198
Conclusions
In this study, we evaluated the relevance of RGB and NIR image products, derived
from UAV images, for discriminating tree species in a Namibian savannah. We
found that data acquired in the NIR wavelength region only were not sufficient or
even necessary, although this conclusion might have been incorrectly drawn because
of co-registration problems between the NIR and RGB imagery. Permanently
marked, well-surveyed ground control points therefore need to be planned for future
image acquisition campaigns. Nevertheless, the OAs achieved with RGB data and
CHM were comparable to other studies that used more expensive hyperspectral data
and LiDAR instruments. This indicated that UAVs have a high potential for future
tree species mapping tasks if areas less than 1 km
2
are to be monitored. However,
the number of species that can be mapped or discriminated seems independent of
the sensor type. The assumption is that hyperspectral data theoretically can outperform RGB-NIR data when a large number of species are present. However, for the
process of training a classifier, such as a Random Forest, the number of training
polygons needs to be at least 30 in order to achieve sufficient and acceptable accuracies. This seems not feasible when rare species (i.e., less than ten individuals per
square kilometre) are present. Hence, a significant future challenge is the task of
mapping species with low abundances.
Practical Application for Nature Conservation
This chapter dealt with the application of UAV-borne consumer grade cameras for
discriminating savannah tree species and has several important messages for practical applications in nature conservation. Firstly, the delta-wing UAV that we employed,
the eBee system (SenseFly 2015), is capable of capturing an area of 1 km
2
during a
single flight when the desired resolution is a 5 cm pixel size or greater. Smaller pixel
sizes, e.g. 2 cm, can only be achieved in several flights (four to five). However, this
this also doubles the disk space required for storing the imagery. Affordable quadcopter systems cannot usually cover 1 km
2
in a single flight. Secondly, we showed
that tree species discrimination based solely on RGB + Canopy Height is possible,
suggesting that a second flight with a NIR camera is potentially unnecessary.
However, we need more studies comparing RGB based spectral indices to NIR based
spectral indices in order to see whether RGB can replace NIR indices in the future.
Finally, we found that ground truthing should take the abundance or frequency of the
species into consideration. We suggest using a minimum of 50 individuals per species for training purposes in order to be successfully mapped. Species from the same
genus, e.g. the different Acacia species in our study, often share similar spectral
properties and thus are very difficult to distinguish. One alternative is to map these at
the genus level, if it is not the users demand to produce species specific map. In conclusion, we have shown that using UAVs to map the individual stems of tree species
could be a cheap and very flexible tool for nature conservation in the near future.
J. Oldeland et al.
Conclusions
In this study, we evaluated the relevance of RGB and NIR image products, derived
from UAV images, for discriminating tree species in a Namibian savannah. We
found that data acquired in the NIR wavelength region only were not sufficient or
even necessary, although this conclusion might have been incorrectly drawn because
of co-registration problems between the NIR and RGB imagery. Permanently
marked, well-surveyed ground control points therefore need to be planned for future
image acquisition campaigns. Nevertheless, the OAs achieved with RGB data and
CHM were comparable to other studies that used more expensive hyperspectral data
and LiDAR instruments. This indicated that UAVs have a high potential for future
tree species mapping tasks if areas less than 1 km
2
are to be monitored. However,
the number of species that can be mapped or discriminated seems independent of
the sensor type. The assumption is that hyperspectral data theoretically can outperform RGB-NIR data when a large number of species are present. However, for the
process of training a classifier, such as a Random Forest, the number of training
polygons needs to be at least 30 in order to achieve sufficient and acceptable accuracies. This seems not feasible when rare species (i.e., less than ten individuals per
square kilometre) are present. Hence, a significant future challenge is the task of
mapping species with low abundances.
Practical Application for Nature Conservation
This chapter dealt with the application of UAV-borne consumer grade cameras for
discriminating savannah tree species and has several important messages for practical applications in nature conservation. Firstly, the delta-wing UAV that we employed,
the eBee system (SenseFly 2015), is capable of capturing an area of 1 km
2
during a
single flight when the desired resolution is a 5 cm pixel size or greater. Smaller pixel
sizes, e.g. 2 cm, can only be achieved in several flights (four to five). However, this
this also doubles the disk space required for storing the imagery. Affordable quadcopter systems cannot usually cover 1 km
2
in a single flight. Secondly, we showed
that tree species discrimination based solely on RGB + Canopy Height is possible,
suggesting that a second flight with a NIR camera is potentially unnecessary.
However, we need more studies comparing RGB based spectral indices to NIR based
spectral indices in order to see whether RGB can replace NIR indices in the future.
Finally, we found that ground truthing should take the abundance or frequency of the
species into consideration. We suggest using a minimum of 50 individuals per species for training purposes in order to be successfully mapped. Species from the same
genus, e.g. the different Acacia species in our study, often share similar spectral
properties and thus are very difficult to distinguish. One alternative is to map these at
the genus level, if it is not the users demand to produce species specific map. In conclusion, we have shown that using UAVs to map the individual stems of tree species
could be a cheap and very flexible tool for nature conservation in the near future.
J. Oldeland et al.