IMAGE CLASSIFICATION USING NON-PARAMETRIC
CLASSIFIERS AND CONTEXTUAL INFORMATION *
F.J. Cortijo and N. Perez de la Blanca
Depto. Ciencias de la Computación e I. A. (DECSAI)
E.T.S. Ingeniería Informática
Universidad de Granada
18071 Granada, Spain
cb@robinson.ugr.es
Commision Ill, Working Group 2
KEY WORDS: Classification, Learning, Algorithms, Combination, Accuracy, Pattern Recognition, Contextual Classification
ABSTRACT
This paper shows some combinations of classifiers that achieve high accuracy classifications. Traditionally it is used the
maximum likelihood classification as the initial classification for the contextual correction. We will show that using non-
parametric spectral classifiers to obtain the initial classification we can improve the accuracy of the classification significatively
with a reasonable computational cost. More specifically we propose to apply the contextual correction performed by the ICM
algorithm to some non-parametric spectral classifications.
1 INTRODUCTION
Supervised classifiers assume the existence of a training set
T composed by n labeled training samples, where the labels
represent informational classes (labels). This information is
used for learning -construction of the classifier- and usually
for testing too. We will note by Q = {w1,w2,...,ws} to the
set of informational classes and by X to the samples used for
learning and classifying. We assume they are d-dimensional
random variables.
Spectral classifiers use only the spectral information related
to the pixel to be classified. The thematic map they give as
output has the overall impression of a “noisy” classification.
This effect is more evident when there is overlapping among
the training sets in the spectral space [Cortijo et al., 1995].
In this case it is necessary a post-processing over the initial
classification because it is expected to find homogeneous re-
gions in the map as they can be found in the Nature. The
straightforward solution consists in incorporating additional
information into the classifier related to the spatial neighbor-
hood -its context- of the pixel to classify. That information
may be the spectral values of the spatial-neighbors pixels,
their labels or both kinds of information combined in some
way. When this kind of information is used for classification
the classifier is known as a contextual classifier.
From a general point of view a contextual classifier can be
seen as a smoothing process over an initial image of labels.
This map is obtained usually by a spectral classifier. lt is
well known that some contextual classifiers achieve a local
optimum [Besag, 1986] determined by the initial classifica-
tion. It is used traditionally the maximum likelihood (ML)
classification as the starting point for the smoothing process.
We have shown [Cortijo & Pérez de la Blanca, 1996a] that
the ML classifier is not the best choice when the training
sets are high-overlapped. In this work we propose the use of
different spectral classifications as initial classifications to a
contextual classifier in order to obtain high-accuracy classi-
fications with a reasonable computational cost.
In order to achieve a higher accuracy it looks reasonable to
*This work has been supported by the Spanish "Dirección General de
Ciencia y Tecnologia” (DGCYT) under grant PB-92-0925-C02-01
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
adopt a high accuracy spectral classification as starting point
to the contextual classifier, given that contextual classifiers
assure convergence to a local maximum. Our proposal con-
sists in adopting different spectral classifications as starting
points with the aim of improving the accuracy of the conven-
tional methodology consisting in contextually correcting the
ML classification. Many others spectral classifiers improve
significatively the results obtained by the ML classifier and
the classifications obtained by them are good candidates to
be the initial classifications for contextual classifiers. Finally,
we must considere the required computational effort to per-
form the global process: spectral classification followed by
the contextual classification. For a particular contextual clas-
sifier it is obvious that the contextual classification effort is
the same for any initial classification, thus the global com-
putational effort is determined by the spectral classification
computing demands.
This paper is organized as follows: In section 2 we describe
the methodology we have adopted in this work together with
a brief description of the classifiers we have used. In section 3
we describe the datasets used in this paper and in section 4
we show the results obtained. Finally, the main conclusions
we have achieved are summarized in section 5.
2 METHODOLOGY
Our objective in this work is to show some combinations of
classifiers that achieve high-accuracy classifications. In order
to determine some interesting combinations of classifiers for
Remote Sensing image classification we have tested a wide
number of families of spectral and contextual classifiers.
2.1 Spectral Classifiers
Spectral classifiers are partitioned in two main categories:
a) parametric classifiers, if they assume the existence of an
underlying probability distribution of the data and b) non-
parametric classifiers, if they do not assume anything about
the probability distribution.
The structure of the Bayes classifier is determined, basically,
by the probability density functions (pdf's) p(X|wi). The
objective in the construction of a supervised parametric clas-
sification rule is to characterize the pattern of each class in
120
terms of its
function, tl
(estimated
class. The
so it is on
each class:
Seis
The maxir
quadratic c
in the repr
variance IY
ratic classi
boundaries
sensitive t
pdf's than
training se
than the t
is well-knc
high dimer
enough to
When ado
ing an ext
aries to th
highly ove
good choi
adjustmen
daries. TI
inant anal
man [Frie
RDA allo
ing the qi
particular
is perform
parameter
situations
future mi:
parameter
nique is n
set size is
In most «
forms of t
istic. The
actually e
ric model:
problems
The only |
sification
assumptic
We can
classifiers
categories
parametr
(nearest
timation
[Devijver
directly tl
classifica
consists
means of
ables inv
opted tw«
and knov
