A Bayesian hierarchical nonhomogeneous hidden Markov model for multisite streamflow reconstructions
AffiliationUniv Arizona, Sch Geog & Dev
MetadataShow full item record
PublisherAMER GEOPHYSICAL UNION
CitationA Bayesian hierarchical nonhomogeneous hidden Markov model for multisite streamflow reconstructions 2016, 52 (10):7837 Water Resources Research
JournalWater Resources Research
Rights© 2016. American Geophysical Union. All Rights Reserved.
Collection InformationThis item from the UA Faculty Publications collection is made available by the University of Arizona with support from the University of Arizona Libraries. If you have questions, please contact us at email@example.com.
AbstractIn many complex water supply systems, the next generation of water resources planning models will require simultaneous probabilistic streamflow inputs at multiple locations on an interconnected network. To make use of the valuable multicentury records provided by tree-ring data, reconstruction models must be able to produce appropriate multisite inputs. Existing streamflow reconstruction models typically focus on one site at a time, not addressing intersite dependencies and potentially misrepresenting uncertainty. To this end, we develop a model for multisite streamflow reconstruction with the ability to capture intersite correlations. The proposed model is a hierarchical Bayesian nonhomogeneous hidden Markov model (NHMM). A NHMM is fit to contemporary streamflow at each location using lognormal component distributions. Leading principal components of tree rings are used as covariates to model nonstationary transition probabilities and the parameters of the lognormal component distributions. Spatial dependence between sites is captured with a Gaussian elliptical copula. Parameters of the model are estimated in a fully Bayesian framework, in that marginal posterior distributions of all the parameters are obtained. The model is applied to reconstruct flows at 20 sites in the Upper Colorado River Basin (UCRB) from 1473 to 1906. Many previous reconstructions are available for this basin, making it ideal for testing this new method. The results show some improvements over regression-based methods in terms of validation statistics. Key advantages of the Bayesian NHMM over traditional approaches are a dynamic representation of uncertainty and the ability to make long multisite simulations that capture at-site statistics and spatial correlations between sites.
NoteFirst published: 12 October 2016; 6 month embargo.
VersionFinal published version
SponsorsScience and Technology grant from Bureau of Reclamation; National Science Foundation [CNS-0821794]; University of Colorado Boulder