2 ANSER TECHNIQUE
21 The Architecture of ANSER
The architecture of an ANSER for satellite-derived
estimation of rainfall can be seen in Fig. 1. There are
three parts of this architecture: (a) ANSER USER
SYSTEM; (b) ANSER TRAINING SYSTEM; (c) ANSER
CENTER SYSTEM. The ANSER USER SYSTEM has
one or more user subsystem(s) based in the IBM PC for
estimates by users. Each user subsystem consists of a
training subsystem, weight base and estimate subsystem.
The ANSER TRAINING SYSTEM has more than one
training subsystems operating on the mainframe computer
NCCF HDS 9000 for training weights of ANSER. Each
training subsystem has first training, re-training and output
result functions. ANSER CENTER SYSTEM receives
satellite data based on the IBM RISC 520 for the experts
and it has six parts: (a) display subsystem for output of
rainfall estimates; (b) explanation subsystem that gives
different classification to different data; (c) a reasoning
network for rainfall estimation based on the input data
and rule, model, and knowledge base; (d) rule bases,
mode bases, and knowledge bases will save rule, model,
and knowledge provided by the expert; (e) a training
subsystem for getting suitable weights for the ANSER; (f)
weight bases for keeping weights of ANSER. ANSER
USER SYSTEM, ANSER TRAINING SYSTEM and
ANSER CENTER SYSTEM communicate with each
other using Ethernet. This architecture will be enhanced
for application to derive estimation of rainfall from
satellite data.
22 Architecture of Reasoning Network
The basic architecture of reasoning network for 1/2 hour
satellite-derived estimation of rainfall can be seen in Fig.
2. This is a 3 layer artificial neural network that includes
7 input linear neurons, 30 hidden nonlinear neurons
(divided into 2 layers) and 1 output nonlinear neuron.
There are 345 weights in this network.
The artificial neuron is a unit that functions similar to the
real neuron of human (in this paper, neuron means
artificial neuron). The ANN is a system which consists of
artificial neurons that are connected to each other by
weights. The function of each nonlinear neuron is sigmoid
and is given by:
N
Yj-M( *exp(-5 ^ (Yi«Wij))..
i=0
where: Yj - the output of the j-th artificial neuron.
Wij - the weight connected the i-th artificial
neuron with the j-th artificial neuron.
Yi - the output of the i-th artificial neuron.
When the ANN was described in this paper, the artificial
neuron numbers of layer are different. However, three
layers of structure are always used. Several basic
structures of ANNs will be connected to each other to
become reasoning networks and the basic structure of the
ANSER system.
3 ANSER PERFORMANCE
3.1. Input and Output of Reasoning Network
The input and output of the reasoning network are as
follows:
Input: G = (cloud top temperature + cloud growth factor)
or (cloud top temperature + strong divergence aloft)
RB = rain burst factor
OS = overshooting top factor
M = merger factor
SE = saturated environment factor
MC = moisture correction
= speed of storm
Output: 1/2 hour Satellite-derived Estimation of Rainfall.
32 Training Algorithm of the Reasoning Network
The training algorithm of the reasoning network for 1/2
hour Satellite-derived Estimates of Rainfall are as follows:
(1) Set random values to all weights of the reasoning
neural network.
(2 A special type or model of rainfall should be
chosen. It means that the reasoning network will be
trained for this special type or model of rainfall
estimation.
(3) One case of rainfall which has the same type or
model rainfall mentioned in (2) is chosen to train
the reasoning network for getting convergence
weights.
(a) Using Scofield/Oliver Technique:
The inputs of reasoning network are
the 7 factors shown in the beginning
of this section. The output of the
reasoning network is: S/O E=[(G or
RB)+OS+M+SE]*MC*S.
(b) The reasoning network is trained.
The convergence weights are the
weights of the 1/2 hour
Scofield/Oliver reasoning network.
(4) The same case of rainfall which has the same type
or model rainfall mentioned in (2) is chosen again
to re-train the reasoning network for getting
convergence weights for this special type or model
case of rainfall.
(a) The weights of the 1/2 hour
Scofield/Oliver reasoning network
will be used firstly.
(b) The inputs of reasoning network are
the 7 factors based on the
Scofield/Oliver Technique.
786