190
A recent, simplified version of the excessive greenness index (Rasmussen et al.
2016) was also calculated, in this study called exG2.
To exploit the additional NIR information, we calculated slope and distance
based vegetation indices (Silleos et al. 2006). Slope based vegetation indices make
use of the difference in the slope of the red and NIR channel; the famous Normalized
Difference Vegetation Index (NDVI, Tucker 1979) belongs here. The SAVI is a
modified NDVI which adjusts for potential effects of bare soil (Huete 1988).
Thiam’s vegetation index improves on the NDVI by multiplying the absolute NIR
and Red band values with their square root (Thiam 1998). Distance based vegetation indices make use of the concept of the so-called “soil-line” (Silleos et al. 2006).
The distances refer to the distance of samples in the two dimensional red-NIR spectral space to the soil line, that describes the lower boundary of pixels in this space,
usually aligning across a clearly visible axis. To determine the soil line parameters
required for the calculation of the distance based vegetation indices, a set of n = 100
bare soil pixels were selected, stratified by the hectare grid of the Biodiversity
Observatory, and the NIR and red values were extracted. Based on these values, a
linear regression (R
2
= 0.89, p < 0.001) was used to estimate the intercept and the
slope of the soil line. The linear regression parameters intercept (alpha = −227.29)
and slope (beta=1.877) were used to calculate the Perpendicular Vegetation Index
III (Qi et al. 1994; Silleos et al. 2006). All indices were calculated and image manipulations were performed with the open source software SAGA-GIS (Conrad et al.
2015). For all individual tree crown polygons, we calculated values for the mean
and standard deviations using the zonal statistics tool in SAGA-GIS.
Image Texture
Richards (2013) suggested that the texture of an image can be described as smooth,
rough or repetitive in terms of the spatial arrangement of grey values. In terms of
canopy cover this would describe whether tree crowns consist of repeating patterns
of shadow and greenness or whether the canopy is closed and thus equal in colour.
Often texture measures will improve remote sensing classifiers (Krefis et al. 2011).
As our main interest was to discriminate between tree species canopies, the greenness (exG) of the canopy seemed to be a good parameter for a texture analysis. We
used the Orfeo Toolbox v.5.6.1 (McInerney and Kempeneers 2015), a free open
source software for remote sensing image analysis, to calculate eight different types
of simple image texture measures. Haralick’s grey level occurrence matrix (GLCM,
Haralick et al. 1973), which is a standard for describing image texture, was the basis
for calculating all of the texture measures. We choose a constant window size of 5×5
pixels and an offset of 1 for x and y. The number of grey levels was set to 16. We
then loaded the calculated image texture measures into SAGA-GIS and extracted
the texture as average and standard deviation for each individual tree crown canopy
polygon.
J. Oldeland et al.
A recent, simplified version of the excessive greenness index (Rasmussen et al.
2016) was also calculated, in this study called exG2.
To exploit the additional NIR information, we calculated slope and distance
based vegetation indices (Silleos et al. 2006). Slope based vegetation indices make
use of the difference in the slope of the red and NIR channel; the famous Normalized
Difference Vegetation Index (NDVI, Tucker 1979) belongs here. The SAVI is a
modified NDVI which adjusts for potential effects of bare soil (Huete 1988).
Thiam’s vegetation index improves on the NDVI by multiplying the absolute NIR
and Red band values with their square root (Thiam 1998). Distance based vegetation indices make use of the concept of the so-called “soil-line” (Silleos et al. 2006).
The distances refer to the distance of samples in the two dimensional red-NIR spectral space to the soil line, that describes the lower boundary of pixels in this space,
usually aligning across a clearly visible axis. To determine the soil line parameters
required for the calculation of the distance based vegetation indices, a set of n = 100
bare soil pixels were selected, stratified by the hectare grid of the Biodiversity
Observatory, and the NIR and red values were extracted. Based on these values, a
linear regression (R
2
= 0.89, p < 0.001) was used to estimate the intercept and the
slope of the soil line. The linear regression parameters intercept (alpha = −227.29)
and slope (beta=1.877) were used to calculate the Perpendicular Vegetation Index
III (Qi et al. 1994; Silleos et al. 2006). All indices were calculated and image manipulations were performed with the open source software SAGA-GIS (Conrad et al.
2015). For all individual tree crown polygons, we calculated values for the mean
and standard deviations using the zonal statistics tool in SAGA-GIS.
Image Texture
Richards (2013) suggested that the texture of an image can be described as smooth,
rough or repetitive in terms of the spatial arrangement of grey values. In terms of
canopy cover this would describe whether tree crowns consist of repeating patterns
of shadow and greenness or whether the canopy is closed and thus equal in colour.
Often texture measures will improve remote sensing classifiers (Krefis et al. 2011).
As our main interest was to discriminate between tree species canopies, the greenness (exG) of the canopy seemed to be a good parameter for a texture analysis. We
used the Orfeo Toolbox v.5.6.1 (McInerney and Kempeneers 2015), a free open
source software for remote sensing image analysis, to calculate eight different types
of simple image texture measures. Haralick’s grey level occurrence matrix (GLCM,
Haralick et al. 1973), which is a standard for describing image texture, was the basis
for calculating all of the texture measures. We choose a constant window size of 5×5
pixels and an offset of 1 for x and y. The number of grey levels was set to 16. We
then loaded the calculated image texture measures into SAGA-GIS and extracted
the texture as average and standard deviation for each individual tree crown canopy
polygon.
J. Oldeland et al.