183
© Springer International Publishing AG 2017
R. Díaz-Delgado et al. (eds.), The Roles of Remote Sensing in Nature Conservation,
DOI 10.1007/978-3-319-64332-8_10
The Potential of UAV Derived Image Features
for Discriminating Savannah Tree Species
J. Oldeland, A. Große-Stoltenberg, L. Naftal, and B.J. Strohbach
Abstract Mapping tree species at the single-tree level is an active field of research
linking ecology and remote sensing. However, the discrimination of tree species
requires the selection of the relevant spectral features derived from imagery. We can
extract an extensive number of image parameters even from images with a low spectral resolution, such as Red-Green-Blue (RGB) or near-infrared (NIR) images.
Hence, identifying the most relevant image parameters for tree species discrimination is still an issue. We generated 42 parameters from very high resolution images
acquired by Unmanned Aerial Vehicles (UAV), such as chromatic coordinates, spectral indices, texture measures and a canopy height model (CHM). The aim of this
study was to compare the relevance of these components for classifying savannah
tree species. We obtained very high (5 cm) pixel resolution RGB-NIR imagery with
a delta-wing UAV in a thorn bush savannah landscape in central Namibia in April
2016. Simultaneously, we gathered ground truth data on the location of 478 individual trees and large shrubs belonging to 16 species. We then used a Random
Forest classifier on single and combined thematic sets of image data, e.g. RGB,
NIR, texture and in combination with CHM. The best average overall accuracy was
0.77 and the best Cohen´s Kappa value was 0.63 for a combination of RGB imagery
and the CHM. Our results are comparable to other studies using hyperspectral data
and LiDAR information. We further found that the abundance of the tree species is
crucial for successful mapping, with only species with a high abundance being classified satisfactorily. Diverse ecosystems such as savannahs could therefore be a
challenge for future tree mapping projects. Nevertheless, this study indicates that
UAV-borne RGB imagery seems promising for detailed mapping of tree species.
J. Oldeland (*)
Biodiversity, Ecology and Evolution of Plants, Biocentre Klein Flottbek and Botanical
Garden, University of Hamburg, Ohnhorststr. 18, Hamburg 22609, Germany
e-mail: jens.oldeland@uni-hamburg.de
A. Große-Stoltenberg
Applied Landscape Ecology and Ecological Planning Institute of Landscape Ecology,
Heisenbergstraße 2, Münster 48149, Germany
e-mail: ags@uni-muenster.de
L. Naftal • B.J. Strohbach
Faculty of Natural Resources and Spatial Sciences, Namibia University of Science and
Technology, P/Bag, 13388, Windhoek, Namibia
e-mail: wexly4one@gmail.com; bstrohbach@nust.na
© Springer International Publishing AG 2017
R. Díaz-Delgado et al. (eds.), The Roles of Remote Sensing in Nature Conservation,
DOI 10.1007/978-3-319-64332-8_10
The Potential of UAV Derived Image Features
for Discriminating Savannah Tree Species
J. Oldeland, A. Große-Stoltenberg, L. Naftal, and B.J. Strohbach
Abstract Mapping tree species at the single-tree level is an active field of research
linking ecology and remote sensing. However, the discrimination of tree species
requires the selection of the relevant spectral features derived from imagery. We can
extract an extensive number of image parameters even from images with a low spectral resolution, such as Red-Green-Blue (RGB) or near-infrared (NIR) images.
Hence, identifying the most relevant image parameters for tree species discrimination is still an issue. We generated 42 parameters from very high resolution images
acquired by Unmanned Aerial Vehicles (UAV), such as chromatic coordinates, spectral indices, texture measures and a canopy height model (CHM). The aim of this
study was to compare the relevance of these components for classifying savannah
tree species. We obtained very high (5 cm) pixel resolution RGB-NIR imagery with
a delta-wing UAV in a thorn bush savannah landscape in central Namibia in April
2016. Simultaneously, we gathered ground truth data on the location of 478 individual trees and large shrubs belonging to 16 species. We then used a Random
Forest classifier on single and combined thematic sets of image data, e.g. RGB,
NIR, texture and in combination with CHM. The best average overall accuracy was
0.77 and the best Cohen´s Kappa value was 0.63 for a combination of RGB imagery
and the CHM. Our results are comparable to other studies using hyperspectral data
and LiDAR information. We further found that the abundance of the tree species is
crucial for successful mapping, with only species with a high abundance being classified satisfactorily. Diverse ecosystems such as savannahs could therefore be a
challenge for future tree mapping projects. Nevertheless, this study indicates that
UAV-borne RGB imagery seems promising for detailed mapping of tree species.
J. Oldeland (*)
Biodiversity, Ecology and Evolution of Plants, Biocentre Klein Flottbek and Botanical
Garden, University of Hamburg, Ohnhorststr. 18, Hamburg 22609, Germany
e-mail: jens.oldeland@uni-hamburg.de
A. Große-Stoltenberg
Applied Landscape Ecology and Ecological Planning Institute of Landscape Ecology,
Heisenbergstraße 2, Münster 48149, Germany
e-mail: ags@uni-muenster.de
L. Naftal • B.J. Strohbach
Faculty of Natural Resources and Spatial Sciences, Namibia University of Science and
Technology, P/Bag, 13388, Windhoek, Namibia
e-mail: wexly4one@gmail.com; bstrohbach@nust.na