ysis have
Baddeley
Rue and
ve of this
e various
vledge of
yr models
es tend to
n object
Lieshout
sist of an
arbitrary
trary set,
in U, and
e that is
s a finite
ective in
ed) true
image y
ut (1995)
rministic
ic effects
age, y.
our prior
id can be
s that we
umber of
(MAP)
function
:d to this
o called
f three-
1 applied
oach can
1ensional
t defined
in space
> LIDAR
rector, 7,
LIDAR
ved data,
ectors at
ceded a
ISPRS Commission III, Vol.34, Part 3A ,,Photogrammetric Computer Vision", Graz, 2002
Figure 1. LIDAR sensing geometry (red stars represent LIDAR
measurements; blue lines represent pulse vectors
composing the scan space T; y, represents a single
range measurement along pulse vector f).
The distribution of tree crowns over the entire scene is then
modeled as an object configuration, x. If individual plants were
actually solid objects (e.g. ellipsoids, spheres, etc.) in object
space U, these LIDAR measurements, y, would represent the
location where each vector f intercepted the surface of the
object. In the terminology introduced above, these
measurements would represent the signal, or the deterministic
influence of the actual configuration of objects x on the series of
LIDAR range measurements that are observed. A more realistic
approach, however, would need to account for the fact that
plants are not solid geometric objects, and LIDAR pulses
actually penetrate a certain distance into the canopy through
foliage gaps.
This would incorporate a stochastic element to the LIDAR
measurements, y, due to the irregular spatial distribution of
foliage elements (leaves, branches, etc.) in the pathway of a
laser pulse as it intersects a tree crown. Therefore, again using
the notation introduced above, the conditional distribution of
individual LIDAR measurements, given the signal, is given by a
family of densities g(y |x). In the context of LIDAR
measurement of tree crowns, these probability densities will be
related to the laser attenuation function, which in the case of
discrete LIDAR systems is directly related to the probability of
reflection. If the values of individual LIDAR measurements
along a pulse vector can be considered conditionally
independent, given the true configuration of objects, x, the
likelihood function, representing the joint probability of the
data, is given by:
I(y|x) » [26,12 Q)
teT
3.3.1 Modelling the distribution of foliage in complex forest
scenes: Previous studies of laser transmission through the
forest canopy have utilized three-dimensional grid models
populated with generalized geometric forms that represent
individual plants (Sun and Ranson, 2000).
In this study, a three-dimensional array, with 0.91-m cell size,
was used to model the distribution of foliage density throughout
a forested area. The spatial distribution of foliage is a function
of individual tree locations, sizes, crown forms, and an average
leaf area density (LAD). Crown forms were represented as
generalized ellipsoids following Sheng et al.. (2001), where the
space occupied by the foliage within an individual tree crown is
determined by four parameters: crown width (cw), crown height
(ch), crown curvature (cc), tree height (hf), and the 2-D
coordinate of the crown top (X, Y; ) (see Figure 2).
Ap
a) ht:47, cw:11, ch:14, cc:1.45 b) At:30, cw:8, ch:18, cc:0.75
Figure 2. Generalized ellipsoid crown models.
The surface of a tree crown is then given by the following
mathematical expression:
cc/2
(Z+ch+2,, Y. i (de. +(Y-%,, ’)
eh ert
zd. (3)
where the elevation of the crown top, Z,,, is determined by
adding the tree height to the elevation of the base as determined
from the DTM. Values for LAD were obtained from previous
research findings (Webb and Ungs, 1993).
3.3.2 Modelling laser-canopy foliage interaction: The
analysis of data acquired from active remote sensing
technologies requires an understanding of the interactions
between the emitted radiation and the physical properties of the
target.
In our model, where the laser footprint (0.4 m) is significantly
less than the cell size (0.91 m), the probability that a direct light
beam that enters a cell exits from the cell without being
intercepted is calculated as a function of the leaf area density
and the off-nadir angle (0) of the laser pulse. The model then
calculates the probability of penetration to the center of any cell
by direct laser energy originating from outside the cell.
Specifically, the probability of a laser beam that enters a canopy
cell, z;, at a specific off-nadir angle, 6,, reflects from this cell
with foliage density LAD; is given by the following function
(Vanderbilt, 1990):
p(reflection) = [1 —exp-1/cos0 - LAD, -G(0,,2;)- dz) | (4)
where LAD; is the leaf area density (m°/ m?) within cell z;, G(0,
Z;) is the projection of the vegetation within cell z; in the
direction of 0, and dz; is the depth of cell z;. Often, a spherical
leaf angle distribution can be assumed, in which case G(6, z;) is