Discriminant analysis application in spatiotemporal evaluation of water quality in South Florida

Surface water quality has become a serious concern for urban planners and managers. The surface water quality in a region is a function of natural factors (precipitation, weather, basin physiography, soil erosion, etc.) and anthropogenic factors (urbanization, industrial and agricultural activities, etc.) (Carpenter et al. 1998; Jarvie et al. 1998). As a result of their impacts, nutrients, toxic substances, and petroleum products enter the rivers, estuaries, lakes and other waterbodies, reducing the quality of water. Anthropogenic factors, such as residential and industrial wastewater, are a constant polluting source in urban areas, whereas natural factors like rainfall, surface runoff, and groundwater level are seasonal phenomenon that are mainly affected by climate (Singh et al. 2004). Seasonal variations in precipitation, surface runoff, ground water flow, and water interception and abstraction have significant effects on river discharge, and subsequently, on the concentration of pollutants in river water (Vega 1998). Therefore, to better investigate and evaluate the water quality of watersheds, a study of temporal variations alongside the spatial variations of water quality seems to be inevitable.

To obtain reliable information about the inherent properties of water quality and to understand the spatial and temporal variations in the hydro-chemical and biological properties of water, continuous and regular monitoring programs are required (Singh et al. 2005). However, the generated databases are large and complex so that their analyses require robust analytical tools. Multivariate statistical techniques are widely used for the evaluation of both temporal and spatial variations and the interpretation of large and complex water quality data sets (Wunderlin et al. 2001; Simeonov et al. 2003; Singh et al. 2004, 2005; Kim et al. 2005; Kowalkowski et al. 2006; Shrestha & Kazama 2007; Schaefer & Einax 2010; Juahir et al. 2011; Mustapha & Aris 2012).

Discriminant analysis (DA) was conducted on the data set of water quality of three selected rivers in the study area to construct the discriminant functions (DFs) on two different standard and stepwise modes to identify the most significant variables that result in water quality spatial and temporal variation. In addition, this step was performed to optimize the monitoring program by decreasing the number of parameters monitored. Wunderlin et al. (2001) obtained a complex data matrix, which was treated using the pattern recognition techniques of cluster analysis (CA), factor analysis/principal component analysis, and DA. They found that the DA technique showed the best results for data reduction and pattern recognition during both temporal and spatial analysis. Singh et al. (2004, 2005) used multivariate statistical techniques to evaluate spatial and temporal variations in the water quality of the Gomti River (India). Zhou et al. (2007) used CA and DA to assess temporal and spatial variations in the water quality of the watercourses in the North Western New Territories of Hong Kong, over a period of five years (2000–2004), using 23 parameters at 23 different sites (31,740 observations). Their results showed that DA allowed a reduction in the dimensionality of the large data set and indicated a few significant parameters that were responsible for most of the variations in water quality. Shrestha et al. (2008) used different multivariate statistical techniques for the evaluation of temporal/spatial variations and the interpretation of a large complex water quality data set of the Mekong River. They also concluded that DA showed the best results for data reduction and pattern recognition during both spatial and temporal analysis. Li et al. (2009) also used DA to evaluate temporal and spatial variations and to interpret a large and complex water quality data set collected from the Songhua River basin in China. The results of DA showed a reduction in the dimensionality of the large data sets, by delineating a few indicator parameters of the water quality. Zhang et al. (2011) applied different multivariate statistical techniques for the interpretation of a complex data matrix obtained between 2000 and 2007 from the watercourses in the Southwest New Territories and Kowloon, Hong Kong. DA provided better results both temporally and spatially. Chen et al. (2015) used the fuzzy comprehensive assessment, CA and DA to assess the water pollution status and analyze its spatiotemporal variation. Furthermore, long-term hydro-chemical data of shallow water bodies were evaluated using DA (Medina-Gomez & Herrera-Silveira 2003; Parinet et al. 2004; Solidoro et al. 2004).

Given the above considerations, a large data matrix obtained over a 15-year (2000–2014) monitoring period at 16 different sites for 12 water quality parameters, and in two wet and dry seasons (approximately 35,000 observations), was subjected to DA to identify the most significant water quality variables responsible for spatial and temporal variations in river water quality.

The hydrography network of the study area, generated using the 1:24,000 national hydrography data set obtained from the South Florida water management district's (SFWMD) geographic information systems (GIS) data catalog, was used to delineate the flow line of the three selected rivers. The most recent (2008–2009) land cover/land use (LCLU) map, provided by the SFWMD, was used in this study. These data were then clipped to fit our study area. The area of each type of land use within each watershed was calculated using an ESRI ArcGIS 10.0 platform. The monitoring stations, downloaded from the same source, were overlaid with the rivers’ map in ArcGIS to design a network of sampling stations that include sufficient historical data to construct a robust statistical database of studied parameters, considering a suitable spatial distribution on the river. Then, the DBHYDRO (environmental database of SFWMD) was used to obtain continuous time series data for 12 selected water quality parameters from 2000 to 2014. This database was then divided into two dry and wet seasons (the wet season from May 15th to October 15th and the dry season from October 16th to May 14th). The selected water quality parameters for investigation in this study include chl-a, dissolved oxygen (DO), total Kjeldahl nitrogen (TKN), total phosphorus (TP), total phosphate, ammonia-N, water temperature (WT), total suspended solids (TSS), turbidity, magnesium, chloride, and sodium.

Temporal variation in water quality was evaluated through DA. Both standard and stepwise modes were applied on the raw data after dividing the whole data set into two seasonal groups (wet and dry seasons). Season was the dependent variable, while all observed water quality parameters were independent variables. The values of Wilks' lambda and the chi-square statistic for each DF were obtained from each standard mode and stepwise mode. As shown in Table 1, the values varied from 0.222 to 0.244 and from 277.3 to 267.5 for Wilks' lambda and the chi-square, respectively. Smaller values of Wilks’ lambda indicate a greater discriminatory ability of the function. The associated chi-square statistic tests the hypothesis that the means of the functions listed are equal across groups. The small significance values (p-level <0.01) indicate that the DF does better than chance at separating the groups. Thus, the temporal DA was credible and effective. The first function in standard DA explained almost all (R = 88.2%) of the total variance in the dependent variables. The stepwise DA had similar results, which indicated that 87% of the total group differences in the data set were explained by its first DF. Therefore, the first DF alone was sufficient to explain the difference of water quality among the two wet and dry seasons.

The stepwise DA identified five variables (chl-a, DO, TKN, TP and WT) as the most important discriminating variables. Table 2 shows that the first function in the stepwise DA was perfectly correlated with temperature (coefficient = 1.000), and then mostly correlated with DO (coefficient = 0.613). Classification functions (CFs) and the classification matrices (CMs) obtained from standard and stepwise modes of DA are shown in Tables 3 and 4. In the standard mode, all 12 variables were included to construct CFs which correctly classified 95.3% of the original grouped cases. In the stepwise mode, the DA correctly assigned 93.2% of the cases using only five discriminating variables. Therefore, the temporal DA results of the stepwise mode suggested that chl-a, DO, TKN, pH, and WT were the most significant parameters for discriminating differences between the wet season and dry season, and could be used to explain most of the expected temporal variations in water quality.

The identified patterns of the most important discriminating variables show that, in general, the average concentrations of the water quality parameters in the Miami Canal are worse than the other two rivers. The Kissimmee River demonstrated even lower average values. Nonetheless, the outliers in Figure 4 are related to the average concentration of the represented variables in two highly polluted sites of S154C and CES03 in the Kissimmee River and the Caloosahatchee River, respectively. However, TP showed that a different pattern and average concentrations of this variable were much higher in the Caloosahatchee River and the Kissimmee River, respectively, in comparison to the Miami Canal. Besides the two mentioned highly polluted sites, TP was found higher at the Caloosahatchee River and the Kissimmee River than the Miami Canal sites. The percentage of agricultural and urbanized areas in the Miami Canal watershed was measured from the LULC map in GIS and was seen to be even more than the other two rivers. Therefore, this could be related to the effectiveness of eco-restoration projects implemented in its watershed and adjacent linked watersheds (the water conservation area-3, WCA-3) in order to decrease the amounts of nutrients.

The first pattern showed clear spatial differences for chl-a and DO as a measure of life's vitality and the activity level of the plants and animals living in rivers. The higher average values of these two variables were found in the Kissimmee River and the Caloosahatchee River, which indicates the dynamism and strength of aquatic lives in this river. Besides, the Miami Canal had lower average concentrations of chl-a and DO, which indicated that organic pollution may play a significant role in the Miami canal, especially in urbanized areas which are under the influence of more domestic and industrial wastewater.

The second pattern showed higher average concentrations of TP in the Caloosahatchee River and the Kissimmee River, which consists of two highly polluted sites of CES03 and S154C, respectively. The asterisks or stars are extreme outliers observed in these two sites that represent cases that have values more than three times the height of the boxes. Previous analysis (Haji-Gholizadeh et al. 2016) indicated that these two highly polluted sites are extremely affected by urbanized areas, and also high-density environmental resource permits with more industrial effluent and domestic sewage.

The third identified pattern of the most important discriminating variables in the dry season showed that the average concentrations of TKN, magnesium, and chloride in the Miami Canal are worse than the other two rivers. The Kissimmee River demonstrated lower average values. Nonetheless, the outliers in Figure 6 are related to the average concentration of the represented variables in two highly polluted sites of S154C and CES03 in the Kissimmee River and the Caloosahatchee River, respectively.

In this study, DA was applied to evaluate spatial and temporal variations in surface water quality of three major rivers of South Florida using 15 years (2000–2014) data sets of 12 water quality variables covering 16 sampling stations, with approximately 35,000 observations. DA, as an important data reduction method, was used to assess the water pollution status and analysis of its spatiotemporal variation. In temporal DA, 12 months of raw data were divided into two seasonal groups (wet and dry season) as the dependent variable, while all observed water quality parameters were independent variables. In the spatial DA, each river was separately considered as one spatial region to evaluate the patterns associated with spatial variations in each river's water quality. The three major rivers of South Florida were the grouping (dependent) variable, while the observed parameters in each season constituted the independent variables.

It was found by the stepwise DA that five variables (chl-a, DO, TKN, TP and WT) are the most important discriminating water quality parameters responsible for temporal variations. In spatial DA, the stepwise mode identified only five variables (TKN, TP, ammonia-N, magnesium, and sodium) and seven variables (hl-a, DO, TKN, TP, ammonia-N, magnesium, and chloride) as the most significant discriminating variables responsible for spatial variations in the wet and dry season, respectively. There were also different patterns associated with spatial variations that were identified, depending on the variables and considered season. However, it was found that the average concentrations of the most significant discriminating variables in both the wet and dry seasons in the Miami Canal was worse than the other two rivers, and the Kissimmee River demonstrated lower average values. Nonetheless, two highly polluted sites of S154C and CES03 in the Kissimmee River and the Caloosahatchee River require more attention and consideration.

This study showed the feasibility and reliability of DA in river water quality research. It is desirable that both state and local agencies pay more attention and consideration in order to improve and protect the vulnerable river quality. Additional studies will be required to assess precisely the unidentified sources of pollution and variation of further water quality parameters that were not analyzed in this study. Furthermore, the conclusion would be beneficial to water environment protection and water resources management in the future. The results of the spatial and temporal variations could be used to select the polluted areas and set the priority areas for the river water quality management in the study area.

The research was funded from Florida International University, Miami, USA. The observational data were obtained from South Florida Water Management District (SFWMD). We also thank the reviewers for providing insightful comments, as well as the conference officials.