This paper focuses on configuring ANN for inundation
areas identification based on various flood
causative factors (Fig 2), which were selected based
on knowledge acquired from during our preliminary
on-site survey [4]. There were indeed many recent
studies that adopted MLP and RBF in flood prediction
[5-9]. What these techniques had in common
is that they built a model which predicts runoff
(or discharge) from rainfall, without taking into
account other flood causative factors. Moreover,
their common deficit is that prediction could only
be made at watersheds and where gauging stations
exist (typically sparse). This limited coverage as a
result constrains the inundation areas being identifiable
merely at locations where water level can be
measured. Nonetheless in their own right, both MLP
and RBF are versatile and have a strong merit as
prediction network if properly implemented. Compared
with other ANN based methods, many studies
[10-13] have confirmed MLP and RBF superior
accuracy over, for instance, LVQ and others. The
Adaptive Resonance Theory (ART) neural network,
despite its simplicity, are not fully developed and currently
criticized for its statistical inconsistency and
high de-pendency on the order in which the data are
processed. Furthermore, we have previously reported
[4] that when benchmarked against prominent learning
based systems, the ANN outperformed GA, SOM
and Fuzzy Logic in terms of flood prediction accuracy
and model generalization. Consider the Self Organizing
Map (SOM) for example; it could predict the
flood no better than an educated guess, and hence
not particularly useful. After closer inspection into
our raw data, it was revealed that in those cases
the over-fitting had overruled model generalization.
More specifically, SOM ranked Terrain Height as almost
the least significant factor due to the fact that
Pathumthani terrain is generally flat; therefore flood
probability was considered by SOM to be more dependent
on other factors. This is however, not the
case in some areas that survived the flood, mostly due
to its relatively raised level. Discussion on shortcomings
of the remaining techniques can also be found in
[4]. Last but not least, research in MLP and RBF
are relatively matured and as such they are currently
widely accessible to hydrological organizations, whose
prolonged learning curve in adopting cutting edge AI
paradigm could cause an adverse effect for unhindered
risk management.