In: Wagner W., Székely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B
MULTITEMPORAL FUZZY MARKOV CHAIN-BASED CLASSIFICATION
OF VERY HIGH RESOLUTION IMAGES OF AN URBAN SITE
G. A. O. P. Costa 2 ' *, R. Q. Feitosa 3 , L. F. G. Rego b
a Department of Electrical Engineering
b Department of Geography
Pontificia Universidade Catolica do Rio de Janeiro Rua Marqués de Säo Vicente 225, Gâvea, Rio de Janeiro, CEP:
22453-900, RJ, Brasil - (gilson, raul)@ele.puc-rio.br, regoluiz@puc-rio.br
KEY WORDS: Multitemporal, Interpretation, Fuzzy Logic, High Resolution, Urban, Land Use, Land Cover
ABSTRACT:
This work discusses the application of the cascade, multitemporal classification method based on fuzzy Markov chains originally
introduced in (Feitosa et al. 2009), over a set of IKONOS images of urban areas within the city of Rio de Janeiro, Brazil. The
method combines the fuzzy, monotemporal, classification of a geographical region in two points in time to provide a single unified
result. The method does not require knowledge of the true class at the earlier date, but uses instead the attributes of the image object
being classified at both the later and the earlier date. A transformation law based on class transition possibilities projects the earlier
classification to the later date before combining both results. While in (Feitosa et al. 2009) the fuzzy Markov chain-based method
was evaluated over a series of medium resolution, LANDSAT images, in this work very high resolution images were processed.
Additionally, while the target area of the previous work was characterized predominantly by agricultural use, in this work an urban
area was the subject of classification. The results showed that the performance of the multitemporal method was consistently superior
to that of the monotemporal classification of the study area, and confirmed the robustness of the fuzzy Markov chain-based method
with respect to sensor characteristics and target sites.
1. INTRODUCTION
Sequences of Remote Sensing images of the same geographical
area acquired at different points represent a valuable source of
information that can be used to improve the accuracy and
reliability of classification-based image analysis.
Most traditional multidate image classification methods can be
regarded as “post-classification” approaches (Weismiller et al.,
1977), which are decisively dependent on the accuracy of the
initial classifications. More powerful alternatives, called
“cascade-classification” approaches (Swain, 1978) use all the
information contained in the image sequence, trying to explore
the correlation contained in the temporal data sets.
Feitosa et al. (2009) presented a detailed overview of the most
relevant efforts towards automatic cascade multitemporal
schemes found in the literature. These attempts include
probabilistic methods, methods based on neural networks and
multi-classifier approaches.
A first attempt towards a fuzzy cascade classification technique
can be found in (Mota et al., 2007). That method is restricted to
applications where the true class of the object being classified at
an earlier time is known. Feitosa et al. (2009) described a new
fuzzy cascade multitemporal classification model, explicitly
based on fuzzy Markov chains, in which object features other
than the true classification are used as the information from the
earlier date. In the later method, before the classifications of
two images at two dates are combined, the fuzzy classification
at the earlier date undergoes a temporal transformation that
projects it onto the later date.
This work discusses the application of the cascade,
multitemporal classification method introduced in (Feitosa et al.
2009), originally applied over an agricultural over a set of
IKONOS II images of urban areas within the city of Rio de
Janeiro, Brazil.
2. FUZZY MARKOV CHAINS
This section describes briefly the concept of Fuzzy Markov
Chain (FMC). A complete and more general presentation about
this technique and the related concepts may be found in
(Avrachenkov and Sanchez, 2002).
In this work we consider images acquired at dates t 0 +tAt, where
t 0 is some stipulated initial time, At is a given time interval, and
t is any integer number. For simplicity the date t 0 +tAt will be
denoted from this point on as time t, and t 0 +(t+l)At as time t+1,
forte Z
Let ft = {co!, co 2 ,•••, co n } be a set of n distinguishable land-
use/land-cover (LULC). A binary fuzzy relation can be defined
on the Cartesian product ftxft represented by a nxn transition
matrix T = {r, ; }. The symbol zy stands for the possibility that an
image object belongs to the class co,- e ft at time t and to the
class co ; e ft at time t+1, with 0 < zy < 1, for ij = 1,...,«.
This can be pictorially described by a class transition diagram
(Figure 1), a weighted directed graph whose nodes correspond
to classes and links to plausible class transitions between t and
t+1. Each link is labeled with the class transition possibility Zy.
For simplicity links with zy = 0 are not drawn.
* Corresponding author.
148