In the preliminary study [4], where prominent
learning based prediction systems, i.e., ANN, GA,
Fuzzy Logic and SOM were benchmarked, generic
ANN was found to be the most superior in terms
of flood predictability. In that study, key factors
which were incorporated into the model were rainfall,
flood plain in the past, height above sea level, water
density, water blockage, sub-basin area, soil drainage
ability, and land utilization. Based on the total of
100 spatial data, which were uniformly sampled from
the flood event recorded in 2011, several statistical
analyses were employed to validate and compared
those learning models [4]. It was reported that, for
instance, ANN and Fuzzy Logic performed equally
well, with the best accuracy of 92%. However, when
considering their generalization ability, using Leaveone-
out cross validation (LOOC) on 6 but 1 district,
and taking into account both false positive and negative,
ANN could predict the event with the least error
of 30.2%, as shown in Table 1.