In: Wagner W., Szfltely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B
spectral bands,
hcompact is compactness heterogeneity change,
^smooth is smoothness heterogeneity change,
w c are the weights associated with each layer,
wcompact is compactness weight (parameter),
1 -wcompact is smoothness weight (parameter),
w is weight (parameter) for overall spectral heterogeneity
change, and
l-w is weight (parameter) for shape heterogeneity
change.
The h c spectral , h compact , and h smoolh are calculated according to the
image grey values within the two neighbor objects in each
spectral bands, whereas the weighs (parameters) w c , w compact (or
1 -wcompact), an d w (or 1 -w) must be given by the user. The user
must also give a scale value (s) as a threshold to stop the
merging. Figure 2 shows the interface of eCognition to allow
users to input the parameters to guide and control the
segmentation process.
h spectral
'• spectral
h c
n spectral
h compact
hsmooth
Raster leaver 1
Raster Layer 2
Raster Layer c
Figure 1. Relationship between the segmentation parameters (user determined weights) in eCognition. Usually, the weights for
individual spectral layers (bands) (wi, w 2 , ... w c ) are set to 1. Users need to give the value for Smoothness weight (1 -w compact ) (or
Compactness weight (w compacl )) and Shape weight (l-w). The weights (1 -w compac t) and (l-w) are used to calculate the Fusion Value
(/). The value / is then compared with a user specified Scale value (5) to estimate whether the two adjacent objects need to be
merged, or not (iff < s 2 , merge the two objects; iff> s 2 , stop the merging).
Figure 2. Interface of eCognition to allow users to input user
defined segmentation parameters (Image Layer weights (w t , w 2 ,
... w c ), Scale parameter (s), Shape (l-w), and Compactness
(compact))
The purposes of the segmentation parameters are (Hofmann,
2001):
(1) Scale parameter: influence the average object size. It
determines the maximal allowed heterogeneity of the
objects. The larger the scale parameter, the larger the
objects become.
(2) Shape/Color: adjust the influence of shape vs. color
homogeneity on the object generation. The higher the
shape value, the less spectral homogeneity influences
the object generation.
(3) Smoothness/Compactness: determine the compactness
or smoothness of the resulting object. With a selected
shape value, the user can influence the compactness or
smoothness of the final object.
(4) Image Layer weights: determine the weight of each
spectral band in the segmentation. It is used to control
the influence of each band on the object generation.
(5) Level settings: determine whether a newly generated
image level will either overwrite a current level or
whether the generated objects shall contain sub- or
super-objects of a existing level. The order of the level
generation affects the objects’ shape (top-down vs.
bottom-up segmentation).
2.3. Difficulty of Segmentation Parameter Selection
The segmentation parameters to be selected by the user are
interrelated to each other. It is impossible to directly find a set
of proper segmentation parameters at one time. Users have to
repeatedly select a set of segmentation parameters and test
them through a trial-and-error process, until a reasonable
segmentation result is achieved or the user does not want to
continue the trial and error any more. The change of any of the
parameters affects the influences of other parameters on the
segmentation, so that it is a tedious and time-consuming
process. The segmentation results directly depend on the
knowledge and experience of the user. The segmentation
process is considered by users as a “black art” (Smith and
Morton, 2010). Normally, only those users who are familiar
with the spectral characteristics of the land-cover objects of
interest and understand the segmentation procedure can select
proper segmentation parameters in a relatively efficient way.
But, this is not always available in practice (Flanders et at.,
691