A reconstruction of Turkey’s potential natural vegetation using climate indicators

Abstract

Turkey, containing three of the world’s biodiversity hotspots, is a hub for genetic biodiversity. However, the vegetation cover has drastically changed in recent decades as a result of substantial transformations in land-use practices. A map of the potential natural vegetation can be used to represent the biodiversity of a country, and therefore a reference to effectively develop conservation strategies. The multinomial logistic regression is used to simulate the probability of different biomes occurring in the country using elevation, climatological data and natural vegetation data. A correlation test was applied to the climatological data to determine which predictors influence vegetation the most. These were temperature, precipitation, relative humidity and cloudiness. The Ordinary Kriging method was employed to transform the data into the format for the multinomial logistic regression model. The model showed that temperature was the most influencing factor with respect to Turkey’s vegetation and distribution follows a similar distribution as the various macroclimates. Broadleaf forests are mostly found in the Black Sea region, which is also the wettest region of the country. The Marmara region is the only other region where there are broadleaf forests. Mixed forests and shrublands are mostly located in Central Anatolia due to the region’s low humidity which favours herbaceous flora. Coniferous forests were dominant in the Aegean and Mediterranean regions, attributed to high temperatures.

Full text

1 23
Journal of Forestry Research
ISSN 1007-662X
J. For. Res.
DOI 10.1007/s11676-018-0855-7
A reconstruction of Turkey’s potential
natural vegetation using climate indicators
Nussaïbah B.Raja, Olgu Aydin, İhsan
Çiçek & Necla Türkoğlu

1 23
Your article is protected by copyright and
all rights are held exclusively by Northeast
Forestry University. This e-offprint is for
personal use only and shall not be self-
archived in electronic repositories. If you wish
to self-archive your article, please use the
accepted manuscript version for posting on
your own website. You may further deposit
the accepted manuscript version in any
repository, provided it is only made publicly
available 12 months after official publication
or later and provided acknowledgement is
given to the original source of publication
and a link is inserted to the published article
on Springer's website. The link must be
accompanied by the following text: "The final
publication is available at link.springer.com”.

ORIGINAL PAPER
A reconstruction of Turkey’s potential natural vegetation using
climate indicators
Nussaı
¨bah B. Raja
1
•Olgu Aydin
2
•I
˙hsan C¸ic¸ek
2
•Necla Tu
¨rkog
˘lu
2
Received: 29 March 2017 / Accepted: 28 November 2017
Northeast Forestry University 2018
Abstract Turkey, containing three of the world’s biodi-
versity hotspots, is a hub for genetic biodiversity. However,
the vegetation cover has drastically changed in recent
decades as a result of substantial transformations in land-
use practices. A map of the potential natural vegetation can
be used to represent the biodiversity of a country, and
therefore a reference to effectively develop conservation
strategies. The multinomial logistic regression is used to
simulate the probability of different biomes occurring in
the country using elevation, climatological data and natural
vegetation data. A correlation test was applied to the cli-
matological data to determine which predictors influence
vegetation the most. These were temperature, precipitation,
relative humidity and cloudiness. The Ordinary Kriging
method was employed to transform the data into the format
for the multinomial logistic regression model. The model
showed that temperature was the most influencing factor
with respect to Turkey’s vegetation and distribution fol-
lows a similar distribution as the various macroclimates.
Broadleaf forests are mostly found in the Black Sea region,
which is also the wettest region of the country. The Mar-
mara region is the only other region where there are
broadleaf forests. Mixed forests and shrublands are mostly
located in Central Anatolia due to the region’s low
humidity which favours herbaceous flora. Coniferous for-
ests were dominant in the Aegean and Mediterranean
regions, attributed to high temperatures.
Keywords Biomes Multinomial logistic regression 
Statistical modelling Turkey Vegetation
Introduction
Turkey is the only country covered almost entirely by three
of the world’s 34 biodiversity hotspots, namely the Cau-
casus, Irano-Anatolian and Mediterranean, and is a hub for
genetic biodiversity, hosting approximately 10,000 plant
and 80,000 animal species (C¸ olak and Rotherham 2006;
Serkercioglu et al. 2011). However, vegetation cover has
drastically changed in recent decades as a result of sub-
stantial transformations in land-use practices, especially
urbanisation by an increasing population (Gu
¨ler et al. 2007;
Evrendilek et al. 2011). Most of Turkey’s eco-regions have
been declared as critically endangered (Hoekstra et al.
2005), which translates into an urgent need for the under-
standing and analysing of the country’s biodiversity to
overcome conservation challenges.
It is essential to first assess the current biological situ-
ation, including the ecological potential of Turkey, to
develop effective conservation strategies to address the loss
of biodiversity (Ricotta et al. 2000; Bohn et al. 2007;
Rosati et al. 2008). Turkey’s vegetation is at high risk due
to the northward expansion of arid areas, possibly resulting
in increased water stress and desertification (Gao and
Giorgi 2008; Anav and Mariotti 2011). A change in veg-
etation may also affect regional climates as well as radia-
tion budgets through the modification of surface albedo and
the hydrological cycle in terms of evapotranspiration,
The online version is available at http://www.springerlink.com
Corresponding editor: Hu Yanbo.
&Olgu Aydin
drolguaydin@gmail.com
1
GeoZentrum Nordbayern, University Erlangen-Nu
¨rnberg,
91054 Erlangen, Germany
2
Department of Geography, Faculty of Humanities, Ankara
University, 06100 Ankara, Turkey
123
J. For. Res.
https://doi.org/10.1007/s11676-018-0855-7
Author's personal copy

precipitation and runoff (Findell et al. 2007; Sen et al.
2013). While there are existing protected area networks in
place for conservation purposes, these do not necessarily
provide an adequate representation of the biodiversity of
the country and hence protection (Anav and Mariotti 2011;
Cox and Underwood 2011).
The concept of ‘‘potential natural vegetation’’ (PNV),
first introduced by Tu
¨xen (1956), refers to the final state of
vegetation that ‘‘would become established if all succes-
sion sequences were completed without interference by
man under present climatic and edaphic conditions (in-
cluding those by man)’’ (Mueller-Dombois and Ellenberg
1974, p. 422). Initially developed for vegetation mapping
purposes in cultural landscapes (Zerbe 1998; Bastian 2000;
Blasi et al. 2000; Carranza et al. 2003; Hemsing 2010),
PNV integrates abiotic factors and phytogeographic infor-
mation combined with structure, dynamics and ecology of
plant communities, and is widely used in landscape plan-
ning and management (del Rio et al. 2005) to develop
climate change scenarios (Lexer et al. 2002), to define
biogeographical classifications at regional levels (Gallizia
Vuerich et al. 2001), and to determine conservation prior-
ities (Anav and Mariotti 2011). Recent developments
regarding the concept of PNV have led to the inclusion of
additional environmental parameters such as topographic
and edaphic features, as compared to the original focus on
a climate-constrained analysis (Gallizia Vuerich et al.
2001). A PNV map can therefore be generated using
existing vegetation as a reference point for potential dis-
tribution at similar sites where such vegetation is absent
(Bryn 2008).
Franklin (2009) provides an extensive description of the
predictive models used for mapping of species distribution.
Traditionally, PNV models were generated on expert-based
manual modelling (EMM), but as a result of the develop-
ments in GIS tools, new methods such as rule-based
envelope modelling (RBM), statistical techniques, and
machine learning methods have been implemented in the
construction of habitat ecological models, including PNV
(Franklin 2009; Hemsing and Bryn 2012). These ‘‘modern’’
methods, coupled with readily available high resolution
species occurrence data, i.e., actual vegetation maps, and
environmental data has made it possible to easily carry out
predictions concerning the potential distributions of dif-
ferent species in a specific geographical area.
RBM divides the geographical area of interest according
to specific attributes. Hemsing (2010), used elevation fea-
tures to separate the study area and hence categorise actual
vegetation types according to elevation, after which other
‘‘rules’’, such as soil properties and human influence were
included in the model to identify specific regions with, for
example, alpine vegetation, wetlands and peatland forests
or anthropogenic vegetation. This, however, introduces
some subjectivity into the study as modelling rules have to
be manually set. A way to reduce this bias is to include
statistical approaches in the modelling of species distribu-
tion. Several methods are employed for this purpose,
namely generalised linear models (GLM), generalised
additive models (GAM), and more recently, multivariate
adaptive regression splines (MARS). GLM, especially
logistic regression which has been used in this study, is
considered to be one of the best- established frameworks
for species distribution modelling, both for flora and fauna
(Franklin 2009). However, even statistical inference
includes some subjectivity, as the distribution of the data is
decided by the analyst and the parameters used for the
model are estimated from the data. Machine- learning
methods, such as Artificial Neural Networks, maximum
entropy (Maxent), or genetic algorithms, on the other hand,
employ several algorithms to inductively determine map-
ping functions or classification rules after a training period
with the data available. Despite their effectiveness, con-
sidering the complex classification problems, there is a
steep learning curve when it comes to machine- learning
methods. Also, some approaches, such as Maxent, can only
be implemented using specific software developed for the
purpose of species distribution modelling.
The purpose of this study is to use the multinomial
logistic regression (MLR) to develop a model for the
reconstruction of the potential natural vegetation of Tur-
key, based on actual vegetation using climate data as pre-
dictors. This would be a starting point for the assessment of
the ecological problems facing by the country. The MLR,
based on the GLM, simulates the probability of different
biomes occurring in Turkey, using elevation, climatologi-
cal and actual natural vegetation data. The influence of
climate on vegetation dynamics has drawn increasing
attention over the past years (Serkercioglu et al. 2011;
Atalay et al. 2014). This method is free of any dynamic
biases or subjective corrections as well as being faster
compared to others, which would concurrently offer an
independent reference for vegetation models (Levavasseur
et al. 2013).
Materials and methods
Study area
Turkey is located between 36–42N latitude and 26–45
E longitude and includes mountain ranges, plateaus with
deep split river valleys with mountains of volcanic origin,
old lake and marine sediments, delta plains expanding at
the mouths of major rivers, as well as tectonic basins
covered with rich alluvium soils (Koc¸man 1993). The
country is unevenly elevated, with an average elevation of
N. B. Raja et al.
123
Author's personal copy

1132 m in mountainous areas. More than 55% of the
landscape is classified as high-field areas (Fig. 1a). The
North Anatolian Mountains and the Taurus Mountains
constitute the ranges along the northern and southern
coasts, respectively. The thrusting effect between these
mountain ranges has led to uneven terrain in the Eastern
Anatolia Region. The Central Anatolia Region, on the
other hand, consists of large, high- elevated plains created
by these mountain ranges that extend towards the Aegean
and Marmara seas. The occurrence of the multi-climatic
regimes, as previously mentioned, is due to these factors
amongst others. The slopes of the mountain ranges
Fig. 1 a Elevation map of Turkey, bdistribution network of meteorological stations across Turkey
A reconstruction of Turkey’s potential natural vegetation using climate indicators
123
Author's personal copy

overlooking the seas receive abundant long-term heavy
precipitation, while the interior slopes receive less with
yearly increases in temperature differences, suggesting that
proximity to the coast represents the first and foremost
effect on continentality (Koc¸man 1993). Figure 1b pro-
vides an overview of the distribution of the meteorological
stations used in this study. A relatively dense network of
stations is found in the Aegean region, compared to sparser
networks in South East Anatolia, South Marmara, and
Central Anatolia. Due to Turkey’s multi-climate regime,
and a heterogeneous and inadequate meteorological net-
work, evaluating the potential natural vegetation (PNV)
based on climate variables can be challenging and limited.
Vegetation
The current distribution of vegetation is based on field and
satellite surveys by the Turkish Ministry of Forestry
(Fig. 2). Despite forest cover having decreased from
60–70% to 26–27%, and steppe cover increased from
10–15% to 24% during the period 1965 and 1988, 90% of
the forest cover is still considered ‘‘original’’ or ‘‘natural’’
(Atalay 1994;C¸ olak and Rotherham 2006; Serkercioglu
et al. 2011). Approximately 42% of the total forested area
consists of conifers, 53% broad-leaved deciduous and 5%
of mixed coniferous and broad-leaved forests (Atalay et al.
2014). The main distribution of broad-leafed deciduous
forests is around the Black Sea coastal belt due to the rich
organic content of the soils and mull humus developed as a
result of the humid and temperate conditions. Amongst the
approximately 10,000 plant species found in Turkey, half
are concentrated in the Black Sea region. Broad-leaved
deciduous forests are located up to 1200 m a.s.l. and, in
some areas, up to 1500 m a.s.l. in this region. There is a
transition to coniferous forests on the north slopes of the
North Anatolian Mountains moving from the Black Sea
region towards Central Anatolia between 1000 and 2000 m
a.s.l. At 1000–1500 m a.s.l., the transition zone between
broadleaved deciduous and coniferous forests is a mixed
forest with several broadleaf species, after which, at
1500–2000 m a.s.l. pure coniferous forests can be found.
These are located around the Mediterranean and the Black
Sea regions from sea level up to 2000 m a.s.l. (Serker-
cioglu et al. 2011; Atalay et al. 2014). The southern slopes
of the Taurus Mountains along the coastal belt of the
Mediterranean consist mainly of coniferous species.
Coniferous forests are also found in mountainous areas of
the Marmara and Aegean regions between 1000 and
1500 m a.s.l. Most of the native vegetation in inner Ana-
tolia has been degenerated and mostly replaced with steppe
vegetation. Remnant native vegetation consists of broad-
leaved species distributed between 1200 and 1500 m a.s.l.
Conifers can be found at 1500 m a.s.l. upwards. Shrublands
and steppes covering around 10% of the Anatolian plains
are common in the Mediterranean phytogeographical
regions, extending up to 400 m s.a.l. in the Marmara
region, 600 m s.a.l. in the Aegean region and 1000 m s.a.l.
in the Mediterranean region.
Fig. 2 Current distribution of vegetation in Turkey
N. B. Raja et al.
123
Author's personal copy

Climate
Turkey’s climate variability has been thoroughly studied
by Koc¸man (1993), Tu
¨rkes¸(2010) and Aydin and C¸ic¸ek
(2015). Turkey is located between mid-latitude temperate
and subtropical climate zones and possesses a diverse
topography. It is characterised as a Mediterranean macro-
climate with several climate regimes (Iyigun et al. 2013).
In addition, the topographic effects associated with the
mountainous terrain greatly influence climate variability.
Precipitation and temperature are dependent on the fre-
quency and intensity of circulating depressions, high
atmospheric western winds, and the polar jet stream as well
as the location of Mediterranean and Polar fronts. Fur-
thermore, anticyclonic pressure patterns such as the Azores
high-pressure system and the Siberian anticyclone are also
important. Variations as a result of local conditions cause
the emergence of differences in spatial precipitation and
temperature (O
¨lgen 2010) as seen in Fig. 3. The presence
of the North Anatolian Mountains in the north and the
Taurus Mountains in the south protects Central Anatolia
from coastal and marine effects—the primary factor
affecting continentality in Turkey (Koc¸man 1993). Annual
mean temperatures vary from 3.6 to 20.1 C depending on
location and elevation. Annual mean precipitation overall
is around 648 mm with an annual variation of
295–2220 mm. Central Anatolia on average receives less
precipitation than other areas; the Mediterranean and Black
Sea regions receive the highest annual precipitation. While
the areas facing the coasts receive abundant precipitation,
interior regions are closed off by the mountains and receive
less precipitation. Other geographical factors such as slopes
and air pressures also significantly affect precipitation
distribution.
Methodology
ArcGIS software was used for preparing data and visual-
ising the results while all statistical computations were
carried out using the R statistical programme (R Devel-
opment Core Team 2016). The grid size was 90 990 m
based on the vegetation data from the Ministry of Forestry.
All interpolations were carried out using the same grid size.
Explanatory variables
Vegetation distribution in Turkey according to Atalay et al.
(2014) is dependent on several factors: topography, aspect,
slope, precipitation, temperature, relative humidity, wind,
and geology. Levavasseur et al. (2013) also consider other
variables such as the diurnal temperature cycle, the number
of wet days, number of frost days, sunshine duration, wind
speed/intensity and total cloudiness. As it would be
computationally intensive to consider all these variables, a
correlation test was applied to the available data to deter-
mine which predictors influenced the vegetation (Table 1).
The following variables had the strongest correlation
with the presence of different types of vegetation
(Table 1):
•Surface air temperature (C)
•Total mean annual precipitation (mm)
•Relative humidity (%)
•Total cloudiness (%)
•Elevation (m)
The data used in this study spans 40 years from 1975 to
2014 and is based on climate data from 202 meteorological
stations throughout Turkey (Fig. 1b). The data were
transformed into the correct format by spatially interpo-
lating the data using the Ordinary Kriging (OK) method to
allow use in the MLR model. The OK method was chosen
as the interpolation method as it surpasses other geosta-
tistical methods, and the OK method is able to assess the
spatial correlation structure between observed values, and
provide results with minimum variance value (Aydin and
C¸ic¸ek 2015). The OK method allows the statistical gener-
ation of optimal spatial predictions (Cressie 1993) using
the weighted averages of the observations:
^
Ys
0
ðÞ¼
X
ns
i¼1
ws
i
ðÞys
i
ðÞ ð1Þ
where ns is the total number of observed points, ^
Ys
0
ðÞthe
interpolation value at location s0,ys
1
ðÞ...ys
i
ðÞobserved
values at locations s1to siand ws
1
ðÞ...ws
i
ðÞthe weights
generated from a model of the spatial correlation structure
of the data, usually a valid variogram model fit using the
observations. The OK interpolations are in Fig. 4.
The multinomial logistic regression (MLR) model
MLR based on generalized linear models (GLM) simulates
the probability of different biomes occurring using eleva-
tion, climatological and actual natural vegetation data. The
GLM examines the relationship between dependent vari-
ables and predictors as well as unifies several statistical
regression models including the Poisson and logistic
regressions. The classical binary logistic regressions com-
putes the occurrence probability of a binary event (e.g.,
vegetation or bare soil), which can take continuous values
between 0 and 1, by fitting data to a logistic function.
However, the logistic regression in this study is used in its
multinomial form (Eq. 2) to compute the occurrence
probability of four biomes, namely, coniferous forests,
broad leaved deciduous forests, mixed forests and shrub-
lands. The MLR estimates the occurrence probability of the
A reconstruction of Turkey’s potential natural vegetation using climate indicators
123
Author's personal copy

explained variables (Y, the vegetation) for each biome j,
taking into account numerical explanatory variables (Xk):
log PY
i¼jðÞ
PY
i¼rðÞ

¼b0jþX
p
k¼1
bk;jXi;k;8 6¼ rð2Þ
where PY
i¼jðÞis the probability of the jth biome, b0is
the intercept for the jth biome, bkare the regression
coefficients for the jth biome, pis the number of predictors,
and iis the grid-cell. To use MLR, a reference category r
(here the coniferous species) should be chosen. The j-1
Fig. 3 Spatial distribution of atotal mean annual precipitation (mm) and bmean annual temperature of (C) for 1976–2010 interpolated using
the Ordinary Kriging method as per the specifications of Aydin and C¸ic¸ek (2015)
N. B. Raja et al.
123
Author's personal copy

relationship are obtained after which the occurrence prob-
abilities of the reference biome can be deduced in each
grid-cell iwith Pm
j¼1PY
i¼jðÞ¼1 (considering mbiomes
including r). An algorithm of likelihood maximisation was
used to compute the MLR in R.
Error metrics
The goodness of model fit was tested with the pseudo-R
2
,
the Nagelkerke 
R2as well as the Brier Score and the area
under curve (AUC) of the receiver operating characteristic
(ROC).
The Nagelkerke 
R2can be interpreted as the proportion
of explained variation in the regression model and there-
fore can be used as a measure of success in predicting the
dependent variable from the independent variables
(Nagelkerke 1991). Nagelkerke 
R2is defined as:

R2¼R2
max R2
ðÞ ð3Þ
where
R2¼1L0ðÞ=L^
b
no
2=nð4Þ
and
max R2

¼1L0ðÞ
2
=nð5Þ
where L^
b
and L0ðÞrepresents the fitted models with the
explanatory variables and the ‘‘null’’ model fitted with only
the intercept.
Table 1 Results of correlation test between vegetation and chosen
parameters
Parameter Correlation coefficient
Elevation 0.358
Precipitation -0.254
Temperature -0.241
Humidity -0.175
Cloudiness -0.168
Wind speed -0.103
Number of cloudy days 0.093
Number of wet days -0.07
Number of frost days 0.06
Number of sunny days -0.021
Diurnal Temperature 0.007
Fig. 4 Kriging results of the chosen climatic parameters namely, aprecipitation; btemperature; chumidity; dcloudiness
A reconstruction of Turkey’s potential natural vegetation using climate indicators
123
Author's personal copy

The Brier Score provides a means of assessing the rel-
ative accuracy and generates the ‘‘error rate’’ of the logistic
regression model (Brier 1950). The formulation of the
Brier score is:
BS ¼1
NX
N
t¼1
ftot
ðÞ
2ð6Þ
where ftis the probability forecasted in Eq. (2), otis the
observed outcome of the event at instance t and N is the
number of forecasting instances. The Brier Score ranges
from 0 to 1 and measures the mean squared difference
between the predicted probability and the observed out-
come. Therefore, a completely accurate forecast would
generate a value of 0.
The ROC analyses the relationship between sensitivity
and specificity of a binary classifier, in this instance the
occurrence of the different biomes. Sensitivity refers to
the proportion of positives correctly classified, i.e., the
proportion of vegetation occurrences correctly identified
when comparing predicted probability and observed
values, whereas specificity measures the proportion of
negative correctly classified, the proportion of vegeta-
tion-free regions correctly identified (Flach 2010). Con-
ventionally, the ROC involves plotting sensitivity against
one minus specificity. The AUC evaluates the perfor-
mance of the model by taking into account the rank order
of the scores. AUC is measured on a scale of 0 and 1.
AUC ¼1 is achieved if every positive is scored higher
than every negative, thus showing a completely accurate
forecast.
The Wald (1941) statistics, which evaluate the signifi-
cance of each coefficient bjin the model, were also cal-
culated by:
Wj¼bj
SEbj
!
2
ð7Þ
where Wjrepresents the Wald test and SEbjthe standard
error of coefficient bjfor the independent variable j.
Results
The logistic regression coefficients are shown in Table 2.
Based on these values, temperature was the first and fore-
most factor influencing vegetation followed by cloudiness.
Elevation, on the other hand, appears to have a negligible
effect.
Following the multinomial logistic regression, the
probabilities of the four biomes occurring were generated.
For simplicity, the dominant biome with the maximum
occurrence probability ([0.5) was considered the
dominant biome in regions where there were overlapping
biomes. Figure 5shows a composite map of the predicted
biomes. Broadleaf forests are dominant along the Black
Sea belt while shrublands and mixed forests are more
likely to occur within Central and Eastern Anatolia. Along
the Aegean and Mediterranean coasts, coniferous forests
are widespread. The actual and potential presence of the
biomes were also compared (Fig. 6). While the regression
model projected the presence of broadleaf forests only in
the Black Sea belt, actual vegetation surveys show that
they are widespread in Turkey. Similarly, shrublands were
predicted to be found only in Central and Eastern
Anatolia.
Previous studies provide several statements concerning
the performance evaluation of logistic regression models,
namely: (1) the significant Wald statistic for independent
variables should be less than 0.05 (Bai et al. 2010; Dahal
et al. 2012); (2) the Nagelkerke 
R2should be greater than
0.2 (Clark and Hosking 1986; Ayalew and Yamagishi
2005); and, (3) the Brier Score should be less than 0.25
(Steyerberg et al. 2010). Based on these criteria, the
regression model generated with 
R2of 0.251, a Brier Score
of 0.152 and Wald statistics of \0.05 for all predictors
except for the temperature parameter of mixed forests, is
considered satisfactory (Table 2).
It can therefore be concluded that the logistic regression
model provides an acceptable representative of the PNV
for Turkey.
Discussion
Overall, the model is consistent with previous studies on
vegetation in Turkey (C¸ olak and Rotherham 2006; Serk-
ercioglu et al. 2011) and shows that temperature was the
most influencing factor on vegetation. Vegetation distri-
bution follows a similar distribution as the country’s
macroclimates. Biltekin (2010), based on Erinc¸(1996)’s
study, identifies four main macroclimate types and nine
sub-macroclimate types (Table 3).
The coniferous and broadleaf forests (Fig. 5) are also
consistent with the potential forest land identified by C¸ olak
and Rotherham (2006). Their study identified the wet cli-
mate of the Black Sea region as the main driving factor for
vegetation in this region compared to the rest of Turkey
where vegetation is mainly determined by temperature and
humidity. The evergreen forests, in the form of broad
leaved forests, are mainly affected by the heavy precipi-
tation and frequent mists occurring year round in the Black
Sea Region. As such, a decrease in broadleaf forests and an
increase in mixed forests is observed from east to west. The
Marmara Region is the only region, aside from the Black
N. B. Raja et al.
123
Author's personal copy

Sea Region where broadleaf forests are found. This can be
attributed to low evaporation in this region, which is suit-
able for this type of vegetation.
Mixed forests and shrublands are mostly found in
Central Anatolia within the Anatolian steppe climate. The
low summer humidity is a limiting factor, favouring mostly
Table 2 Performance results
and logistic regression
coefficients of the model
Regression coefficient SE Wald Chi square ZðÞ Pr [ZjjðÞ
Broad leaved forests
(Intercept) -6.239 0.118 -52.699 0.000
Elevation 0.000 0.000 3.201 0.001
Precipitation 0.008 0.000 87.518 0.000
Temperature -0.086 0.006 -13.346 0.000
Humidity -0.001 0.000 -1.977 0.048
Cloudiness 0.048 0.003 15.878 0.000
Coniferous forests
(Intercept) -8.397 0.100 -83.644 0.000
Elevation 0.001 0.000 37.920 0.000
Precipitation 0.005 0.000 64.068 0.000
Temperature 0.170 0.005 34.150 0.000
Humidity 0.008 0.000 22.297 0.000
Cloudiness -0.014 0.004 -3.164 0.002
Mixed forests
(Intercept) -6.598 0.122 -53.891 0.000
Elevation 0.000 0.000 5.016 0.000
Precipitation 0.006 0.000 66.646 0.000
Temperature 0.001 0.006 0.167 0.868
Humidity 0.003 0.000 6.988 0.000
Cloudiness 0.034 0.003 10.751 0.000
Shrubs
(Intercept) -4.227 0.075 -56.663 0.000
Elevation 0.001 0.000 64.346 0.000
Precipitation 0.001 0.000 16.390 0.000
Temperature 0.058 0.004 14.368 0.000
Humidity 0.003 0.000 11.328 0.000
Cloudiness 0.013 0.004 3.383 0.001
Fig. 5 Composite map of the
predicted biomes in Turkey
A reconstruction of Turkey’s potential natural vegetation using climate indicators
123
Author's personal copy

herbaceous and suffruticose flora with the exception of
some conifers, thus explaining the presence of mixed for-
ests and shrubs in the PNV mostly in Continental Anatolia
of this study. Eastern Anatolia, which also experiences a
continental climate with arid or semi-arid characteristics,
also consists mostly of mixed forests and shrubs. Southeast
Anatolia, on the other hand, consists mainly of coniferous
forests, attributed to high temperatures (above 30during
summer), of the Southeast Anatolian steppe. Similarly,
high temperatures along the Aegean and Mediterranean
coasts account for the presence of mostly coniferous forests
in this region.
On the other hand, the presence of coniferous forests
along the Black Sea belt may be explained by the change of
elevation from the coast to further inland. At lower alti-
tudes, the forest is mainly broadleaved but at higher levels,
the presence of conifers, resembling those in the mountains
in Central Europe, increase to become dominant. Accord-
ing to C¸ olak and Rotherham (2006), Central Anatolia
remains an enigma as a result of the difficulties in identi-
fying several genera playing an important role.
Similarly, Serkercioglu et al. (2011)’s reconstruction of
the ‘‘original vegetation cover’’ of Turkey shows broad-
leaved forests in the Black Sea region and coniferous in the
Mediterranean region. They identified Central Anatolia to
be a mix of steppes, shrublands, pine forests (coniferous)
and oak (deciduous), hence consistent with the results of
this study. They highlight the presence of coniferous for-
ests in the Black Sea region but state that the difference
between those in the Mediterranean region range from sea
level to 2000 m while those in the Black Sea region are
mainly found between 1000 and 1200 m. The steppe
grasslands located from Central to Southeast Anatolia are
Fig. 6 Actual and potential presence of different biomes in Turkey
Table 3 Macroclimate characteristics of Turkey and corresponding vegetation
Macroclimate Sub-macroclimate Characteristics Vegetation type
Steppe Climate Ia—Anatolian steppe
climate
Hot summers (20–25 C) and cold winters (0–3 C) Mixed forests
and shrubs
Ib—Southeastern Anatolian
steppe climate
Hot summers ([30 C) and cold winters (0–5 C) with high evaporation
observed (1000–2000 mm annually)
Coniferous
Black Sea
Climate
IIa—Eastern Black Sea
climate
High precipitation all year round, with temperate winters Broad-leaf
IIb—Central Black Sea
climate
Moderate precipitation all year round Broad-leaf and
mixed forests
IIc—Western Black Sea
climate
Low precipitation with cooler summers and winters Broad-leaf and
mixed forests
Mediterranean
Climate
IIIa—Mediterranean
climate
Very hot summers, small amount of snow observed in winter Coniferous
IIIb—Marmara region
climate
Very cold winters, low evaporation Broad-leaf and
mixed forests
Eastern
Anatolian
Climate
IVa—All seasons with
precipitation
Continental climate regime Mixed forests
and shrubs
IVb—Arid summer type High precipitation in winter and spring, low precipitation and high
evaporation during summer and winter
Mixed forests
and shrubs
N. B. Raja et al.
123
Author's personal copy

mainly due to the dry climate, are mostly attributed to the
overharvesting of timber over previous millennia.
Due to the model specifications, only the main regions
were identified where the different biomes were located.
For example, broadleaf forests and shrublands were located
in specific regions. This is attributed to the cut-off value
(0.5) used for maximum probability of occurrence. This
suggests that the model may have determined the presence
of different biomes in other regions of Turkey, but deter-
mined that they were less likely to occur there. Habitat
models are mostly based on the assumption that the species
composition of the vegetation (plant community, vegeta-
tion type) depends on habitat conditions (environmental
predictors). This relation is called ‘‘sociological determin-
ism’’ (Braun-Blanquet 1928). Other concepts such as
mosaic cycles and positive feedback models should also be
taken into consideration (Fischer et al. 2013). The actual
state of the vegetation not only depends on environmental
indicators but also on history and random events in the
past. The importance of other variables can thus only be
evaluated after they are incorporated in a model.
The above highlights the necessity of investigating
vegetation beyond the basic classification used in this
study. The vegetation types represented by the model
captured only regional-scale patterns, hence finer-scaled
variations that are caused, for example, by a mosaic of soil
types, are not represented. It also excluded vegetative areas
such as wetlands, agricultural areas and heavily managed
forests. As stated by Serkercioglu et al. (2011), there is a
distinct difference between coniferous forests in the Black
Sea Region and those in the Mediterranean Region. In
addition, vegetation in Inner Anatolia can only be under-
stood through a local investigation of the different species.
C¸ olak and Rotherham (2006) noted that there may be
significant unrecorded and unrecognised biodiversity in
Turkish forests, highlighted by the poor resolution of the
current vegetation data, especially in the Central Anatolia
region. The results also indicate the potential for forest
management as they show the optimal environments for
different vegetation types. Forest reserve managers could
therefore be able to introduce tree species in certain loca-
tions to allow for near-natural regeneration. However the
results should only be treated as scenarios and not pre-
dictions as such models cannot include all aspects of spe-
cies ecology. Furthermore, there exists considerable
uncertainty in climate projections. Therefore, any adapta-
tion strategy must be flexible, dynamic and responsive to a
changing climate and new research on its impact.
Conclusion
With forests subject to expansive deforestation, it is clear
that effective conservation management in an urgent
necessity. Forestry activities, such as ‘‘close-to-nature’’
silvicultural operations and restoration, can only be effec-
tive through understanding the natural vegetation and can
assist in the prediction, at a landscape scale, of variations,
deviations or conservation risks. This understanding can be
achieved through an analysis of the potential natural veg-
etation of any region, which shows the final state of veg-
etation by considering the current vegetation without any
future anthropogenic influences. The aim of this study was
to use multinomial logistic regression to develop a model
for the reconstruction of the potential natural vegetation
based on actual vegetation using climate data as predictors.
While the actual vegetation may not always mean native, in
the case of Turkey the current data is representative of
approximately 90% of the remaining forests which are
near-natural, semi-natural or only partly altered (C¸ olak and
Rotherham 2006).
Through this study, regions needing data to improve the
potential natural vegetation distribution, such as Central
Anatolia, can be targeted. The modelled PNV by the
multinomial logistic regression model highly depends on
the abundance and geographical distribution of data points.
Therefore, if a biome is absent or over/underrepresented,
this will have a significant impact on the model.
Nonetheless, the model appears consistent with current
climatic patterns in Turkey. However, this should be taken
with caution as climatic indicators alone are not sufficient
to distinguish a dominant biome. Other indicators such as
soil properties should also be included as predictors in the
model. Human influences are also important factors to
consider. The multinomial logistic regression model is
mainly based on climatological data between 1974 and
2013, impacted by human activities through climate
change. As such, it could be relevant to also calibrate the
model using data for a longer time period going back to the
Palaeolithic era when land-use was limited (Levavasseur
et al. 2013).
Accounting for all observations and statistical results,
the multinomial logistic regression model provided a
realistic potential natural vegetation distribution for Tur-
key. The method is a fast and robust alternative in vege-
tation modelling with several advantages. The potential
natural vegetation map is: (1) directly, and only, based on
vegetation and climatological data; (2) not subjective and
independent of any vegetation model; and, (3) easily
updated as additional data is available.
A reconstruction of Turkey’s potential natural vegetation using climate indicators
123
Author's personal copy

References
Anav A, Mariotti A (2011) Sensitivity of natural vegetation to climate
change in the Euro-Mediterranean area. Clim Res 46(3):277–292
Atalay I (1994) Vegetation of Turkey. Aegean University Press,
Izmir, pp 17–30
Atalay I, Efe R, O
¨ztu
¨rk M (2014) Ecology and classification of forests
in Turkey. Procedia Soc Behav Sci 120:788–805
Ayalew L, Yamagishi H (2005) The application of GIS-based logistic
regression for landslide susceptibility mapping in the Kakuda-
Yahiko Mountains, Central Japan. Geomorphology
65(1–2):15–31
Aydin O, C¸ic¸ek I
˙(2015) Geostatistical interpolation of precipitation
in Turkey. Lambert Academic Publishing, Saarbrucken,
pp 195–197
Bai SB, Wang J, Lu
¨GN, Zhou PG, Hou SS, Xu SN (2010) GIS-based
logistic regression for landslide susceptibility mapping of the
Zhongxian segment in the Three Gorges area, China. Geomor-
phology 115:23–31
Bastian O (2000) The assessment of landscape and vegetation
changes: the case study upper lusatian heath and pond landscape.
Probl Landsc Ecol 6:31–53
Biltekin D (2010) Vegetation and climate of North Anatolian and
North Aegean region since 7 Ma according to pollen analysis
(Doctoral dissertation). Universite
´Claude Bernard - Lyon I,
France, pp 15–17
Blasi C, Carranza ML, Frondoni R, Rosati L (2000) Ecosystem
classification and mapping: a proposal for Italian landscapes.
Appl Veg Sci 3(2):233–242
Bohn U, Zazanashvili N, Nakhutsrishvili G (2007) The map of the
natural vegetation of Europe and its application in the Caucasus
Ecoregion. Bull Ga Acad Sci 175:112–121
Braun-Blanquet J (1928) Pflanzensoziologie: grundzu
¨ge der vegeta-
tionskunde. Springer, Berlin, p 17
Brier GW (1950) Verification of forecasts expressed in terms of
probability. Mon Weather Rev 78(1):1–3
Bryn A (2008) Recent forest limit changes in south-east Norway:
effects of climate change or regrowth after abandoned utiliza-
tion. Nor Geogr Tidsskr - Nor J Geogr 62:251–270
Carranza ML, Ricotta C, Fortini P, Blasi C (2003) Quantifying
landscape change with actual vs. potential natural vegetation
maps. Phytocoenologia 33(4):591–601
Clark WA, Hosking PL (1986) Statistical methods for geographers.
Wiley, New York, pp 447–500
C¸ olak AH, Rotherham ID (2006) A review of the forest vegetation of
Turkey: its status past and present and its future conservation.
Biol Environ 106B(3):343–354
Cox RL, Underwood EC (2011) The importance of conserving
biodiversity outside of protected areas in Mediterranean Ecosys-
tems. PLOS 6(1):e14508
Cressie NAC (1993) Statistics for spatial data. Wiley, New York,
pp 52–58
Dahal RK, Hasegawa S, Bhandary NP, Poudel PP, Nonomura A,
Yatabe R (2012) A replication of landslide hazard mapping at
catchment scale. Geomat Nat Hazards Risk 3(2):161–192
del Rio S, Penas A, Perez-Romero R (2005) Potential areas of
deciduous forests in Spain (Castile and Leon) according to future
climate change. Plant Biosyst 139:222–233
Erinc¸ S (1996) Klimatoloji ve Metodları[Climatology and its
methods]. Istanbul University Press, Istanbul, p 538
Evrendilek F, Berberoglu S, Karakaya N, Cilek A, Aslan G, Gungor
K (2011) Historical spatiotemporal analysis of land-use/land-
cover changes and carbon budget in a temperate peatland
(Turkey) using remotely sensed data. Appl Geogr
31(3):1166–1172
Findell KL, Shevliakova E, Milly PCD, Stouffer RJ (2007) Modeled
impact of anthropogenic land cover change on climate. J Clim
20(14):3621–3634
Fischer HS, Winter S, Lohberger E, Jehl H, Fischer A (2013)
Improving transboundary maps of potential natural vegetation
using statistical modeling based on environmental predictors.
Folia Geobot 48(2):115–135
Flach PA (2010) ROC analysis. In: Sammut C, Webb GI (eds)
Encyclopedia of Machine Learning. Springer, New York,
pp 869–875
Franklin J (2009) Mapping species distributions. Cambridge Univer-
sity Press, Cambridge, pp 113–203
Gallizia Vuerich L, Poldini L, Feoli E (2001) Model for the potential
natural vegetation mapping of Friuli-Venezia Giulia (NE Italy)
and its application for a biogeographic classification of the
region. Plant Biosyst 135:319–336
Gao X, Giorgi F (2008) Increased aridity in the Mediterranean region
under greenhouse gas forcing estimated from high resolution
simulations with a regional climate model. Glob Planet Change
62(3–4):195–209
Gu
¨ler M, Yomralıog
˘lu T, Reis S (2007) Using landsat data to
determine land use/land cover changes in Samsun, Turkey.
Environ Monit Assess 127(1):155–167
Hemsing LO (2010) GIS-modelling of potential natural vegetation
(PNV): a methodological case study from south-central Norway
(Master’s thesis). Norwegian University of Life Sciences,
Norway, pp 8–13
Hemsing LO, Bryn A (2012) Three methods for modeling potential
natural vegetation (PNV) compared: a methodological case study
from south-central Norway. Nor Geogr Tidsskr - Nor J Geogr
66:11–29
Hoekstra JM, Boucher TM, Ricketts TH, Roberts C (2005) Con-
fronting a biome crisis: global disparities of habitat loss and
protection. Ecol Lett 8:23–29
Iyigun C, Tu
¨rkes¸ M, Batmaz I
˙, Yozgatligil C, Purutc¸uog
˘lu V, Koc¸
EK, O
¨ztu
¨rk MZ (2013) Clustering current climate regions of
Turkey by using a multivariate statistical method. Theor Appl
Clim 114(1):95–106
Koc¸man A (1993) Climate of Turkey. Ege University Faculty for
Literature Publications, Izmir, pp 56–61
Levavasseur G, Vrac M, Roche DM, Paillard D, Guiot J (2013) An
objective methodology for potential vegetation reconstruction
constrained by climate. Glob Planet Change 104:7–22
Lexer MJ, Ho
¨nninger K, Scheifinger H, Matulla C, Groll N, Kromp-
Kolb H, Schadauer K, Starlinger F, Englisch M (2002) The
sensitivity of Austrian forests to scenarios of climatic change: a
large-scale risk assessment based on a modified gap model and
forest inventory data. For Ecol Manag 162:53–72
Mueller-Dombois D, Ellenberg H (1974) Aims and methods of
vegetation ecology. Wiley, New York, p 422
Nagelkerke NJD (1991) A note of a general definition of the
coefficient of determination. Biometrika 78(3):691–692
O
¨lgen KM (2010) Tu
¨rkiye’de yıllık ve mevsimsel yag
˘ıs¸
deg
˘is¸ kenlig
˘inin alansal dag
˘ılımı[Spatial distribution of annual
and seasonal precipitation variability in Turkey]. Aegean Geogr
J 19(1):85–95
R Development Core Team (2016) The R project for statistical
computing. Retrieved from http://www.r-project.org/. Accessed
01 Dec 2016
Ricotta C, Carranza ML, Avena G, Blasi C (2000) Quantitative
comparison of the diversity of landscapes with actual vs.
potential natural vegetation. Appl Veg Sci 3(2):157–162
Rosati L, Marignani M, Blasi C (2008) A gap analysis comparing
natura 2000 vs national protected area network with potential
natural vegetation. Community Ecol 9:147–154
N. B. Raja et al.
123
Author's personal copy

Sen OL, Bozkurt D, Vogler JB, Fox J, Giambelluca TW, Ziegler AD
(2013) Hydro-climatic effects of future land-cover/land-use
change in montane mainland southeast Asia. Clim Change
118(2):213–226
S¸ erkerciog
˘lu CH, Anderson S, Akc¸ay E, Bilgin R, Can O
¨E, Semiz G,
Tavs¸ anog
˘lu C¸ , Yokes¸ MB, Soyumert A, I
˙pekdal K, Sag
˘lam I
˙K,
Yu
¨cel M, Dalfes HN (2011) Turkey’s globally important
biodiversity in crisis. Biol Conserv 144(12):2752–2769
Steyerberg EW, Vickers AJ, Cook NR, Gerds T, Gonen M,
Obuchowski N, Pencina MJ, Kattan MW (2010) Assessing the
performance of prediction models: a framework for traditional
and novel measures. Epidemiology 21:128–138
Tu
¨rkes¸ M (2010) Climatology and meteorology. Kriter Publishing,
Istanbul, pp 379–387
Tu
¨xen R (1956) Die heutige potenzielle natu
¨rlich vegetations als
gegenstand der vegetationskartierung [The current potential
natural vegetation as the object of vegetation mapping].
Pflanzensoziologie 13:5–42
Wald A (1941) Asymptotically most powerful tests of statistical
hypotheses. Ann Math Stat 12:1–19
Zerbe S (1998) Potential natural vegetation: validity and applicability
in landscape planning and nature conservation. Appl Veg Sci
1(2):165–172
A reconstruction of Turkey’s potential natural vegetation using climate indicators
123
Author's personal copy