Global natural catchment classification based on hydrological similarity

Abstract

Catchments are important in hydrology, and a catchment classification framework helps to understand the catchment hydrological behaviors, explain the differences between catchments, and predict in ungauged catchments. However, no research has yet established a global catchment classification framework. We selected a group of hydrological signatures to represent catchment hydrological behaviors, and used the fuzzy clustering method to classify natural catchments. To explain the classification rules and the catchment attributes dominating the classification, we used decision tree and random forests, respectively. The results show that: the global natural catchments are divided into six classes by the fuzzy clustering method, most of the classes are extreme in at least one hydrological behavior, and the selected hydrological signatures can distinguish the catchment groups; The decision tree gives explicit classification rules, with an accuracy rate of over 93%, which reasonably explains the fuzzy clustering results and facilitates the judgment of catchment classes; The precipitation characteristics, aridity index and the lowest altitude of catchments are considered to be the dominant catchment attributes for catchment classification, among which the average daily precipitation is the most important; Compared with physiography, land cover, soil and geological factors, the relative importance of climate factors in catchment classification exceeds 50%; The global catchment classification pattern output by random forests is a comprehensive reflection of hydrological signatures and can better reflect the hierarchical differences in hydrological behavior among catchments in contrast to climate classification. The validity of the proposed global classification pattern is supported by its consistency with regional studies conducted in Europe, the United States, and Australia. Furthermore, about 64.1% classification accuracy of catchment class and 62.0% simultaneous hit rates of eight hydrological signatures can be achieved by the random forests model, demonstrating the ability of proposed catchment classification in estimating the hydrological behavior of ungauged catchments. As the first step towards global catchment classification, this study developed a natural catchment classification method based on hydrological similarity using data-driven approaches, obtained a global distribution map, and laid the foundation for establishing a generally accepted global catchment classification framework.