Next Article in Journal
The Use of Algae and Fungi for Removal of Pharmaceuticals by Bioremediation and Biosorption Processes: A Review
Previous Article in Journal
Treatment of High-Strength Animal Industrial Wastewater Using Photo-Assisted Fenton Oxidation Coupled to Photocatalytic Technology
 
 
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Modeling and Investigating the Mechanisms of Groundwater Level Variation in the Jhuoshui River Basin of Central Taiwan

1
State Key Laboratory of Eco-Hydraulics in Northwest Arid Region of China, Xi’an University of Technology, Xi’an 710048, China
2
Department of Civil and Environmental Engineering, Pennsylvania State University, University Park, PA 16802, USA
3
Institute of Hydrology and water resources, Zhejiang University, Hangzhou 310058, China
4
Department of Bioenvironmental Systems Engineering, National Taiwan University, Taipei 10617, Taiwan
5
Department of Water Resources and Environmental Engineering, Tamkang University, New Taipei City 25137, Taiwan
6
Water Resources Agency, Ministry of Economic Affairs, Taipei 10651, Taiwan
*
Author to whom correspondence should be addressed.
Water 2019, 11(8), 1554; https://doi.org/10.3390/w11081554
Submission received: 11 May 2019 / Revised: 18 July 2019 / Accepted: 24 July 2019 / Published: 27 July 2019
(This article belongs to the Section Hydrology)

Abstract

:
Due to nonuniform rainfall distribution in Taiwan, groundwater is an important water source in certain areas that lack water storage facilities during periods of drought. Therefore, groundwater recharge is an important issue for sustainable water resources management. The mountainous areas and the alluvial fan areas of the Jhuoshui River basin in Central Taiwan are considered abundant groundwater recharge regions. This study aims to investigate the interactive mechanisms between surface water and groundwater through statistical techniques and estimate groundwater level variations by a combination of artificial intelligence techniques and the Gamma test (GT). The Jhuoshui River basin in Central Taiwan is selected as the study area. The results demonstrate that: (1) More days of accumulated rainfall data are required to affect variable groundwater levels in low-permeability wells or deep wells; (2) effective rainfall thresholds can be properly identified by lower bound screening of accumulated rainfall; (3) daily groundwater level variation can be estimated effectively by artificial neural networks (ANNs); and (4) it is difficult to build efficient models for low-permeability wells, and the accuracy and stability of models is worse in the proximal-fan areas than in the mountainous areas.

1. Introduction

Due to industrial and economic development along with rapid population growth in Taiwan in recent years, more water sources are required to satisfy the increasing water demands from the development of human civilization and livelihoods. Due to its easy accessibility, low cost, and stable quality and temperature, groundwater is often considered an important water source. The interactive mechanism of groundwater level variations is rather complex and includes precipitation, tides, atmospheric pressure, and even earthquakes. Previous research has shown that rainfall can be regarded as a leading indicator of groundwater levels by applying time series to analyze the relationship between rainfall and groundwater levels [1,2,3,4,5,6,7]. However, the temporal relationship between groundwater and rainfall might vary widely in different regions. In this study area, Taiwan, some research has suggested that time lags between rainfall and groundwater level variation might occur [8]. Chen [9] identified that different durations of rainfall affected groundwater levels and that response time spans range from a few hours to a few days.
Artificial neural networks (ANNs) are recognized as an effective technique for modeling complicated systems (nonlinear issues) and have been widely adopted in various hydro system problems in past decades. The advantages of ANNs include their ability to learn from datasets, their flexible noise tolerance, and their ability to generalize. Additionally, ANNs can extract significant and meaningful features from complex data structures and can learn the implicit relationship between inputs and outputs based on sufficient training data [10,11]. Pradhan and Lee [12] used a back-propagation neural network (BPNN) coupled with a geographic information system (GIS) to analyze the risk of landslides. Karthikeyan et al. [13] predicted ground water levels in India by using two networks, feed forward neural networks (FFNN) and recurrent neural network (RNN), and adjusting the parameters according to the related input factors revealed that the hydro-meteorological parameters affect the models. Sreekanth et al. [14] predicted groundwater levels to deal with the problems of water resource usage through FFNN-LMB applied to groundwater level forecasting and verified that ANN is a promising tool. Li et al. [15] executed the back-propagation (BP)coupled with a sensitivity analysis to determine that the most sensitive anthropogenic factor, coal mining drainage, has a major effect on groundwater levels in the Jinci Spring Basin in northern China. Mohanty et al. [16] used various methods to construct models of groundwater systems and found that ANN models provided better short-term groundwater level prediction than MODFLOW-based numerical models without sufficient parameters, data and boundary conditions. New techniques such as the neuro-fuzzy inference system that integrates ANNs and fuzzy logic methods have also proven powerful. The neuro-fuzzy inference system has the potential to incorporate the strengths of both ANN and fuzzy logic in one framework. The adaptive neuro-fuzzy inference system (ANFIS) was introduced to provide a new approach on the reservoirs’ optimal operations [17] and rainfall-runoff simulations [18].
Some studies have focused on the interactive mechanism of surface water and groundwater in mountainous areas. The purpose of this study is to explore the interactive mechanism of surface and subsurface water by analyzing and modeling groundwater level variations using statistical and artificial intelligence techniques. The approaches consist of correlation analysis between precipitation and groundwater level variation and the estimation of groundwater level variations. The Jhuoshui River basin in Central Taiwan is selected as a study area.

2. Methodology

This study investigates the interactive mechanism of groundwater level variations through artificial neural networks, statistical analyses, the Gamma test and sensitivity analysis. The methodologies used in this research are briefly addressed as follows:

2.1. Pearson Correlation Coefficient

The Pearson correlation coefficient [19] is a quantitative method used to measure the linear correlation between two variables. In this study, it is used to determine the temporal relationship between groundwater level variations and precipitation as well as stream flow. The Pearson correlation coefficient ranges between [−1,1], in which 1 denotes a positive exact linear relationship, −1 indicates a negative perfect linear relationship, and 0 indicates that no linear relationship exists between the two variables.
For variables of P (precipitation) and G (groundwater) with n datasets (each dataset can be denoted by pi and gi, where i = 1, 2, …, n), the correlation coefficient can estimate the Pearson correlation (R) between P and G. The calculation of the correlation coefficient is given as:
R P , G = i = 1 n ( p i P ¯ ) ( g i G ¯ ) i = 1 n ( p P ¯ ) 2 i = 1 n ( g i G ¯ ) 2
where P ¯ and G ¯ are the means of P and G accordingly.

2.2. Gamma Test (GT)

The Gamma test (GT) is designed to estimate noise variance by measuring the smooth relationship between the input-output dataset. The GT was initially proposed by Končar and Aðalbjörn [20,21] to determine the best input combination for a neural network [22,23]. This study implements the GT for the extraction of nonlinearity between groundwater level variations and precipitation, along with the identification of important factors affecting groundwater level variations.
Given an input-output pair denoted by (X, y) = ((x1, …, xm), y), where X is the input items and scalar y is the output items, the observations can be described as Equation (2).
{ ( X i , y i ) , 1 i M }
where, XiRm are the m-dimensional input variables with a dataset length of M, which is constrained to a closed bounded set CRm. The corresponding outputs yiR are scalars. The underlying relationship of the data set can be denoted:
y = f ( x 1 , x m ) + r
where, f is a smooth function, and r denotes a random variable for noise. The Gamma statistic (Γ) is a variance estimation of noise. Let Xi,k denote the kth nearest neighbor to Xi in terms of Euclidean distance. The delta function is defined:
δ M ( k ) = 1 M i = 1 M | X i , k       X i | ( 1 k p )
where, |⋯| is the Euclidean distance, and p is the number of neighbors. For computing the Γ, the line of least squares regression is built for p points:
γ = A δ + Γ
where, A is the gradient. The steeper the gradient is, the more complex is the model. The Γ is an index to evaluate the model performance of an input combination. A Γ closer to 0 implies a more appropriate input combination.

2.3. Back-Propagation Neural Network (BPNN)

The back-propagation neural network (BPNN), developed by Rumelhart et al. [24], is used to build a model for estimating the groundwater level variations in this study. Figure 1 shows the BPNN structure in this study.
The input items comprise the streamflow and accumulated rainfall. The node number of the hidden layer is determined by trial and error. The model target is groundwater level variation. The BPNN applies the gradient steepest descent method for adjusting the neuron weights for minimizing the output error. In the learning process (calculation iterations), the neuron weights can be adjusted by an error convergence technique to approximate the model target based on the given input set. The output layer error may propagate backward to the input layer. Further details of the BPNN algorithm are found in Rumelhart et al. [24].

2.4. Adaptive Neuro-Fuzzy Inference System (ANFIS)

The adaptive neuro-fuzzy inference system (ANFIS) was proposed by Jang [25] who used the fuzzy inference system as an essential core in combination with the artificial neural network. The ANFIS preserves the learning process of ANNs for projecting input features to an output space and retains the fuzzy advantages, including if-then rules (rule layer), for describing the regional behavior of such projection. Then, the results of the inference system are acquired through the reasoning capability of fuzzy logic. The ANFIS was shown to have powerful modeling abilities in comprehensive fields such as motor fault detection and diagnosis [26], power systems dynamic load [27], forecasting systems for the demand of teachers’ human resources [28] and real-time reservoir operations [29].
This study applies the Takagi-Sugeno fuzzy model [30] in a fuzzy rule layer. Here, 2 input variables, p (precipitation) and s (streamflow), and 1 output variable, g (groundwater), are taken as an example for describing the rule layer. The rule sets of the rule layer can be expressed as:
  • Rule 1: If p is A1 and s is B1 then g1 = k1*p + t1*s + r1
  • Rule 2: If p is A2 and s is B2 then g2 = k2*p + t2*s + r2
where k, t and r are linear parameters in the consequent part (then-part) of the first-order Takagi-Sugeno fuzzy model.
The typical ANFIS includes 5 layers (Figure 2). The input layer—in this layer, each node produces membership degrees that belong to each of the fuzzy sets by using appropriate membership functions. The rule layer—the AND operator is used to obtain one output that represents the result of the antecedent for that rule. The average layer—the ratio of each ith rule’s firing strength to the sum of all the rules’ firing strength is calculated. The consequent layer—the node function determines the contribution of each ith rule’s toward the total output. The output layer—the defuzzification process transforms each rule’s fuzzy results into the model output. The details of the ANFIS are found in Chang and Chang [31].

3. Case Study

This study investigates the interactive mechanism of groundwater level variations and precipitation in mountainous areas in the Jhuoshui River basin by using statistical methods. The Jhuoshui River basin is located in Central Taiwan with a maximum elevation of 3200 m. The watershed area of Jhuoshui River is approximately 3157 km2. Figure 3a shows the distribution of the gauge stations and the hydrogeological information of groundwater monitoring wells in the study area. There are 2 main regions in this study area: The mountain area and the alluvial fan. The alluvial fan can be roughly divided into 3 districts: The fan-top district; the fan-mid district; the fan-tail district (Figure 3a). The G1~G4 monitoring wells belong to the mountain area, and the G5~G8 monitoring wells belong to the fan-top of the alluvial fan. As shown in Figure 3b, the fan-top district is the only district without a significant confined aquifer. In general, the hydrogeological feature of the fan-top district in this study area is recognized as a high potential groundwater recharge area because the main geological components are sandstone, phyllite and slate (the gravel of the geological structure) (Figure 3b). According to investigations of core drilling samples, the soil thickness of this study area (the depth to bedrock, including the soil layer, colluvium layer and saprolite) is deep [32]. The Central Geological Survey (CGS) (Ministry of Economic Affairs (MOEA), R.O.C.) indicated that the hydrogeological structure of the study area can be divided into several strata. Based on the depth from the land surface (Figure 3b), these strata are as follows: Aquifer 1 (F1), Aquitard 1 (T1), Aquifer 2 (F2), Aquitard 2 (T2) and Aquifer 3 (F3). The average thicknesses of Aquifer 1, Aquifer 2 and Aquifer 3 are 42 m, 95 m and 86 m, respectively. The thickness of the gravel and sand strata in the fan-top areas can reach more than 130 m [33]. In this study area, some groundwater monitoring wells include two different depth monitoring records (in a different aquifer), such as G1, G4, G7 and G8 (Figure 3c). However, according the hydrogeological information provided by the CGS (Figure 3b), it is recognized that the shallow groundwater well in G7 belongs to Aquifer 3 (F3). The rainfall is distributed unevenly, which mainly occurs from May to September due to the unique topographical terrain and location. The total rainfall in wet periods comprises 75% of the annual rainfall, which implies rainfall differs significantly between wet and dry periods.
The groundwater level data at twelve groundwater monitoring wells and the flow data at two streamflow gauging stations were collected from the Water Resources Agency in Taiwan during 2001 and 2011, and rainfall data at seventeen rainfall gauging stations were collected from the Water Resources Agency, Central Weather Bureau and Taiwan Power Company in Taiwan in 2001 and 2011 (Table 1, Table 2 and Table 3). The missing data were infilled by using linear regression techniques based on the data of the surrounding stations. The flowchart of this study is shown in Figure 4. First, this study conducts rainfall duration analyses by investigating the effective duration of accumulated rainfall on groundwater level variations. Second, this study conducts effective rainfall analysis by identifying significant accumulated rainfall amount thresholds affecting groundwater level variations, and therefore the correlation between accumulated rainfall amounts and groundwater level variations can be estimated. Third, this study constructs estimation models of the groundwater level variations by using the BPNN and ANFIS based on rainfall, streamflow and groundwater level variation data, in which the GT is used to determine the proper input variables associated with the rainfall gauging stations as well as the streamflow gauging stations (S1), and rainfall durations for each model at individual groundwater monitoring stations. Finally, the performance of the models is compared.
The Pearson correlation coefficient and the root mean square error (RMSE) are used to evaluate the performance of the estimation model. The calculation of RMSE is given below:
RMSE = 1 N i = 1 N ( y i d i ) 2
where, yi is the model estimation and di is the observation, and N is the number of datasets. The RMSE is used to evaluate the accuracy of the estimations of the groundwater level variations. The lower the RMSE value is, the better is the model’s performance.

4. Results and Discussion

4.1. Duration of Accumulated Rainfall Analysis

First, this study uses the Pearson correlation coefficient to investigate the relationship between the duration of accumulated rainfall (1–10 days) and the groundwater level variation. A high correlation coefficient indicates a strong relationship between the accumulated rainfall and the groundwater level variation. To generate a comprehensive data set, rainfall levels at each rainfall gauging station can be determined by the average rainfall over the basin obtained from the Thiessen polygon method [34]. Figure 5 shows the time series of groundwater levels at the groundwater monitoring wells and the rainfall data at R2. This study also analyzed the relationship between the groundwater level variation and the rainfall at a 1-day lag (from the current day to the previous day), and further compared the results with a duration analysis. At all groundwater monitoring stations, the results show that the correlation coefficient of groundwater level variations and accumulated rainfall with different durations (2–10 days) is higher than the rainfall at the current day and 1-day lag. In other words, the accumulated rainfall has a stronger positive relationship with the groundwater level variation than the rainfall of the previous day at all wells.
As shown in Figure 6, the highest correlation between the groundwater level variation and two-day accumulated rainfall occurs at G4(1), G4(2) and G2, which might result from the gravel of the geological structure at G4 and G2. The highest correlation between groundwater level variations and three-day accumulated rainfall occurs at G1(1) and G3, which may be because of the shallow well depth of G1(1) and G3. The highest correlation between the groundwater level variation and a five-day accumulated rainfall occurs at G1(2), which might be caused by the deep well depth of G1(2). Therefore, the accumulated rainfall over different durations is selected as an input variable in the ANN models for estimating the groundwater level variations. In addition, the differences in geological structure and depth at each groundwater monitoring well produces different relationships between the groundwater level variations and the accumulated rainfall of different durations.

4.2. Effective Rainfall Analysis

The different quantities of rainfall may have different impacts on the groundwater level variation. This study considers the average rainfalls associated with different thresholds for evaluating the rainfall amount impacts on the groundwater level variations at G1 to G4 by the use of the Pearson correlation coefficient. The duration of the accumulated rainfall adopts the result of Section 4.1 Duration of accumulated rainfall analysis: G4(1), G4(2) and G2—two days; G1(1) and G3—three days; G1(2)—five days. This study includes two types of screening:
(1) Lower bound screening: Delete the data sets below the rainfall threshold
The rainfall data (a total more than 3986) were screened at thresholds from 1–50 or 1–100 mm (with 1 mm increasing). Figure 7 shows the Pearson correlation between the groundwater level variations and the accumulated rainfall at different lower bound thresholds. In Figure 7a, the correlation patterns of G4(1) and G4(2) change slightly, and the correlation curve rises at the 13 mm threshold of the accumulated rainfall. This might be due to the greater structural permeability of G4 (gravel), as well as the location of G4 close to the Jhuoshui river, so that the groundwater level at G4 is affected by rainfall and the stream flow at the same time. The correlation curve of G2 rises obviously at the accumulated rainfall threshold of 26 mm. It is suspected that this causes saturation when the accumulated rainfall is 26 mm, and furthermore, the excess rainfall causes percolation to recharge the groundwater layer. Therefore, the effective rainfall thresholds affecting the groundwater level variation at G2 can be identified as 26 mm. In Figure 7b, the correlation pattern of G1(1) and G3 are uniform. At G1(1), the correlation increases gradually at the accumulated rainfall threshold of 34 mm, and the highest correlation coefficient is 0.8 at the accumulated rainfall threshold of 63 mm. In Figure 7c, the correlation pattern of G1(2) is similar to G2. The correlation curve of G1(2) rises obviously at the threshold of 48 mm of the accumulated rainfall, which causes effective groundwater recharge activity when the accumulated rainfall rises to 48 mm in five days.
(2) Upper bound screening: deleting the data sets over the rainfall threshold
Figure 8 demonstrates the Pearson correlation between the groundwater level variations and the accumulated rainfall at different upper bound thresholds. All the groundwater monitoring wells show the same correlation patterns well.
With the higher upper bound rainfall threshold (which retain more heavy rainfall data), the correlation coefficient increases. This indicates that the groundwater level variations increase with more accumulated rainfall. This study can identify the lower bound threshold of the accumulated rainfall affecting the groundwater level effectively, but the authors cannot identify the upper bound threshold of accumulated rainfall.

4.3. Estimation of Groundwater Level Variations

This study built two estimation models (BPNN and ANFIS) for the groundwater level variations for each groundwater monitoring well. The learning and training efficiency of ANN models decrease if the data from all hydrology information is incorporated into the models. Therefore, the GT was used to determine the critical input factors (i.e., rainfall and streamflow) that correlated strongly with the groundwater level variations for ANN models. According to Chang et al. [35], it is better to use rainfall gauging stations as well as streamflow gauging stations as simultaneous input factors while constructing the groundwater level variation models at G1 to G4. Therefore, the study adopts the GT to evaluate the critical rainfall stations from G1 to G4, and streamflow gauging stations S1 and S2 are used as two other inputs for the ANN models. However, due to uncertainty as to whether or not the streamflow information is the critical factor of the groundwater level in the proximal-fan areas, this study adopts the GT to determine the critical rainfall gauging stations and streamflow gauging stations at G5 to G8. The GT is also used to detect rainfall durations (2–5 days) that have the highest correlations with the groundwater level variations at all wells. Table 4 shows the results of the GT analysis.
The node numbers of the hidden layer in BPNN and the rule numbers in ANFIS were determined by trial and error. The optimal model structure is determined by the lowest RMSE value and the highest correlation coefficient in the validation phases at each groundwater monitoring station. The observations were divided into the training, validation and testing phases in the ratio of 5:2.5:2.5 or 6:2:2. Table 5 shows the estimation results of BPNN and ANFIS at each groundwater monitoring well (G1 to G8), and Table 6 shows the statistics of the observed data.
The results indicate that the ANN models provide an effective simulation and prediction of the groundwater level variation because the observations and the estimations have high correlations and small RMSE values. When comparing Table 5 and Table 6, the RMSEs are smaller than the standard deviation at G1 to G4. When comparing the two types of models, the estimations from BPNN are better than those from the ANFIS model. In particular, BPNN shows more than 23% improvement in terms of the correlation coefficient at G1(1), G2 and G5. Therefore, when comparing the two areas of this study, it is easier to predict the groundwater level variation with adequate reliability and accuracy in mountainous areas than in proximal-fan areas.

5. Conclusions

Water resource deficiency is a global problem, especially under severe climate change. In Taiwan, groundwater is considered an important alternative water resource due to its low cost and convenient accessibility. The improper groundwater development may result in disasters coupled with environmental and economic losses, and strategic development for groundwater conservation is therefore a critical issue. Precipitation is the main source of groundwater recharge, so this study used statistical methods to explore the relationship between precipitation (duration and amount) and the groundwater level variation. The statistical results demonstrate that the connection of precipitation and groundwater can be described by the hydrogeological features. The geological structure (potential high-porosity and high-permeability) of the gravel and the thin soil thickness (shallow monitoring wells) resulted in a high correlation between the groundwater level variation and the short precipitation duration. The results also show that the groundwater level variation in the high-permeability geological structures would have minor effects on accumulated rainfall. In the other words, the high-permeability area can be considered to be an area sensitive to precipitation.
After identifying the relationship between precipitation and the groundwater level variations, this study built an estimation model of the groundwater level variation by artificial intelligence techniques. This study demonstrates that limited data can be used to build a robust estimation model for the groundwater level variation in the potential groundwater recharge area by artificial intelligence techniques. In other words, a model estimating the groundwater level variations with less parameter bias can be built since this model needs only the streamflow and the rainfall variables (which are easier to measure) as input items. This model feature is helpful when applied to complex hydrogeological systems. Furthermore, from the authors’ viewpoint, the groundwater level variation is an important issue for water resource management, especially in the potential recharge area. If information can be accurately provided about the variation in the groundwater level to water resources decision makers, they can develop more appropriate water resource allocations or water management policies. Overall, the model shows better performance for high-permeability areas (mountain regions) than the low-permeability area (the proximal-fan region). This study also adopted the Gamma test to deal with the complex correlation between the surface water and the groundwater. The model results show that neural network models can provide accurate performance with suitable data preprocessing. In summary, the results indicate that neural network-based estimation models perform well. These results can also provide valuable information for the prevention and treatment of land subsidence and can serve as an effective reference for water resource management in the alluvial fan and mountain areas. The results also demonstrate that the application of artificial neural networks in the study of groundwater is a promising avenue.

Author Contributions

T.B.: Literature search, study design, figure design, data interpretation and writing; W.-P.T.: Literature search, study design, figure design, data analysis, data interpretation and writing; Y.-M.C.: Literature search, study design and data interpretation; F.-J.C.: Literature search, study design and data interpretation; W.-Y.C.: Literature search, data collection and data analysis; L.-C.C.: Literature research, study design and data interpretation; K.-C.C.: Literature research and data interpretation.

Funding

National Key R & D Program of China (2017YFC0405900); Water Resources Agency, Taiwan, ROC (Grant number: MOEAWRA1010425).

Acknowledgments

The streamflow and groundwater level data provided by the Water Resources Agency, Taiwan, and the rainfall data provided by the Water Resources Agency, Central Weather Bureau and Taiwan Power Company, Taiwan, are very much appreciated. In addition, the authors are indebted to the reviewers for their valuable comments and suggestions.

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. Changnon, S.A.; Huff, F.A.; Hsu, C.F. Relations between precipitation and shallow groundwater in illinois. J. Clim. 1988, 1, 1239–1250. [Google Scholar] [CrossRef]
  2. Cheng, J.D.; Hong, H.N.; Chou, L.H. Use Time Series to Analysis Rainfall and Groundwater Level. J. Chin. Soil Water Conserv. 2003, 35, 47–56. (In Chinese) [Google Scholar]
  3. Chen, Z.; Grasby, S.E.; Osadetz, K.G. Relation between climate variability and groundwater levels in the upper carbonate aquifer, southern Manitoba, Canada. J. Hydrol. 2004, 290, 43–62. [Google Scholar] [CrossRef]
  4. Lee, L.; Lawrence, D.; Price, M. Analysis of water-level response to rainfall and implications for recharge pathways in the Chalk aquifer, SE England. J. Hydrol. 2006, 330, 604–620. [Google Scholar] [CrossRef]
  5. Nayak, P.C.; Rao, Y.R.S.; Sudheer, K.P. Groundwater level forecasting in a shallow aquifer using artificial neural network approach. Water Resour. Manag. 2006, 20, 77–90. [Google Scholar] [CrossRef]
  6. Yoon, H.; Jun, S.C.; Hyun, Y.; Bae, G.O.; Lee, K.K. A comparative study of artificial neural networks and support vector machines for predicting groundwater levels in a coastal aquifer. J. Hydrol. 2011, 396, 128–138. [Google Scholar] [CrossRef]
  7. Tsai, W.P.; Chiang, Y.M.; Huang, J.L.; Chang, F.J. Exploring the Mechanism of Surface and Ground Water through Data-Driven Techniques with Sensitivity Analysis for Water Resources Management. Water Resour. Manag. 2016, 30, 4789–4806. [Google Scholar] [CrossRef]
  8. Chen, T.S.; Jan, C.D.; Chen, S.H.; Tseng, C.M. Correlation between Groundwater Level Variations and Rainfall Depths. Taiwan Water Conserv. 2005, 53, 1–12. (In Chinese) [Google Scholar]
  9. Chen, T.S. Study on Rainfall-Induced Groundwater Water Level Variation–Case study at the Naba, Liujar, and Donher Well Stations. Ph.D. Thesis, National Cheng Kung University, Tainan, Taiwan, 2006. (In Chinese). [Google Scholar]
  10. Chang, F.J.; Kao, L.S.; Kuo, Y.M.; Liu, C.W. Artificial neural networks for estimating regional arsenic concentrations in a blackfoot disease area in Taiwan. J. Hydrol. 2010, 388, 65–76. [Google Scholar] [CrossRef]
  11. Chang, L.C.; Chen, P.A.; Chang, F.J. Reinforced two-step-ahead weight adjustment technique for online training of recurrent neural networks. IEEE Trans. Neural Netw. Learn. Syst. 2012, 23, 1269–1278. [Google Scholar] [CrossRef]
  12. Pradhan, B.; Lee, S. Regional Landslide Susceptibility Analysis Using Back-Propagation Neural Network Model at Cameron Highland, Malaysia. Landslides 2010, 7, 13–30. [Google Scholar] [CrossRef]
  13. Karthikeyan, L.; Kumar, D.N.; Gaur, S. Prediction of Ground Water Levels in the Uplands of a Tropical Coastal Riparian Wetland Using Artificial Neural Networks. Water Resour. Manag. 2013, 27, 871–883. [Google Scholar] [CrossRef]
  14. Sreekanth, P.D.; Geethanjali, N.; Sreedevi, P.D.; Ahmed, S.; Kumar, N.R.; Jayanthi, P.D.K. Forecasting Groundwater Level Using Artificial Neural Networks. Curr. Sci. 2009, 96, 933–939. [Google Scholar]
  15. Li, X.; Shu, L.; Liu, L.; Yin, D.; Wen, J. Sensitivity analysis of groundwater level in Jinci Spring Basin (China) based on artificial neural network modeling. Hydrogeol. J. 2012, 20, 727–738. [Google Scholar] [CrossRef]
  16. Mohanty, S.; Jha, M.K.; Kumar, A.; Panda, D. Comparative evaluation of numerical model and artificial neural network for simulating groundwater flow in Kathajodi–Surua Inter-basin of Odisha, India. J. Hydrol. 2013, 495, 38–51. [Google Scholar] [CrossRef]
  17. Chiang, Y.M.; Chang, L.C.; Tsai, M.J.; Wang, Y.F.; Chang, F.J. Auto-control of pumping operations in sewerage systems by rule-based fuzzy neural networks. Hydrol. Earth Syst. Sci. 2011, 15, 185–196. [Google Scholar] [CrossRef] [Green Version]
  18. Talei, A.; Chua, L.H.C.; Quek, C. A novel application of a neuro-fuzzy computational technique in event-based rainfall–runoff modeling. Expert Syst. Appl. 2010, 37, 7456–7468. [Google Scholar] [CrossRef]
  19. Pearson, K. Mathematical contributions to the theory of evolution. III.; Regression, Heredity and Panmixia. Philos. Trans. R. Soc. Lond. Ser. 1896, 187, 253–318. [Google Scholar] [CrossRef]
  20. Stefánsson, A.; Končar, N.; Jones, A.J. A note on the Gamma test. Neural Comput. Appl. 1997, 5, 131–133. [Google Scholar] [CrossRef]
  21. Končar, N. Optimisation Methodologies for Direct Inverse Neurocontrol. Ph.D. Thesis, University of London, London, UK, 1997. [Google Scholar]
  22. Moghaddamnia, A.; Gousheh, M.G.; Piri, J.; Amin, S.; Han, D. Evaporation estimation using artificial neural networks and adaptive neuro-fuzzy inference system techniques. Adv. Water Resour. 2009, 32, 88–97. [Google Scholar] [CrossRef]
  23. Noori, R.; Hoshyaripour, G.; Ashrafi, K.; Araabi, B.N. Uncertainty analysis of developed ANN and ANFIS models in prediction of carbon monoxide daily concentration. Atmos. Environ. 2010, 44, 476–482. [Google Scholar] [CrossRef]
  24. Rumelhart, D.; Hinton, G.; Williams, R. Learning Internal Representations by Error Propagation. Read. Cogn. Sci. 1988, 323, 399–421. [Google Scholar]
  25. Jang, J.S. ANFIS: Adaptive-network-based fuzzy inference system. IEEE Trans. Syst. Man Cybern. 1993, 23, 665–685. [Google Scholar] [CrossRef]
  26. Altug, S.; Chen, M.Y.; Trussell, H. Fuzzy inference systems implemented on neural architectures for motor fault detection and diagnosis. IEEE Trans. Ind. Electron. 1999, 46, 1069–1079. [Google Scholar] [CrossRef]
  27. Oonsivilai, A.; El-Hawary, M.E. Power System Dynamic Load Modeling Using Adaptive-Network-Based Fuzzy Inference System. In Proceedings of the Engineering Solutions for the Next Millennium 1999 IEEE Canadian Conference on Electrical and Computer Engineering (Cat. No. 99TH8411), Edmonton, AB, Canada, 9–12 May 1999; pp. 1217–1222. [Google Scholar]
  28. Liao, H.P.; Su, J.P.; Wu, H.M. An application of ANFIS to modeling of a forecasting system for the demand of teacher human resources. J. Educat. Psychol. 2001, 24, 1–17. [Google Scholar]
  29. Chang, L.C.; Chang, F.J. Intelligent control for modelling of real-time reservoir operation. Hydrol. Process. 2001, 15, 1621–1634. [Google Scholar] [CrossRef]
  30. Takagi, T.; Sugeno, M. Fuzzy Identification of Systems and Its Applications to Modeling and Control. Read. Fuzzy Sets Intell. Syst. 1993, 15, 387–403. [Google Scholar]
  31. Chang, F.-J.; Chang, Y.-T. Adaptive neuro-fuzzy inference system for prediction of water level in reservoir. Adv. Water Resour. 2006, 29, 1–10. [Google Scholar] [CrossRef]
  32. Chang, W.Y. Investigating the Mechanisms of Groundwater Level Variation and Recharge at Zhuoshui River Basin. Master’s Thesis, National Taiwan University, Taipei, Taiwan, 2013. (In Chinese). [Google Scholar]
  33. Tsai, J.P.; Chen, Y.W.; Chang, L.C.; Kuo, Y.M.; Tu, Y.H.; Pan, C.C. High Recharge Areas in the Choushui River Alluvial Fan (Taiwan) Assessed from Recharge Potential Analysis and Average Storage Variation Indexes. Entropy 2015, 17, 1558–1580. [Google Scholar] [CrossRef] [Green Version]
  34. Thiessen, A.H. Precipitation averages for large areas. Mon. Weather. Rev. 1911, 39, 1082–1089. [Google Scholar] [CrossRef]
  35. Chang, F.J.; Lin, C.H.; Chang, K.C.; Kao, Y.H.; Chang, L.C. Investigating the Interactive Mechanisms between Surface Water and Groundwater over the Jhuoshuei River Basin in Central Taiwan. Paddy Water Environ. 2014, 12, 365–377. [Google Scholar] [CrossRef]
Figure 1. Architecture of the back-propagation neural network (BPNN) model.
Figure 1. Architecture of the back-propagation neural network (BPNN) model.
Water 11 01554 g001
Figure 2. Architecture of the adaptive neuro-fuzzy inference system (ANFIS) model.
Figure 2. Architecture of the adaptive neuro-fuzzy inference system (ANFIS) model.
Water 11 01554 g002
Figure 3. The distribution of the gauge stations and hydrogeological information of the groundwater monitoring wells.
Figure 3. The distribution of the gauge stations and hydrogeological information of the groundwater monitoring wells.
Water 11 01554 g003aWater 11 01554 g003b
Figure 4. Implementation flowchart of this study.
Figure 4. Implementation flowchart of this study.
Water 11 01554 g004
Figure 5. Time series of the groundwater level at the groundwater monitoring wells and the rainfall data at R2.
Figure 5. Time series of the groundwater level at the groundwater monitoring wells and the rainfall data at R2.
Water 11 01554 g005
Figure 6. The correlation coefficient between the groundwater level variation and rainfall durations at different groundwater monitoring wells.
Figure 6. The correlation coefficient between the groundwater level variation and rainfall durations at different groundwater monitoring wells.
Water 11 01554 g006
Figure 7. The correlation between the groundwater level variations and the accumulated rainfall amount at different low thresholds (Lower bound screening).
Figure 7. The correlation between the groundwater level variations and the accumulated rainfall amount at different low thresholds (Lower bound screening).
Water 11 01554 g007aWater 11 01554 g007b
Figure 8. The correlation between the groundwater level variations and the accumulated rainfall amount at different up thresholds (Upper bound screening).
Figure 8. The correlation between the groundwater level variations and the accumulated rainfall amount at different up thresholds (Upper bound screening).
Water 11 01554 g008aWater 11 01554 g008b
Table 1. The statistics of seventeen rainfall stations.
Table 1. The statistics of seventeen rainfall stations.
Rainfall StationElevation (m)Rainfall (mm)
MeanSD 1Annual Rainfall
R14008.1829.162985
R22318.4330.643076
R32036.2420.772278
R42156.3521.332320
R52965.9322.962166
R63935.7620.992103
R77248.1437.452970
R83225.6421.602059
R916667.3126.502669
R104858.9129.343250
R1122007.8430.562863
R1211356.0425.602203
R1312007.3826.332695
R1415207.0127.072558
R1523035.4622.401991
R16825.1519.301880
R171104.4418.331620
1 Standard deviation.
Table 2. The basic statistics of twelve groundwater monitoring wells.
Table 2. The basic statistics of twelve groundwater monitoring wells.
Monitoring WellWell Depth (m)Elevation (m)Groundwater Level (m)
MeanSD
G1(1) [shallow]102.6151.2141.71.86
G1(2) [deep]199.3151.2142.71.46
G2150113.3109.04.63
G324.1179.3169.11.03
G4(1) [shallow]78.2151.1137.62.17
G4(2) [deep]193.2151135.51.45
G5112.782.438.753.35
G69672.338.482.84
G7(1) [shallow]14049.535.972.38
G7(2) [deep]26949.634.022.31
G8(1) [shallow]38.746.634.731.55
G8(2) [deep]97.546.534.731.52
Table 3. The basic information of two streamflow gauging stations.
Table 3. The basic information of two streamflow gauging stations.
Stream Flow StationElevation (m)Discharge (cm s)
MeanSD
S1107.17136.51401.22
S2279.09113.59272.19
Table 4. The results of the Gamma test at the groundwater monitoring wells.
Table 4. The results of the Gamma test at the groundwater monitoring wells.
Groundwater Monitoring WellInput Type: StreamflowInput Type: RainfallNumber of Datasets
Rainfall Duration (days)Rainfall Gauging Station SelectedAverage Rainfall (mm)Standard Deviation (mm)
G1(1)S1; S2threeR2; R4130141236
G1(2)S1; S2fourR3; R13146152169
G2S1; S2twoR3; R66079372
G3S1; S2threeR310298243
G4(1)S1; S2twoR4; R7; R988134312
G4(2)S1; S2twoR7; R9; R1388135364
G5S1twoR4; R126494251
G6S1twoR3; R126293330
G7(1)S1fourR17145120139
G7(2) fourR6; R7171252154
G8(1)S1fourR17139120142
G8(2)S1fourR17138119150
Table 5. The estimation performance of the BPNN and the ANFIS at the groundwater monitoring wells.
Table 5. The estimation performance of the BPNN and the ANFIS at the groundwater monitoring wells.
WellANN Model: BPNN 1 ANFIS 2RMSE (m)Correlation Coefficient
TrainingValidationTestingTrainingValidationTesting
G1(1)BPNN (4-3-1)0.0910.1080.110.8440.6960.724
ANFIS (4-2-1)0.0850.1120.1330.8670.6660.578
G1(2)BPNN (4-4-1)0.1150.1370.1490.4810.3040.221
ANFIS (4-2-1)0.1140.1380.1490.4910.2750.22
G2BPNN (4-4-1)0.0550.0910.1120.3420.3940.274
ANFIS (4-2-1)0.0560.0920.1150.3080.3120.121
G3BPNN (3-6-1)0.0580.0830.1280.8440.8460.682
ANFIS (3-2-1)0.0630.1070.1290.8060.7530.667
G4(1)BPNN (5-9-1)0.0630.0890.1090.8950.8060.893
ANFIS (5-3-1)0.0580.0940.1440.9120.7790.775
G4(2)BPNN (5-10-1)0.0480.0670.0690.9280.880.865
ANFIS (5-2-1)0.0450.0640.0620.9380.8870.89
G5BPNN (3-3-1)0.0770.0840.1640.5220.4070.355
ANFIS (3-2-1)0.0760.0860.1680.5410.3650.289
G6BPNN (3-3-1)0.060.0790.1260.4380.3680.167
ANFIS (3-3-1)0.0490.0650.1240.6770.7050.236
G7(1)BPNN (2-3-1)0.120.1210.1380.7520.7150.734
ANFIS (2-3-1)0.0760.1260.1470.9070.7030.662
G7(2)BPNN (2-2-1)0.1320.1860.1860.6190.6250.582
ANFIS (2-2-1)0.1240.1920.2130.6750.5990.504
G8(1)BPNN (2-3-1)0.0930.0960.1460.8470.8430.841
ANFIS (2-3-1)0.0550.1210.2070.9490.6920.841
G8(2)BPNN (2-3-1)0.1030.1120.1180.8740.8590.832
ANFIS (2-2-1)0.070.1220.1260.9440.8410.909
1 (number of input-number of nodes in the hidden layer-number of output); 2 (number of input-number of rules-number of output).
Table 6. The statistics of the observed data at the groundwater monitoring wells.
Table 6. The statistics of the observed data at the groundwater monitoring wells.
WellStatistics (m)
MeanSDMaximumMinimum
G1(1)0.150.191.190.01
G1(2)0.080.110.760.01
G20.160.657.950.01
G30.150.221.670.01
G4(1)0.270.452.780.01
G4(2)0.110.161.260.01
G50.070.10.860.01
G60.060.070.80
G7(1)0.060.070.370.01
G7(2)0.080.130.760.01
G8(1)0.080.090.450.01
G8(2)0.080.090.450.01

Share and Cite

MDPI and ACS Style

Bai, T.; Tsai, W.-P.; Chiang, Y.-M.; Chang, F.-J.; Chang, W.-Y.; Chang, L.-C.; Chang, K.-C. Modeling and Investigating the Mechanisms of Groundwater Level Variation in the Jhuoshui River Basin of Central Taiwan. Water 2019, 11, 1554. https://doi.org/10.3390/w11081554

AMA Style

Bai T, Tsai W-P, Chiang Y-M, Chang F-J, Chang W-Y, Chang L-C, Chang K-C. Modeling and Investigating the Mechanisms of Groundwater Level Variation in the Jhuoshui River Basin of Central Taiwan. Water. 2019; 11(8):1554. https://doi.org/10.3390/w11081554

Chicago/Turabian Style

Bai, Tao, Wen-Ping Tsai, Yen-Ming Chiang, Fi-John Chang, Wan-Yu Chang, Li-Chiu Chang, and Kuang-Chih Chang. 2019. "Modeling and Investigating the Mechanisms of Groundwater Level Variation in the Jhuoshui River Basin of Central Taiwan" Water 11, no. 8: 1554. https://doi.org/10.3390/w11081554

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop