92
were found to match well with those in literature, the data were used directly instead
of a forward simulation of a LOP model (Feret et al. 2008). This was done to reduce
the number of parameters and associated uncertainties. The broadleaf species composition used for spectral mixing was derived from the forest inventory information
and is dominated by beech (about 50%), with lesser contributions from maple, elm,
linden, and ash.
One particular issue of the Laegeren site is its large variation in the spectral
background. Because we had multitemporal full-waveform lidar data available for
the Laegeren site, we used this information to classify the ground into distinct
classes (gravel, litter, soil) and assigned matching spectra from our field measurements to these classes (Leiterer et al. 2013). As Schneider et al. (2014) showed,
using several understory classes instead of a homogenous (black) background
makes simulated top-of-canopy (TOC) and top-of-atmosphere (TOA) reflectance
values more realistic.
4.4.1.2 3-D Reconstruction
Two different approaches for 3-D reconstruction of the vegetation structure were
implemented and tested. The first approach relied on a single-tree identification and
the second one on a direct computation of plant area index (PAI) values inside a
voxel cell. Voxels are basically 3-D pixels, dividing the 3-D space into equal-sized
cubes. The single-tree detection (individual tree crown, ITC) method used was
based on Morsdorf et al. (2004), which derives tree location, height, and crown
diameter to reconstruct the forest in 3-D based on simple geometric primitives like
rotational paraboloids. However, as with most local maxima detection-based ITC
methods, its performance within the mixed forest stands of the Laegeren site was
suboptimal, with tree detection rates of only 50–70%. This is much lower than what
can be expected for conifer forests, where rates of up to 90% can be achieved
(Kaartinen et al. 2012; Wang et al. 2016). Conifers generally have conical crowns
with one distinct peak (treetop), greatly facilitating their detection as local maxima
in a digital surface model (DSM). The main difference between the voxel-grid and
ITC approaches is the added level of semantics (Morsdorf et al. 2018) in the singletree case, which might be relevant for some species- and individual-focused experiments (i.e., when trying to link EO-based traits with genetic information of the
individual tree). If the aim of the 3-D reconstruction is an accurate simulation of the
radiative regime, single-tree identification adds a layer of unnecessary complexity,
so the voxel-grid approach led to better results (Schneider et al. 2014) and was subsequently used for upscaling of the trait information (Schneider et al. 2017).
4.4.1.3 Linking Field and RS Data
The perspective of forest inventory is from within or beneath the canopy and the
main sampling unit is the tree, quantified as diameter at breast height (DBH). RS,
on the other hand, has a top-down perspective on the canopy, and the sampling unit
F. Morsdorf et al.
were found to match well with those in literature, the data were used directly instead
of a forward simulation of a LOP model (Feret et al. 2008). This was done to reduce
the number of parameters and associated uncertainties. The broadleaf species composition used for spectral mixing was derived from the forest inventory information
and is dominated by beech (about 50%), with lesser contributions from maple, elm,
linden, and ash.
One particular issue of the Laegeren site is its large variation in the spectral
background. Because we had multitemporal full-waveform lidar data available for
the Laegeren site, we used this information to classify the ground into distinct
classes (gravel, litter, soil) and assigned matching spectra from our field measurements to these classes (Leiterer et al. 2013). As Schneider et al. (2014) showed,
using several understory classes instead of a homogenous (black) background
makes simulated top-of-canopy (TOC) and top-of-atmosphere (TOA) reflectance
values more realistic.
4.4.1.2 3-D Reconstruction
Two different approaches for 3-D reconstruction of the vegetation structure were
implemented and tested. The first approach relied on a single-tree identification and
the second one on a direct computation of plant area index (PAI) values inside a
voxel cell. Voxels are basically 3-D pixels, dividing the 3-D space into equal-sized
cubes. The single-tree detection (individual tree crown, ITC) method used was
based on Morsdorf et al. (2004), which derives tree location, height, and crown
diameter to reconstruct the forest in 3-D based on simple geometric primitives like
rotational paraboloids. However, as with most local maxima detection-based ITC
methods, its performance within the mixed forest stands of the Laegeren site was
suboptimal, with tree detection rates of only 50–70%. This is much lower than what
can be expected for conifer forests, where rates of up to 90% can be achieved
(Kaartinen et al. 2012; Wang et al. 2016). Conifers generally have conical crowns
with one distinct peak (treetop), greatly facilitating their detection as local maxima
in a digital surface model (DSM). The main difference between the voxel-grid and
ITC approaches is the added level of semantics (Morsdorf et al. 2018) in the singletree case, which might be relevant for some species- and individual-focused experiments (i.e., when trying to link EO-based traits with genetic information of the
individual tree). If the aim of the 3-D reconstruction is an accurate simulation of the
radiative regime, single-tree identification adds a layer of unnecessary complexity,
so the voxel-grid approach led to better results (Schneider et al. 2014) and was subsequently used for upscaling of the trait information (Schneider et al. 2017).
4.4.1.3 Linking Field and RS Data
The perspective of forest inventory is from within or beneath the canopy and the
main sampling unit is the tree, quantified as diameter at breast height (DBH). RS,
on the other hand, has a top-down perspective on the canopy, and the sampling unit
F. Morsdorf et al.