images when dealing with a detailed urban mapping. There is no
universally accepted method to determine an optimal scale level to
segment objects. Moreover, a scale level may not be suitable for all
classes in an image classification. The best approach to select which
bands to be considered for the membership function and which scale
level to employ for a particular class would be to identify the class with
different options and qualitatively analyze them on the display screen
as a generate and test approach. The output map needs to be checked
carefully throughout the image. We note that some membership
functions seemed to work very well for a particular class in one part of
the study area, but it did not perform well for the same class in other
parts. The analyst needs to check a zoom-in version of the output at all
possible spots before observing other classification options at different
level of scales. It would be a good idea to treat individual classes
separately to explore which method and scale level with which feature
space is potentially good to extract a class. However, in some cases,
classification of a few classes together in multispectral bands with the
use of the nearest neighbor option could be better than individual
classes separately using a membership function with a set of expert
system rules. Even though there were seven classes in our classification,
we kept several different classes or many different training
samples of the same classes to perform the classification. The threshold
values given in this study may not be applicable to other urban
mapping using the same satellite data (QuickBird), even though the
classification system employs the same classes in a similar urban
environment. However, similar or considerably different threshold
values with a slight modification of the parameters can be expected to
be effective for urban mapping in different environmental settings.
Since both classifiers available in the object-based approach are
non-parametric rules, they are independent of the assumption that
data values need to be normally distributed. This is advantageous,
because most data are not normally distributed in many real world
situations. One of the other advantages of the object-based approach
is that it allows additional selection or modification of new objects
(training samples) each time, after performing a nearest neighbor
classification quickly until the satisfactory result is obtained. There
are many possible combinations of different functions, parameters,
features, and variables available with the object-based approach. The
successful use of the object-based paradigm largely relies on
repeatedly modifying training objects, performing the classification,
observing the output, and/or testing different combinations of
functions as a trial-and-error process.
Our experience was that Definiens or eCognition software was not
able to perform many features or bands at many different scale levels
for image segmentation and classification. This was simply because
the computer memory needs to be used extensively to segment
tremendous numbers of objects from many different bands, especially
when requiring smaller scale parameters (larger scale segmentation).
We used different computer hardware and experienced numerous
computer breakdowns and freezes during the segmentation, even
though our study area is a small part of the whole Phoenix
metropolitan area. This should be considered a limitation especially
when dealing with a large dataset (finer resolution data for a
relatively large area). Nonetheless, the object-based classification
system is a better approach than the traditional per-pixel classifiers in
urban mapping using high-resolution imagery