Fig. 1.
Study Area and Long-Term Hydrological Data Variations.
Fig. 2.
CNN-BP model structure
Fig. 3.
Random Forest results of factor screening. (a) Feature importance of 20 rainfall stations. (b) Spatial distribution of 14 chosen and 6 unchosen rainfall stations.
Fig. 4.
Radar charts of R
2
values associated with CNN-BP and BPNN models at T
+
1 and T
+
3 for 25 groundwater monitoring stations.
Fig. 5.
Time delays (
T
delay
) in forecasts of BPNN and CNN-BP models at testing stages for individual groundwater monitoring station (G1
–
G25) based on CCF.
Fig. 6.
Performance comparison between CNN-BP and MLR models during testing stages for groundwater level forecasting at stations G3, G16 and G20.
Fig. 7.
Forecast biases of the CNN-BP model in regional groundwater levels across four seasons at the testing stage. STD stands for standard deviation.
Fig. 8.
Analysis between regional average lithology (Tr) and groundwater level forecast biases of the CNN-BP model at the testing stage. (a) Log(Tr). (b) Forecast biases.
Fig. 9.
Regression of CNN-BP
’
s forecast bias (m) and the logarithm of transmissivity (Log(Tr)).
Fig. 10.
Spatial interpolation maps for Typhoons Haitang and Maria.
“
Potential Recharge
”
is defined as the differences in groundwater level before and after the typhoon period. (a)
&
(d) Accumulated rainfall. (b)
&
(e) Observed groundwater
“
Potential Recharge
”
. (c)
&
(f) Forecasted groundwater
“
Potential Recharge
”
obtained from the CNN-BP model at T
+
3.
Fig. 11.
Analysis between CNN-BP
’
s groundwater forecasts and their social impact. (a) Population density of Zhuoshui River basin. (b) Number of months with SPI≤
|