The influence of methodological procedures on hydrological classification performance
EstadísticasView Usage Statistics
Full recordShow full item record
AuthorPeñas Silva, Francisco Jesús; Barquín Ortiz, José; Snelder, T. H.; Booker, D. J.; Álvarez Díaz, César
ABSTRACT: Hydrological classification has emerged as a suitable procedure to disentangle the inherent hydrological complexity of river networks. This practice has contributed to determining key biophysical relations in fluvial ecosystems and the effects of flow modification. Thus, a plethora of classification approaches, which agreed in general concepts and methods but differed largely in specific procedures, have emerged in the last decades. However, few studies have compared the implication of applying contrasting approaches and specifications over the same hydrological data. In this work, using cluster analysis and modelling approaches, we classify the entire river network covering the northern third of the Iberian Peninsula. Specifically, we developed classifications of increasing level of detail, ranging from 2 to 20 class levels, either based on raw and normalized daily flow series and using two contrasting approaches to determine class membership: classify-then-predict (ClasF) and predict-thenclassify (PredF). Classifications were compared in terms of
their statistical strength, the hydrological interpretation, the ability to reduce the bias associated with underrepresented parts of the hydrological space and their spatial correspondnece. The results highlighted that both the data processing and the classification strategy largely influenced the classification outcomes and properties, although differences among procedures were not always statistically significant. The normalization of flow data removed the influence of flow magnitude and generated more complex classifications in which a wider range of hydrologic characteristics were considered. The application of the PredF strategy produced, in most of the cases, classifications with higher discrimination ability and presented greater ability to deal with the presence of distinctive gauges in the data set than using the ClasF strategy.