(5)
(6)
(7)
(c) The first output of reasoning network
is the output of the 1/2 hour
Scofield/Oliver reasoning network.
(d) The next output of the reasoning
network is a function of previous
outputs of the reasoning network and
observed data of this case.
(e) The reasoning network is trained.
The convergence weights can be used
as the weights for the 1/2 hour
reasoning network for this special
type or model of rainfall.
Another case of rainfall which has the same type or
model rainfall mentioned in (2) is chosen to retrain
the reasoning network for getting convergence
weights for this special type or model of rainfall.
(a) The weights of the 1/2 hour
reasoning network derived previously
will be used first.
(b) The inputs of the reasoning network
are the 7 factors based on the
Scofield/Oliver Technique.
(c) The output of the reasoning network
is a function of previous output of
the reasoning network and
observational data for this case.
(d) This reasoning networks then trained.
the new convergence weights can be
used as the new weights of 1/2 hour
reasoning network for this special
type or model of rainfall.
Repeat (5) until all the training samples are used
for training this special type or model of rainfall
reasoning network. The final trained result of the
weights are the weights of 1/2 hour satellite-derived
reasoning neural network for this special type or
model of rainfall.
Using the 1/2 hour satellite-derived reasoning
neural network for testing the estimation of rainfall.
(a) A test case which has same type or
model of rainfall has be chosen.
(b The 1/2 Hour Satellite-derived
Reasoning Network is run using the
weights from (6) to obtain the
estimation result.
789
(c) If the testing result is not good, the
error is more than 10%, go back to
(5) and the testing case will become
another training case.
(d) If the testing result is good, error is
less than or equal to 10%, go to (7)
and test again.
4. RESULTS OF ESTIMATION
4.1
Experimental Results of the 1/2 Hour Mesoscale
Convective Complex (MCC) reasoning Network
The experiment results of a 1/2 hour the MCC type
reasoning network for the estimation of rainfall can be
seen in Table 1. In this case, on July 19, 1985, a MCC
located in IOWA (IA), USA. The observed rainfall was
9.5 inches.
The values of column S/O E are the results from the
Xie/Scofield study (Xie and Scofield, 1988). The sum of
the 1/2 hour estimates was 18.64 inches. The error (the
difference between the observed data and the sum of the
1/2 hour satellite estimates) was + 96.2%.
The values of column MCC E are the 1/2 hour estimates
result from the 1/2 hour MCC reasoning network of the
ANSER system. The sum of the 1/2 hour estimation data
was 9.43 inches. The error is only -0.74%. In this case,
after all information had been received, the satellite-
derived estimation of rainfall only required 2 seconds of
HDS 9000 CPU time to execute. Therefore the weights of
the 1/2 hour MCC Reasoning Network of ANSER are
very good for this type of event.
4.2 Experimental Results of the 1/2 Hour Multi-
Clustered Linear (MCL) reasoning Network
The experiment results of a 1/2 hour the MCL type
reasoning network for the estimation of rainfall can be
seen in Table 2. In this case, on August 12, 1987, a MCL
located in Kansas (KS), USA. The observed rainfall was
8.7 inches.
The values of column S/O E are the results from the
Xie/Scofield study (Xie and Scofield, 1988). The sum of
the 1/2 hour estimates was 6.038 inches. The error (the
difference between the observed data and the sum of the
1/2 hour satellite estimates) was -30.6%.
The values of column MCL E are the 1/2 hour estimates
result from the 1/2 hour MCL reasoning network of the
ANSER system. The sum of the 1/2 hour estimation data
was 8.53 inches. The error is only + 1.92%. In this case,
after all information had been received, the satellite-
derived estimation of rainfall only required 2 seconds of
HDS 9000 CPU time to execute. Therefore the weights of
the 1/2 hour MCL Reasoning Network of ANSER are
very good for this type of event.
