3. Istanbul 2004 
tional to their 
he affinity, the 
ulation M and 
to be kept as 
net. 
CATION 
e involves two 
m with a set of 
by selecting the 
ied to train the 
remote sensing 
a set of sample 
ecting region of 
, The training 
that can be 
Ab, represent 
f remaining Ab. 
domly choosing 
ory cells Ab, 
the training set 
nd present it to 
üns the affinity 
nt investigation, 
of affinity. The 
(1) 
(2) 
age bands. 
compose a new 
| to Ag, and 
MC 
match * 
genic affinities, 
tigenic affinity, 
r each of the n 
senerated N, is 
  
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
  
where = a multiplying factor 
N 7 the toutal number of Ab's 
round(-) 7 the operator that rounds its argument toward 
the closet integer. 
5. Allow each Ab's in clone set C! the opportunity to produce 
mutated offspring C 7 The higher the affinity, the smaller the 
mutation rate. Where mutate procedure and function mutate(x) 
are defined in Figure 2. In Figure 2, the function Irandom() 
returns a random value in the range [0,1] and Lrandom returns a 
random value in the range [-1,1].Function I(t, y) is defined in 
equation(4) as follows: 
Lt 
AG. y)eyq-r f ) (4) 
where  /-the iteration number 
T = the maximum of iteration number 
r = a random value in the range [0,1] 
A= a parameter to decide the nonconforming degree 
  
mutate(x) 
{ 
foreach(x.vi in x.v) 
do 
ai = minvi 
bi = maxvi 
rd mr 7 Irandom() 
rd to = Lrandom() 
if(rd_mr < mutation_rate) 
if(rd to>=0) 
xvi= x.vi+ 0, bi- xvi) 
else 
Xvi= x.vi- (t, x.vi- ai) 
done 
return x 
} 
  
  
  
Fig.2. Mutation 
* pk 
6. Calculate the affinity aff, of the matured clones C^ in 
relation to antigen Ag; 
p* 
7. Select the highest affinity from the set of C" in relation 
to À g; as the candidate memory cell, MC andidate . t0 enter 
the set of memory antibodies Ab mn} : 
that was 
8. Decide whether the mc replaces mC 
candidate match 
previously identified. If MC has more affinity by the 
candidate 
training antigen, ag, The candidate memory cell is added to the 
set of memory cells Ab, , and replace mc 
{m} match * 
9. Replace the d lowest affinity Ab's from Ab, : 
j 
10. A stopping criterion is calculated at this point. It is met if 
the average affinity for Ab's is above a threshold value. If the 
stopping criterion is met, then training on this one antigen stops. 
If the stopping criterion has not been met, repeat, beginning at 
step 3. 
This process continues until all antigens have been training. 
4.2 CLASSIFICATION 
After training has completed, the evolved memory cells Ab, 
are available for the use for classification. Each memory cell is 
presented with a data item. By calculating the affinity between 
memory cell and image data, the image is classified into the 
class that has the maximum affinity. 
5. EXPERIMENTAL RESULTS 
5.1 Data 
The study area of this research is in WUHAN city in China. 
The TM images (400x400 pixels) used were acquired in Oct.26 
1998. Fig.3.shows the image. The classification patterns 
adopted here are five classes: Changjiang River, lake, 
vegetation, road and building. In the experiment, five regions of 
interests representing the five classes respectively were selected 
for training regions and every training region had 100 ground 
reference sample points. 
  
5.2 Results 
In this case, the running parameters were n = 10, d=5 and = 10. 
Fig.4 illustrates the classification result using Artificial Immune 
classifier. In order to compare the classification result, Fig.5 
illustrates the classification result using maximum likelihood 
classifier. Tablel shows the classification accuracy and the 
Table 2 is the accuracy of the AIS method. From the table 2, it 
is found that AIS approach produces better classification results 
than the Maximum Likelihood method. In order to check the 
results in more detail, we show confusion matrices in Table 1 
and Table2. As shown in Table 2, the AIS approach improved 
overall classification accuracy from 85.0% to 89.8%(4.8% 
improvement). For each class, the vegetation has the largest 
improvement from 59% to 76%(17% improvement), followed 
by road (5% improvement), building (4% improvement). The 
reason for this is that the maximum likelihood approach works 
well only when the underlying assumptions are satisfied and 
poor performance may be obtained if the true probability 
density functions are different from those assumed by the model, 
while AIS are nonlinear models, which make them flexible in 
modeling real world complex relationships.