Science of the Total Environment 598 (2017) 828–838

Contents lists available at ScienceDirect

Science of the Total Environment

j ourna l homepage: www.e lsev ie r .com/ locate /sc i totenv

Conservation of groundwater from over-exploitation—Scientific analysesfor groundwater resources management

Fi-John Chang a,⁎, Chien-Wei Huang a, Su-Ting Cheng a, Li-Chiu Chang b

a Department of Bioenvironmental Systems Engineering, National Taiwan University, Taiwan, ROCb Department of Water Resources and Environmental Engineering, Tamkang University, New Taipei City, 25137, Taiwan, ROC

H I G H L I G H T S G R A P H I C A L A B S T R A C T

• Provide scientific analyses of thegroundwater systems for futuregroundwater management plans

• Relate spatial-temporal patterns ofgroundwater with surface water basedon large datasets

• PCA classified the groundwater systemsinto eastern, western and transitionzones

• SOM visibly explored regional ground-water variations and inter-relationsamong variables

• Build inter-relation between surfacewater with groundwater mechanismsat a watershed-scale context

⁎ Corresponding author at: No. 1, Sec. 4, Roosevelt RoaE-mail address: changfj@ntu.edu.tw (F.-J. Chang).

http://dx.doi.org/10.1016/j.scitotenv.2017.04.1420048-9697/© 2017 Elsevier B.V. All rights reserved.

a b s t r a c t

a r t i c l e i n f o

Article history:Received 13 February 2017Received in revised form 14 April 2017Accepted 19 April 2017Available online 27 April 2017

Editor: D. Barcelo

Groundwater over-exploitation has produced many critical problems in the southern Taiwan. The accumulatedstresses and demandsmake groundwatermanagement a complex issue that needs innovative scientific analysesfor deriving better water management strategies. In this study, we aimed to provide scientific analyses of thegroundwater systems in the Pingtung Plain through soft-computing techniques to explore its spatial-temporaland hydro-geological characteristics for the elaboration of future groundwater management plans and in deci-sion-making process. We conducted a study to assess the essential features of the groundwater systems basedon the long-term large datasets of regional groundwater levels by using the principal component analysis(PCA), and the self-organizingmap (SOM)with regression analysis. The PCA results demonstrated that two lead-ing components couldwell present the spatial characteristics of the groundwater systems and classify the regioninto eastern,western and transition zones. The SOM results could visibly explore the behavior of regional ground-water variations in various aquifers and themulti-relations among climate and hydrogeological variables. Resultsrevealed that the potential of groundwater rechargemade by precipitation or river flowwas higher in the easternzone than in thewestern zone. Analysis results further showed an increase of the groundwater levels in thewest-ern zone after year 2006, while therewere no obvious increases of the groundwater levels in the eastern or tran-sition zones. Based on the investigated characteristics, we suggest that a sound groundwater management planshould consider zonal difference of the groundwater systems to achieve groundwater conservation.

© 2017 Elsevier B.V. All rights reserved.

Keywords:Self-organizing map (SOM)Principal component analysis (PCA)Regional groundwaterGroundwater–surface water interactionsPingtung Plain

d, Da-An District, Taipei, 10617, Taiwan, ROC.

829F.-J. Chang et al. / Science of the Total Environment 598 (2017) 828–838

1. Introduction

While very uneven rainfall and steep-sloped rivers constrain thestorage and use of surface water resources in Taiwan, groundwater be-comes a substitute for surface water in many areas (Liang et al., 2016).Especially in the southern Taiwan, exploitation of groundwater hasbeen extracted not only for drinking purposes but for irrigation and cul-tivation in agricultural and aqua-cultural sectors (Hu et al., 2006; Jang etal., 2016). A year-round domestic water demand and aqua- and agricul-tural operations have resulted in intensive groundwater exploitation(Jang et al., 2016). The over exploitation of groundwater has led tomany problems, such as the contamination of groundwater that causedthe arsenic intrusion in drinking water, which greatly increase the risksof diabetes and cancers (Das et al., 2016; Lamm et al., 2004; Liu et al.,2003; Liang et al., 2016). Another serious problem caused by groundwa-ter over-exploitation is land subsidence and groundwater depletion(Burbey, 2008; Hung et al., 2012; Wada et al., 2012). Along the coastalregion, a serious vertical land movement was observed in the range of−0.7 to −2.8 cm per year in the Pingtung Plain (Hsieh et al., 2011)leading to the other serious groundwater salinization problems (Chenet al., 2007).

With the purposes to remediate these serious problems caused byintensive groundwater exploitation, an amendment of the Regulationson Groundwater Conservation (RGC) was passed in year 2002 for regu-lating and limiting the use of groundwater in Taiwan. Under this amend-ment, the Water Resources Agency and the associated governmentalagencies in Taiwan are endowed with the power to determine and regu-late where and how much groundwater can be exploited. Since 2002,both previously and newly constructed wells in the regulatory zonesneed to apply for licenses and renewals. At present, the RGC has beenenacted formore than ten years. In addition, theWater Resources Agencyhas launched groundwater level monitoring programs in the regulatoryzones. Under better protection of the groundwater resources, an increaseof the groundwater storage in the regulatory zones is expected. As such,this study was conducted as a part of a regional groundwater assessmentandmodeling study requested by theWater Resources Agency of Taiwanin the context of sustainable groundwater management.

Groundwater aquifers are complex and heterogeneous systems,which pose great challenges in the modeling of groundwater systemsas well as systematically quantification for sustainable water resourcesmanagement (Alley and Leake, 2004). The level of groundwater is af-fected by a combination of several natural and anthropogenic factorslike precipitation, geology, slope and water withdrawal (Redwan andMoneim, 2016). Prediction of the regional groundwater level often in-volves using precipitation and/or stream flow as primary sources tothe recharging mechanisms of groundwater aquifers (Tsai et al.,2016). However, the interactions vary among different hydrogeologicalsystems (Krause et al., 2007), and the distribution of precipitation dif-fers temporally and spatially. These spatial-temporal features requirelocal- and regional-scale studies (Cheng and Wiley, 2016) to catch thechanges in the groundwater system and realize the dynamic balanceof the abstraction, replenish, and storage process in space and time(Bekesi et al., 2009). Without enough understanding of the mechanismsand forces that control the level of groundwater in the Pingtung Plain,plus insufficient data to account for the actual magnitude of groundwaterwithdrawn or the real precipitation distribution, sustainable manage-ment of groundwater resources in this area has been very challenging.To pursue sustainable use of groundwater, scientific investigations ofthe groundwater recharge/discharge mechanisms and classifications ofthe complex spatial-temporal groundwater features in the PingtungPlain are needed. As such, the comprehensive characterization of the re-gion from a groundwater perspective can provide useful elaboration forfuture groundwater management plans and in decision-making process.

In this paper, we developed a groundwater modeling study to ex-plore the changes of the groundwater level in the Pingtung Plain andevaluated the effectiveness of the RGC amendment that has been

executed for more than 10 years. The objective of this study was tomake a preliminary assessment of the groundwater responses to meetthe need for decision making on groundwater resources managementwith a focus on groundwater conservation from over exploitation. Wefirst applied a principal component analysis (PCA) to the groundwaterlevel data collected during dry seasons to investigate the characteristicsof the groundwater systems. Following that, we performed the self-or-ganizing map (SOM) based on year-round groundwater level data toclassify the patterns of groundwater levels. Thenwe employed a regres-sion analysis between groundwater level and precipitation as well asriverflow.We coupled all the results to discuss spatial and temporal dif-ferences associated with the hydro-geophysical characteristics in theplain. Last, we provided a comprehensive understanding of the ground-water systems for sustainable groundwater resourcesmanagement andthe evaluation results for the effectiveness of the RGC amendment.

2. Methods

2.1. Study area

The Pingtung Plain is located in southwestern Taiwan with a totaldrainage area of 1210 km2. It consists of alluvial fans of threemajor riv-ers, including the Kaoping River, the Tungkang River and the LinbianRiver. The rivers are originated from the Central Mountain Range inthe northeast, progress through the plain, and eventually drain intothe Taiwan Strait (Fig. 1).

The geological setting of the main stratigraphic compositions in thePingtung Plain was investigated by several drilling studies and strati-graphic analyses since 1995 (Liang et al., 2016). The investigation re-sults revealed that the plain were comprised by multiple overlappingsequences, and the Central Geological Survey in Taiwan defined it intofour aquifers A, B, C, and D by depths of 0–70, 40–130, 90–180, and160–250 m, respectively (Fig. 2). Daily groundwater level data were re-corded by 126 groundwatermonitoring sensors installed in a total of 55observation wells spreading across the four aquifers (Fig. 1). Daily pre-cipitation records and daily flow dataweremonitored at 15 and 3 gaug-ing stations in the plain, respectively (Fig. 1).

2.2. Data collection and preprocessing

Provided by theWater Resources Agency in Taiwan, 14-year (1999–2012) time-series data of daily groundwater levels, along with precipi-tation andflow recordswere aggregated for the Pingtung Plain. To com-pare the variation of groundwater level in the region, we transformedgroundwater elevation data into relative groundwater levels prior tothe analysis process. The relative groundwater level (RGL) was calculat-ed by subtracting the observed minimum groundwater elevation fromthe observed groundwater elevation (Eq. (1)).

RGLi ¼ OGLi−MGLi ð1Þ

where RGLi is the relative groundwater level of the ith groundwater sen-sor, OGLi is the observed groundwater elevation of the ith groundwatersensor, and MGLi is the observed minimum groundwater elevation ofthe ith groundwater sensor. In this way, the variation of groundwaterlevel can be captured and standardized for making comparisonsamong different groundwater sensors.

2.3. Computational methods

2.3.1. Principal component analysis (PCA)The PCA is a widely used analysis that helps reduce the dimension-

ality of the data sets but retain most of the variations for easier data ex-ploration (Noori et al., 2010; Azid et al., 2014; Zahra et al., 2014). Weused the PCA to extract principal components in the order of their sig-nificance representing the features of the groundwater levels among

Fig. 1. Locations of the Pingtung Plain and the monitoring stations of groundwater, precipitation, and flow.

830 F.-J. Chang et al. / Science of the Total Environment 598 (2017) 828–838

the historical measurements of the 126 groundwater level sensors inthe Pingtung Plain.We ran the PCA by usingMATLAB based on historicaldaily RGL data. The input data matrix of the PCA was aggregated andcompiled from daily RGL of 126 sensors recorded during dry seasons(fromNovember to the next April) in years 1999–2012.We intentional-ly used data from dry seasons in order to simplify the complications de-riving from the influence of precipitation. As such, a total of 2538datasets were used in the PCA. Datasets were first transformed into Zscores before running the PCA. Next, we decided which principal com-ponents would be included or ignored. With the chosen principal com-ponents, there were loadings assigned to each groundwater sensor forrepresenting its interrelations associated with the component. We cal-culated themetric scores bymultiplying the eigenvectors by the originaldatasets for the chosen principal components.

After that, we used the loading assigned to each sensor of the chosenprincipal component as a spatial attribute and derived loading contoursby using the “linear variogram” kriging gridding method (Delhomme,1978) within Arc-GIS. The map was constructed for the purpose ofdisplaying the regional characteristics variance of the groundwater sys-tems in the Pingtung Plain. We determined the criteria to form separa-tion lines based on loadings derived from the selected PCA components.These lines roughly divided the groundwater systems into zonation.

2.3.2. Self-organizing map (SOM)The SOM (Kohonen, 1982) is a powerful clustering tool, which im-

plements an ordered dimensionality-reducing mapping of input vari-ables to generate 2-dimensional topologically ordered maps. Weclassified the complex high-dimensional data sets based on the

Fig. 2. Geological setting of the main stratigraphic compositions in the Pingtung Plain of Taiwan (modified from Lai et al., 2002).

831F.-J. Chang et al. / Science of the Total Environment 598 (2017) 828–838

topological maps of SOM to make a comprehensive analysis of ground-water variations in the four aquifers and to present their spatio-tempo-ral features. The regional monthly average groundwater levels duringthe whole year period from 1999 to 2012 (i.e., a total of 168 datasetscompiled from 126 monitoring sensors) were used to configure theSOM. The trained SOM was used to identify output responses

resembling the same topological order as the features of the datasetsand to capture the variance of groundwater systems under differenttemporal and spatial conditions (Kohonen, 1982; Tsai et al., 2017).

We also decomposed the data structure of the SOM feature map tobetter understand how the monthly datasets of the 126 groundwatersensors configure the SOM feature map. Assuming that signals received

Table 1PCA results of the first ten principal components with the individual and cumulative per-centages of variance explained by each extracted component.

Component Variance explained (%) Cumulative variance explained (%)

1 63.0 63.02 18.1 81.23 4.3 85.54 2.8 88.35 1.7 90.06 1.2 91.27 1.1 92.48 1.0 93.49 0.9 94.310 0.6 94.9

832 F.-J. Chang et al. / Science of the Total Environment 598 (2017) 828–838

from similar groundwater level patterns were automatically groupedinto the same neurons of the SOM, the decomposition of the datastructure could help discover the groundwater recharge mechanisms(Chang et al., 2014).

Lastly, we employed the regression analysis to assess the relation-ship between RGL and river flow (similarly for precipitation) based onthe classified PCA zonation. Monthly averages across years 1999–2012were used to compute the correlation between RGL and river flow (sim-ilarly for precipitation). At the end, we coupled all the results to explorethe temporal and spatial characteristics of groundwater systems andprovide comprehensively science-based information for groundwaterresources management.

Fig. 3. Locations of 55 groundwater wells installed with a total of 126 sensors for four differentmethod in Arc-GIS. The two lines separating the characteristics of groundwater recharge wereSensors were then clustered into zonation: eastern zone (dot); western zone (triangle); and tra

3. Results

The PCA extracted principal components in the order of their signif-icance representing the total variance of 2538 datasets (each dataset in-cluded 126 sensors' daily RGL). Results showed that the 1st componentaccounted for 63.0% of the total variance. A combination of the 2nd com-ponent and the 1st one explained 81.2% of the total variance (Table 1).The explanatory power dropped to 4.3% for the 3rd component and con-tinued decreasing for the other components thereafter (Table 1). There-fore, we chose the 1st and 2nd components for the classification of thespatial characteristics of the RGL in the plain.

Since the loading assigned to each sensor of the chosen principalcomponent was considered as a spatial attribute, the loading contoursrevealed the spatial interrelations of the groundwater systems in thePingtung Plain that the loadings of the 1st component showed aneast-to-west distribution, in contrast to a west-to-east distributionfrom those of the 2nd component (Fig. 3). These two lines roughly di-vided the groundwater systems into three zones. The first separationline (yellow line in Fig. 3) was obtained from the 1st component witha loading of 0.1 (i.e., most of the loadings in the eastern sites wereover 0.1, and then decreased to −0.02 in the western boundary),while the second separation line (red line in Fig. 3) was determinedby the 2nd component with a loading of 0 (the loading from +0.18 inthe west decreasing to −0.13 in the east). We grouped those allottedto the eastern part as the “eastern zone”, and vice versa (i.e., the “west-ern zone”). When a well experiencing mixed loadings due to multipleaquifers, it might be grouped into eastern or western zone for different

aquifers in the Pingtung Plain of Taiwan. The loading contours were derived by the krigingdetermined by using the loadings for the 1st and the 2nd PCA components, respectively.nsition zone (blue dashed circles for thosemismatched sensors based on the two criteria).

833F.-J. Chang et al. / Science of the Total Environment 598 (2017) 828–838

aquifers. If by either criterion the inconsistency occurred, the wellwould be grouped into the “transition zone”. Based on these criteria,we had 37 groundwater sensors (in 18 wells) being grouped into theeastern zone, 38 groundwater sensors (in 17 wells) being groupedinto the western zone, and 51 groundwater sensors (in 20 wells) withmixed loadings or grouped inconsistently by the two criteria beinggrouped into the transition zone (blue dashed circles in Fig. 3).

Plotting the metric scores of the PCA over time showed two distincttrends associated with the 1st and the 2nd components (Fig. 4). Thescores represented the multiplication of component loadings by the Zscores of each sensor. Therefore, results showed that the scores of the1st component appeared an annual periodic fluctuation over time, asopposed to the scores of the 2nd component that transitioned fromneg-ative to positive scores in year 2006. According to the metric scores ofthe 2nd component, we found an increasing pattern of groundwaterlevel after year 2006, indicating a better status of groundwater conser-vation. As the loading contours of the 2nd component diminishedfrom the west to the east, the increasing pattern of the metric scoresof the 2nd component also indicated that the increase of the groundwa-ter level was more obvious in the western side than in the eastern sideof the plain.

Asmentioned above, the 126 groundwatermonitoring sensors weredistributed in four heterogeneous aquifers. This 3-dimensional problemwasdifficult towell present in a 2-dimensional SOMmap. Consequentlywe made each aquifer a 2-dimensional SOM map. Results of the SOMshowed 16 (4*4) classifications for each aquifer of A, B, C and D (Fig.5). With the implementation of the Kriging method, the colors of thecontour lines representing the values of RGL in the SOM clearly showedgreat variations among different neurons, different zones in a neuron, aswell as different aquifers. It was easy to tell that (1) the relative ground-water level (RGL) in the eastern zonewasmuch higher than those in thewestern zone, and (2) the RGL in the eastern neurons was in generalhigher than those of the western neurons for all the four aquifers.

To excavate the temporal-spatial features inside the SOM, the timeand average RGL in each neuron were analyzed. We discovered a topo-logical feature shift in years 2005/2006 through the decomposition ofthe SOM data structure, and the similarities of datasets could be orga-nized by year or by month. When organizing by year, neurons 9, 10,11, 13, 14 and 15 consisted of data collected from 1999 to 2005 (orangeareas in Fig. 6)while neurons 1, 2, 3, 4, 8 and 12 consisted of data collect-ed mostly from 2006 to 2012 (green areas in Fig. 6). Some neuronsconsisted of data collected during only 2002 and 2006 (e.g., neurons 5,6 and 7), but some contained data collected in multiple years (e.g., neu-ron 16).

In terms of month, the results showed that the groundwater systemcould be distinguished between the months with lower groundwaterlevels (from January to July) and those with higher groundwater levels(from June to December), in which some variations occurred in themonths with persisting groundwater levels (Fig. 6). This could befound at neurons 1, 2, 3, 9 and 13, involving months between January/February/March and June/July (yellow areas in Fig. 6) and at neurons

Fig. 4.Trends showing themetric scores of the 1st and the 2nd PCA components over time.

4, 8, 11, 12, 14, 15 and 16, involving months between June/July and Oc-tober/November/December (blue areas in Fig. 6).

The post-calculation of the weights assigned to each neuron in theSOM revealed that average groundwater levels were different at theeastern, western and transition zones (Table 2). According to the yearlyclassifications of the SOM, the results showed that the average RGL in thewestern zone increasedmore after 2006 (i.e., neurons of 1, 2, 3, 4, 8 and12) than years before 2005 (i.e., neurons 9, 10, 11, 13, 14 and 15). Forexample, in Aquifer A the weights of the neurons for the western zoneranged from 1.0 to 2.6 before 2005, but ranged from 1.8 to 2.8 after2006. Increasing trends could be found in Aquifers B, C and D for thewestern zone. Nonetheless, trends for the eastern and the transitionzones were not readily distinguishable before or after year 2005/2006,as compared with those of the western zone.

Regarding themonthly classifications of the SOMduring the seasonswith lower groundwater levels (i.e., neurons 1, 2, 3, 9 and 13), the low-est groundwater levels occurred, in general, in thewestern zone, follow-ed by the transition zone and the eastern zone. Exceptions did exist forcertain neurons, such as neurons 1 and 2, where Aquifers B, C and D ofthe western and/or transition zones exhibited higher groundwaterlevels than those of the eastern zone. This might be considered as an in-dication of a better groundwater conservation condition during seasonswith lower groundwater levels for thewestern zone. In contrast, duringseasons with higher groundwater levels, the eastern zone almost al-ways had much higher groundwater levels than the transition zone,followed by the western zone.

Based onmonthly averages across all the investigated years, ground-water levels represented greater changes at the eastern zone than at thetransition or western zones (Fig. 7), indicating that the groundwatersystem at the eastern zone made the rapidest response to precipitationor river flow, followed by those of the transition zone and the westernzone. Fig. 7 also showed that for all three zones, the lowest groundwaterlevel occurred in May while the highest in September.

According to the regression analyses based on themonthly data setsin various sub-regions (zones), different trends between zonal ground-water levels and river flow or precipitation were found. The fitted re-gression correlation coefficient (r) showed distinct correlations amongdifferent zonation that groundwater levels were more related to riverflow at the eastern zone (RGL = 0.0002·Flow + 5.020, r = 0.56, p =0.060) than at the transition (RGL = 0.0001·Flow + 3.674, r = 0.53,p = 0.079) or the western (RGL = 0.00003·Flow + 2.464, r = 0.48,p = 0.115) zones. Regarding the relation of groundwater level to pre-cipitation, its correlations were lower than those of groundwater levelto river flow, but analysis results also showed a decreasing trendstarting from the eastern zone (RGL = 0.004·Precipitation + 6.185,r = 0.32, p = 0.315) to the transition (RGL = 0.002·Precipita-tion + 4.207, r = 0.28, p = 0.376) and the west (RGL = 0.0005·Precip-itation + 2.660, r = 0.23, p = 0.474) zones.

These results suggested that the classified zonation represented dif-ferent features of the groundwater systems, with temporal changesamong aquifers. However, the correlation values with river flow andprecipitation were neither high nor significant, indicating that thegroundwater systems may be affected by the great variability of spa-tial-temporal hydro-geological features and/or by other factors, so thatthe relationships among precipitation or river flow cannot be well rep-resented by simple linear regressions. Consequently, a sustainablegroundwatermanagement should consider the different spatial-tempo-ral zonal characteristics of the groundwater system framedbyour scien-tific analysis.

4. Discussion

The characteristics of the groundwater systems are complex tempo-rally and spatially.With the analyses performed,wewere able to extractimportant characteristics about the groundwater systems in thePingtung Plain as a scientific basis for groundwater resources

Fig. 5. The SOM classified the relative groundwater levels into 16 groups (aka. neurons) for aquifers A, B, C and D in the Pingtung Plain.

834 F.-J. Chang et al. / Science of the Total Environment 598 (2017) 828–838

management. Findings included: (1) based on the loadings, the PCA re-vealed an east-to-west distribution for the 1st component, in contrast toawest-to-east distribution for the 2nd component; (2) the PCA scores ofthe 1st component appeared a pattern similar to the pattern of ground-water levels at the wells in the eastern side of the plain, while the PCA

scores of the 2nd component displayed an increasing trend after 2006,reflecting an increase of the groundwater level at the wells in the west-ern side; (3) based on the PCA results, we classified the groundwatersystems in the Pingtung Plain into zonation – the eastern, western andtransition zones; (4) SOM results also showed that the increase of the

Fig. 6.Decomposition of the data structure configured by neurons 1–16 of the SOM. (Neurons 9, 10, 11, 13, 14 and 15 consisted of data collected from 1999 to 2005 (orange areas), whileneurons 1, 2, 3, 4, 8 and 12 consisted of data collected from 2006 to 2012 (green areas)).

835F.-J. Chang et al. / Science of the Total Environment 598 (2017) 828–838

groundwater levels at the wells in the western side of the plain waslarger during periods after 2006 than during periods before 2005; (5)the trained SOM map revealed that the rate of groundwater rechargemade by precipitation or river flow was faster in the eastern zone

Table 2Weights (i.e., RGL) of each neuron in different aquifers based on the PCA zonation.

Aquifer A

Neurons 1 2 3 4 5 6 7 8East 3.2 4.2 5.8 7.3 3.9 6.3 8.2 8.7Transition 2.0 2.4 3.1 3.7 2.4 3.4 4.2 4.3West 1.8 1.9 2.1 2.5 1.5 1.9 2.4 2.7

Aquifer BNeurons 1 2 3 4 5 6 7 8East 3.2 4.4 6.3 7.9 4.2 7.0 9.0 9.5Transition 2.2 2.8 3.7 4.5 2.6 3.9 4.9 5.2West 3.6 3.7 4.0 4.4 2.6 3.2 4.1 4.7

Aquifer CNeurons 1 2 3 4 5 6 7 8East 2.4 3.4 5.0 6.5 3.3 5.8 7.6 8.0Transition 3.1 3.7 4.7 5.5 3.7 5.2 6.1 6.3West 3.0 3.2 3.6 3.9 2.5 3.1 3.8 4.2

Aquifer DNeurons 1 2 3 4 5 6 7 8East 2.7 3.7 5.3 6.7 3.5 5.9 7.5 8.0Transition 3.4 3.9 4.7 5.4 3.7 5.0 5.8 6.0West 3.1 3.3 3.5 3.8 2.6 3.1 3.7 4.0

than in the western zone; and (6) the regression results showed thatthe correlation gradients between groundwater level and precipitation(or river flow) started from the eastern zone to the western and thetransition zones, indicating their correspondence with the

9 10 11 12 13 14 15 164.0 6.1 9.2 10.7 5.5 8.7 11.0 12.02.5 3.3 4.6 5.3 3.1 4.4 5.4 5.91.0 1.3 2.0 2.8 1.2 1.7 2.6 3.2

9 10 11 12 13 14 15 164.4 6.8 10.2 11.6 6.1 9.6 11.9 12.82.6 3.7 5.3 6.2 3.4 5.0 6.3 6.91.6 1.9 2.9 4.3 1.7 2.2 3.6 4.6

9 10 11 12 13 14 15 163.4 5.6 8.8 10.1 4.9 8.1 10.4 11.33.9 5.0 6.7 7.3 4.7 6.3 7.4 7.81.8 2.2 3.1 4.1 2.0 2.6 3.7 4.5

9 10 11 12 13 14 15 163.6 5.6 8.5 9.7 4.9 7.9 9.9 10.73.7 4.7 6.1 6.7 4.4 5.8 6.7 7.22.0 2.2 2.9 3.8 2.0 2.4 3.4 4.0

Fig. 7. Trends showing the monthly averages of RGL in the eastern, western and transition zones and those of precipitation and flow across all the investigated years (1999–2012).

836 F.-J. Chang et al. / Science of the Total Environment 598 (2017) 828–838

characteristics of the groundwater systems extracted by the PCA or theSOM. These findings are important information for making sustainableuse of groundwater resources in the study area.

In this study, we found that the PCAwas a useful tool for groundwa-ter resources management. The PCA would place the principal compo-nents in order of the variability explained. In our analysis, the first twocomponents accounted for a high percentage (more than 80%) of thetotal variance in the datasets; therefore, the two leading componentshelped extract the spatially distributed characteristics of groundwatersystemswithout losing toomuch information given by the other princi-pal components. With a high representation of the data featuresmade by the first two principal components, the PCA greatly de-creased the redundancy (multicollinearity) of the data dimensions(Grossman et al., 1991; Monk et al., 2007). Nonetheless, the repre-sentation of the selected components was not necessarily one-di-mensional. With the assistance from the PCA, we were able toidentify the spatial similarities of the groundwater systems in thePingtung Plain and to group them into zonation—the eastern, west-ern and transition zones.

It was noted that the information given by the PCA for estimating thecorrelations among each groundwater sensors was based on the linearprinciples (Brosse et al., 2001). Consequently, the spatial correlation ofthe 126 simultaneous groundwater level recordswas taken into account“linearly”. However, correlations between groundwater level with pre-cipitation or river flow were not high suggesting that the relationshipsamong groundwater and the surrounding environments might be non-linear. As such, a nonlinear SOM technique could act as a complement to

give a comprehensive investigation of the groundwater systems andprovide great potentials for solving real life water resources manage-ment problems by system solutions (Kalteh et al., 2008; Park et al.,2015; Chang et al., 2016a & Chang et al., 2016b; Hong et al., 2016).

Based on the SOM serving as another classifier, the time series fea-tures of groundwater levels were extracted nonlinearly. SOM results re-vealed different patterns of the groundwater systems between theeastern and thewestern sides of the plain (Fig. 5). Apart from the differ-ences in linear or nonlinearmethodologies,we used different time-scaledatasets for the PCA and the SOM (i.e., the PCA used daily datasets in dryperiods, as opposed to the SOM using monthly datasets). The two dis-tinct techniques presented very similar features showing spatial andtemporal variations. As such, our results suggested that both linearand nonlinear approaches could achieve feature separation if distinctiondid exist within the source signals.

The zonation suggested by the PCA was found to match with theconfiguration of the river network and the distribution pattern of pre-cipitation. The structure of tributaries and mainstream is known to beshaped by the interaction of geomorphological and hydrological pro-cesses (Cheng et al., 2016). In the Pingtung Plain, the topography ofthe eastern side of the plain is relatively higher than that of the westernpart. Tributaries are mostly located at the eastern side of the plain,whereas themainstream is located at thewestern side (Fig. 1). The pre-cipitation records used for our study were only available at 15 gaugingstations outside the plain, which may not reflect the real distributionof the precipitation in the study area. But the spatial features of ground-water systems extracted by the PCA or the SOM still relatively imitated

837F.-J. Chang et al. / Science of the Total Environment 598 (2017) 828–838

the distribution features of the river network settings and the spatialpattern of precipitation.

Based on the PCA zonation, we also found dissimilar trends for dif-ferent zones or aquifers, reflecting the geology of the Pingtung Plain.The geological map showed that the main stratigraphic composition ofthe eastern side was simpler than that of the western side in thePingtung Plain and appeared a compositional gradient ranging fromthe eastern to the western sides (Fig. 2). By linking the geological infor-mation with data mining analysis of groundwater records, we derivedthat the recharge patternswere significantly associatedwith the perme-ability of soil types. As the eastern side mostly consisted of gravels(Fig. 2), the topographic and geological arrangements resulted inmore rapid and more direct processing mechanisms for the easternzone. In contrast, the western side resided in layers of gravels, sands,mud and metamorphic rocks (Fig. 2), and our analyses captured a dis-tinct feature of the groundwater systems resulted from a lower topogra-phy with more complex soil layers/types compositions leading to theslower accumulations of water resources from precipitation or riverflow for the western zone (Fig. 7).

In addition, we discovered the response time of the groundwaterlevel change associated with the geological conditions. As known,areas of stratigraphic units with higher rates of permeability, such asgravels, would involve shorter response time between the changes ingroundwater level and a precipitation event. Conversely, areas withmore complicated structural features or composed with layers of lowpermeability would experience longer response time. Our analysis re-sults confirmed these facts and showed a trend of longer responsetime starting from the western to the eastern zones. The transitionzone lying from the east to the west at the extension of the geologicalterrain happened to have different aquifers. Subsequently, the “zona-tion” not only reflected the characteristics of the groundwater systemsbut might be considered as a consequence of the interactions betweenthe climatic and hydrogeological conditions. As such, the geomorphicand hydrogeological settings along with precipitation were crucial fac-tors dominating the dynamics of the hydrogeological processing withinthe groundwater systems.

Based on our analysis, groundwater levels have increased in thewestern zone of the plain after year 2006. Since one of the main factorscontributing to the groundwater variations is the withdrawal volumefrom abstraction wells, especially during the drought events, the incre-ments of groundwater level compared to previous years in the westernzone could serve as an indication of the effectiveness of the execution ofthe RGC amendment. However, results showed that groundwater levelsin the eastern and the transition zones did not appear increasing trends.Regarding the lower groundwater level and longer response time in thecoastal (western) zone, compared to those of the eastern zone, moreconservative groundwater exploitation should be considered in themanagement plan. Moreover, there is a need to establish differentthresholds of groundwater withdrawal based on zonal difference forgroundwater sustainability. Furthermore, studies relating the mecha-nisms of precipitation and surface water to groundwater or relatingaquifer connectivity to rock types/compositions are needed.

5. Conclusion

The over exploitation of groundwater in the Pingtung Plain in thesouthern Taiwan has led to many serious problems, such as groundwa-ter contamination, land subsidence and groundwater depletion. To re-mediate these problems, a sound groundwater management is crucialbut challenging. In this paper, we developed a groundwater modelingstudy to explore the changes of the groundwater level based on large-scale long-term groundwater records in the region. We demonstratedthat the soft-computing techniques, including the PCA and the SOMwith regression analysis, could be useful tools to extract the majorhydro-geological mechanisms and identify the major features of thecomplex regional groundwater systems. The PCAwith only two leading

components were able to classify the spatial similarities of the ground-water systems in the Pingtung Plain and to group them intozonation—the eastern, western and transition zones. The SOM resultsvisibly indicated an increase of the groundwater levels at the wells inthewest of the plain after 2006, and revealed that the rate of groundwa-ter recharge made by precipitation or river flow was faster in the east-ern zone than in the western zone. Overall, analysis results found ageneral trend following precipitation, river configuration and geologicalconditions closely, which was similar to the classified zonation. Thesefindings well reflected the hydrogeological features of the groundwatersystems and could be considered as a consequence of the interactionsbetween the climatic and hydrogeological conditions. This is crucial in-formation for making sustainable use of groundwater resources in thestudy area. Our analyses also supported the effectiveness of the execu-tion of the amendment tagged to the Regulations on Groundwater Con-servation. Nonetheless, we also argued that a sound groundwaterresources management should consider the spatial and temporal differ-ences of groundwater characteristics. The determination of groundwa-ter "zonation" could be a useful classification representing differentinteractive processing among surface hydrology, geomorphology andgroundwater systems. Therefore, we suggest that zonal groundwatermanagement is required to achieve the conservation of groundwaterfrom over exploitation.

Acknowledgements

This study was supported by the Water Resources Agency, Taiwan,ROC. (Grant number: MOEAWRA1040062). It is very much appreciatedthat streamflow and groundwater level data were provided by theWa-ter Resources Agency, rainfall datawere provided by the CentralWeath-er Bureau, Taiwan, ROC, and the geological map was provided by theCentral Geological Survey, MOEA, Taiwan, ROC. The authors would liketo thank the editors and anonymous reviewers for their review andvaluable comments related to this manuscript.

References

Alley,W.M., Leake, S.A., 2004. The journey from safe yield to sustainability. GroundWater42 (1), 12–16.

Azid, A., Juahir, H., Toriman, M.E., Kamarudin, M.K.A., Saudi, A.S.M., Hasnam, C.N.C., Aziz,N.A.A., Azaman, F., Latif, M.T., Zainuddin, S.F.M., Osman, M.R., Yamin, M., 2014. Predic-tion of the level of air pollution using principal component analysis and artificial neu-ral network techniques: a case study in Malaysia. Water Air Soil Pollut. 225, 2063.

Bekesi, G., McGuire, M., Moiler, D., 2009. Groundwater allocation using a groundwaterlevel response management method—Gnangara groundwater system, Western Aus-tralia. Water Resour. Manag. 23, 1665–1683.

Brosse, S., Giraudel, J.L., Lek, S., 2001. Utilisation of non-supervised neural networks andprincipal component analysis to study fish assemblages. Ecol. Model. 146 (1–3),159–166.

Burbey, T.J., 2008. The influence of geologic structures on deformation due to groundwater withdrawal. Ground Water 46 (2), 202–211.

Chang, L.C., Shen, H.Y., Chang, F.J., 2014. Regional flood inundation nowcast using hybridSOM and dynamic neural networks. J. Hydrol. 519, 476–489.

Chang, F.J., Chang, L.C., Huang, C.W., Kao, I.F., 2016a. Prediction of monthly regionalgroundwater levels through hybrid soft-computing techniques. J. Hydrol. 541,965–976.

Chang, F.J., Chen, P.A., Chang, L.C., Tsai, Y.H., 2016b. Estimating spatio-temporal dynamicsof stream total phosphate concentration by soft computing techniques. Sci. Total En-viron. 562, 228–236.

Chen, C.T., Hu, J.C., Lu, C.Y., Lee, J.C., Chan, Y.C., 2007. Thirty-year land elevation changefrom subsidence to uplift following the termination of groundwater pumping andits geological implications in the metropolitan Taipei Basin, Northern Taiwan. Eng.Geol. 95 (1–2), 30–47.

Cheng, S.T., Wiley, M.J., 2016. A reduced parameter stream temperature model (RPSTM)for basin-wide simulations. Environ. Modell. Softw. 82, 295–307.

Cheng, S.T., Herricks, E.E., Tsai, W.P., Chang, F.J., 2016. Assessing the natural and anthropo-genic influences on basin-wide fish species richness. Sci. Total Environ. 572, 825–836.

Das, S., Liu, C.C., Jean, J.S., Liu, T.L., 2016. Dissimilatory arsenate reduction and in situ mi-crobial activities and diversity in arsenic-rich groundwater of Chianan Plain, South-western Taiwan. Microb. Ecol. 71 (2), 365–374.

Delhomme, J.P., 1978. Kriging in the Hydrosciences. Adv. Water Resour. 1 (5), 251–266.Grossman, G.D., Nickerson, D.M., Freeman, M.C., 1991. Principal component analyses of

assemblage structure data: utility of tests based on eigenvalues. Ecology 72 (1),341–347.

838 F.-J. Chang et al. / Science of the Total Environment 598 (2017) 828–838

Hong, N., Hama, T., Suenaga, Y., Aqili, S.W., Huang, X., Wei, Q., Kawagoshi, Y., 2016. Appli-cation of a modified conceptual rainfall–runoff model to simulation of groundwaterlevel in an undefined watershed. Sci. Total Environ. 541, 383–390.

Hsieh, C.S., Shih, T.Y., Hu, J.C., Tung, H., Huang, M.H., Angelier, J., 2011. Using differentialSAR interferometry to map land subsidence: a case study in the Pingtung Plain ofSW Taiwan. Nat. Hazards 58 (3), 1311–1332.

Hu, J.C., Chu, H.T., Hou, C.S., Lai, T.H., Chen, R.F., Nien, P.F., 2006. The contribution to tec-tonic subsidence by groundwater abstraction in the Pingtung area, southwestern Tai-wan as determined by GPS measurements. Quat. Int. 147, 62–69.

Hung, W.C., Hwang, C., Liou, J.C., Lin, Y.S., Yang, H.L., 2012. Modeling aquifer-system com-paction and predicting land subsidence in central Taiwan. Eng. Geol. 147, 78–90.

Jang, C.S., Chen, C.F., Liang, C.P., Chen, J.S., 2016. Combining groundwater quality analysisand a numerical flow simulation for spatially establishing utilization strategies forgroundwater and surface water in the Pingtung Plain. J. Hydrol. 533, 541–556.

Kalteh, A.M., Hiorth, P., Bemdtsson, R., 2008. Review of the self-organizing map (SOM) ap-proach in water resources: analysis, modelling and application. Environ. Model.Softw. 23 (7), 835–845.

Kohonen, T., 1982. Self-organized formation of topologically correct feature maps. Biol.Cybern. 43 (1), 59–69.

Krause, S., Bronstert, A., Zehe, E., 2007. Groundwater–surface water interactions in aNorth German lowland floodplain–implications for the river discharge dynamicsand riparian water balance. J. Hydrol. 347 (3), 404–417.

Lai, T.-C., Eri, L.-Y., Chiang, C.-J., 2002. The Characteristics of Groundwater Zones in Tai-wan. Central Geological Survey, MOEA, Taiwan.

Lamm, S.H., Engel, A., Kruse, M.B., Feinleib, M., Byrd, D.M., Lai, S.H., Wilson, R., 2004. Arse-nic in drinking water and bladder cancer mortality in the United States: an analysisbased on 133 US counties and 30 years of observation. J. Occup. Environ. Med. 46(3), 298–306.

Liang, C.P., Wang, S.W., Kao, Y.H., Chen, J.S., 2016. Health risk assessment of groundwaterarsenic pollution in southern Taiwan. Environ. Geochem. Health 38, 1271–1281.

Liu, C.W., Lin, K.H., Kuo, Y.M., 2003. Application of factor analysis in the assessment ofgroundwater quality in a blackfoot disease area in Taiwan. Sci. Total Environ. 313(1–3), 77–89.

Monk, W.A., Wood, P.J., Hannah, D.M., Wilson, D.A., 2007. Selection of river flow indicesfor the assessment of hydroecological change. River Res. Appl. 23 (1), 113–122.

Noori, R., Karbassi, A.R., Sabahi, M.S., 2010. Evaluation of PCA and Gamma test techniqueson ANN operation for weekly solid waste prediction. J. Environ. Manag. 91 (3),767–771.

Park, Y., Cho, K.H., Park, J., Cha, S.M., Kim, J.H., 2015. Development of early-warning pro-tocol for predicting chlorophyll-a concentration using machine learning models infreshwater and estuarine reservoirs, Korea. Sci. Total Environ. 502, 31–41.

Redwan, M., Moneim, A.A.A., 2016. Factors controlling groundwater hydrogeochemistryin the area west of Tahta, Sohag, Upper Egypt. J. Afr. Earth Sci. 118, 328–338.

Tsai, W.P., Chiang, Y.M., Huang, J.L., Chang, F.J., 2016. Exploring the mechanism of surfaceand ground water through data-driven techniques with sensitivity analysis for waterresources management. Water Resour. Manag. 30 (13), 4789–4806.

Tsai, W.P., Huang, S.P., Cheng, S.T., Shao, K.T., Chang, F.J., 2017. A data-mining frameworkfor exploring the multi-relation between fish species and water quality through self-organizing map. Sci. Total Environ. 579 (1), 474–483.

Wada, Y., Beek, L.P., Sperna Weiland, F.C., Chao, B.F., Wu, Y.H., Bierkens, M.F., 2012. Pastand future contribution of global groundwater depletion to sea-level rise. Geophys.Res. Lett. 39 (9), L09402.

Zahra, A., Hashmi,M.Z., Malik, R.N., Ahmed, Z., 2014. Enrichment and geo-accumulation ofheavy metals and risk assessment of sediments of the Kurang Nallah—feeding tribu-tary of the Rawal Lake Reservoir, Pakistan. Sci. Total Environ. 470, 925–933.