Performance and Reliability of Probabilistic Scenarios of Interpolated Rainfall Based on Geostatistics

Abstract

Common probabilistic scores are used to evaluate the performance and reliability of interpolated rainfall scenarios based on geostatistics. The kriging variance is used jointly with the best linear unbiased estimator at the interpolation sites to render an ensemble of rainfall depth at each time step, providing dynamic information about the uncertainty of the interpolated field. Tests are conducted with rainfall data, extending from June 1, 2000, to October 31, 2005, from 50 stations located in the Saint Lawrence Valley, Canada. Three categories of variograms are examined: automated variograms that change at each time step; variable variograms that associate an average variogram to a specific interval of mean areal rainfall; and unique variograms, which are averages over the whole series. Simple, ordinary, and universal kriging are also examined. Rainfall scenarios constructed from automated variograms, simple kriging, and 30 neighboring stations presented acceptable sharpness and overall performance, including reliable exceedance probability and confidence intervals. An optimal choice of neighborhood is critical to performance and reliability.