extraction, the Quantum
GIS
environment was employed
(Q
GIS
Development Team, 2014). As the aim was to extract
rooftops based on their spectral and geometric properties,
spectrally-homogeneous areas should at first be detected,
which could later be refined into rooftops considering their
shape as well. It was determined that three segmentation
levels were required for the extraction of buildings. Each level
aimed to represent different scale objects. In implementing
a mixed bottom-up and top-down approach, larger objects
were created in higher levels of the hierarchy and finer scale
objects were created in the lower levels. Details concerning
the classes, properties and class definitions developed, can be
found in the uploaded ontology in the following link:
https://
ontohub.org/repositories/geobia-ontologies/ontologies/9893/
.
Given that vegetated, shadowed and water areas (e.g.,
pools) can be detected based on their spectral properties, a
small scale should be employed for proper extraction. Thus,
Level 1
was designed on the lower level of the multi-reso-
lution analysis (Scale = 5, Shape = 0.5 and Compactness =
0.5). First, the classes
VegetationL1
and
WaterL1
were created
and defined by large values of the Normalized Difference
Vegetation Index (
NDVI
) and Normalized Difference Water
Index (
NDWI
), respectively. The rest of the areas were classi-
fied as
OtherAreaL1
(Figure 4a).
ShadowL1
was created as
subclass of
OtherAreaL1
, defined by low intensity values of
Hue Saturation Intensity (HSI) transform of infrared, red, and
green bands. All other areas were classified as
NotShadowL1
(Figure 4a). As buildings and roads have similar spectral sig-
natures, roads were also extracted to avoid confusion with the
buildings. As roads are elongated objects, it was required the
design of a new level with appropriate objects.
Thus, a coarser level (
Level 3
) (Scale = 45, Shape = 0.8,
Compactness = 0.0) was created above
Level 1
. Classification
results from
Level 1
were projected onto
Level 3
, thus the
classes
VegetationL3
,
WaterL3
, and
ShadowL3
were devel-
oped and defined (Figure 4c). To demonstrate the projection
from lower to upper levels the definition of class
VegetationL3
is presented:
VegetationL3 EquivalentTo:
Level3 and (has_RS_0_1_RelativeAreaToSubObjects some VegetationL1)
.
This definition reads as follows:
Objects were assigned to
class VegetationL3 if they were belonging to the class Level3
and had relative area with the objects classified as Vegeta-
tionL1, greater than 50%
. All other areas were classified as
OtherAreaL3
.
RoadL3
was created as subclass of
OtherAr-
eaL3
, defined by large values of the
asymmetry
, low values of
the
density
, and relatively large values of the
length
proper-
ties. The results from the classification process of
Level 3
were
projected onto
Level 1
. Thus, the class
RoadL1
was defined
as subclass of
NotShadowL1
(Figure 4a). To demonstrate the
projection from upper to lower levels, the definition of class
RoadL1
is presented:
RoadL1 EquivalentTo:
NotShadowL1 and (is_RS_0_1_OverlappedBy some RoadL3)
.
This definition reads as follows:
objects were assigned
to class RoadL1 if they were belonging to class NotShad-
owL1 and they were overlapped by objects of class RoadL3.
As
NotRoadL1
were defined the segments not classified as
RoadL1
. Four classes (
WhiteSurfaceL1
,
OrangeSurfaceL1
,
DarkGreySurfaceL1
, and
BrightGreySurfaceL1
) were designed
to represent objects with spectral properties similar to those
of the rooftops. After defining, each class (e.g.,
WhiteSurfa-
ceL1
) a complement class was defined (e.g.,
NotWhiteSurfa-
ceL1
). The next class (e.g.,
OrangeSurfaceL1
) and its comple-
ment (
NotOrangeSurfaceL1
) were defined as subclasses of the
aforementioned class (Figure 4a). This approach was applied
to all four spectral classes related to rooftop areas. A set of
spectral indices were employed, to describe the spectral prop-
erties of each of these classes.
The objects, classified as
WhiteSurfaceL1
,
OrangeSurfa-
ceL1
,
DarkGreySurfaceL1
, and
BrightGreySurfaceL1
were spa-
tially merged and a new level;
Level 2
was created between
Level 1
and
Level 3
. Classification results from
Level 1
were
projected onto
Level 2
, thus the classes
BrightGreySurfaceL2
,
DarkGreySurfaceL2
,
OrangeSurfaceL2
,
RoadL2
,
ShadowL2
,
VegetationL2
,
WaterL2
and
WhiteSurfaceL2
were defined
through a set of proper topologic/spatial properties (Figure
4b). To exclude areas which were misclassified at
Level 1
additional spectral and texture properties were applied to
reduce noise.
Based on geometric properties (such as the rectangular fit,
length/width ratio and the area) the classes
BrightGreySurfa-
ceL2
,
DarkGreySurfaceL2
,
OrangeSurfaceL2
, and
WhiteSurfa-
ceL2
were refined into the final classes representing the roofs
of the buildings:
BrightGreyRoofL2
,
DarkGreyRoofL2
,
Orang-
eRoofL2
, and
WhiteRoofL2
. To demonstrate the geometric re-
finement, the definition of the class
WhiteRoofL2
is presented:
WhiteRoofL2 EquivalentTo:
WhiteSurfaceL2
and (areaM2 some fuzzy_TRP_20_40_360_380)
and (rectangularFit some fuzzy_rs_0.6_0.7)
.
This definition reads as follows:
Objects were assigned
to class WhiteRoofL2 if they were belonging to class White-
SurfaceL2 and had area between 30m
2
and 370m
2
and had
rectangular fit greater than 0.65
. In Figure 5 the results of the
building extraction process are presented. Some omission er-
rors are shown in ellipses, while some commission errors are
shown in rectangles. A visual examination showed satisfac-
tory results, as there are few areas misclassified as rooftops,
while the majority of the rooftops were correctly classified.
Accuracy Assessment
Classification results were compared with human interpreted
ground data and were evaluated with two methods. At first,
the number of ground data rooftops was compared to the
number of extracted building rooftops. From the total of 191
rooftops, 166 (87 percent) were detected, 25 (13 percent) were
omitted, and 16 (8 percent) were committed. On a second
step, the areas correctly detected (True Positive -
TP
), omitted
(True Negative -
TN
) or committed (False Positive -
FP
) were
computed by comparison to the ground data. Based on these
areas the Completeness, Correctness and Quality indices were
computed as follows (Agouris
et al
. 2004):
Completeness
TP
TP FN
=
+
=
. %75 0
(4)
Correctness
TP
TP FP
=
+
=
. %
79 4
(5)
Quality
TP
TP FP TN
=
+ +
=
. %62 8 .
(6)
Based on fuzzy logic, the classification stability index was
computed which is the difference between the two largest
membership values for each segment (Trimble, 2011). Table
1 contains the results of this index, for each of the building
classes of Level 2.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
June 2015
495
