ALBERT

Description: [1]  This paper explores the use of a parametric geostatistical model for combining rainfall characteristics derived from raingauge data with the same characteristics derived from remote-sensed data-sets. Hypotheses can then be tested about which predictors significantly increase precision of an estimated characteristic. Although applicable wherever ground-level data and remote-sensed data are to be combined, the statistical procedure set out in the paper is developed for two examples of rainfall characteristics: (i) G , the mean annual rainfall at an ungauged site, conditional on knowledge of two predictor variables T (the mean annual rainfall calculated from the TRMM 3B42 data-set for 1998–2009 ), and C (mean annual rainfall derived from the CMORPH data-set for 2003–2009 ); (ii) the mean annual maximum one-day rainfall H , interpolated using the same modelling procedure with predictor variables T and C derived from annual maximum one-day rainfalls in the same remote-sensed data-sets. Prediction errors showed no bias, skewness of distribution, or spatial heterogeneity. The model's generality means that it could be used with any predictors other than T and C , possibly derived from other satellite data-sets or radar. Provided that predictor variables are correlated with the variable to be predicted, it is not necessary for the model relating them to be fitted using data from identical periods, nor for the grid spacing of T and C to be identical. Model performance was evaluated by using a “leave-one-site-out” procedure, which showed that the RMSE of model predictions at omitted sites was smaller than RMSEs obtained from five other well-known spatial predictors.