Analysis of scale dependence of quantitative precipitation forecast verification: A case-study over the Mackenzie river basin

Six-hour rainfall accumulations derived from radar observations collected during a 3-day summertime precipitation event over central Alberta (Canada) are used to assess the performance of a regional Canadian numerical weather prediction system for quantitative precipitation forecast verification. We show that radar data provide a simple and efficient way to significantly reduce model phase errors associated with misplacement of predicted precipitation patterns. Using wavelet analysis, we determine that the limiting spatial scale of predictability of the model is about six times its grid resolution for 6 h accumulated fields. The use of longer accumulation periods is shown to smooth out forecast errors that may have resulted from slight phase or time shift errors but does not change the limiting scale of predictability. The scale decomposition of the mean-square forecast error also reveals that scales which cannot be accurately reproduced by the model account for about 20% of the total error. Using classical continuous and categorical scores, we show that significantly better model performance can be achieved by smoothing out wavelengths that cannot be predicted.