a, CA, 9-11 Nov. 1999
> to find the difference between
and the underlying interpolated
rve the preciseness of raw laser
nitation, we chose the simpler
vements in the near future will
data for vegetation and using a
round laser points.
pectral videography (initially in
) the laser altimetry images
) overlay on the Canopy Height
SULTS
ned between ground-measured
ight on. The mean of the two
Id for these trees was regressed
ht read from the CHM for 36
wood). The linear model yielded
-0.01). The scatter of points in
linear trends, an observation that
iat the linear model fit gave the
International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 3W14, La Jolla, CA, 9-11 Nov. 1999
best results. The predictive model (true height from laser
height) is given by:
Mean ground truth height = 4.24 + 0.91 * laser height (1)
30
25
n
na
u
20 d Ha
£
*2 n
= n n
= 15 Eon
E Pa
o po
5 :
10 a
n a
o na?
5 uir f
0
0 5 10 15 20 25 30
Mean ground-truth height (m)
Figure 3 - Comparison of mean ground-truth heights with laser
heights.
Error assessment
Tree heights predicted by the linear regression model in Figure
3 were compared to the mean of the two ground measurements,
yielding an absolute difference of 1.42 m with a standard
deviation of 1.15 m. The average relative error is 11 % with a
standard deviation of 9 %. The average difference between the
two eround measurements for each tree is 1.52 m (SD), and the
relative error (the absolute difference between the two ground
measurements divided by the average of the two measurements)
is 10% (SD = 8%). The data thus suggests that after the laser
absolute measurements of tree height are corrected by a linear
model, the laser prediction have an accuracy comparable to that
of ground measurements. Caution must be used in interpreting
these results since the number of trees is relatively small. Also,
“truc” tree height, or tree height measured with, say, a
centimetric accuracy is unknown. Rather, the accuracy of the
laser predicted heights is evaluated from error—ridden ground-
truth data. The R? of equation 1 would probably be higher if the
error associated with ground measurements could be reduced. A
manual photogrammetic approach will be tested in trying to
alleviate this problem. Moreover, the comparison of the level of
laser-predicted height errors relative to the level of ground
errors is only indicative since the error source cannot be
established with certainty. If most of the error was eventually
attributed to ground measures, and R? proven to be higher, the
use of laser predictions could then be considered as being of
higher accuracy than ground measures.
The main factors determining accuracy of raw laser
measurements, and thus of the prediction of tree height are
believed to be laser spot density relative to crown surface and
laser penetration in vegetation. Indeed, the correlation between
Crown surface (obtained by modeling the crown as a circle and
by using the average of the four crown radii measured on the
ground) and relative laser tree height prediction error is —0.76.
This indicates that the probability of a laser hit falling directly
on the tree top (maximum height) is directly proportional to the
crown projected surface. However, even trees that were well
covered by laser hits, such as large hardwoods, gave a raw laser
height estimate that was systematically lower than ground
measured height, i.e. often two to three meters less than true
height. Light penetration in the first layers of leaves or needles
is the most probable explanation for the fact that laser raw
heights are systematically lower than ground heights. For
example, when the height of a conifer is measured, the top of
the tree is defined as the tip of the highest branch. The volume
of needles and branches in the tip of the tree is very low and
does not constitute a very good interceptor of light, even in the
case of a very small footprint falling exactly on the center of the
tree. This also applies to the somewhat ragged tree top surface
of hardwoods. The first return from the top of the tree probably
comes from a level lower than the height that is measured on
the ground. Moreover, our study area is located in an old forest.
The forest floor is often covered by dense regeneration growing
across numerous debris such as fallen branches or trees,
boulders and rotten tree trunks. Even in the field, the ground
level is in some cases difficult to see. Therefore, it is safe to
assume that the last return could come from objects above true
ground level.
The shape of the crown, determined by species, can
theoretically influence the probability that the laser beam hits at
a level close to the maximum tree height simply because the
hardwoods have a much more rounded crown top. However,
when absolute and relative error are compared by hardwood
versus softwood categories using ANOVA, no statistically
significant differences are revealed. This conclusion must be
interpreted with caution since only 8 hardwood trees and 14
softwood trees were used in this analysis. Finally, the TIN
interpolation and the conversion from a TIN to a grid could also
induce some errors. Indeed, the height of a grid cell normally
corresponds to that of the center point of the grid cell projected
on the TIN rather than the maximum TIN height observed
inside the grid cell. Since the apex of trees defines a convex
shape, the center point of any grid cell covering the apex of the
tree will most probably be projected at a lower level than the
maximum height observed in that cell.
6 CONCLUSION
By comparing tree height measured on the ground with tree
height measured by a small foot print high density airborne
lidar, we observe that the accuracy of laser measurements show
an accuracy comparable to that of ground measurements. The
errors can be mostly attributed to the size of the crown since
larger crown have a higher probability of being hit at a level
close to that of maximum tree height. Many improvements are
considered: increasing the number of tree in the sample, using a
better algorithm for interpolating ground points and preserving
the raw first return altitude instead of using the gridded version.
Also, we are looking at fitting three dimensional models of tree
crowns (ellipsoid, cones and bullet shaped solids) to laser first
returns to try to better capture maximum tree height and to
