where y,is the actual output of the node À, and f is the 
activation function. The most commonly used activation 
function ( f ) is the sigmoid function. The sigmoid function 
most often used for ANNs (in this study also) is the logistic 
function such as below. 
1 ; 
= u,) —————— 5) 
yı = od 1+exp(-u,) : 
The measure of error is the sum squared error (SEE) of learning 
function that is calculated by equation (6), which should be 
minimized. 
SSE- Y 3 (d, - y 6) 
Dick 
where d pk and y pk are the teaching and actual outputs of 
output node Kk on pattern D respectively. The back 
propagation algorithm attempts to minimize iteratively the SSE. 
2.3 The Method 
The ANN used in present study, is SNNS (Stuttgart Neural 
Network Simulator) developed by university of Stuttgart 
(Martin et. al 1998). The Feed-Forward ANN with back- 
propagation learning algorithm, SEE for error estimation and 
randomize weights function (The weights and biases initialize 
with distributed random values) are used. To investigate the 
best architecture of ANN for the development of the 
classification algorithm, the 400 patterns have been randomly 
divided into two parts, for the training, and testing. In all data- 
sets, as indicated previously, the inputs T, are normalized 
between 0 and 1 using equation (2) and for both training and 
testing pattern files the teaching output are represented by 0.0 
and 0.9. If the output value of pattern D for class l is2 0.5, 
the pattern is considered as class / otherwise it is considered as 
not classified. 
The network training consists of 20000 iterations through the 
training patterns. At the beginning of training, all weights and 
biases are set to independent random values between + 1 . This 
procedure is repeated several times by assuming a different set 
of random values of weights and biases for initializing the 
network and different learning parameter (between 0.2-0.9). 
Note that in these runs, the number of nodes in each hidden 
layer is kept constant. The objective of that is to reach the 
global minimum or deepest possible local minimum of the error 
surface. To avoid over-fitting and arrive at the optimal network 
architecture, this procedure is repeated for the different 
architectures of network. The architecture, which yields the 
lowest SSE magnitude on training and testing datasets, is 
considered to be optimal for the chosen number of hidden 
nodes and hidden layers. In our case, the optimal network 
configuration, which error is the globally smallest, had 2 hidden 
IAPRS & SIS, Vol.34, Part 7, “Resource and Environmental Monitoring”, Hyderabad, India,2002 
layers with 8 nodes in each layer (5-8-8-8). The program was 
run for 20000 iterations and learning has been done by all 200 
training patterns. For over all performance, it is tested with 200 
test patterns, which network never seen them before. The SSE, 
in training case, drops fairly rapidly from first few hundred 
iterations, and then the performance improve more slowly. It 
gives some indication of iteration effect on SSE. The SSE for 
all 200 testing patterns is 43.4. 
2.4 Results 
The brightness temperatures at five frequencies of AMSU-B are 
selected as input for discriminate and classification of the eight 
classes of weather features, using the ANN. The test performed 
using the ANN classifier resulted in 162 of 200, samples being 
correctly classified. The accuracy of classification and 
classification of the test patterns as a function of classes, i 
details, is presented in table 1 and Fig 2. : 
A significant portion of the improper classifications occur 
between class3 (Rnf) and class 4 (Snf). For example, from 4 
improper class patterns in class 4, all are being improper 
classified as class 3. The improper classification may be due to 
large similarity between rainfall and snowfall classes in the 
microwave region. The most of improper classified samples in 
class 2 are also being misclassified as class 3. It can be 
Table 1: the classes (col. 1), and corresponding test patterns of 
each class per AMSU-B channel (col.2), number of correct, 
improper and not classified patterns (col.3-5). 
  
  
  
  
  
  
  
  
Classes “TTP! Cc NPC? NC* 
1. Trs 25 (1-26) 25 0 0 
2.Hvr | 25 (26-51) 14 6 5 
3.Rnf | 25(51-76) 19 3 3 
4.Snf | 25(76-101) 14 4 7 
5.CII | 25 (101-126) 16 3 6 
6..Crl-. 25(126-151) 24 1 0 
7.Cls d 25 051-176) 25 0 0 
8.Crs | 25 (176-201) 25 0 0 
Total 200 162 (81%) | 17(8.5%) | 21(10.5%) 
  
  
"TTP = total test patterns, ^ CC = Correct Classified, * NPC = 
not proper Classified, ^ NC - not Classified, and ? ' the pattern 
numbers are indicated in the bracket (see Fig 2). 
seen from Fig 2 that more improper classification appeared for 
classes of 2 to 5. Not classified patterns, in Fig 2, are also 
presented as class nine. 
Otherwise of our ANN results could only be compared and 
verified using other (IMO and meteorological observations). 
There is no other model available for comparison. Results in 
table 1 and Fig 2, however indicate that the only five 
frequencies data of AMSU-B, is not able to give full 
justification directly for data analysis for all the reported eight 
IMO classes. There are probably some other input parameters, 
which are essential for the purpose (probably like ground 
temperature). 
   
    
   
     
  
  
   
     
    
   
    
    
       
      
     
     
  
    
   
    
     
   
    
  
   
   
    
    
  
   
  
    
Fig. 
Ac 
pro 
Sci 
faci 
ack 
Bar 
inr 
Bar 
Net 
Vo 
Bu 
Ret 
SS] 
101 
Chi 
19% 
Mi 
Oce