A comparison of three geostatistical procedures for rainfall network optimization
Résumé
This paper presents the methodology used for the optimization of a precipitation observation network by predicting monthly averaged rainfall. The predicted rainfall maps were obtained by comparing different geostatistical algorithms such as kriging with external drift, regression-kriging and cokriging which integrates the altitude as secondary variable. The experimental simple semi-variograms and cross-variograms are constructed and fitted to estimate the rainfall levels and the estimation variance and to develop corresponding contour maps. The estimation variance is utilized to locate the regions where new stations must be added to obtain less important error estimation. The techniques are illustrated using monthly rainfall averaged observations measured at 140 climatic stations in Tunisia over an area of 164150 km2. The three geostatistical algorithms were compared. Contour diagrams for kriging with external drift and regression-kriging were similar and exhibited a pattern corresponding more closely to local topographic features, while cokriging showed a smooth zonal pattern. Smaller prediction errors are obtained for the regression-kriging algorithm. The estimation variance varies with seasons and regions. In fact, the interpolation system seems more sensitive, mainly in autumn, when adding new stations in the south of the country. Whereas, in the north, the network stations gives a good representation of the spatial rainfall repartition.