4.1. Methodological considerations
As hypothesized, we observed high correlations among the 315 calculated indices. This was particularly evident between metrics that describe duration and frequency and those related to magnitude (Fig.2). This result is not surprising considering the high redundancy and multicollinearity among hydrological variables . As a consequence, statistical techniques for selecting a reduced set of non-redundant indices are necessary to represent the hydrological gradients . In our study, we first selected 171 indices with the highest loading coefficient associated with the first dimension of the PCA (Table C1), and many of them were also correlated (Figs.AC1-AC7). Similar processes to reduce the number of indices have been reported in previous studies for hydrological classifications (e.g., . Considering that PCA analyses provided many indexes for each attribute, we refined the selection using cluster analysis and repetition of metrics in the first 5 dimensions of PCA. However, these results need to be interpreted with caution, since we prioritized the selection of annual indices, such as number of days or percentage of months without flow per year, mean daily of annual flows, coefficient of variation of Julian date of the annual start of zero flow and annual rise rate.
As also hypothesized, our results indicate that the PCA and the correlations between the indices are analogous for all thresholds 0, 1, 2, and 5 l/s to define zero-flow conditions. Similar results were obtained by , who found no significant change in zero-flow indices for thresholds of 1, 2 and 5 l/s. However, they found that upper thresholds (i.e. 10 l/s) were not suitable for characterizing periods of low flow. They applied different thresholds considering that river gauge stations may have different measuring tools and some thresholds may not be ecologically relevant. In fact, described the causes of zero-flow stream gauge measures and their consequences (e.g., ecological status, statistical classification, or hydrological modelling) for understanding hydrology, biochemical, and ecological processes in NPRS.
We found that number of days per year without flow and annual percentage of months without flow (identified in the first dimension of the PCA) contain the greatest degree of variation for all the thresholds used to define zero flow conditions (Table 1). These metrics, related to duration and frequency attributes, express stream drying events and are common to describe the flow temporality . They indicate the unavailability of water during the dry phase and act as an environmental bottleneck for species of these ecosystems . Furthermore, indices of magnitude were found mainly in the second dimension of the PCA. Mean daily annual flows is repeated in all the zero-flow thresholds and has been reported as a major variable for hydrological characterization in perennial rivers . Surprisingly, we found that timing and rate of change attributes were identified in subsequent PCA dimensions; CV of Julian date of the annual start of zero flow in the third dimension and annual rise rate only in the fifth, both described by . Taken together, these results suggest that duration, frequency and magnitude indices (mostly selected in the first and second components) are more important than timing and rate of change indices (selected from the third component). These findings indicate that these attributes were best at predicting the temporal variability on the flows; duration and frequency in the dry phase and magnitude in the wet, reinforcing the importance of these two phases for the structure and composition of the aquatic fauna . However, in this study we have retained at least one metric for each of the five attributes for a better characterization of the variability of the NPRS, both in dry and wet periods, and to facilitate comparison with other available classifications.