Identifying sediment-discharge event types with a data-based clustering approach

Suspended sediment poses a risk to human and natural systems in terms of compromising water quality, flood hazard, hydropower production, and aquatic habitats. In many rivers the bulk of annual suspended sediment yield is mobilised and transported during (extreme) episodic runoff. Understanding such sediment-discharge events, including their drivers, may inform management strategies aimed at mitigating potential detrimental effects. Event-based analysis of local time series of suspended sediment transport has become a common approach to infer the dominant drivers and processes of sediment dynamics at the catchment scale. The increasing availability of detailed and continuous monitoring time series data enables us to use machine-learning techniques to identify groups of similar events, i.e. event types, and test whether and how these groups reflect similar catchment conditions and hydro-meteorological drivers. We present an approach which automatically detects, characterises and clusters sediment-discharge events. Hydrograph separation is used to automatically detect events, which are then filtered based on suspended sediment magnitude. The detected events are subsequently characterised with a selection of metrics, which are transformed into uncorrelated event characteristics with principal component analysis. Based on these characteristics events are clustered using a Gaussian mixture model. Finally, the identified event types are interpreted using catchment metrics describing antecedent conditions, hydrometeorological forcing, and catchment freezethaw state and snowcover.Applying our approach to a high alpine, glaciated catchment we find that the event regime in the catchment is mainly defined by event magnitude, hysteresis and event shape complexity. However, for the clustering suspended sediment and streamflow magnitude, and event shape complexity are the most important factors, whereas sediment discharge hysteresis is less relevant. The four identified event types are attributed to (1) compound rainfall-melt extremes, (2) glacier and seasonal snow melt, (3) freezethaw-modulated snow-melt and precipitation events, and (4) late season glacier melt. Our approach enables event-based analysis of riverine sediment fluxes, by detecting and grouping similar events together, which can in turn be interpreted to understand under which conditions episodic sediment fluxes occur in the target catchment.