195
and NGRDI occurred in the All and Frequent species subsets. Highest importance
values were achieved by the RGB indices exG2 and chrR both with 32% and 20%
for Frequent and All species subsets respectively. In the case of the infrequent species, two texture measures appear in the ten most important variables. However,
their variable importance did not reach values higher than five percent, rendering all
parameters for the infrequent species redundant.
In summary, the frequent species dataset together with RGB+CHM image
parameters provided the highest accuracy in the discrimination of tree species, used
the lowest number of predictors and provided the smallest confidence intervals.
Discussion
We evaluated the relevance of different parameters derived from very high resolution RGB-NIR imagery for the discrimination of savannah tree species. We could
confirm the commonly found pattern that information based on the visual part of the
spectrum is important for discriminating tree species (Fassnacht et al. 2016). In particular, we found that RGB-based spectral indices (Meyer and Neto 2008; Rasmussen
et al. 2016) and simple chromatic coordinates (Woebbecke et al. 1995) in combination with a canopy height model (CHM) achieved the best results. By contrast, the
performance of the NIR image parameters was weak and deteriorated when combined with the CHM. Our overall accuracy of 77% (on average, maximum was 0.83)
is comparable to the results of the recent review of Fassnacht (see Fig. 3 in Fassnacht
et al. (2016)) who analysed 129 case studies on tree species mapping.
Most case studies have used a combination of hyperspectral and/or LiDAR for
tree species mapping and have typically achieved overall accuracies of between 75
and 90%. Of these, three were carried out in southern Africa, all in Kruger National
Park, and all used hyperspectral image data and height information derived from
LiDAR sensors. Naidoo et al. (2012) achieved 82% with four hyperspectral indices
(including NDVI) and height information; whilst Cho et al. (2012) also used hyperspectral data but resampled them to seven World View 2 multispectral bands and
combined them with LiDAR based height information. They achieved OA values of
between 63 and 81%. Colgan et al. (2012) used LiDAR-based height information
and BRDF corrected reflectance values for the VIS-NIR region. The bidirectional
reflectance distribution function (BRDF) is a function describing the change in
reflectance values due to view angle and sun position during image assessment. The
BRDF correction improved the hyperspectral information and thus led to OAs of
between 70 and 78%. Hence, our UAV based tree species discrimination approach,
requiring only a ≈100$ RGB camera, performed equally well, when compared to
the technically more sophisticated, and also much more expensive, hyperspectral
and LiDAR sensors.
VIS-NIR imagery is by far the most commonly used data source for generating
spectral indices. Pure RGB based spectral information has been used less often in
remote sensing studies that have focused on the discrimination and mapping species.
The Potential of UAV Derived Image Features for Discriminating Savannah Tree Species
and NGRDI occurred in the All and Frequent species subsets. Highest importance
values were achieved by the RGB indices exG2 and chrR both with 32% and 20%
for Frequent and All species subsets respectively. In the case of the infrequent species, two texture measures appear in the ten most important variables. However,
their variable importance did not reach values higher than five percent, rendering all
parameters for the infrequent species redundant.
In summary, the frequent species dataset together with RGB+CHM image
parameters provided the highest accuracy in the discrimination of tree species, used
the lowest number of predictors and provided the smallest confidence intervals.
Discussion
We evaluated the relevance of different parameters derived from very high resolution RGB-NIR imagery for the discrimination of savannah tree species. We could
confirm the commonly found pattern that information based on the visual part of the
spectrum is important for discriminating tree species (Fassnacht et al. 2016). In particular, we found that RGB-based spectral indices (Meyer and Neto 2008; Rasmussen
et al. 2016) and simple chromatic coordinates (Woebbecke et al. 1995) in combination with a canopy height model (CHM) achieved the best results. By contrast, the
performance of the NIR image parameters was weak and deteriorated when combined with the CHM. Our overall accuracy of 77% (on average, maximum was 0.83)
is comparable to the results of the recent review of Fassnacht (see Fig. 3 in Fassnacht
et al. (2016)) who analysed 129 case studies on tree species mapping.
Most case studies have used a combination of hyperspectral and/or LiDAR for
tree species mapping and have typically achieved overall accuracies of between 75
and 90%. Of these, three were carried out in southern Africa, all in Kruger National
Park, and all used hyperspectral image data and height information derived from
LiDAR sensors. Naidoo et al. (2012) achieved 82% with four hyperspectral indices
(including NDVI) and height information; whilst Cho et al. (2012) also used hyperspectral data but resampled them to seven World View 2 multispectral bands and
combined them with LiDAR based height information. They achieved OA values of
between 63 and 81%. Colgan et al. (2012) used LiDAR-based height information
and BRDF corrected reflectance values for the VIS-NIR region. The bidirectional
reflectance distribution function (BRDF) is a function describing the change in
reflectance values due to view angle and sun position during image assessment. The
BRDF correction improved the hyperspectral information and thus led to OAs of
between 70 and 78%. Hence, our UAV based tree species discrimination approach,
requiring only a ≈100$ RGB camera, performed equally well, when compared to
the technically more sophisticated, and also much more expensive, hyperspectral
and LiDAR sensors.
VIS-NIR imagery is by far the most commonly used data source for generating
spectral indices. Pure RGB based spectral information has been used less often in
remote sensing studies that have focused on the discrimination and mapping species.
The Potential of UAV Derived Image Features for Discriminating Savannah Tree Species