bul 2004
——
| extents,
ast) and
mages in
andoura
4.8.2000
’Goussa
).12.2000
8x5
built-up,
adequate
e visual
sing two
tic values
Accuracy
OA), the
1 Overall
accuracy
found in
ied out in
test) is to
rules for
to other
he second
nentioned
table and
ssification
ssification
ssification
ithin cach
es of each
ges. The
ying the
sting rules
ssary add
nage) the
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
Resulting from this transferability test, there are a number of
potential classification rules for specific feature classes and the
knowledge which rules were transferable to other images of
each geographic region.
4.0 Base rule set
Further investigations deal with the question: which image
object properties are transferable or in other words which
properties are stable in different images? Therefore potential
object properties which resulted from the transferability test
(section 4.1) were compared in all test sites of the geographic
region coast to come up with a set of knowledge-based
classification rules. For this comparison, the four images of the
coast region were combined into one big image. Then each of
the potential object properties was displayed in eCognition and
the ranges used in the initial classification were tested to find if
applicable in all images or not. If not, different ranges were
tested until either a good range of values is found or the
property considered as unsuitable.
a A age 3 €. ç
E cid m u^ ve
Figure 2: Texture property of Jijel
1061
Figure 1 and 2 show an example of an examined image object
property (Grey Level Co-Occurrence Matrix -GLCM-
Homogeneity in the NIR channel (Haralick 1973)). The images
are displayed in grey values. Dark coloured image objects
posses a small value whereas brighter segments have a high
value for this texture property. From the figures, it is clear that
this texture property has similar values for the built-up areas of
Mostaganem and Jijel. It is also visible that using solely this
property is not sufficient to classify built-up areas because other
features like quarry and bush land also fulfil it.
All other potential properties were analysed in the same manner
to produce the base rule set for coast (Table 3). In this set, non
built-up areas which comprise different kinds of vegetation like
agricultural land and forest, are classified using the Normalized
Difference Vegetation Index (NDVI) and the GLCM Variance
in NIR. Similar textural and spectral properties characterize the
class built-up. The rules for the linear classes roads and river
contain shape properties like length to width, ratio and density
(the area covered by the image object divided by its radius
which describes the compactness of an image object). Water is
defined by the Ratio NIR (the NIR mean value of an image
object divided by the sum of all spectral channel mean values).
Lake and sea are child classes from water so that they inherit the
property Ratio NIR from water. The feature lake is classified
using shape properties area and length and sea is characterised
by the Ratio NIR like the parent class water.
The resulting so called base rule set which contains spectral,
texture, and shape properties was applied to all the test sites in
the coast region and accuracy assessment was performed.
Table 3: Base rule set for Coast
Class Property Range
NDVI >0
Non-Built up GLCM
Variance NIR 9
Length/ =
Vidth nhá
Roads Wic
Density < 0.9
Water Ratio NIR > 0.26
Lake Area > 6500
(child of
water) |Length < 310
Sea
(child of |Ratio NIR <0.19
water)
GLCM ASM
< 0.002
NIR 0.00
Buil Mean Green >90
ilt-
naw Ow S
Homog. NIR +
GLCM 5
Contrast Red =n
Mean NIR 25-50
River Length/ E dn
sy > 5
Width
