Key points are not available for this paper at this time.
Precipitation patterns exist on a continuum of scales where generally larger scale features facilitate longer useful prediction times at the expense of coarser resolution. Favorable measurement range and resolution make weather radar observations an attractive choice for input to automated short-term weather prediction (nowcasting) systems. Previous research has shown that nowcasting performance can be improved by spatially filtering radar observations and considering only those precipitation scales that are most representative of pattern motion for prediction or filtering those scales from predicted fields deemed unpredictable by remaining past their lifetimes. It has been shown that an improvement in nowcasting performance can be obtained by first applying a nonlinear elliptical spatial filter to observed Weather Surveillance Radar 88 Doppler vertically integrated liquid water fields to predict motion of larger scale features believed to better represent the motion of the entire precipitation pattern for forecast lead times up to 1 h. It has also been shown in the literature that wavelet transform can be used to develop measures of predictability at each scale and adaptive wavelet filters can be designed to remove perishable scales from predicted continental-scale reflectivity fields according to prediction lead time. This paper investigates the adaptation of both of these approaches and Fourier filtering to evaluate the effects of scale filtering on nowcasting performance using a Fourier-based nowcasting method and high-resolution Engineering Research Center for Collaborative Adaptive Sensing of the Atmosphere radar reflectivity data. A maximum improvement of approximately 18% in terms of Critical Success Index was observed by applying Fourier filtering in the context of truncating Fourier coefficients within the prediction model to the observed sequence of reflectivity fields used for assimilation. In addition, applying Fourier filtering to the resulting predictions showed a maximum reduction in mean absolute error of approximately 14%.
Ruzanski et al. (Fri,) studied this question.