Quantitative Evaluation of AutoNowCaster Convective Likelihood

National Weather Service (NWS) Weather Forecast Offices rely on meteorological observations, numerical model outputs, satellite data, and radar data to forecast the weather.  Because it is practically impossible for a forecaster to review and assimilate all of the relevant information needed to forecast the weather, a number of automatic data integration systems which support the weather forecast process have been developed.  One such system is the National Center for Atmospheric Research’s AutoNowCaster (ANC), which is being made operational by the NWS Meteorological Development Laboratory.  ANC integrates overlapping areas of convective instability, surface convergence, trigger mechanisms such as convergence boundaries, and large-scale forcing in order to produce nowcasts of convective likelihood which delineate areas where storms are likely to initiate and be sustained.

ANC’s nowcasts of convective likelihood will become available to forecasters in September, 2016.  Therefore, it is important to evaluate these nowcasts quantitatively in order to provide statistical scores so that forecasters can have confidence using this product.  However, evaluating gridded nowcasts of thunderstorm initiation is challenging because it is hard to delineate convective initiation objectively and distinguish it from existing storms.  In addition, it is difficult to evaluate gridded nowcasts statistically because of the high sensitivity of statistical scores to small errors in location and/or timing, in part because areas with active convection are lumped together with areas with no convection.  For this study, an objective method was used to classify radar observations as being new, ongoing, or decaying storms.  To reduce the high sensitivity of statistical scores to small errors in location, a relaxation technique was applied to both the forecast and the truth fields.  This paper presents the results of ANC’s performance using data within the May, 2012 to September, 2012 time frame.