Verification statistics spanning three forecast seasons (2015-2017) are presented. The UDFCD’s network of over 150 tipping bucket rain gages is used to supplement the commonly used NOAA Stage IV precipitation estimates for a spatially well-sampled Quantitative Precipitation Estimate of high rainfall rates.
The following six findings are presented. First, ensemble guidance could discriminate the locations of heavy rainfall threats surprisingly well, even during summer “weak flow” regimes. This is hypothesized to be partially because of the importance of the region’s diurnal topographically-induced circulation and may not apply elsewhere. Second, models are generally found to underestimate maximum 1-hr QPF. This bias was removed using quantile mapping. Third, each ensemble member performs roughly in line with the rest (after bias correction); in other words, no evidence is found to support a weighting of one model over another. Fourth, the inter-ensemble maximum 1-hour QPF placed a reasonable bound on that day’s heavy rainfall amount, allowing us to term this the “realistic worst-case scenario”. Fifth, post-processing the ensemble distribution using a logistic regression resulted in better performance than using raw ensemble data (even after bias correction). Lastly, despite correcting systematic biases using quantile mapping, additional flow dependent biases remained. For example, in cases with anomalously low mid-level (i.e. steering) wind speed, QPF underestimates were noted. It is shown that wind speed forecasts can be included in the post-processing to further improve performance.
Overall, we argue that a post-processed, ensemble based system is the key to improving heavy rainfall QPF moving forward.