In: Wagner W., Szekely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B
Multi Keioluitan 6«gm*n(atttl> (iplimi/ilton t IS
Open:
Iteration
Ittrstion: fT* NtimSubobj::? Scale: jin'
Target Object Inlormation
Segmentation Parameter«
Smoothness Weight: U.'j
Shape: In. i
Compaetnes« Weight: ¡0.6
TexMre: ¡21.07
RetSangolar Fit;
[.9349
Stability: ¡73.17 Brightness: j?14.21
Compactness: [l.28«i
Ares: |Si2~“
Subojecl*
1 Texture ! Stshility
I QrfaMft***
t
21.41
44.22
193.51
14
2
8.641
24.4
190.24
137
3
9,021
22.71
208.96
313
4
9.473
36.72
234.81
143
5
4.644
16.04
226.62
115
6
19.72
80.46
241.46
66
7
8.SB/
36.95
224.43
34
Select Folder., >
Help
Figure 4. Interface of FbSP optimizer. The Segmentation
Parameters, Target Object Information and Sub Object
Information are inputted into the system to train FbSP
optimizer to estimate the optimal Segmentation Parameters for
the Target Object through fuzzy logic analyses.
In Figure 4, the Segmentation Parameters (Scale, Shape, and
Smoothness weight) for performing a preliminary over
segmentation are initially selected and inputted by user. They
are then updated by the FbSP optimizer in the next iteration
according to the information of the target object ant its sub
objects. The Texture, Stability, Brightness and Area for the
target object and sub objects are calculated according to the
pixel grey values within the target object and individual sub
objects, respectively. The Compactness and Rectangular Fit are
calculated according to the shape of the target object.
4. EXPERIMENT AND RESULTS COMPARISION
4.1. Data Sets
Pan-sharpened QuickBird MS image and pan-sharpened Ikonos
MS image over Fredericton, Canada, and pan-sharpened
QuickBird MS image of Oromocto, Canada, were used to test
the FbSP optimizer. The UNB-Pansharp was used to fuse the
Pan and MS images. The four pan-sharpened multispectral
bands were used as input bands.
4.2. Segmentation Process and Results
For small objects, the FbSP optimizer can find optimal
segmentation parameters through one or two iteration(s) of
segmentation parameter estimation (Figure 3 and Figure 5). For
example, for small buildings shown in Figure 5, the FbSP
optimizer used just one iteration to find the optimal
segmentation parameters for the object of interest—small
buildings. If the initial parameter selection by the user for the
initial over segmentation is counted as one iteration, two
iterations of parameter selection in total were needed, one by
the user and one by the FbSP optimizer (Figure 5.a and 5.c, and
Table 1).
Table 1 shows the segmentation parameters selected by the user
for initial over segmentation (Iteration 1), and the parameters
estimated by the FbSP optimizer in the first loop (Iteration 2).
Table 2 shows the feature information of the sub objects
(Figure 5.a, red) selected to form a target object. Table 3 lists
the feature information of the target object (Figure 5.b, red).
(a)
JEL
(£)
■SR
Figure 5. Segmentation process and final segmentation result
using the FbSP optimizer for small buildings (pan-sharpened
QuichBird MS, Oromocto). (a) Initial over segmentation using
user selected initial segmentation parameters (first iteration)
and the sub objects (red) selected to form a target object, (b)
Target object formed by the sub objects for training the FbSP
optimizer, (c) Segmentation result achieved using the
parameters estimated by FbSP optimizer in the first loop
(second iteration in total), (d) Final segmentation result using
the parameters estimated in (c).
According to the feature information in Table 2 and Table 3,
the FbSP optimizer estimated the optimal segmentation
parameters (Table 1, Iteration 2) for the target object. Using the
parameters in Iteration 2 of Table 1 to segment the input image,
the segmentation result shown in Figure 5.c was achieved. The
feature information (Table 4) of the resulting segment (Figure
5.c, red) is almost identical to that of the target object (Table 3
and Figure 5.b, red), i.e. the resulting segment converges with
the target object (Figure 3, step 11), so that the segmentation
parameters estimated by FbSP optimizer in the first loop
(Iteration 2 of Table 1) was accepted as the optimal
segmentation parameters for small buildings.
For large objects, more sub objects need to be selected to form
a target object, so that more iterations are usually needed to
reach the convergence between the target object and the
resulting segment. For example, four iterations were needed to
reach the convergence for large buildings shown in Figure 6.
4.3. Result Evaluation
Figure 7 shows the segmentation result of large buildings using
trial-and-error approach for segmentation parameter selection.
The result was achieved by a very experienced operator
through approximately two hours of parameter selection and
test. Comparing the result from FbSP optimizer (Figure 6) and
that from trial and error (Figure 7), we can see that the FbSP
693