On the Use of Data Uncertainty in Hydro-Climatic Modeling: A Bayesian Approach

Sea surface temperature (SST) is an important link between global climate and regional hydrology, and consequently its use is ubiquitous in statistical and physical models for hydrologic prediction. Historical observations of SST are based on in situ measurements from ships and buoys. These measurements have large associated uncertainties that have been archived with the data, but are vastly ignored in the literature either because of lack of appropriate tools or because the benefits of engaging uncertainties are not well understood. In this study, a Bayesian framework is developed to incorporate data uncertainty when performing supervised (regression analysis) and unsupervised (principal component analysis) learning. The developed methods are applied to predict all Indian summer monsoon rainfall (AISMR) using SST as inputs. The benefits of engaging data uncertainties, which include better assessment of predictions skills, are discussed.