33. Istanbul 2004 International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
  
ion time, results 
ok more time in 
  
  
.3 have opposite 
tection ability in 
it seems the 
ake the network 
n is examined in 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
stance as input te^ Ys" 
E 4 e 
| 9 pixels in t Cu NH t^ * wl fi 
pixels in the * e M mn 
pixels are added Pu E ; x 
le input layer is > : 
ly 9 normalized | Maximum Likelihood Method Hidden Neurodes: 10 Hidden Neurodes: 15 
36 neurodes are RCC: 55.93 Iteration: 5000 Iteration: 1500 
k's structure and BCC: 70.15 RCC: 73.31 RCC: 75.53 
RMSE: 0.3347 BCC: 88.87 BCC: 95.59 
Overall Acc.: 91.23 RMSE: 0.2238 RMSE: 0.2012 
Overall Acc.: 94.66 Overall Acc.: 95.18 
  
  
  
  
  
  
  
  
Figure7: Left image: Maximum-Likelihood result, Middle image: Simple BNN with 10 neurods in hidden layer from section 4.1, 
Right image: Improved BNN with 15 neurods in hidden layer from section 4.4. 
  
  
  
5. NETWORK'S FUNCTIONALITY ON QUICK-BIRD 
IMAGES 
In this section a part of an RGB Quick-bird image from 
Bushehr harbour in Iran is chosen as input image to evaluate 
network's behaviour on this kind of images. 
Figure 8 shows the original image and its manually produced 
reference map which is used in accuracy assessment. 
Two input parameter types are implemented. In the first case 
only spectral values are used in input vector formation and 
therefore three neurodes are designed in input layer. In the 
second case the suggested input parameter set, which is made 
up of spectral values and normalized distances of all pixels in 
surrounding window, is implemented and therefore 36 neurodes 
  
ixels with their 
meters 
listance as input 
  
  
  
  
  
  
  
  
  
ASE Overall ; ; 
: Acc. are designed in input layer. 
008 95.13 A variety of networks with different neurode numbers in 
s "ie hidden layer are used and each network was trained with 
1999 95.18 multiple iteration times to discover best iteration time when the 
network is not over-trained. 
2012 95.18 The optimum network structures and iteration times are 
à selected considering computed accuracy assessment parameters. 
2006 95.19 
Obtained results and their accuracy assessment parameters are 
shown in Figure 9. 
.2, 3 shows that 
could improve 
ound detection. 
ning stage more 
nsated to some 
ze and iteration 
  
thods, obtained 
on method and 
ther in Figure 7 
  
  
  
(b) 
Figure (8) a) RGB Quick-Bird Image from Bushehr harbour in 
Iran. b) Manually produced reference map