The Use of Ensemble Modeling of Suspended Sediment to Characterize Uncertainty

Climate and land cover changes have the potential to intensify rates of soil erosion and sedimentation in large watersheds, which can adversely affect aquatic life and poses a critical challenge for water treatment and reservoir management. The goal of this research is to develop a modeling ensemble for estimating sediment transport within large-scale mountainous catchments (>1000 km²). The results from four sediment modules inserted into a common hydrologic framework are presented to quantify uncertainty and improve predictability. The ensemble includes empirical modules: monovariate rating curve (MRC) and the modified universal soil loss equation (MUSLE), a stochastic module: multi-variate regression using a generalized linear model (GLM) and a physically-based module: distributed hydrology soil vegetation model (DHSVM) sediment model. Calibration results from a multi-objective optimization routine are presented that optimize parameters and identify performance tradeoffs. The GLM module had the overall highest performance for both daily (NSE=0.84) and event-based (NSE=0.99) predictions. The MRC module performed well under both time steps (NSE=0.64, NSE=0.91). The MUSLE module had the highest performance in percent bias (−0.56%) of all the modules, though it performed poorly in timing and variability for both time steps (NSE=0.22, NSE=0.49). The DHSVM module performed the poorest under the daily simulation (NSE=−0.24), but the skill was greatly enhanced for event-based predictions (NSE=0.96) reflecting the influence of temporal discretization. This work highlights the tradeoffs in sediment prediction across a range of model structures with key differences in daily versus event-based model performance.