Geology Reference
In-Depth Information
arti
cial intelligence models for application in hydrology. They have reviewed
more than 43 journal papers in hydrology and pointed out that in most cases the
inputs were chosen arbitrarily without any scienti
c reasoning and some studies
used a trial and error approach or validation data. A study by Bowden et al. [
19
]
gives an extensive review of the background and methodology adopted in input
determination for neural network models in water resource applications. They have
classi
ed major attempts in input data selection in water resources into
ve
categories:
1. Relying on prior knowledge of the system
ASCE Task Committee on Application of Arti
cial Neural Networks in
Hydrology [
9
] has pointed out the predominant use of a priori knowledge of the
system as an indicator of model selection in hydrology. Studies such as
Campolo et al. [
20
] and Jayawardena et al. [
38
] have used modelers
'
expert
knowledge on the system and the study condition to select the in
uencing
inputs. Some studies have bene
ted from the combination of a priori knowledge
and analytical approaches [
47
,
48
]. Factors such as large dependency on an
expert
s knowledge, very subjective nature, and case dependency are considered
as disadvantages of such methods.
2. Based on linear cross-correlation
Cross-correlation methods are the most common and popular analytical tech-
niques for selecting appropriate inputs [
35
,
36
,
67
]. In this approach research
normally depends on linear cross-correlation analysis values to determine the
strength of the relationship between the input time series and the output time
series at various lags [
31
]. The disadvantage associated with this method is its
inability to capture any nonlinear dependence that may exist between the inputs
and the output. The cross-correlation method works on linear dependence
between two variables, so there is a good chance of the omission of important
inputs that are related to the output in a nonlinear fashion.
3. Based on heuristic approach
In this approach, various models are trained using different subsets of inputs. In
this method, some researchers often employ stepwise selection of inputs such as
forward selection and backward elimination to avoid total enumeration [
75
]. The
concepts of these two approaches are self-explanatory from the name itself.
Forward selection is the most common approach, in which we try to
'
nd the best
single input and select it for the
final model [
48
]. Backward elimination works in
the opposite way; it starts modeling with a set of all inputs, and sequentially
removes the input set which reduces performance least. Most of the heuristic
approaches are computationally intensive trial and error procedures and there is
no guarantee that they will
find the globally best subsets.
4. Methods that extract knowledge contained within trained ANNs
In this type of method, researchers mostly depend on sensitivity analyses to
extract information from a trained ANN [
45
,
61
]. Abrahart et al. [
5
] used a novel
concept known as saliency analysis to disaggregate a neural network solution in
