Geology Reference
In-Depth Information
terms of its forecasting inputs. The saliency analysis was achieved by setting
one input data stream at a time to zero and then performing the modeling,
replacing the input data stream after the computation, then repeating this process
on the next data set, and so on. This approach determines the relative importance
of each input by examining the change in forecasting error and the plots from
the
flood hydrograph. Abrahart et al. [
5
] claimed superiority of the saliency
approach over sensitivity analysis, as sensitivity analysis does not investigate
the rate of change of one data variable with respect to the change in another.
Bowden et al. [
19
] suggests the disadvantages of their approach are (1) lack of
retraining the ANN after removing each input and (2) the possibility of pro-
ducing nonsensical outputs due to the presence of zero inputs.
5. Methods that use various combinations of the above four approaches
Some studies have used effective combinations of the above-mentioned methods
in data selection [
4
,
61
,
67
]. Abrahart et al. [
4
] used a genetic algorithm (GA)-
based approach to optimize the inputs to an ANN model used to model runoff.
Approaches such as Pearson correlation, stepwise forward regression analysis,
and sensitivity analysis were used by Schleiter [
61
] to select appropriate inputs
for water quality modeling.
The above-mentioned approaches are widely used, even in multiple linear
regression models, although many disadvantages are associated with them [
46
].
Another possible approach associated with models with nodes is the method cited
in the previous section, the
There are very powerful pruning
algorithms available which are used effectively for input variable selection in other
“
pruning approach.
”
fields of engineering [
74
]. The study by Livingstone et al. [
46
] has pointed out the
relevance of selection of effective modeling of the responsive variables (data series)
to the success of nonlinear models, which is dictated by the data.
Bowden et al. [
19
] proposed state-of-the-art methods such as the Partial mutual
information algorithm (applied for calculation of dependence in the case of multiple
inputs), the Self-organizing map (SOM) (used to reduce the dimensionality of the
input space and obtain independent inputs), the GA, and the General regression
neural network (GRNA) (applied to determine which inputs have a signi
cant
relationship with the output (dependent) variable) for input selection of ANN
models.
2.4 Redundancy in Input Data and Model
Hydrologists often face challenges in identifying redundant input data during the
preprocessing period as the sets of possible inputs into a hydrological system are
huge. This process becomes more challenging in the modeling of some hydro-
logical processes, as all measurable variables are highly nonlinear in dynamics and
have multiple interrelations. The normal practice for data-based model practitioners
