In: Wagner W., Székely, В. (eds.): ISPRS ТС VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
DEVELOPMENT OF A SUPERVISED SOFTWARE TOOL FOR AUTOMATED 
DETERMINATION OF OPTIMAL SEGMENTATION PARAMETERS FOR 
ECOGNITION 
Y. Zhang* a , T. Maxwell, H. Tong, V. Dey 
“University of New Brunswick, Geodesy & Geomatics Eng., 15 Dineen Dr., E3B 5A3, Fredericton, Canada 
Technical Commission VII Symposium 2010 
KEY WORDS: Land Cover, Classification, Abstraction, Image, Software, High resolution, Advancement 
ABSTRACT: 
Image segmentation is one of the most important steps in object-based classification. The commercial software eCognition has been 
proven to be the most advanced software tool for object-based classification of high resolution remote sensing imagery. However, its 
segmentation process still relies on trial and error to find proper segmentation parameters. The segmentation process is very time 
consuming and the segmentation quality directly depends on the experience of the operator. To overcome this problem, a supervised 
software tool—Fuzzy-based Segmentation Parameter optimizer (FbSP optimizer)—was developed to determine the optimal 
segmentation parameters through a training process and a fuzzy logic analysis. The optimal segmentation parameters are then used in 
eCognition to segment the entire image, achieving an optimal segmentation result. The FbSP optimizer can radically increases the 
efficiency of segmentation parameter selection, and achieve improved segmentation results. It also reduces the influence of the 
operator’s experience on the quality of segmentation results. 
1. INTRODUCTION 
Since the successful launch of the very high resolution (VHR) 
Ikonon satellite in 1999, object-based classification has quickly 
become the mainstream technology for land cover classification 
of VHR remote sensing images, such as Ikonos, QuickBird, 
GeoEye-1, WorldView-2 and airborne digital imagery (Smith 
and Morton, 2010; Blaschke, 2010). In object-based 
classification, image segmentation is a crucial process which 
directly influences the efficiency of the classification process 
and quality of the classification result. To date, eCognition 
software developed by Definiens has proven to be the most 
effective technique for object-based classification among a 
variety of object-based classification techniques (Lavigne et al., 
2006). 
However, trial and error is still a standard approach of 
eCognition to finding proper segmentation parameters for 
achieving a proper segmentation of objects of interest. In the 
segmentation, operator’s knowledge of the image and 
experience of the segmentation process play an important role 
for the success of the segmentation. In addition, the 
segmentation process is time consuming. These drawbacks 
have significantly limited the potential of eCognition for a 
broad range of practical applications. 
To overcome the limitation of eCognition in finding proper 
segmentation parameters for image segmentation, a software 
tool has been developed in the CRC-AGIP Lab (Canada 
Research Chair Laboratory in Advanced Geomatics Image 
Processing) at the University of New Brunswick, based on 
previous work done in the lab (Maxwell, 2005; Zhang and 
Maxwell, 2006). The software tool, named Fuzzy-based 
Segmentation Parameter optimizer (FbSP optimizer), can 
automatically determine optimal segmentation parameters for 
eCognition through a supervised training process and fuzzy 
logic analysis. Using the FbSP optimizer in combination with 
eCognition, the segmentation of an object of interest can be 
achieved within minutes, instead of hours by solely using 
eCognition. In addition, the segmentation result can be 
significantly improved. 
This paper will first introduce the general concept of image 
segmentation used in eCognition and the role of segmentation 
parameters. It will then introduce the concept and process of 
the developed of the FbSP optimizer for identifying optimal 
segmentation parameters for eCognition. The experiment 
results and the comparison between the segmentation qualities 
and time used in the segmentation processes will also be given 
to allow readers to judge the improvement made by the 
supervised software tool— FbSP optimizer. 
2. SEGMENTATION TECHNIQUE OF ECOGNITION 
2.1. Region Merging 
To find the boundary of an image object or segment an object, 
eCognition implemented a region merging approach to 
segmentation called “Fractal Net Evolution” approach (Baatz 
and Schape, 1999). This technique starts with individual 
adjacent pixels as initial objects, and then measures (1) the 
spectral heterogeneity change, h spectrah and (2) the shape 
heterogeneity change, h shape , between the two neighbor pixels 
(objects) to determine whether they need to be merged 
together, or not. Once the two pixels are merged into one 
object, the region of the object grows one step. This 
measurement and merging process continues iteratively until a 
user defined threshold is reached. Then, the region of the object 
stops growing; resulting in one image segment. The region 
merging and region growing process was designed with the 
view to meeting six aims including the (Baatz and Schape, 
2000): 
a. Production of homogeneous image object-primitives; 
b. Adaptability to different scales; 
c. Production of similar segment sizes for a chosen scale; 
d. Applicability to a variety of data sets; 
e. Reproducibility of segmentation results; and 
f. Requirement for reasonably fast performance. 
2.2. Role of Segmentation Parameters 
Figure 1 illustrates the relationship between spectral 
heterogeneity change, h spectra i, the shape heterogeneity change, 
h sh ape, and the corresponding segmentation parameters. Where 
■ h c spectral is spectral heterogeneity change of individual 
690