Evaluation of river water quality variations using multivariate statistical techniquesSava River (Croatia): a case study

For the evaluation of seasonal and spatial variations and the interpretation of a large and complex water quality dataset obtained during a 7-year monitoring program of the Sava River in Croatia, different multivariate statistical techniques were applied in this study. Basic statistical properties and correlations of 18 water quality parameters (variables) measured at 18 sampling sites (a total of 56,952 values) were examined. Correlations between air temperature and some water quality parameters were found in agreement with the previous studies of relationship between climatic and hydrological parameters. Principal component analysis (PCA) was used to explore the most important factors determining the spatiotemporal dynamics of the Sava River. PCA has determined a reduced number of seven principal components that explain over 75 % of the data set variance. The results revealed that parameters related to temperature and organic pollutants (CODMn and TSS) were the most important parameters contributing to water quality variation. PCA analysis of seasonal subsets confirmed this result and showed that the importance of parameters is changing from season to season. PCA of the four seasonal data subsets yielded six PCs with eigenvalues greater than one explaining 73.6 % (spring), 71.4 % (summer), 70.3 % (autumn), and 71.3 % (winter) of the total variance. To check the influence of the outliers in the data set whose distribution strongly deviates from the normal one, in addition to standard principal component analysis algorithm, two robust estimates of covariance matrix were calculated and subjected to PCA. PCA in both cases yielded seven principal components explaining 75 % of the total variance, and the results do not differ significantly from the results obtained by the standard PCA algorithm. With the implementation of robust PCA algorithm, it is demonstrated that the usage of standard algorithm is justified for data sets with small numbers of missing data, nondetects, and outliers (less than 4 %). The clustering procedure highlighted four different groups in which the sampling sites have similar characteristics and pollution levels. The first and the second group correspond to relatively low and moderately polluted sites while stations which are located in the middle of the river belong to the third and fourth group and correspond to highly and moderately polluted sites.