Data-Based Evapotranspiration Modeling - Hydrological Data Driven Modelling - page 225

to give a consistent performance with less variability with change in inputs. The
better performance of LLR indicates that even a less complex model with a good
scienti

c base can replicate the natural processes effectively if the learning process
is proper and effective.

7.8 Discussions and Conclusions

As stated earlier, this case study was performed using data sets from three major
catchments: the Brue catchment, UK, the Santa Monica station, USA, and the
Chahnimeh reservoirs region, Iran. In Sect. 7.1 we have concentrated on the tra-
ditional Penman empirical equations for evapotranspiration estimation for different
time steps using the data from the Brue catchment and Santa Monica Station, and
have compared their modeling results with MLP arti cial neural networks. The

analysis used Penman mathematical models such as FAO56-PM, ASCE-PM, CI-
MIS-PM, and a new empirical equation (Copais Approach) for analysis. The study
emphasized the capabilities of the FAO56-PM equation in ET 0 modeling on data
rich catchments and concluded that ANN can be successfully applied in both data
rich and scare situations for modeling evapotranspiration. It has been found that the
relatively less data intensive model Copais Approach could perform well in all
seasons, except summer, in different time scales of analysis, ranging from hourly to
annual. In the second major section, we have used the daily evaporation data from
Chahnimeh reservoirs region of the Iran.
This case study extensively evaluated the performance of GT in identifying the
input structure and training data length in the context of evaporation modeling. The
study has used different model selection approaches such as the AIC, BIC, entropy
theory, and traditional approaches such as data splitting and cross correlation
methods on the daily data from the Chahnimeh reservoirs region of the Islamic
Republic of Iran, to check the authenticity of GT analysis. The analysis with the GT
identi

ed three different combinations — [W, RH, Ed], [W, T, Ed], and [W, T, RH,

Ed] with training data lengths of 2413, 2327, and 1325, respectively, considering

different attributes such as gradient, Gamma statistics, and error. This highlights the
need to evaluate the GT results through modeling and effective comparison to
locate the best input structure and training length, rather than merely trust on the
lowest Gamma static value. A collaborative work with this author [ 21 ] trusted the
[W, RH, Ed] input structure and length of 2413 for further modeling without any
authenticity check. Moghaddamnia et al. [ 19 ] have used the same GT suggested
inputs for modeling with different ANN and ANFIS techniques. The GT gave
results in terms of lowest value of Gamma static which contradicted many estab-
lished research projects which highlight

—

the importance of air temperature on

''

evaporation modeling [ 12 , 16 It has

been found that the relative importance of inputs is W > Ed > RH > T. The
signi

− 17 , 24 ]. Moghaddamnia et al. [ 19 ] noted that

cance of the daily mean temperature data was relatively small when com-
pared with other weather variables since the elimination of this input made small