International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
Segmentation (MS). The segmentation algorithm is a bottom-up
region-merging technique. MS starts by considering each pixel
as a separate object. Subsequently, adjacent pairs of image
objects are merged to form bigger segments. The merging
decision is based on local homogeneity criterion, describing the
similarity between adjacent image objects. The pair of image
objects with the smallest increase in the defined criterion is
merged. The process terminates when the smallest increase of
homogeneity exceeds a user-defined threshold. Therefore a
higher threshold will allow more merging and consequently
bigger objects, and vice versa. The homogeneity criterion is a
combination of colour (spectral values) and shape properties (a
combination of smoothness and compactness) Applying
different thresholds and colour/shape combinations, the user is
able to create a hierarchical network of image objects which is
necessary to extract different types of objects (area and linear
objects) because they require varied segmentation parameters
(eCognition User Guide, 2004). Darwish et al. (2003) report
about a research work about finding optimum segmentation
parameters. Up till now, finding the optimum segmentation
parameters is a trial and error process which is usually carried
out by experienced analysts.
As already mentioned MS is a region-merging algorithm which
depends strongly on the heterogeneity of the image data and on
local contrasts. Therefore the image content may influence the
transferability of segmentation parameters. The segmentation
result of transferred segmentation parameters to another image
of the same sensor in a similar geographic region can be
evaluated visually or by evaluating the classification result
which is based on the segmentation. A comparison of average
image object sizes between different images could also serve as
an indicator for the transferability of segmentation parameters.
2.2 Knowledge-base
Following image segmentation, the next step is describing the
output classes. This is achieved using a knowledge-base (Class
hierarchy) which defines the properties of the classes to be
extracted. Each class description is composed of fuzzy
expressions (Membership functions) that include logical
operations and hierarchical class descriptions. These
descriptions may include not only spectral properties but also
shape and size characteristics, context, and texture information.
Within this class hierarchy it is possible to inherit image object
properties from a super-class to a sub-class and also to group
classes semantically.
3. TEST SITES
3.1 Image Data
In this study, images from the Indian Remote Sensing Satellite
(IRS) are used. The three multispectral bands green, red, and
Near infrared (NIR) were merged with the panchromatic band
to produce a pan-sharpened image with a pixel size of 5m. The
images are not radiometrically corrected. The test sites belong
to two geographic regions in Algeria. The first is the coastal
region (Northern part of the country) and the second is the
desert region (four test sites in each region). The image contents
comprise built-up areas, agricultural land, different kinds of
vegetation, desert areas, irrigated fields, several water bodies
(sea, lake, river), and lines of communication like roads. The
distance between the most western and most eastern two test
sites in the coast region is approximately 600km. Table 1 and 2
list the used images with their dates of acquisitions and extents,
The dates of acquisition spread over nine months (coast) and
two years (desert). However it has to be noted that the images in
the desert region were all taken in December.
Table 1: Test Sites Coast
Mosta-
Test Site Algiers iie As :
ganem giers Jijel Mandoura
Date of
acquisi- 25.11.1999 30.1.2000 11.3.2000 14.8.2000
tion
Extent x10 38 x 19 17 x 10 7x6
[km x km]
Table 2: Test Sites Desert
ad Northeast
4 Hassi-El- : eS
Test Site ; of Ouargla | N'Goussa
Frid
Ouargla
Date of
acquisi- 25.12.1998 6.12.1998 19.12.2000 19.12.2000
tion
Extent 7x10 12x15 RYE 8x5
[km x km]
3.2 Output Classes and Evaluation
The output classes comprise four base classes (water, built-up,
non built up, and roads). There is no ground truth nor adequate
reference data available for Algeria. Therefore visual
interpretations of the images are used as reference data.
The evaluation of the classification is done using two
assessment tools, Error Matrix (EM) and Kappa statistic values
(K). While the former reports three values: Producers Accuracy
(PA), Users Accuracy (UA) and Overall Accuracy (OA), the
later reports a single K value for each class and an Overall
Kappa value (OK). More detailed information about accuracy
assessment from remotely sensed images can be found in
Congalton and Green (1999).
4. PRACTICAL INVESTIGATIONS
Practical investigations in this research have been carried out in
two steps. The objective of the first (transferability test) Is to
come up with a number of potential classification rules for
specific feature classes and test their transferability to other
images of each geographic region. The objective of the second
step (base rule set) is to analyse the above mentioned
classification rules to define a set of rules that is stable and
transferable and achieves the highest possible classification
accuracies.
4.1 Transferability test
After the definition of segmentation parameters, classification
rules were set-up for each test site. Following the classification
of each image accuracy assessment was performed. Within each
of the geographic regions, the sets of classification rules of each
image were transferred to the other three images. The
classification results were evaluated after applying the
transferred sets and also after adapting (adjusting existing rules
-changing image object property values- and if necessary add
new rules or feature classes specific for an image) the
transferred sets to the new image.
1060
Inte
Res
pot
kno
eacl
4.
Furt
obje
prof
obje
(sec
regi
clas:
coas
the |
the 1
appl
teste
prop
