196
In their review on tree species mapping, Fassnacht et al. (2016) also state that the
VIS region (350–650 nm) contains the most often selected features for tree species
mapping, without mentioning the relation to RGB. A recent study comparing spectral indices derived from RGB and NIR camera images to multispectral imagery
(Rasmussen et al. 2016) showed that cheaper RGB / NIR cameras are equal in performance for mapping barley biomass in agricultural fields. Another study (Fischer
et al. 2012) compared an NDVI calculated from high spectral resolution field spectrometer (i.e. ASD Field Spec 3) with an NDVI derived from a modified Olympus
consumer-grade camera for mapping the spatial variability of NDVI in biotic soil
crusts; they found strong correlations with R
2
of 0.91.
In this study, we found that RGB bands were an important predictor but the least
important was texture. This is surprising as we thought that texture would have a
high potential for describing crown properties related to shadow patterning or variation in greenness. Fassnacht et al. (2016) list several studies that applied texture to
improve tree species classification by 10–15%. However, using texture also creates
problems that make its use seem clumsy and time consuming. First of all, the idea of
texture is a multiscale problem. The relevant scale (i.e., the size of the window in
which a texture value is calculated) has to be identified empirically. Therefore, different window sizes have to be compared with these usually being 3×3, 5×5, 7×7,
9×9 and so on. When UAV data volumes are large (>1 GB), this quickly becomes
unwieldy. Also, with very high resolution images, larger window sizes are needed
but these slow processing times. Other options such as the offset and the number of
grey levels considered, provide further opportunities to optimize the results; yet
leading to a seemingly endless endeavour in finding the right parameter settings.
Second, different species might require different window sizes. This seems logical
but is difficult to realize technically. Third, the large number of available texture
measures makes it difficult to select those that are optimal or most appropriate. This
is complicated by the typically high correlation between the different texture measures. In our study, all parameters were kept stable (i.e., a window size of 5×5, offset
of 1×1 and 16 grey levels). Not testing different settings might explain the poor
performance of the texture parameters. We also used only a small fraction of the
available texture measures. Other texture measures related to tree crown shape and
size could have been considered (Fassnacht et al. 2016). The Orfeo Toolbox provides
around 40 different texture measures in total. Hence, texture measures derived from
UAV imagery require more studies on selecting and optimising the best measures
and optimal window sizes for tree species discrimination or tree crown analyses.
Tree species mapping through remote sensing data can become an efficient tool
in biodiversity monitoring. However, the nature of biodiversity is that communities
under study almost always consist of common and rare species (Magurran and
McGill 2011). The occurrence of rare tree species (rare equal to less than ten individuals overall) severely affected the potential to classify the whole tree species
pool. Out of 16 species, seven species were considered as too rare to be used in the
classification. Another five infrequent species, i.e. with less than 30 tree crowns for
training and testing, led to poor classification results due to the small amount of
training data. However, the infrequent species were impossible to classify correctly,
J. Oldeland et al.
In their review on tree species mapping, Fassnacht et al. (2016) also state that the
VIS region (350–650 nm) contains the most often selected features for tree species
mapping, without mentioning the relation to RGB. A recent study comparing spectral indices derived from RGB and NIR camera images to multispectral imagery
(Rasmussen et al. 2016) showed that cheaper RGB / NIR cameras are equal in performance for mapping barley biomass in agricultural fields. Another study (Fischer
et al. 2012) compared an NDVI calculated from high spectral resolution field spectrometer (i.e. ASD Field Spec 3) with an NDVI derived from a modified Olympus
consumer-grade camera for mapping the spatial variability of NDVI in biotic soil
crusts; they found strong correlations with R
2
of 0.91.
In this study, we found that RGB bands were an important predictor but the least
important was texture. This is surprising as we thought that texture would have a
high potential for describing crown properties related to shadow patterning or variation in greenness. Fassnacht et al. (2016) list several studies that applied texture to
improve tree species classification by 10–15%. However, using texture also creates
problems that make its use seem clumsy and time consuming. First of all, the idea of
texture is a multiscale problem. The relevant scale (i.e., the size of the window in
which a texture value is calculated) has to be identified empirically. Therefore, different window sizes have to be compared with these usually being 3×3, 5×5, 7×7,
9×9 and so on. When UAV data volumes are large (>1 GB), this quickly becomes
unwieldy. Also, with very high resolution images, larger window sizes are needed
but these slow processing times. Other options such as the offset and the number of
grey levels considered, provide further opportunities to optimize the results; yet
leading to a seemingly endless endeavour in finding the right parameter settings.
Second, different species might require different window sizes. This seems logical
but is difficult to realize technically. Third, the large number of available texture
measures makes it difficult to select those that are optimal or most appropriate. This
is complicated by the typically high correlation between the different texture measures. In our study, all parameters were kept stable (i.e., a window size of 5×5, offset
of 1×1 and 16 grey levels). Not testing different settings might explain the poor
performance of the texture parameters. We also used only a small fraction of the
available texture measures. Other texture measures related to tree crown shape and
size could have been considered (Fassnacht et al. 2016). The Orfeo Toolbox provides
around 40 different texture measures in total. Hence, texture measures derived from
UAV imagery require more studies on selecting and optimising the best measures
and optimal window sizes for tree species discrimination or tree crown analyses.
Tree species mapping through remote sensing data can become an efficient tool
in biodiversity monitoring. However, the nature of biodiversity is that communities
under study almost always consist of common and rare species (Magurran and
McGill 2011). The occurrence of rare tree species (rare equal to less than ten individuals overall) severely affected the potential to classify the whole tree species
pool. Out of 16 species, seven species were considered as too rare to be used in the
classification. Another five infrequent species, i.e. with less than 30 tree crowns for
training and testing, led to poor classification results due to the small amount of
training data. However, the infrequent species were impossible to classify correctly,
J. Oldeland et al.