Abstract
One of the most important parameters for spring runoff forecasting is the snow water equivalent on the watershed, often estimated by kriging using in situ measurements, and in some cases by remote sensing. It is known that kriging techniques provide little information on uncertainty, aside from the kriging variance. In this paper, two approaches using Bayesian hierarchical modelling are compared with ordinary kriging; Bayesian hierarchical modelling is a flexible and general statistical approach that uses observations and prior knowledge to make inferences on both unobserved data (snow water equivalent on the watershed where there is no measurements) and on the parameters (influence of the covariables, spatial interactions between the values of the process at various sites). The first approach models snow water equivalent as a Gaussian spatial process, for which the mean varies in space, and the other uses the theory of Markov random fields. Although kriging and the Bayesian models give similar point estimates, the latter provide more information on the distribution of the snow water equivalent. Furthermore, kriging may considerably underestimate interpolation error. Copyright © 2006 Environment Canada. Published by John Wiley & Sons, Ltd.