 2 To al numbe o aining and alida ion a eas conside ed in he supe ised classi ica ion o
each habi a class and yea .
Habi a class T aining Valida ion T aining Valida ion
2003 2016
Sand dunes 38 30 35 29
Tidal a eas 48 29 49 31
Fo es 64 61 76 43
Reedbed 21 22 20 23
Sea dunes o A lan ic coas 46 29 56 29
Na u al and semi-na u al g asslands 31 55 44 39
Sal ma shes and meadows 34 33 33 26
Bu ned a eas 0 0 39 23
TOTAL 282 259 346 243
Valida ion p ocedu e
The accu acy o habi a maps was assessed om con usion ma ices based on he numbe o
pixels co ec ly (and inco ec ly) classi ied pe class, and by compa ing he esul s ob ained
om di e en classi ica ion algo i hms. The main quali y pa ame e s we e he o e all
accu acy (%), he p oduce ’s and use ’s accu acies, and he Kappa index o ag eemen .
Regos and Domínguez (2018), Pee J, DOI 10.7717/pee j.4540 6/19
We used McNema ’s es s o e alua e s a is ical signi icance o he di e ence in accu acy
be ween each pai o algo i hms. This is a non-pa ame ic es ha is based on con usion
ma ices collapsed o wo by wo con ingency ables (Foody, 2004;De Leeuw e al., 2006).
P- alues om McNema ’s es s we e ep esen ed wi h hea maps o help isualizing
s a is ical signi icance o he di e ence be ween all possible compa isons. These p- alues
we e used o suppo he selec ion o algo i hms o he subsequen ensemble p ocedu es.
The eby, classi ica ion algo i hms wi h s a is ically lowe accu acies we e no included in
he ensemble p ocedu es (McNema ’s es s, p<0.05).
Da a impo a ion, p e-p ocessing, spec al indices, image classi ica ion and g aphical
display we e pe o med using he oolse a ailable in RS oolbox package, e sion 0.1.5
(Wegmann, Leu ne & Dech, 2016;Leu ne & Ho ning, 2017) (see h p:// pubs.com/
ARegos/359655 o R code and o ma ed ou pu s).
Change analysis
We quan i ied he spa ial ex en (in ha) o each habi a class pe yea (2003 and 2016) om
each classi ica ion algo i hm and ensemble app oach. Boxplo s we e cons uc ed using he
R package ggplo 2 (Wickham, 2009). The con ibu ion o each habi a class o he habi a
change (i.e., con e sion om one habi a class o ano he ) was showed h ough a ansi ion
ma ix ob ained by c oss- abula ion o he habi a maps de i ed om he wo ensemble
classi ica ion app oaches. T ansi ion ma ices we e compu ed wi h he R package lulcc
.1.0.2 (Moulds, 2017) (see h p:// pubs.com/ARegos/359655 o R code and o ma ed
ou pu s).
RESULTS
Accu acy assessmen
The habi a maps wi h he highes accu acy (up o 95%) in 2003 we e ob ained using
suppo ec o machines and disc iminan analysis, wi h he ‘amdai’ classi ie p o iding
sligh ly be e esul s (Fig. 2). Fo 2016, he highes accu acy was ob ained by applying
suppo ec o machines wi h linea ke nel (Fig. 2). Howe e , McNema ’s es did no
show s a is ical signi icance o he di e ence in accu acy be ween indi idual classi ica ion
algo i hms (p>0.05; Fig. 3), excep o ‘pls’, ‘a NNe ’ (p<0.01; Fig. 3), ‘gbm and
‘s mRadial’ (p<0.05; Fig. 3). These algo i hms showed limi a ions o speci ic habi a
classes ha ha e led o unde - and o e es ima ions o hei ex en (Fig. 4). Fo ins ance,
‘pls’ showed e y low use ’s accu acies o he hema ic class ‘ o es ’, while ‘s mRadial’
ma kedly o e es ima ed he habi a class ‘ idal a ea’ (see low use ’s accu acy and high
p oduce ’s accu acy alues, i.e., low omission e o s and high commission e o s in Fig. 4).
The eby, ‘pls’ and ‘a NNe ’ o bo h yea s and ‘gbm’ and ‘s mRadial’ o yea 2016 we e
inally no conside ed du ing he ensemble p ocedu es.
The habi a maps de i ed om he ensemble app oaches (majo i y and weigh ed o e)
showed an o e all accu acy o 94% o he 2003 da a (Kappa index o 0.93) and o 95% o
he 2016 da a (Kappa index o 0.94) (Fig. 2) wi h no s a is ical signi icance o he di e ence
be ween hem (p>0.05). Change analysis was he e o e pe o med using he wo ensemble
me hods.
Regos and Domínguez (2018), Pee J, DOI 10.7717/pee j.4540 7/19
Accu acy
Kappa
2003 2016 2003 2016
0.6
0.7
0.8
0.9
Land co e class
Yea
alg
amdai
a NNe
Ens_SV
Ens_WV
gbm
knn
mda
pls
s mLinea
s mPoly
s mRadial
yea
2003
2016
Figu e 2 Accu acy o habi a maps (o e all accu acy and Kappa coe icien ) pe yea and classi ica ion
me hod. amdai (Adap i e Mix u e Disc iminan Analysis), a NNe (Model A e aged Neu al Ne wo k),
gbm (S ochas ic G adien Boos ing), knn (k-Nea es Neighbou s), mda (Mix u e Disc iminan Analysis),
pls (Pa ial Leas Squa es), (Random Fo es ), s mPoly (Suppo Vec o Machines wi h Polynomial Ke -
nel), s mRadial (Suppo Vec o Machines wi h Radial Basis Func ion Ke nel), s mLinea (Suppo Vec-
o Machines wi h Linea Ke nel), simply o ing (‘Ens_SV’) and weigh ed o ing (‘Ens_WV’) ensemble
app oaches. The boxplo s display he median, he 50% (box) and 95% (whiske s) con idence in e als.
Full-size DOI: 10.7717/pee j.4540/ ig-2
Change analysis
The change analysis e ealed impo an empo al dynamics be ween 2003 and 2016 o he
habi a classes iden i ied in he s udy a ea (Fig. 5,Table 3). Howe e , he changes depended
on he classi ica ion algo i hm used (Fig. 5). Fo example, alues o wa e -dependen
habi a classes anged om a ound 60 ha wi h mos o he classi ica ion algo i hms, o
almos 7 imes his alue wi h he ‘s mRadial’ classi ie , clea ly indica ing o e es ima ion
o his uni (Figs. 5–7). The co e age es ima ed o habi a class domina ed by sal ma shes
and meadows in 2016 anged om alues close o 52 ha wi h he ‘s mRadial’ classi ie o
mo e han 260 ha wi h he ‘gbm’ classi ie (Fig. 5).
The habi a maps ob ained using he wo ensemble classi ica ion app oaches show a
educ ion in habi a classes domina ed by sal ma shes and meadows (24.6–26.5%), na u al
and semi-na u al g asslands (25.9–26.5%) o sand dunes (20.7–20.9%) and an inc ease in
o es (31–34%) and eed bed (60.7–67.2%) in he s udy a ea (Fig. 5). In pa icula , he
Regos and Domínguez (2018), Pee J, DOI 10.7717/pee j.4540 8/19
1 0.68
1
1
1
1
0.45
0.22
0.1
0.22
1
0.68
0.25
0.62
0.72
0.01
0.01
0.01
0.01
0.01
0.01
0.37
0.68
1
0.62
0.1
0.13
0.01
0.68
1
1
1
0.18
0.37
0.01
1
0.01
0.01
0.01
0.01
0.1
0.05
0.01
0.01
0.01
s mPoly
s mRadial
s mLinea
knn
gbm
pls
mda
amdai
s mRadial s mLinea knn gbm pls mda amdai a NNe
0.00 0.25 0.50 0.75 1.00
McNema Tes (p− alue)
Yea 2003
0.13 1
0.05
1
0.05
0.48
0.05
1
1
0.22
1
0.1
0.1
0.05
0.01
0.01
0.01
0.01
0.01
0.01
1
0.13
1
1
0.48
0.22
0.01
0.62
0.37
0.25
0.25
0.13
0.68
0.01
0.48
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
s mPoly
s mRadial
s mLinea
knn
gbm
pls
mda
amdai
s mRadial s mLinea knn gbm pls mda amdai a NNe
0.00 0.25 0.50 0.75 1.00
McNema Tes (p− alue)
Yea 2016
A) B)
Figu e 3 P- alue om McNema ’s es s o each pai o classi ica ion algo i hm. amdai (Adap i e
Mix u e Disc iminan Analysis), a NNe (Model A e aged Neu al Ne wo k), gbm (S ochas ic G adien
Boos ing), knn (k-Nea es Neighbou s), mda (Mix u e Disc iminan Analysis), pls (Pa ial Leas Squa es),
(Random Fo es ), s mPoly (Suppo Vec o Machines wi h Polynomial Ke nel), s mRadial (Suppo
Vec o Machines wi h Radial Basis Func ion Ke nel), s mLinea (Suppo Vec o Machines wi h Linea
Ke nel). Whi e colou s indica e p- alues lowe han 0.01, blue colou in ensi y inc eases wi h he p- alue.
Full-size DOI: 10.7717/pee j.4540/ ig-3
Sand dunes
Tidal a eas
Fo es
Reedbed
Sea dunes
G asslands
Sal ma shes
Bu ned a eas
2003
P A
2016
P A
2003
U A
2016
U A
0.85
0.90
0.95
1.00
0.6
0.7
0.8
0.9
1.00.4
0.6
0.8
1.0
0.00
0.25
0.50
0.75
1.00
0.00
0.25
0.50
0.75
1.00
0.6
0.8
1.0
0.00
0.25
0.50
0.75
1.00
0.7
0.8
0.9
1.0
amdai
a NNe
gbm
knn
mda
pls
s mPoly
s mRadial
s mLinea
Ens_SV
Ens_WV
amdai
a NNe
gbm
knn
mda
pls
s mPoly
s mRadial
s mLinea
Ens_SV
Ens_WV
amdai
a NNe
gbm
knn
mda
pls
s mPoly
s mRadial
s mLinea
Ens_SV
Ens_WV
amdai
a NNe
gbm
knn
mda
pls
s mPoly
s mRadial
s mLinea
Ens_SV
Ens_WV
Accu acy
Classi ica ion me hod
alg
amdai
a NNe
Ens_SV
Ens_WV
gbm
knn
mda
pls
s mLinea
s mPoly
s mRadial
Figu e 4 P oduce ’s (P A) and use ’s (U A) accu acy pe yea , habi a class and classi ica ion me hod.
amdai (Adap i e Mix u e Disc iminan Analysis), a NNe (Model A e aged Neu al Ne wo k), gbm
(S ochas ic G adien Boos ing), knn (k-Nea es Neighbou s), mda (Mix u e Disc iminan Analysis), pls
(Pa ial Leas Squa es), (Random Fo es ), s mPoly (Suppo Vec o Machines wi h Polynomial Ke nel),
s mRadial (Suppo Vec o Machines wi h Radial Basis Func ion Ke nel), s mLinea (Suppo Vec o
Machines wi h Linea Ke nel), simply o ing (‘Ens_SV’) and weigh ed o ing (‘Ens_WV’) ensemble
app oaches. See Table 1 o habi a classes.
Full-size DOI: 10.7717/pee j.4540/ ig-4
Regos and Domínguez (2018), Pee J, DOI 10.7717/pee j.4540 9/19
C acknell MJ, Reading AM. 2014. Geological mapping using emo e sensing da a: a
compa ison o i e machine lea ning algo i hms, hei esponse o a ia ions in
he spa ial dis ibu ion o aining da a and he use o explici spa ial in o ma ion.
Compu e s and Geosciences 63:22–33 DOI 10.1016/j.cageo.2013.10.008.
Da idson NC. 2014. How much we land has he wo ld los ? Long- e m and ecen
ends in global we land a ea. Ma ine and F eshwa e Resea ch 65:934–941
DOI 10.1071/MF14173.
Delalieux S, Some s B, Haes B, Koois a L, Müche CA, Vanden Bo e J. 2010.
Moni o ing hea hland habi a s a us using hype spec al image classi ica ion
and unmixing. In: 2nd Wo kshop on hype spec al image and signal p ocess-
ing: e olu ion in emo e sensing, WHISPERS 2010–wo kshop p og am. 7–10
DOI 10.1109/WHISPERS.2010.5594895.
De G oo R, S uip M, Finlayson M, Da idson N. 2006. Valuing we lands: guidance
o aluing he bene i s de i ed om we land ecosys em se ices. Ramsa echnical
epo no 3. Ramsa Con en ion Sec e a ia , Gland.
De Leeuw J, Jia H, Yang L, Liu X, Schmid K, Skidmo e AK. 2006. Compa ing accu acy
assessmen s o in e supe io i y o image classi ica ion me hods. In e na ional Jou nal
o Remo e Sensing 27:223–232 DOI 10.1080/01431160500275762.
Domínguez J, O e o M, Vidal M. 2006. Guía de las a es del Pa que Na u al do complexo
duna de Co ubedo e lagoas de Ca egal e Vixán. San iago de Compos ela: Con-
selle ía de Medio Ambien e e Desen ol emen o Sos ible.
Du P, Xia J, Zhang W, Tan K, Liu Y, Liu S. 2012. Mul iple classi ie sys em o emo e
sensing image classi ica ion: a e iew. Senso s 12:4764–4792 DOI 10.3390/s120404764.
Foody GM. 2004. Thema ic map compa ison: e alua ing he s a is ical signi icance o
di e ences in classi ica ion accu acy. Pho og amme ic Enginee ing & Remo e Sensing
70:627–633 DOI 10.14358/PERS.70.5.627.
Gallan AL. 2015. The challenges o emo e moni o ing o we lands. Remo e Sensing
7:10938–10950 DOI 10.3390/ s70810938.
Gao BC. 1996. NDWI—a no malized di e ence wa e index o emo e sensing o
ege a ion liquid wa e om space. Remo e Sensing o En i onmen 58:257–266
DOI 10.1016/S0034-4257(96)00067-3.
Gon¸
cal es J, Hen iques R, Al es P, Sousa-Sil a R, Mon ei o AT, Lomba Â, Ma cos
B, Hon ado J. 2016. E alua ing an unmanned ae ial ehicle-based app oach o
assessing habi a ex en and condi ion in ine-scale ea ly successional moun ain
mosaics. Applied Vege a ion Science 19:132–146 DOI 10.1111/a sc.12204.
Gonzalez-Ma ínez XI. 2014. Nue os da os sob e xeno lo a en el é mino municipal de
Ribei a (A Co uña). No a Ac a Cien i ica Compos elana 21:65–76.
González-Ma ínez XI. 2017. Con ibución al conocimien o de la lo a alóc ona de
Galicia (NO Península Ibé ica, España). Bo anica Complu ense 41:53–67.
Ho wi z P, Finlayson CM. 2011. We lands as se ings o human heal h: inco po a ing
ecosys em se ices and heal h impac assessmen in o wa e esou ce managemen .
BioScience 61:678–688 DOI 10.1525/bio.2011.61.9.6.
Regos and Domínguez (2018), Pee J, DOI 10.7717/pee j.4540 16/19
Hube -Moy L, Co onnec A, Le Du L, Cha din A, Pe ez P. 2001. A compa i-
son o pa ame ic classi ica ion p ocedu es o emo ely sensed da a applied
on di e en landscape uni s. Remo e Sensing o En i onmen 75:174–187
DOI 10.1016/S0034-4257(00)00165-6.
Kha ami R, Moun akis G, S ehman S. 2016. A me a-analysis o emo e sensing
esea ch on supe ised pixel-based land-co e image classi ica ion p ocesses: gene al
guidelines o p ac i ione s and u u e esea ch. Remo e Sensing o En i onmen
177:89–100 DOI 10.1016/j. se.2016.02.028.
Kuhn M. 2017. ca e : classi ica ion and eg ession aining. R package e sion 6.0-76.
A ailable a h ps://CRAN.R-p ojec .o g/package=ca e .
K is L, Ponnikas S, Belda EJ, Encabo I, Ma ínez E, On ubia A, He nández JM, Ve a P,
Ne o JM, Mon ós JS. 2011. Endange ed subspecies o he Reed Bun ing (Embe iza
schoeniclus wi he byi and E. s. lusi anica) in Ibe ian Peninsula ha e di e en gene ic
s uc u es. Jou nal o O ni hology 152:681–693 DOI 10.1007/s10336-011-0646-9.
Lehmann JRK, P inz T, Zille SR, Thiele J, He inge G, Mei a-Ne o JAA, Bu scha d
TK. 2017. Open-sou ce p ocessing and analysis o ae ial image y acqui ed wi h a
low-cos unmanned ae ial sys em o suppo in asi e plan managemen . F on ie s
in En i onmen al Science 5:1–16 DOI 10.3389/ en s.2017.00044.
Leu ne B, Ho ning N. 2017. RS oolbox: ools o emo e sensing da a analysis. R
package e sion 0.1.8. A ailable a h ps://CRANR-p ojec .o g/package=RS oolbox.
Leyequien E, Ve els J, Slo M, Schaepman-S ub G, Hei könig IMA, Skidmo e
A. 2007. Cap u ing he ugi i e: applying emo e sensing o e es ial animal
dis ibu ion and di e si y. In e na ional Jou nal o Applied Ea h Obse a ion and
Geoin o ma ion 9:1–20 DOI 10.1016/j.jag.2006.08.002.
Lu D, Weng Q. 2007. A su ey o image classi ica ion me hods and echniques o
imp o ing classi ica ion pe o mance. In e na ional Jou nal o Remo e Sensing
28:823–870 DOI 10.1080/01431160600746456.
Ma caccio JV, Ma kle CE, Chow-F ase P. 2015. Unmanned ae ial ehicles p oduce
high- esolu ion, seasonally- ele an image y o classi ying we land ege a ion.
In e na ional A chi es o he Pho og amme y, Remo e Sensing and Spa ial In o ma ion
Sciences 40:249–256 DOI 10.5194/isp sa chi es-XL-1-W4-249-2015.
Moulds S. 2017. lulcc: land use change modelling in R. R package e sion 1.0.2. A ailable
a h ps://c an. -p ojec .o g/package=lulcc.
Muchoney DM. 2008. Ea h obse a ions o e es ial biodi e si y and ecosys ems.
Remo e Sensing o En i onmen 112:1909–1911 DOI 10.1016/j. se.2008.01.003.
Ng WT, Rima P, Einzmann K, Immi ze M, A zbe ge C, Ecke S. 2017. Assessing he
po en ial o sen inel-2 and pléiades da a o he de ec ion o p osopis and achellia
spp. in Kenya. Remo e Sensing 9(1):74 DOI 10.3390/ s9010074.
O’Conno B, Secades C, Penne J, Sonnenschein R, Skidmo e A, Bu gess ND,
Hu on JM. 2015. Ea h obse a ion as a ool o acking p og ess owa ds he
Aichi Biodi e si y Ta ge s. Remo e Sensing in Ecology and Conse a ion 1:19–28
DOI 10.1002/ se2.4.
Regos and Domínguez (2018), Pee J, DOI 10.7717/pee j.4540 17/19
Pe ou ZI, Manakos I, S a haki T. 2015. Remo e sensing o biodi e si y moni o -
ing: a e iew o me hods o biodi e si y indica o ex ac ion and assessmen o
p og ess owa ds in e na ional a ge s. Biodi e si y and Conse a ion 24:2333–2363
DOI 10.1007/s10531-015-0947-z.
Ramil-Rego P. 2007. E aluación del es ado de conse ación de las lagunas cos e as de Gali-
cia: p opues as de conse ación y uso sos enible. San iago de Compos ela: Fundación
A ao.
Ramil-Rego P, Rod íguez-Gui ián MA, Fe ei o da Cos a J, Rubinos-Román M,
Gómez-O ellana Rod íguez L, De Nó oa Fe nández B, Hinojo-Sánchez BA,
Ma ínez-Sánchez S, Cille o-Cas o C, Albe o Díaz-Va ela R, Rod íguez-González
PM, Muñoz-Sob ino C. 2008. Os Hábi a s de In e ese Comuni a io en Galicia. Fichas
desc i i as. Monog a ías do Ibade . Lugo: Uni e sidade de San iago de Compos ela.
Regos A, Ninye ola M, Mo é G, Pons X. 2015. Linking land co e dynamics wi h
d i ing o ces in moun ain landscape o he No hwes e n Ibe ian Peninsula.
In e na ional Jou nal o Applied Ea h Obse a ion and Geoin o ma ion 38:1–14
DOI 10.1016/j.jag.2014.11.010.
Romão C. 1996. In e p e a ion manual o Eu opean Union habi a s. Ve sion EUR 15.
B ussels: Eu opean Commission. A ailable a h p://aei.pi .edu/33244/1/A3.pd .
Rouse JW, Hass RW, Schell JA, Dee ing DH, Ha lan JC. 1974. Moni o ing he e nal
ad ancemen and e og ada ion (g een wa e e ec ) o na u al ege a ion. G eenbel :
NASA/GSFC.
Skidmo e AK, Pe o elli N, Coops NC, Gelle GN, Hansen M, Lucas R, Müche CA,
O’Conno B, Paganini M, Pe ei a HM, Schaepman ME, Tu ne W, Wang T,
Wegmann M. 2015. En i onmen al science: ag ee on biodi e si y me ics o ack
om space. Na u e 523:403–405 DOI 10.1038/523403a.
S ellmes M, Röde A, Udelho en T, Hill J. 2013. Mapping synd omes o land change
in Spain wi h emo e sensing ime se ies, demog aphic and clima ic da a. Land Use
Policy 30:685–702 DOI 10.1016/j.landusepol.2012.05.007.
S a oulias D, Balz e H, Sykio i O, Zlinszky A, Tó h VR. 2015. E alua ing sen inel-
2 o lakesho e habi a mapping based on ai bo ne hype spec al da a. Senso s
15:22956–22969 DOI 10.3390/s150922956.
Tho sell J, Le y R, Siga y T. 1997. A global o e iew o we land and ma ine p o ec ed
a eas on he Wo ld He i age lis . Gland: IUCN. A ailable a h ps://www.iucn.o g/
con en /global-o e iew-we land-and-ma ine-p o ec ed-a eas-wo ld-he i age-lis .
Tu ne W, Rondinini C, Pe o elli N, Mo a B, Leidne A, Szan oi Z, Buchanan G,
Dech S, Dwye J, He old M, Koh LP, Leimg ube P, Taubenboeck H, Weg-
mann M, Wikelski M, Woodcock C. 2015. F ee and open-access sa elli e da a
a e key o biodi e si y conse a ion. Biological Conse a ion 182:173–176
DOI 10.1016/j.biocon.2014.11.048.
Valkama E, Lyy inen S, Ko iche a J. 2008. The impac o eed managemen on wildli e:
a me a-analy ical e iew o Eu opean s udies. Biological Conse a ion 141:364–374
DOI 10.1016/j.biocon.2007.11.006.
Regos and Domínguez (2018), Pee J, DOI 10.7717/pee j.4540 18/19
Vázquez-Paz M, Pé ez-Albe i A. 2002. Análisis dinámico y medioambien al de un
espacio p o egido en la cos a a lán ica gallega (Pa que Na u al de Co ubedo, A
Co uña). Xeog á ica 2:155–176.
Wegmann M, Leu ne B, Dech S. 2016. Remo e sensing and GIS o ecologis s: using open
sou ce so wa e. Exe e : Pelagic Publishing L d.
Wickham. 2009. ggplo 2: elegan g aphics o da a analysis. New Yo k: Sp inge -Ve lag
New Yo k.
Zo illa-Mi as P, Palomo I, Gómez-Bagge hun E, Ma ín-López B, Lomas PL, Mon es
C. 2014. E ec s o land-use change on we land ecosys em se ices: a case s udy in
he Doñana ma shes (SW Spain). Landscape and U ban Planning 122:160–174
DOI 10.1016/j.landu bplan.2013.09.013.
Regos and Domínguez (2018), Pee J, DOI 10.7717/pee j.4540 19/19
