In: Wagner W., Székely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
692 
2003; Hay et al., 2003; Maxwell, 2005; Li et al., 2009; Smith 
and Morton, 2010). 
3. SUPERVISED DETERMINATION OF 
SEGMENTATION PARAMETERS 
3.1. Design of the Segmentation Tool 
To address the issues existing in the trial-and-error selection of 
segmentation parameters, a tool for supervised segmentation 
parameter determination should meet the following 
requirements (Maxwell, 2005): 
a. Each execution of the tool is aimed at extracting one land 
cover type and results in one level of the object 
hierarchy; 
b. Segmentation must be controlled and refined in an 
iterative manner based on an object model; 
c. The tool must rely on an initial segmentation as a start 
state; 
d. Scale, shape, and smoothness parameters must be 
determined; 
e. Parameter selection must be reproducible; and 
f. The tool must demonstrate reasonably fast and efficient 
performance. 
The segmentation of an input image is performed on a number 
of different levels to permit objects of different scales to be 
extracted on their own level. By using this approach, objects 
can be classified on the level where the segments are the most 
meaningful and best represent the object of interest. This infers 
that the user must have a specific land cover class in mind 
when segmenting the image so that the parameters can be best 
estimated and then refined through iteration. As a result, the 
tool must aim to extract one particular land cover type each 
time it is executed. By running the tool a number of times, a 
hierarchy of object levels can then be developed. 
3.2. Workflow of FbSP Optimizer 
To meet the design requirements of the software tool, the 
workflow for the supervised fuzzy-based determination of 
segmentation parameters, i.e. the FbSP optimizer, was 
developed as shown in Figure 3. To train the FbSP optimizer, 
the input image needs to be initially segmented achieving an 
over segmentation, i.e. the segments are smaller than the 
objects of interest (see Figure 5.a). The small segments, also 
called sub objects, can be selected to form a meaningful target 
segment/object. The information of the target object and its sub 
objects is then used to train the FbSP optimizer to determine 
the optimal segmentation parameters for the target object 
(Figure 4). 
Figure 3. Workflow of the proposed FbSP optimizer. The values of the current Segmentation Parameters {Smoothness ( 1 -w compact ), 
Shape (1-w) and Scale (s)), Sub objects information (Texture, Stability, Brightness, and Area) and Target Object information 
(Texture, Stability, Brightness, Area, Rectangular Fit, and Compactness) are inputted into FbSP optimizer to train the FISs (Fuzzy 
Inference Systems) to estimate the optimal Segmentation Parameters (1 -w compacl , 1-w, and 5) for the Target Object in an iterative 
process.