Geoscience Reference
In-Depth Information
Table 2.
Performance of ANNs and MLR models for Longchuang River.
Calibration
Validation
RMSE
∗
R
2
RMSE
∗
R
2
Model
Group I
1
252.07
0.6200
215.42
0.6624
2
229.47
0.6856
182.59
0.7575
3
226.26
0.6943
186.83
0.7461
4
212.96
0.7292
183.01
0.7564
ANN
Group II
5
244.93
0.6305
202.31
0.7023
6
206.15
0.7382
169.31
0.7915
7
213.61
0.7189
170.60
0.7883
8
206.75
0.7367
171.86
0.7852
Group III
9
190.70
0.7829
133.77
0.8696
10
192.20
0.7794
134.10
0.8692
11
190.88
0.7824
145.54
0.8459
12
188.18
0.7886
139.23
0.8590
MLR
A
233.07
0.6760
234.19
0.6011
B
225.89
0.6950
226.74
0.6260
C
200.58
0.7600
186.92
0.7458
∗
Unit (kg/s).
ANN 9, respectively, were calibrated and validated. Their RMSE and
R
2
are given in Table 2.
The selection of the input variables plays an important role in the accu-
racy of the network. The networks in Group I show that a successful sim-
ulation of sediment discharge cannot be made by using averaged rainfall
and temperature only as inputs (Table 2). When the variables representing
the storm event were added as inputs in Group II, the performance of the
network increased significantly. Adding water discharge as input in Group
III can further increase the performance of the models. In addition, adding
input variables at previous time steps to the network could improve the
prediction. However, the degree to which the information from previous
months should be involved depended on the physical relationship between
the input variable and the output.
The linear relationships between the observed and the predicted sedi-
ment discharges by ANNs are better than those by MLR models (Fig. 2).
Furthermore, ANNs generate more reasonable predictions for the points
with low values where MLP models may give negative predictions, due to
the nonlinear transformation process involved (Fig. 2).
