ground scenes. An ontology was developed based on the imag-
ery annotations, and a common sense ontology was employed
to extract additional concepts that linked the concepts defined
by the annotations. A multi-stage reasoning approach was de-
veloped to extract the annotations for each scene.
In the field of remote sensing, ontologies have been em-
ployed to represent expert knowledge for the representation of
the properties and relations of objects in an image. Kohli
et al
.
(2012) developed a generic slum ontology based on input from
50 experts, covering 16 countries. Bertrand
et al
. (2013) devel-
oped an ontology framework based on
OWL
2 and the Seman-
tic Web Rule Language (
SWRL
) employing expert knowledge,
to describe urban elements and their spatial organization.
Ontologies have been also applied in the field of
GEOBIA
. Du-
rand
et al
. (2007) developed a methodology for object recogni-
tion, from very high-resolution imagery, based on an ontology
developed by machine learning techniques and experts. The
ontology was based on spectral and geometric properties of
the objects. In Forestier
et al
. (2012) the matching algorithm
introduced by Durand
et al
. (2007) was employed to map
the observations extracted from the image with the domain
nomenclature (linguistic notions). The method was employed
to extract image domain elements from QuickBird Imagery.
In Belgiu and Lapoltshammer (2013), information contained
in visual interpretation keys was modeled into an ontology to
perform extraction from Very High Resolution imagery. An on-
tology was designed, defining a class- subclass hierarchy based
on the United Nations Land Cover Classification System. Each
class was defined by employing spectral and geometric features
existing in visual interpretation keys. Belgiu
et al
. (2014) em-
ployed random forest and ontologies to extract buildings from
ALS data by employing topographic and geometric features.
The above studies demonstrated the applicability of
ontologies in
GEOBIA
. However, there still remains the need
to incorporate fuzzy reasoning with spatial relationships and
multi-scale analysis in an ontology-based
GEOBIA
approach.
Indeed, Belgiu
et al
. (2013) stated the need for fuzzy reason-
ing in
GEOBIA
ontologies. Furthermore, the importance of
spatial relationships has already been stated in early
GEO-
BIA
studies (Baatz and Sch pe, 2000; Burnett and Blaschke,
2003). Although, spatial reasoning with ontologies for image
analysis have been examined (Hudelot
et al
., 2008; Forestier
et al
., 2012; Bannour and Hudelot, 2014), there still remains
the need for spatial reasoning in a multi-scale
GEOBIA
ontolo-
gies. In Arvor
et al
. (2013) and Belgiu
et al
. (2014) a limitation
was stated for current Description Logic (
DL
)-based ontology
reasoners concerning the required processing time, when
dealing with large number of classes and objects, which is
very common in
GEOBIA
studies.
To address the above stated needs, this research aids in the
development of an ontological reasoner for
GEOBIA
, named
SPatial Ontology Reasoner (
SPOR
).
SPOR
was developed to pro-
vide representation of fuzzy spectral, geometric, and spatial
relationships for the development of a
GEOBIA
ontology. Spa-
tial relationships were incorporated to express relationships
between single or multiple levels of analysis. Considering the
language employed to represent ontologies,
OWL
2 was se-
lected, which is a World Wide Web Consortium (
W3C
) recom-
mendation (Motik
et al
. 2012).
OWL
2 ensures the integration
and exchange of
GEOBIA
ontologies with current semantic web
technologies. In Hay and Castilla (2008) it was noted the need
for
GEOBIA
applications to incorporate Open
GIS
Standards
(Open Geospatial Consortium, 2014) to ensure collaboration
with current remote sensing and
GIS
software. Thus,
SPOR
was
designed to incorporate technologies which already adopted
these standards (e.g., PostgreSQL) and also provide for time
efficiency in processing large amount of data. A final objective
was to suggest guidelines for the syntax of
GEOBIA
ontologies
to increase the readability by remote sensing experts.
This paper is organized as follows: the next Section pres-
ents the developed methodological framework, followed by
building extraction through ontologies. Then, a discussion of
results and comparison with other studies is presented fol-
lowed by conclusions and future prospects.
Methodology and Implementation
Bobillo and Stracia (2011) enhanced
OWL
2 with fuzzy rep-
resentations encoded as
OWL
2 metadata (annotation proper-
ties). Thus, it was adopted as ontological language. The mem-
bership values were computed based on Zadeh semantics
(Zadeh, 1965). To store the segmentation results compatibly
with Open
GIS
Standards and perform spatial computations,
it was decided to employ PostgreSQL (PostgreSQL Global
Development Group, 2014). In the following section, the
implementation of
SPOR
is presented.
Fuzzy OWL 2 Ontologies for GEOBIA
Based on the
OWL
2 specification,
OWL
2 ontologies enable
the design of classes, individuals, properties, datatypes, and
annotations (Motik
et al
. 2012).
Classes
represent groups
of things, while
Individuals
represent actual objects of
the domain (segments), which in
SPOR
were stored within
PostgreSQL. Data properties represent relationships between
an
individual
and data values.
Object properties
are entities
that connect pairs of
individuals
. As spatial relationships are
interpreted as topologic links between two segments,
object
properties
were employed to represent spatial relationships
between the objects, such as the hasRelativeAreaToSubOb-
jects.
Datatypes
are entities that refer to sets of data values.
Annotation
properties are employed to encode metadata for
the ontology itself or the declarations within the ontology.
Axioms
, or
expressions
are statements that are asserted to be
true in the domain being described.
To restrict
Data properties
,
fuzzy datatypes
are employed
(Bobillo and Stracia, 2011). As an example, the partial defi-
nition of Vegetation class is presented in Manchester
OWL
Syntax (Motik
et al.
, 2012) as follows:
ndvi some mediumToHighValuesOf
NDVI
.
(1)
The mediumToHighValuesOf
NDVI
is a
fuzzy datatype
. The
fuzzy information, regarding the membership function and
its borders is encoded, within an
annotation property
called
fuzzyLabel
, as follows:
<Datatype type=”rightshoulder” a=”0.2” b=”0.4” /> . (2)
To design
fuzzy object properties
capable of defining
topologic/spatial relationships, the relative annotation should
include information regarding the type of the spatial relation-
ship, the membership function, and its borders. Thus, in the
restriction
isSurroundedBy some Water
, the isSurroundedBy
notion was designed to be the
fuzzy object property
which
represents the spatial relationship of the segments with the
ones classified as Water. Thus, the
fuzzy object property
was
designed to have an annotation property as it follows:
<Role datatype=”highRelBrdr” spatial_relationship=”relative_border”/>.
(3)
The highRelBrdr is a
fuzzy datatype
with similar definition
to the one presented earlier. The
spatial_relationship
attribute
describes the type of the spatial relationship represented by
the
fuzzy object property
. The developed spatial relation-
ships express links between objects at the same level (relative
border, length of common border, distance of centroids, and
distance from the outer border of a class), between an object
and objects of lower levels (relative area and overlaps) and be-
tween an object and objects of higher levels (overlapped by).
In Figure 1, the expressions supported by
SPOR
are presented.
The examples are given in Manchester
OWL
2 Syntax. For
each
OWL
2 entity a suggested syntax example is given.
492
June 2015
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
