International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
n-l n-l
Mean i= SNR m
LJ
i=0 j=0
n-l n-l 1/
St deviation = S: S| , (i — Mean _ iy [2
E00
n-l n-l
Dissimilarity — SS SS P - j
i=0 j=0
where Pi; is the normalized grey level co-occurrence matrix
(GLCM) with n x n size, where original n = 256 for 8bit data is
pre-scaled to n=16 such that
n-l
SS pij)
i, j-0
2.2 Object-oriented analysis
Problem with wide range of digital values representing one
thematic class and overlapping values for individual classes
were partly solved by the previous step — new channel
calculations. This step has not solved all problems with
overlapping values of individual classes. That was image
segmentation what helped to divide image data into large
regions representing important parts of land use. This
segmentation allowed separating of urban parts from
agricultural parts and from forest areas.
The segmentation performed by object-oriented analysis
defined in eCognition software simplified thematically
complicated image data content. The object-oriented analysis
comprises two parts. The first one prepares image data by
creating segments from them and the second one allows their
classification.
The segmentation is based on heterogeneity evaluation. The
heterogeneity is characterized either by spectral heterogeneity,
or spatial heterogeneity and their mutual combination. Higher
influence of spectral heterogeneity is accomplished by lower
spatial heterogeneity while their sum is equal to |. The spatial
heterogeneity compares either the compactness taking into
account segment length and its number of pixels, or the
smoothness expressing relation between segment length and the
shortest segment length defined by a rectangle circumscribing
the segment.
2.3 Class definition
Two-level class definition was the result of the presented
methodology. The higher-level segmentation served to
classification into basic regions:
e old forest,
e young forest
e agricultural area,
eurban area.
The lower-level classification comprised higher number of
-lasses. Each of them belonged into the only higher-level class.
The following list shows classes for the lower-level (more
456
detailed classification). Class names are created from three
parts. The first part part (F = forest, NF = nonforest, A=
agricultural area and U urban area) defines the real situation of
segments derived from the lower-level classification. The
second part determines belonging to higher-level classification
where OF means old forest, YF represents young forest, AA is
an abbreviation for agricultural area and UA urban areas. The
third part of class names indicates the lower-level class. The
complete lower-level class names are:
eF OF coniferous forest,
eF OF. deciduous forest,
eF OF forest older than 7 years,
eNF OF forest up to 7 years,
eF YF coniferous forest,
eF YF deciduous forest,
eF YF forest older than 7 years,
eNF. YF forest up to 7 years
eU YF road,
eF AA tree,
*A AA field,
eU AA house
eU AA road,
eF UA tree,
eNF UA green area,
eU UA light house,
eU UA dark house,
eU UA house,
eU UA road.
The first part of names shows possible regrouping of certain
classes, which were originally classified into thematically
wrong higher-level classes. This process of regrouping brings
new improvements into the analysis. The forest class is formed
by three classes from old forest, three classes from young forest
and one class from agricultural areas and one class from urban
areas. Class definitions and their list were adapted to real
situation in the image data and can change from to region to
region.
2.4 Segmentation
The segmentation called multiresolution segmentation allows
segmentation in more levels. The higher-level segmentation for
image data division into thematically simple regions was
calculated for high scale value (250) in the first image-
processing phase. The original orthophoto and the channel
calculated by median filter with 5x5 kernel size formed the
input image data for the segmentation processes in the higher-
level segmentation.
The lower-level segmentation being the second image-
processing phase was a repetition of the first one for lower scale
value (35 — 50 according to image data). The segmentation
process used the same channels.
The influence of spectral heterogeneity varied from higher
value for the higher segmentation level to lower value for more
detailed classification where spatial heterogeneity played more
important role.
2.5 Classification
After the segmentation, segments can be classified into classes.
The segmented image data classification. was done by the
Intern
neare:
used
calcul
differ
last p
as tex
stand:
calcul
and 2
signat
The h
cases
proce:
done |
shows
classi
The a
sampl
accurz
f
forest
dec
coni
2 of
ER
Produ
User