GC53B-1209
Improving predictive certainty and system understanding with watershed hydrology models

Friday, 18 December 2015
Poster Hall (Moscone South)
Christa Kelleher, Duke University, Durham, NC, United States, Brian L McGlynn, Duke University, Nicholas School of the Environment, Durham, NC, United States and Thorsten Wagener, University of Bristol, Civil Engineering, Bristol, United Kingdom
Abstract:
Modeling at the intersection of climate variability and hydrology is complicated by uncertainties that make predicting physical behavior a challenge. Environmental models used to simulate how climate will impact hydrology are typically complex, demand many spatial and temporal data inputs, contain numerous parameters, and can be computationally expensive. Distributed models in particular complicate the assessment of how uncertainty in the model framework, inputs, parameters, and observations impact predictive uncertainty. In addition, future climate perturbations may alter the magnitude of these uncertainties. Here, we focus on model parameters as a key source of uncertainty. Identifying those model parameters that most influence the predictions at a particular place can reduce a complex, multidimensional problem to a simpler form. We demonstrate how sensitivity analysis in the absence of observational streamflow can be used to identify sensitive model parameters by conditioning a model on climate data and a priori parameter ranges. We apply this approach to five headwater catchments in the Tenderfoot Creek Experimental Forest located in central Montana using the Distributed Hydrology-Soil-Vegetation Model. Across these five sub-catchments, climate clearly organizes parameter sensitivities. To further explore the relationship between parameter sensitivities and climate, we assess how parameter sensitivities change when meteorological forcing data is perturbed to reflect natural variability at the site. This general approach can support uncertainty reduction. However, parameter equifinality will still impact finer scale predictions of any environmental variable in space and time. As such, improving our certainty in environmental predictions should evaluate point predictions as well as simulations of internal catchment behavior, and must not only rely on our use of computational methods but on our basic understanding of system functioning.