Systematic uncertainty reduction strategies for developing streamflow forecasts utilizing multiple climate models and hydrologic models

Recent studies show that multimodel combinations improve hydroclimatic predictions by reducing model uncertainty. Given that climate forecasts are available from multiple climate models, which could be ingested with multiple watershed models, what is the best strategy to reduce the uncertainty in streamflow forecasts? To address this question, we consider three possible strategies: (1) reduce the input uncertainty first by combining climate models and then use the multimodel climate forecasts with multiple watershed models (MM-P), (2) ingest the individual climate forecasts (without multimodel combination) with various watershed models and then combine the streamflow predictions that arise from all possible combinations of climate and watershed models (MM-Q), (3) combine the streamflow forecasts obtained from multiple watershed models based on strategy (1) to develop a single streamflow prediction that reduces uncertainty in both climate forecasts and watershed models (MM-PQ). For this purpose, we consider synthetic schemes that generate streamflow and climate forecasts, for comparing the performance of three strategies with the true streamflow generated by a given hydrologic model. Results from the synthetic study show that reducing input uncertainty first (MM-P) by combining climate forecasts results in reduced error in predicting the true streamflow compared to the error of multimodel streamflow forecasts obtained by combining streamflow forecasts from all-possible combination of individual climate model with various hydrologic models (MM-Q). Since the true hydrologic model structure is unknown, it is desirable to consider MM-PQ as an alternate choice that reduces both input uncertainty and hydrologic model uncertainty. Application on two watersheds in NC also indicates that reducing the input uncertainty first is critical before reducing the hydrologic model uncertainty.