197
as can be seen from the non-significant or low quality models (Table 2, Fig. 3). Thus
out of 16 species, only the four frequent species could be classified to acceptable
levels of accuracy. The pattern that only frequent species can be mapped with sufficient accuracy is confirmed by many other studies (Naidoo et al. 2012; Immitzer
et al. 2012; Cho et al. 2012; Baldeck et al. 2015). The review of Fassnacht et al.
(2016) reports the number of species that were classified in the analysed studies
ranging from two to seventeen with an average of five or six. This finding has important implications for future biodiversity monitoring that should be based on tree
species mapping leading to a complete census. Mapping rare tree and shrub species
becomes a challenge when too few individuals can be found for training and testing
a classifier. Thus, in future studies, more emphasis should be put on high quality
ground truth data gathering an equal number of ground truth tree locations per species. In the study of Colgan et al. (2012), three species made up 30% of the landscape while the category “other” also had 20% of all occurring tree crowns. Thus,
rare species can make a large fraction of tree crowns in a savannah but are represented by a small number of individuals per species. Common trees however, bear
different challenges. For example, a high genus – species ratio (i.e. where many
species of the same genus occur as in the genus Acacia or Combretum) means these
species are sometimes lumped together into a single tree category at genus level
(Naidoo et al. 2012). The species abundances in (semi-)natural ecosystems are
much more complex than in temperate forests and will require special considerations for an operative tree species mapping based on remote sensing imagery.
Although our study was successful in discriminating selected savanna tree species with a UAV-borne RGB camera, the limitations of UAVs in comparison to airborne or satellite-borne sensors requires discussion. In our case, the largest obstacle
was the mismatch in the co-registration of NIR and RGB imagery, which had to be
corrected manually. Better results could be achieved when using multispectral cameras or even lightweight hyperspectral cameras. The spatial mismatch could have
been the reason why the averaged tree crown parameters were worse for NIR than
for RGB. Digitization of the tree crowns was also undertaken manually using only
the RGB imagery. Hence, it is possible that the NIR imagery parameters contained
a higher shadow fraction or parts of neighbouring tree crowns. Although manual
digitization seems straightforward, it is also error prone and could be avoided by
using specifically designed algorithms or software packages, e.g. TIDA (Culvenor,
2002), JSEG (Kang et al. 2016) or ITCsegment (Dalponte and Coomes 2016). Other
serious problems connected to light and shadowing effects that can occur when
using UAV imagery are discussed by Rasmussen et al. (2016). In our study, the different flight directions during the drone overflight affected the brightness pattern.
Rasmussen et al. (2016) also mentioned that BRDF effects affect the outcome of a
study when not taken into consideration. These issues, co-registration and changing
light conditions (including BRDF effects), seem to compromise the utility of UAV
imagery. Ground control points should be essential for proper image co-registration,
however, often these require expensive differential-GPS equipment. Further
improvement of technical equipment or standardized procedures for UAV image
acquisition should bring remedies in the future.
The Potential of UAV Derived Image Features for Discriminating Savannah Tree Species
as can be seen from the non-significant or low quality models (Table 2, Fig. 3). Thus
out of 16 species, only the four frequent species could be classified to acceptable
levels of accuracy. The pattern that only frequent species can be mapped with sufficient accuracy is confirmed by many other studies (Naidoo et al. 2012; Immitzer
et al. 2012; Cho et al. 2012; Baldeck et al. 2015). The review of Fassnacht et al.
(2016) reports the number of species that were classified in the analysed studies
ranging from two to seventeen with an average of five or six. This finding has important implications for future biodiversity monitoring that should be based on tree
species mapping leading to a complete census. Mapping rare tree and shrub species
becomes a challenge when too few individuals can be found for training and testing
a classifier. Thus, in future studies, more emphasis should be put on high quality
ground truth data gathering an equal number of ground truth tree locations per species. In the study of Colgan et al. (2012), three species made up 30% of the landscape while the category “other” also had 20% of all occurring tree crowns. Thus,
rare species can make a large fraction of tree crowns in a savannah but are represented by a small number of individuals per species. Common trees however, bear
different challenges. For example, a high genus – species ratio (i.e. where many
species of the same genus occur as in the genus Acacia or Combretum) means these
species are sometimes lumped together into a single tree category at genus level
(Naidoo et al. 2012). The species abundances in (semi-)natural ecosystems are
much more complex than in temperate forests and will require special considerations for an operative tree species mapping based on remote sensing imagery.
Although our study was successful in discriminating selected savanna tree species with a UAV-borne RGB camera, the limitations of UAVs in comparison to airborne or satellite-borne sensors requires discussion. In our case, the largest obstacle
was the mismatch in the co-registration of NIR and RGB imagery, which had to be
corrected manually. Better results could be achieved when using multispectral cameras or even lightweight hyperspectral cameras. The spatial mismatch could have
been the reason why the averaged tree crown parameters were worse for NIR than
for RGB. Digitization of the tree crowns was also undertaken manually using only
the RGB imagery. Hence, it is possible that the NIR imagery parameters contained
a higher shadow fraction or parts of neighbouring tree crowns. Although manual
digitization seems straightforward, it is also error prone and could be avoided by
using specifically designed algorithms or software packages, e.g. TIDA (Culvenor,
2002), JSEG (Kang et al. 2016) or ITCsegment (Dalponte and Coomes 2016). Other
serious problems connected to light and shadowing effects that can occur when
using UAV imagery are discussed by Rasmussen et al. (2016). In our study, the different flight directions during the drone overflight affected the brightness pattern.
Rasmussen et al. (2016) also mentioned that BRDF effects affect the outcome of a
study when not taken into consideration. These issues, co-registration and changing
light conditions (including BRDF effects), seem to compromise the utility of UAV
imagery. Ground control points should be essential for proper image co-registration,
however, often these require expensive differential-GPS equipment. Further
improvement of technical equipment or standardized procedures for UAV image
acquisition should bring remedies in the future.
The Potential of UAV Derived Image Features for Discriminating Savannah Tree Species