130 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 50, NO. 1, JANUARY 2012
A Genetic Fuzzy-Rule-Based Classifier for Land
Cover Classification From Hyperspectral Imagery
Dimitris G. Stavrakoudis, Georgia N. Galidaki, Ioannis Z. Gitas, and John B. Theocharis, Member, IEEE
Abstract—This paper proposes the use of a genetic fuzzy-rule-
based classification system for land cover classification from
hyperspectral images. The proposed classifier, namely, Feature
Selective Linguistic Classifier, is constructed through a three-stage
learning process. The first stage produces a preliminary fuzzy rule
base in an iterative fashion. During this stage, a local feature selec-
tion scheme is employed, designed to guide the genetic evolution,
through the evaluation of deterministic information about the rel-
evance of each feature with respect to its classification ability. The
structure of the model is then simplified in a subsequent postpro-
cessing stage. The performance of the classifier is finally optimized
through a genetic tuning stage. An extensive comparative analysis,
using an Earth Observing-1 Hyperion satellite image, highlights
the quality advantages of the proposed system, when compared
with nonfuzzy classifiers, commonly employed in hyperspectral
classification tasks.
Index Terms—AdaBoost, evolutionary algorithms (EAs),
genetic fuzzy-rule-based classification systems (FRBCSs)
(GFRBCSs), genetic tuning, hyperspectral image classification,
local feature selection, remote sensing.
I. I NTRODUCTION
T
HE GROWING development and availability of hyper-
spectral sensor technologies over the past two decades
have provided new possibilities in improving the accuracy of
land cover classification from satellite images. Hyperspectral
sensors collect several (typically 200 or more) narrow spectral
bands from the visible to the short-wave infrared portions of
the electromagnetic spectrum, providing an almost continuous
spectral reflectance signature. It has been reported that hyper-
spectral data are capable of producing both genus- and species-
level classifications, whereas multispectral data are well suited
for genus-level classifications [1]. In particular, in land cover
classification of forests, where typically different species of
the same genus coexist, it has been shown that hyperspectral
Manuscript received June 17, 2010; revised February 22, 2011; accepted
May 22, 2011. Date of publication July 29, 2011; date of current version
December 23, 2011.
D. G. Stavrakoudis and J. B. Theocharis are with the Department of
Electrical and Computer Engineering, Division of Electronics and Computer
Engineering, Aristotle University of Thessaloniki, 54124 Thessaloniki, Greece
(e-mail: jstavrak@auth.gr; theochar@eng.auth.gr).
G. N. Galidaki was with Laboratory of Forest Management and Remote
Sensing, School of Forestry and Natural Environment, Aristotle University of
Thessaloniki, 54124 Thessaloniki, Greece. She is now with the Biology Depart-
ment, University of Trieste, 34127 Trieste, Italy (e-mail: galidaki@for.auth.gr).
I. Z. Gitas is with the Laboratory of Forest Management and Remote
Sensing, School of Forestry and Natural Environment, Aristotle University of
Thessaloniki, 54124 Thessaloniki, Greece (e-mail: igitas@for.auth.gr).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TGRS.2011.2159613
satellite imagery can significantly increase the classification
accuracy [2].
The high number of features encountered in remote-sensing
classification tasks from hyperspectral images has rendered
the use of simple statistical classifiers—commonly employed
in multispectral image classification—ineffective. Therefore,
more sophisticated classifiers must be considered in order to
exploit the rich spectral information offered by hyperspectral
data. In this direction, various methods from the field of pattern
recognition and artificial intelligence have been proposed for
handling high-dimensional remote-sensing data: neural net-
works (NNs) [3], fuzzy clustering algorithms [4], decision trees
[5], genetic algorithms (GAs) [6], kernel-based techniques [7],
and combinations of them [8], [9]. In particular, support vector
machines (SVMs) have been extensively used because of their
good classification performance in high-dimensional data sets
[10]–[12]. These proposals focus primarily on the accuracy
of the resulting thematic map, and they commonly apply a
feature selection preprocessing step in order to reduce the high-
dimensional feature space [6], [12].
Another approach taken by the respective research commu-
nity focuses on the interpretability of the considered classifi-
cation model. Indeed, most of the aforementioned classifiers
act as black boxes, providing only the class label for each
pixel. However, in many cases, it is desirable to derive a better
understating of the underlying physical relations of the classi-
fication problem, which is useful from an operational remote-
sensing perspective. Toward this direction, the use of rule-based
classifiers [13] and decision trees [14]–[16] has been consid-
ered. Because of the inherent interpretability of these methods,
an analyst can examine the resulting model and identify the
important factors that distinguish classes from one another.
In the context of interpretable classification models, fuzzy-
rule-based classification systems (FRBCSs) have arguably the
most intuitive representation, represented by simple rules with a
linguistic meaning. Moreover, the fuzzy partition of the feature
space can describe the naturally overlapping spectral signatures
of closely related vegetation species more accurately as com-
pared to crisp (nonfuzzy) classifiers. However, their application
to remote-sensing classification tasks has been rather limited,
with the existing proposals employing simple deterministic
learning methods and applying them to multispectral data only
[17]–[20]. The generation of FRBCSs for data sets comprising a
high number of features is a challenging task, since the number
of possible fuzzy rules increases exponentially with the increase
of the feature space. In the latter direction, the enhanced search
capabilities of GAs [21] have been extensively used in the
derivation of fuzzy-rule-based systems (FRBSs), giving rise to
0196-2892/$26.00 © 2011 IEEE