European Journal of Remote Sensing

ISSN: (Print) 2279-7254 (Online) Journal homepage: https://www.tandfonline.com/loi/tejr20

Land cover classification in Romanian Carpathians

and Subcarpathians using multi-date Sentinel-2
remote sensing imagery

Marina-Ramona Rujoiu-Mare, Bogdan Olariu, Bogdan-Andrei Mihai,

Constantin Nistor & Ionuț Săvulescu

To cite this article: Marina-Ramona Rujoiu-Mare, Bogdan Olariu, Bogdan-Andrei Mihai,

Constantin Nistor & Ionuț Săvulescu (2017) Land cover classification in Romanian Carpathians
and Subcarpathians using multi-date Sentinel-2 remote sensing imagery, European Journal of
Remote Sensing, 50:1, 496-508, DOI: 10.1080/22797254.2017.1365570

To link to this article: https://doi.org/10.1080/22797254.2017.1365570

© 2017 The Author(s). Published by Informa Published online: 29 Aug 2017.

UK Limited, trading as Taylor & Francis
Group.

Submit your article to this journal Article views: 3468

View related articles View Crossmark data

Citing articles: 3 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=tejr20
EUROPEAN JOURNAL OF REMOTE SENSING, 2017
VOL. 50, NO. 1, 496–508
https://doi.org/10.1080/22797254.2017.1365570

Land cover classification in Romanian Carpathians and Subcarpathians using

multi-date Sentinel-2 remote sensing imagery
Marina-Ramona Rujoiu-Marea, Bogdan Olariua, Bogdan-Andrei Mihaib, Constantin Nistorb
and Ionuț Săvulescub
a
University of Bucharest, Faculty of Geography, Simion Mehedinți Doctoral School, Bucharest, Romania; bUniversity of Bucharest, Faculty
of Geography, Department of Geomorphology-Pedology-Geomatics, Bucharest, Romania

ABSTRACT ARTICLE HISTORY

In this article, we processed Sentinel-2 images in order to obtain high accuracy land cover Received 15 January 2017
maps for two complementary study areas. The first is represented by the Romanian Revised 3 July 2017
Subcarpathians, a hilly highly fragmented area with heterogeneous land cover pattern and Accepted 7 August 2017
the second by Romanian Carpathians, a mountain area with homogenous structure of KEYWORDS
vegetation cover. The aim of this article is to evaluate the potential of a singledate in Sentinel-2; land cover; multi-
comparison with multi-date images for which a complete calibration and an iterative process date images; supervised
of supervised classification using Maximum Likelihood (ML) and Support Vector Machine classification; support vector
(SVM) algorithms were applied for the both study areas. The results show that in the case machine; maximum
of Subcarpathian area, the SVM classification on multi-date images has better accuracy due to likelihood
high complexity of the land cover pattern and spectral similarities between classes, while in
the Carpathians, the ML returns good accuracy, consequence of high spectral separabilities
between compact features. The validation process based on ground reference data shows
good accuracies, about 92.41% for the Subcarpathians and 98.65% for the Carpathians. It is
clearly noticed that the land cover pattern determines the use of different algorithms and the
multi-date images enhance the overall accuracy of the classification.

Introduction et al., 2008) and GLC2000 (Bartholomé & Belward,

2005; Hansen, Defries, Townshend, & Sohlberg,
Land cover maps still remain one of the first products
2000).
from remote sensing image analysis in Earth
Land cover was also the subject of many scientific
Observation application projects (Chuvieco, 2016).
approaches, including change detection analysis at
Since the end of the 1970s and 1980s, the US
different scales, from European to National levels.
Geological Survey Land Use Land Cover
(Feranec, Hazeu, Christensen, & Jaffrain, 2007;
Classification system (Anderson, Hardy, Roach, &
Feranec, Hazeu, Jaffrain, & Cebecauer, 2007;
Witmer, 1976) and later the European
Feranec, Jaffrain, Soukup, & Hazeu, 2010)
Environmental Agency’s Corine Land Cover
Remote sensing data selection for land cover
(Büttner, Feranec, & Jaffrain, 2000) vector data
mapping need to be adapted to the objectives of
(1990–2012) provided the first data coverage at
mapping and to the study area features (Campbell
National and Continental levels, based on the inter-
& Wynne, 2011; Robin, 1998). Romanian
pretation of different available imagery (Landsat, IRS,
Carpathians and Subcarpathians landscapes feature
SPOT etc.). Food and Agriculture Organization of the
a complex land cover pattern, with an average
United Nations started in 1990 the development of
minimum mapping unit which cannot be reached
the Land Cover Classification System (LCCS), first
with medium resolution imagery like Landsat
evaluated in Africa (Di Gregorio & Jansen, 1997),
Operational Land Imager (OLI) (Rujoiu-Mare &
and later implemented as a software platform in
Mihai, 2016).
2000, LCCS(v1) (Di Gregorio & Jansen, 2000).
The recently available Sentinel-2 MultiSpectral
LCCS and the Land Cover Meta-Language were
Instrument (MSI) imagery represents a new oppor-
adopted as standards in 2009 and 2012 (ISO-19144-
tunity in finer scale land cover mapping, especially
1, 2009; ISO-19144-2, 2012). There can also be men-
for the most complex landscapes, with higher
tioned the Global level land cover databases, as GLC-
degree of fragmentation. It is the typical case of
SHARE (Latham, Cumani, Rosati, & Bloise, 2014),
Romanian Carpathians and Subcarpathians, with a
GlobeLand30 (Chen et al., 2015), GlobCover (Arino
diversified landscape pattern, introduced by a high

CONTACT Marina-Ramona Rujoiu-Mare rujoiumarina@yahoo.com University of Bucharest, Faculty of Geography, Simion Mehedinți Doctoral
School, Bucharest 010041, Romania
© 2017 The Author(s). Published by Informa UK Limited, trading as Taylor & Francis Group.
This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits
unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

anthropogenic transformation of the primary land
cover: forest fragmentation, deforestation, intensive
grazing, fragmented pattern of agricultural plots
around villages and town built-up areas. The 13
spectral bands of Sentinel-2 MSI sensor with fine
spatial resolution (10, 20, 60 m) (Drusch et al.,
2012) and spectral resolution and several near-
infrared wavelength intervals (Delegido, Verrelst,
Alonso, & Moreno, 2011) enable an advanced
image analysis and land cover mapping at local
levels.
The aim of our article is to evaluate available
Sentinel-2 MSI imagery in mapping land cover at local
scale, after the complete calibration of the data sets from
complementary areas: Prahova Subcarpathians, a hilly
area with a highly complex land cover pattern, frag-
mented by valley corridors (about 200–1100 m a.s.l. –
above sea level) and Bucegi Massif (700–2500 m a.s.l.),
featuring all the main vegetation zones in the
Carpathians and an intensive anthropogenic influence
by traffic and tourism.
The second objective of our study is to evaluate
the use of multi-date images for the classification in
order to be able to distinguish the land cover fea-
tures with similar spectral response and to decrease
the classification errors. Old land cover maps gen-
erated from a single-date image have low thematic
accuracies, especially where the feature classes were
complex and heterogeneous. This technique of
using multi-date images in classification was used
and assessed in many recent studies (Parmentier &
Eastman, 2014; Sexton, Urban, Donohue, & Song,
2013; Shao, Lunetta, Wheeler, Iiames, & Campbell,
2016), and it had been demonstrated that maps
generated using temporal series approaches have
higher accuracies than using single image methods
(Gomez, White, & Wulder, 2016; Khatami,
Mountrakis, & Stehman, 2016). The two main fac-
tors that influence the classification accuracy and
that can be controlled by the analyst are the input
data and the classification algorithm (Khatami
et al., 2016). The multi-date imagery could be
defined as the use of multiple images for a single
classification images (Khatami et al., 2016). Using
multi-date images in classification instead of single-
date data, multiple spectral variables are captured
and used for an optimal detection and discrimina-
tion between the land cover types. Furthermore, it
has been demonstrated that including multiple
spectro-temporal variables in the analysis improves
the classification overall accuracy with 6.9%
(Khatami et al., 2016) and also increases the feature
space (Gomez et al., 2016). We applied this techni-
que to different areas with various complexities, in
order to produce a single-date land cover map from
multiple observations acquired within a year.
EUROPEAN JOURNAL OF REMOTE SENSING 497
Materials and methods
Study areas
For this study, we have chosen two areas from
Romanian Subcarpathians and Carpathians. These
regions have different land cover features, and for
this reason, it was necessary to use distinct algorithms
and processing techniques for classifying the images.
The Subcarpathian region between Prahova (West)
and Teleajen (East) Valleys is limited by the Ploiești
piedmont Plain on the southern part and Baiu and
Grohotiș Mountains on the northern part (Figure 1).
The average altitude of this area is about 616 m,
ranging from 185 to 1047 m (Vârful Frumos Peak).
It is characterized by complex and different land
cover features with natural processes, such as erosion
and landslides, and also by human activities, includ-
ing quarries, mining and oil gas platforms from the
last 200 years. This complex of heterogeneous land
cover features often generates errors in classifications
because of the spectral similarities of some classes, as
well as bare land, built-up areas and croplands or
deciduous forest and sparsely vegetated areas.
Bucegi Mountains are situated in the Southern
Carpathians, West from Prahova Valley (Figure 1).
The altitude of the massif varies from 700 to 2505 m
(Omu Peak), with rocky abrupt and suspended pla-
teaus, assuring high biodiversity (Sârbu et al., 2013)
on its slopes. It includes a Natural Park designated as
V IUCN category, for protecting fir and spruce for-
ests, associated with subalpine grassland and high
specific biodiversity of flora and fauna. The area is
characterized by a complex land cover on the plateau,
with buildings, roads and erosional areas, but rela-
tively homogenous in forest areas.
Data sets used in the analysis
For this analysis, we have used multiple clouds-free and
snow-free Sentinel-2 imagery from different months
over a year period (Table 1). All the granules used for
both study sites were included in the same Sentinel-2
scene product. We integrated different ancillary data for
enhancing the satellite images, for validating the classi-
fications and for mapping the results (Figure 2).
Comparing with Landsat 8 OLI data (Table 2),
each granule of Sentinel-2A product covers 100 km2
and contains 13 spectral bands. The temporal resolu-
tion of Sentinel-2 images is 10 days now, but it will
increase up to 5 days, when Sentinel-2B MSI sensor
will be active (Wulder et al., 2015). The new bands
around the red-edge region at 20 m spatial resolution,
centred at 705, 740, 783 and 865 nm, are very useful
for vegetation discrimination and land cover produc-
tion (Delegido et al., 2011).
498 M.-R. RUJOIU-MARE ET AL.

Figure 1. Location map of the study areas.

Table 1. Data sets and data sources used in the analysis.

Spatial resolution
Data sets /scale Date of acquisition Source
Sentinel-2 satellite 10, 20, 60 m 8 August 2015 Global Monitoring for Environment and Security (GMES) – Copernicus
images 28 August 2015
23 December 2015
4 April 2016
13 July 2016
Elevation DEM 10 m 1981 Contours from Topographic Military Maps 1:25 000, Military Mapping
Directorate
Orthophotos 0.5 m 2012 National Agency for Cadastre and Land Registration (ANCPI, Bucharest,
Romania)

Images pre-processing techniques processor to read the parameters (Müller-Wilm,

2015) for each sequential band.
The Sentinel-2 data, level 1C available for download
The Top of the Atmosphere reflectance L1C input
are geometrically and radiometrically corrected
images was transformed in the Bottom of the
images, containing Top of the Atmosphere reflec-
Atmosphere reflectance at L2A processing level, includ-
tances in cartographic geometry (UTM projection,
ing the Surface Reflectance bands, resampled and gen-
WGS 84 datum). For land cover classification pro-
erated for an equal resolution (10 m), an Atmospheric
duction, calibrated data in terms of topographic
Optical Thickness (AOT) image and a Water Vapour
ground reflectance are needed. For this reason, it is
map (WV). The AOT and WV images were calculated
important to apply atmospheric and illumination
with the help of the Coastal /Aerosols band (B1) and
corrections on the L1C images.
Atmospheric/Water Vapour band (B9). The band 10
Atmospherically corrected data were obtained
(Cirrrus) is omitted from the L2A output as it does not
using the Sen2Cor plugin available on SNAP soft-
contain surface information (Müller-Wilm, 2015).
ware. This application depends on the calculation of
For the illumination corrections of the images, we
the radiative transfer functions for different sensors
have used the C-correction formula, proposed by
and solar geometries, ground elevations and atmo-
Teillet, Guindon, and Goodenough (1982) in order
spheric parameters (Müller-Wilm, 2015). It is based
to remove the shadows effects on the image and to
on the libRadtran library (Mayer & Kylling, 2005),
unify the spectral response of the similar land cover
which contains the calculations of solar and thermal
features.
radiation on the Earth’s atmosphere. This library  
s þ cλ
generates the Look-up Tables necessary for the ρλ;h;i ¼ ρλ;i  cosθ
cosγ þ cλ cλ ¼ bλ =mλ )
i
 EUROPEAN JOURNAL OF REMOTE SENSING 499

Figure 2. The workflow of the analysis.

Table 2. Comparison between Sentinel-2A MSI (ESA, 2015) and Landsat 8 OLI bands (USGS, 2016).
Sentinel-2A MSI data Landsat 8 OLI data
Band Spatial Central wavelengths Band Spatial Central wavelengths
number Band name resolution (nm) number resolution (nm)
B1 Coastal/aerosols 60 m 443 B1 30 m 443
B2 Visible Blue 10 m 490 B2 30 m 482
B3 Visible Green 10 m 560 B3 30 m 561
B4 Visible Red 10 m 665 B4 30 m 654
B5 NIR 2 20 m 705 – – –
B6 NIR 3 20 m 740 – – –
B7 NIR 4 20 m 783 – – –
B8 NIR 1 10 m 842 B5 30 m 864
B8A NIR 5 20 m 865 – – –
B9 Atmospheric /water vapour 60 m 945 – – –
B10 Cirrus 60 m 1375 B9 30 m 1 373
B11 SWIR 1 20 m 1610 B6 30 m 1608
B12 SWIR 2 20 m 2190 B7 30 m 2200

where from Military Topographic Map (1:25 000, with 5 m

ρλ,h,i = surface reflectance, topographically cor-
rected bands
ρλ,i = surface reflectance of the spectral bands
θs = solar zenith angle (in radians)
γi = illumination angle (in radians)
cλ = empirical constant for each band (generated
from linear regression plot between the illumination
angle and each spectral band).
This algorithm needs the solar zenith angle, avail-
able in the metadata file and a digital model of
elevation co-registered with the images to be cor-
rected (the same cover extent, spatial resolution, geo-
metries references and correlation between pixels).
For this reason, we used the contour lines extracted
equidistance) in order to create a digital elevation
model (DEM) at 10 m spatial resolution. Using the
DEM and the sun azimuth and sun elevation angles
for the Sentinel-2 images, we generated the angle of
incidence, which is complementary to the angle of
illumination (Hantson & Chuvieco, 2011; Riano,
Chuvieco, Salas, & Aguado, 2003; Vanonckelen,
Lhermitte, & Van Rompaey, 2013).
These atmospheric and illumination corrections
were applied on each multi-date granules used in
the analysis for the both study areas of Bucegi and
Prahova-Teleajen Subcarpathians. Thus, the new
resulted images at L2A level of processing contain
topographic corrected surface reflectances (Figure 3).
500 M.-R. RUJOIU-MARE ET AL.

Figure 3. Calibration workflow results of the Sentinel-2 images for Prahova-Teleajen Subcarpathians (A) and Bucegi Mountains (B).

Training sample collection and evaluation The largest ones reflect a lot of variability in the spec-
tral reflectance characteristics of a ROI class. The error
The supervised classification process starts with train-
bars of two or more ROI classes that overlap show
ing areas creation for each land cover class from the
spectral similarities between those classes and poten-
study areas. Being a fragmented territory, the
tial errors generated in the classification.
Subcarpathian area contains more land cover classes
Secondly, we calculated the ROI spectral separability
than the mountainous region, presented in Table 3.
using the Jeffries–Matusita and Transformed
The training areas were statistically validated in
Divergences algorithms. These techniques showed how
correlation with the associated input data. Firstly, we
well separated were the land cover classes, comparing
computed a series of graphs showing the mean spectral
them in pairs. These algorithms for quantitative expres-
signature (Figure 4a) and the standard deviation
sions of the ROI separability are based on the covariance
(Figure 4b) for each land cover class in every spectral
and weighted distance between the mean values of the
band from the input image. These graphs evaluate the
classes (Chuvieco, 2016). The result is a table showing
spectral separability between the training areas to see if
the statistics separability of pairs. The highest value is
they are statistically different. The smallest standard
2000 and it shows that those classes are clearly separated
deviation bars from the spectral dispersion graph
on the input image (the weighted distance between
(Figure 4b) show that the reflectances within a region
classes is high). The values lower than 2.00 indicate
of interest (ROI) class in a spectral band are similar.

Table 3. Land cover classes for each study area and their description.
Land cover classes
Prahova-Teleajen Subcarpathians Bucegi Mountains
(number of ROI pixels) (number of ROI pixels) Description
1 Coniferous forest (750) Coniferous forest (324) Abies alba, Picea abies, Pinus sylvestris, Larix decidua, Pinus cembra
2 Deciduous forest (3080) Deciduous forest (144) Fagus sylvatica, Querqus petraea, Querqus robur, Robinia Pseudacacia
3 – Dwarf pine (356) Pinus mugo
4 – Mixed forest (208) Picea abies, Abies alba and Fagus sylvatica
5 Sparsely vegetated areas (including – Orchards, shrubs and different broadleaves species with rare densities grown
orchards and shrubs) (264) close to settlements and roads
6 Grasslands (3714) Grasslands (304) Herbaceous area used for pastures or hays
7 Croplands (1518) – Agricultural lands (cereals or vegetables) cultivated in the depressions or at
the contact with the plain (South from Subcarpathian area)
8 Built-up area (950) – Settlements, transport and hydrotechnical facilities, industry
9 Bare land (961) – Mainly riverbeds, but also bare soil and bare rock
10 – Bare land/built-up area (80) Constructions, roads, tracks, as well as bare soil/rock or degraded grassland
11 Water bodies (1153) Water bodies (208) Reservoirs, lakes and rivers

Figure 4. Evaluation of the spectral separability between the training areas for a single-date image from 28 August 2015,
Subcarpathian area: (a) spectral signature plot; (b) spectral dispersion graph.
that confusions between the land cover classes and these
errors will be propagated in the classification.
For improving the separability between the land
cover classes, we added more spectral bands as input
data for the classification. For the Subcarpathian area,
we have used three intra-annual Sentinel-2 scenes,
from different vegetation seasons (summer, winter
and spring), in order to eliminate the confusions
between the land cover classes and distinguish the
spectral response. Using multi-date images in the
classification generates changes in the training sam-
ples data, which collects a large spectral variability
(Sexton et al., 2013). Thereby, the training sets
include various spectral signatures of the land cover
classes captured across time and space.
For Bucegi Mountains, we have used two granules
for the classification because the land cover classes
are more compact and clearly separated. Both
selected images were from the summer season, for
as there were no clouds-free or no snow-free images
available from the other seasons for the mountai-
nous area.
Multi-date image classifications
Choosing the right classification algorithm is also a
difficult task because of the land cover classes and their
spectral signatures. maximum likelihood (ML) is a
parametric widely used algorithm and it evaluates the
EUROPEAN JOURNAL OF REMOTE SENSING 501
variance and covariance of the spectral response for
each class (Chuvieco, 2016). It computes the probabil-
ity of each pixel value to belong to each land cover
classes and assigns it to the class with the high prob-
ability value (Chuvieco, 2016; Lillesand, Kiefer, &
Chipman, 2015; Schowengerdt, 2007). This algorithm
manages well with compact land cover classes, but
there are necessary large and pure training samples
(Gomez et al., 2016). It returns good results on the
mountainous area where the land cover classes are
more compact and homogeneous, but in the
Subcarpathian region, it generates lots of confusions
between built-up area, bare land and croplands or
deciduous forests and sparsely vegetated areas class.
Support vector machine (SVM) is a non-parametric
classification algorithm that yields good results from
complex and noisy data (Chuvieco, 2016) because it
separates the land cover categories using a decision
surface that maximizes the margins of the classes (M.
Hansen, 2012) and the spectral response of the classes
does not over-fit (Gomez et al., 2016). It manages well
large feature space, and there are not necessary large
training samples, but they need to contain a wide
range of spectral variability of the classes (Gomez
et al., 2016). It includes a penalty parameter that con-
trols the margins between classes where the training
data are non-separable, allowing a certain degree of
misclassification. This algorithm was tested on various
feature space, compared with other classification
502 M.-R. RUJOIU-MARE ET AL.
algorithms as ML and neural networks and it returned
good results with lower errors (Khatami et al., 2016).
Both algorithms were applied for each study area and
the best results were selected for further processing and
analysis. As mentioned earlier, the SVM managed well
the heterogeneous classes of the Subcarpathians, while
the ML classified well the compact land cover structure.
Accuracy assessments of the classification results
Each classification was assessed with a set of ground
truth collected on the orthophotos (2012) and from field
observations, for both study areas. For the
Subcarpathian region, we used 11,206 sample points
with a near-uniform spatial distribution and an average
density of 7.6 points per square kilometre. For the
Carpathian region, we used 446 sample points with an
average density of 7.5 points per square kilometre. The
confusion matrices were computed to compare the rela-
tionship between the known reference data and the
results from the classifications (Lillesand et al., 2015).
These methods show the overall accuracies, kappa coef-
ficients and the various errors of commission and omis-
sion related to user’s and producer’s accuracies.
Results
The spectral signature plot (Figure 4a) for a single-
date image from 28 August 2015 for Subcarpathian
area shows overlaps for deciduous forest with spar-
sely vegetated and built-up areas with croplands. The
spectral dispersion graph for the same image shows
the variance in the reflectances of each category.
Deciduous, coniferous and sparsely vegetated areas
Figure 5. Jeffries–Matusita ROI separability values for pairs of land cover features applied on different input data for
Subcarpathian area (CF – coniferous forest; DF – deciduous forest; SP – sparsely vegetated area; G – grasslands; C – croplands;
B – built-up area; BR – bare land; W – water bodies).
have low error bars in the visible spectrum; therefore,
the spectral signatures of each category are similar.
Otherwise, in red edge bands, near-infrared and
short-wave infrared spectrum, the error bars for
deciduous, coniferous and sparsely vegetated areas
show a high spectral variability in the spectral signa-
tures. The built-up area and bare land have high
variability in the visible spectrum. On the other
hand, the error bars of build-up area and croplands
are overlapping which means that those classes are
not statistically different and they will be mismatched
in the classification process. The same problems
appear between deciduous forest and sparsely vege-
tated area classes and for bare land with croplands
and built-up area.
For improving the spectral separability between
land cover classes, we used multi-date images as
input data. The results of the separability of the train-
ing areas (Figure 5) associated with multi-date images
were increased, the lowest value from single-date being
1.790, whereas for multi-date images being 1.993.
The land cover map for Subcarpathian area,
obtained using SVM classification, algorithm has
eight classes. The confusion issues were solved by
using multi-date images (Figure 6a); therefore, the
land cover categories are clearly presented on the
map. The coniferous forest occupies small areas, situ-
ated in the northern part, close to the mountainous
area. The settlement structure was well mapped,
showing their spread and shape. Also, other built-up
features are well represented, as the national road and
the railway or industrial area from the south of study
area. The sparsely vegetated class, which include
orchards, shrubs and rare broadleaves trees, is located

Figure 6. Differences between single-date (1) and multi-date image classification (2) for Subcarpathian (1) and Carpathian (2)
study areas.
near the settlements and it is well represented on the
map. The riverbeds of Prahova and Teleajen have a
large extend at the contact of the Subcarpathians with
the plain and in depressions (Figure 7).
For Bucegi Massif, part of Bucegi Natural Park, the
land cover map (Figure 8) obtained from Sentinel-2
images contains of 7 classes. The most complex cate-
gory to differentiate was the mixed forest class
(between fir, spruce and beech). Using two images
for multi-date analysis (only from August 2015 and
July 2016) helped for a better differentiation of this
class (Figure 6b). Due to the same spectral reflec-
tance, the built-up area had to be merged with the
bare land surfaces in one category. This includes all
the constructions, roads, tracks, as well as bare soil
/rock or degraded grassland used for intensive graz-
ing during the summer season. Compared with the
orthophotos, the classification showed good results
and a fine extraction of all tracks and details that
could be seen on the plateau. The forest differentia-
tion included coniferous (Picea abies, Abies alba,
Larix decidua), beech (Fagus sylvatica), mixed forest
and dwarf pine (Pinus mugo). Coniferous and dwarf
pine covered areas were extracted very well, but some
errors were encountered when differentiating beech
and mixed forest. These are described in the confu-
sion matrix in Table 6.
Discussions
Using multi-date images in classification enhances
the spectral separability of the land cover classes
and increases the overall accuracy of the land cover
EUROPEAN JOURNAL OF REMOTE SENSING 503
map. The image pre-processing time increases,
because the atmospheric and topographic corrections
are applied on each data set.
The selection of training samples is also a difficult
task, as they have to be statistically spectral different
for generating a good accuracy of the classification. It
is important to select samples from the field area or
to validate them with expert knowledge.
The classification algorithms have to be chosen
according to the land cover characteristics. We have
chosen ML algorithm for the mountainous area
where the land cover features were more compact
and with clearly spectral behaviour. Otherwise, for
Subcarpathian area, we used SVM, as the land cover
classes were more fragmented and a lot of classes had
similar spectral response. The resulted land cover
maps have a high level of detail as they are produced
from Sentinel-2 images at 10 m resolution (Table 4).
The overall accuracy of Subcarpathian land cover
map of 92.41% (Table 5) shows a good correlation
between the results of the classification and the refer-
ence data. The biggest omission affects sparsely vege-
tated class because of their spectral similarities with
the grasslands on the fields covered with grasslands
and shrubs. Other omission errors are affecting bare
lands and croplands, where some pixels are com-
mitted in built-up area class. The user’s accuracy
shows high values for water bodies, coniferous, decid-
uous and croplands. The lowest value of user’s accu-
racy refers to sparsely vegetated areas and bare lands.
For Bucegi Massif, the overall accuracy of the land
cover map is 98.65% (Table 6). The highest scores for
omission affects grassland and dwarf pine, mostly
504 M.-R. RUJOIU-MARE ET AL.
Figure 7. Land cover map for Subcarpathian area, generated from multi-date Sentinel-2 images, using SVM classification
algorithm.
because of similar spectral response in some situations
(degraded patches of grassland or isolated dwarf pines).
Commission error, however, shows in some cases a
significant difference between the classification and
the reference data. For instance, the bare land/built-up
area may have included some areas that were not neces-
sarily in that category, because in many cases bare land
areas are not clearly delimited along the Prahova
abrupt, a very complex area. The second case represents
the mixed forest areas, which may often include small
compact patches of coniferous or deciduous forest that
are generalized into one category, the resolution of the
sensor being also a limitation.
We compared the results with orthophotos, but we
also validated the land cover maps with field observa-
tions. In Figure 9, the images reflect the complexity of
land cover features in the Subcarpathians. In
Figure 9a, a settlement situated in the depression
with pasture field nearby and deciduous forest
(Fagus sylvatica) and also coniferous (Picea abies)
on the interfluves is observed. In Figure 9b, the spar-
sely vegetated area class is well represented with fruit
trees in blossom and other species of trees or shrubs
situated on hill slopes, close to settlements. The
Subcarpathian area is very fragmented by erosion
and landslides that affect the slopes with soft rocks
and no forested areas (Figure 9c).
A problem in the case of Bucegi Massif was to
distinguish the built-up area from the bare land
patches and that they had to be merged into one
category. In Figure 10a, details from land cover map
of Bucegi Massif are presented. Figure 10a illustrates
Babele area with erosional areas due to touristic traf-
fic and anthropogenic activities. This class, bare land
and built-up area has the same spectral response and
includes also areas like in Figure 10b with abrupt and
rocky slopes (Caraiman abrupt). We mention the fact
that for the Bucegi Natural Park management, the
land cover map that includes the degraded areas
represents a good solution for constant monitoring
of the environmental evolution in the reservation.
The third situation (Figure 10c) represents the Pinus
mugo-covered area, protected inside the reservation,
which was mostly mapped on the images except for
the isolated patches/singular trees. Although highly
fragmented, the plateau area from Bucegi Massif was
well represented on the land cover map by each
category with more details.
 EUROPEAN JOURNAL OF REMOTE SENSING 505

Figure 8. Land cover map for Bucegi area, generated from multi-date Sentinel-2 images, using ML classification algorithm.

Table 4. Comparative table with statistical accuracies between ML and SVM for Subcarpathian and Carpathian area.
Subcarpathian area Carpathian area
Producer’s accuracy User’s accuracy Producer’s accuracy User’s accuracy
ML SVM ML SVM ML SVM ML SVM
Land cover classes Coniferous 98.26 97.32 100.00 98.78 Coniferous 100.00 90.22 100.00 89.25
Deciduous 98.54 97.55 95.33 98.43 Deciduous 98.46 100.00 100.00 98.48
Sparsely vegetated area 21.24 45.05 93.96 57.88 Mixt forest 100.00 97.28 99.32 90.51
Grasslands 98.27 97.1 83.03 85.46 Dwarf pine 96.25 68.75 100.00 82.09
Cropland 91.86 76.43 98.05 97.75 Grassland 94.59 97.30 100.00 97.30
Built-up area 99.74 96.52 57.46 94.08 Bare land/Built-up area 100.00 94.74 79.17 94.74
Bare land 67.19 92.23 92.65 74.15 water bodies 100.00 100.00 100.00 100.00
Water bodies 98.98 100.00 100.00 99.83
Overall accuracy ML 84.47% SVM 92.41% Overall accuracy ML 98.65% SVM 91.03%
Kappa Coefficient ML 0.8177 SVM 0.9098 Kappa Coefficient ML 0.9829 SVM 0.8854

Conclusions images acquired in different seasons, the accuracy of

the final result enhances as the spectral separability
The input spectral bands, the training sets and classi-
between land cover features increase. The land cover
fication algorithm represent key factors for land cover
maps generated from supervised classification of
map production. The training samples created on
intra-annual multi-date images show higher levels of
multi-date images collect various spectral signatures
accuracies (92.41% for Subcarpathians and 98.65%
of the land cover features in different seasons from
for Carpathians) than the single-date classification.
the same year. Using multiple spectral bands from
In the case of the hilly area with wide variety of
506 M.-R. RUJOIU-MARE ET AL.

Table 5. Confusion matrix for Prahova–Teleajen Subcarpathians.

Ground reference data User’s Commission
Classes CF G BA W C SV C BL DF Total accuracy % error
CLASSIFIED Coniferous (CF) 97.32 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.30 6.47 98.78 1.22
IMAGE Grasslands (G) 0.00 97.10 0.00 0.00 2.28 54.95 0.00 0.00 0.00 16.22 85.46 14.54
Built-up area (BA) 0.00 0.00 96.52 0.00 3.78 0.00 0.00 7.77 0.00 17.39 94.08 5.92
Water bodies (W) 0.00 0.00 0.10 100.00 0.00 0.00 0.00 0.00 0.00 10.33 99.83 0.17
Croplands (C) 0.00 0.92 0.62 0.00 76.43 0.00 0.00 0.00 0.00 10.57 97.75 2.25
Sparsely vegetated 0.00 1.42 0.00 0.00 3.91 45.05 0.00 0.00 1.85 2.90 57.88 42.12
areas (SV)
Clouds (C) 0.00 0.00 0.00 0.00 0.00 0.00 100.00 0.00 0.30 1.42 94.41 5.59
Bare land (BL) 0.00 0.55 2.75 0.00 11.85 0.00 0.00 92.23 0.00 8.31 74.15 25.85
Deciduous (DF) 2.68 0.00 0.00 0.00 1.76 0.00 0.00 0.00 97.55 26.38 98.43 1.57
Producer’s accuracy % 97.32 97.10 96.52 100.00 76.43 45.05 100.00 92.23 97.55
Omission error 2.68 2.90 3.48 0.00 23.57 54.95 0.00 7.77 2.45
Overall accuracy = 92.41 Kappa coefficient = 0.9098

Table 6. Confusion matrix for Bucegi Massif.

Ground reference data User’s
Classes CF DP DF MF G WB B Total accuracy % Commission error
CLASSIFIED IMAGE Coniferous forest (CF) 100.0 0.00 0.00 0.00 0.00 0.00 0.00 20.63 100.0 0.00
Dwarf pine (DP) 0.00 96.25 0.00 0.00 0.00 0.00 0.00 17.26 100.0 0.00
Deciduous forest (DF) 0.00 0.00 98.46 0.00 0.00 0.00 0.00 14.35 100.0 0.00
Mixed forest (MF) 0.00 0.00 1.54 100.0 0.00 0.00 0.00 33.18 99.32 0.68
Grassland (G) 0.00 0.00 0.00 0.00 94.59 0.00 0.00 7.85 100.0 0.00
Water bodies (WB) 0.00 0.00 0.00 0.00 0.00 100.0 0.00 1.35 100.0 0.00
Bare land/built-up area (B) 0.00 3.75 0.00 0.00 5.41 0.00 100.0 5.38 79.17 20.83
Producer’s accuracy % 100.0 96.25 98.46 100.0 94.59 100.0 100.0
Omission error 0.00 3.75 1.54 0.00 5.41 0.00 0.00
Overall accuracy = 98.65 Kappa coefficient = 0.9829

Figure 9. SVM classification for Subcarpathian area, validated with orthophotos (2012) and field observations: (a) built-up area
with pasture and forest, (b) sparse vegetation including fruit trees (plum and apple trees) and dispersed trees (acacia, hornbeam
and willow), (c) sparse vegetation (shrubs) with pastures and bare lands.

land uses (discontinuous built-up areas, crops, orch- date approach is necessary to separate the spectral
ards, together with sparsely wooden vegetation and signatures of the similar classes (for example, bare
deciduous and coniferous forests), a seasonal multi- lands and croplands). In comparison, in the

Figure 10. ML classification for Carpathian area, validated with orthophotos (2012) and field observations: (a) bare land with
pasture and constructions, (b) rocky abrupt, (c) bare land with dwarf pine (Pinus mugo).
mountain area with compact and spectral-different
features, the analysis is not necessary to include
images from different seasons.
Choosing the processing algorithm depends on the
characteristics of the land cover types. In this article,
we have used two different algorithms for as our
study regions were complementary: Prahova–
Teleajen Subcarpathians with fragmented and hetero-
geneous features and Bucegi Massif with homoge-
neous vegetation layers.
The land cover maps that were obtained from
Sentinel-2 calibrated images offer great level of
detail with more than seven categories. They were
validated in field, with ground truth data and
orthophotos, being a good solution for spatial plan-
ning, hazard evaluation and protected areas
management.
Using multi-date Sentinel-2 images involves time-
consuming and important hardware resources, mak-
ing it a difficult approach for casual uses. Also, in
some situations, differentiating land cover categories
with spectral similarities implies some issues due to
spatial and spectral resolution.
Sentinel-2 MSI offers new possibilities for map-
ping land cover with greater precision that can be
used for environmental monitoring, agriculture and
forestry management, urban planning and decision
making at local and regional scales.
Disclosure statement
No potential conflict of interest was reported by the
authors.
EUROPEAN JOURNAL OF REMOTE SENSING 507
References
Anderson, J., Hardy, E.E., Roach, J.T., & Witmer, R.E.
(1976). A land use and land cover classification system
for use with remote sensing data. U.S. Geological Survey
Professional Paper, No. 964 USGS, Washington DC.
Arino, O., Bicheron, P., Achard, F., Latham, J., Witt, R., &
Weber, J.-L. (2008). The most detailed portrait of Earth.
Eur. Space Agency, 136, 25–31.
Bartholomé, E., & Belward, A.S. (2005). GLC2000: A new
approach to global land cover mapping from Earth obser-
vation data. International Journal of Remote Sensing, 26(9),
1959–1977. doi:10.1080/01431160412331291297
Büttner, G., Feranec, J., & Jaffrain, G. (2000). Corine land
cover update 2000. technical guidelines. Copenhagen,
Denmark: European Environment Agency.
Campbell, J.B., & Wynne, R.H. (2011). Introduction to
remote sensing (5th ed.). New York; London: Guilford.
Chen, J., Chen, J., Liao, A.P., Cao, X., Chen, L.J., Chen, X.
H., ... Mills, l. (2015). Global land cover mapping at 30 m
resolution: A POK-based operational approach. ISPRS
Journal of Photogrammetry and Remote Sensing, 103, 7–
27. doi:10.1016/j.isprsjprs.2014.09.002
Chuvieco, E. Fundamentals of satellite remote sensing: An
environmental approach. Second Edition. Boca Raton/
London/New York: CRC Press Taylor & Francis. 2016
Delegido, J., Verrelst, J., Alonso, L., & Moreno, J. (2011).
Evaluation of Sentinel-2 Red-Edge Bands for empirical
estimation of green LAI and chlorophyll content.
Sensors, 11(7), 7063–7081. doi:10.3390/s110707063
Di Gregorio, A., & Jansen, E. (1997). A new concept for a
land cover classification system. Paper presented at the
Earth Observation and Environmental Information 1997
Conference, Alexandria, Egipt.
Di Gregorio, A., & Jansen, L.J.M. (2000). Land cover classi-
fication system (LCCS). Classification concepts and user
manual for software version 1.0. Rome: FAO.
Drusch, M., Del Bello, U., Carlier, S., Colin, O., Fernandez,
V., Gascon, F., ... Bergellini, P. (2012). Sentinel-2: ESA’s
optical high-resolution mission for GMES operational
508 M.-R. RUJOIU-MARE ET AL.
services. Remote Sensing of Environment, 120, 25–36.
doi:10.1016/j.rse.2011.11.026
ESA. (2015). Sentinel-2 user handbook. ESA Standard
Document, 1(2). Retrieved January 2017 from https://
sentinels.copernicus.eu/web/sentinel/user-guides/docu
ment-library/-/asset_publisher/xlslt4309D5h/content/sen
tinel-2-user-handbook
Feranec, J., Hazeu, G., Christensen, S., & Jaffrain, G. (2007).
Corine land cover change detection in Europe (case stu-
dies of the Netherlands and Slovakia). Land Use Policy,
24(1), 234–247. doi:10.1016/j.landusepol.2006.02.002
Feranec, J., Hazeu, G., Jaffrain, G., & Cebecauer, T. (2007).
Cartographic aspects of land cover change detection
(over- and underestimation in the I&CORINE Land
Cover 2000 Project). Cartographic Journal, 44(1), 44–
54. doi:10.1179/000870407X173869
Feranec, J., Jaffrain, G., Soukup, T., & Hazeu, G. (2010).
Determining changes and flows in European landscapes
1990-2000 using CORINE land cover data. Applied
Geography, 30(1), 19–35. doi:10.1016/j.
apgeog.2009.07.003
Gomez, C., White, J.C., & Wulder, M.A. (2016). Optical
remotely sensed time series data for land cover classifi-
cation: A review. Isprs Journal of Photogrammetry and
Remote Sensing, 116, 55–72. doi:10.1016/j.
isprsjprs.2016.03.008
Hansen, M. (2012). Classification trees and mixed pixel
trening data. In C. Giri (Ed.), Remote sensing of land
use and land cover. priciples and applications (pp. 127–
136). Boca Raton, London, New York: Taylor&Francis
Group, CRC Press.
Hansen, M.C., Defries, R.S., Townshend, J.R.G., &
Sohlberg, R. (2000). Global land cover classification at
1 km spatial resolution using a classification tree
approach. International Journal of Remote Sensing, 21
(6–7), 1331–1364. doi:10.1080/014311600210209
Hantson, S., & Chuvieco, E. (2011). Evaluation of different
topographic correction methods for Landsat imagery.
International Journal of Applied Earth Observation and
Geoinformation, 13(5), 691–700. doi:10.1016/j.jag.2011.
05.001
ISO-19144-1. (2009). Geographic information —
Classification systems — Part 1: Classification system
structure (Vol. ISO 19144-1). Prepared by FAO,
Published by the International Organisation for
Standardization at https://www.iso.org/standard/32562.
html
ISO-19144-2. (2012). Geographic information - data. Canadian Journal of Remote Sensing(8), 84–106.
Classification systems – Part 2: Land Cover Meta
Language (LCML) (pp. 107): ISO/TC 211 geographic
information/geomatics.Prepared by FAO, Published by
the International Organisation for Standardization at
https://www.iso.org/standard/44342.html
Khatami, R., Mountrakis, G., & Stehman, S.V. (2016). A
meta-analysis of remote sensing research on supervised
pixel-based land-cover image classification processes:
General guidelines for practitioners and future research.
Remote Sensing of Environment, 177, 89–100.
doi:10.1016/j.rse.2016.02.028
Latham, J., Cumani, R., Rosati, I., & Bloise, M. (2014). FAO
global land cover (GLC-SHARE) beta-release 1.0
Database. Rome: FAO. Retrieved January 2017 from
www.fao.org/uploads/media/glc-share-doc.pdf
Lillesand, T.M.A., Kiefer, R.W.A., & Chipman, J.W.A.
(2015). Remote sensing and image interpretation. 7th
Edition, New York, NY: John Wiley & Sons
Mayer, B., & Kylling, A. (2005). Technical note: The
libRadtran software package for radiative transfer calcu-
lations - description and examples of use. Atmospheric
Chemistry and Physics, 5, 1855–1877. doi:10.5194/acp-5-
1855-2005
Müller-Wilm, U. (2015). Sentinel-2 MSI - Level-2A proto-
type processor installation and user manual. Germany:
Telespazio VEGA Deutschland GmbH.
Parmentier, B., & Eastman, J.R. (2014). Land transitions
from multivariate time series: Using seasonal trend ana-
lysis and segmentation to detect land-cover changes.
International Journal of Remote Sensing, 35(2), 671–
692. doi:10.1080/01431161.2013.871595
Riano, D., Chuvieco, E., Salas, J., & Aguado, I. (2003).
Assessment of different topographic corrections in
Landsat-TM data for mapping vegetation types.
Transactions on Geoscience and Remote Sensing, 41(5),
1056–1061. doi:10.1109/TGRS.2003.811693
Robin, M. (1998). La Télédetection. Paris: Editions Nathan.
Rujoiu-Mare, M.R., & Mihai, B. (2016). Mapping land
cover using remote sensing data and GIS techniques: A
case study of Prahova Subcarpathians. Elsevier Procedia
Environmental Sciences, 32, 244–255. doi:10.1016/j.
proenv.2016.03.029
Sârbu, A., Anastasiu, P., Smarandache, D., Pascale, G.,
Lițescu, S., & Mihai, D.C. (2013). Habitate cu valoare
conservativă din Parcul Natural Bucegi. Bucharest: Ceres
Publisher House.
Schowengerdt, R.A. (ed.). (2007). Remote sensing: Models
and methods for image processing (3rd ed.). London:
Academic Press.
Sexton, J.O., Urban, D.L., Donohue, M.J., & Song, C.H.
(2013). Long-term land cover dynamics by multi-tem-
poral classification across the Landsat-5 record. Remote
Sensing of Environment, 128, 246–258. doi:10.1016/j.
rse.2012.10.010
Shao, Y., Lunetta, R.S., Wheeler, B., Iiames, J.S., &
Campbell, J.B. (2016). An evaluation of time-series
smoothing algorithms for land-cover classifications
using MODIS-NDVI multi-temporal data. Remote
Sensing of Environment, 174, 258–265. doi:10.1016/j.
rse.2015.12.023
Teillet, P.M., Guindon, B., & Goodenough, D.G. (1982).
On the slope-aspect correction of multispectral scanner
doi:10.1080/07038992.1982.10855028
USGS. (2016). Landsat 8 (L8) data users handbook Version
2.0. Department of the Interior U.S. Geological Survey.
Retrieved December 2016 from https://landsat.usgs.gov/
…/Landsat8DataUsersHandbook.pdf
Vanonckelen, S., Lhermitte, S., & Van Rompaey, A. (2013).
The effect of atmospheric and topographic correction
methods on land cover classification accuracy.
International Journal of Applied Earth Observation and
Geoinformation, 24, 9–21. doi:10.1016/j.jag.2013.02.003
Wulder, M.A., Hilker, T., White, J.C., Coops, N.C., Masek,
J.G., Pflugmacher, D., Crevier, Y. (2015). Virtual con-
stellations for global terrestrial monitoring. Remote
Sensing of Environment, 170, 62–76. doi:10.1016/j.
rse.2015.09.001