185
coverage and a restricted carrying capacity which renders UAVs unsuitable for
heavy hyperspectral or LiDAR sensors and thus being restricted to consumer-grade
cameras or small multispectral cameras. Only a few studies have evaluated RGB
spectral indices for species discrimination, e.g. Rasmussen et al. (2013), who tested
the potential of RGB indices for site-specific weed management, Dvořák et al.
(2015) used RGB imagery for invasive plant detection, and Rasmussen et al. (2016)
studied the performance of RGB indices to measure barley biomass. Hence, the
question remains open as to whether very high spatial resolution imagery taken by
a standard UAV can successfully discriminate tree species. If so, nature conservation could make use of a very flexible image acquisition platform for monitoring
small areas, i.e. covering several square kilometres with a small number of flights.
Furthermore, the question how consumer-grade cameras with RGB or a NIR-filter
perform in such a task needs to be addressed. Is NIR really necessary or are observations in RGB sufficient?
Most of the studies reviewed by Fassnacht et al. (2016) had two things in common: they used hyperspectral imagery in combination with LiDAR data and were
undertaken in temperate or boreal forests. Bunting and Lucas (2006) and Lucas
et al. (2008) established the use of CASI and HYMAP hyperspectral data for discriminating tree species in open woodlands and forests in Queensland, Australia,
confirming that differences in the mean spectra from crown objects increased the
accuracy of discrimination. However, only a few studies set out to classify savannah
tree species in southern Africa (Naidoo et al. 2012; Cho et al. 2012; Colgan et al.
2012). These studies classified between six and 15 tree species. Cho et al. (2015)
also tested the suitability of very high resolution satellite imagery for this purpose,
but only used three out of ten dominant canopy species. It seems that the abundance
of a tree species also contributes to its capability for being mapped precisely. We are
of the opinion that this issue has not been sufficiently highlighted in the literature
(but see comments in Fassnacht et al. 2016).
Study Area
The study was part of the Biodiversity Observatory S05 of the BIOTA Africa project
(www.biota-africa.org), which is a cross-country biodiversity monitoring project
with a standardized monitoring approach performing monitoring in southern, western, and northern Africa (Jürgens et al. 2012). The observatory is located on the
cattle farm Erichsfelde (coordinates: 16.935° E 21.597° S) in central Namibia. The
Biodiversity Observatory spans 1 km
2
and is divided into 100 ha from which 20
were selected in the year 2001 for permanent annual monitoring of vegetation and
animal diversity. The vegetation monitoring was undertaken within plots of
20 × 50 m, which were situated at the mid-point of a selected hectare. The vegetation consisted of typical Thornbush savanna sensu Giess (1998), dominated by
Acacia mellifera subsp. detinens and Boscia albitrunca. Other Acacia species also
occurred, in particular A. hebeclada subsp. hebeclada, A. tortilis, A. reficiens and A.
The Potential of UAV Derived Image Features for Discriminating Savannah Tree Species
coverage and a restricted carrying capacity which renders UAVs unsuitable for
heavy hyperspectral or LiDAR sensors and thus being restricted to consumer-grade
cameras or small multispectral cameras. Only a few studies have evaluated RGB
spectral indices for species discrimination, e.g. Rasmussen et al. (2013), who tested
the potential of RGB indices for site-specific weed management, Dvořák et al.
(2015) used RGB imagery for invasive plant detection, and Rasmussen et al. (2016)
studied the performance of RGB indices to measure barley biomass. Hence, the
question remains open as to whether very high spatial resolution imagery taken by
a standard UAV can successfully discriminate tree species. If so, nature conservation could make use of a very flexible image acquisition platform for monitoring
small areas, i.e. covering several square kilometres with a small number of flights.
Furthermore, the question how consumer-grade cameras with RGB or a NIR-filter
perform in such a task needs to be addressed. Is NIR really necessary or are observations in RGB sufficient?
Most of the studies reviewed by Fassnacht et al. (2016) had two things in common: they used hyperspectral imagery in combination with LiDAR data and were
undertaken in temperate or boreal forests. Bunting and Lucas (2006) and Lucas
et al. (2008) established the use of CASI and HYMAP hyperspectral data for discriminating tree species in open woodlands and forests in Queensland, Australia,
confirming that differences in the mean spectra from crown objects increased the
accuracy of discrimination. However, only a few studies set out to classify savannah
tree species in southern Africa (Naidoo et al. 2012; Cho et al. 2012; Colgan et al.
2012). These studies classified between six and 15 tree species. Cho et al. (2015)
also tested the suitability of very high resolution satellite imagery for this purpose,
but only used three out of ten dominant canopy species. It seems that the abundance
of a tree species also contributes to its capability for being mapped precisely. We are
of the opinion that this issue has not been sufficiently highlighted in the literature
(but see comments in Fassnacht et al. 2016).
Study Area
The study was part of the Biodiversity Observatory S05 of the BIOTA Africa project
(www.biota-africa.org), which is a cross-country biodiversity monitoring project
with a standardized monitoring approach performing monitoring in southern, western, and northern Africa (Jürgens et al. 2012). The observatory is located on the
cattle farm Erichsfelde (coordinates: 16.935° E 21.597° S) in central Namibia. The
Biodiversity Observatory spans 1 km
2
and is divided into 100 ha from which 20
were selected in the year 2001 for permanent annual monitoring of vegetation and
animal diversity. The vegetation monitoring was undertaken within plots of
20 × 50 m, which were situated at the mid-point of a selected hectare. The vegetation consisted of typical Thornbush savanna sensu Giess (1998), dominated by
Acacia mellifera subsp. detinens and Boscia albitrunca. Other Acacia species also
occurred, in particular A. hebeclada subsp. hebeclada, A. tortilis, A. reficiens and A.
The Potential of UAV Derived Image Features for Discriminating Savannah Tree Species