Interpretable Unsupervised Classification of River Catchments with Network Science

Abstract

The classification of river catchments has been an active field of study for decades and the recent surge in hydrological and environmental datasets promotes the formulation of new approaches to this endeavor. We present a novel method for catchment classification based on physical traits similarity using network science, where the relationship among the catchments is represented by the edges of a network. Under this framework we leverage the capability of networks to capture collective behaviors to find clusters of catchments with similar physical traits. The use of networks allows the adoption of similarity metrics other than the common euclidean distance, which is subjected to quality degradation in high dimensions but is still required in many traditional clustering algorithms. Also, a network of traits is built to investigate their similarity patterns and condense this information into a small number of interpretable traits categories. Such categories are used to provide a characterization of each cluster of catchments. The method has been tested on over 9000 river catchments across the contiguous United States, each one accompanied by traits such as climate or vegetation coverage, and anthropogenic features such as land use or proximity to developed areas. The resulting classification shows a remarkable geographical coherence supported by the characteristic traits categories. Additionally, we find that when hydrological indices (like statistics on streamflow or water temperature) are aggregated according to the clusters of catchments, different clusters show different hydrologic behaviors. This, along with the information from cluster characterization, allows us to establish a connection between hydrological behaviors and physical traits. Finally, this framework can be applied at multiple scales, from continental to regional. When tested on a regional scale, the method automatically modifies the network topology to reflect the traits patterns relevant to the area under investigation.