PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
October 2014
963
Object-Based Hyperspectral Classification of
Urban Areas by Using Marker-Based
Hierarchical Segmentation
Davood Akbari, Abdolreza Safari, and Saeid Homayouni
Abstract
An effective approach to spectral-spatial classification has
been achieved using Hierarchical SEGmentation (
HSEG
) by
Tarabalka et al. (2009 and 2010). Our goal is to improve
this approach to the classification of hyperspectral images
in urban areas. The first step of our proposed method is to
segment the spectral images using a novel marker-based
HSEG
method. The spatial features from segmented images
are then extracted. Spatial information such as the area,
entropy, shape, adjacency, and relation features constitute
the components of feature space. Last, using both spectral
and spatial features, the image objects are classified by a
support vector machine (
SVM
) classifier. Three image data-
sets were used to test this method. The results of our ex-
periment indicate that the main advantage of the proposed
method, compared to the previous
HSEG
-based approach,
is that it increases classification accuracy by selecting the
appropriate contextual features of different objects.
Introduction
Imaging spectroscopy, also known as hyperspectral imaging,
is concerned with the measurement, analysis, and inter-
pretation of spectra acquired from either a given scene or
a specific object at a short, medium, or long distance by a
satellite sensor (Shippert, 2004). Recent improvements in the
resolution of spatial, spectral, and radiometric spectrometer
images have led to the development of new methods for land
cover classification. Among the many studies that have been
published on this topic, two main categories of techniques
have been established: the spectral (i.e., pixel-based) tech-
niques and the spectral-spatial (i.e., object-based) techniques.
The problem with the first category of techniques is that these
techniques are only usable to spectral patterns. Many scien-
tists devoted to multidimensional data analysis have, for good
reason, emphasized the importance of analyzing both spatial
and spectral patterns. These techniques have been studied
from various perspectives. For instance, Landgrebe (2003)
discussed several ways to refine the results obtained through
pixel-based techniques in multispectral imaging. This is nor-
mally achieved by adding a second step in which a particular
spatial context is used. Such contextual classification can also
be applied to hyperspectral images (Jimenez
et al
., 2005).
Because complex and diverse land cover types are preva-
lent within urban environments, the classification of high-res-
olution hyperspectral imagery is a difficult task (Lu
et al
.,
2010). For example, because the Meadow
and Tree
classes are
spectrally similar to one another, they are often misclassified.
Traditional classification methods, which only take spectral
information into account, are unable to differentiate between
these classes with a high degree of accuracy. Methods that can
extract spatial information in addition to spectral information
are crucial for producing more accurate land cover maps in
urban areas (Carleer and Wolff, 2006; Jensen, 2004; Shackel-
ford and Davis, 2003).
In early studies on spectral-spatial image classification,
the spectral information extracted from neighborhoods,
defined by either fixed windows (Camps-Valls
et al
., 2006;
Paneque-Galvez
et al
., 2013) or morphological profiles (Fauvel
et al
., 2008; Huang and Zhang, 2011), was used to classify
and label each pixel. Segmentation techniques are a powerful
tool for defining spatial dependencies. Segmentation is an
exhaustive partitioning of the input image into homogeneous
regions (Gonzalez and Woods, 2002). The advantages of
using segmentation for distinguishing spatial structures from
one another are also discussed by Tarabalka
et al
. (2010 and
2010b). Different unsupervised techniques, such as watershed
techniques, partitioning clustering, and Hierarchical SEGmen-
tation
(
HSEG
), have been evaluated for their capacity to accu-
rately segment hyperspectral data (Tarabalka
et al
., 2010a).
The
HSEG
method is state-of-the-art for hyperspectral image
analysis (Tilton, 2010). It successfully integrates the spatial
and spectral information in a two-step procedure. In the first
step, the homogenous areas are segmented at their maximum
details, and then, by grouping the spectrally similar but spa-
tially disjointed regions, larger and more uniform objects are
created (Tilton, 1998).
An alternative way to achieve accurate segmentation is to
perform a marker-based segmentation (Gonzalez and Woods,
2002; Soille, 2003). The idea behind this approach is to select
either one or several pixels that belong to each spatial ob-
ject. Each spatial object is often referred to as either a region
seed, or a marker of the corresponding region. These regions
then grow from the selected seeds such that every region
in the resulting segmentation map is associated with one
region seed. Marker-based segmentation significantly reduces
over-segmentation and has, as a result, led to a better accura-
cy rate. Tarabalka
et al
. (2010) proposed using the estimation
probabilities obtained by the pixel-based
SVM
classification
for selecting the most reliable classified pixels as markers (i.e.,
as seeds of spatial regions). In another study, Tarabalka
et al
.
(2010) used three classifiers for the marker selection
Davood Akbari and Abdolreza Safari are with the Department
of Surveying and Geomatics Engineering, College of Engineer-
ing, University of Tehran, Tehran, Iran (
.
Saeid Homayouni is with the Department of Geography, Uni-
versity of Ottawa, Ottawa, Canada.
Photogrammetric Engineering & Remote Sensing
Vol. 80, No. 10, October 2014, pp. 963–970.
0099-1112/14/8010–963
© 2014 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.80.10.963
