A Fuzzy-rule Based Ontology for Urban Object Recognition

Stella Marc-Zwecker, Khalid Asnoune and C

edric Wemmert

ICube Laboratory, BFO team, University of Strasbourg, CNRS, Illkirch Cedex, Strasbourg, France

Keywords:

Ontologies, OWL, SWRL, Fuzzy Logic, Urban Object Recognition, Satellite Images.

Abstract:

In this paper we outline the principles of a methodology for semi-automatic recognition of urban objects from

satellite images. The methodology aims to provide a framework for bridging the semantic gap problem. Its

principle consists in linking abstract geographical domain concepts with image segments, by the means of

ontologies use. The imprecision of image data and of qualitative rules formulated by experts geographers

are handled by fuzzy logic mechanisms. We have deﬁned fuzzy rules, implemented in SWRL (Semantic

Web Rule Language), which allow classiﬁcation of image segments in the ontology. We propose some fuzzy

classiﬁcation strategies, which are compared and evaluated through an experimentation performed on a VHR

image of Strasbourg region.

1 INTRODUCTION

In the domain of knowledge representation for image

recognition, we outline the principles of a method-

ology for semi-automatic extraction of urban objects

from Very High Resolution (VHR) satellite images.

This methodology relies on the design and implemen-

tation of ontologies, which are an effective tool for

domain’s knowledge formalization and exploitation

(Gruber, 1993) and for the implementation of reason-

ing. We have deﬁned two ontologies, the domain on-

tology and the image ontology. The domain ontology

(Cravero et al., 2012) describes urban objects at high

level of abstraction (houses, roads, vegetation, rivers,

etc.). The image ontology describes the image itself,

and its components (image objects), which are assem-

blies of segments resulting from the image segmen-

tation process. The ontologies are implemented in

OWL2 (Ontology Web Language) under the Prot

OWL2 editor (Grau et al., 2008).

The aim of our approach is to propose an

ontology-based framework for addressing the seman-

tic gap problem (Smeulders et al., 2000), which char-

acterizes the lack of concordance between the seman-

tic interpretation of image objects, and the pixel nu-

merical values describing them. Actually, ontologies

have been widely used in the context of image recog-

nition. We give here a brief overview of some sig-

niﬁcant work in this domain. (Maillot and Thon-

nat, 2008) proposed an ontology-based object learn-

ing and recognition system involving different aspects

of cognitive vision. Their approach relies on an on-

tology of visual concepts, such as colour and texture,

which can be seen as an intermediate layer between

domain knowledge and image processing procedures.

However, this kind of learning system requires that

the expert produces examples for each of the concepts

he is looking for.(Athanasiadis et al., 2007) present a

framework for both image segmentation and object

labeling using an ontology in the domain of multime-

dia analysis. In the ﬁeld of remote sensing, (Fonseca

et al., 2002) propose the construction of ontologies

at different levels of Geographic Information System

(GIS). The ontologies are seen as components, coop-

erating to fulﬁll the system’s objectives. Indeed, in

GIS’s conception, multiple ontologies are commonly

used to represent different levels of knowledge, but

this leads to complex systems which are difﬁcult to

understand as a whole. (Forestier et al., 2012) pointed

that although ontologies are useful to describe hierar-

chies of concepts and meta-data about image repre-

sentation, they often fail to propose an operable rep-

resentation of the knowledge that can be effectively

used for image interpretation in the domain of re-

mote sensing. They proposed a solution, based on the

construction of a knowledge base, that can be con-

sidered as a starting point for our approach, in the

sense we propose to use the ontologies in an applica-

tive way, in order to provide a semantic labelisation of

the data extracted from the image. The implemented

ontologies have the goal to help to automatically link

image objects (characterized by quantitative values)

153

Marc-Zwecker S., Asnoune K. and Wemmert C..

A Fuzzy-Rule Based Ontology for Urban Object Recognition.

DOI: 10.5220/0005026601530160

In Proceedings of the International Conference on Knowledge Engineering and Ontology Development (KEOD-2014), pages 153-160

ISBN: 978-989-758-049-9

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

with domain objects (characterized by qualitative val-

ues). The problem of mapping between qualitative

and quantitative values is called the symbol anchor-

ing problem (Coradeschi and Safﬁotti, 2003).

Our approach also copes to the imprecision of data

due to different types of sensors, sampling of data,

etc. The use of fuzzy logic allows to take this impreci-

sion into account and also facilitates the translation of

expert’s rules which are usually qualitatively formu-

lated in natural language. Our methodology is based

on a multi-level analysis, and implements fuzzy rules

for the classiﬁcation of the image segments in the on-

tology. These rules are written with SWRL (Seman-

tic Web Rule Language) (Horrocks et al., 2004). The

originality of our approach lies in the fact that, despite

the popularity of fuzzy logic, few works integrate in

the ontologies a fuzzy reasoning in the context of real

world applications (Ghorbel et al., 2010), and partic-

ularly in the domain of urban object recognition (Bel-

giu et al., 2013).

This paper is organized as follows: in section 2,

we show an overview of the proposed ontologies. In

section 3, we outline the principles of a multi-level

methodological framework, which uses fuzzy rules

for the classiﬁcation of urban objects in the ontology.

In section 4, we present a partial implementation of

the methodology, which is illustrated by an experi-

mental evaluation of the image segments fuzzy clas-

siﬁcation, based on their spectral properties. This ex-

perimentation is performed on a VHR urban image of

Strasbourg city. We ﬁnally conclude in section 5.

2 ONTOLOGIES OVERVIEW

Ontologies are a natural way to express a hierarchy of

concepts and their properties. They incorporate rea-

soning mechanisms, which allow the classiﬁcation of

individuals in the most appropriate class. We present

here two ontologies -domain and image ontologies-

that we have implemented to assist the task of semi-

automatic urban object recognition.

2.1 The Domain Ontology

This ontology has been deﬁned in collaboration with

experts geographers who created a dictionary of urban

objects (de Bertrand de Beuvron et al., 2013).

At the higher level of the hierarchy, the objects can

be either (see ﬁgure 1):

• single objects, which belong to elementary classes

(building, vegetation, etc.)

• aggregate objects, which are composed of single

objects.

Figure 1: Dictionary of urban objects.

Regarding the spatial relationship, ﬁve classical

relations used in GIS have been selected (adjacency,

inclusion, composition, alignment, distance).

Figure 2 shows a global view of the domain ontol-

ogy. For sake of simplicity, we only show the hierar-

chy’s higher concepts. In particular, relations between

concepts, such as composition or inclusion (which

correspond to OWL Object Properties) are omitted in

the ﬁgure.

Spatial Relationship concepts correspond to:

• the set of RCC8 (Region Connection Calculus)

spatial relationship, which have been reiﬁed to al-

low a consistent computation of the composition

table (Marc-Zwecker et al., 2013).

• the distance relation, which is also reiﬁed to con-

nect two objects with a value.

The Geo Object property concept allows the def-

inition of all the attributes which characterize a

Geo Object (size, shape, texture, resolution). The ex-

perts geographers usually associate these attributes to

qualitative values, e.g. large, medium, or small for

the size, rectangular or circular for the shape, homo-

geneous or heterogeneous for the texture.

2.2 The Image Ontology

Let us note that, as for the domain ontology, only the

hierarchy’s higher concepts are shown in the ﬁgure

KEOD2014-InternationalConferenceonKnowledgeEngineeringandOntologyDevelopment

154

Figure 2: Domain ontology’s higher level concepts.

3. Transversal relations between concepts are also

omitted for sake of simplicity. The higher level con-

cepts of the image ontology have the following mean-

ing: a satellite image (instance of Image concept), is

composed by a set of disjoint segments (instances of

Image Segment concept). The aim of our image pro-

cessing is to correctly assemblate the Image Segment

individuals in order to form an Image Object indi-

vidual. Actually, an ideal segmentation could allow

the fusion of the concepts Image Segment and Im-

age Object. Unfortunately, in practice, within a VHR

urban image, the object of interest is generally asso-

ciated to a set of segments (e.g. a roof is divided in

(at least) two segments, corresponding respectively to

the sunny and shadowy slopes).

The Image concept is described by the Im-

age Property concept, which characterizes the image

itself:

• acquisition properties, such as the type of sensor

used (e.g. LANDSAT, SPOT);

• location properties, such as georeferenced param-

eters, or the geographical region (e.g. Strasbourg,

South of France, etc.).

• temporal properties, such as the date and the hour,

the season, etc.

• spatial properties, such as resolution.

These meta-data are very important for selecting

the rules that will be suitable. In particular, contex-

tual knowledge can play an essential role, such as for

instance the region (e.g. roofs in the South of France

are mainly orange), or the season (the vegetation is

green in spring and yellow in autumn).

The Image Segment concept is subdivided in three

classes : Built Segment, Natural Segment and Hy-

brid Segment. A Natural Segment belongs to a pri-

mary class (water, shadow, vegetal, bare soil). A seg-

ment which is not Natural is a Built Segment (road

or building). The Hybrid Segment is a fuzzy concept

used to characterize segments which could belong to

different classes (e.g. a shadow and vegetal segment).

The Image Segment concept is characterized by OWL

Data Properties, such as spectral and spatial values

and indices.

The Image Object concept is described par Im-

age Object Property concepts, which are global

properties (size, texture, etc.) characterizing a group

of segments.

Figure 3: Image ontology’s higher level concepts.

3 INTEGRATION OF FUZZY

RULES IN THE ONTOLOGY

3.1 Fuzzy Concepts

In the area of urban image recognition, fuzzy logic

mechanisms are used to cope with the imprecision

of acquired data (Shackelford and Davis, 2003; Sui,

1992). On the other hand, fuzzy logic’s formalism

is adequate for the expression of qualitative concepts

(e.g. small, medium, and large) and allows to take

into account the vagueness that is inherent to human

natural language’s descriptions. Consequently it is

well-suited for handling rules formulated by experts

(Dubois and Prade, 2006). We brieﬂy recall the main

fuzzy logic principles that are used in our approach.

3.1.1 Fuzzy Set Theory Elements

Fuzzy set theory was proposed by (Zadeh, 1965) and

aims to address vagueness and imprecise knowledge,

by relaxing the notion of membership to a set. For-

mally, if X is the reference set, a fuzzy subset A of X

is deﬁned by the membership function f

(x), which

AFuzzy-RuleBasedOntologyforUrbanObjectRecognition

155

assigns to every x ∈ X, a value in the real interval [0,1].

As in the classical set theory case, 0 corresponds to

non-membership, and 1 to full-membership.

Membership functions are represented by fuzzy

intervals. The most popular ones have a trapezoidal,

triangular, left or right form (Straccia, 2005). The

trapezoidal function trz(x;a, b, c, d) is deﬁned as fol-

lows :

trz(x; a, b, c, d) =

⎧

⎪

⎨

⎪

⎩

0 i f x ≤ a

(x − a)/(b − a) i f x ∈ [a, b]

1 i f x ∈ [b, c]

(d − x)/(d − c) i f x ∈ [c, d]

0 i f x ≥ d

3.1.2 Fuzzy Rules

When we consider a fuzzy rule of the form: ”if X is A

then Y is B”, we need to quantify the degree of inﬂu-

ence between the premise ”X is A” and the conclusion

”Y is B”.

The fuzzy implication operator can be deﬁned

as: f

(x, y) = φ( f

(x), f

(y)), and several gener-

alizations of classical logic implication exist. We

have adopted Mamdani’s inference (Mamdani, 1977),

which is widely used in decision systems : f

(x, y) =

min( f

(x), f

(y)).

3.2 Fuzzy Rules Implementation in the

Ontology

We use Semantic Web Rule Language (SWRL) under

Protege OWL2 to implement ontology’s fuzzy rules,

with Pellet Reasoner. Indeed, (Bobillo and Strac-

cia, 2010) proposed a fuzzy ontology plugin under

Prot

e, but fuzzy concepts are modelled as anno-

tations, and hence fuzzy classiﬁcation of concepts is

not straightforward. We have adopted the approach

of (Fudholi et al., 2009), where fuzzy intervals are

directly implemented through SWRL rules, thus al-

lowing us to control the fuzzy classiﬁcation process

within the ontology.

We show in ﬁgure 4 an example of the fuzzy trape-

zoidal intervals associated to shadow and vegetal seg-

ments, for spectral band 4 (Near Infra Red) values.

In section 4 we will present the approach for learning

such spectral values.

SWRL rules that calculate the membership func-

tion associated to the vegetal segment in ﬁgure 5 are

directly deduced from the formulae of the trapezoidal

function tr (x; 800, 1000, 1500, 1600) from ﬁgure

4 (a=800, b=1000, c=1500, d=1600). The SWRL

syntax is used, and vegetal_band4(?s, ?x) repre-

sents, for the analyzed segment ?s, the value ?x of

Figure 4: Fuzzy intervals associated to shadow and vegetal

segments for NIR (Near Infra Red) values.

its membership to a vegetal segment, with relation to

the NIR (Near Infra Red) band value. Below 800 and

above 1600, the inferred membership value is equal

to zero (the corresponding rules are not shown).

Image-Segment(?s), band4(?s, ?b4),

greaterThanOrEqual(?b4, 800.0f),

lessThan(?b4, 1000.0f),

subtract(?sub, ?b4, 800.0f),

divide(?div, ?sub, 200.0f)

-> vegetal_band4(?s, ?div)

Image-Segment(?s), band4(?s, ?b4),

greaterThanOrEqual(?b4, 1000.0f),

lessThanOrEqual(?b4, 1500.0f)

-> vegetal_band4(?s, 1.0f)

Image-Segment(?s), band4(?s, ?b4),

greaterThan(?b4, 1500.0f),

lessThanOrEqual(?b4, 1600.0f),

subtract(?sub, 1600.0f, ?b4),

divide(?div, ?sub, 100.0f)

-> vegetal_band4(?s, ?div)

3.3 Fuzzy Classiﬁcation Strategies

The inference mechanism is based on Mamdani’s in-

ference (Mamdani, 1977), with two alternative strate-

gies for the calculation of the membership value asso-

ciated to the rule’s premises:

• in the ﬁrst strategy, the calculated membership

value corresponds to the minimum of membership

values of the premise’s elementary fuzzy propo-

sitions. We call this strategy CMI, for Classical

Mamdani Inference ;

• in the second strategy, the calculated member-

ship value corresponds to the weighted average

of membership values of the premise’s elementary

fuzzy propositions. We call this strategy WAMI,

for Weighted Average Mamdani Inference.

In our approach, the conclusion function trivially

corresponds to the searched membership degree (e.g.

the degree for ”is vegetal”). Therefore the conclusion

can be merely modelled by the identity function with

KEOD2014-InternationalConferenceonKnowledgeEngineeringandOntologyDevelopment

156

a membership degree of 1. Hence, the Mamdani’s in-

ference simply consists in copying the membership

degree that has been computed for the rule’s premise

(by any of the strategies) into the conclusion’s mem-

bership degree.

The defuzziﬁcation process will consist in apply-

ing the ”Smallest of Maximum” method, thus return-

ing the value of the membership degree calculated by

the Mamdani’s inference.

The ﬁnal crisp decision (e.g. vegetal segment)

is then submitted to a threshold (e.g. 0.7). Within

our method, a segment can be classiﬁed into different

classes (e.g. vegetal and shadow) with distinct mem-

bership values, since this corresponds to the reality

(e.g. the shadow on a meadow).

The following SWRL rule illustrates the WAMI

strategy, where we consider a simple average (i.e.

the same weight for all the rule’s premisses). The

predicates vegetal_band3, vegetal_band4, and

vegetal_ndvi_index, respectively give the values

of blue band, NIR band, and Normalized Difference

Vegetation Index for a given segment.

Image-Segment(?s),

vegetal_band3(?s, ?vb3),

vegetal_band4(?s, ?vb4),

vegetal_ndvi_index(?s,?vnd),

add(?add1,?vnd, ?vb4),

add(?add,?add1,?vb3),

divide(?div,?add,3)

-> is_vegetal_segment(?s, ?div)

3.4 A Multi-level Methodology for

Urban Object Classiﬁcation in the

Ontology

Our methodology of semi-automatic urban object

recognition is under implementation. It based on the

following steps:

• the satellite image to analyse is ﬁrst segmented,

using an image segmentation algorithm. The

obtained segments with their attributes are then

loaded in the image ontology under the Prot

OWL2 editor. Every segment is entered as an in-

dividual in the Image Segment class, and its at-

tributes are associated to OWL Data Properties

(see ﬁgure 5).

• all the individuals corresponding to image seg-

ments are then classiﬁed in the image ontology,

with the type of fuzzy SWRL rules that we have

presented in the previous section. This ﬁrst classi-

ﬁcation is based on the segments’ spectral proper-

ties, and gives the primary class (vegetation, wa-

ter, shadow, etc.).

• the adjacent segments which belong to the same

primary class after the spectral classiﬁcation are

then grouped, to constitute Image Object individ-

uals. At this stage, some spatial criteria can be

used. For instance, the adjacency to segments of

shadow allows the detection of buildings.

• ﬁnally, the obtained Image Object individuals are

compared to the domain ontology’s Geo Object

concepts, in order to improve the classiﬁcation.

For example, object’s attributes such as shape or

size will allow to verify if an Image Object in-

dividual classiﬁed as a house, corresponds to a

house description according to the domain ontol-

ogy.

As we have stated before, the result of the fuzzy

classiﬁcation can propose one or several Geo Object

classes for the same Image Object individual, with

distinct membership degrees. This methodology is

currently under development, and so far, we have par-

tially implemented all the steps except the last one.

Figure 5: A segment’s description under Prot

e editor.

4 EXPERIMENTATION AND

EVALUATION RESULTS

4.1 Learning of the Fuzzy Rules

In these experiments, we used a Pleiades

image

of the city of Strasbourg (France) acquired in 2012

and having a resolution of 0.7m/pixel and a size of

9211x11275 pixels. We extracted an area which

presents all the interesting thematic classes we wanted

to identify. Figure 6 presents the segmentation of the

area that was used to learn the fuzzy rules (400x650

pixels).

The spectral fuzzy rules were built based on the

standard attributes and spectral indexes found in liter-

ature, such as NDVI (Normalized Difference Vegeta-

tion Index) or NDWI (Normalized Difference Water

Pleiades: Dual optical system for metric resolution ob-

servations (CNES, France)

AFuzzy-RuleBasedOntologyforUrbanObjectRecognition

157

Figure 6: Segmentation used for the learning of the fuzzy

rules.

Index) (Sebari and He, 2013; Bouziani et al., 2010).

To get better rules, more adapted to our type of im-

ages, we tried to combine experiments with theoreti-

cal values. We obtained the experimental values us-

ing a set of samples, extracted by geographical ex-

perts. The images were segmented using the Mean-

shift algorithm (Comaniciu and Meer, 2002). Then,

all the segments were labeled using six ground truth

classes: vegetal, shadow, water, building, road and

soil. The hybrid and unknown class contains the over-

lapped and non-identiﬁed segments. Table 1 resumes

the number of labeled segments for each class.

Table 1: Number of labeled segments for each ground truth

class used.

Vegetal 567

Shadow 279

Water 144

Road 86

Soil 113

Building 253

Building (Wall) 44

Hybrid and unknown 477

Total 1963

4.2 Validation

To validate our approach, we have tested the fuzzy

rules that we implemented in SWRL language under

Prot

e 4.3. We have carried out a two-stage valida-

tion. The ﬁrst stage points out the effectiveness of the

fuzzy rules in term of f-measure compared to crisp

rules. Moreover, we compared the weighted average

Mamdani inference (WAMI) to the classical Mam-

dani inference (CMI) for the considered example. In

the second stage, the behavior of the fuzzy classiﬁ-

cation was visually analyzed in different types of ar-

eas, which were extracted of the same global image

of Strasbourg region.

For the ﬁrst scenario, the precision, recall and f-

measure criteria have been calculated.

precision =

+ f

(1)

recall =

+ f

(2)

f − measure =

2. precision.recall

precision+recall

(3)

where t

stands for true positive (i.e. the number

of items correctly labeled as belonging to the positive

class), f

for false positive (i.e. number of items in-

correctly labeled as belonging to the class) and f

for

false negative (i.e. items which were not labeled as

belonging to the positive class but should have been).

4.2.1 Comparison of Crisp and Fuzzy

Classiﬁcation Rules

In this ﬁrst experimentation, we have used four types

of rules:

• rules based on the small crisp interval (core)

• rules based on the large crisp interval (support).

• rules based on the fuzzy WAMI strategy with a

threshold of 0.7,

• rules based on the fuzzy CMI strategy with a

threshold of 0.7.

Table 2 shows that the precision obtained with

small interval is obviously superior to the precision

obtained with large and fuzzy intervals but with the

lowest recall (and vice versa). The fuzzy interval ob-

tained by applying the WAMI strategy gives the best

f-measure which represents the best compromise be-

tween precision and recall. However, in the partic-

ular case of shadow class, the CMI strategy gives

the best f-measure, while the WAMI strategy gives

the best recall, which seems more interesting for the

classiﬁcation. Indeed, we observed that shadow seg-

ments belonging to other classes, e.g. vegetal, are

detected as being simultaneously shadow and vege-

tal segments. Thus we assume that during the next

step of the methodology, the use of spatial rules will

enhance the precision. Moreover, as we have already

pointed, getting the maximum recall for the shadow

segments is useful for buildings detection. Conse-

quently, in this experimentation, the WAMI strategy

is the best suited to the classiﬁcation of image seg-

ments.

KEOD2014-InternationalConferenceonKnowledgeEngineeringandOntologyDevelopment

158

Table 2: Results of the comparison among fuzzy rules (WAMI and CMI strategies with threshold=0.7) and crisp rules.

Small interval Large interval Fuzzy interval WAMI Fuzzy interva CMI

Pr Rec F-meas Pr Rec F-meas Pr Rec F-meas Pr Rec F-meas

Vegetal 0.988 0.155 0.267 0.763 0.966 0.852 0.876 0.864 0.87 0.963 0.47 0.63

Shadow 0.864 0.433 0.576 0.257 1.0 0.408 0.572 0.967 0.718 0.92 0.63 0.74

Water 0.99 0.71 0.82 0.73 0.965 0.831 0.969 0.881 0.922 0.984 0.861 0.917

Soil 0.368 0.371 0.369 0.148 0.991 0.257 0.376 0.672 0.482 0,42 0.292 0.344

4.2.2 Visual Evaluation

To conﬁrm the quality of the results, we also proposed

a visual evaluation. Figure 7 presents the result of the

detection obtained applying the fuzzy rules with the

WAMI strategy to an urban area, similar to the one

used for learning (375x399 pixels). We have super-

posed over the original image the result of our clas-

siﬁcation : green areas have been classiﬁed as veg-

etation, yellow areas have been classiﬁed as shadow,

and red areas correspond to unclassiﬁed areas (mostly

vegetation or shadow).

Figure 7: Classiﬁed urban area.

5 CONCLUSIONS

We have presented an original work, that attempts to

show the effectiveness of ontologies’ use in the do-

main of urban object recognition from VHR satellite

images. The proposed methodology is under develop-

ment and aims to address the semantic gap and sym-

bol anchoring problems, by providing explicit cor-

respondence mechanisms between an abstract ontol-

ogy, which qualitatively describes the domain’s con-

cepts, and a concrete ontology, which quantitatively

describes the image objects.

Our work takes into account the uncertainty that

is inherent to the acquired data by the implementation

of fuzzy rules in the ontology.

So far, our methodology’s implementation allows

the fuzzy classiﬁcation of image segments into pri-

mary classes (vegetal, shadow, water, bare soil, build-

ing). When the spectral classiﬁcation does not affect a

segment to a primary class, the system deduces that it

belongs to an artiﬁcial class (building or road). The

distinction between building and road will be done

in a later stage, by using the adjacency to shadow (a

building is adjacent to a shadow, a road is not).

The grouping of segments of the same primary

class into image objects and the buildings detection

are being currently developed. The step of matching

between the domain ontology and the image ontology

will be ﬁrst applied to the buildings construction. In-

deed, at the current stage, the adjacency to shadow al-

lows the detection of some segments belonging to the

buildings. But unlike the case of objects belonging to

primary classes (vegetation, water, etc.), the assembly

of segments belonging to a building is very difﬁcult

because roofs are dissimilar and their spectral prop-

erties are not stable. It will then be essential to ﬁnd

the general characteristics of the houses (shape, size,

etc.) in the domain ontology to help the grouping of

buildings’ segments.

A medium term perspective of our work is to in-

tegrate the fuzziness in all stages of the methodology,

and particularly in the image segmentation process.

ACKNOWLEDGEMENTS

All the data sets were produced by UMR7263 LIVE

(Anne Puissant, Simon Rougier, Gregory Skupinski),

for the research project COCLICO (ANR-12-MONU-

001), which funded this work.

REFERENCES

Athanasiadis, T., Mylonas, P., Avrithis, Y., and Kollias, S.

(2007). Semantic image segmentation and object la-

beling. Circuits and Systems for Video Technology,

IEEE Transactions on, 17(3):298–312.

AFuzzy-RuleBasedOntologyforUrbanObjectRecognition

159

Belgiu, M., Lampoltshammer, T. J., Hofer, B., et al.

(2013). An Extension of an Ontology-Based Land

Cover Designation Approach for Fuzzy Rules, volume

2013. Verlag der

Osterreichischen Akademie der Wis-

senschaften.

Bobillo, F. and Straccia, U. (2010). Fuzzy ontology repre-

sentation using owl 2. CoRR, abs/1009.3391.

Bouziani, M., Go

ıta, K., and He, D.-C. (2010). Automatic

change detection of buildings in urban environment

from very high spatial resolution images using exist-

ing geodatabase and prior knowledge. ISPRS Journal

of Photogrammetry and Remote Sensing, 65(1):143–

153.

Comaniciu, D. and Meer, P. (2002). Mean shift: A robust

approach toward feature space analysis. Pattern Anal-

ysis and Machine Intelligence, IEEE Transactions on,

24(5):603–619.

Coradeschi, S. and Safﬁotti, A. (2003). An introduction

to the anchoring problem. Robotics and Autonomous

Systems, 43:85–96.

Cravero, M., de Beuvron, F. d. B., Zanni-Merk, C., and

Marc-Zwecker, S. (2012). A description logics geo-

graphical ontology for effective semantic analysis of

satellite images. In KES, pages 1573–1582.

de Bertrand de Beuvron, F., Marc-Zwecker, S., Puissant,

A., and Zanni-Merk, C. (2013). From expert knowl-

edge to formal ontologies for semantic interpretation

of the urban environment from satellite images. Inter-

national Journal of Knowledge-based and Intelligent

Engineering Systems, 17(1):55–65.

Dubois, D. and Prade, H. (2006). La logique ﬂoue. REE.

Revue de l’

electricit

e et de l’

electronique, (8):35–41.

Fonseca, F. T., Egenhofer, M. J., Agouris, P., and C

amara,

G. (2002). Using ontologies for integrated geographic

information systems. Transactions in GIS, 6(3):231–

257.

Forestier, G., Puissant, A., Wemmert, C., and Ganc¸arski,

P. (2012). Knowledge-based region labeling for re-

mote sensing image interpretation. Computers, Envi-

ronment and Urban Systems, 36(5):470–480.

Fudholi, D. H., Maneerat, N., Varakulsiripunth, R., and

Kato, Y. (2009). Application of prot

e, swrl and

sqwrl in fuzzy ontology-based menu recommenda-

tion. In Intelligent Signal Processing and Commu-

nication Systems, 2009. ISPACS 2009. International

Symposium on, pages 631–634. IEEE.

Ghorbel, H., Bahri, A., and Bouaziz, R. (2010). Fuzzy

ontologies building method: Fuzzy ontomethodology.

In Fuzzy Information Processing Society (NAFIPS),

2010 Annual Meeting of the North American.

Grau, B., Horrocks, I., Motik, B., Parsia, B., Patel-

Schneider, P., and Sattler, U. (2008). Owl 2: The next

step for owl. Web Semantics: Science, Services and

Agents on the World Wide Web.

Gruber, T. R. (1993). Toward Principles for the Design

of Ontologies Used for Knowledge Sharing. To-

ward Principles for the Design of Ontologies Used for

Knowledge Sharing, 43:907–928.

Horrocks, I., Patel-Schneider, P. F., Boley, H., Tabet, S.,

Grosof, B., and Dean, M. (2004). SWRL: A seman-

tic web rule language combining OWL and RuleML.

W3c member submission, World Wide Web Consor-

tium.

Maillot, E. and Thonnat, M. (2008). Ontology based com-

plex object recognition. Image and Vision Computing,

26(1):102–113.

Mamdani, E. H. (1977). Application of fuzzy logic to ap-

proximate reasoning using linguistic systems. IEEE

Transactions on Computers, 26:1182–1191.

Marc-Zwecker, S., De Beuvron, F. D. B., Zanni-Merk, C.,

Le Ber, F., et al. (2013). Qualitative spatial rea-

soning in rcc8 with owl and swrl. In KEOD 2013-

International Conference on Knowledge Engineering

and Ontology Development.

Sebari, I. and He, D.-C. (2013). Automatic fuzzy object-

based analysis of vhsr images for urban objects extrac-

tion. ISPRS Journal of Photogrammetry and Remote

Sensing, 79:171–184.

Shackelford, A. K. and Davis, C. H. (2003). A combined

fuzzy pixel-based and object-based approach for clas-

siﬁcation of high-resolution multispectral data over

urban areas. IEEE T. Geoscience and Remote Sens-

ing, 41(10):2354–2363.

Smeulders, A., Worring, M., Santini, S., Gupta, A., and

Jain, R. (2000). Content-based image retrieval at the

end of the early years. IEEE Transactions on Pattern

Analysis and Machine Intelligence, 22.

Straccia, U. (2005). A fuzzy description logic for the se-

mantic web.

Sui, D. Z. (1992). A fuzzy gis modeling approach for urban

land evaluation. Computers, environment and urban

systems, 16(2):101–115.

Zadeh, L. A. (1965). Fuzzy sets. Information and Control,

8:338–353.

KEOD2014-InternationalConferenceonKnowledgeEngineeringandOntologyDevelopment

160