AMSU-B 
WORK 
ssify of cloud- 
ıres of weather 
hannels data of 
nples for each 
-65E, 22-45N), 
orm, 2) heavy 
[hey are based 
emperatures in 
dditional input 
s using multi- 
classifying of 
thunderstorm 
. (Snf), cloudy 
1dy sky on the 
n found using 
22-45N). The 
the 400 
ing data from 
of the ANN 
| from the 20 
erent location 
uary 2001 to 
temperatures 
we call them 
ng synoptic 
Organization 
AMSU-B data 
emperatures. 
(1) 
in, frequency . 
radiance in 
juency f and 
temperature T (in Kelvin). The corresponding T of areas, 
where weather features (table 1, col. 1), were reported by IMO, 
are selected. The validation of the weather features of class 1 
through 5 is carried out with the hourly report of IMO. The 
clear sky on the land and the clear or cloudy sky conditions on 
the sea cases, were defined by collocated IR-Meteosat and 
AMSU-B images. Each T is normalized between 0 and 1 
according to equation (2). 
2 (T, Tu) Q) 
E x uu E 
(Cua. = T yi, ) 
where Ts is normalized brightness temperature, T, is 
observed brightness temperature calculated by eq. 
(0, T o and T are common maximum and minimum 
brightness temperatures for all channels frequency. Note that 
T is related to which channel that T, is. For network training 
and testing, the 400 input patterns are divided into two sets, 200 
for training and 200 for testing. 
2. 2 Artificial Neural Networks (ANNs) 
The most commonly used form of ANN is the Feed-forward 
neural network. A schematic diagram of the type of ANN that 
was used for this study is presented in Fig. la. It consists of an 
IAPRS & SIS, Vol.34, Part 7, “Resource and Environmental Monitoring”, Hyderabad, India,2002 
input layer of Mnodes (to which an input vector X; = 
(X, X5 ,***, X, ) is applied), one or more hidden layer, and an 
output layer of Knodes with output vector Y; = 
(V1 > V2»**"> Y, ). In our case, the number of input nodes is 
the five Ty corresponding to five frequencies of AMSU-B 
(n = 5), and the number of nodes in output layer is the eight 
weather features (k =8 ). The number of hidden layers and 
the number of nodes in each hidden layer must be determined 
by trial and error. In this study, the best accuracy was achieved 
using 2 hidden layers and eight nodes in each layer. Each of the 
input nodes is connected to all 772 nodes in the first hidden 
layer and each node in any hidden layer is connected to all 
nodes in neighborhoods layer. 
A node (neuron) is an information-processing unit that is 
fundamental to the operation of an ANN. The three basic 
elements of the neuronal model are identified in Fig.1b: i) A set 
of synapses or connecting links. A signal X jt the input of 
synapse J is connected to node k , is multiplied by the synaptic 
weight Wy ; ii) An adders for summing the input signals, 
weighted by the respective synapse of the node and iii) An 
activation function for limiting the amplitude of the output of a 
node. 
Fig.1: (a) schematic of a feed forward ANN architecture. The components of input vector X, output vector Y, along with the weight 
and bias links. The circles symbolize the nodes (neurons). (b) a sample node with the three basic elements 
  
A mk EX — [biases] (a) 
  
  
output vector 
  
  
  
  
  
   
  
    
   
  
  
    
   
   
   
    
    
      
   
  
   
   
  
   
  
  
  
   
   
  
  
   
  
  
(b) 
M 1 X 
activation function 
  
  
  
    
  
— tà , 
  
ai 
  
  
inputs 
  
  
  
   
n : 
vector input layer hidden layer output layer 
   
summing junction 
    
  
INPUS  sunoptic weights 
   
  
  
  
  
  
The node model of Fig.1b also includes an externally applied 
bias denoted by b,. The bias bi. depending on whether it is 
positive or negative, has the effect of increasing or lowering the 
net input of the activation function. In mathematical terms, a 
neuron may describe as following equation: 
n 
U, = yx, à) 
j=0 
where W,0% = b, are called the bias, Xj and Wy are the 
input signal and the synaptic weight and, is the linear 
combiner output due to the bias and input signal of the node k. 
The weights (W kj ) are determined during the training process. 
In the present study, weights are obtained using back 
Propagation algorithm. It adjusts the weights iteratively to 
reduce the difference between teaching outputs and actual 
outputs calculated by the network using the input values. The 
effective incoming signal U, is passed through a nonlinear 
activation function to produce the outgoing signal ( y, ) of the 
node. 
y, = fu) (4)