Automatic fuzzy object-based analysis of VHSR images for urban objects extraction

Imane Sebari
a,
, Dong-Chen He
b
a
Filire de Sciences Gomatiques et Ingnierie Topographique, IAV Hassan II, Rabat, Morocco
b
Centre dapplications et de recherches en tldtection (CARTEL), Universit de Sherbrooke, Sherbrooke, Qubec, Canada
a r t i c l e i n f o
Article history:
Received 15 September 2012
Received in revised form 5 February 2013
Accepted 5 February 2013
Available online 25 March 2013
Keywords:
Automatic object extraction
Object Based Image Analysis (OBIA)
Fuzzy rule base
VHSR satellite images
Urban areas
a b s t r a c t
We present an automatic approach for object extraction from very high spatial resolution (VHSR) satellite
images based on Object-Based Image Analysis (OBIA). The proposed solution requires no input data other
than the studied image. Not input parameters are required. First, an automatic non-parametric coopera-
tive segmentation technique is applied to create object primitives. A fuzzy rule base is developed based
on the human knowledge used for image interpretation. The rules integrate spectral, textural, geometric
and contextual object proprieties. The classes of interest are: tree, lawn, bare soil and water for natural
classes; building, road, parking lot for man made classes. The fuzzy logic is integrated in our approach in
order to manage the complexity of the studied subject, to reason with imprecise knowledge and to give
information on the precision and certainty of the extracted objects. The proposed approach was applied
to extracts of Ikonos images of Sherbrooke city (Canada). An overall total extraction accuracy of 80% was
observed. The correctness rates obtained for building, road and parking lot classes are of 81%, 75% and
60%, respectively.
2013 International Society for Photogrammetry and Remote Sensing, Inc. (ISPRS) Published by Elsevier
B.V. All rights reserved.
1. Introduction
Mapping impervious surface from remote sensing imagery is
signicant to a range of issues especially to sustainable develop-
ment of urban areas (Demarchi et al., 2012; Gao et al., 2012;
Longley et al., 2005; Holland et al., 2006; Donnay et al., 2001). Very
high spatial resolution (VHSR) images and advanced image pro-
cessing algorithms both driven the technologic advance in remote
sensing of impervious surfaces (Xu, 2013). One of the emerging
trends in this eld is OBIA, object-based image analysis (Weng,
2012). OBIA is considered as a powerful tool for classication and
analysis of VHSR images compared to the traditional per-pixel
classiers (Blaschke, 2010; Navulur, 2007; Blaschke et al., 2000).
The advantage of OBIA is that it does not use individual pixels
but adjacent pixel groups that can be characterized by spectral,
textural, geometric and contextual information. Taking this infor-
mation into account through the object-based approach allow
obtaining enhanced results (Campbell, 2007).
Object based image analysis has been dened as a new disci-
pline at the rst international conference on Object-Based Image
Analysis:
Object-Based Image Analysis (OBIA) is a sub-discipline of
GIScience devoted to partitioning remote sensing (RS) imagery into
meaningful image-objects, and assessing their characteristics
through spatial, spectral and temporal scale. (Hay and Castilla,
2006).
The OBIA, also called object oriented image analysis, aims to
replicate and/or to surpass the human interpretation of images
automatically or semi-automatically (Hay and Castilla, 2006).
Two main stages can form the OBIA process: (1) creation of image
objects and (2) classication of image objects. Usually, the rst
step is performed through a segmentation technique (Lang and
Blaschke, 2006; Jensen, 2005). This step is a crucial since it pro-
vides the basic units (image objects) on which later process will
be applied. Therefore, the success of OBIA approach is related to
segmentation quality. The second stage, classication, tries to cre-
ate real objects from image objects. The classication method is
chosen with relation to the desired goal, to the studied image, and
also to its ability to integrate ancillary information. Several meth-
ods can be used at the two stages of the OBIA approach. The chosen
algorithms strongly inuence the nal results (Lu and Weng, 2007;
Caloz and Pointet, 2003).
The rst known reference that used the object-based approach
was Kettig and Landgrebe (1976). They proposed a classication
approach of multispectral images by extracting and classifying
homogenous objects. Their approach consists rst in subdividing
the image in spectrally homogenous pixel groups. These groups
are then classied through supervised technique (maximum likeli-
hood). They applied their approach to aerial and satellite images
0924-2716/$ - see front matter 2013 International Society for Photogrammetry and Remote Sensing, Inc. (ISPRS) Published by Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.isprsjprs.2013.02.006

Corresponding author. Tel.: +212 650632611; fax: +212 37680180.

E-mail addresses: [email protected], [email protected] (I. Sebari).
ISPRS Journal of Photogrammetry and Remote Sensing 79 (2013) 171184
Contents lists available at SciVerse ScienceDirect
ISPRS Journal of Photogrammetry and Remote Sensing
j our nal homepage: www. el sevi er. com/ l ocat e/ i spr sj prs
Downloaded from http://www.elearnica.ir
(Landsat). This method is still used under the name of ECHOs
(Extraction and Classication of Homogeneous Objects). It is avail-
able in an open-source environment Multispec (Biehl and Land-
grebe, 2002). Lee et al. (2003) applied it to extract shapes and
positions of building from Ikonos image. The ECHO approach has
been used by Jimnez et al. (2005) under a non-supervised version
(UnECHO) to extract homogenous regions from hyperspectral
images.
The use of OBIA became more widespread with the advent in
2000 of eCognition, the rst commercially available, object based
image analysis software (Blaschke, 2010). It is today known as
Deniens. The object image creation step is conducted by a multi-
resolution segmentation based on the Fractal Net Evolution Ap-
proach (Baatz and Schpe, 2000). The segmentation algorithm is
a bottom-up region-growing technique. The growing decision is
based on local homogeneity criteria describing the similarity of
adjacent image objects in terms of size, distance, texture, spectral
similarity and form (Baatz and Schpe, 2000). User-dened thresh-
olds are interactively used to decide whether objects are merged
into larger objects or not. For image object classication, two
methods are proposed: the nearest neighbor and fuzzy rules base
(Benz et al., 2004). Several attributes (spectral, geometric and con-
textual) can be integrated in the classication process. Signicant
studies have used this software in different applications (forestry,
urban, agriculture, coastal zone, etc.). Regarding VHSR satellite
images of the urban area, they attempted to interpret either the
whole image (Myint et al., 2011; Kux and Arajo, 2008; Marchesi
et al., 2006; Caprioli and Tarantino, 2003; Mittelberg, 2002; Kress-
ler et al., 2001; Meinel et al., 2001), or to extract specic objects
such as buildings (Hofmann, 2001a), roads (Repaka et al., 2004;
Nobrega et al., 2008), and private gardens (Mathieu et al., 2007)
or informal settlements (Hofmann et al., 2008). Recently, other im-
age analysis software has developed OBIA modules like Feature
Analyst (Tsai et al., 2011) or ENVI Feature Extraction (Hu and
Weng, 2011).
Compared to the pixel-based approach, the extracted objects by
an OBIA approach are more homogeneous than by pixels based ap-
proach and are closer to a visual human interpretation (Huiping
et al., 2003). The OBIAs results (extracted objects) can be inte-
grated within vector GIS more easily than classied raster maps
(Walter, 2004). The OBIA can be applied on different satellite
images. However, it has proven to be particularly appropriate for
VHSR imagery especially in urban areas (Jacquin et al., 2008;
Campbell, 2007; Mo et al., 2007; Herold et al., 2003; Shackelford
and Davis, 2003; Rego and Koch, 2003; Bauer and Steinnocher,
2001; Hofmann, 2001b). The classication of VHSR images of urban
environment can show some imprecision due to the nature of
these images and the studied objects. In fact, satellite images pix-
els can correspond to several objects of different natures. This mix-
ture causes imprecision in the classication of these pixels. With
VHSR images, the problem is less important but still present. Also,
some different objects can yield close or similar spectral responses.
This problem is greater in urban areas. Several objects of different
classes yield the same spectral responses. This is due to the use of
the same building materials and/or by the low spectral resolution
of the images (Herold et al., 2004).
The researches on object extraction from VHSR images in urban
areas which applied the OBIA approach are various. They differ
according to the objects to extract and also to the methods used
in each stage. The extraction can concern only one object class (like
building (Lee et al., 2003), Roads (Repaka et al., 2004)) or try to
interpret the whole image (Shackelford and Davis, 2003). The used
classication methods can vary from the more conventional, using
only spectral information (Myint et al., 2011), to the more complex,
based on external knowledge (Forestier et al., 2012; Bouziani et al.,
2010). In OBIA approach, the integrated knowledge can be related
to the objects characteristics (spectral, geometric and contextual
properties, relationships between objects, etc.), to information ini-
tially relative to the extraction (object models, constraints, etc.) or
to the used data (date and position of the sensor) (Baltsavias,
2004). Adopting a classication method based on knowledge
would give the opportunity to take more information into account.
This allows better discrimination between object classes and ef-
cient extraction of objects (Campbell, 2007). Rule-based systems
belong to knowledge-based methods that simulate the human rea-
soning mechanism and translate knowledge through decision rules
(Tso and Mather, 2001). Fuzzy logic can also be integrated in clas-
sication methods in order to resolve knowledge representation
and classication problems in complex environments (Han et al.,
2005). Thus, fuzzy rules consist of a set of fuzzy expressions allow-
ing the evaluation of specic attributes. In comparison with a clas-
sical rule, the response to a fuzzy rule is given with a degree that
expresses the satisfaction of this rule, simultaneous application
of several rules is allowed and an object can have different mem-
bership degrees to the studied classes (Dubois et al., 2007). The -
nal decision can be taken according to the rule to which the
membership degree is maximal.
Several studies have used fuzzy rule base in OBIA approach to
extract urban object from VHSR. A review of these studies can be
found in Weng (2012) and Blaschke (2010). They have shown high
extraction precisions. However, during their process, values of
parameters and thresholds were set manually. One of the recent
OBIA research is directed towards the automation of image pro-
cessing (Blaschke, 2010). This should concern both segmentation
and classication methods used during OBIA process. For segmen-
tation, automatic technique should allow the creation of image ob-
jects without setting any parameters or homogeneity threshold.
For a fuzzy rule base, an automatic solution will allow the general-
ization of decision rules on other images of different types without
reformulating other rules more adapted to the new context. For-
mulating rules based on human expert knowledge is not always
easy. Ascertaining thresholds and weights of rules is usually left
to the human user or based on training data (Walter, 2004; Ishibu-
chi et al.,1992; Puissant et al., 2006).
In this paper, we propose a methodology for automatic extrac-
tion of urban objects from VHSR satellite images through an ob-
ject-based image analysis approach. The proposed solution
requires no input data other than the studied image. Not input
parameters are required. Segmentation is conduct by a nonpara-
metric cooperative technique. The extraction is based on a fuzzy
rule base adapted for interpreting VHSR images in urban areas.
The extracted objects are organized in layers with information on
the precision and certainty of their extraction. This paper is orga-
nized as follows. We describe rst the proposed methodology.
We present then the test data as well as obtained results. A last
section will present the analysis and discussion of these results.
2. Proposed approach
We proposed a new approach for automatic extraction of ob-
jects from VHSR images. The objects of interest concern principal
urban object and are presented in Table 1. The proposed approach
adopts object-based image analysis principle and is constitutes of
two principal steps: creation of primitives from pixels and creation
of objects from primitives. The primitive corresponds to an inter-
mediate state between the pixel and the object to extract. It con-
sists of a group of homogenous adjacent pixels. The rst step is
conducted by segmentation technique and the second by fuzzy
rule base. These two steps are performed automatically without
need to introduce parameters. The extracted objects are organized
in layers by classes. Individual layers are overlaid to produce a nal
172 I. Sebari, D.-C. He / ISPRS Journal of Photogrammetry and Remote Sensing 79 (2013) 171184
output map of extracted urban objects. Information on the quality
of the extraction is also provided. Fig. 1 presents the concept of our
approach.
2.1. Creation of primitives from pixels
Segmentation technique is used to create the primitives from
pixels. The image segmentation algorithm used in this study fol-
lows the approach given in Sebari and He (2009) where the seg-
mentation requires no parameters and no input data other than
the images to be processed. It is based on cooperation between re-
gion-growing segmentation and edge segmentation. The segmen-
tation adopts sequential region-edge cooperation. The edge
segmentation is performed rst on panchromatic band and inte-
grated into multispectral region growing segmentation as addi-
tional criteria for seeds selection and for segmentation criteria
denition. The approach uses a spectral homogeneity criterion
whose threshold is adaptive. It varies across the image depending
on the object to be segmented and its neighborhood. It is more
appropriate than a single threshold to apply to the entire image
especially for complex images like VHSR images. The Fig. 2 pre-
sents the principle of the adopted segmentation.
The threshold of spectral homogeneity is calculated automati-
cally for each new segment in every spectral band during the pro-
cess of segmentation: Once the seed is chosen and before
aggregating the pixels, a window of analysis is centered on it and
the spectral values of pixels contained in this window are consid-
ered to elaborate a frequency histogram for differentiated values.
For this purpose, a band differentiation algorithm is used to deter-
mine the absolute maximum difference between a pixel and its
neighbors for each pixel:
dvi; j max
16k61
16l61
jvi; j vi k; j lj 1
With dv(i,j) is the differentiated value at the pixel (i,j) and v(i,j)
is the value of pixel (i,j). In homogenous areas, dv has small values,
whereas dv takes larger values in the boundaries between regions.
The overall shape of the histogram is bimodal (Fig. 3): Since more
pixels are inside objects than boundaries, the rst peak corre-
sponds to pixels inside the objects and the second represents the
pixels at the boundaries. If dv is considered as the spectral homo-
geneity factor h, the threshold T is considered as the valley be-
tween the two peaks. The automatic detection of this valley is
done according to a modied technique of Zack et al. (1977). The
valley corresponds to the point in the histogram that is farthest
from the straight line joining the two peaks of the histogram.
A pixel may be assigned to a segment if it satises in the n
bands of the image the following condition:
fh
1
< T
1
AND. . . AND h
b
< T
b
AND. . . ANDh
n
< T
n
g;
with; b f1; . . . ; ng 2
After the segmentation, the primitives are transformed into vec-
tor format and represented by polygons. Our choice of the vector
mode is justied by many reasons:
- Easy denition of geometric properties.
- Explicit topology between the different objects to dene con-
textual properties.
- Possibility to describe a segment through several attributes
stored in a database.
- Easier to integrate the extracted objects layers in existing geo-
graphic database.
- Possibility of overlaying the polygons layer with other layers in
raster or vector format.
- Easier to compare to others geographic database.
2.2. Creation of objects from primitives
The step of passing from primitives to object is conducted by
applying a fuzzy rule base. This fuzzy rule base contains a knowl-
edge used by a human photo-interpreter to identify urban objects.
Table 1
Objects of interest.
Level I Level II Level III
Natural classes Vegetation Lawn
Forest
Tree
Bare soil Bare soil
Water River
Lake
Man-made classes Road network Road
Parking lot
Building Building
E
x
t
r
a
c
t
e
d
i
n
f
o
r
m
a
t
i
o
n
P
r
o
p
o
s
e
d

A
u
t
o
m
a
t
i
c
O
b
j
e
c
t

e
x
t
r
a
c
t
i
o
n

a
p
p
r
o
a
c
h
I
n
p
u
t

d
a
t
a
B G R NIR
Multispectral
segmentation
Fuzzy rule base
Object classes M
e
m
b
e
r
s
h
i
p

d
e
g
r
e
e
Vegetation Building Road Parking lot
Segmented image
PAN
Information on
extraction quality
Fig. 1. Proposed automatic object extraction approach.
I. Sebari, D.-C. He / ISPRS Journal of Photogrammetry and Remote Sensing 79 (2013) 171184 173
The methodology adopted to establish the rules is as follow:
knowledge modeling, fuzzy rule base creation and assignment of
objects to classes.
2.2.1. Knowledge modeling
The purpose of this step is to model the knowledge that an ex-
pert uses in order to identify urban objects. Five photo-interpreta-
tion keys (Paine and Kiser, 2003) were used to dene objects
properties: size, shape, color (spectral response), texture and sha-
dow. We have considered spectral, textural, geometric and contex-
tual properties to describe studied objects. Then, a quantiable
attribute is associated to each object property. A quantiable attri-
bute is dened by a mathematical formulation. The choice of attri-
butes is based upon those used in the literature and validated by
tests. The Table 2 presents the quantied attributes adopted for
each property.
Then, we have done a discriminative analysis which purpose is
to dene, for each attribute, a discriminative threshold upon which
the attribute will characterize the associated property. The analysis
consists in studying the mathematic formulation and variation of
each attribute. Afterwards, two approaches were followed to
dene attributes discriminative thresholds depending on whether
the studied attribute is dependent or not on the image. For the
attributes that are independent on the image (like elongation in-
dex, compactness index, etc.), the thresholds are determined based
on their use in the literature and on tests. For the attributes
depending on the image, methodologies are proposed to automat-
ically dene the corresponding discriminative thresholds. The con-
cerned attributes are the brightness index (shadow property) and
area (large property).
The methodology used to automatically determine the bright-
ness index threshold is based on the histogram of frequencies of in-
dexs values in the studied image. The threshold is chosen as the
rst valley in the histogram (Fig. 3). This is justied by the fact that
shadow areas in VHSR image present low spectral values in the
four bands. We consider that, for an urban area image, and on
the frequency histogram of brightness index, the rst peak reects
shadow areas. It reects also water or low albedo materials. But,
since the objects will be described by multicriteria rules, the use
of other attributes will allow the discrimination between shadow
and the others objects.
The automatic extraction of the brightness index threshold is
performed by applying the proposed algorithm: (1) Searching
automatically rst for the value corresponding with the rst peak
(rst maximal value), (2) searching automatically thereafter for
the value of the threshold that satises the criteria: its brightness
index value is higher to the one of the rst peak and it is comprised
within two values whose frequencies are higher than the fre-
quency of this value.
The area index is used to describe large size objects. Dening
area index threshold is based on the knowledge that, on an urban
VHSR image, there are more objects of small and medium size than
large-size objects. Thus, the frequency of an images objects area
values histogram looks like described in Fig. 4. The automatic
extraction of this threshold is performed by using a modied ver-
sion of the triangle technique proposed by Zack et al. (1977). It is
performed through an algorithm that searches for the value corre-
sponding to the farthest point on the line joining the greatest peak
and the last point in the histogram with the maximum area value.
Automatic determination
of spectral homogeneity
thresholds for n bands
Add to the region's
boundary
New segment
For each pixel in the list
Yes
No
Segment
Look for the 8
unassigned
neighboring pixels
Add to the list of pixels to check
T
1
, ,T
n
Add the pixel to the
region
Seed
Check the spectral homogeneity criterion and the edge criterion

Look for the 8 unassigned

neighboring pixels
Check the adjacency criterion
h
b
< T
b
| b={1,,n}
edge pixel
Fig. 2. Adopted segmentation approach (Sebari and He, 2009).
Differentiated
values (dv)
Threshold selected
Objects
Edges
P
i
x
e
l
s
Fig. 3. General shape of the histogram of the differentiated values.
174 I. Sebari, D.-C. He / ISPRS Journal of Photogrammetry and Remote Sensing 79 (2013) 171184
The equation of the straight line connecting the two points
P
1
(A
1
, f
1
) and P
2
(A
2
, f
2
) is:
a x b y c 0
with: a
f
1
f
2
A
1
A
2
, b 1 and c f
1
a A
1
The distance D
P
i
between each point P
i
of the histogram and this
line is calculated:
D
P
i

a x
P
i
b y
P
i
c

a
2
b
2
_

The histogram valley corresponds to the point of the histogram

whose distance D
P
i
is maximal. This technique offers the advantage
of valley detection being unaffected by histogram irregularities.
Also, we reminder that the detected threshold will be used in fuzzy
rule which will take in account the imprecision related to its
determination.
Adopting these methodologies to dene attributes thresholds is
according to the purpose of our approach: extract object from im-
age without set up of any parameter. The possible imprecision
associated to these methodologies will be take in count by using
fuzzy logic during rules formulation.
2.2.2. Fuzzy rule base creation
The use fuzzy logic is considered appropriate for problems re-
lated to urban areas images (Longley et al., 2001): uncertainty
associated to the objects boundaries, fuzziness in the denition
of classes, possibility to belong to more than one class. One of its
benets is the possibility to express the membership degrees of ob-
ject to different classes and the uncertainty related to objects
extraction process.
Several reasons justify our choice for fuzzy rule base: in our
case, no a priori knowledge related to the extracted object is used.
The rules are formulated based on photo-interpretations descrip-
tion; the fuzzy logic allows the formulation of knowledge given
in natural language with vague and imprecise expressions. Also,
it avoids the use of rigid limits for class denitions and introduces
thresholds uncertainty; what will reduce the imprecision related
to the adopted threshold determination methods. With this a lack
of knowledge, the extracted objects should be provided with infor-
mation of the precision of their determination; it is possible with
fuzzy logic which allows the use of membership functions and of
degree of truth attributed to each extracted object (Bouchon-Meu-
nier and Marsala, 2003; Dubois et al., 2007).
Each objects property is described by a fuzzy proposition and
considered as a fuzzy set characterized by a membership function.
The membership degree is determined according to the value of
the attribute associated to the property. Each objects class is de-
scribed by a fuzzy rule that expresses the relationship between
the class and the properties that describe it. The satisfaction degree
of a class rule is determined according to the membership degrees
of properties.
2.2.2.1. Fuzzy propositions on properties. Fuzzy proposition on prop-
erty correspond to a fuzzy formulation of the condition that the
quantiable attribute must satisfy in order to discriminate the ob-
ject property. The formulation of a fuzzy proposition on an object
property starts by the description of the property by a fuzzy subset.
Generally, a fuzzy proposition is like X is P, with X being a vari-
Table 2
Correspondence between properties and quantiable attributes.
Property Quantiable attribute Mathematical formulation
Spectral attributes Vegetation Normalized Difference Vegetation
Index (NDVI)
NDVI
NIRR
NIRR
Shadow Brightness index I
br

1
6
B V 2R 2NIR
Water Near-infrared (NIR) spectral
response
NIR
Textural attributes Texture Homogeneity index
HOM

n
i0

n
i0
1
1 i j
2
h
c
i; j
Geometric
attributes
Elongated shape Elongation index I
elongation

4pA
P
2
A and P are the area and the perimeter of the exterior border
Compact shape Compactness index
Icompactness
4pAHull
P
2
Hull
A
Hull
and P
Hull
are the area and the perimeter of the convex hull polygon
Convex shape Convexity index I
convexity

A
AHull
A is the objects area and A
Hull
is the area of the convex hull polygon
Large size Area A = Area of the object
Contextual
attributes
Neighboring object Adjacency Determines the neighboring objects that share a part of their limits
Direction between two
objects
Azimuth
Az
12
arctg
X2X1
Y2Y1
_ _
(X
1
, Y
1
) coordinates of the centroid of object 1
(X
2
, Y
2
) coordinates of the centroid of object 2
Elevated object Adjacency + Shadow + Azimuth Determines if object is adjacent to shadow zones in the direction of the
sun
Frequencies (f)
f
1
A
2
f
2
Area
(A)
Area Threshold
D
Pi
P
1
P
2
A
1
Fig. 4. Automatic detection of areas threshold.
I. Sebari, D.-C. He / ISPRS Journal of Photogrammetry and Remote Sensing 79 (2013) 171184 175
able and P is the fuzzy subset studied. The subset is dened by a
membership function l
p
that associates to each value x of the attri-
bute a real value l
p
(x) in the interval [0, 1]. This value l
p
(x) trans-
lates the satisfaction degree of the property P by an object. The
higher the membership degree is, the more the property is satis-
ed, and vice versa.
For the determination of the satisfaction degree of the opposite
property, the complement operator is used. For example, if l
c
(x)
is the membership degree for the compact shape property, the
membership degree to the non-compact shape property l
non c
(x)
is equal to:
l
non c
x 1 l
c
x 3
There are many types of membership functions: continuous
(e.g. Gaussian or sigmoid functions) or piecewise linear (e.g. trian-
gular or trapezoidal functions) (Zadeh, 2003). The latest type is a
simple form which contains straight line segments. We have
choose this type of membership functions since it will allows to
represent the knowledge collected during the modeling step. Thus,
the shapes of membership functions are dened with relation to
information related to attributes mathematic formulations and
thresholds values. The shape of the membership function is spe-
cic for each property.
The methodology starts by specifying the intervals for which
the associated attribute values characterize the property, as well
as those for which the property is not validated. Between those
intervals, a transition interval, dened with relation to the attri-
butes threshold, reects a gradual membership and minimizes
the imprecision related the attributes thresholds determination.
Fig. 5 represents the overall shape of the used membership
function.
2.2.2.2. Fuzzy rules on object classes. An object class rule constitutes
a conjunction of propositions on the properties describing this
class. Generally, it follows this form:
IF X
1
is P
1
AND . . . AND X
i
is P
i
AND . . . X
n
is P
n
THEN Y is C
i
4
where X
i
is P
i
is a fuzzy proposition on a property and Y is C
i
the
conclusion translating membership to the class C
i
. The degree of
satisfaction of a rule is obtained by using a fuzzy aggregation oper-
ator which aggregates the membership degrees like a logical AND.
We adopt the following operator:
l
C
i
x

a
i1
l
P
i
x
_ _
1=a
5
where l
C
i
x is the membership degree of object x to class C
i
,
l
P
i
x is the satisfaction degree of object x to property P
i
and a is
the number of fuzzy properties propositions that constitute the
class rule. We have chosen this operator because since the aggrega-
tion of the rules is performed in a conjunctive way, if one of the
rules concludes that the membership degree to the class C
i
is 0,
then no other rule can change this conclusion, and the class C
i
will
be considered impossible. In other terms, if an object does not sat-
isfy an object class property, it will not be affected to this class. Its
satisfaction degree will be null. Also, this operator veries the nor-
malization condition (Eq. (6)); the comparison of the satisfaction
degrees of the various class rules is possible.
sup
x2A
px 1 6
The membership function of each class is represented by a sin-
gleton. The membership degree is a real value in the interval [0, 1].
Fig. 6 presents a simplied example that illustrates the principle of
determination of class membership degree. Class A is described by
two properties elongated shape and non-compact shape. The
fuzzy rule of this class is:
IF object shape is long AND object shape is non-compact THEN
object belongs to class A
The studied object presents values for the attributes associated
to these properties: the elongation index for the elongated shape
propriety and the compactness index for the non-compact shape
propriety. First, the satisfaction degrees of the object to the two
fuzzy proprieties propositions are determined, l
L
(x) and l
NC
(x).
The membership degree of the object to the class A is obtained
by aggregating the two degrees with the fuzzy operator obtained
through the Eq. (3): l
Road
(x) = (l
L
(x). l
NC
(x))

.
The following rules were used:
WATER _RULE = {(RS is water)}
TREE_RULE = {(RS is vegetation) AND (texture is rough)}
LAWN_RULE = {(RS is vegetation) AND (texture is non rough)}
BARE_SOIL_RULE = {(RS is bare soil)}
BUILDING_RULE = {(RS is non-vegetation) AND (RS is non-bare
soil) AND (RS is non-water) AND (elevated object) AND (Shape
is compact) AND (shape is convex)}
ROAD_RULE = {(RS is non-vegetation) AND (RS is non-bare soil)
AND (RS is non-water) AND (Object non-shadow) AND (Object
non-elevated) AND (shape is Elongated) AND (size is large)}
PARKING_LOT _RULE = {(RS is non-vegetation) AND (RS is non-
soil) AND (RS is non-water) AND (Object non-shadow) AND
(Object non-elevated) AND (large size) AND (shape is
compact) AND (shape is convex)}
2.2.2.3. Auto-learning process. Since no knowledge on the real pro-
prieties of objects is used during rules formulation, we integrate an
auto-learning process to automatically retrieve information about
potential objects. The retrieved knowledge is integrated into new
rules in order to extract more objects. This process starts after
the application of the dened fuzzy class rules. It retrieves infor-
mation from potential objects that are dened as objects having re-
sponded to the fuzzy class rules with high membership degrees.
They are considered as representative of their class and conse-
quently the values of their attributes characterize the objects of
this class. The auto-learning process retrieves spectral and geomet-
ric attribute values and integrates them in a library created for this
purpose. The spectral information is used to evaluate the spectral
membership of each object to the class spectral library. The geo-
metric information is integrated in order to evaluate the geometric
conformity with potential objects. Contextual information is used
with the retrieved information to dene new class rules. This pro-
cess is applied for man made classes: building, road and parking lot.
For the road class, the new rule is dened in order to connect the
extracted road network. Thus, the object, in addition to having
membership to the road spectral library, must be adjacent to an al-
ready extracted road object. For the parking lot class, the corre-
0
1
0
Threshold
1
Attribute

A
Fig. 5. General shape of the used membership function.
176 I. Sebari, D.-C. He / ISPRS Journal of Photogrammetry and Remote Sensing 79 (2013) 171184
sponding rule includes membership to the spectral and geometric
parking lots libraries. The new rules for the building class include
membership to the spectral and geometric buildings librarys and
shapes compactness condition. In order to take more advantage
of the collected spectral library, the primitives that do not belong
to the building, road or parking lot classes, and have membership
to the spectral libraries, are extracted as impervious surfaces.
2.2.3. Assignment of objects to classes
After application of the fuzzy rule base, each object has a mem-
bership degree that varies from 0 to 1 to each class.
After applying the rule base, an object has membership degrees
to the all studied classes (varying from 0 to 1). In order to interpret
theses results and to take a decision, we have adopted some no-
tions of the possibility theory. The possibility theory is linked with
the fuzzy sets theory and allows to reason on imprecise or vague
knowledge.
If C is the set of classes, a possibility distribution p is dened as
a function that attributes a possibility coefcient p
x
(C
i
) to each ele-
ment of C. This coefcient represents the possibility that an object
x belongs to a class C
i
. If p
x
(C
i
) = 0 then the object cannot belong to
the class C
i
. If p
x
(C
i
) = 1 then it is possible (but not certain) that the
object belongs to the class C
i
.
The interest of using possibility distribution is the denition of
measures that will estimate the quality of the extraction. We have
used possibility and necessity measures. These measures quantify
the imprecision and the uncertainty of the extraction. A possibility
measure is elaborated by considering, for any part of A, the coef-
cients of the elements of A that compose it (Dubois et al., 2007):

A sup
u2A
p
x
u 7
The value of G(A) corresponds to the element(s) of A having the
greatest possibility degree according to p
x
. It satises the following
max-decomposability characteristic property (Dubois et al., 2007):
PA [ B maxPA; PB 8
The possibility measure of an event estimates to what extent it
is possible. But it is insufcient to inform if an event will be real-
ized. Information about the complement event is useful. The neces-
sity measure can be dened as the impossibility of A

, the
complement of A (not A):
NA 1

A inf
uRA
1 p
x
u 9
Thus, P(A) corresponds to the degree with which it is possible
that A is true, N(A) as the degree with which it is certain that A is
true. The necessity measure is an evaluation of the certainty (Bou-
chon-Meunier and Marsala, 2003). By using these two measures,
the imprecision and the uncertainty of the object extraction pro-
cess is quantied and consequently its quality is estimated.
The following example claries these two measures of possibil-
ity and necessity. If C is the set of classes {Building; Road; Parking
lot}. We suppose that, after applying the fuzzy rule base, an object
x presents the following membership degrees {1/Building; 0.1/
Road; 0.5/prkg_lot}. The object x presents the possibility distribu-
tion illustrated by Fig. 7. The object is assigned to the building class
because it presents the greatest possibility degree. The certainty
degree of this assignment is obtained through the necessity
measure:
NC
Building
1

C
Building

1 max

C
Road
;

C
prkg lot

_ _ _ _
10
NC
Building
1

C
prkg lot
1 0:5 0:5 11
So, the certainty that the object x belong to the building class is
0.5. Thus, the object x is a building with a precision degree of 1 and
a certainty degree of 0.5. It is 50% certain that the object x is 100% a
building.
The conict between two classes regarding an object is dened
by the difference between its membership degrees to two classes.
The smaller is the difference, the higher is the conict. Ideally, a
well-extracted object will present very high precision and cer-
tainty degrees and a low to null conict with the other classes.
The following cases can be found:
- If a unique class C
i
such as p
X
(C
i
) = 1 exists, the class C
i
is the
most certain since N(C
i
) > 0 and N(C
j
) = 0 for all j i.
- If several C
i
are such as p
X
(C
i
) = 1, no decision can be taken
between these classes: the rule base does not have enough
knowledge to discriminate. However, if the classes with possi-
bility degrees equal to 1 correspond to a signicant class subset,
then the certainty of this subset is strictly positive.
- If p
X
(C
i
) = 1 everywhere, then there is a total uncertainty on the
class.
This information on certainty and precision of the object extrac-
tion is given with the object extraction results. Indeed, the pro-
Object x I
Comp
(x)
Road
F
u
z
z
y
a
g
g
r
e
g
a
t
i
o
n
o
p
e
r
a
t
o
r
Elongation
Index

Road
(object x)

NC

class A

L
(x)

NC
(x)
Non-compact shape Elongated shape
Compactness
Index
I
Elong
(x)
Class : Road Property :
Fig. 6. Objects membership degree to a class.
Building Road
Parking lot

1
0
Fig. 7. Possibility distribution.
I. Sebari, D.-C. He / ISPRS Journal of Photogrammetry and Remote Sensing 79 (2013) 171184 177
posed approach allows the organization of the extracted objects in
layers by class. Each object is described by a set of spectral, geo-
metric and contextual attributes and also by precision and cer-
tainty information related to the extraction quality.
2.3. Evaluation of the results
The evaluation of extracted objects was performed by compar-
ison with reference data representing ground truth, according to:
extraction accuracy assessment and geometric quality assessment.
For accuracy assessment, two indices were used rst to appre-
ciate the overall objects extraction with relation to reference data:
completeness and correctness. The completeness index represents
the percentage of the reference objects area that is extracted. The
correctness index is the percentage of the extracted objects area
that is correctly extracted.
Completeness
S
CE
S
R
12
Correctness
S
CE
S
E
13
With S
CE
being the objects area correctly extracted, S
R
the total area
of the reference objects, and S
E
being the total area of the extracted
objects.
We also produced error matrices according to percentage of the
area of objects. Correctness, completeness, Producers accuracy and
Users accuracy were generated. For man made classes, the accu-
racy assessment has also been performed according to the number
of objects for buildings and parking lots classes, and according to
object length for road class.
For geometric quality assessment, two indexes are used to ana-
lyze the geometric quality of extracted objects: area ratio and po-
sition error. The area ratio (R
A
) yields information on the
percentage of the area accurately extracted for each object. It is de-
ned as follows:
R
A

AReal Objet \ Extracted Objet
AReal Objet
14
where A is the area. The position error (E
p
) is related to the mean
distance between extracted objects and their corresponding objects
in a reference data. It is determined according to characteristic
points in the two data. For buildings, the characteristic points cor-
respond to building corners, and for roads, they are chosen along
road axes. The position error (E
p
) is determined by the formula:
E
p

n
k1

x
ext
x
ref

2
y
ext
y
ref

2
_
_ _
n
15
where n is the number of the characteristic points, (x
ext
, y
ext
) corre-
spond to extracted points coordinates and (x
ref
, y
ref
) are coordinates
of the corresponding characteristic point on the reference layer.
3. Application
3.1. Data and study area
The proposed approach was applied to Ikonos image of the city
of Sherbrooke (Canada) acquired on 20 May 2001. We selected
three subsets from Ikonos image presenting different land use (ob-
jects size and density) in order to evaluate the approach on differ-
ent urban environment. In addition, we selected another Ikonos
image acquired on November 2004 over Sherbrooke (Canada) to
evaluate the effectiveness of the approach and its transferability
on the future. We intentionally selected an unchanged zone from
the two dates to examine if the approach was able to extract the
same objects. This way, the eventual observed changes in the re-
sults will be attributed to the approach. The reference data (ground
truth) was generated by photo-interpretation of the studied image.
The Ikonos image contains four spectral bands (blue, green, red
and near-infrared) and a panchromatic band with spatial resolu-
tions of 4 m and 1 m, respectively. In order to mutually benet
from multispectral information and very high spatial resolution,
the four bands have been individually fused with the panchromatic
image. The image fusion method used is the one proposed by He
et al. (2004). This method allows preserving faithfully the spectral
aspect of the low resolution image while integrating the spatial
information of the high resolution image.
3.2. Results and discussion
Fig. 8 shows the results on three extracts. 80% of the total refer-
ence area was extracted (completeness index) (Table 5). The cor-
rectness rate was of 78% (Table 3). The confusion matrices of the
three study sites are given by Tables 46 according to classes of le-
vel II of Table 1. The results show that the correctness rates for the
natural classes are high between 84% and 90%. This is due to their
distinct spectral responses. The spectral rules were the most
important to discriminate the natural classes and to differentiate
them from the man made classes. The main source of errors is re-
lated to the photo-interpretation conducted to create the ground
truth: it was difcult during photo-interpretation to identify some
vegetation areas due to their low contrast. For the man made clas-
ses, the results are encouraging with 80% as mean correctness rate.
For building class, 80% of the total reference area was correctly ex-
tracted. According to the number of extracted building, the results
analysis has revealed completeness rates of 93% for site 1, 77% for
site 2 and 96% for site 3, and a mean of correctness rate of 93% (Ta-
ble 7). The auto-learning process has contributed to reach these
values. It has allowed the extraction of more building (For example,
8 correct buildings were extracted in site 2 through the auto-learn-
ing rules). Some buildings have been omitted by the approach be-
cause they not have the properties dened for the building class.
Even with the auto-learning process, their spectral responses did
not correspond to those recorded in the retrieved library. Another
reason of this omission is that the distinction of these buildings
with relation to their neighborhood was difcult during the stage
of the creation of the primitives (segmentation). They were fused
with their neighborhood and thus did not have the characteristics
of the building class objects. The objects incorrectly extracted as
buildings correspond to articial areas that have similar geometric
and contextual properties. The complete and accurate extraction of
buildings objects faces two major problems: the confusion of
buildings with the neighborhood and the heterogeneity of the
buildings roofs.
Considering the road network class, the correctness rates are
about 75%. The false classied areas from the road network repre-
sent less than 4%. An analysis of the results of the two separate
subclasses road and parking lot is performed (Table 7). For the road
class, 80% of the total length of the roads for the studied sites has
been extracted. Their elongated shape is an interesting property for
their extraction. For a homogenous road, the approach allows its
quasi-complete extraction. Some omissions are due to the hetero-
geneity of roads surface and to projected shadows on roads either
by trees or by buildings.
For parking lot class, the confrontation with the ground truth
has showed that the quality of the extraction depends on the stud-
ied site. For some sites, all lots were extracted while for some oth-
ers, some parking lots have been omitted (site 2). Two reasons can
be identied. First, the denition of the objects of this class in the
rule base uses the property large size object. The approach allowed
178 I. Sebari, D.-C. He / ISPRS Journal of Photogrammetry and Remote Sensing 79 (2013) 171184
the extraction of the objects respecting this characteristic, but in
site 2, parking lots are located in residential areas and are of small
sizes. The second reason is the segmentations results. In some
case, due to presence of cars, the shape of the segmented parking
lot becomes elongated instead of compact and can be classied
as road. Also, two adjacent road and parking lot objects can form
one segment since they are built with the same material and ex-
tracted as one object.
The objects of the building and road classes are extracted with
an average of 80% of their area (Fig. 9). The objects of the parking
lot class have less important rates (average of 60%). This is due to
the complexity and the heterogeneity of these objects, as well as
by the presence of details (for example cars). For the position error,
the buildings present high spreads (higher than 11 m in the case of
site 3). For roads, a maximal value of 6 m was observed. These
spreads are due to: in the case of road objects, the central axe of
road can be offset due to the modied boundaries by shadows of
trees or buildings; for buildings, the corners of extracted objects
Fig. 8. Application of the proposed approach on Sherbrookes Ikonos image extracts (Site 1, Site 2, Site 3) (a) image extracts (b) extracted objects (c) reference map.
Table 3
Extraction quality index.
Indices Site 1 Site 2 Site 3
Completeness (%) 79 83 78
Correctness (%) 81 83 73
I. Sebari, D.-C. He / ISPRS Journal of Photogrammetry and Remote Sensing 79 (2013) 171184 179
can be offset due to incorrect delimitation by the approach and the
confusion between some roofs and fronts.
3.2.1. Shadow extraction
In order to evaluate the shadow detection performance, we
have compared extracted shadow areas to reference shadow areas
(created by photo-interpretation) for the extract 3 (Fig. 10). 92% of
shadow areas were well extracted. Producers accuracy (about 28%)
Table 5
Confusion matrix site 2.
Extracted objects Reference data
Vegetation Bare Soil Water Building Road network Correctness rate Producers accuracy
Vegetation 84 0 3 2 2 81 19
Bare soil 0 0 0 0 0
Water 0 0 90 0 0 99 1
Building 0 0 0 78 0 81 19
Road network 2 0 0 4 71 72 28
Completeness 84 90 78 71
Users accuracy 16 10 21 29
The values in the error matrix are presented in % of objects area.
Table 4
Confusion matrix site 1.
Extracted objects Reference data
Vegetation Bare soil Water Building Road network Correctness rate Producers accuracy
Vegetation 84 0 0 0 1 84 16
Bare soil 0 0 0 0 0
Water 0 0 0 0 0
Building 1 0 0 80 0 84 16
Road network 3 0 0 2 74 78 21
Completeness 84 80 74
Users accuracy 16 20 26
The values in the error matrix are presented in % of objects area.
Table 6
Confusion matrix site 3.
Extracted objects Reference data
Vegetation Bare Soil Water Building Road network Correctness rate Producers accuracy
Vegetation 85 0 0 0 0 71 29
Bare soil 0 0 0 0 0
Water 0 0 0 0 0
Building 0 0 0 77 2 78 22
Road network 0 0 0 1 81 75 25
Completeness 85 0 0 77 81
Users accuracy 15 0 0 23 19
The values in the error matrix are presented in % of objects area.
Table 7
Road and parking lot classes objects extraction quality indices.
Indices Road Parking lots
Site 1 Site 2 Site 3 Site 1 Site 2 Site 3
Completeness (%) 84 96 90 100 56 70
Correctness (%) 90 86 75 100 56 77
% of road length % of number of objects
0
20
40
60
80
100
Site 1 Site 2 Site 3 Site 1 Site 2 Site 3
A
r
e
a

R
a
t
i
o

(
%
)

Building Objects
0
2
4
6
8
10
12
P
o
s
i
t
i
o
n

E
r
r
o
r

(
m
)
Building objects

Fig. 9. Geometric quality assessment.
Extracted shadow Real shadow (by photo-interpretation)
Fig. 10. Comparison between extracted and reference shadow areas.
180 I. Sebari, D.-C. He / ISPRS Journal of Photogrammetry and Remote Sensing 79 (2013) 171184
is due in major part to the self shadows. The Users accuracy is esti-
mated to 8%. It is due to some low albedo materials. The advantage
of using fuzzy logic is that these areas present membership degrees
to shadow and also to the real class. The use of others attributes
(spectral, geometric and contextual attributes) minimizes the er-
rors of urban objects extraction. Even though detecting shadow
areas was useful for the identication of elevated objects (specially
building), some problems were observed: the projected shadow
has prevented the extraction of some parts of roads and parking
lots; the self shadow (the part of the object that is not illuminated,
i.e., the faade of the building) is not always detected, which has
cause in some cases the incorrect extraction of the geometric form
of objects.
3.2.2. Membership degrees
The objects membership degrees to nal classes are generally
high (greater than 0.6). The mean degrees are about 0.8 for the veg-
etation class, 0.8 for the building class, 0.85 for the road class and
0.65 for the parking lot class. The membership conict analysis has
revealed that the objects of natural classes present membership
degrees only to their corresponding class, whereas non-natural ob-
ject classes can present membership degrees to more than one
class. The objects that have a higher membership degree belong
to the appropriate class. The analysis of the precision and certainty
degrees of their extraction has revealed that more than 60% of the
objects extracted have a precision degree higher than 0.6 and a cer-
tainty degree higher than 0.8 (Fig. 11). For the certainty degrees,
they are higher than 0.8 and can reach a value of 1. Typical ob-
jects present high degrees of precision and certainty and low or
null conicts with others classes.
For all the classes, there are objects that present membership
degree equal to 1 which means that the class rule is descriptive
of the studied object. Also, no object has presented membership
degree equal to 1 for all the classes. We can conclude that the rule
base contains enough knowledge to discriminate between the
objects.
3.2.3. Thresholds determination
In order to evaluate the adopted methodology for automatically
determinate attributes thresholds, we have conducted a sensitiv-
ity analysis by introducing variations to the determined thresholds
values. The corresponding rule is then applied and the correctness
rate, Producers and users accuracies are determined. Table 8 pre-
sents the example of results of this analysis in the case of bright-
ness index (S) used to detect shadow areas. Variations DS of 10
were added and subtracted from the thresholds value.
The analysis shows that the use of obtained threshold yields the
best results (correctness rate of 92% with producers accuracy of
28%). Using S + DS, the correctness rate is the same but the pro-
ducers accuracy is greater (40%) while with S DS, even if the
producers accuracy is lower, the Users accuracy is still important
(23%).
The determination of thresholds is very sensitive task. We
think that using threshold in rigid way is not adequate to our case
(for example, object with a area of 700 m
2
is considered large and
another with an area of 701 is not?). So, using fuzzy threshold is in
our opinion more interesting. The proposed approach uses many
thresholds which are automatically derived. But, we should precise
that the automatically determined thresholds are used as approx-
imate values which will be useful to establish the membership
functions. The proposed methodologies for automatic detection
are not perfect. The imprecision related to the determination of
threshold is considered by fuzzy logic. Indeed, the imprecision of
adopted thresholds is considered by the used membership func-
tions: each attribute threshold is considered as the crossover point
of the corresponding membership function. A transition interval,
dened with relation to the attributes threshold, reects a gradual
membership and minimizes the imprecision related the attributes
thresholds determination (Fig. 5).
All the thresholds are important because they are useful to
establish the membership functions. During rule formulation, the
incorrect determination of thresholds may be cause the incorrect
classication of an object if the difference between the correct
and the incorrect values of threshold is important. If not, the object
will have a low membership degree to its real class.
3.2.4. Transferability test
We have also conduct a transferability test in order to verify if,
for the same zone, the approach will extract the same objects with-
out any changes in the rule base. The transferability of the extrac-
tion approach was evaluated on two subsets from two different
Ikonos images over the same zone (Fig. 12). The same rule base
was applied on two subsets. The decision rules are not modied
between the two images and no new rule was used. The thresholds
of the attributes dependant of the image are automatically deter-
mined by the approach. The objects were well extracted on both
images. They have the same assignations to the classes on the
2001 and 2004 extracts. The only differences are observed at object
areas and also at shadow zones.
3.2.5. Some limitations
For very large image, the application of the proposed approach
will give satised results when extracting typical objects. But, it
Site 1 Site 2 Site 3
0.0
0.2
0.4
0.6
0.8
1.0
D
e
g
r
e
e

o
f

p
r
e
c
i
s
i
o
n

Site 1 Site 3 Site 2
0.0
0.2
0.4
0.6
0.8
1.0
D
e
g
r
e
e

o
f

C
e
r
t
a
i
n
t
y

Fig. 11. Statistics on (a) the precision degrees and (b) certainty degrees of extracted objects on the three study sites.
Table 8
Sensitivity analysis of the brightness index thresholds.
Real shadow (reference data)
Correctness rate (%) Users accuracy (%) Producers accuracy (%)
Extracted shadow
S 92 8 28
S DS 77 23 12
S + DS 93 7 40
I. Sebari, D.-C. He / ISPRS Journal of Photogrammetry and Remote Sensing 79 (2013) 171184 181
can present some limitations according to the difculty and the
complexity of the urban environment. The typical objects can
be extracted without problems (except the exact geometric shape).
A typical object can be for example a building that presents a com-
pact form and a cast shadow. Some buildings in dense zone are
contiguous and it will be difcult to extract individual building.
Also, for roads, on some images, they are hided by building and
shadows. Their extraction is not going to be exhaustive and com-
plete. The extraction results will depend on the segmentation re-
sults. The used algorithm used an adaptive threshold when
segmenting each segment. The algorithm presents satisfactory re-
sults. But, in dense environment, the algorithm may confuse some
objects presenting same spectral responses and consequently pro-
vide one segment.
4. Conclusion
This paper presents a solution for automatic extraction of geo-
graphic urban information from VHSR multispectral images. The
proposed approach is object-based image analysis. It does not
use no auxiliary data or parameter to introduce and does not re-
quire any examples for learning process. The extraction is based
on simple, objective, and transferable rules. Formulating the rules
is performed through the translation of the knowledge used by
the photo-interpreters in order to interpret urban objects from
VHSR image. It used different types of properties. The thresholds
of the attributes, which are dependent on the images and used in
the rules, are automatically determined. Integrating the fuzzy logic
and the possibility theories into the rule base has contributed to
manage the complexity of the subject studied. It allowed to reason
with imprecise and uncertain knowledge and to give information
on the quality of the extracted objects. The used fuzzy rule base
has the benet of not requiring any order of application for the
rules. The objects respond to the different rules with satisfaction
degrees ranging from 0 to 1. The order of application has no effect
on the nal results. The nal decision is taken at the end according
to the precision, certainty and conict degrees observed on the dif-
ferent classes. The results obtained translate the pertinence of the
rules. A global extraction rate of 80% was observed. The correctness
rates obtained for the building, road and parking lot classes are of
81%, 75% and 60%, respectively.
The challenge was great given the complexity of the studied
area and the particularities of the used images. The automatic
extraction approach proposed has displayed a transferability
power. One of the limitations of this approach is the conformity
of the geometric shape of the extracted objects to the reality.
Enhancing this aspect would be useful for the future integration
of the objects extracted in spatial databases. It would also be inter-
esting to study the application of the approach on high spectral sa-
tellite images in order to study the effect of introducing additional
bands on the improvement of mapping accuracy.
Acknowledgments
This work was supported by the NSERC (the Natural Sciences
and Engineering Research Council of Canada) and the PCBF (Prog-
amme Canadien de Bourses de la Francophonie) of CIDA (Canadian
International Development Agency).
References
Baatz, M., Schpe, M., 2000. Multiresolution segmentation an optimization
approach for high quality multi-scale image segmentation. In: Strobl, J.,
Fig. 12. Transferability study of the approach (a) Ikonos image subset of 2001, (b) Ikonos image subset of 2004, (c) extracted objects from Ikonos 2001, (d) extracted objects
from Ikonos 2004.
182 I. Sebari, D.-C. He / ISPRS Journal of Photogrammetry and Remote Sensing 79 (2013) 171184
Blaschke, T., Griesebner, G. (Eds.), Angewandte Geographische Informations-
Verarbeitung XII. Wichmann Verlag, Karlsruhe, pp. 1223.
Baltsavias, E.P., 2004. Object extraction and revision by image analysis using
existing geodata and knowledge: current status and steps towards operational
systems. ISPRS Journal of Photogrammetry and Remote Sensing 58 (34), 129
151.
Bauer, T., Steinnocher, K., 2001. Per-parcel land use classication in urban areas
applying a rule-based technique. GIS 6 (1), 2427.
Blaschke, T., Lang, S., Lorup, E., Strobl, J., Zeil, P., 2000. Object-oriented image
processing in an integrated GIS/remote sensing environment and perspectives
for environmental applications. In: Cremers, A., Greve, K. (Eds.), Environmental
Information for Planning, Politics and the Public, vol. 2. Metropolis Verlag,
Marburg, pp. 555570.
Blaschke, T., 2010. Object-based image analysis for remote sensing. ISPRS Journal of
Photogrammetry and Remote Sensing 65 (1), 216.
Benz, U.C., Hofmann, P., Willhauck, G., Lingenfelder, I., Heynen, M., 2004. Multi-
resolution, object-oriented fuzzy analysis of remote sensing data for GIS-ready
information. ISPRS Journal of Photogrammetry and Remote Sensing 58 (34),
239258.
Biehl, L., Landgrebe, D., 2002. MultiSpec: a tool for multispectral hyperspectral
image data analysis. Computers and Geosciences Archive 28 (10), 11531159.
Bouchon-Meunier, B., Marsala, C., 2003. Logique oue, principes, aide la dcision.
Hermes science publications, Paris.
Bouziani, M., Gota, K., He, D.-C., 2010. Automatic change detection of buildings in
urban environment from very high spatial resolution images using existing
geodatabase and prior knowledge. ISPRS Journal of Photogrammetry and
Remote Sensing 65 (1), 143153.
Caloz, R., Pointet, A., 2003. Analyse comparative de la classication contextuelle et
du maximum de vraisemblance: synthse et cas dtude. Tldtection 3 (23
4), 311322.
Campbell, J.B., 2007. Introduction to Remote Sensing. Guildford Press, New York.
Caprioli, M., Tarantino E., 2003. Urban features recognition form VHR satellite data
with an object-oriented approach. International Symposium of Remote Sensing
(ISPRS) Commission IV Joint Workshop Challenges in Geospatial Analysis,
Integration and Visualization II. <http://www.igf.uniosnabrueck.de/
mitarbeiter/schiewe/papers/24.pdf>.
Demarchi, L., Chan, J.C.-W., Ma, J., Canters, F., 2012. Mapping impervious surfaces
from superresolution enhanced CHRIS/Proba imagery using multiple
endmember unmixing. ISPRS Journal of Photogrammetry and Remote Sensing
72, 99112.
Donnay, J.-P., Barnsley, M.J., Longley, P.A., 2001. Remote Sensing and Urban
Analysis. Taylor and Francis, London.
Dubois, D., Esteva, F., Godo, L., Prade, H., 2007. Fuzzy-set based logicsan history-
oriented presentation of their main developments. In: Gabbay, D.M., Woods, J.
(Eds.), Handbook of the History of Logic. Elsevier, pp. 325449.
Gao, F., Brown de Colstoun, E., Ma, R., Weng, Q., Masek, J.G., Chen, J., Pan, Y., Song, C.,
2012. Mapping impervious surface expansion using medium-resolution
satellite image time series: a case study in the Yangtze River Delta, China.
International Journal of Remote Sensing 33 (24), 76097628.
Forestier, G., Puissant, A., Wemmert, C., Ganarski, P., 2012. Knowledge-based
region labeling for remote sensing image interpretation. Computers,
Environment and Urban Systems 36 (5), 470480.
Han, J., Chi, K., Yeon, Y., 2005. Land cover classication of IKONOS multispectral
satellite data: neuro-fuzzy, Neural network and maximum likelihood methods.
In: S

le zak, D. et al. (Eds.), Proc. 10th International Conference of Rough Sets,

Fuzzy Sets, Data Mining, and Granular Computing, Regina, Canada, 31 August3
September, pp. 251262.
Hay, G.J., Castilla, G., 2006. Object-based image analysis : strengths, weaknesses,
opportunities and threats (SWOT). In: 1st International Conference on Object-
based Image Analysis (OBIA 2006), Salzburg, Austria, July 45, 2006.
International Archives of the Photogrammetry, Remote Sensing and Spatial
Information Science. Com VI, WG VI/4, 5p.
He, D.-C., Wang, L., Amani, M., 2004. A new technique for multi-resolution image
fusion. In: Proc. IGARSS-International Geoscience and Remote Sensing
Symposium, Anchorage, Alaska, 1926 September, 4 p.
Herold, M., Gardner, M.E., Roberts, D.A., 2003. Spectral resolution requirements for
mapping urban areas. IEEE Transactions on Geoscience and Remote Sensing 41
(9), 19071919.
Herold, M., Roberts, D.A., Gardner, M.E., Dennison, P.E., 2004. Spectrometry for
urban area remote sensing development and analysis of a spectral library
from 350 to 2400 nm. Remote Sensing of Environment 91 (34), 304319.
Hofmann, P., 2001a. Detecting buildings and roads from IKONOS data using
additional elevation information. GIS 6 (1), 2833.
Hofmann, P., 2001b. Detecting urban features from IKONOS data using an object-
oriented approach. In: Proc. 1st Annual Conference of the Remote Sensing and
Photogrammetry Society (RSPS2001), London, 1214 September, 13 p.
Hofmann, P., Strobl, J., Blaschke, T., Kux, H.J., 2008. Detecting informal settlements
from QuickBird data in Rio de Janeiro using an object-based approach. In:
Blaschke, T., Lang, S., Hay, G.J. (Eds.), Object Based Image Analysis. Springer,
Heidelberg, Berlin, New York, pp. 531553.
Holland, D., Boyd, D.S., Marshall, P., 2006. Updating topographic mapping in Great
Britain using imagery from high-resolution satellite sensors. ISPRS Journal of
Photogrammetry and Remote Sensing 60 (3), 212223.
Hu, X., Weng, Q., 2011. Impervious surface area extraction from IKONOS imagery
using an object-based fuzzy method. Geocarto International 26 (1), 320.
Huiping, H., Bingfang, W., Jinlong, F., 2003. Analysis to the relationship of
classication accuracy segmentation scale image resolution. In: Proc. IGARSS
IEEE, Toulouse, July 2003, 3p.
Ishibuchi, H., Ken Nozaki, K., Tanaka, H., 1992. Pattern Classication by Distributed
Representation of Fuzzy Rules. In: Proc. IEEE International Conference on Fuzzy
Systems, San Diego, 812 Mars, pp. 506515.
Jacquin, A., Misakova, L., Gay, M., 2008. A hybrid object-based classication
approach for mapping urban sprawl in periurban environment. Landscape and
Urban Planning 84 (2), 152165.
Jensen, J.R., 2005. Introductory Digital Image Processing. Pearson Prentice Hall,
Upper Saddle River, New Jersey.
Jimnez, L.O., Rivera-Medina, J.L., Rodrguez-Daz, E., Arzuaga-Cruz, E., Ramrez-
Vlez, M., 2005. Integration of spatial and spectral information by means of
unsupervised extraction and classication for homogenous objects applied to
multispectral and hyperspectral data. IEEE Transactions on Geoscience and
Remote Sensing 43 (4), 844851.
Kettig, R.L., Landgrebe, A., 1976. Classication of Multispectral Image Data by
Extraction and Classication of Homogeneous Objects. IEEE Transactions on
Geoscience Electronics GE-14 (1), 1926.
Kressler, F.P., Bauer, T.B., Steinnocher, K.T., 2001. Object-oriented per-parcel land
use classication of very high resolution images. In: Proc. IEEE/ISPRS Joint
Workshop on Remote Sensing and Data Fusion Over Urban Areas, Rome, Italy,
89 November, 4p.
Kux, H.J.H., Arajo, E.H.G., 2008. Object-based image analysis using QuickBird
satellite images and GIS data, case study Belo. Horizonte (Brazil). In: Blaschke,
T., Lang, S., Hay, G.J. (Eds.), Object Based Image Analysis. Springer, Heidelberg,
Berlin, New York, pp. 571588.
Lang, S., Blaschke, T., 2006. Bridging remote sensing and GIS what are the main
supporting pillars? International Archives of Photogrammetry, Remote Sensing
and Spatial, Information Sciences XXXVI-4/C42.
Lee, D.S., Shan, J., Bethel, J.S., 2003. Class-Guided building extraction from Ikonos
imagery. Photogrammetric Engineering and Remote Sensing 69 (2), 143150.
Longley, P.A., Barnsley, M.J., Donnay, J.-P., 2001. Remote sensing and urban analysis:
a research agenda. In: Donnay, J.-P., Barnsley, M.J., Longley, P.A. (Eds.), Remote
Sensing and Urban Analysis. Taylor and Francis, London, New York, pp. 243
255.
Longley, P.A., Goodchild, M.F., Maguire, D.J., Rhind, D.W., 2005. Geographic
Information Systems and Science. Wiley J. & Sons, Chichester.
Lu, D., Weng, Q., 2007. A survey of image classication methods and techniques for
improving classication performance. International Journal of Remote Sensing
28 (5), 823870.
Marchesi, A., Colombo, R., Valentini, P., 2006. Application of high spatial resolution
satellite imagery for urban environment mapping. In: Proc. 1st International
Conference on Object-based Image Analysis (OBIA 2006), Salzburg, Austria, July
45, 2006, 4p.
Mathieu, R., Freeman, C., Aryal, J., 2007. Mapping private gardens in urban areas
using object-oriented techniques and very high-resolution satellite imagery.
Landscape and Urban Planning 81 (3), 179192.
Meinel, G., Neubert, M., Reder, J., 2001. The potential use of very high resolution
satellite data for urban areas rst experiences with Ikonos data, their
classication and application in urban planning and environmental monitoring.
In: Jrgens, Carsten (Eds.): Remote Sensing of Urban Areas, Regensburg, pp.
196205.
Mittelberg, B., 2002. Pixel versus object: a method comparison for analyzing urban
areas with VHR data. eCognition Application Note 3 (2), 12.
Mo, D.-K., Lin, H., Li, J., Sun, H., Xiong, X.-J., 2007. Design and implementation of a
high spatial resolution remote sensing image intelligent interpretation system.
Data Science Journal 6, S445S452.
Myint, S.W., Gober, P., Brazel, A., Grossman-Clarke, S., Weng, Q., 2011. Per-pixel vs.
object-based classication of urban land cover extraction using high spatial
resolution imagery. Remote Sensing of Environment 115 (5), 11451161.
Navulur, K., 2007. Multispectral Image Analysis using the Object-Oriented
Paradigm. CRC Press, Boca Raton, FL.
Nobrega, R.A., OHara, C.G., Quintanilha, J.A., 2008. An object-based approach to
detect road features for informal settlements near Sao Paulo, Brazil. In:
Blaschke, T., Lang, S., Hay, G.J. (Eds.), Object Based Image Analysis. Springer,
Heidelberg, Berlin, New York, pp. 589607.
Puissant, A., Weber, C., Sheeren, D., Wemmert, C., Gancarsk, P., 2006. Amlioration
des connaissances sur lenvironnement urbain : intrt de lintgration de
rgles dans les procdures de classications. In: Proc. Colloque International
Interactions Nature-Socit : analyses et modles, La Baule, France, 36 may,
pp. 15.
Paine, D.P., Kiser, J.D., 2003. Aerial Photography and Image Interpretation. Wiley, J.,
and Sons, Inc., Hoboken, New Jersey.
Rego, L.F.G., Koch, B., 2003. Automatic classication of land cover with high
resolution data of the Rio De Janeiro City Brazil comparison between pixel and
object classication. The International Archives of the Photogrammetry, Remote
Sensing and Spatial Information Sciences XXXIV-7/W9, 153157.
Repaka, S.R., Truax, D.D., Kolstad, E., OHara, C.G., 2004. Comparing spectral and
object based approaches for classication and transportation feature extraction
from high resolution multispectral imagery. In: Proc. of the American Society of
Photogrammetry and Remote Sensing, Denver, May 2004, 11p.
Sebari, I., He, D.-C., 2009. Approach to nonparametric cooperative multiband
segmentation with adaptive threshold. Applied Optics, Optical Society of
America 48 (20), 39673978.
I. Sebari, D.-C. He / ISPRS Journal of Photogrammetry and Remote Sensing 79 (2013) 171184 183
Shackelford, A.K., Davis, C.H., 2003. A combined fuzzy pixel-based and object-based
approach for classication of high-resolution multispectral data over urban
areas. IEEE Transactions on Geoscience and Remote Sensing 41 (10), 2354
2363.
Tsai, Y.H., Stow, D., Weeks, J., 2011. Comparison of object-based image analysis
approaches to mapping new buildings in Accra, Ghana using multi-temporal
QuickBird satellite imagery. Remote Sensing 3 (12), 27072726.
Tso, B., Mather, P.M., 2001. Classication Methods for Remotely Sensed Data. Taylor
and Francis, New York.
Walter, V., 2004. Object-based classication of remote sensing data for change
detection. ISPRS Journal of Photogrammetry and Remote Sensing 58 (34), 225
238.
Weng, Q., 2012. Remote sensing of impervious surfaces in the urban areas:
requirements, methods, and trends. Remote Sensing of Environment 117, 34
49.
Xu, H., 2013. Rule-based impervious surface mapping using high spatial resolution
imagery. International Journal of Remote Sensing 34 (1), 2744.
Zack, G., Rogers, W., Latt, S., 1977. Automatic measurement of sister chromatid
exchange frequency. Journal of Histochemistry and Cytochemistry 25 (7), 741
753.
Zadeh, L., 2003. Fuzzy logic, neural networks, and soft computing. Fuzzy Systems 37
(3), 7784.
184 I. Sebari, D.-C. He / ISPRS Journal of Photogrammetry and Remote Sensing 79 (2013) 171184