International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004
(GLCM) homogeneity of height (texture measure), 2) mean
value of the segment in the red channel of the aerial image, and
3) standard deviation of length of edges in a ‘shape polygon’
created on the basis of the segment. Fuzzy membership
functions for classifying buildings and trees were formed on the
basis of the distributions of the attributes. In classification, the
three membership values for each segment were combined by
calculating their mean value.
The first classification result was improved by using the size of
the segments and contextual information on the classes of
neighbouring segments. For example, small segments classified
as buildings but mainly surrounded by trees or ground became
classified as trees. This classification step was useful to correct
very small, misclassified segments.
All neighbouring segments classified as buildings were merged
using the classification-based segmentation operation. After
this, each building segment corresponded to one entire building.
The new segments were classified on the basis of the previous
result but also using the three attributes discussed above for
buildings and trees. By this means, a membership value to class
building, calculated on the basis of the three attributes, was
obtained for each building segment.
3.2 Change detection
The change detection step was conducted in Matlab, and it was
based on simple comparisons between building segments found
in building detection and building segments derived from the
old map (the map in raster format was segmented in eCognition
to obtain a segmentation in which each building is represented
by one segment). Building segments detected from the laser
scanner and aerial image data were divided into four classes
using the following rules:
e Under 10% of the building segment is covered with
buildings in the map — New building
eo Membership value to building in classification
was > 0.75 — Certain detection
e Membership value to building in classification
was € 0.75 — Uncertain detection
e 10 — THR% of the building segment is covered with
buildings in the map —> Enlarged building
e Over THR% of the building segment is covered with
buildings in the map —> Old building
The threshold value THR was selected separately for each area
and was 80% for the industrial arca, 70% for the apartment
house area and 60% for the small-house area. Buildings of the
old map were divided into three classes on the basis of the
building detection result:
e Over 80% of the building is covered with buildings in
the classification result — Old building detected
e 10 - 80% of the building is covered with buildings in
the classification result —> Old building partly
detected |
e Under 10% of the building is covered with buildings
in the classification result — Old building not
detected
The final change detection results consist of two separate
segmentations (new segments based on the DSM and old
building segments derived from the map) with associated
classifications. For visualization, a change image was formed by
first plotting new and enlarged buildings from the classification
result and then overlaying buildings of the old map classified as
detected, partly detected or not detected (see Figures | and 2).
The image thus shows the shape and location of old buildings as
they appear in the map and the shape and location of new
buildings as they were detected in building detection. In the
study, the segmentation results were treated in raster format, but
they can also be easily converted into vector polygons.
4. RESULTS AND DISCUSSION
4.1 Building detection
Building detection results for the test areas are shown in the
upper part of Figure 1. Tables 1 and 2 show the accuracy of the
results compared with the reference map (a small part of the
apartment house area was not covered with the reference map
and was thus excluded). Results in Table | were obtained by
comparing the classification results and reference map pixel by
pixel. The accuracy measures calculated were:
; HCB&MB. | io,
e Interpretation accuracy = ————— 100% and
"yp
Rep gms 0;
e Object accuracy = — 100%,
Neg
where neg & mp iS the number of pixels labelled as buildings
both in the classification result and in the map, nyg is the total
number of pixels labelled as buildings in the map, and ncg is the
total number of pixels labelled as buildings in the classification
result.
Table 1. Accuracy of building detection estimated pixel by
pixel (I. is industrial area, A. apartment house area
and S. small-house area, see Figure 1).
Accuracy estimate Area
L A. S: All
Interpretation 96.7% | 94.9% | 91.7% | 94.2%
accuracy
Object accuracy 84.3% | 86.1% | 72.4% | 80.1%
Buildings classified 1.6% 2.1% 5.5% 3.2%
as trees
Buildings classified | 1.6% 3.0%
as ground ;
2.8% 2.6%
—d
Table 2 shows building-based accuracy estimates. In this
estimation, a given overlap calculated as the percentage of the
building's area was required for correct detection (e.g. over
70% of a building in the map had to be labelled as building in
the classification result, or over 50% of a building in the
classification result had to be labelled as building in the map).
Comparisons were made with different threshold values.
Comparisons were also made separately for large and small
buildings (threshold value 200 m?) Some comparisons were
made by considering only ‘certain’ buildings of the
classification result (membership value to building over 0.75). It
should be noted that classification and accuracy estimation was
conducted in six parts (one for the industrial area, one for the
apartment house area and four for the small-house area). If a
building was located on the boundary of the parts, it became
considered as two (or more) separate buildings in the building-
based accuracy estimation.
436
Interna
Building
Building
Figure
*
Figure
As sho
Was ac
area. T
(96.794
accurac
72.4%
area. A
not ex:
data. ai
