International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
F:Qx Xx r — Q is the stochastic state
transition function
O:Q — Y is the stochastic output function
Q(t + 1) = T(Q(0) is the reinforcement scheme.
C = {c¢j, 0 <j < o} is the penalty probability
distribution.
r= {r;,0 <j<c00}is the reinforcement signal.
5. USING CELLULAR LEARNING ATOMATA FOR
POSTCLASSIFICATOIN
In order to use cellular learning automata for improving
classification accuracy, a cellular learning with 8 neighbour
structures is considered, and the following steps which include
choosing an action by automata, compute penalty probability by
environment, updating neighbour functions and updating inner
state are considered.
5.1.1 Action: Action @; is choosing one of two classes
which have more probability; at initial state it choose randomly
by automata.
5.1.2 Penalty probability: penalty probability c; is
associated with action a; which is chosen by environment. The
environment considers two criteria for evaluating action
automata: pixel entropy for local optimization and omission
error for global optimization of each class. Once the automata
choose an action that lead to increase the entropy of pixel,
environment gives it penalty. After each iteration if the
omission error decreased the environment will give reward to
the automata's action. Amount of reward and penalty is
compute as follows:
C= a*C,i+b*Cy (1 1)
0<a,b<l , atb=l
where C2i =omission error
a pera, (x)
C UT is AL
The amount of C; maps to 0 and 1 as follows:
If (Ci « 0.5 ) then B=1 else B=0 (13)
5.1.3 Neighbour function between automata: in order to
compute the inner state of automata it should compute
neighbour function between automata. We use equation 5 in
which way that the C; affects on neighbour function between
automata.
5.1.4 Computing inner state of automata: in this stage, at
first the local probabilities of pixels based on two stage of
percipience memory of neighbour pixel which refers to penalty
probability are computed. Then, an updating probability role
which depends on local probabilities, initial inner state and
neighbor function was introduced. After that, inner state of
automata is computed by probability role.
The algorithm executes the steps mentioned already and
continues until reach to a best situation; the best situation is a
state where pixels have less entropy with the classes having less
omission error.
6. EVALUATION AND EXPERIMENT RESULTS
In order to evaluate the algorithm of post processing, a subset
image (Figure 3) which is a portion of the Airborne
Visible/Infrared Imaging Spectrometer (AVIRIS) of
hyperspectral data is used. This image was taken over an
agricultural area of California, USA in 1994. This data has 220
spectral bands about 10 nm apart in the spectral region from 0.4
to 2.45 um with a spatial resolution of 20 m. The subset image
is 145 by 145 pixels and its corresponding ground truth map is
shown in Figure 4 .the image area has 12 classes.
aur ™ REA
Figure 4. Grand truth of area with 12 classes
At first some noisy bands were put away. In order to separate
noise, and to extract original signal from image bands the
minimum noise fraction transform was performed. Based on
eigenvalue of components we chose components which had
high variance; therefore the original image dimension Wäs
reduced. We used 46 components which contain high percent of
994
Intern
origin:
classif
endem
were
accura
the im
The ré
fractio
We in
and pi
used 1
classif
were C
Classi
Alg
P
proc
algc
Ov
acc!
Noticii
image:
adapte
probat
that th
togeth
overco
Hyper:
fractio
useful
advant
classif
probal
probab
indepe
param
tries {
optimi
satisfic
equatic
depenc
prospe
enviror
action
way. C