Sensitivity of Precipitation Forecast Skill Scores to Bilinear Interpolation and a Simple Nearest-Neighbor Average Method on High-Resolution Verification Grids

Abstract

Grid transformations are common postprocessing procedures used in numerical weather prediction to transfer a forecast field from one grid to another. This paper investigates the statistical effects of two different interpolation techniques on widely used precipitation skill scores like the equitable threat score and the Hanssen?Kuipers score. The QUADRICS Bologna Limited Area Model (QBOLAM), which is a parallel version of the Bologna Limited Area Model (BOLAM) described by Buzzi et al., is used, and it is verified on grids of about 10 km (grid-box size). The precipitation analysis is obtained by means of a Barnes objective analysis scheme. The rain gauge data are from the Piedmont and Liguria regions, in northwestern Italy. The data cover 243 days, from 1 October 2000 to 31 May 2001. The interpolation methods considered are bilinear interpolation and a simple nearest-neighbor averaging method, also known as remapping or budget interpolation, which maintains total precipitation to a desired degree of accuracy. A computer-based bootstrap technique is applied to perform hypothesis testing on nonparametric skill scores, in order to assess statistical significance of score differences. Small changes of the precipitation field induced by the two interpolation methods do affect skill scores in a statistically significant way. Bilinear interpolation affects skill scores more heavily, smoothing the maxima, and smearing and increasing the minima of the precipitation field over the grid. The remapping procedure seems to be more appropriate for performing high-resolution grid transformations, although the present work shows that a precipitation edge-smearing effect at lower precipitation thresholds exists. Equitable threat score is more affected than Hanssen?Kuipers score by the interpolation process, since this last score weights all kind of successes (hits and correct no-rain forecasts). Correct no-rain forecasts at higher thresholds often outnumber hits, misses, and false alarms, reducing the sensitivity to false alarm changes introduced by the interpolation process.