In: Wagner W., Szgkely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010,1 APRS, Vol. XXXVIII, Part 7B
3.3 Particularization of the classification model
FMC models may be built using any t-norm and s-norm
composition. We favor the max-product composition, since it
leads to a simple multitemporal classification model with an
intuitive interpretation,. Thus Equation (1) takes the form:
,+1 p ; . = max{'a,.T, y } (11)
for i,j = 1
The defuzzyfication step is carried out by a hardening function
H that selects the fuzzy set with the highest membership grade,
formally:
\w h ...,w n ] =w= H(p)= H([p/,..., p„]), where
lforp,. = max{p 1 ,...,p„} (12)
0 otherwise
The aggregation function F is the product of corresponding
elements of the input fuzzy vectors. Thus,
,+ V =F(' +/ a, <+/ p)= [ ,+ V + %-, ,+k a n ,+/ p„] (13)
Putting it all together, the multitemporal classifier assigns the
image object to the class co, e il at time t+1, for which
"V- = m fx{' +1 a, maxf'a^j} = max{'a,T tt ' +1 a t } (14)
holds.
3.4 Estimating transition possibilities
The estimation of transition possibilities basically selects the set
of possibility values that maximizes the classification accuracy
computed upon a given training set.
The estimation procedure consists of finding the set of
transition possibility values T={iy} that maximizes the selected
accuracy function G, for the selected image objects S of image
objects at the later date (the training set) and for the selected
monotemporal classifiers. This is formally expressed by:
T = argmax{G(s, £ C/-C,f)} (15)
f
The computation of transition possibilities defined in Equation
(17) involves an optimization procedure. The classification
model introduced in the previous sections is actually not bound
to any particular parameter optimization technique. In this
work, the average class accuracy was used as accuracy function,
and a Genetic Algorithm was the optimization technique used to
estimate transition possibilities (Schmiedle et al., 2002).
4. EXPERIMENTS
The experiments described in the following sections were
designed to evaluate the proposed method over a set of high
resolution IKONOS II images. The experiments aimed at
comparing the outcome of the multitemporal classification with
that of the monotemporal classification of the later image. We
also investigated the performance of the method tuning the
performance of the earlier monotemporal classifier, as described
in Section 4.4. The idea was to investigate how the method
behaves with different (increasing) performances of the earlier
monotemporal classifier.
The data set used in all experiments is described in the next
section. Afterward, the design of the particular monotemporal
classifier that composes the multitemporal scheme in the
experiments is presented. The following two sections describe
respectively the monotemporal classifier design and the
optimization technique used to estimate transition possibilities.
4.1 Description of the data set
The test site corresponds to a 14.4 km 2 area, situated on the
north face of the Tijuca National Park, within the city of Rio de
Janeiro, an important Atlantic Forest reminiscent. The test-site
is a subset of the area of interest of the PIMAR Project (Remote
Environmental Monitoring Program), which aims at monitoring
the suppression of rainforest on conservation units inside the
municipality of Rio de Janeiro through high resolution optical
remote sensing images (PIMAR, 2010).
This area was selected as test-site because of the noticeable
sprawl of informal dwellings over legally protected natural
areas. Moreover, the area contents various instances of all land
cover classes considered by the PIMAR Project.
Two pan-shaped, orthorectified IKONOS II images were used
in the experiments. The images actually take part of two stereo
pairs, each pair acquired on the same orbit, with different
elevation angles. The orthorectification was performed using
digital elevation models derived from each stereo pair. For each
year the image with the highest elevation angle was chosen and
submitted to orthorectification.
It is important to note that because of the time of the year the
images were acquired - March 2008 (late summer) and June
2009 (late autumn) respectively -, they present quite different
illumination conditions, with an important presence of clouds in
the 2009 image. No radiometric correction or equalization was
performed over the orthorectfied images.
Figure 2. Area of the test-site in Rio de Janeiro municipality.
The orthorectified IKONOS II images were segmented using
the multiresolution segmentation algorithm proposed by Baaz
and Shape (2000), through the Definiens Developer 7 software.
The parameters chosen for the segmentation procedure were