3. Istanbul 2004
ongly labeled as
oT
ments.
for detecting the
The proposed
g and utilizes the
ast shadows. The
lows can provide
d buildings. The
tory. The overall
he 79 collapsed
providing 79.7%
[lapsed buildings,
nt between the
for un-collapsed
ildings, 142 were
’s accuracy.
RV data analysis
in Izmit, Turkey.
24(12), pp. 2439-
itomatic detection
:Ms created from
ial photographs.
tion of collapsed
urkey earthquake
rial photographs.
THE AUTOMATIC CLASSIFICATION OF B&W AERIAL PHOTOS
L. Halounova
Remote Sensing Laboratory, Faculty of Civil Engineering, Czech Technical University Prague
Thakurova 7, 166 29 Prague 6, Czech Republic, halounov@fsv.cvut.cz
KEY WORDS: Land Cover, Photography, Texture, Fuzzy Logic, Segmentation, Aerial, Multiresolution
ABSTRACT:
The paper shows possibility to classify B&W aerial orthophotographs and other monochromatic remote sensing data using image
enhancement phase and object-oriented analysis phase for the automatic classification.
The first phase enlarges the spectral signature space by channels calculated by image filtering (median filter and Gauss filter) and by
texture measures. The combination of various filter sizes (texture measures) and kernel sizes (filters) enlarges the signature space
allowing the following image segmentation and classification in two or three scale levels. The at least two level classification
simplifies thematically complex aerial orthophotographs by dividing the photo into thematically more homogenous areas in the
higher level. The lower level brings the final sought classes. which can be slightly corrected in the third (lowest) level. The
eCognition software was used for the image segmentation and for the automatic classification. It is the first method of automatic
classification of land cover of monochromatic remote sensing data bringing accuracy better than 80 per cent.
1. INTRODUCTION
The paper author was a responsible person for analyzing one
part of a project of the Czech Ministry of Agriculture. The aim
of the project was to find solutions for automatic information
extraction from B&W aerial orthophotographs. The scale of
these aerial photographs was 1: 23 000. Each othophoto was a
result of mosaicking. The classical way of automatic
classification is based on close spectral signatures or other
signatures of pixels representing the same classes. This
assumption is not valid in case of B&W photographs where
different areas are formed by pixels with the same values and on
the other side one class is formed by wide range digital values.
The signature space for individual classes overlapped and could
not have been used for their distinguishing in this state. No
references were found to show solution for the similar
monochromatic data type. Known image enhancement methods
were applied to be obtained more separable class signature
space for individual classes. The automatic pixel-by-pixel
classification was excluded from the analysis and replaced by
the object-oriented classification performed for segmented
image data.
Segmentation has been used by several specialists who applied
various interpretation keys (Borisov et al., 1987, Jagtap et al.,
1994, Naesset, 1996, Zihlavník, Palaga, 1995). The
segmentation used in this project was the Fractal Net Evolution
Approach (FNEA) commercially introduced by Baatz and
Schäpe (1999) incorporated in commercial software
eCognition.
The proposed method can be applied for large number of with
relatively high level of automation. That was the project goal.
The result of the project should be applied for the whole coun-
try and therefore their image processing operability was neces-
sary.
2. METHODOLOGY
There were two main tasks in the image processing. To enlarge
signature space of individual classes to be separable was the
first task. To perform the automatic classification formed the
second task.
2.1 Signature space enlargement
Signature space enlargement was done by using two ways of
new channel calculation. One way was the image filtering by
low-pass filters where Median filter and Gauss filter were
applied. They supressed local image unhomogeneities. They
showed high correlation with original image data. That was the
reason why they did not assure sufficient class separability.
Channels calculated from two kernel sizes and repeatable
filtering were used within the project. Channels filtered by
Gauss filter were calculated for standard deviation equal to 2, 3,
and 4.
Another tool was necessary for the successful solution. Haralick
functions were the tool as functions characterizing textures.
There are more Haralick functions used in image processing.
Different numbers of them were used for different level of
classification detail. Haralick functions were tested for several
window sizes. Window sizes depended on individual class
member sizes. Smaller window sizes were useful for small
resulting class members sizes. The tests showed that there were.
no prevailing trends in a certain direction and that was the
reason why all directions where relations between pixels are
determined, were taken into account. All directions meant that
differences between a pixel and a reference pixel were
calculated for directions 0°, 45°, 90°, and 135°. These
differences were used for GLCM (Grey Level Co-occurrence
Matrix) and GLDV (Grey Level Difference Vector) calculations
from Haralick functions. Mean, standard deviation and
dissimilarity functions were chosen for the orthophoto
classification defined in following expressions. Their window
size were 5x5 pixels, 11x11 pixels, and 21 x 21 pixels.