An object-based approach to the verification of convective forecasts has the potential to provide detailed insight into forecast quality and highlight strengths and weaknesses where standard scalar verification measures fall short. In this approach, a convolution and thresholding procedure is applied to both the forecast and observation fields to define objects. The object definition parameters are tuned to each application to create objects which maintain the intended characteristics of the forecast. A fuzzy logic algorithm is applied to merge clusters of objects within each field and match those clusters of objects between the two fields. Based on the results of the merging/matching process, meaningful differences between the forecast and observation fields can be discerned. By analyzing large volumes of data using an object-based verification approach, forecast quality trends come into focus.
This paper reviews the application of an object-based approach to the verification of one-hour convective forecasts for the 2005 Dallas/Fort Worth AutoNowcaster demonstration project. The forecast domain for this demonstration project consists of a region approximately one third the size of Texas and centered over the Dallas/Fort Worth area. The forecast is generated with a forecaster-in-the-loop (FIL) who enters boundaries and consists of both a growth and decay component as well as an initiation component forecasting three levels of likelihood for convective initiation. The observation field for this evaluation is based on a mosaic of radar observations in this region.
The skill of the growth and decay component is assessed by comparing it to the quality of an extrapolation-only forecast with no growth and decay and no FIL. To analyze the initiation component, the initiation likelihood is thresholded at each of three levels and combined with the growth and decay component into a single forecast. The parameters used in the definition of the corresponding observation objects and in the fuzzy logic object merging/matching scheme are varied as appropriate for each initiation level. The impact on forecast quality of incorporating the initiation component will be discussed.