Geocarto International
ISSN: 1010-6049 (Print) 1752-0762 (Online) Journal homepage: http://www.tandfonline.com/loi/tgei20
Evaluating the potential of the red edge channel
for C3 (Festuca spp.) grass discrimination using
Sentinel-2 and Rapid Eye satellite image data
Charles Otunga, John Odindi, Onisimo Mutanga & Clement Adjorlolo
To cite this article: Charles Otunga, John Odindi, Onisimo Mutanga & Clement Adjorlolo
(2018): Evaluating the potential of the red edge channel for C3 (Festuca spp.) grass
discrimination using Sentinel-2 and Rapid Eye satellite image data, Geocarto International, DOI:
10.1080/10106049.2018.1474274
To link to this article: https://doi.org/10.1080/10106049.2018.1474274
Accepted author version posted online: 08
May 2018.
Published online: 24 May 2018.
Submit your article to this journal
Article views: 15
View related articles
View Crossmark data
Full Terms & Conditions of access and use can be found at
http://www.tandfonline.com/action/journalInformation?journalCode=tgei20
Geocarto InternatIonal, 2018
https://doi.org/10.1080/10106049.2018.1474274
Evaluating the potential of the red edge channel for C3 (Festuca
spp.) grass discrimination using Sentinel-2 and Rapid Eye satellite
image data
Charles Otungaa
, John Odindia, Onisimo Mutangaa and Clement Adjorlolob
a
School of agricultural, earth & environmental Sciences, Discipline of Geography, University of KwaZulu-natal,
Pietermaritzburg, South africa; bSouth african national Space agency (SanSa) – earth observation, Pretoria, South
africa
ABSTRACT
ARTICLE HISTORY
Integrating the Red Edge channel in satellite sensors is valuable for plant
species discrimination. Sentinel-2 MSI and Rapid Eye are some of the new
generation satellite sensors that are characterized by finer spatial and
spectral resolution, including the red edge band. The aim of this study was
to evaluate the potential of the red edge band of Sentinel-2 and Rapid Eye,
for mapping festuca C3 grass using discriminant analysis and maximum
likelihood classification algorithms. Spectral bands, vegetation indices and
spectral bands plus vegetation indices were analysed. Results show that the
integration of the red edge band improved the festuca C3 grass mapping
accuracy by 5.95 and 4.76% for Sentinel-2 and Rapid Eye when the red edge
bands were included and excluded in the analysis, respectively. The results
demonstrate that the use of sensors with strategically positioned red edge
bands, could offer information that is critical for the sustainable rangeland
management.
received 6 December 2017
accepted 27 april 2018
KEYWORDS
red edge band; c3 Festuca;
grass species mapping;
discriminant analysis
Introduction
Grasslands cover approximately 30% of the Earth’s surface and provide a range of critical socio-economic and ecological services (Eriksson et al. 1995; Cousins and Eriksson 2002; Gardi et al. 2002;
Critchley et al. 2004).These include; provision of grazing, wildlife habitat, control of soil erosion,
biological regulation of pests, preservation of species and genetic diversity, recreation, water and
climate change regulation through carbon fixation (Parton et al. 1996; Daily et al. 1997; Leadley et al.
1999; Conant et al. 2001; Flanagan et al. 2002; Li et al. 2004; t Mannetje 2006; Corcoran 2010; O’Mara
2012). In South Africa, grassland is the second largest biome, covering approximately 28.4% of the
country. The country’s grasslands are mainly found in the interior Highveld and higher elevation, and
in patches along the Eastern Cape and KwaZulu-Natal coast (Matsika 2008; O’Connor and Kuyler
2009; Bond and Parr 2010; van Wilgen et al. 2012; Everson and Everson 2016).
Generally, sustainable delivery of the above named services is highly influenced by the stability of
grassland ecosystems. To ensure such sustainability, an understanding of key drivers to their productivity, as well as approaches to determine such productivity is necessary (Black 2006; He 2008). However,
most of these grasslands are commonly characterized by complex grass species stand structure and
CONTACT charles otunga
pesatoango@yahoo.ca, charlesotunga95@gmail.com, 212558378@stu.ukzn.ac.za
© 2018 Informa UK limited, trading as taylor & Francis Group
2
C. OTUNGA ET AL.
abundance, posing a spatial mapping challenge (Mutanga and Rugege 2006; Everson and Everson
2016).
C3 and C4 grasses represent a unique species functional type that influence the functioning and
provision of ecosystem services. The geographic distribution of these grass species is found in the
Southern and Eastern Africa, Middle East and the former Soviet Union, the Java, Aldabra Atoll and
North and Central America (Ellis et al. 1980; Pearcy et al. 1981; Hattersley 1983). They are an important
source of forage for livestock and wildlife and are valuable environmental and biodiversity resources
(Niu et al. 2008; Mansour et al. 2012; Pau and Still 2014; Shoko et al. 2016). The C3/C4 grass photosynthetic pathways represent a unique functional type of species that share physical and chemical
processes, phenological, annual or seasonal and structural properties. Studies have recently highlighted
that changes in climate and land use/land cover continue to impose substantial changes in the C3 and
C4 spatial distribution (Adjorlolo 2013; Xia et al. 2014). These anticipated changes will most likely
result in a shift in their distribution, abundance, forage availability, palatability and composition (Diaz
and Cabido 1997; Adjorlolo et al. 2016). Such changes would likewise affect the intake potential of
livestock, hence influence livestock production and by extension food security in a country (Knox et
al. 2011; Singh et al. 2017).
Studies (e.g. Low and Rebelo 1998; Van Wyk and Van Oudtshoorn 1999; Terry et al. 2008; Botha
et al. 2012) show that the C3 Festuca costata grass species, with low grazing value, continues to proliferate in the high value C4 rangelands in the eastern higher altitudes of South Africa. In the end,
such proliferation would represent a significant threat to rangeland biodiversity, as well as local and
regional ecosystem function (D’Antonio and Vitousek 1992; Rossiter et al. 2003; Zhu et al. 2007; Thomas
2013). Such infiltration may also compromise rural livelihoods as a large proportion of the population
in the region depend on grasslands as the main forage source for livestock (Meadows and Hoffman
2002). Hence, determination of C3 (Festuca spp.) spatial distribution is valuable for understanding
their spatial distribution and optimization of rangeland management (Marshall 2013). Such spatial
information is critical hence required for decision-making and resource planning by stakeholders.
There has been an increased adoption of remotely sensed data in mapping grass vegetation distribution (Ullah et al. 2012; Nestola et al. 2016). This has however not been optimally utilized as adoption of these data in heterogeneous landscapes are limited by their coarse spatial resolution and the
mixed pixel phenomenon. Consequently, low temporal resolution of sensors which do not capture the
variability in the reflectance of different vegetation species (Franke et al. 2012; Darvishzadeh 2015),
further impedes species mapping. Even though hyper spectral data, with higher mapping accuracy
remain popular for C3 Festuca grasses mapping (Liu and Cheng 2011), their high cost per unit area
still remains an impediment to wide adoption as repetitive monitoring is necessary for determining
C3 Festuca grass spatial distribution (Marshall 2013). Therefore, relatively affordable and/or freely
available high spatial multispectral remotely sensed data are necessary for mapping grass species
distribution in rangeland ecosystems at both local and regional levels.
Recently, a suite of new generation sensors like GeoEye, WorldView-2/3, Landsat 8 Operational
Land Imager (OLI), SPOT-7, TerraSAR-X, ALOS-2, Rapid Eye and Sentinel-2 MSI, with mediumto-higher spatial and spectral resolution have emerged. Such sensors, with the additional strategically positioned red edge channels have opened up additional opportunities for improved vegetation
mapping (Cheng and Sustera 2009). Specifically, these new sensors, with finer spatial resolutions and
improved spectral characteristics offer a great opportunity for discriminating vegetation species in
heterogeneous landscapes, hence is seen as a trade-off between benefits offered by multispectral and
hyperspectral remotely sensed data (Adam et al. 2010; Ramoelo et al. 2015; Tarantino et al. 2016; Dhau
et al. 2017; Dube et al. 2017; Sibanda et al. 2017).
Whereas, remotely sensed data and associated techniques have been critical in C3/C4 grassland
mapping, there is paucity in literature that seek to evaluate the influence of the red edge channel in
high spatial resolution multispectral sensors (e.g. Rapid Eye, WorldView-2/3 and Sentinel-2 MSI) for
discrimination and mapping of C3(Festuca spp.) grass species distribution. The value of the additional
Rapid Eye and Sentinel-2 MSI bands, for instance, has been demonstrated in a number of diverse studies
GEOCARTO INTERNATIONAL
3
Figure 1. location of Fort nottingham commonage rangeland study area in in KwaZulu-natal Province, shown with bands 3, 2 and
1 from the rapid eye satellite imagery as background.
that include, mapping alien and indigenous vegetation (Odindi et al. 2016), classifying insect defoliation
levels in an African savanna (Adelabu et al. 2014), land use classification (Schuster et al. 2012), estimation of nitrogen in a savanna ecosystem (Ramoelo et al. 2012) and crop assessment (Eitel et al. 2007;
Forkuor et al. 2017). Hence, these applications provide scope for testing its value in mapping vegetation
species in heterogeneous landscapes. In this study, we compare the performance of a commercial and
a freely available sensor in delineating C3/C4 grass species. Specifically, we evaluate the value of both
Sentinel-2 MSI’s and the Rapid Eye’s red edge channel and associated vegetation indices. The enhanced
spectral (i.e. three red-edge bands) and spatial (10–20-m resolution) capabilities of Sentinel-2 MSI were
tested to ascertain their contribution in improving mapping and discrimination accuracy of C3 Festuca
and coexisting C4 grass species in Fort Nottingham commonage rangeland using DA algorithm, due
to its ability to discriminate and project mapping and classification that is based on performance of
continuous variables in separating specified categories in the classification process.
Materials and methods
Study area
The study was conducted in Fort Nottingham Commonage, Kwazulu-Natal province, South Africa
(Figure 1). The area is between 1519 and 1750 m above sea level and is characterized by a sub-tropical
humid climate with about 950 mm of rain per year, mainly received between the summer months of
October and March (Mucina and Rutherford 2006; Stears and Shrader 2015). The area has moderately
cool temperatures, with mean annual temperature ranging from 14.2 to 15.1 °C. Heavy frosts occur in
34 days (Kruger and Shongwe 2004; Mucina and Rutherford 2006; Boothway et al. 2012). Vegetation
in the study area is homogeneous, characterized by an extensive coverage of the Eastern Mist belt
4
C. OTUNGA ET AL.
Table 1. training and validation/test/evaluation data-sets used.
Training data-set
2/3 (n = 80) of the data
Presence (1) 51 (34)
absence (0) 69 (46)
total 80 (66.7%)
Test/evaluation data-set
1/3 (n = 40) of the data
17
23
40 (33.3%)
Forest and Drakensberg Foothill Moist Grassland that supports a number of rare and threatened
species (Edwards 1983; Mucina and Rutherford 2006). Specifically, vegetation in the area has been
classified as northern KwaZulu-Natal moist Grassland, which is characterized by a forb-rich grassland, dominated by short bunch grasses that include; Themeda triandra, Hyparrhenia hirta, Tristachya
leucothrix, Sporobulus africanus and patches of indigenous bushclumps (Killick 1990; Mucina and
Rutherford 2006; Shaw and Escott 2011; Stears and Shrader 2015). Additionally, the shallow-rooted
cool-season C3 grass species consists of Festuca costata, which coexists with other C3 grass species that
include Poa binate, Agrostis barbulegera and Merxmuellera macowanii. Generally, the Festuca costata
and Merxmuellera macowanii remain green during winter as they have lower water and nitrogen use
efficiency, hence lower radiation use efficiency (Pearcy and Ehleringer 1984). In this regard, Festuca
grass species, like most C3 grasses have shown increased encroachment due to elevated CO2 levels
resulting from high emission of greenhouse gases (Norby and Luo 2004; Langley and Megonigal 2010).
Soils with red and yellow brown apedal B-horizons and lithocutanic B-horizons are common within
the study area (Boothway et al. 2012). As noted by King (2002), these soils are primarily formed on
fine-grained sandstone, shale, siltstone and mudstone which alternate in horizontal progressions.
The commonage rangeland area provides suitable habitat for endemic and priority species such as
Oribi, African Crowned crane, Blue Crane, Samango monkey, Long toed tree frog, Tree hyrax, Grey
Rhebuck, Cape and Bearded Vultures, Serval, Midlands Dwarf Chameleon and Amur falcon (Botha et
al. 2012). A number of grazing paddocks also exist that are utilized for grazing by wild herbivore such
Oribi antelope, similarly the commonage rangeland support cattle grazing from individual farmers
through a lease arrangement that also determine various rangeland or land use intensities. The area
forms part of the Tugela Macro Ecological Corridor that is an important biodiversity corridor linking
the protected areas in the uKhahlamba Drakensberg World Heritage Site, hence critical in protecting
the integrity of the rich fauna and flora within the area and KwaZulu-Natal as a whole. The study area
is also recognized for its local significance to the C3/C4 grass species diversity, hence ideal for the
current study.
Field data collection
A random stratified sampling strategy as explained by Hirzel and Guisan (2002)) was adopted for this
study. Well-distributed random points (n = 120) within the grazing paddocks of the entire commonage
(Figure 1) were generated on an existing GIS layer using ArcMap 10.4 software. The generated points
and their attributes were then randomly divided into 2/3 (66.7%) and 1/3 (33.3%) for training the
models and accuracy assessment, respectively (Table 1).
Field data collection was conducted from 27 April to 5 May 2015 to record the field locations of
C3/C4 grasses and to compliment the remotely sensed data. During data collection, the generated
random sample points were input into a sub-metre handheld Trimble GPS device, which was used
to navigate to the 120 random point locations within the Commonage. At each random point a
10 m × 10 m plot was demarcated. Within this10 m × 10 m demarcated plots, a 1 m × 1 m quadrant
was randomly used to capture the plots grass species variability. A total of 360, 1 m × 1 m subplots
were thus sampled for presence or absence of C3 (Festuca spp.) grass species. Based on the previous
study by Botha et al. (2012), we recognized three dominant C3/C4 grass species types (Festuca costata,
Themeda triandra & Tristachya leucothrix) within the 11 grazing paddocks visited. All the grass species
GEOCARTO INTERNATIONAL
5
that were found to be growing on the random sampling plot were recorded on the basis of a simple
ordinal abundance scale ranging from 0 to 1, where 1 indicated that the C3 Festuca grass species was
present at that plot and 0 that other coexisting C3/C4 grass species were occurring. In each of the
visited random plots, Festuca grass species, e.g. Festuca costata identified was considered presence
(1), while other co-occurring C3/C4 grass species (which included Themeda triandra & Tristachya
leucothrix), observed was considered absence (0). Each 1 m × 1 m sampling site was considered to
have absence or presence of Festuca grass when the identified grass species covered approximately
four-fifths of the target sample plot.
The data collection exercise also involved recording a number of metadata which included; species
count, percentage species count, sample plot’s Geographic or X-, Y-Coordinates, topography (topographic position, slope, aspect and elevation) which were generated from the 5-m digital elevation
model (DEM).
Satellite image data
In this study, we utilized Sentinel-2 MSI and results compared to those that were produced when
Rapid Eye satellite imagery was used. The Sentinel-2 MSI satellite was chosen as it delivers freely and
readily available higher resolution, high-quality deca-metric images that incorporates the red edge
band (Immitzer et al. 2016). The Sentinel-2 image was freely downloaded from the Scientific Data Hub
website (https://scihub.copernicus.eu/).The data were acquired in 13 spectral bands, spanning from
the visible through the near infrared (NIR) and the additional strategically positioned red edge, to
the short wave infra-red (SWIR) at 10, 20 and 60 spatial resolutions (Cole et al. 2014). The Sentinel-2
MSI has four 10-m spatial resolution bands that include Band 2 (490 nm), Band 3 (560 nm), Band 4
(665 nm) and Band 8 (842 nm). Bands acquired at 60-m (coastal aerosol band 1, water vapour band
9 and cirrus band 10) spatial resolution are dedicated for detecting atmospheric features and were
therefore excluded from the analysis (Drusch et al. 2012). We re-sampled the Sentinel-2 20-m bands
to 10-m. The red-edge bands are centred at 705 and 740 nm with a band width of 15 nm and a spatial
resolution of 20 m (Clevers et al. 2017).
A cloud-free commercial multi-spectral Earth observation mission Rapid Eye Ortho rectified (Level
3A product) satellite imagery was acquired on 2 May 2015 to coincide with ground data collection.
Rapid Eye is a multispectral Earth observation mission that includes a constellation of five mini
identical satellites providing multispectral images in the blue, green, red, red edge and near-infrared
spectral domains (Tyc et al. 2005). The mission provides high-resolution multispectral imagery along
with an operational GIS (Geographic Information System) service on a commercial basis. Its ability
to combine attributes of high spatial and temporal resolution sensors make it unique amongst other
Earth observation platforms (Eitel et al. 2007). The rapid Eye sensor image has five optical bands in the
400–850-nm range, including a unique Red Edge band 4 (690–730 nm) Blue band 1 (440–510 nm),
Green band 2 (520–590 nm), Red band 3 (630–685 nm) and NIR band 5 (760–850 nm) (Table 2) and
provides 5-m pixel size at nadir (RapidEye 2016) critical for vegetation characterization. Rapid Eye’s
red edge band is particularly sensitive to chlorophyll content, hence valuable for vegetation characterization. Detailed characteristics of the images used for analysis are presented in Table 2.
Image data pre-processing
Sentinel-2 MSI image was atmospherically corrected using the SEN2COR procedure available in the
Sentinel-2 SNAP (Sentinel Application Platform) toolbox, that converted top-of-atmosphere (TOA)
reflectance into top-of-canopy (TOC) reflectance (Louis et al. 2016). SEN2COR performs a pre-processing of TOA image data and applies a scene classification for atmospheric correction. Atmospheric
bands of Sentinel-2 are used to derive maps of aerosols, water vapour and cirrus clouds (Louis et
al. 2016). The Sentinel-2 atmospheric correction was based on the ATCOR algorithm (Richter and
Schläpfer 2002).
6
C. OTUNGA ET AL.
Table 2. rapid eye and Sentinel-2 MSI satellite image spectral characteristics.
Satellite Image
Sentinel-2 MSI (VIS–nIr)
Sentinel-2 MSI (red-edge–SWIr)
rapid eye
Band
Spatial
Spectral
Wavelength
Band
Description
Blue 2
Green 3
red 4
nIr 8
red-edge 5
red-edge 6
red-edge 7
red-edge 8a
Blue 1
Green 2
red 3
red-edge 4
nIr 5
Resolution (m)
10
10
10
10
20
20
20
20
5
5
5
5
5
Range (nm)
458–523
543–578
650–680
855–875
698–713
733–793
773–793
855–875
440–510
520–590
630–690
690–730
760–850
Centre (nm)
490
560
665
842
705
740
783
865
Width (nm)
20
65
35
30
70
70
55
40
90
The Rapid Eye satellite image was provided pre-processed for radiometric, sensor and geometric
anomalies by the data providers using the digital terrain elevation data (DTED) level 1 Shuttle Radar
Terrain Mission (SRTM). The image was geo-registered to the Universal Transverse Mercator UTM
zone 36 South projection and the geo-rectification validated using orthophotos derived from the
study area’s aerial photograph, 5-m digital elevation model (DEM) and 14 ground truthing points.
The digital numbers of the Rapid Eye satellite image data were converted to radiance and then to
top of atmosphere (TOA) reflectance (see Barsi et al. 2003; Chander et al. 2009). To retrieve surface
reflectance, atmospheric correction was executed using the Improved Dark Object Subtraction (DOS)
technique as described by Chavez (1996) and Chavez (1988). According to Hadjimitsis et al. (2003),
the DOS approach has been found to provide a reasonable correction in the visible bands. It is similarly recommended over more sophisticated techniques that require the acquisition of atmospheric or
other auxiliary data for a cloudless image scene. The pre-processed image was then utilized to derive
the spectral signatures for statistical analysis.
Spectral data analysis
To determine the value of the red edge bands, supervised classification of both Sentinel-2 MSI and
Rapid Eye satellite image was performed using five different sets of spectral feature sets. Apart from
the standard Sentinel-2 MSI and Rapid Eye sensor spectral bands, the analysis also comprised the
usual vegetation index incorporation; the normalized difference vegetation index (NDVI) and the
normalized difference vegetation index – Red Edge (NDVI-RE).
For both Sentinel-2 MSI and Rapid Eye, only the reflective bands in the visible, NIR and SWIR
sections of the EM and red edge were subsequently used in the analysis. This is because as noted by
Seelig et al. (2008), Ceccato et al. (2002), Danson and Bowyer (2004), the reflectance in these bands
varies with different species of vegetation. We also extracted some vegetation indices (VIs) that are
normally relevant for vegetation studies; normalized difference vegetation index (NDVI), and red
edge-dependent NDVI (NDVI-RE) (see Huete et al. 2011; Förster et al. 2012; Schuster et al. 2012;
Kim and Yeom 2015). The analysis was executed in four stages (a), (b), (c) and (d) shown in Table 3.
Discriminant analysis (DA) and maximum likelihood classification (MLC)
Observed C3 (Festuca spp.) grass species distribution
We utilized Discriminant analysis (DA) algorithm to evaluate the capability of Rapid Eye and Sentinel-2
MSI satellite image in distinguishing the reflectance of C3 (Festuca spp.) from other coexisting C4
grass species. This algorithm is one of the data mining techniques that is used to discriminate a single
Table 3. Satellite image spectral bands and computed vegetation indices.
Applied variables
Spectral bands
chosen bands
Vegetation indices
Spectral bands and nDVIs
Variables utilized
rapid eye: Blue band 1 (440–510 nm), Green band 2 (520–590 nm), red band 3 (630–685 nm), red edge band 4 (690–730 nm) and nIr band 5
(760–850 nm)
Sentinel-2 MSI: Blue band 2 (458–523 nm); Green band 3 (543–578 nm); red band 4 (650–680 nm); red-edge band 5 (698–713 nm); red-edge
band 6 (733–793 nm); red-edge band 7 (733–793 nm); nIr band 8 (855–875 nm)
rapid eye: Blue band 1 (440–510 nm), red band 3 (630–685 nm), red edge band 4 (690–730 nm) and nIr band 5 (760–850 nm)Sentinel-2
MSI: Blue band 2 (458–523 nm); Green band 3 (543–578 nm); red band 4 (650–680 nm); red-edge band 5 (698–713 nm); red-edge band 6
(733–793 nm); red-edge band 7 (733–793 nm); nIr band 8 (855–875 nm)
normalized vegetation indices (nDVIs): nDVI; nDVI-re
rapid eye’s & Sentinel-2 MSI bands + nDVIs
Analysis process
(a)
(b)
(c)
(d)
GEOCARTO INTERNATIONAL
7
8
C. OTUNGA ET AL.
classification variable using multiple attributes (Fernandez 2002). It is a powerful tool for analysing
and describing group differences and for classifying cases into groups formed on the basis of their
similarities and differences on multiple variables. The analysis also assigns observations to one of
pre-defined groups based on the knowledge of a number of attributes (Brown and Wicker 2000).
The algorithm was chosen due to its ability to discriminate and project classification based on the
performance of continuous variables in separating specified categories in the classification process
(Sibanda et al. 2015; Matongera et al. 2017). A number of studies have recently successfully used the
DA in discriminating different land covers (Beck et al. 1997; Fraley and Raftery 2002; Ju et al. 2003;
Sibanda et al. 2015).
The analysis aggregates wavebands into components, also known as latent factors (Dong et al. 2013;
Sibanda et al. 2016). The impact of these latent factors is measured using Eigen vectors or variable
scores. Consequently, the most effective latent factors in discriminating grass reflectance in this study
were those that had the highest scores.
To determine the potential of integrating the red edge band of Sentinel-2 MSI and Rapid Eye satellite
sensors’ spectral configuration in mapping and discrimination of C3 Festuca and coexisting C4 grasses,
we utilized the Discriminant Analysis (DA). Before we applied the DA, the data was randomly split
into 70 and 30% for training/classification and accuracy assessment, respectively. As noted by Adelabu
et al. (2014), Adjorlolo and Mutanga (2013), Sibanda et al. (2015) among others, this partitioning is a
standard requirement for all machine learning algorithms. The training sample trains the DA algorithm
in mapping the grass species, whereas the test sample validates the performance of the model. In this
regard, the DA function was utilized to: (a) determine whether the integration of the red edge channel
of the two satellite sensors would increase the discrimination accuracy of C3 Festuca and coexisting
C4 grass species. (b) identify the best (most influential) spectral bands of the two sensors’ spectral
configuration that are useful for mapping and discriminating the two grass species and (c) identify
the most relevant variables among standard bands, vegetation indices and their combination thereof
in the mapping and discriminating C3 Festuca and other coexisting C4 grass species.
Next, we conducted a Wilks’s Lambda (Rao’s approximation) and chi-Square (chi-square asymptotic
approximation) tests using variable importance in the projections (VIPs), to examine the magnitude of
variation within the C3 (Festuca spp.) and coexisting C4 grass species class covariance matrices. The
measures evaluate the statistical significance of the discriminatory power of the discriminant function.
In this regard, the ratio of within-group variability to total variability on the discriminator variables,
is an inverse measure of the importance of the functions (Brown and Wicker 2000; Walde 2014).
Consequently, values close to 1 indicate that almost all of the variability in the discriminator variables
is due to within-group differences (differences between cases in each group); while values close to 0
indicate that almost all of the variability in the discriminator variables is due to group differences (Betz
1987; Brown et al. 1995). The DA was executed within XLSTAT for Microsoft Excel 2013 platform and
Statistical Product and Service Solutions – SPSS (PASW-predictive analytic software, Statistics 18 for
Windows) (XLSTAT 2013). Discriminant analysis thus aggregated the wavebands into components
or latent factors (Fernandez 2002). As noted by Sibanda et al. (2017, the influence of these underlying
factors is measured using Eigen vectors or variable scores which points at the discrimination power
of the function. Furthermore, the most effective latent factors in discriminating C3 (Festuca spp.) and
other coexisting C3/C4 grass species reflectance were taken to be those having the highest scores.
Overall, the DA results were presented as classification and cross-validated results. In this regard, the
variable screening ensures the identification of a few, most influential input variables that can further
enhance the model predictive accuracy. This process is critical as demonstrated in literature that the
relative individual variable influence is important because not all model input variables are equally
relevant in any modelling process (Ramoelo et al. 2015). The model thus computes the individual variable’s relative influence based on the contribution of each variable in reducing overall model deviance.
Maximum likelihood classification (MLC) was also utilized to further show C3 Festuca and other
coexisting C4 grass species spatial distribution given that the DA has no capability of map generation
hence was only utilized to compare improvement in classification accuracy. The ML algorithm is a
GEOCARTO INTERNATIONAL
9
robust and simple parametric approach in remote sensing image classification (Jensen 1996; Foody and
Dash 2007; Dixon et al. 2014). The method is based on the assumption that the data may be modelled
by a set of multivariate normal distributions. It therefore takes into account both the spectral variability
within and between classes (Fahsi et al. 2000; Liu et al. 2002; Weng 2002). According to DiPietro et
al. (2002 using training sets from the same image in MLC provides an excellent method for mapping
vegetation classes. In this study, we have not apply a post-classification processes like majority filtering
for the enhancement of our classification results as the study focused on the result of the individual
pixel classification to evaluate the Red Edge effect on classification accuracy. For the description and
illustration of results, accuracy assessment was performed for each of the classification algorithms
using confusion/error matrix.
Accuracy assessment
Classification accuracy for the C3 (Festuca spp.) and other coexisting C3/C4 grass species discrimination was assessed using the summary parameters as well as classification overall accuracies – quantity
disagreement and allocation disagreement (Pontius and Millones 2011). A confusion/error matrix
was used to compare the true class with the class assigned by the classifier, and to calculate the user’s
accuracy (UA), the producer’s accuracy (PA) and overall accuracy. According to Congalton (1991), PA
is calculated as the total number of correctly classified cases divided by the total number of cases in that
class, indicated by the estimation data. Furthermore, UA is calculated as the total number of correctly
classified cases of one category, divided by the total number of cases classified in that category (see
Story and Congalton 1986; Bobbe et al. 2001). In that regard, the PA gives a measure of commission
errors that correspond to those pixels belonging to the class of interest that the classifier has failed
to detect, while the UA gives a measure of commission errors that corresponds to those pixels from
other classes that the classifier has assigned to a class of interest (Bhaskaran et al. 2010). As noted by
Zhou et al. (2014), in the current study, the commission error stems from the incorrect inclusion of
the species in the other grass species class or category, whereas omission error occurs when a sample
of a particular grass species class is excluded from the class under consideration (Liu et al. 2007).
Consequently, to compare the accuracy change of inclusion and exclusion of red edge band and
also of Sentinel-2 and Rapid Eye sensor, a McNemar test was run in addition to the confusion/error
matrix. McNemar test is a nonparametric test used to compare classification abilities. It has demonstrated robustness as well as a higher precision in comparing accuracy assessments in classification
studies (Zhang and Xie 2012; Rodriguez-Galiano and Chica-Rivas 2014). The McNemar’s test is a better
statistic for comparing accuracies of classification methods than the kappa index and it is simple to
compute (Petropoulos et al. 2012; Adelabu et al. 2013). The kappa chi-squared requires that independent data are used to assess accuracies, but in this study the same points are used in all methods thus the
McNemar’s test was more appropriate as it is also more precise and sensitive (Manandhar et al. 2009).
The test is based on a chi-square (χ2) statistic, which is calculated from two error matrices stated as
(
) (
)2 (
)
𝜒 2 = f12 −f21 ∕ f12 + f21
where f12 is the number of classes that are incorrectly classified by the first algorithm and are appropriately classified by the second algorithm, whereas f21 is the number of classes that are properly classified
by the first algorithm and incorrectly classified by the second classifier algorithm (Kumar et al. 2002;
de Leeuw et al. 2006; Manandhar et al. 2009).
It is therefore a strong test that has been successfully utilized to compare classification accuracies
(Sibanda et al. 2016; Shoko and Mutanga 2017a). Hence, at a confidence of 95%, a McNemar test
result (indicated by z score) above 1.96 indicate that the classification accuracies derived from the
different sensors and associated variables are significantly different (Rodriguez-Galiano and ChicaRivas 2014; Sibanda et al. 2016). In the current study, a McNemar test was run to compare the change
10
C. OTUNGA ET AL.
Figure 2. overall classification accuracies achieved from the Mlc when: (a) rapid eye and (b) Sentinel-2 variables were used in the
analysis process.
in classification accuracy of inclusion and exclusion of red edge band and also when Sentinel-2 and
Rapid Eye were used in mapping and discrimination of C3 (Festuca spp.) and coexisting C4 grasses.
Results
The influence of sensors spectral bands on C3 Festuca and coexisting C4 grass species
mapping and discrimination
Using variables scores, derived from the DA model, Sentinel-2 MSI sensor provided more bands which
have great potential (indicated by the high variable scores) in mapping and discriminating C3 Festuca
and coexisting C4 grass species. The red edge bands (5, 6,7), notably centred at 705-, 740-, 783- and
865-nm, the VIS–NIR bands (2 and 8) centred at 490- and 842-nm spectral bands of the Sentinel 2
were the most influential in mapping C3 (Festuca costata) and coexisting C4 (Tristachya leucothrix)
grasses. Relatively, for the Rapid Eye sensor, the blue (between 450 and 510 nm) and red edge bands
between 705 and 745 nm were also the most influential bands.
The DA analysis thus brought forth the five most important spectral bands to be used in subsequent
analysis. We found Band2 and Band8 (in the VIS–NIR), Band5, band6 and Band7 (in the red-edge
channel) of Sentinel-2 MSI sensor, to have great potential (indicated by the high variable scores) in
mapping and discriminating C3 Festuca and coexisting C4 grass species. Consequently, compared to
the Rapid Eye sensor, a set of best (most influential) spectral bands was also identified. In particular,
the significant spectral channels in Rapid Eye image were identified as Band3, Band4 and Band5. Figure
2, thus, shows the influence of the spectral bands of the two sensors on discriminating between C3
Festuca and coexisting C4 grass species, using variables scores, derived from the DA model. We however did not utilize all bands of the Sentinel-2 sensor in the analysis. Similar to Féret et al. (2015) and
Immitzer et al. (2016), we considered bands 1 (coastal aerosol), 9 (water vapour) and 11 (SWIR cirrus)
for the Sentinel-2 to be irrelevant to vegetation analysis, hence not useful to the study (see Figure 3).
Maximum likelihood classification results are shown in Figure 4 below. C3 (Festuca spp.) and other
coexisting C4 grass species were distributed heterogeneously across the study site. Results show that the
integration of the red edge band improved the C3 (Festuca spp.) mapping and discrimination accuracy
from (72.5–77.5%) – representing 5.00% for Sentinel-2 MSI, (75.00–80.00%) – representing 5.00% for
Rapid Eye when the red edge bands were included and excluded in the analysis respectively. (Table
4). In this regard, the overall accuracy increased by 5% when the red edge band was incorporated in
the MLC analysis. Generally, the inclusion of the red edge band has led to a general improvement of
the classification accuracy results.
GEOCARTO INTERNATIONAL
Figure 3. Most influential variables (bands) in mapping and discrimination of c3 Festuca and coexisting c4 grass using both Sentinel-2 MSI and rapid eye image bands and Da.
11
12
C. OTUNGA ET AL.
Figure 4. Grass species distribution maps generated when: Standard bands of rapid eye image were included (a), (b) the red edge band
was excluded, (c) Standard bands of Sentinel-2 MSI image were included and (d) the red edge band was excluded in the Mlc analysis.
Outstanding accuracy increases in terms of producer’s and user’s accuracies are also observed for
C3 (Festuca spp.) and other coexisting C4 grass species, with respect to analysis a, b, c and d when
Sentinel-2 MSI, Rapid Eye red edge bands and NDVI-RE are included and excluded, respectively
(Figures 4, 5 and Tables 4–6.).
Mapping and discrimination of C3 Festuca and other coexisting C4 grass species using
Sentinel-2 MSI and the Rapid Eye optimal spectral bands versus analysis excluding and
excluding red-edge
Results from our analysis demonstrate that the integration of the red edge band improved the C3
(Festuca spp.) mapping and discrimination accuracy from (77.38–83.33%) – representing 5.95% for
Sentinel-2 MSI, (79.76–84.52%) – representing 4.76 for Rapid Eye when the red edge bands were
included and excluded in the analysis, respectively. Consequently, mapping and discrimination
accuracy improvement was realized from (72.62–75.00%) – representing 2.38% for Sentinel-2 MSI,
(70.24–76.19%) – representing 5.95% for Rapid Eye, when the red edge indices were incorporated and
excluded in the analysis, respectively (see Tables 6, 7(a)–(d) and Figure 5 below).
Furthermore, McNemar’s test results from our analysis have also shown that there were statistically significant differences (z > 1.96) in classification accuracy changes when the red edge band of
Sentinel-2 and Rapid Eye sensors were included and excluded in the analyses. For example, the z-scores
were 2.78 and 2.32 when the red edge band of Sentinel-2 were included and excluded, respectively,
while the z-scores were 2.56 and 2.24 when the red edge band of Rapid Eye sensors were included
and excluded, respectively.
GEOCARTO INTERNATIONAL
13
Table 4. confusion error matrix derived from the Mlc results when: (a) the rapid eye’s standard optimal bands were included and
(b) the red edge channel band was excluded in the mapping and discrimination process.
Observed
(a)
classified
(b)
classified
Shrubs/
C3
Thickets
Festuca C3/C4
Other Coexisting
Shrubs/thickets
2
1
1
c3 Festuca
1
17
2
other coexisting
1
2
13
c3/c4
total
4
20
16
omission error
50
15
18.75
Producer’s accuracy 50
85
81.25
the overall accuracy is (2 + 17 + 13)/40 = 32/40 = 80.00%
Shrubs/thickets
3
1
0
c3 Festuca
0
14
2
other coexisting
1
6
13
c3/c4
total
4
21
15
omission error
25
33.33
13.33
Producer’s accuracy 75.00
66.67
86.67
the overall accuracy is (3 + 14 + 13)/40 = 30/40 = 75.00%
Commission
User’s
Overall
error
accuracy
accuracy
50
15
18.75
50
85
81.25
80.00
25
12.5
35
75.00
87.5
65
75.00
Total
4
20
16
40
4
16
20
40
note: the bold values represents the percentage accuracies.
Figure 5. overall classification accuracies achieved from the Da when: (a) the standard optimal bands of rapid eye and Sentinel-2
were included (b) the red edge channel band was excluded (c) the nDVI-re was included and (d) when the nDVI-re was excluded
in the analysis process.
Discussion
User’s accuracy forms a guide to the reliability of the map, and as a predictive device, it tells the user
of the classified map generated that in our assessment, of the area labelled C3 Festuca on the map
for example, in Table 6(a) 76.92% actually corresponds to C3 Festuca on the ground. Consequently,
Producer’s accuracy in our case informs us that, of the actual coexisting C3/C4 grass species, for
example, 85.71% was correctly classified. Commission error (%) = 1 – User’s accuracy, while Omission
error (%) = 1 – Producer’s accuracy (%) (Congalton and Green 2008). Comparing the results of our
analysis; a, b, c and d have revealed that the additional red edge channel has influenced the overall
classification accuracy. This is supported by the results from the analyses based on inclusion as well as
exclusion, of the red edge band in both DA and MLC algorithms we utilized. The overall classification
14
C. OTUNGA ET AL.
Table 5. confusion error matrix derived from the Mlc results when: (a) the Sentinel-2 MSI’s standard bands were included and (b)
the red edge channel bands were excluded in the mapping and discrimination process.
Shrubs/
Observed
(a)
classified
(b)
classified
Thickets
C3
Other
Festuca
Coexisting C4
Commission
User’s
Overall
Error
Accuracy
Accuracy
Total
Shrubs/thickets
2
1
1
c3 Festuca
0
16
4
coexisting c4
0
3
13
total
2
20
18
omission error
0
20
27.78
Producer’s accuracy 100
80
72.22
overall accuracy is (2 + 16 + 13)/40 = 31/40 = 77.50
4
20
16
40
27.78
50
20
18.75
50
80
81.25
77.5
Shrubs/thickets
3
1
1
c3 Festuca
0
12
3
coexisting c4
0
6
14
total
3
19
18
omission error
43
36
37
Producer’s accuracy 66.7
64
63
overall accuracy is (3 + 12 + 14)/40 = 29/40 = 72.50
5
15
20
40
43
40
33
57
60
67
72.5
note: the bold values represents the percentage accuracies.
Table 6. confusion error matrix derived from the Da when: (a) the standard most influential bands of Sentinel-2 MSI were included
(b) the red edge channel bands were excluded (c) the VIs (nDVI, nDVI-re) were included and (d) when the VIs (nDVI, nDVI-re) were
excluded in the analysis process.
C3
(a)
classified
(b)
classified
(c)
classified
(d)
classified
Other
Commission
User’s
Overall
Error
Accuracy
Accuracy
Observed
Festuca
Coexisting C4
Total
c3 Festuca
coexisting c4
total
omission error
Producer’s accuracy
30
5
35
14.29
85.71
9
40
49
18.37
81.63
39
45
84
23.08
11.11
76.92
88.89
83.33
c3 Festuca
coexisting c4
total
omission error
Producer’s accuracy
29
9
38
23.68
76.32
10
36
46
21.74
78.26
39
45
84
25.64
20
74.36
80
77.38
c3 Festuca
coexisting c4
total
omission error
Producer’s accuracy
32
14
46
30.43
69.57
7
31
38
18.42
81.58
39
45
84
17.95
31.11
82.05
68.89
75.00
c3 Festuca
coexisting c4
total
omission error
Producer’s accuracy
24
8
32
25.00
75.00
15
37
52
28.85
71.15
39
45
84
38.46
17.78
61.54
82.22
72.62
note: the bold values represents the percentage accuracies.
accuracy improvement can be linked to the additional information provided when the red edge band
was included in the analysis. This represented an overall mapping accuracy increase of 5.95 and 4.76%
for Sentinel-2 MSI and Rapid Eye when DA algorithm was utilized, respectively. Improvement was also
realized when the vegetation indices were incorporated in the analysis, i.e. the accuracy increased by
GEOCARTO INTERNATIONAL
15
Table 7. confusion error matrix derived from the Da when: (a) the standard influential bands of rapid eye image were included
(b) the red edge channel band was excluded (c) the VIs (nDVI, nDVI-re) were included and (d) when the VIs (nDVI, nDVI-re) were
excluded in the analysis process.
(a)
classified
(b)
classified
(c)
classified
(d)
classified
C3
Other
Observed
Festuca
Coexisting C4
Total
Commission
User’s
Overall
Error
Accuracy
Accuracy
c3 Festuca
coexisting c4
total
omission error
Producer’s accuracy
32
6
38
15.79
84.21
7
39
46
15.22
84.78
39
45
84
17.95
13.33
82.05
86.67
84.52
c3 Festuca
coexisting c4
total
omission error
Producer’s accuracy
29
7
36
19.44
80.56
10
38
48
20.83
79.17
39
45
84
25.64
15.56
74.36
84.44
79.76
c3 Festuca
coexisting c4
total
27
8
35
12
37
49
39
45
84
30.77
35.56
69.23
64.44
76.19
omission error
Producer’s accuracy
22.86
77.14
24.49
75.51
c3 Festuca
coexisting c4
total
omission error
Producer’s accuracy
25
11
36
30.56
69.44
14
34
48
29.17
70.83
39
45
84
35.90
24.44
64.10
75.56
70.24
note: the bold values represents the percentage accuracies.
2.38% for Sentinel-2 MSI (72.62–75.00%) and 5.95% for Rapid Eye (70.24–76.19%), when the vegetation indices (NDVI-RE) was incorporated and excluded in the analysis, respectively. The classification
accuracy produced when using the Sentinel 2 Red Edge bands did not differ significantly (z < 1.96)
from those produced when Rapid Eye Red Edge bands were included and excluded in the analyses. Our
results hence confirm the earlier reported potential of the strategically positioned and/or additional
bands of different sensors in enhancing their ability in discriminating vegetation species (Immitzer
et al. 2016; Laurin et al. 2016; Shoko and Mutanga 2017b). Both the C3 Festuca and coexisting C4
grass species seemed to be affected by an accuracy improvement resulting from the incorporation of
the red edge band in the analysis.
The influence of the red edge band in discriminating different C3/C4 grasses could be attributed to
the fact that, it is insensitive to the soil background effect (Vane et al. 1993; Datt and Paterson 2000;
Clevers et al. 2001; Schumacher et al. 2016). Furthermore, literature shows that the Red Edge region
of the electromagnetic spectrum is highly associated with vegetation characteristics, such as chlorophyll and LAI, which directly affect the reflectance of vegetation (Curran et al. 1990, 1995; Clevers
and Gitelson 2013; Delegido et al. 2013). In this study, C3 (Festuca costata) and other coexisting C4
(Tristachya leucothrix) grass species were satisfactorily discriminated at optimal overall accuracies
based on the Red Edge waveband.
The improvement of classification accuracy obtained when Vis (NDVI-RE, REP and REIP) were
used could be explained by the fact that the wavebands are more sensitive to vegetation traits when
compared to individual bands only. Furthermore, VIs are highly sensitive to vegetation spectral and
temporal characteristics, while reducing the sensitivity from atmospheric noise, view/Sun angle, soil
background and topographic effects when compared to individual bands (Thenkabail et al. 2011). In
16
C. OTUNGA ET AL.
that regard, the satisfactory performance of VIs in discriminating C3 Festuca grasses could be attributed to their ability to reduce noise while being sensitive to the different C3/C4 grass species spectral
signatures. Furthermore, the plausible performance of red edge-associated VIs could be attributed to
the fact that these indices are derived from a region in the electromagnetic spectrum that is highly
sensitive to variations in grass vegetation chlorophyll.
Conclusion
The purpose of this study was to evaluate the value of the Rapid Eye’s and Sentinel-2 MSI’s Red Edge
channel for mapping and discriminating C3 Festuca grass species.
Based on these findings, we conclude that:
• The study confirmed the high value of the red-edge and infrared bands from new relatively
affordable or freely available multispectral sensors; Rapid Eye and Sentinel-2 MSI for grass vegetation mapping.
• The study has demonstrated the potential of Sentinel-2 MSI’s and Rapid Eye’s red edge bands in
improving mapping and discrimination accuracy of C3 Festuca grass species
• The use of new generation sensors (Rapid Eye and Sentinel-2 MSI) with additional channels
containing the red edge band is promising for C3 Festuca grass species mapping.
• Sentinel-2 MSI provides a great opportunity for global vegetation monitoring due to its enhanced
spatial, spectral and temporal characteristics compared with the commercial Rapid Eye image
data.
• To ensure that there is a wider applicability of the procedures, the methods we have presented in
this study need to be tested on other C3/C4 grass species spatial distribution in different locations.
• Studies should now move towards embracing the readily and freely available new generation
multispectral sensors with strategically positioned bands more so in sub-Saharan Africa where
data availability still remains a challenge
Overall, results have indicated the potential of the new generation sensors, with strategically positioned spectral bands such as the red edge in mapping C3 Festuca grass species. The current study
results open up prospects for discriminating C3 Festuca grass using multispectral sensors having
strategically positioned bands. However, we do recommend that more studies need to be carried out
to investigate the feasibility of using other freely and readily available multispectral sensors; Landsat
8 OLI to detect, map and predict the spatial distribution of C3 Festuca costata in other rangelands.
Disclosure statement
No potential conflict of interest was reported by the authors.
ORCID
Charles Otunga
http://orcid.org/0000-0002-0153-6275
References
Adam E, Mutanga O, Rugege D. 2010. Multispectral and hyperspectral remote sensing for identification and mapping
of wetland vegetation: a review. Wetlands Ecol Manage. 18(3):281–296.
Adelabu S, Mutanga O, Adam E, Cho MA. 2013. Exploiting machine learning algorithms for tree species classification
in a semiarid woodland using RapidEye image. J Appl Remote Sens. 7(1):073480–073480.
Adelabu S, Mutanga O, Adam E. 2014. Evaluating the impact of red-edge band from Rapideye image for classifying
insect defoliation levels. ISPRS J Photogrammet Remote Sens. 95:34–41.
Adelabu S, Mutanga O, Adam E, Sebego R. 2014. Spectral discrimination of insect defoliation levels in mopane woodland
using hyperspectral data. IEEE J Selected Topics Appl. 7(1):177–186.
GEOCARTO INTERNATIONAL
17
Adjorlolo C. 2013. Remote sensing of the distribution and quality of subtropical C3 and C4 grasses. Pietermaritzburg:
University of KwaZulu-Natal.
Adjorlolo C, Mutanga O. 2013. Integrating remote sensing and geostatistics to estimate woody vegetation in an African
savanna. J Spatial Sci. 58(2):305–322.
Adjorlolo C, Mashiyi N, Mangara P, Mutanga, O. 2016. Trends in seasonal composition of plant functional types:
retrieving fraction of winter-green grass cover in the Drakensberg Mountains using long-term satellite observations
(Platform). 51st Annual Congress of the Grassland Society of Southern Africa; George, South Africa
Barsi JA, Barker JL, Schott JR. 2003. An atmospheric correction parameter calculator for a single thermal band earthsensing instrument. In: Geoscience and Remote Sensing Symposium, 2003 IGARSS’03 Proceedings 2003 IEEE
International. IEEE. p. 3014–3016.
Beck LR, Rodriguez MH, Dister SW, Rodriguez AD, Washino RK, Roberts DR, Spanner MA. 1997. Assessment of a
remote sensing-based model for predicting malaria transmission risk in villages of Chiapas, Mexico. Am J Trop
Med Hyg. 56(1):99–106.
Betz NE. 1987. Use of discriminant analysis in counseling psychology research. J Counsel Psychol. 34(4):393.
Bhaskaran S, Paramananda S, Ramnarayan M. 2010. Per-pixel and object-oriented classification methods for mapping
urban features using Ikonos satellite data. Appl Geogr. 30(4):650–665.
Black SC. 2006. Estimation of grass photosynthesis rates in mixed-grass prairie using field and remote sensing approaches.
Saskatoon: University of Saskatchewan.
Bobbe T, Lachowski H, Maus P, Greer J, Dull C. 2001. A primer on mapping vegetation using remote sensing. Int J
Wildland Fire. 10(4):277–287.
Bond WJ, Parr CL. 2010. Beyond the forest edge: ecology, diversity and conservation of the grassy biomes. Biol Cons.
143(10):2395–2404.
Boothway G, Martindale G, Gillings K. 2012. Fort Nottingham nature reserve: Combined management plan. Version
10. Pietermaritzberg: Ezemvelo KwaZulu-Natal Wildlife.
Botha JO, Boothway G, Keith M. 2012. An assessment of the rangelands of the Fort Nottingham townlands.
Pietermaritzburg: Cedara, Department of Agriculture and Environmental Affairs Natural Resources Macro Planning
Section
Brown MT, Wicker LR. 2000. Discriminant analysis. In: Tinsley EA, Brown SD, editors. Handbook of applied multivariate
statistics and mathematical modeling. San Diego, CA: Academic Press; p. 209–235.
Brown JH, Mehlman DW, Stevens GC. 1995. Spatial variation in abundance. Ecology. 76(7):2028–2043.
Ceccato P, Gobron N, Flasse S, Pinty B, Tarantola S. 2002. Designing a spectral index to estimate vegetation water content
from remote sensing data: Part 1. Remote Sens Environ. 82(2–3):188–197.
Chander G, Markham BL, Helder DL. 2009. Summary of current radiometric calibration coefficients for Landsat MSS,
TM, ETM+, and EO-1 ALI sensors. Remote Sens Environ. 113(5):893–903.
Chavez PS. 1988. An improved dark-object subtraction technique for atmospheric scattering correction of multispectral
data. Remote Sens Environ. 24(3):459–479.
Chavez PS. 1996. Image-based atmospheric corrections-revisited and improved. Photogram Eng Remote Sens.
62(9):1025–1035.
Cheng P, Sustera J. 2009. Using RapidEye data without Ground Control: automated high-speed high-accuracy.
GEOinformatics. 12:36–40.
Clevers JG, Gitelson AA. 2013. Remote estimation of crop and grass chlorophyll and nitrogen content using red-edge
bands on Sentinel-2 and-3. Int J Appl Earth Obs Geoinf. 23:344–351.
Clevers JG, de Jong SM, Epema GF, van der Meer F, Bakker WH, Skidmore AK, Addink EA. 2001. MERIS and the rededge position. Int J Appl Earth Obs Geoinf. 3(4):313–320.
Clevers JG, Kooistra L, van den Brande MM. 2017. Using Sentinel-2 data for retrieving LAI and leaf and canopy
chlorophyll content of a potato crop. Remote Sens. 9(5):405.
Cole B, Balzter H, Smith G, Morton D, King S. 2014. Delivering the Copernicus land monitoring service, production
of the CORINE Land Cover Map in the UK. A forward looking perspective to the Sentinel-2 mission. EGU General
Assembly Conference Abstracts.
Conant RT, Paustian K, Elliott ET. 2001. Grassland management and conversion into grassland: effects on soil carbon.
Ecol Appl. 11(2):343–355.
Congalton RG. 1991. A review of assessing the accuracy of classifications of remotely sensed data. Remote Sens Environ.
37(1):35–46.
Congalton RG, Green K. 2008. Assessing the accuracy of remotely sensed data: principles and practices. New York:
CRC Press.
Corcoran E. 2010. Dead planet, living planet: Biodiversity and ecosystem restoration for sustainable development: a
rapid response assessment. Birkeland, Norway: UNEP/Earthprint.
Cousins SA, Eriksson O. 2002. The influence of management history and habitat on plant species richness in a rural
hemiboreal landscape. Landscape Ecol. 17(6):517–529.
Critchley C, Burke M, Stevens D. 2004. Conservation of lowland semi-natural grasslands in the UK: a review of botanical
monitoring results from agri-environment schemes. Biol Conserv. 115(2):263–278.
18
C. OTUNGA ET AL.
Curran PJ, Dungan J, Gholz H. 1990. Exploring the relationship between reflectance red edge and chlorophyll content
in slash pine. Tree Physiol. 7(1–4):33–48.
Curran PJ, Windham WR, Gholz HL. 1995. Exploring the relationship between reflectance red edge and chlorophyll
concentration in slash pine leaves. Tree Physiol. 15(3):203–206.
D’Antonio CM, Vitousek PM. 1992. Biological invasions by exotic grasses, the grass/fire cycle, and global change. Annu
Rev Ecol Syst. 23:63–87.
Daily GC, Alexander S, Ehrlich PR, Goulder L, Lubchenco J, Matson PA, Mooney HA, Postel S, Schneider SH, Tilman D.
1997. Ecosystem services: benefits supplied to human societies by natural ecosystems. Washington (DC): Ecological
Society of America.
Danson F, Bowyer P. 2004. Estimating live fuel moisture content from remotely sensed reflectance. Remote Sens Environ.
92(3):309–321.
Darvishzadeh RR. 2015. Multivariate statistical analysis for estimating grassland leaf area index and chlorophyll content
using hyperspectral data. Msc., University of Twente.
Datt B, Paterson M. 2000. Vegetation-soil spectral mixture analysis. In: Geoscience and Remote Sensing Symposium,
2000 Proceedings IGARSS 2000. IEEE 2000 International. IEEE. p. 1936–1938.
Delegido J, Verrelst J, Meza C, Rivera J, Alonso L, Moreno J. 2013. A red-edge spectral index for remote sensing estimation
of green LAI over agroecosystems. Eur J Agron. 46:42–52.
Dhau I, Adam E, Mutanga O, Ayisi K, Abdel-Rahman EM, Odindi J, Masocha M. 2017. Testing the capability of spectral
resolution of the new multispectral sensors on detecting the severity of grey leaf spot disease in maize crop. Geocarto
Int. (just-accepted), 1–14.
Diaz S, Cabido M. 1997. Plant functional types and ecosystem function in relation to global change. J Veget Sci. 8:463–474.
DiPietro D, Ustin SL, Underwood E. 2002. Mapping the invasive plant Arundo donax and associated riparian vegetation
using AVIRIS. Proceedings 11th Airborne visible/infrared image spectrometer (AVIRIS) Workshop: Jet Propulsion
Laboratory, Pasadena, CA CD-ROM.
Dixon A, Faber-Langendoen D, Josse C, Morrison J, Loucks C. 2014. Distribution mapping of world grassland types.
J Biogeograp. 41(11):2003–2019.
Dong W, Ni Y, Kokot S. 2013. A near-infrared reflectance spectroscopy method for direct analysis of several chemical
components and properties of fruit, for example, Chinese Hawthorn. J Agric Food Chem. 61(3):540–546.
Drusch M, Del Bello U, Carlier S, Colin O, Fernandez V, Gascon F, Hoersch B, Isola C, Laberinti P, Martimort P. 2012.
Sentinel-2: ESA’s optical high-resolution mission for GMES operational services. Remote Sens Environ. 120:25–36.
Dube T, Mutanga O, Sibanda M, Bangamwabo V, Shoko C. 2017. Testing the detection and discrimination potential of
the new Landsat 8 satellite data on the challenging water hyacinth ( Eichhornia crassipes ) in freshwater ecosystems.
Appl Geogr. 84:11–22.
Edwards D. 1983. A broad-scale vegetation classification of vegetation for practical purposes. Bothalia. 14(3&4):705–712.
Eitel J, Long D, Gessler P, Smith A. 2007. Using in‐situ measurements to evaluate the new RapidEye ™ satellite series for
prediction of wheat nitrogen status. Int J Remote Sens. 28(18):4183–4190.
Ellis RP, Vogel J, Fuls A. 1980. Photosynthetic pathways and the geographical distribution of grasses in South West
Africa/Namibia. South Afr J Sci. 76(7):307–314.
Eriksson Å, Eriksson O, Berglund H. 1995. Species abundance patterns of plants in Swedish semi-natural pastures.
Ecography. 18(3):310–317.
Everson CS, Everson T. 2016. The long-term effects of fire regime on primary production of montane grasslands in
South Africa. Afr J Range Forage Sci. 33(1):33–41.
Fahsi A, Tsegaye T, Tadesse W, Coleman T. 2000. Incorporation of digital elevation models with Landsat-TM data to
improve land cover classification accuracy. Forest Ecol Managem. 128(1):57–64.
Féret J-B, Corbane C, Alleaume S. 2015. Detecting the phenology and discriminating mediterranean natural habitats
with multispectral sensors – an analysis based on multiseasonal field spectra. IEEE J Sel opics Appl Earth Obs Remote
Sens. 8(5):2294–2305.
Fernandez GC. 2002. Discriminant analysis, a powerful classification technique in data mining. In: Proceedings of the
SAS Users International Conference. p. 247–256.
Flanagan LB, Wever LA, Carlson PJ. 2002. Seasonal and interannual variation in carbon dioxide exchange and carbon
balance in a northern temperate grassland. Glob Change Biol. 8(7):599–615.
Foody GM, Dash J. 2007. Discriminating and mapping the C3 and C4 composition of grasslands in the northern Great
Plains, USA. Ecol Inf. 2(2):89–93.
Forkuor G, Dimobe K, Serme I, Tondoh JE. 2017. Landsat-8 vs. Sentinel-2: examining the added value of Sentinel-2’s
red-edge bands to land-use and land-cover mapping in Burkina Faso. In: GIScience Remote Sens. p. 1–24.
Förster M, Schmidt T, Schuster C, Kleinschmit B. 2012. Multi-temporal detection of grassland vegetation with RapidEye
imagery and a spectral-temporal library. In: Geoscience and Remote Sensing Symposium (IGARSS), 2012 IEEE
International. IEEE. p. 4930–4933.
Fraley C, Raftery AE. 2002. Model-based clustering, discriminant analysis, and density estimation. J Am Stat Assoc.
97(458):611–631.
GEOCARTO INTERNATIONAL
19
Franke J, Keuck V, Siegert F. 2012. Assessment of grassland use intensity by remote sensing to support conservation
schemes. J Nature Conserv. 20(3):125–134.
Gardi C, Tomaselli M, Parisi V, Petraglia A, Santini C. 2002. Soil quality indicators and biodiversity in northern Italian
permanent grasslands. Eur J Soil Biol. 38(1):103–110.
Hadjimitsis D, Clayton C, Retalis A. 2003. Darkest pixel atmospheric correction algorithm: a revised procedure for
environmental applications of satellite remotely sensed imagery. In: Proceedings 10th international symposium on
remote sensing. p. 12.
Hattersley P. 1983. The distribution of C3 and C4 grasses in Australia in relation to climate. Oecologia. 57(1–2):113–128.
He Y. 2008. Modeling grassland productivity through remote sensing products. Doctor of Philosophy, University of
Saskatchewan.
Hirzel A, Guisan A. 2002. Which is the optimal sampling strategy for habitat suitability modelling. Ecological Modelling.
157(2):331–341.
Huete A, Didan K, van Leeuwen W, Miura T, Glenn E. 2011. MODIS vegetation indices. Land Remote Sensing and
Global Environmental Change. 7(26):579–602.
Immitzer M, Vuolo F, Atzberger C. 2016. First experience with Sentinel-2 data for crop and tree species classifications
in Central Europe. Remote Sens. 8(3):166.
Jensen JR. 1996. Introductory digital image processing: a remote sensing perspective, Columbus, GA: Prentice-Hall Inc.
Ju J, Kolaczyk ED, Gopal S. 2003. Gaussian mixture discriminant analysis and sub-pixel land cover characterization in
remote sensing. Remote Sens Environ. 84(4):550–560.
Killick DJB. 1990. A field guide to the flora of the Natal Drakensberg. Johannesburg: Jonathan Ball and Ad. Donker
200p.-illus., col. illus., map. ISBN: 947464352.
Kim H-O, Yeom J-M. 2015. Sensitivity of vegetation indices to spatial degradation of RapidEye imagery for paddy rice
detection: a case study of South Korea. GIScience Remote Sens. 52(1):1–17.
King G. 2002. An Explanation of the 1: 500 000 General Hydrogeological Map: Durban 2928. Durban: Directorate
Geohydrology, Department of Water Affairs and Forestry.
Knox NM, Skidmore AK, Prins HH, Asner GP, van der Werff H, de Boer WF, van der Waal C, de Knegt HJ, Kohi EM,
Slotow R. 2011. Dry season mapping of savanna forage quality, using the hyperspectral Carnegie Airborne observatory
sensor. Remote Sens Environ. 115(6):1478–1488.
Kruger AC, Shongwe S. 2004. Temperature trends in South Africa: 1960–2003. Int J Climatol. 24(15):1929–1945.
Kumar L, Schmidt K, Dury S, Skidmore A. 2002. Imaging spectrometry and vegetation science. Imaging Spectrometry.
4(1):111–155.
Langley JA, Megonigal JP. 2010. Ecosystem response to elevated CO2 levels limited by nitrogen-induced plant species
shift. Nature. 466(7302):96–99.
Laurin GV, Puletti N, Hawthorne W, Liesenberg V, Corona P, Papale D, Chen Q, Valentini R. 2016. Discrimination of
tropical forest types, dominant species, and mapping of functional guilds by hyperspectral and simulated multispectral
Sentinel-2 data. Remote Sensing of Environment. 176:163–176.
Leadley PW, Niklaus PA, Stocker R, Körner C. 1999. A field study of the effects of elevated CO2 on plant biomass and
community structure in a calcareous grassland. Oecologia. 118(1):39–49.
de Leeuw J, Jia H, Yang L, Liu X, Schmidt K, Skidmore A. 2006. Comparing accuracy assessments to infer superiority
of image classification methods. Int J Remote Sens. 27(1):223–232.
Li T, Grant RF, Flanagan LB. 2004. Climate impact on net ecosystem productivity of a semi-arid natural grassland:
modeling and measurement. Agric Forest Meteorol. 126(1):99–116.
Liu L, Cheng Z. 2011. Mapping C3 and C4 plant functional types using separated solar-induced chlorophyll fluorescence
from hyperspectral data. Int J Remote Sens. 32(24):9171–9183.
Liu X-H, Skidmore A, Van Oosten H. 2002. Integration of classification methods for improvement of land-cover map
accuracy. ISPRS J Photogramm Remote Sens. 56(4):257–268.
Liu C, Frazier P, Kumar L. 2007. Comparative assessment of the measures of thematic classification accuracy. Remote
Sens Environ. 107(4):606–616.
Louis J, Debaecker V, Pflug B, Main-Knorn M, Bieniarz J, Mueller-Wilm U, Cadau E, Gascon F. 2016. Sentinel-2 Sen2Cor:
L2A Processor for Users Proceedings Living Planet Symposium 2016. Spacebooks Online. p. 1–8.
Low AB, Rebelo AG. 1998. Vegetation of South Africa, Lesotho and Swaziland: a companion to the vegetation map of
South Africa, Lesotho and Swaziland.
Manandhar R, Odeh IO, Ancev T. 2009. Improving the accuracy of land use and land cover classification of landsat data
using post-classification enhancement. Remote Sens. 1(3):330–344.
t Mannetje L. 2006. Grasslands of the world. Grass Forage Sci. 61(2):200–201.
Mansour K, Mutanga O, Everson T. 2012. Remote sensing based indicators of vegetation species for assessing rangeland
degradation: Opportunities and challenges. Afr J Agric Res. 7(22):3261–3270.
Marshall V. 2013. Mapping, modelling and remote sensing buffel grass (Cenchrus ciliaris) infestations in arid Australia.
Doctor of Philosophy, University of Adelaide.
Matongera TN, Mutanga O, Dube T, Sibanda M. 2017. Detection and mapping the spatial distribution of bracken fern
weeds using the Landsat 8 OLI new generation sensor. Int J Appl Earth Obs Geoinf. 57:93–103.
20
C. OTUNGA ET AL.
Matsika R. 2008. Land-cover change: threats to the grassland biome of South Africa. MSc., University of Witwatersrand.
Meadows M, Hoffman M. 2002. The nature, extent and causes of land degradation in South Africa: legacy of the past,
lessons for the future? Area. 34(4):428–437.
Mucina L, Rutherford MC. 2006. The vegetation of South Africa, Lesotho and Swaziland. South African National
Biodiversity Institute.
Mutanga O, Rugege D. 2006. Integrating remote sensing and spatial statistics to model herbaceous biomass distribution
in a tropical savanna. Int J Remote Sens. 27(16):3499–3514.
Nestola E, Calfapietra C, Emmerton CA, Wong C, Thayer DR, Gamon JA. 2016. Monitoring grassland seasonal carbon
dynamics, by integrating MODIS NDVI, proximal optical sampling, and eddy covariance measurements. Remote
Sens. 8(3):260.
Niu S, Liu W, Wan S. 2008. Different growth responses of C3 and C4 grasses to seasonal water and nitrogen regimes
and competition in a pot experiment. J Exp Botany. 59(6):1431–1439.
Norby RJ, Luo Y. 2004. Evaluating ecosystem responses to rising atmospheric CO2 and global warming in a multi-factor
world. New Phytol. 162(2):281–293.
O’Connor T, Kuyler P. 2009. Impact of land use on the biodiversity integrity of the moist sub-biome of the grassland
biome, South Africa. J Environ Manage. 90(1):384–395.
O’Mara FP. 2012. The role of grasslands in food security and climate change. Ann Botany. 209:1–8.
Odindi J, Mutanga O, Rouget M, Hlanguza N. 2016. Mapping alien and indigenous vegetation in the KwaZulu-Natal
Sandstone Sourveld using remotely sensed data. Bothalia 46(2):9.
Parton W, Coughenour M, Scurlock J, Ojima D, Gilmanov T, Scholes R, Schimel D, Kirchner T, Menaut J, Seastedt T.
1996. Global grassland ecosystem modelling: development and test of ecosystem models for grassland systems. In:
Scope-Scientific Committee on Problems of the Environment International Council of Scientific Unions. Nairobi.
Pau S, Still CJ. 2014. Phenology and Productivity of C3 and C4 Grasslands in Hawaii. PLoS ONE. 9(10):e107396.
Pearcy R, Ehleringer J. 1984. Comparative ecophysiology of C3 and C4 plants. Plant Cell Environ. 7(1):1–13.
Pearcy RW, Tumosa N, Williams K. 1981. Relationships between growth, photosynthesis and competitive interactions
for a C3 and C4 plant. Oecologia. 48(3):371–376.
Petropoulos GP, Kalaitzidis C, Vadrevu KP. 2012. Support vector machines and object-based classification for obtaining
land-use/cover cartography from Hyperion hyperspectral imagery. Comput Geosci. 41:99–107.
Pontius RG Jr, Millones M. 2011. Death to Kappa: birth of quantity disagreement and allocation disagreement for
accuracy assessment. Int J Remote Sens. 32(15):4407–4429.
Ramoelo A, Skidmore AK, Cho MA, Schlerf M, Mathieu R, Heitkönig IM. 2012. Regional estimation of savanna grass
nitrogen using the red-edge band of the spaceborne RapidEye sensor. Int J Appl Earth Obs Geoinf. 19:151–162.
Ramoelo A, Cho M, Mathieu R, Skidmore AK. 2015. Potential of Sentinel-2 spectral configuration to assess rangeland
quality. J Appl Remote Sens. 9(1):094096–094096.
Ramoelo A, Cho MA, Mathieu R, Madonsela S, van de Kerchove R, Kaszta Z, Wolff E. 2015. Monitoring grass nutrients
and biomass as indicators of rangeland quality and quantity using random forest modelling and WorldView-2 data.
Int J Appl Earth Obs Geoinf. 43:43–54.
RapidEye A. 2016. Satellite imagery product specifications. Satellite imagery product specifications. Version.
Richter R, Schläpfer D. 2002. Geo-atmospheric processing of airborne imaging spectrometry data. Part 2: Atmospheric/
topographic correction. Int J Remote Sens. 23(13):2631–2649.
Rodriguez-Galiano VF, Chica-Rivas M. 2014. Evaluation of different machine learning methods for land cover mapping
of a Mediterranean area using multi-seasonal Landsat images and Digital Terrain Models. Int J Digital Earth. 7(6):492–
509.
Rossiter NA, Setterfield SA, Douglas MM, Hutley LB. 2003. Testing the grass-fire cycle: alien grass invasion in the
tropical savannas of northern Australia. Diversity and distributions. 9(3):169–176.
Schumacher P, Mislimshoeva B, Brenning A, Zandler H, Brandt M, Samimi C, Koellner T. 2016. Do red edge and
texture attributes from high-resolution satellite data improve wood volume estimation in a semi-arid mountainous
region? Remote Sens. 8(7):540.
Schuster C, Förster M, Kleinschmit B. 2012. Testing the red edge channel for improving land-use classifications based
on high-resolution multi-spectral satellite data. Int J Remote Sens. 33(17):5583–5599.
Seelig H-D, Hoehn A, Stodieck L, Klaus D, Adams W, Emery W. 2008. Relations of remote sensing leaf water indices to
leaf water thickness in cowpea, bean, and sugarbeet plants. Remote Sens Environ. 112(2):445–455.
Shaw CS, Escott B. 2011. KwaZulu–natal vegetation type description document for vegetation map 2011.
kznveg05v2_1_11_wll. shp. 2(1).
Shoko C, Mutanga O, Dube T. 2016. Progress in the remote sensing of C3 and C4 grass species aboveground biomass
over time and space. ISPRS J Photogram Remote Sens. 120:13–24.
Shoko C, Mutanga O. 2017a. Examining the strength of the newly-launched Sentinel 2 MSI sensor in detecting and
discriminating subtle differences between C3 and C4 grass species. ISPRS Journal of Photogrammetry and Remote
Sensing. 129:32–40.
GEOCARTO INTERNATIONAL
21
Shoko C, Mutanga O. 2017b. Seasonal discrimination of C3 and C4 grasses functional types: An evaluation of the
prospects of varying spectral configurations of new generation sensors. International Journal of Applied Earth
Observation and Geoinformation. 62:47–55.
Sibanda M, Mutanga O, Rouget M. 2015. Examining the potential of Sentinel-2 MSI spectral resolution in quantifying
above ground biomass across different fertilizer treatments. ISPRS J Photogram Remote Sens. 110:55–65.
Sibanda M, Mutanga O, Rouget M, Odindi J. 2015. Exploring the potential of in situ hyperspectral data and multivariate
techniques in discriminating different fertilizer treatments in grasslands. J Appl Remote Sens. 9(1):096033–096033.
Sibanda M, Mutanga O, Rouget M. 2016. Discriminating Rangeland Management Practices Using Simulated HyspIRI,
Landsat 8 OLI, Sentinel 2 MSI, and VENμS Spectral Data. IEEE J Sel Topics Appl Earth Obs Remote Sens. 9:3957–3969.
Sibanda M, Mutanga O, Rouget M. 2017. Testing the capabilities of the new WorldView-3 space-borne sensor’s rededge spectral band in discriminating and mapping complex grassland management treatments. Int J Remote Sens.
38(1):1–22.
Singh L, Mutanga O, Mafongoya P, Peerbhay KY. 2017. Multispectral mapping of key grassland nutrients in KwaZuluNatal, South Africa. J Spatial Sci. 1–18.
Stears K, Shrader AM. 2015. Increases in food availability can tempt oribi antelope into taking greater risks at both large
and small spatial scales. Anim Behav. 108:155–164.
Story M, Congalton RG. 1986. Accuracy assessment: a user’s perspective. Photogram Eng Remote Sens. 52(3):397–399.
Tarantino C, Adamo M, Lucas R, Blonda P. 2016. Detection of changes in semi-natural grasslands by cross correlation
analysis with WorldView-2 images and new Landsat 8 data. Remote Sens Environ. 175:65–72.
Terry E, M, E. & JCO, D. T. 2008. Report on Fort Nottingham Vegetation Assessment. Pietermaritzburg: University of
KwaZulu-Natal.
Thenkabail PS, Lyon JG, Huete A. 2011. Hyperspectral remote sensing of vegetation. New York: CRC Press.
Thomas S. 2013. Predicting the spatial distribution of an invasive plant species and modeling tolerance to herbivory using
Lythrum salicaria L as a model system. Doctor of Philosophy Dissertation, Iowa State University.
Tyc G, Tulip J, Schulten D, Krischke M, Oxfort M. 2005. The RapidEye mission design. Acta Astronaut. 56(1):213–219.
Ullah S, Si Y, Schlerf M, Skidmore AK, Shafique M, Iqbal IA. 2012. Estimation of grassland biomass and nitrogen using
MERIS data. Int J Appl Earth Obs Geoinform. 19:196–204.
Van Wyk E, Van Oudtshoorn F. 1999. Guide to grasses of Southern Africa. Pretoria: Briza.
Vane G, Green RO, Chrien TG, Enmark HT, Hansen EG, Porter WM. 1993. The airborne visible/infrared imaging
spectrometer (AVIRIS). Remote Sens Environ. 44(2–3):127–143.
Walde J. 2014. Discriminant analysis. December.
Weng Q. 2002. Land use change analysis in the Zhujiang Delta of China using satellite remote sensing, GIS and stochastic
modelling. J Environ Managem. 64(3):273–284.
van Wilgen BW, Forsyth GG, Le Maitre DC, Wannenburgh A, Kotzé JD, van den Berg E, Henderson L. 2012. An
assessment of the effectiveness of a large, national-scale invasive alien plant control strategy in South Africa. Biol
Conserv. 148(1):28–38.
Xia J, Liu S, Liang S, Chen Y, Xu W, Yuan W. 2014. Spatio-temporal patterns and climate variables controlling of biomass
carbon stock of global grassland ecosystems from 1982 to 2006. Remote Sens. 6(3):1783–1802.
XLSTAT. 2013. Complete data analysis software system and statistics add-in for MS Excel, Version 2014. 2nd ed. Belmont,
CA: Statistical Innovations.
Zhang C, Xie Z. 2012. Combining object-based texture measures with a neural network for vegetation mapping in the
Everglades from hyperspectral imagery. Remote Sens Environ. 124:310–320.
Zhou P, Huang J, Pontius RG, Hong H. 2014. Land classification and change intensity analysis in a coastal watershed of
Southeast China. Sensors. 14(7):11640–11658.
Zhu L, Sun OJ, Sang W, Li Z, Ma K. 2007. Predicting the spatial distribution of an invasive plant species (Eupatorium
adenophorum) in China. Landscape Ecol. 22(8):1143–1154.