, CA, 9-11 Nov. 1999
s obtained and geometrically
multispectral imagery. We then
its were correlated with laser
e a good hit coverage. Factors
ors are then discussed.
REGION
developed and tested for the
f Lake Duparquet (TRFLD), a
t in the Abitibi region, Quebec
part of the Forest Ecosystem
oreal forest is the largest forest
terrestrial land. Its floristic
indeed, only nine species are
landscape at the TRFLD is
od, softwood and mixed stands
years growing on a part of the
at 382 m. Common species
lus tremuloides), White spruce
etula paperifera), Balsam fir
'ked by spruce budworm, Jack
cedar (thuya occidentalis), and
he study site was commercially
generation areas. This test area
ape and habitat diversity,
cal forest, the wide availability
collaborative effort between
s and the socio-economic
QUISITION
'try data
as acquired on June 28th 1998
rument on a Piper Navajo plane
es were needed for vegetation
tion (single pass) respectively.
was carried out by the data
software. Flight, laser and GPS
Table 1 and a sample of the
| in Figure 1. The impulse
owest sustainable flight speed
achieve the desired laser hit
To alleviate this problem, each
n effort to double point density
However, since some first pass
ss hits, the effective hit density
oubled. This does not constitute
| because of the existence of
and the possibility of using à
ve consequences on this study's
good database to carry out à
accuracy (not presented in this
Multispectral imagery
The area covered by laser data had been surveyed the year
before (September 27th, 1997) using a Super VHS video
camera, functioning in zoom mode, mounted onboard an
airplane. The videotape was later converted to a series of
multispectral digital images having a 50 cm resolution. The
characteristics of the multispectral digital video survey are
presented in table 2. A sample of the digital images is presented
in figure 2. Tree growth between the end of September 1998
(multispectral data) and the end of the following month of June
(laser data) is minimal at these latitudes (48.48 degrees north),
allowing for direct comparison of the geometry of trees (size,
shape) between the two dates.
Ground measures
The height of individual trees was measured on the ground
using a standard clinometer and distance tape method. Two
measures were taken from different vantage points separated by
at least 90 degrees to insure independence between the two
measures. Trees for which the two height measures differed by
more than 3 meters or by more than 1596 were discarded so that
errors in comparing laser heights to acfual heights can mostly
be attributed to the laser methodology. These two heights for all
well measured trees were later used to assess the accuracy of
ground measurements. The study focused on two species:
Trembling Aspen (Populus tremuloides) and White Spruce
(Picea glauca) but some other species mentioned in section 2
were measured. Of the approximately 200 trees measured, only
40 had been positioned on the laser and multispectral imagery
at the time of paper submittal. After ground error filtering, 36
trees remained. Other ground measures include: diameter at
breast height, crown radius measured in the four cardinal
directions, species, and GPS positioning. A Corvallis Alto GPS
was used as a base station continuously downloading
differential data on a computer while a Corvallis MC-GPS
served as the mobile unit to position trees. Three hundred
epochs were obtained for each GPS point allowing for a
maximum error after differential correction that does not exceed
three meters for most trees.
4 METHODOLOGY
Creation of the Canopy Height model
The Canopy Height Model (CHM) was obtained by subtracting
the interpolated terrain altitudes from the interpolated canopy
altitudes. Triangulated irregular network (TIN) interpolation of
the X,Y,Z points was converted into a 50 cm resolution grid
using Surfer 6.04.
A TIN interpolation assumes that altitudes vary linearly
between points, a fact that is not necessarily observed in reality.
However, the choice of a different interpolation model would
have to rely on appropriate knowledge of the three-dimensional
International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 3W14, La Jolla, CA, 9-11 Nov. 1999
geometry of the canopy in the case of the vegetation, which can
vary with species and was not available. An outlook of future
research presented at the end of this paper discusses this matter
further. Subsequent work will compare interpolation
alternatives for the terrain hits in their ability to better tree
height predictions by providing a more accurate description of
terrain altitudes. The three stages of the creation of the CHM
are presented in Figure 1.
Multispectral image registration
The multispectral imagery was rectified to allow for accurate
overlay on the CHM. A minimum of five control points were
acquired for each 740 x 540 pixel multispectral images (approx.
370 m x 270 m) covering areas corresponding to ground tree
measures sites. These control points were obtained by visually
matching the center of small tree crowns that could be
recognized on the CHM and the corresponding crowns found in
the multispectral imagery. On both types of imagery, the
control points were positioned at the center of the crown. The
multispectral images were rectified by cubic convolution
resampling guided by a first order polynomial. The resulting
RMS residues of the polynomial are typically between one and
two meters. The fact that the raw multispectral images were
acquired in zoom mode (narrow view angle), yielding quasi-
orthographic images, helped in achieving low RMS. After
rectification, the multispectral images were overlaid on the
CHM allowing for 2D and 3D rendering (figure 2). A total of
40 images were rectified. All image processing operations were
carried out using ER-Mapper 5.5 software.
Establishing
correspondence
ground position to image position
Tree location on the imagery was established either by marking
the tree on a printout of the multispectral imagery based on
observations made during the ground survey, or by plotting a
differentially corrected GPS point on the imagery. The first
method is unpractical in that it asks for extreme care in visually
matching trees observed in the field to trees appearing on the
printouts. Trees near a lakefront can be identified after careful
examination, but trees a few dozen meters from the lakefront
and further can only be identified by progressing from one
outstanding tree to another, thus rendering walking in the forest
a very slow and tiring process. The other method of locating
trees used GPS technology. The X, Y coordinates of each tree
was plotted on the laser and rectified multispectral imagery.
The color of the tree (that relates to species) and the crown
diameters in the four cardinal directions (describing crown
shape and size) Were used as collateral data to insure that the
correct tree was identified by the GPS approach. In some
instances, a tree having the correct color and shape could not be
identified with certainty within a 5 meter radius of the GPS
point because two or more candidate trees with similar
characteristics were present in the vicinity of the plotted GPS
point. These trees were discarded from the study to avoid error.
