object based methods for identification of crop parcels using
variety of images having range of spatial resolution between
.5 m and 30 m. Their results show that
OBC
provided best
classification accuracy values for 10 m and finer spatial reso-
lution images. The Object-Based Classification process starts
with segmentation step (Herold
et al
., 2003). Multi-resolution
segmentation algorithm combines pixels or existing image
objects with each other based on homogeneity criteria which
is combination of shape and compactness criteria; also, scale
parameter defines object sizes that obtained with segmenta-
tion process. These segmentation parameters are determined
by testing segmentation results with various parameters.
In Urban Atlas, the urban area related classes are classi-
fied with a minimum mapping unit of 0.25 ha. Object sizes
have been taken into account in the segmentation process.
Before the segmentation process, Center and Safranbolu
districts divided into 2 regions within themselves based on
administrative borders for rural and urban areas. As these
areas have different properties, the segmentation and clas-
sification process varied for each one. For urban areas, the
most appropriate values identified as scale factor 100, shape
0.8, and compactness 0.6. To increase the accuracy and obtain
sharper-lined segments, especially for residential parcels and
roads, vector data prepared from the 1:1000 scaled zoning
plan and the railway vector data from OpenstreetMap were
added as thematic layers and included in the segmentation
phase. In the segmentation process for rural areas, Openstreet-
Map used for roads and railways and vector data for forest
areas which were obtained from related official institution
used as thematic layer. For rural areas, the scale parameter
was chosen as 150, shape 0.8, and compactness 0.8. The clas-
sification process was conducted by using thematic layers,
image based indices and functions by using objects’ texture
and spectral information. Table 2 shows the indices and func-
tions that were used in the classification process.
For both Center and Safranbolu urban areas, 1:1000 zone
maps derived vector data was used to assist the object-based
classification and assignment of residential areas, industrial,
commercial, public, military, and private units, green urban
areas, sport and leisure facilities. Elliptic fit function was
used for classification of the roads.
NDVI
values used to dis-
cern forest areas and agricultural, semi natural, wetland areas
and assigned to related class. For determination of urban fab-
ric subclasses,
NDVI
values used. The
NDVI
values calculated
in the interval -1 -1 are converted to 8-bit data ranging from
0-255 to be compatible with other satellite image data sets.
Then, threshold values are determined for the five subclasses
to be created and assigned to the related subclasses from the
“urban fabric class”. Classification results of the west part of
Safranbolu urban area is shown in Figure 2.
The
NDVI
values of Center urban area were found to be
distributed between 131.94 and 212.81. The
NDVI
values of
the urban fabric classes were found to be distributed between
139.13 and 208.28 values. Five subclasses of urban fabric
were divided into 5 ranges and 5 categories of urban fabric
that were classified each interval. According to this classifica-
tion, Figure 3 shows some examples of urban fabric classes
Figure 2. West part of Safranbolu urban area: (a) Segmentation result, and (b) classification result.
Table 2. Definitions of indices and functions used in
classification process.
NDVI
Indicator used to determine green vegetation NDVI=
(NIR-RED) / (NIR+RED)
Elliptic Fit measure how objects look like ellipse
Area
Calculates the number of pixels contained in the image
objects.
Shape
Index
Defines the smoothness of the edges of an image object.
As the smoothness increases, this value decreases.
Calculated by, dividing the edge length of image object
by the square root of the 4th area of image object.
Standard
Deviation
Calculated from the image band density values of all the
pixels forming the image object
Brightness The arithmetic average of image band values
714
November 2018
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
