The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
1214
According to Smet’s work [Smet et al., 2000], the efficiency of
rainfalling watershed segmentation is superior remarkably to
watershed algorithms based on immersion simulation. For this
reason, the rainfalling watershed algorithm is chosen to initially
segment an image into a set of so called sub feature units, which
are the initial parcels with smaller sizes.
With these sub feature units, the spectrums, shapes (area,
perimeter, etc.) and neighbourhood topology should be recorded
to serve the following merging processes. These are fulfilled
with Region adjacency graph (RAG) and nearest neighbour
graph (NNG).
Figure 1. RNG & NNG of a parcel graph
As illustrated in Figure 1, RAG is an undirected map which can
be expressed as G = (V, E), in which V={1,2is the set of
nodes, E is the set of links, and ^ v v . Every parcel is a
node of the map, and a link exists if two nodes are neighbours.
where, c is the band count, w c is user specified weights for
every band (1.0 by default).
The shape heterogeneity is the combination of compactness and
smoothness heterogeneity:
Given a specified RAG and its merging cost function S, its
corresponding NNG can be expressed as G m = (V m , E m ), where
V m is just similar to V in RAG, and E m only records the
minimal merging cost of every node, which indicates NNG is a
directed map. In particular, if the begin and end nodes of a link
are superposed, there exists a cycle. NNG improves the merging
efficiency than RAG because it obviously decreases the storage
and calculation of links.
2.2 MERGING CRITERION
h , = w . x h
shape cmpct cmpct
+ 0-w cmpct ) x K m o,hO)
in which compactness heterogeneity is calculated as:
h — u
cmpct Merge
Merge
~{n x
l
+ n,
/,
1 /— ' 2
* Merge v ^1
and smoothness heterogeneity is calculated as:
) (4)
Based on the sub feature units, a merge cost function integrating
spectral heterogeneity and shape heterogeneity is designed to
guide the merging of parcels. The use of shape is to make the
merged parcels more regular in shapes. With experiments, the
merging cost function is similar to [Baatz et al., 2007]:
/ = wx h color + (1 - w) X h shape (1)
In which w is the weight for spectral heterogeneity falling in the
interval [0, 1]. A generally suitable weight for colour is 0.9, and
0.1 for shape. Too large shape weight will bring unreasonable
segmentation.
h = tq
smooth Merge
fnc- (ni .L + „ 2 .L) (J)
V Merge 0 2
where / is the perimeter of a parcel, n is the pixels, b is the
perimeter of its bounding box. A commonly suitable setting of
Wcmpct IS 0.5.
The merged parcel variance can be got with Formula 6 to avoid
redundant calculation:
The spectral heterogeneity is the variance of the parent parcel
minus the sum of the variances of the two child parcels,
weighted with their respective areas:
h
color
Z W e (” Merge 17 ~ («1 f + n 2°l )) < 2 >
(7
Merge
((«, -1)0-, + (n, - l)<f) !(n merge -1)
+ n,n 2 (m,-m 2 f /(n Merge )(n Merge ~ 1)
(6)
where m\, m 2 are the means of the two child parcels.