P
gap
than a frequency based estimate. As techniques for describ-
ing vertical structure are less well described,
COVVES
, a metric
that characterizes vertical canopy arrangement by estimating
the number of canopy layers from the second derivative of
P
gap
was also included (Wilkes
et al
., 2014). Figure 2 compares the
range of values for each of the four metrics at each study area.
To ascertain the benefit of increasing pulse density when
charactering forest structure, metrics derived using thinned da-
tasets were compared to a dataset with a common density of 10
pl m
-2
(Næsset, 2009). Mean values were computed from the nine
plotwise realizations at six pulse densities from which difference
and root mean square difference were calculated. Coefficient of
determination values were also calculated where the thinned da-
tasets were the independent variable, and the dependent variable
was the metric computed from the high pulse density dataset.
Furthermore, descriptive statistics for the nine plotwise realiza-
tions were calculated to ascertain intra-plot variability.
Results
Airborne Laser Scanning (
ALS
) data for three hundred plots
were extracted across six vegetation systems that characterize
diverse forested landscapes in Australia. For each plot, veg-
etation metrics representing three primary descriptors of veg-
etation structure (Kane
et al
., 2010; Lefsky
et al
., 2005) were
computed at to six different simulated pulse densities (0.05 to
4 pl m
-2
). Results for each descriptor are presented below.
Canopy Height
Differences in canopy height estimates are low for high pulse
densities and increase with decreasing sampling frequency
(Figure 4A). For example, at pulse densities
≥
0.5 pl m
-2
, Root
Mean Square Difference (
RMSD
) is <0.5 m for the
RF
,
CR
,
ZZ
, and
LI
study areas
,
<1 m at
WC
and <1.5 m at
RC
. At the
structurally simple and homogenous
RF
study area (Figure
2) where pulse density was simulated at 4 pl m
-2
,
RMSD
in
height estimation is less than the quoted vertical accuracy of
the bare earth ground surface as stated by the data provider.
When pulse density <0.5 pl m
-2
,
RMSD
is generally less than 2
m, however for the
RC
study area error can be >10 m (Figure
4A). Furthermore, at low pulse densities, differences in height
estimation is not necessarily a simple systematic offset; this
is highlighted by a decrease in coefficient of determination
values with decreasing pulse density (Figure 5A). As noted in
previous studies, canopy height estimates decrease with de-
creasing pulse density as a result of the incomplete sampling
of crown apexes (Goodwin
et al
., 2006; Morsdorf
et al
., 2008).
However, in this investigation the largest errors occurred as a
result of the misidentification of the ground surface; this re-
sulted in a poor
TIN
model from which to calculate vegetation
height (Figure 4E) (Takahashi
et al.
, 2008). This is particularly
evident at the
RC
study area where high vegetation density in
the upper canopy leads to attenuated penetration through the
canopy profile.
Intra-plot variance in canopy height estimation increases
with decreasing pulse density for all study areas (Figure 6A).
Standard deviation in canopy height estimates derived from
different realizations of the same dataset can exceed 4 m. Calcu-
lation of canopy height Coefficient of Variation (
C
v
) normalizes
for the large differences in canopy height when comparing plots
and study areas (Figure 2A). Variance in
C
v
of canopy height is
greatest for plots in the
LI
and
CR
study areas, particularly at
low pulse densities (Figure 6B). This is attributed to clumped
vegetation and low tree density where successive realizations
capture a significantly different proportion of vegetation.
Canopy Cover
Difference in canopy cover estimates as a function of sample
density are smaller when compared with canopy height. For
example the difference in cover estimates is close to zero for
all point densities across all areas with the exception being
RC
(Figure 4B). The larger error at
RC
at pulse densities <0.5
pl m
-2
is again attributed to the poor identification of ground
surface. This has shifted the height of the ground datum and
therefore altered the proportion of vegetation returns included
relative to the nominal height threshold. The trend is further
reflected in the low coefficient of determination values at the
RC
area (
R
2
<0.2) (Figure 5B). Replication of cover estimates
are also robust to diminishing pulse densities (Figure 6C),
where standard deviation of plotwise estimates is <10% of
total cover for 0.05 pl m
-2
.
Vertical Canopy Structure
Difference in
C
v
of return height are close to zero for all areas
when pulse density
≥
0.5 pl m
-2
. When pulse density decreases
to <0.5 pl m
-2
the difference becomes positive for the sparsely
vegetated areas and negative for other areas (Figure 4C). The
largest differences are seen at the
LI
and
CR
areas where
canopy cover is the lowest (Figure 2B). Here it is suggested
that large positive differences are caused by the under repre-
sentation of vegetation in the return height profile at lower
pulse densities. The opposite effect is seen to a lesser degree
at the
RC
study area where returns cluster towards the top of
the canopy. Analysis of
COVVES
values suggests the canopy
height profile is generally well represented when pulse den-
sity is
≥
0.5 pl m
-2
. This is reflected in an error of <1 canopy
layer across all areas (Figure 6D). Pulse density <0.5 pl m
-2
leads to an overestimation in number of canopy layers as lay-
ers appear increasingly fragmented; this trend becomes more
apparent with increasing canopy height (Figure 2).
Standard deviation of plotwise
C
v
of return height in-
creases with decreasing pulse density. Similar to canopy
height estimates, variance in
C
v
of return height is greatest for
plots with low canopy cover at pulse densities of <0.5 pl m
-2
(Figure 2B). The
C
v
of return height for plots where canopy
cover >20% have a relatively small standard deviation which
suggests the dispersion of returns through the canopy is
constant with each realization. Standard deviation of plotwise
COVVES
is also low for pulse densities >0.5 pl m
-2
where
intra-plot variation is <0.5 canopy layers (Figure 6E). In forest
types where vertical structure is relatively simple (Figure 2D),
standard deviation remains low for lower pulse densities.
Characteristics of Thinned Point Clouds
For the original datasets used in this investigation, mean pulse
density was ~22 pl m
-2
. Generating different realizations of
point clouds from an original dataset introduces the potential
for shared points between subsets. At pulse density of 4 pl m
-2
,
~85 percent of returns were shared between
≤
2 of 9 subsets.
This is suggested as the upper limit for simulation when apply-
ing this technique to a dataset with an original pulse density of
20 pl m
-2
. Simulations at a pulse density of 0.01 pl m
-2
were also
attempted. However, for forest types where canopy cover was
high, poor ground identification limited the number of success-
ful simulations (even within plots), and therefore these results
were disregarded. Pulse densities for thinned datasets were an
average of 10 percent less than the prescribed density, which
is a result of grid points with no returns within the search
window. This is caused by irregular pulse spacing which is
attributed to transmission losses (Korpela
et al
., 2012), oblique
viewing angles (Lovell
et al
., 2005) or gaps between flight line
overlap. An iterative approach could have been used to dy-
namically alter search window dimensions until the prescribed
pulse density was achieved (Næsset, 2009), although this
would have increased the risk of return duplication in subsets.
Examples of thinned point clouds for the six forest types
are presented in Figure 7, which illustrates the differences in
vegetation cover density and homogeneity apparent between
study areas. It is clear that the fidelity of canopy scale features
is reduced with decreasing pulse density particularly at the
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
August 2015
629
