96
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
Hence, the density of C given the covariance matrix £ is defined
by
MC| S ) = OCr-«exp(-n. t r(S-C)) (4)
| S |n )/2 Y\T(n-k+l)
k= 1
where | • | is the determinant and fr(-) is the trace of a matrix,
r( ) is the standard gamma function and q is the dimensionality
of s.
2.2 Wishart-based distance measure
An often used distance measure for polarimetric SAR data is
based on the Wishart distribution and is defined in (Lee et al.,
1997):
d w { C,£) = -ilnp(C|£) (5)
= In (|£|) + ME" 1 -C)+c (6)
where
In
Tj-gte- 1 )/ 2 ]TJ r(n — /c + 1)
k=1
(7)
The constant term c in (6) is class independent and can be omit
ted, if this distance is used as in (Lee et al., 1997) to decide if the
data point C more probably belongs to class ci represented by
the covariance matrix £i instead of belonging to class C2 repre
sented by £2. The distance measure simplifies to:
d w (C, £) = In ( I £ I ) + fr(£ _1 • C)
(8)
Although this distance measure is not a metric, because it is nei
ther homogeneous, nor symmetric and does not fullfill the tri
angle inequality, it is often used and has shown its effectiveness
in practice. Because of this and its direct relation to the density,
function it will be used in a slightly modified version in this wjork.
3 WEIGHTED PYRAMID LINKING
3.1 Pyramid construction and initialisation
The basic structure used in this approach is a multiresolution im
age pyramid. While the original data (fully-polarimetric SAR
data, multi-look complex covariance matrices) is forming the base,
the higher levels are versions of the image with subsequent re
duced resolutions. The height of the pyramid (the number of lev
els without the bottom level 1=0) shall be noted by L. Each pixel
is represented by a node in this pyramid. The value v\(n l (x, y))
of the node n at position (x, y) at level l is simply an average of
a window with certain size s at the previous level / — 1:
VZ E [1, L\ : vAn ) = V2(n l ) = - • v\(n)
s
n'€des(n*)
(9)
The difference between the covariance matrices vi (n l ) and V2(n l )
will be explained in section 3.3. Just note, that they are set to
the same value during the initialisation. Each node n at level l
(0 < l < L) is therefore connected with a set of nodes at level
l — 1, called descendents des{n) and a set of nodes in level l -f 1
called parents par(n). Nodes at the bottom level Z = 0 have
only parents, while nodes at the top level Z = L have only de
scendents. Only vertical connections between nodes at adjacent
levels and no horizontal relations between nodes of the same level
are used.
The windows overlap by a predefined amount o of pixels
(0 < o < s). The size s of the window and the overlap o define
the decrease in resolution of the next level. The parameters used
in this paper are a quadratic window size s of 4x4 = 16 and an
overlap o of two pixels in x- and y-direction. Given this setting
of o and s there will not be enough pixels at the border of a level
for a whole window, if the dimensions of this level are not even.
In that case, the level is simply extended with as many pixels as
needed. These additional pixels have the same value as the bor
der pixels. Due to this manipulation the border pixels gain greater
influence on the pixels at the next level. However, this effect is
insignificant as experiments have shown.
3.2 Weight adjustment
The most important part of this approach is the introduction of
link strengths w between nodes on adjacent levels of the pyra
mid. Instead of using only the descendent with the largest degree
of association, all descendents contribute accordingly to their link
strength to the node value at the next level. The link strength
w(n,n') between node n and its descendent n' E des(n) is
based on proximity and similarity:
w(n,n) = exp(-dspec(V2(,n),Vl(n)))
■ exp (—var(n))
•exp (-dspat(ri' ,n)
(10)
The first factor has the most crucial role. It measures the spec
tral distance between two nodes at adjacent levels in the image
pyramid. Any proper distance measure can be used here. As
mentioned above an often used distance measure for polarimet
ric SAR data is (8), which is based on the Wishart distribution.
Note, that the link strength is used to define the contribution of
a descendent to the value of the current node in comparison to
all other descendents of this node. Furthermore, the values of
neighbouring nodes at one level are unlikely to be equal and the
number of looks can be different, too. All nodes in the pyramid
at the same level will have the same number of looks merely after
the initialisation. That is why c in (6) cannot be omitted and (8)
cannot be used here. Therefore d S p e c(v2(n'), vi(n)) is defined
as:
dspec (V2(n'),vi(n)) - ~\np(v2(n')\vi(n))
(ID
where p(v2(n')\v\{ri)) is the density of the Wishart distribution
defined in (4).
Within the second factor the euclidian distance d S pat{n\n) be
tween the spatial positions of the two nodes is used. The spatial
distance within the 4 x 4 neighbourhood is defined as:
/ y/9 V5 V5 V9 \
\/5 vT vT Vb
s/h \/ï y/l \/5
V V9 Vò Vò V9 )
(12)
As the link strength now depends on geometric closeness the re
gions tend to be more compact, whereas they would have more
irregular shapes without this factor.
The third factor represents the variability of the descendent of
node n. Since the goal is to segment the image into homoge
neous regions, nodes that represent segments with high variabil
ity should get a lower link strength than nodes representing more