Attribute Profiles on Derived Features for
Urban Land Cover Classification
Bharath Bhushan Damodaran, Joachim Höhle, and Sébastien Lefèvre
Abstract
This research deals with the automatic generation of 2D land
cover maps of urban areas using very high resolution multi-
spectral aerial imagery. The appropriate selection of classifier
and attributes is important to achieve high thematic accura-
cies. In this paper, new attributes are generated to increase the
discriminative power of auxiliary information provided by re-
mote sensing images. The generated attributes are derived from
the vegetation index and elevation information using morpho-
logical attribute profiles. The extended experimental evaluation
and comparison of attribute profile-based mapping solutions
is conducted to derive the optimal combinations of attributes
required for classification and to understand the genericity of
attributes on a range of classifiers, i.e., various combinations
of attributes and classifiers. Experimental results with two high
resolution images show that the proposed attributes derived on
auxiliary information outperform the existing attribute profiles
computed on original image and its principal components.
Introduction
Nowadays, due to the advances in the sensor technology, it is
possible to acquire very high-resolution satellite or airborne
multispectral imagery to generate accurate urban land cover
maps. However, analyzing very high-resolution imagery is
complex, mainly due to the diversity in the size and shape
of objects, high spectral and spatial variations, and objects
composed of similar type of materials (Foody, 2000). Urban
areas are of special interest because they change all the time
and the maps should be updated in short intervals (Cao
et al
.,
2015). Also, small objects like walls, hedges, and trees should
be detected. Machine learning techniques or supervised clas-
sifiers are used to translate images into useful information in
form of a thematic map. The thematic accuracy of land cover
maps depends very much on the type of landscape, the input
data, the class attributes, and on the type of the classification
method used (Damodaran and Nidamanuri, 2014a; Thomas
et
al
., 2003). The type of class attributes has a significant impact
on the performance of the underlying classifier. Therefore,
derivation and appropriate use of both simple and advanced
attributes are of prime importance for generating accurate 2D
thematic maps from very high-resolution imagery.
The image classification is generally carried out using the
spectral characteristics of remotely sensed images. Howev-
er, this information alone is not sufficient to obtain accurate
thematic maps. On the other hand, it is essential to derive the
spatial contextual attributes to incorporate the neighboring re-
lation among the pixels (Blaschke, 2010; Salehi
et al
., 2012b;
Damodaran
et al
., 2015). Several spatial-based attributes such
as texture and geometrical attributes are derived to account for
the neighborhood information in high-resolution imagery. The
combination of these attributes has significantly increased the
classification accuracy. The commonly used texture attribute
is described by means of gray-level co-occurrence matrix,
mean, standard deviation, entropy, and contrast (Trias-Sanz
et
al
., 2008). The geometrical attributes are derived based on the
contours (boundaries) and regions in the image. The common-
ly used attributes are convexity, perimeter, compactness, and
area (De Martinao
et al
., 2003; Du
et al
., 2015; Inglada, 2007).
Recently, operators based on multi-scale modeling by mathe-
matical morphology (
MM
) have been employed to extract the
geometrical informative attributes from high-resolution urban
imagery (Dalla Mura
et al
., 2010; Du
et al
., 2015). The attri-
bute profiles (
AP
) are built from morphological operators that
provide multi-level or multi-scale geometrical characteriza-
tion of very high-resolution imagery (Dalla Mura
et al
., 2010;
Aptoula
et al
., 2016). These attribute profiles are a powerful
model to increase the discrimination between the land cover
classes. The concept of the APs with all its modifications and
generalizations is explained in Ghamisi
et al
. (2015). Several
other studies were carried out to demonstrate the potential of
attribute profiles with multi- or hyperspectral imagery using
single classification methods (Benediktsson
et al
., 2005; Fau-
vel et
al
., 2008; Pedergnana
et al
., 2010; Ghamisi
et al
., 2014).
Apart from the intensities of the original images or ortho-
images, auxiliary information such as the normalized differ-
ence vegetation index (
NDVI
), the digital surface model (
DSM
)
or the normalized digital surface model (
nDSM
) are considered
as important (widely used) attributes for remote sensing image
classification (Elshehaby and Taha, 2009; Salehi
et al
., 2012a;
Höhle and Höhle, 2013). Several studies in literature high-
lighted the importance of
DSM
and
NDVI
attributes to provide
discriminative information to distinguish between the classes
in very high-resolution imagery (Salehi
et al
., 2012a; Sampath
and Shan, 2007). The height above ground (
nDSM
) derived
from filtering the
DSM
was considered as even more important
in high-resolution urban imagery (Höhle, 2013). However, the
performance improvement of the thematic map using only
the above information might be limited, and sometimes it
might not be as effective as expected. Recently, Tokarczyk
et
al
. (2015) showed that the
NDVI
attribute was not a useful at-
tribute for the classification of high resolution urban imagery,
since it does not provide any additional information. Thus,
it is essential and necessary to provide an alternative way to
utilize the
NDVI
attribute for high resolution urban image clas-
sification, which could benefit many urban mapping applica-
tions. In this paper, we propose a methodological framework
to utilize the
NDVI
attribute for high-resolution urban image
classification by incorporating the multi-level spatial con-
textual information from the
NDVI
attribute. The multi-level
characterization of the
NDVI
attribute is generated by means of
Bharath Bhushan Damodaran & Sébastien Lefèvre are with Univ-
er-sité de Bretagne-Sud, UMR 6074, IRISA, Vannes 56000, France.
Joachim Höhle is with Aalborg University, Department of
planning, Skibbrogade 3, DK 9000 Aalborg, Denmark
(
)
Photogrammetric Engineering & Remote Sensing
Vol. 83, No. 3, March 2017, pp. 183–193.
0099-1112/17/183–193
© 2017 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.83.3.183
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
March 2017
183
