RMSE
was lower in forest clearings than under mature
canopy for both
NCALM
and
WMNF
datasets. This result is ex-
pected, as the density of lidar ground returns used during
DEM
interpolation should increase as vegetative cover decreases.
Higher lidar ground return density facilitates the interpola-
tion of more accurate
DEMs
, and thus clearings are presumed
to allow more accurate representations of the ground surface
than regions with a canopy and therefore exhibit a lower
RMSE
. Greater
RMSE
for RG5 compared with other survey
sites may have been caused by dense beech saplings. Young
beech thickets have a higher leaf-area index (LAI) (Genet
et
al.
, 2009) and perhaps a more complex branch architecture
than mature stands. Limbs can cause lidar pulses to reflect
in a direction other than to the sensor resulting in spurious
data points known as multipath errors. Furthermore, when a
pulse encounters an object and records a return, there is a lag
distance of approximately three vertical meters before another
return can be recorded. Hobblebush and other herbaceous
understory plants found in WS3 growing to about one meter
in height did not appear to interfere with lidar accuracy. How-
ever, the midstory beech thicket likely increased multipath
errors and decreased the number of recorded ground returns,
resulting in reduced accuracy and/or misclassification of
off-terrain features as ground. Overall, our results suggest that
DEM
elevation error under a mature forest canopy of northern
hardwoods in leaf-off conditions and in steep terrain may
include as much 0.3 m to 0.5 m more error in elevation com-
pared to open areas.
The two lidar datasets had different ground return densities
in WS3; the
NCALM
dataset had a density of 3.27 ppsm while
the
WMNF
dataset had a density of 1.16 ppsm (Table 1). Yet the
datasets generated similar
DEMs
in terms of raw elevations, and
the
DEMs
produced similar catchment boundaries, especially
when coarsened to 3 m and 5 m resolutions. Furthermore, dis-
tributions of topographic metrics computed from filtered/un-
filtered coarser
DEMs
derived from the original 1 m resolution
for each dataset were similar (Table 2; Figure 4). Density of
nominal post-spacing has been shown to affect
DEM
accuracy
(e.g., Aguilar
et al.
, 2005; Hodgson
et al.
, 2004; Hu
et al.
, 2009)
although several studies also indicated that in some situations
lidar sampling density can be reduced by up to 50 percent
with no degradation of
DEM
accuracy (Anderson
et al.
, 2006;
Liu
et al.
, 2007). Denser post-spacing can be achieved using
a higher pulse rate, lower altitude over-flight, narrower scan
angle, or some combination of these variables (Raber
et al
.,
2007). The
WMNF
dataset contained lower nominal post-spac-
ing and was collected from a higher altitude than the
NCALM
dataset. Similarity of
DEM
elevations, computed topographic
metrics, and delineated catchment boundaries for each dataset
investigated in this study suggest that post-spacing, within a
modest range of variation, may not be the most limiting factor
to the quality of
DEMs
for such watershed studies.
Topographic Metric Variation with DEM Resolution
TWI
distributions tended to increase with
DEM
coarsening and
filtering, consistent with previous investigations of coarser
(10 m and greater)
DEMs
(e.g., Quinn
et al
., 1991; Sorensen and
Seibert, 2007; Wolock and Price, 1994) but extended to a finer
scale in this study. Observed increases in median
UAA
values
with
DEM
coarsening (Figure 4) area also was consistent with
previous observations (Zhang and Montgomery, 1994). The
large differences in mean
TWI
values (Table 2) computed from
unfiltered 1 m and filtered 10 m
DEMs
derived from
NCALM
and
WMNF
lidar datasets, respectively, are significant from the per-
spective of the hydrological modeler. This disparity is equiva-
lent to approximately two orders of magnitude difference in
simulated subsurface flow when following TOPMODEL theory
(Beven and Kirkby, 1979) and using the hydraulic conductivity
distribution with depth observed for WS3 (Detty and McGuire,
2010). The parameterization of TOPMODEL and other hydro-
logic models is often independent of topography where it is
static and taken directly from available elevation data. How-
ever, small differences (less than 10 m) in
DEM
resolution have
implications for computed topographic metric values (e.g.,
the mean) that affect hydrologic quantities derived from such
models.
Increasing
TWI
values with
DEM
coarsening appeared to be
controlled by increases in
UAA
values. Minimum slope values
increased and mean/median slope values decreased only
slightly (Figure 2; Figure 4) with
DEM
coarsening and therefore
cannot explain the observed increases in
TWI
values. How-
ever,
UAA
boxplots (Figure 4) and
UAA
maps (Plate 1) demon-
strated increasing values with
DEM
coarsening. Mean values
varied over an order of magnitude from a low of 161 m
2
for
the
NCALM
filtered 1 m
DEM
to a high value of 3,411 m
2
for the
NCALM
filtered 10 m
DEM
, and must therefore offset observed
changes in minimum, mean, and median slope values when
calculating
TWI
. Maximum slope value decreases, in addition
to
UAA
increases, may at least partially explain
TWI
maximum
value increases with
DEM
coarsening.
UAA
values computed from the 3 m and 5 m
DEMs
best dif-
ferentiated topographic variation seen on the catchment map
with a 3 m contour interval (Plate 1). Summits and convex
shoulder slopes aligned with smallest
UAA
values while
drainages and toeslopes aligned with greatest
UAA
values, and
backslopes aligned with moderate
UAA
values. In contrast,
the 1 m
DEMs
generated small
UAA
values throughout much
of the catchment, with much of the landscape in a variety of
topographic positions having a
UAA
of <100 m
2
(Plate 1) This
may be due to fine topographic detail and surface irregulari-
ties, which our total station survey suggests are not well
represented, that interfere with the computation of
UAA
. On
the other hand, the 10 m
DEMs
generated
UAA
values >1,000
m
2
throughout many parts of the catchment, including some
areas best described as planar backslopes (Plate 1).
An example of the implications of variation in topographic
metrics derived from varying
DEM
scale in our study area is
provided by Bailey
et al
. (2014) who found that variation soil
horizon sequences and thickness, and groundwater fluctua-
tions were best correlated with
TWI
and
UAA
values derived
from 3 m to 5 m
DEM
resolution. Five soil map units were de-
lineated based on these hydropedological variations and oc-
curred at hillslope positions that can be predicted by interpre-
tation of the 3 m contour interval topographic map. Further
investigation of topographic metrics calculated from
DEMs
at
this scale is warranted as a digital soil mapping tool in this re-
gion and highlights the importance of careful consideration of
DEM
resolution used to compute topographic metrics, as small
resolution differences can yield dramatically different results
in metric values for a given point on a landscape.
Conclusions
Total station surveys suggested that small scale terrain fea-
tures such as boulders and fallen trees are smoothed when a
DEM
is generated from lidar data, and that
DEM
elevation ac-
curacy increases in the absence of vegetative cover. However,
even under mature forest canopy and in rough terrain, lidar
still can produce
DEMs
useful for soil and hydrologic analyses.
The similarities we observed in watershed boundaries, topo-
graphic metrics, and
RMSE
values computed from each set of
DEMs
suggest that differences in lidar collection methods and
ground return densities we studied were not sufficient to cre-
ate tangible
DEM
accuracy differences. Methods for increasing
accuracy also increase the cost of lidar acquisition and data
processing/storage requirements. Our findings suggest that
these costs can be reduced while generating
DEMs
as accurate
as those developed with greater monetary and time inputs.
394
May 2015
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING