242
0.8
0.7
~
0.6
0.5
~ 0.4
CI:)
0.3
f-;
"0
0.2
• o
•
•
•• o
M. Gevrey . S. Lek . T. Oberdorff
(a)
0.1
•
o~~~·~~·~·~~·r·--~--~----~·---r------~
o
1000
2000
3000
4000
0.8
o ••
(b)
0.6
•
!
•
0.4
•
[g 0.2
0
• • 0"
"0
•
0
h_ o
~
. .. ...
' " .... '
-0.2
0
500
1000
1500
2000
2500
3000
NPP
Fig 12.6. Partial derivatives of the ANN model response (TSR) with respect to
each independent variables (Pad algorithm, Derivatives Profile) (a) SAD, (b) NPP.
12.5
Discussion
The power of Artificial Neural Networks is verified by a very high determination
coefficient between the observed values and the estimated values for both models
(r 2 =0.92 for ESR prediction and r 2 =0.86 for TSR prediction). The results are in
agreement with the literature, in which ANN performances have repeatedly been
reported to surpass those of more traditional methods such as MLR (cf. reference
cited in the introduction). This may point to the predominantly non-linear
relationships between the studied variables on the one hand, and on the other
hand, the ability of ANNs to take direct1y into accounts any non-linear
relationships between the dependent variables and each independent variable (Lek