Bias-corrected nonparametric correlograms for geostatistical radar-raingauge combination

Abstract

Geostatistical methods have been widely used for quantitative precipitation estimation (QPE) based on the combination of radar and raingauge observations. They are flexible and accurate and allow for radar-raingauge combination in real-time. Even within the area of geostatistical methods, however, a wide range of choices have to be made when planning for a particular application. These choices regard, for example, the actual combination method (e.g., kriging with external drift, cokriging), the kriging neighbourhood (global vs. local), the technique used to estimate the parameters of the geostatical model (e.g., least-squares, maximum-likelihood estimation), and the transformation of the precipitation variable. In addition to these issues, there are a number of options for modeling spatial dependencies in the precipitation data. Correlograms (variograms) for kriging are customarily one-dimensional, but two- or higher-dimensional correlation maps are also used and are one way of taking spatial anisotropy into account. Furthermore, correlograms can be parametric or nonparametric, they can be obtained from the radar or the raingauge data, and they can be estimated flexibly on a case-by-case basis or with data from a longer period of time. Recently, nonparametric correlograms based on spatially complete radar rainfall fields have been used in combining radar and raingauge data [1]. Here, we compare the estimation of nonparametric correlograms with the estimation of parametric semivariogram models conventionally used in geostatistical applications. We identify and explain a bias of the nonparametric correlograms towards too low ranges, and suggest a correction for this bias.