International Journal of Applied Earth Observation and Geoinformation 52 (2016) 361–370

Contents lists available at ScienceDirect

International Journal of Applied Earth Observation and

Geoinformation
journal homepage: www.elsevier.com/locate/jag

Mapping Brazilian savanna vegetation gradients with Landsat time

series
Marcel Schwieder a,∗ , Pedro J. Leitão a , Mercedes Maria da Cunha Bustamante b ,
Laerte Guimarães Ferreira c , Andreas Rabe a , Patrick Hostert a,d
a
Humboldt-Universität zu Berlin, Geography Department, Unter den Linden 6, 10099, Berlin, Germany
b
Universidade de Brasília—UNB, Campus Universitário Darcy Ribeiro, Asa Norte, 70910-900, Brasília, DF, Brazil
c
Universidade Federal de Goiás—UFG, Campus II, LAPIG, 74001-970, Goiânia, GO, Brazil
d
Humboldt-Universität zu Berlin, Integrated Research Institute on Transformations of Human-Environment Systems (IRI THESys), Unter den Linden 6,
10099, Berlin, Germany

a r t i c l e i n f o a b s t r a c t

Article history: Global change has tremendous impacts on savanna systems around the world. Processes related to climate
Received 17 February 2016 change or agricultural expansion threaten the ecosystem’s state, function and the services it provides. A
Received in revised form 15 June 2016 prominent example is the Brazilian Cerrado that has an extent of around 2 million km2 and features high
Accepted 21 June 2016
biodiversity with many endemic species. It is characterized by landscape patterns from open grasslands
Available online 19 July 2016
to dense forests, defining a heterogeneous gradient in vegetation structure throughout the biome. While
it is undisputed that the Cerrado provides a multitude of valuable ecosystem services, it is exposed to
Keywords:
changes, e.g. through large scale land conversions or climatic changes. Monitoring of the Cerrado is thus
Cerrado
Physiognomy mapping
urgently needed to assess the state of the system as well as to analyze and further understand ecosystem
Landsat time series responses and adaptations to ongoing changes. Therefore we explored the potential of dense Landsat
Land surface phenology time series to derive phenological information for mapping vegetation gradients in the Cerrado. Frequent
data gaps, e.g. due to cloud contamination, impose a serious challenge for such time series analyses. We
synthetically filled data gaps based on Radial Basis Function convolution filters to derive continuous
pixel-wise temporal profiles capable of representing Land Surface Phenology (LSP). Derived phenological
parameters revealed differences in the seasonal cycle between the main Cerrado physiognomies and
could thus be used to calibrate a Support Vector Classification model to map their spatial distribution.
Our results show that it is possible to map the main spatial patterns of the observed physiognomies based
on their phenological differences, whereat inaccuracies occurred especially between similar classes and
data-scarce areas. The outcome emphasizes the need for remote sensing based time series analyses at
fine scales. Mapping heterogeneous ecosystems such as savannas requires spatial detail, as well as the
ability to derive important phenological parameters for monitoring habitats or ecosystem responses to
climate change. The open Landsat and Sentinel-2 archives provide the satellite data needed for improved
analyses of savanna ecosystems globally.
© 2016 Elsevier B.V. All rights reserved.

1. Introduction goods and services such as food, pollinators and carbon storage
(Marchant, 2010). They occur in tropical and sub-tropical cli-
Savanna ecosystems cover an estimated 20% of the global terres- mate zones across all continents, but are particularly prevalent
trial surface (Lehmann et al., 2011), providing essential ecosystem in Australia, Africa and the Americas (Solbrig, 1996). Global land
use change processes related to a growing demand for natural
resources led to a conversion of approximately 50% of the global
savannas with a direct impact on biodiversity and carbon storage
∗ Corresponding author. (Foley et al., 2011). Climate change, invasive species and fertilizer
E-mail addresses: [email protected] (M. Schwieder), pollution rapidly impact savanna biodiversity (MEA, 2005), while
[email protected] (P.J. Leitão), [email protected] global change is likely to alter the global distribution of savannas
(M.M. da Cunha Bustamante), [email protected] (L.G. Ferreira),
and might even lead to a change of biome states (Staver et al., 2011).
[email protected] (A. Rabe), [email protected]
(P. Hostert).

http://dx.doi.org/10.1016/j.jag.2016.06.019
0303-2434/© 2016 Elsevier B.V. All rights reserved.
362 M. Schwieder et al. / International Journal of Applied Earth Observation and Geoinformation 52 (2016) 361–370

The Brazilian savanna, better known as Cerrado, is a promi- Using high temporal resolution satellite data time series not
nent example of global change impacts on savanna ecosystems. It only allows to discriminate vegetation types but also to describe
is the second largest eco-region in Brazil with an extent of about different phenological vegetation phases throughout a season
2 million km2 (Ratter et al., 1997). Landscape formations that range (Zhang et al., 2003). Ferreira and Huete (2004) assessed the sea-
from open grass and shrub dominated lands to dense forests are sonal dynamics of the Cerrado vegetation using time series of
characteristic (Oliveira-Filho and Ratter, 2002) and form a gradient AVHRR vegetation indices. In spite of the drawbacks inherent to
of vegetation structure and biomass, which is often differentiated the data (e.g. coarse spatial resolution, sensor uncertainties and
into discrete structural physiognomy classes (Ribeiro and Walter, broad bandwidths) they were able to derive phenological pat-
2008). The climate of the Cerrado is characterized by a wet and a terns capable of depicting the seasonal cycle and which allowed
dry season, which influence the vegetation’s spatial and temporal to distinguish between savanna formations, pastures, croplands
dynamics. It’s harsh environmental conditions led to a high floris- and forests. Ratana et al. (2005) explored the potential of MODIS
tic diversity and a variety of phenological adaptation strategies 16-day composite time series to analyze phenological patterns of
(Ferreira and Huete, 2004). The Cerrado is thus considered as the different Cerrado physiognomies, revealing their distinct responses
biodiversity-richest savanna globally (Silva and Bates, 2002) with to seasonal contrasts with a 250 m spatial resolution. Even though
approximately 160,000 species of fungi, flora and fauna (Furley, both studies provided valuable insights on phenological differences
1999). However, a weak land conservation status has led to large- between Cerrado vegetation formations, they did not aim at map-
scale conversions from natural to agricultural land that already ping their spatial distribution. As the analysis of phenology derived
affected more than 40% of the Cerrado, which is likely to aggravate from remote sensing data is based on measures of temporal changes
in the future (Ferreira et al., 2012; Sano et al., 2010). This conver- in surface reflectance at the pixel scale (Hanes et al., 2014), it usu-
sion of primary vegetation threatens the stability of the ecosystem ally relates to a signal mixture of different canopy or understory
and related services provided, such as carbon sequestration and cli- layers and depicts rather a vegetation community instead of a sin-
mate regulation. It further impacts its biodiversity, rendering the gle species’ phenology. The sensor’s spatial resolution is therefore
Cerrado as an under-researched global biodiversity hotspot in need critical for the potential detail of the derived information, which is
of in-depth monitoring as basis for profound conservation planning particularly important for analyzing and mapping the complex veg-
(Myers et al., 2000). Thus, accurate mapping and monitoring of the etation gradients in the Brazilian Cerrado. The analysis of Landsat
temporal and spatial dynamics of Cerrado vegetation is essential to data with a spatial resolution of 30 m is promising, but has mostly
understand ecosystem properties and responses to ongoing change been restricted to multi-temporal imagery with low temporal res-
processes to support decision makers (Rocha et al., 2011; Sano et al., olution, which is not sufficient to derive continuous phenological
2010). information. However, the opening of the extensive data holdings
Field based mapping and ecological assessments that rely on of the Landsat archive (Wulder et al., 2012) allows to combine
established classification schemes are indispensable for a detailed data from the Landsat Thematic Mapper (TM) with Enhanced The-
analysis of local processes, but at the same time costly and thus matic Mapper (ETM+) and Operational Land Imager (OLI) sensors.
restricted to relatively small areas. Remote sensing data, on the This improves the temporal resolution to potentially eight days,
other hand, have successfully been used to map the Cerrado across rendering Landsat a system with capabilities for detailed phenolog-
large areas and in inaccessible terrain. Sano et al. (2010) created ical information retrieval and offers a great opportunity to observe
a land cover map of the entire Cerrado based on a mosaic of 170 vegetation gradients over long time periods with sufficient spatial
Landsat scenes from 2002. They differentiated anthropogenic and resolution.
natural (grass-, shrub- and forestlands) land cover classes by image In order to deepen the knowledge on Cerrado vegetation, we aim
segmentation and visual interpretation with an overall accuracy to derive spatially explicit phenological information, using state
of 71%. Spatially less extended studies investigated the usability of the art remote sensing techniques. We focus on the applica-
of combined radar and optical data (Sano et al., 2005) or spectral bility and limitations of the combined use of Landsat time series
unmixing of Landsat data (Ferreira et al., 2007) to discriminate Cer- and an established physiognomy classification scheme (Ribeiro and
rado vegetation physiognomies (structural classes). Other studies Walter, 2008), which enables the expansion of field based ecolog-
have shown that analyzing multi-temporal imagery is advanta- ical assessments to broader scale studies. Therefore we evaluate
geous over single-date images for the distinction of spectrally how i) phenological parameters can be derived at a 30 m spatial
similar vegetation types (e.g. Mesquita Junior, 2000; Mueller et al., resolution, ii) which phenological differences between the main
2015). Cerrado vegetation physiognomies can be revealed based on these

Fig. 1. Cerrado physiognomies describing gradients of vegetation height and density (adapted from Mesquita Junior (2000)). Hereafter only the Portuguese physiognomy
names will be used. The above ground biomass values of woody vegetation were comprehended from several studies by de Miranda et al. (2014).
 M. Schwieder et al. / International Journal of Applied Earth Observation and Geoinformation 52 (2016) 361–370 363

parameters and iii) if the derived parameters are beneficial for map- distinguishes the physiognomies based on the vegetation height
ping these physiognomies’ spatial distribution. and density (Fig. 1).
Our analysis focused on the classification of the main Cer-
rado physiognomies, i.e. campo limpo (open grassland), campo
sujo (grassland with small shrub patches), campo cerrado (shrub-
2. Data and methods
lands with some trees), cerrado sensu stricto (wooded savanna with
shrubs and trees), cerrado denso (woodland with dense shrubs and
2.1. Study sites
trees), cerradão (dense woodland with high trees) and mata de gale-
ria (gallery forests with dense vegetation and high trees along rivers
The Brazilian Cerrado has a unique appearance, characterized by
and waterbodies).
a mixture of xeromorphic vegetation formations. Variations in abi-
Three study sites within one Landsat footprint were analyzed
otic environmental factors such as soil traits and fertility, long term
that are located in the central Cerrado around Brasília, Federal Dis-
climate fluctuations, and fire events have shaped the landscapes of
trict (Fig. 2). All sites lie within protected areas, thus we could
the Cerrado. The vegetation is influenced by strong climatic sea-
ensure that the anthropogenic interference on the land cover is
sonality with a dry season between May and September in which
negligible. The first site covers the whole extent (approx. 29,000 ha)
most of the Cerrado plant species flower, renew their leaves and
of the Brasília National Park (PNB), which encompasses the major
spread their seeds, before they germinate at the beginning of the
savanna formations encountered in the Cerrado, including the
wet season (de Faria et al., 2012), which lasts from October to
transitions from the dominant herbaceous stratum (campo cer-
April. However, many of the woody plant species in the Cerrado
rado) to the more complex, woody dominated stratum (cerrado
are evergreen with less seasonal variations that develop new leaves
sensu stricto). The second study site is the Gama Cabeça-de-Veado
at the end of the dry season and reach their peak in the wet sea-
Environmental Protection Area (APAGCV), which covers around
son (Ratana et al., 2005). These environmental conditions result
20,000 ha of the Cerrado south of Brasília. Even though parts of
in landscape patterns composed of open grass, shrub dominated
APAGCV’s natural vegetation have already been converted, it con-
lands and dense forests, which define a gradient of vegetation den-
tains undisturbed areas in the Brazilian Institute of Geography and
sity, growth form, vertical structure and biomass (Oliveira-Filho
Statistics (IBGE) Ecological Reserve, Brasília’s Botanical Garden and
and Ratter, 2002; Ribeiro and Walter, 2008). Several authors have
an experimental farm that belongs to the University of Brasília.
attempted to define a range of classification schemes that enable
These areas are characterized by the main Cerrado grassland, shrub-
to describe the vegetation gradient in distinct physiognomy classes
land and woodland physiognomies including patches of cerrado
(Oliveira-Filho and Ratter, 2002). In this study we follow the clas-
denso and cerradão. The third study site is the Águas Emendadas
sification scheme proposed by Ribeiro and Walter (2008), which

Fig. 2. Study sites surrounding Brasília, Federal District. Brasília National Park (PNB), Águas Emendadas Ecological Station (ESECAE) and Gama Cabeça de Veado Environmental
Protection Area (APAGCV) depicted in dark blue. The overview map shows the national territory of Brazil (green) located in South America (grey) as well as the extent of
the Brazilian Cerrado (light green). The square in the overview map depicts the location of the three study sites. (For interpretation of the references to colour in this figure
legend, the reader is referred to the web version of this article.)
364 M. Schwieder et al. / International Journal of Applied Earth Observation and Geoinformation 52 (2016) 361–370

Fig. 3. Schematic workflow of the mapping approach. Landsat pre-processing and TC transformation, followed by RBF convolution filter for outlier detection and a RBF
convolution filters ensemble to fill data gaps. Phenological parameters were derived for the 2009/10 season and used as input for a SVM classification.

Ecological Station (ESECAE), a protected area northeast of Brasília (TCbright ) and wetness (TCwet ) components using the TC coefficients
with an extent of approximately 8000 ha. Next to open Cerrado for top-of-atmosphere reflectance, recommended for applications
types (mostly shrubland) and denser forest formations, large areas where atmospheric correction approaches are not feasible (Huang
of cerrado sensu stricto characterize the study site, which is the et al., 2002b). The transformation into TC components has been
most prevalent and complex Cerrado physiognomy, with tree cover shown to be useful for phenology-based classification approaches
ranging from 20% to 70% (Ribeiro and Walter, 2008). (e.g. Dymond et al., 2002).
To approximate the given clear TC observations into a dense 8-
days-sampled time series without data gaps, we used a pixel based
2.2. Time series and phenological patterns Gaussian convolution filter approach (Fig. 3). This approach is com-
parable to a moving window average filter (in time), whereas the
In order to create a dense Landsat time series, we combined all weights are given by a Gaussian function:
available Level 1 Terrain (L1T) geometric corrected TM, ETM+ and
 x− 2
OLI data over the study area (path: 221, row: 071) acquired between 1 −1 
2000 and 2014 with a cloud cover of up to 90%, resulting in 431 f (x) = √ e 2 (1)
 2
potential observations in an 8-day interval. However, the climatic
seasonality in the Cerrado leads to cloud contamination especially with ␴ defining the shape of the kernel function. After an expert
in the wet season and thus to a low clear data availability (Sano et al., driven visual inspection we used a smooth Gaussian kernel with
2007). Another limiting factor for a dense time series is enforced by ␴ = 5 to detect outliers in our observations, which are mostly related
sensor errors, such as the scan line corrector failure in the case of the to clouds or cloud shadows. The kernel includes all coefficients
Landsat 7 ETM+. To overcome these drawbacks, we used a weighted of the +/− 3 × ␴ interval, i.e. an overall size of 31 observations
ensemble of Radial Basis Function (RBF) convolution filters to detect (15 to the left and 15 to the right), resulting in a kernel size
outliers and approximate missing data in a Landsat time series. of 31 × 8 days = 248 days. All observations that were more dis-
Landsat data were converted to top-of-atmosphere reflectance val- tant than one standard deviation from the fitted function were
ues and clouds and cloud shadows were masked using the FMASK defined as outliers and eliminated. To derive the final pheno-
algorithm (Zhu and Woodcock, 2012). The converted data were logical profiles, we used an ensemble of three Gaussian kernel
transformed to tasseled cap (TC) greenness (TCgreen ), brightness based convolution approximations (␴ = 3 × 8, 5 × 8, 7 × 8 days). The
 M. Schwieder et al. / International Journal of Applied Earth Observation and Geoinformation 52 (2016) 361–370 365

Fig. 4. Averaged TC phenological profiles for each physiognomy in the season 2009–10.

final approximation is the weighted average of the three tempo- algorithm, whose underlying principle is to define an optimal
ral convolution filters, whereas weights are dependent on clear hyperplane that separates sample points based on the given train-
observations within the given kernels. To describe the course of ing points (Huang et al., 2002a), in our case a set of phenological
the smoothed continuous TC profiles on a per-pixel basis, we trun- parameters, and is then applied to the area of interest. The SVM
cated the approximations and derived phenological parameters for transforms the input data into a high-dimensional feature space by
the season 2009–2010. For each of the three TC time series 9 sea- applying a kernel-function to solve a non-linear classification prob-
sonal parameters were calculated using the TIMESAT 3.1 software lem. The SVM model was optimized by defining the Gaussian kernel
(Jönsson and Eklundh, 2004). Parameters included day-of-the-year width (␥) and the regularization parameter C via a cross-validated
(DOY) or start, mid, end, and length of season and phenological grid search. We used the imageSVM implementation based on LIB-
proxies like peak and base value, seasonal amplitude or rate of SVM (Chang and Lin, 2011), as available in the EnMAP-Box Version
increase and decrease. Detailed information on the calculation of 2.03 (van der Linden et al., 2015). The SVM model was trained with a
TIMESAT parameters are found in Jönsson and Eklundh (2004). set of stratified random sampled pixel, accounting for feature vari-
ability in all classes (Fig. 3). A physiognomy map for the PNB site
(Ferreira et al., 2007), lastly updated in 2012 by visual interpretation
2.3. Mapping vegetation gradients
of high resolution imagery from 2009, and a map of the IBGE Eco-
logical Reserve, were aggregated to the classes of interest and used
To map the spatial patterns of Cerrado physiognomies, we
as a base map for reference data. For the remaining areas of APAGCV
applied a support vector machine classification (SVM) algorithm
and ESECAE, we used a stratified green peak TCgreen vegetation
(Vapnik, 1998). The SVM is a non-parametric machine learning

Fig. 5. Boxplots showing the mean, 25 and 75 percentile of selected TCgreen parameters (base value, end-of-season, rate of increase and decrease) for five different physiog-
nomies based on the training pixels. Whiskers have the length of 1.5 times the inter-quartile range, beyond which all other data are considered outliers.
366 M. Schwieder et al. / International Journal of Applied Earth Observation and Geoinformation 52 (2016) 361–370

Fig. 6. Classification results based on seasonal phenological parameters for A) PNB, B) IBGE parts of APAGVC for which a reference map was available and C) ESCEA. D) and
E) show the physiognomy patterns in the reference maps.

map from January 2010 as a proxy for vegetation density (Leitão est in TCbright ) and were followed by denser physiognomies, with
et al., 2015). All sampled pixels were manually labelled to a phys- cerradão showing the highest values (lowest in TCbright ).
iognomy class by visual interpretation of high resolution Google All profiles had a shift of the phenological peaks between differ-
Earth imagery. Pixel that were partly covered by “non-natural” fea- ent physiognomies. This is particularly evident in the TCgreen profile,
tures such as roads or which could not be adequately labelled e.g. where phenological peaks shifted around 3 months between dif-
due to missing imagery from that time period were excluded from ferent physiognomies. Denser physiognomies reached their peak
the training set, which resulted in a total of 1897 trainings pixels earlier than sparse ones and created flatter profiles, as there was
(campo limpo: 230; campo sujo: 451; campo cerrado: 425; cerrado less phenological variation due to e.g. larger degrees of evergreen
sensu stricto: 401; cerrado denso: 181; cerradão: 64; mata de gale- vegetation. These observations revealed the expected patterns
ria: 145). The classification accuracy was assessed through a 10-fold while being in line with findings of other studies using data with
cross-validation. coarser spatial resolution (Ferreira and Huete, 2004; Ratana et al.,
2005). Our fitted temporal profiles allow the reconstruction of Land
Surface Phenology (LSP) in savanna ecosystems at the Landsat spa-
3. Results and discussion
tial resolution, which has to our knowledge not been shown before.
3.1. Phenological profiles
3.2. Phenological metrics
We derived pixel-wise TCgreen, TCbright and TCwet phenologi-
cal profiles from the filtered Landsat time series for the period We used the derived profiles to calculate pixel-wise pheno-
2009–2010. Fig. 4 shows averaged profiles for each Cerrado phys- logical parameters using TIMESAT for the season 2009/2010 to
iognomy based on the training pixels and reveal a general seasonal describe the course of the phenological profiles. The distribution
trend, resembling the dry season from May to September and the of these parameters in each physiognomy stressed their pheno-
rainy season from October to April, which was especially prominent logical differences. Fig. 5 shows an example of the distribution
in the TCgreen profiles. In general the TCgreen, TCbright and TCwet pro- of selected TCgreen phenological parameters for the main Cerrado
files differed in shape and amplitude, but they all showed the same physiognomies based on the set of trainings pixels. In most cases,
order from high to low values between different physiognomies, the physiognomies can be clearly distinguished. The base value,
with TCbright showing the opposite direction of TCgreen and TCwet . which was defined as the average of the two seasonal minima
Physiognomies with sparse vegetation densities i.e. campo limpo (Jönsson and Eklundh, 2004), revealed a clear trend of generally
and campo sujo (lighter colors) produced the lowest values (high- higher minima with denser vegetation physiognomies. This pat-
 M. Schwieder et al. / International Journal of Applied Earth Observation and Geoinformation 52 (2016) 361–370 367

Fig. 7. Close-ups of the PNB classification map and three TCgreen parameters in RGB that depict gradual transitions between physiognomies.

tern in the observed LSP signal can be related to less above ground though the start-of-season is an important phenological parameter,
biomass in the open landscape formations such as campo limpo integrated LSP derived from an 8-daily remote sensing time series
and campo sujo, which naturally increases with denser vegeta- are not necessarily in line with in-situ phenological observations
tion physiognomies (de Miranda et al., 2014). Highest values can (Hanes et al., 2014).
be observed for cerradão that can consist of up to 15 m tall trees
with crown covers of around 90% (Ribeiro and Walter, 2008). Sim-
ilar trends were observable for the rates of increase and decrease 3.3. Vegetation mapping
calculated as the ratio of the difference between 20% and 80% at
the beginning (greening-up) and end (browning) of the phenolog- Based on the set of phenological parameters we trained a SVM
ical curve and the related difference in time (Jönsson and Eklundh, classification model to map the spatial distribution of the main
2004). The differences between each physiognomy were particu- Cerrado physiognomies. The broad-scale spatial patterns of all
larly pronounced in the rates of decrease, which can be related to observed physiognomies were well depicted and in line with avail-
less amplitude in dense physiognomies stressing the buffer effect able reference maps (Fig. 6). Large areas with sparse vegetation
of woody vegetation to seasonality (Ratana et al., 2005). physiognomies (from campo limpo to campo cerrado) were preva-
The TCgreen start-of-season parameter, defined as the point in lent in PNB with gallery forests following the river flows. Areas with
time when the phenological curve has reached 10% of its range dense physiognomy classes like cerrado denso and cerradão were
since the minimum (Jönsson and Eklundh, 2004), did not follow almost not present. The physiognomies in APAGCV showed simi-
this clear trend and showed less differences between the phys- lar patterns with gallery forests along the water bodies, sparsely
iognomies (Fig. 5). The woodland physiognomy cerradão has the vegetated areas in the center and denser physiognomies at the
earliest start of the season at around DOY 263 (20th of September border of the study area. Also small patches of cerrado denso and
2009), whereas the open landscape campo limpo season starts later cerradão occurred. In contrast to the other two study sites ESECAE
around DOY 279 (6th of October 2009). The seasonal start of the was characterized by denser vegetation. Most of the area was clas-
other physiognomies falls somewhere in between these dates and sified as cerrado sensu stricto and cerrado denso, respectively. Areas
differs mainly in the interquartile range. The median lies around of sparsely vegetated campo limpo and campo sujo were located in
DOY 271 (28th of September 2009) for all physiognomies. Even the center as well as in the western parts of ESECAE. Small patches
of cerradão were classified within the cerrado denso regions.
368 M. Schwieder et al. / International Journal of Applied Earth Observation and Geoinformation 52 (2016) 361–370

Table 1
Confusion matrix resulting from the 10-fold cross validation, including users (UA), producers (PA) and overall accuracies (OA). The numbers in parentheses show the soft
validation results, allowing misclassifications in adjacent classes.

Reference Class
Map class Campo limpo Campo sujo Campo cerrado Cerrado sensu Cerrado denso Cerradão Mata de galeria Sum UA Area (ha)
stricto

Campo limpo 101 (222) 49 (0) 2 1 0 0 1 154 (226) 0.66 (0.98) 5907
Campo sujo 121 (0) 290 (433) 80 (0) 16 0 1 0 508 (450) 0.57 (0.96) 16189
Campo cerrado 5 94 (0) 269 (421) 81 (0) 2 0 2 453 (430) 0.59 (0.98) 12611
Cerrado sensu stricto 2 18 72 (0) 264 (377) 61 (0) 2 7 426 (406) 0.62 (0.93) 12906
Cerrado denso 0 0 2 32 (0) 109 (174) 26 (0) 2 171 (178) 0.64 (0.98) 3748
Cerradão 0 0 0 3 4 (0) 29 (61) 8 (0) 44 (64) 0.66 (0.95) 450
Mata de galeria 1 0 0 4 5 6 (0) 125 (133) 141 (143) 0.89 (0.93) 3823
Sum 230 451 425 401 181 64 145 1897

PA 0.44 (0.97) 0.64 (0.96) 0.63 (0.99) 0.66 (0.94) 0.6 (0.96) 0.45 (0.95) 0.86 (0.92) OA 0.63 (0.96) 55634

Linear artefacts occurred in all three study sites but were most (MSI) onboard Sentinel-2a already provide high quality data for
prominent in ESECAE, predominantly in regions of SLC-off errors. time series analysis. Their combined use, especially when includ-
The results suggest that not only the total amount of available cloud ing future Sentinel-2b data, will potentially allow to generate dense
and error free data is relevant for our mapping approach, but also times series with an approximate 5-day temporal resolution, mak-
its seasonal distribution. This is in line with findings of other studies ing synthetic data gap filling obsolete. Issues like linear artefacts,
that also distinguished Cerrado land cover classes based on Landsat as currently present in the classification results, will then be over-
time series (e.g. Mueller et al., 2015; Rufin et al., 2015). It seems to come. However, the combination of different data types is still
be rather important that the peak of the season is covered by data, challenging, e.g. due to different spatial and spectral resolutions
as it directly influences the derivation of phenological parameters, (Hostert et al., 2015; Wulder et al., 2015). Especially, the red-edge
which in turn determine the classification outcome. However, this bands of Sentinel-2 (ca. 705, 740 and 783 nm) with a spatial resolu-
effect is likely to be class-dependent, as it did not lead to misclas- tion of 20 m hold potential for vegetation analyses (Verrelst et al.,
sification throughout all physiognomies. 2012) and further research is required to better comprehend this
A detailed look at the classification results revealed that not potential. The assets of narrow spectral bands for the analysis of
many regions in all three study sites are classified as pure physiog- heterogeneous ecosystems have also been demonstrated in several
nomies and occasionally include pixels from the adjacent class. This studies that used for example simulated satellite data based on air-
“salt and pepper” effect (Figs. 6 and 7) reflects the heterogeneity borne hyperspectral imagery (e.g. Roberts et al., 2015; Schwieder
within the complex ecosystem and is likely related to mixed pixel et al., 2014) or hyperspectral data from experimental satellites such
between adjacent physiognomy classes, which from a remote sens- as Hyperion (e.g. Leitão et al., 2015). These data enable the deriva-
ing perspective are different shares of understory and canopy cover tion of biophysical vegetation parameters (e.g. Souza et al., 2010) or
on a pixel level. This effect is particularly prominent when looking the spectral unmixing of sub-pixel fractions of photosynthetic and
at gradients of vegetation density in heterogeneous landscapes at non-photosynthetic active vegetation (e.g. Miura et al., 2003). On
30 m spatial resolution. Fig. 7 shows examples of spatially-explicit the one hand, this additional information can improve classification
phenological differences and gradual transitions between physiog- outputs, e.g. in combination with phenological data. On the other
nomies and how those translate into classes according to a standard hand, continuous biophysical parameters can be quantified through
classification scheme for Cerrado landscapes used in Brazil (Ribeiro regression approaches, if appropriate reference data are available.
and Walter, 2008). Such strategies allow overcoming the limitations of hard classifi-
The final map was validated with a 10-fold cross-validation cations in heterogeneous ecosystems. With the advent of future
resulting in an overall accuracy of 63% (Table 1). This relatively spaceborne hyperspectral sensors such as EnMAP (Guanter et al.,
weak accuracy measure does not agree with the visual comparison 2015) or HyspIRI (Lee et al., 2015), the potential of regression-based
of the reference map and the classification outcome. It is a chal- approaches will unfold.
lenging task to map and validate a classification stemming from
a field-based classification scheme with a classification at Land-
4. Conclusion
sat resolution. We therefore applied a softer validation scheme
to accommodate for spectral-temporal class similarities between
Dense Landsat time series were used to analyze phenological
neighboring classes along the vegetation gradient (e.g. between
differences between the main physiognomies of the Brazilian Cer-
campo sujo and campo limpo). The overall accuracy then reached
rado. Data gaps within the time series were synthetically filled
surprisingly high 96% (Table 1; numbers in parentheses). This
using a weighted ensemble of radial basis convolution filters. The
pattern was also reflected in the SVM class probabilities, which
available amount of data was sufficient to derive comprehensible
revealed that approximately 90% of all pixel had the second highest
phenological profiles and parameters, which enabled depicting dif-
probability in the thematically adjacent class. Emphasizing the dif-
ferences between the main Cerrado vegetation physiognomies in
ficulties in discrete classifications for mapping vegetation gradients
terms of their response to climatic-dependent phenology. Based
with a high spatial resolution.
on these differences, it was possible to map the main spatial pat-
terns of the Cerrado physiognomies using a SVM classification
3.4. Outlook approach. Linear artefacts within the classification result could be
related to sparse data availability (i.e. systematic errors within the
Our results showed that the proposed approach allows map- data) and its seasonal distribution. This drawback will be over-
ping heterogeneous savanna ecosystems. Cloud and error free data come in the near future when e.g. Landsat 8 OLI and recently
are a mandatory prerequisite to generate the dense time series available Sentinel-2 MSI data will be combined. Inaccuracies in
needed to derive phenological information. Recently launched sen- the final estimation were especially prominent between themat-
sors such as OLI on Landsat 8 and the Multispectral Instrument ically adjacent classes along the vegetation gradient. Incorporating
 M. Schwieder et al. / International Journal of Applied Earth Observation and Geoinformation 52 (2016) 361–370 369

(multi-temporal) spaceborne hyperspectral data, or derived prod- Guanter, L., Kaufmann, H., Segl, K., Foerster, S., Rogass, C., Chabrillat, S., Kuester, T.,
ucts, in the analysis will further be beneficial to overcome this Hollstein, A., Rossner, G., Chlebek, C., Straif, C., Fischer, S., Schrader, S., Storch,
T., Heiden, U., Mueller, A., Bachmann, M., Mühle, H., Müller, R., Habermeyer,
limitation. This will become even more feasible once upcoming M., Ohndorf, A., Hill, J., Buddenbaum, H., Hostert, P., van der Linden, S., Leitão,
sensors such as EnMAP or HyspIRI are in orbit and routinely P.J., Rabe, A., Doerffer, R., Krasemann, H., Xi, H., Mauser, W., Hank, T., Locherer,
and repeatedly provide spectral high-resolution data. Our results M., Rast, M., Staenz, K., Sang, B., 2015. The EnMAP spaceborne imaging
spectroscopy mission for earth observation. Remote Sens. 7, 8830.
emphasize the usefulness of utilizing long-term Landsat time series Hanes, J., Liang, L., Morisette, J., 2014. Land surface phenology. In: Hanes, J.M. (Ed.),
for the analysis of heterogeneous ecosystems, where spatial detail Biophysical Applications of Satellite Remote Sensing. Springer, Berlin,
is as much needed as the ability to derive important phenologi- Heidelberg, pp. 99–125.
Hostert, P., Griffiths, P., van der Linden, S., Pflugmacher, D., 2015. Time series
cal parameters. Such time series analyses offer new opportunities
analyses in a new era of optical satellite data. In: Kuenzer, C., Dech, S., Wagner,
e.g. for monitoring trends in habitat changes and related biodiver- W. (Eds.), Remote Sensing Time Series. Springer International Publishing, pp.
sity or ecosystem responses to climate change. Further research 25–41.
Huang, C., Davis, L.S., Townshend, J.R.G., 2002a. An assessment of support vector
may accordingly focus on the combined use of open data archives
machines for land cover classification. Int. J. Remote Sens. 23, 725–749.
and upcoming spaceborne sensor technologies, which enable deep- Huang, C., Wylie, B., Yang, L., Homer, C., Zylstra, G., 2002b. Derivation of a tasselled
ening our understanding of ongoing processes in the Cerrado and cap transformation based on Landsat 7 at-satellite reflectance. Int. J. Remote
comparable heterogeneous ecosystems. Nevertheless, our results Sens. 23, 1741–1748.
Jönsson, P., Eklundh, L., 2004. TIMESAT—a program for analyzing time-series of
also suggest that the use of “hard” classification approaches is satellite sensor data. Comput. Geosci. UK 30, 833–845.
not always adequate for mapping vegetation gradients in hetero- Lee, C.M., Cable, M.L., Hook, S.J., Green, R.O., Ustin, S.L., Mandl, D.J., Middleton, E.M.,
geneous ecosystems, due to a high structural and spectral class 2015. An introduction to the NASA Hyperspectral InfraRed Imager (HyspIRI)
mission and preparatory activities. Remote Sens. Environ. 167, 6–19.
similarity. This was shown to be particularly challenging when dis- Lehmann, C.E.R., Archibald, S.A., Hoffmann, W.A., Bond, W.J., 2011. Deciphering the
tinguishing adjacent (i.e. most similar) classes on the structural distribution of the savanna biome. New Phytol. 191, 197–209.
gradient along the different Cerrado physiognomies. Depending Leitão, P.J., Schwieder, M., Suess, S., Okujeni, A., Galvão, L., Linden, S., Hostert, P.,
2015. Monitoring natural ecosystem and ecological gradients: perspectives
on the purpose of the derived results and data availability, quan- with EnMAP. Remote Sens. 7, 13098–13119.
titative approaches avoiding classes, e.g. regression approaches, MEA, 2005. Ecosystems and Human Well-being Millenium Ecosystem Assessment.
would offer a solution to overcome shortcomings of classification Island Press, Washington, DC.
Marchant, R., 2010. Understanding complexity in savannas: climate, biodiversity
approaches. Even though these approaches often require a more
and people. Curr. Opin. Environ. Sustain. 2, 101–108.
costly training, continuous results will improve e.g. biomass or Mesquita Junior, H.N., 2000. NDVI measurements of neotropical savanna
carbon retention estimations. physiognomies a gradient of: biomass, structure and phenology changes. Int.
Arch. Photogramm. Remote Sens. 33, 93–100.
Miura, T., Huete, A.R., Ferreira, L.G., Sano, E.E., 2003. Discrimination and
Acknowledgments biophysical characterization of Cerrado physiognomies with EO-1
Hyperspctral Hyperion. In: Green, R.E. (Ed.), 12th JPL Airborne Earth Science
Workshop, NASA, Jet Propulsion Laboratory, California Institute of Technology.
This study is part of the research activities of the EnMAP Sci- Passadena, CA, pp. 207–212.
entific Advisory Group (EnSAG), and was funded by the German Mueller, H., Rufin, P., Griffiths, P., Siqueira, A.J.B., Hostert, P., 2015. Mining dense
Landsat time series for separating cropland and pasture in a heterogeneous
Aerospace Centre (DLR) — Project Management Agency, granted by Brazilian Savanna landscape. Remote Sens. Environ. 156, 490–499.
the Ministry of Economics and Technology (BMWi grant 50EE1309). Myers, N., Mittermeier, R.A., Mittermeier, C.G., da Fonseca, G.A.B., Kent, J., 2000.
The authors would like to thank Manuel Ferreira and Betânia Goes Biodiversity hotspots for conservation priorities. Nature 403, 853–858.
Oliveira-Filho, A.T., Ratter, J.A., 2002. Vegetation physiognomies and woody flora of
for providing the reference maps and Philippe Rufin for inspiring
the Cerrado biome. In: Oliveira, P.E.M., Marquis, R.J. (Eds.), The Cerrados of
and helpful discussions. This research also contributes to the Global Brazil—Ecology and Natural History of a Neotropical Savanna. Columbia
Land Project (http://www.globallandproject.org) and the Land- University Press, New York.
Ratana, P., Huete, A.R., Ferreira, L., 2005. Analysis of cerrado physiognomies and
sat Science Team (http://landsat.usgs.gov/Landsat Science Team
conversion in the MODIS seasonal-temporal domain. Earth Interact. 9, 1–22.
2012-2017.php). Ratter, J.A., Ribeiro, J.F., Bridgewater, S., 1997. The Brazilian Cerrado vegetation and
threats to its biodiversity. Ann. Bot. 80, 223–230.
Ribeiro, J.F., Walter, B.M.T., 2008. As principais fitofisionomias do bioma Cerrado.
Appendix A. Supplementary data In: Sano, S.M., Almeida, S.P., Ribeiro, J.F. (Eds.), Cerrado: Ecologia e Flora.
Embrapa Cerrados, Brasilia, DF, pp. 153–212.
Roberts, D.A., Dennison, P.E., Roth, K.L., Dudley, K., Hulley, G., 2015. Relationships
Supplementary data associated with this article can be found, in between dominant plant species, fractional cover and land surface temperature
the online version, at http://dx.doi.org/10.1016/j.jag.2016.06.019. in a Mediterranean ecosystem. Remote Sens. Environ. 167, 152–167.
Rocha, G.F., Ferreira, L.G., Ferreira, N.C., Ferreira, M.E., 2011. Detecção De
Desmatamentos No Bioma Cerrado Entre 2002 E 2009: Padrões, Tendências E
References Impactos. Rev. Bras. de Cartografia 63, 341–349.
Rufin, P., Muller, H., Pflugmacher, D., Hostert, P., 2015. Land use intensity
Chang, C.-C., Lin, C.-J., 2011. LIBSVM, 2. ACM Transactions on Intelligent Systems trajectories on Amazonian pastures derived from Landsat time series. Int. J.
and Technology, pp. 1–27. Appl. Earth Obs. Geoinf. 41, 1–10.
Dymond, C.C., Mladenoff, D.J., Radeloff, V.C., 2002. Phenological differences in Sano, E.E., Ferreira, L.G., Huete, A.R., 2005. Synthetic aperture radar (L band) and
tasseled cap indices improve deciduous forest classification. Remote Sens. optical vegetation indices for discriminating the Brazilian savanna
Environ. 80, 460–472. physiognomies: a comparative analysis. Earth Interact. 9, 1–15.
Ferreira, L.G., Huete, A.R., 2004. Assessing the seasonal dynamics of the Brazilian Sano, E.E., Ferreira, L.G., Asner, G.P., Steinke, E.T., 2007. Spatial and temporal
Cerrado vegetation through the use of spectral vegetation indices. Int. J. probabilities of obtaining cloud-free Landsat images over the Brazilian tropical
Remote Sens. 25, 1837–1860. savanna. Int. J. Remote Sens. 28, 2739–2752.
Ferreira, M.E., Ferreira, L.G., Sano, E.E., Shimabukuro, Y.E., 2007. Spectral linear Sano, E.E., Rosa, R., Brito, J.L., Ferreira, L.G., 2010. Land cover mapping of the
mixture modelling approaches for land cover mapping of tropical savanna tropical savanna region in Brazil. Environ. Monit. Assess. 166, 113–124.
areas in Brazil. Int. J. Remote Sens. 28, 413–429. Schwieder, M., Leitão, P.J., Suess, S., Senf, C., Hostert, P., 2014. Estimating fractional
Ferreira, M.E., Ferreira, L.G., Miziara, F., Soares-Filho, B.S., 2012. Modeling shrub cover using simulated EnMAP data: a comparison of three machine
landscape dynamics in the central Brazilian savanna biome: future scenarios learning regression techniques. Remote Sens. 6, 3427–3445.
and perspectives for conservation. J. Land Use Sci. 8, 403–421. Silva, J.M.C., Bates, J.M., 2002. Biogeographic patterns and conservation in the
Foley, J.A., Ramankutty, N., Brauman, K.A., Cassidy, E.S., Gerber, J.S., Johnston, M., south American Cerrado: a tropical Savanna Hotspo. BioScience 52, 225–234.
Mueller, N.D., O’Connell, C., Ray, D.K., West, P.C., Balzer, C., Bennett, E.M., Solbrig, O.T., 1996. The diversity of the Savanna ecosystem. In: Solbrig, O.T.,
Carpenter, S.R., Hill, J., Monfreda, C., Polasky, S., Rockstrom, J., Sheehan, J., Medina, E., Silva, J.F. (Eds.), Biodiversity and Savanna Ecosystem Processes.
Siebert, S., Tilman, D., Zaks, D.P., 2011. Solutions for a cultivated planet. Nature Springer-Verlag, Berlin, Heidelberg, pp. 1–27.
478, 337–342. Souza, A.A., Galvao, L.S., Santos, J.R., 2010. Relationships between hyperion-derived
Furley, P.A., 1999. The nature and diversity of neotropical savanna vegetation with vegetation indices, biophysical parameters, and elevation data in a Brazilian
particular reference to the Brazilian Cerrados. Global Ecol. Biogeogr. 8, savannah environment. Remote Sens. Lett. 1, 55–64.
223–241.
370 M. Schwieder et al. / International Journal of Applied Earth Observation and Geoinformation 52 (2016) 361–370

Staver, A.C., Archibald, S., Levin, S.A., 2011. The global extent and determinants of Zhu, Z., Woodcock, C.E., 2012. Object-based cloud and cloud shadow detection in
savanna and forest as alternative biome states. Science 334, 230–232. Landsat imagery. Remote Sens. Environ. 118, 83–94.
Vapnik, V., 1998. Statistical Learning Theory. Wiley, New York. de Faria, R.A.P.G., Coelho, M.d.F.B., de Figueiredo e Albuquerque, M.C., de Azevedo,
Verrelst, J., Muñoz, J., Alonso, L., Delegido, J., Rivera, J.P., Camps-Valls, G., Moreno, J., R.A.B., 2012. Phenology of Species in the Cerrado of Mato Grosso State,
2012. Machine learning regression algorithms for biophysical parameter Brazil-Brosimum Gaudichaudii Trécul (Moraceae). INTECH Open Access
retrieval: opportunities for sentinel-2 and -3. Remote Sens. Environ. 118, Publisher.
127–139. de Miranda, M., Bustamante, S.d.C., Palace, M., Hagen, S., Keller, M., Ferreira, L.G.,
Wulder, M.A., Masek, J.G., Cohen, W.B., Loveland, T.R., Woodcock, C.E., 2012. 2014. Regional variations in biomass distribution in Brazilian Savanna
Opening the archive: how free data has enabled the science and monitoring Woodland. Biotropica 46, 125–138.
promise of Landsat. Remote Sens. Environ. 122, 2–10. van der Linden, S., Rabe, A., Held, M., Jakimow, B., Leitão, P.J., Okujeni, A., Schwieder,
Wulder, M.A., Hilker, T., White, J.C., Coops, N.C., Masek, J.G., Pflugmacher, D., M., Suess, S., Hostert, P., 2015. The EnMAP-Box—A Toolbox and Application
Crevier, Y., 2015. Virtual constellations for global terrestrial monitoring. Programming Interface for EnMAP Data Processing. Remote Sens. 7, 11249.
Remote Sens. Environ. 170, 62–76.
Zhang, X., Friedl, M.A., Schaaf, C.B., Strahler, A.H., Hodges, J.C.F., Gao, F., Reed, B.C.,
Huete, A., 2003. Monitoring vegetation phenology using MODIS. Remote Sens.
Environ. 84, 471–475.