Water quality assessment using multivariate statistical techniques in Río Tercero Reservoir, Argentina

Water quality is a matter of serious concern today, being considered the main factor controlling health and the state of disease in both humans and animals. Anthropogenic influences (urban, industrial and agricultural activities increasing consumption of water resources) as well as natural processes (change in precipitation inputs, erosion, weathering in crustal material) degrade surface waters and impair their use for drinking, industrial, agricultural, recreation or other purposes (Simeonov et al. 2003; Li et al. 2009).

Because lakes, reservoirs and rivers constitute the main inland water resources for different purposes, it is imperative to prevent and control water pollution and to have reliable information on water quality (Varol et al. 2012; Bonansea & Fernandez 2013). Therefore, regular monitoring programs are required for representative and reliable estimates of surface water quality (Kazi et al. 2009; Fernandez et al. 2012). However, due to spatial and temporal variations in water quality, such programs need to include a large number of physico-chemical parameters taken at different times and from many sites involving huge financial inputs and resulting in a large and complex data matrix which is often difficult to interpret towards drawing meaningful conclusions. Thus, there is a need to optimize the monitoring networks, reducing the number of water quality parameters, sampling sites and monitoring periods to representative ones without losing useful information (Singh et al. 2004).

Multivariate statistical techniques are presented as an effective tool for the interpretation of this complex data matrix, allowing a better understanding of water quality and ecological status of the studied reservoir (Wunderlin et al. 2001; Shrestha & Kazama 2007). The application of different multivariate statistical approaches, such as principal component analysis (PCA), factor analysis (FA), discriminate analysis and cluster analysis (CA) have been widely used to achieve great efficiency of data compression from the original data and to interpret natural associations between samples or variables, highlighting the information which is not available at first glance (Chen et al. 2007). Furthermore, the multivariate treatment of environmental data is widely successfully used to identify possible pollution sources that influence water systems and offers a valuable technique for reliable management of water resources as well as rapid solutions on pollution problems (De Bartolomeo et al. 2004; Kalamaras et al. 2010; Wang et al. 2012; Ledesma et al. 2013).

The objective of the present study was to analyse the spatial and temporal patterns of water quality of Rio Tercero Reservoir (Argentina). The large data matrix obtained during an 8-year (2003–2010) monitoring program was subjected to the PCA and CA multivariate techniques to obtain information about the similarities and dissimilarities between sampling sites and seasons, and to identify water quality variables responsible for variations in surface water quality. Additionally, the influence of the pollution sources on the water quality parameters was determined. Thus, we illustrated the usefulness of the multivariate statistical analysis to improve the understanding of the surface water system.

Río Tercero Reservoir (32° 11′ S, 64° 23′ W) is located in the province of Córdoba, Argentina (Figure 1). This reservoir, which is the largest artificial reservoir in the province, has an area of 46 km², a volume of 10⁷m³ and maximum and mean depths of 46.5 and 12.2 m, respectively (Mariazzi et al. 1992). The watershed area is 3,300 km² and rainfall is strongly seasonal, with dry winters and heavy rains during spring and summer (MacDonagh et al. 2009).

Río Tercero Reservoir is divided into two basins linked by a passage. The left basin has three branches where tributary rivers flow. The right basin supplies water to the Tercero River.

The reservoir has multiple purposes, such as water supply for three cities located on its shores (Embalse, Villa del Dique and Villa Rumipal with a combined population of 20,000 inhabitants approximately), recreational activities, flood control, irrigation and power generation. In 1986 a nuclear power plant (Central Nuclear Embalse, (CNE)) was installed. Water for cooling the nuclear reactor is taken from the middle section of the reservoir. In the cooling system, water temperature (WT) increases 7.0 °C and is returned to the reservoir by a 5-km-long open-sky channel (Mariazzi et al. 1992; Lamaro et al. 2013).

Water sampling was conducted once during each climatic season (defined in the southern hemisphere as winter: July–September; spring: October–December; summer: January–March; and fall: April–June) from winter 2003 to spring 2010. Six sampling sites were surveyed (Figure 1). The locations of these sites, which were selected in order to reasonably represent the water quality of the reservoir, were recorded using a GPS device. Water samples were collected at 20 cm depth. Preservation and transportation of the water samples to the laboratory were performed according to Standard Methods (APHA-AWWA-WEF 2000). In total, the samples were analysed for 21 parameters. In situ, WT, pH, dissolved oxygen (DO) and electrical conductivity (EC) were measured using portable electronic instruments. Secchi disk transparency (SDT) was measured using a standard 20 cm diameter Secchi disk with alternating black and white quadrants. In the laboratory, chlorophyll-a concentration (Chl-a), total dissolved solids (TDS), total phosphorus (TP), total nitrogen (TN), total Kjeldahl nitrogen (TKN), nitrate nitrogen (NO₃-N), total hardness (T-Hard), total alkalinity (T-Alk), bicarbonates (HCO₃⁻), sulphate (SO₄), chloride (Cl), sodium (Na), potassium (K), calcium (Ca), magnesium (Mg) and fluoride (F) were determined following standard analytical methods (APHA-AWWA-WEF 2000; WHO 2006).

All data analyses were performed using the InfoStat software package (Di Rienzo et al. 2009). One-way analysis of variance (ANOVA) was employed as a first approach to analyse the significant spatial and temporal differences of water quality variables (p < 0.05; least-significance difference, LSD test). To find relationships among these variables, a correlation matrix using the Pearson's correlation coefficient (r) was carried out with statistical significance set at p < 0.05.

Multivariate statistical analysis can help to simplify and organize large data sets to provide meaningful insight (Chen et al. 2007). In this study, two multivariate methods were applied to the water quality data set: the PCA and the CA. This data set was previously standardized using z-scale transformation in order to avoid misclassifications arising from the different orders of magnitude of both numerical value and variance of the parameters analysed (Wunderlin et al. 2001; Singh et al. 2004). According to Güler et al. (2002) and Cloutier et al. (2008), standardization is commonly used in multivariate statistical analysis.

The PCA is a data transformation technique used to find out associations between variables, thus reducing the dimensionality of the data set (Helena et al. 2000; Chen & Hong 2012). This technique extracts the eigenvalues and eigenvectors from the covariance matrix of original variables. The principal components (PCs) are the new uncorrelated (orthogonal) variables which are obtained by multiplying the original correlated variables with the eigenvector, which is a list of coefficients (loadings or weightings). Thus, the PCs are weighted linear combinations of the original variables (Varol et al. 2012). There are as many PCs as original variables. However, the first principal component loading explains the most variance and each subsequent component explains progressively less. As a result, a small number of factors usually account for approximately the same amount of information as the much larger set of the original observations do (Chen et al. 2007). Thus, PC provides information on the most meaningful parameters, which describe the whole data set affording data reduction with minimal loss of information. The PC loadings can be examined to provide further insight into the processes that are responsible for the similarities in the trace element concentrations in the water samples (Helena et al. 2000; Wunderlin et al. 2001; Farnham et al. 2003; Singh et al. 2004). PCA was applied to extract significant PCs summarizing the statistical correlation among components in the water samples.

CA is an unsupervised pattern recognition technique that uncovers intrinsic structure or underlying behaviour of a data set without making a priori assumption about the data, in order to classify the objects of the system into categories or clusters based on their nearness or similarity (Vega et al. 1998). This technique classifies objects, so that each object is similar to the others in the cluster with respect to a predetermined selection criterion. The resulting clusters should then exhibit high internal (within-cluster) homogeneity and high external (between clusters) heterogeneity (Shrestha & Kazama 2007). Hierarchical clustering is the most common approach in which clusters are formed sequentially, by starting with the most similar pair of objects and forming higher clusters step-by-step. The result can be displayed as a dendrogram (tree diagram), which provides a visual summary of the clustering process, presenting a picture of the groups and their proximity, with a dramatic reduction in dimensionality of the original data (Wunderlin et al. 2001). The Euclidean distance usually gives the similarity between two samples and a distance can be represented by the difference between analytical values from the samples (Otto 1998). In this study, hierarchical CA was performed using the Ward's method and Euclidean distances as a measure of similarity. The Ward's method uses an analysis of variance approach to evaluate the distances between clusters, attempting to minimize the sum of squares of any two clusters that can be formed at each step. CA was applied to the water quality data set to group the similar sampling sites (spatial variability) and seasonal (temporal) similarity among the samples, resulting in spatial and temporal dendrograms. The linkage distance is reported as Dlink/Dmax, which represents the quotient of the linkage distances for a particular case divided by the maximal distance, multiplied by 100, as a way to standardize the linkage distance represented on the y-axis (Varol et al. 2012). Finally, ANOVA was performed (p < 0.05) to analyse the significant differences between the clusters obtained by the spatial CA. Whereas, as the temporal CA generated only two clusters, differences between clusters were tested for statistical differences using a Student T-test (p < 0.05).

The basic statistics of the 21 water quality parameters collected from six samples sites during the study period are presented in Table 1. The results, which were based on 180 total water samples, were compared with the maximum limits recommended by the World Health Organization (WHO 2006), revealing that the concentrations of all analysed parameters were below the prescribed maximum limits by WHO guidelines for drinking water. This confirmed that all water samples collected in Río Tercero Reservoir corresponded to low polluted regions.

Most water quality parameters except WT and SDT did not show significant spatial variations (p > 0.05). Higher values of WT and lower values of SDT were recorded in site 3, which was related to the cooling channel of CNE and rivers influxes. The lowest values of SDT registered in the left basin can be explained by river inputs which provide greatest loads of suspended materials and dissolved solids that decrease the penetration of light. In contrast, sites 5 and 6, located in the right basin of the reservoir where deeper waters and sedimentation processes were more active, exhibited the highest SDT values. Similar results were found by Giardino et al. (2001) and Guan et al. (2011) studying different water bodies. On the other hand, WT, pH, EC, TP, T-Hard, T-Alk, Cl, K and Mg showed significant temporal differences (p < 0.05). Among these, WT and TP showed higher values in summer and spring, while pH, EC, T-Hard, T-Alk, Cl, K and Mg showed higher values in fall and winter.

The Pearson correlation matrix of the 21 analysed variables was calculated (Table 2). Only those correlation values higher than 0.5 were considered. High and positive correlations were observed between EC, TDS, T-Hard, T-Alk, HCO₃⁻, SO₄, Cl, Na, K, Ca, Mg and F (r = 0.502 to 0.998). Most of these associations are responsible for water mineralization, being directly related to hydrochemical characteristics of the zone. Chl-a was positively correlated with PT, TN and TKN, indicating that these variables are responsible for phytoplankton growth.

PCA was performed on the normalized data to compare the compositional pattern between the water samples and to identify the factors influencing each one. Table 3 summarized the PCA results including the loadings and the eigenvalues of each PC. There are several criteria to identify the number of PCs to be retained in order to understand the underlying data structure (Chen et al. 2007). In the present study, all the PCs with eigenvalues equal to or greater than 1.0 were used. Four PCs were retained explaining about 96.7% of the total variance in the water quality data set. The first component (PC1), which accounted for over 54.8% of the total variance, had positive loadings for EC, TDS, HCO₃⁻, SO₄, Cl, Na, K, Ca, Mg, F, T-Hard and T-Alk. This PC can be interpreted as the mineral component of the surface water of the reservoir, being related to natural sources of the ionic groups of salts in the basin from inflows, soil weathering and runoff. Some of these results agree with those obtained with the Pearson test (Table 2). The high level of major cations (Na, Ca and Mg) and concentration of major anions (Cl, F, CHO₃⁻ and SO₄) in reservoir water increased EC and TDS, being consistent with other studies (Reghunath et al. 2002; Kazi et al. 2009). Furthermore, results indicated that EC and major ions were higher in winter and lower in summer, due to the accumulation on these ions in the deeper water during summer stagnation (Varol et al. 2012).

The second component (PC2) explained 21.2% of the total variance and was characterized by loadings of Chl-a, TP, NO₃-N, TN and TKN. This PC represents the effects of stream input, agricultural runoff and erosion from upland areas of the basin. The results reveal that nutrients were highest in summer and lowest in winter according to rainfall recorded in the region. Córdoba is a province where farming and cattle raising are common activities. Thus, fertilizers are widely used and contribute to high levels of soil nitrogen and phosphorus that can be incorporated into the reservoir by streams and runoff after rainfall, increasing eutrophication processes which often lead to a reduction in the supply of ecosystem services (Hein 2006; Granéli et al. 2008; Saaremäe et al. 2013). PC2 also showed the effect of nutrients on Chl-a, which is one of the most important parameters since it is used to estimate the primary productivity of water ecosystems (Dall'Olmo et al. 2005). This parameter registered the highest values in spring of 2009 (133.5 μg/L) agreeing with Mancini et al. (2010) who detected a bloom of Anabaena spiroides y Microcystis aeruginosa in Piedras Moras Reservoir, located 11 km downstream of Río Tercero Reservoir. These harmful cyanobacterial blooms, which have increased worldwide in fresh and marine waters, are a primary concern for water management and may cause a variety of water quality problems, including fish kills, aesthetic nuisances (e.g., odours, scums, unsightliness), and unpalatable and possibly unsafe drinking water (Torbick et al. 2008). Such problems can severely limit aquatic habitat quality, recreational activities, fisheries and use of a water body as a potable water resource (Vincent et al. 2004). Subsequent nutrient abatement strategies implemented by environmental agencies or water authorities could be largely successful resulting in a reduction in algal biomass and greater reservoir transparency.

The third component (PC3) showed 12.6% of the total variance and had positive loadings on WT, Chl-a and TP, whereas PC4, explaining the lowest variance (8.0%), had positive loadings on pH, NO₃⁻N, TN and TKN and negative loading on DO and SDT. Both these PCs point to sources of nutrient of the reservoir being related to stream inputs, agricultural runoff, soil erosion, municipal effluents and atmospheric deposition. The contribution of WT to PC3 can be considered to be a result of the constant input of heat from the cooling channel of CNE nuclear power plant. Thus, WT, which showed a very characteristic annual cycle with highest values in summer and lowest in winter, showed the highest values in site 3 and lowest in the right basin of the reservoir. From the temperature difference, changes are expected in the DO. According to Shrestha & Kazama (2007), the inverse relationship between WT and DO is a natural process because warmer water becomes more easily saturated with oxygen. This presumption is strengthened by the fact that site 3 showed the lowest DO concentrations. The monitoring of oxygen concentration in an aquatic system is an important subject, as the biological, chemical and physical processes involved in the increase or decrease of oxygen in a lake are so numerous and complex that there is no model that can be used without a careful analysis of local characteristics (Kazi et al. 2009).

The results from the PCA suggested that most of the variation in water quality was explained by a set of soluble salts (natural). A substantial contribution came from the nutrient group of pollutants (point sources such as municipal effluents and non-point such as erosion and agricultural runoff) and organic pollutants (anthropogenic). The rest of the contributions arose from physical parameters (natural and the effect of nuclear power plant on WT). PCA served as a means to identify those parameters that had the greatest contribution to variation in water quality and suggested possible sets of pollution sources in the watershed.

CA was used to detect the similarity groups between the sampling sites and the climatic seasons. The spatial CA rendered a dendrogram where the six sampling sites were grouped into three statistically significant clusters at (Dlink/Dmax) × 100 < 50 (Figure 2(a)), whereas the temporal dendrogram grouped the four climatic seasons into two clusters (Figure 2(b)). As we expected, in both cases the clustering procedure generated groups in a very convincing way, as the elements in these groups had similar characteristic features and natural background source types.

The spatial dendrogram, which represents geographical locations of sampling sites, generated clusters that correspond to regions of relatively high (cluster 1), moderate (cluster 2) and low pollution (cluster 3). The term ‘relatively polluted’ was used because, as previously mentioned, the parameters measured were always below the prescribed maximum limits by WHO guidelines for drinking water (WHO 2006). Cluster 1 (site 3) receives polluted effluents from different sources such as stream inputs, agricultural runoff and soil erosion by the Quillinzo and La Cruz rivers as shown in Figure 1. This cluster is also directly influenced by the effluent of CNE which, as we have previously mentioned, raises the temperature of water. Sites 1 and 2 make one group as cluster 2. Although this cluster receives effluents by the Grande, Amboy and Santa Rosa rivers, their watersheds are related to areas of natural vegetation and low anthropic activities, corresponding to a relatively moderate polluted region. However, in the last years these watersheds have reported an increase in agricultural activities and a decrease in native forest. Hence it is possible that in the near future this cluster will be more related to cluster 1. To fully understand water chemistry, a study of variability in river discharge patterns needs to be considered for an optimal management plan. Cluster 3 (sites 4 to 6) was generated by low relatively polluted sampling sites located in the central region of the reservoir and in the right basin of the dam, where general water quality conditions were better. The results from one-way ANOVA used to compare these clusters showed that WT, TDS, SO₄, Na, Ca, F and T-Hard of cluster 1 were significantly higher (p < 0.05) than clusters 2 and 3 (Table 4), whereas SDT was significantly higher in cluster 3 than in the rest of the groups.

The temporal dendrogram obtained by CA generated two clusters. Cluster 1 corresponded to the wet season including spring and summer. Fall and winter were grouped in cluster 2 corresponding to the dry season. Although we could expect this group combination, CA technique was able to corroborate this assumption. The results from Student's T-test showed significant difference (p < 0.05) in WT, pH, EC, TDS, Cl, Na, K, Ca, Mg T-Hard and T-Alk. Except for WT, these parameters were higher in the dry season.

The results obtained with the spatial CA suggested that for rapid assessment of water quality, only one site in each cluster may serve as a good spatial assessment of the water quality of the whole network. While the temporal CA implied that sampling during only two seasons (wet and dry) in a year may suffice for assessment of temporal variations in water quality. According to Singh et al. (2004) and Kazi et al. (2009), the CA technique is also useful in offering reliable classification of surface waters in the whole region and will make it possible to design a future sampling strategy in an optimal manner. Thus, sampling frequency, number of sites and cost in the monitoring network will be reduced without missing much information. The same aspects have been successfully applied in different water quality programmes (Wunderlin et al. 2001; Kim et al. 2005; Varol et al. 2012). The use of similar methodology in other reservoirs, either natural lakes or artificial dams, could be of great interest for further investigations.

Water quality monitoring programs generate complex multidimensional data sets that need multivariate statistical treatment for their analysis and interpretation. In this paper, different multivariate statistical techniques were used to evaluate temporal and spatial variations in surface water quality of Río Tercero Reservoir. PCA was used to identify the factors or sources responsible for water quality variations. The main parameters responsible for water quality variation are mainly related to soluble salts (natural), organic pollution and nutrients (point and non-point), and physical parameters (natural). The nuclear power plant was also among the major factors responsible for reservoir water quality. Based on similarities in water quality characteristics, hierarchical CA grouped the six sampling sites into three clusters (relative high, moderate and low pollution) and classified climatic seasons in two clusters (wet and dry season). This technique could also facilitate the design of an optimal future monitoring strategy in the reservoir, which could reduce the number of sampling sites, monitoring frequency and associated costs. The multivariate statistical techniques used in this study served as an excellent tool for the analysis and interpretation of complex data sets, identifying pollution sources and understanding temporal and spatial variations in water quality for effective water quality management.

The authors thank the reviewers for their useful comments and helpful suggestions to improve the paper.