4 
| 2004 International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
eeded | significantly affects the computational time, due to the network were images. Thus, several images were generated 
SNNS increased size and complexity of the neural network. An ERS from the original SAR, each one presenting a texture or 
‘hen a | scene (120Mb approximately) requires 5 hours for processing geometry key-feature. Five images were selected according to 
' unit while an image window of 4-16Mb size, requires 2-5 minutes. their performance in oil spill classification (Topouzelis et al, 
d. | The method was applied on image windows of 4-16Mb, to test 2002): the original SAR image, the shape texture, the 
ler | its performance in terms of time requirements and result asymmetry, the mean difference to neighbours and the power to 
| quality. 
mean images (Figure 5). Shape texture image is referred to the 
| texture which is based on spectral information provided by the 
The main aspects considered for oil spill detection using neural original image layer and calculated as the standard deviation of 
| networks were data preparation, network architecture decision, the different mean values of image objects already produced. 
| parameters estimation and network performance accession. The Asymmetry can be expressed as the ratio of the lengths of 
| general overflow of the method developed for both networks minor and major axes of an ellipse which can be approximated 
> | (MLP and RBF) is illustrated in Figure 4. for image objects. Mean difference to neighbours can be 
: expressed as the mean difference for each neighbouring object 
In a previous study, a detailed examination of the features multiplied by the shared border length of the object concerned. 
| contribution to oil spill detection has been performed Power to mean ratio is defined as the ratio of the standard 
| (Topouzelis et al, 2002). Features, which have been led to deviation and the mean value of the objects. 
| successful oil spill detection, were extracted from the SAR 
| image. A preparation was necessary in order these features to 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
    
  
    
be functional to neural network. Moreover, an initial network 
| was chosen and trained for each network. Image results were 
qe | compared with reference data to assess method accuracy. 
Network architecture continuously changed, adding a node 
ristics, | (input layer or neuron) and re-evaluating the method. Figure 5 
ns and | presents the methodology used. 
licated ; iTS 
power | et + P - 
nguish | SAR Feature Extraction à = ri : 
he two Image wd b) Shape texture image 
lem is ' | t 7 
rder to D 
s from | y pe. 
fferent | Pre-processing 
ion are | 
sent a | : GEI M DA M ME E 
etwork | = 
| > Initial NN Topology 
| Initial Weights 
: | | 
| Add ‘ dites ux us LONE 
| node Training Net €) Power to mean image 
  
  
  
  
  
  
  
y Figure 5. Inputs to Neural Networks 
Create Image Results 
  
  
  
  
4.3 Network Topology 
  
  
  
  
Calculate Accuracy For both MPL and RBF neural networks an initial network 
topology was selected. The selection of the best suited topology 
for each NN was designed through the hill-climbing approach, 
which for a search point uses a solution created from a previous 
Figure 4. Methodology used topology. The contractive algorithm was used, in which initial 
IN topology was the simplest one and nodes were added 
afterwards. The performance of each topology was evaluated 
and the process was repeated iteratively until a predetermined 
  
  
  
  
  
  
  
4.2 Preparing the data 
ings Preparing the data involves feature extraction and normalize the Hopping CrICHON was 3 ghigved. Cansiictive als vas 
image d parma red inta vali à ie: [O. according to-the chosen among other hill-climbing algorithms (e.g. pruning) 
strong as b p Intemva (ar Sa h 5 f because it was very easy to specify the initial NN topology and 
ontains mum ard maximum values of he feature: The purpose 9 it was significantly faster in terms of training time. 
"(local feature extraction is to map the image to a feature space that | 
neni ous Pie es inc basis Tor Shih divis scouts In the present study, was chosen a layer with one input node - I 
inputs 1997). In order the neural network to be functional an s the original SAR image - and one output node as initial network ' 
classification procedure to be simple, the inputs of the neural 
727