error in metric estimation increases with decreasing pulse
density (Takahashi
et al
., 2010); this is illustrated by a simul-
taneous decrease in coefficient of determination values with
pulse density (Figure 5).
Systematically applying an offset to the point-grid al-
lowed robust estimates of vegetation metrics independent
of sampling location origin. This is particularly useful when
analyzing low pulse densities where large variation is evident
between realizations (Figure 6). Analysis of intra-plot results
suggests that variance in metrics is captured at pulse densities
≥
0.5 pl m
-2
. For lower pulse densities, difference between real-
izations can be more significant. For example at
WC
difference
in height estimations between realizations have a standard
deviation ~3 m at the lowest pulse density, this drops to <1 m
at a pulse density of 0.5 pl m
-2
(Figure 6A). Intra-plot variance
was lowest for plots where canopy cover was homogenous
and structurally simple, e.g., a single canopy (Figures 2 and
7). For example standard deviation of canopy height remained
<1 m for the
RF
study area, even when considering the lowest
pulse density simulated (Figure 6A). Conversely at the sparse-
ly vegetated
LI
study area where vegetation is clumped, the
C
v
of canopy height and
C
v
of return height is highly variable
(Figure 6B and 6D, respectively). This is due to vegetation
cover being heterogeneous and therefore subsequent plotwise
realizations capture significantly different proportions of veg-
etation, resulting in dissimilar interpretations of the canopy
profile. When comparing four overlapping flight lines ac-
quired on the same day with the same capture specifications,
Bater
et al
. (2011) found a similar difference between height
estimates of <1 m at a pulse density of 2 pl m
-2
. The authors
also noted significant differences in metrics calculated from
last-returns, e.g., describing the ground and lower region of
the canopy. This is again less apparent in this study owing to
the utilization of a multi-return recording instrument.
The technique introduced in this paper has allowed a
robust comparison of estimating vegetation structure metrics
simulating different capture pulse densities across a range
of vegetation types. The inclusion of a
TIN
model calculation
is seen as an important step that is not always included in
thinning simulations.
TIN
modeling highlighted that utilizing
a pulse density <0.5 pl m
-2
in dense forests is unlikely to iden-
tify the ground surface adequately to accurately determine
Figure 7. Example point clouds for each forest type displayed at four different pulse densities. Points are classified into either ground
(black) or non-ground (grey), e.g., vegetation. Heights are relative to the Australian Height Datum. Point size is shown to increase with
decreasing density for visualization purposes only.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
August 2015
633
