We propose an alternate solution where the forecast and observed fields (for example, precipitation) are decomposed into sets of patterns or features. Pattern recognition techniques can be used to classify the dominant feature found within a region (e.g., convective lines, cells, stratiform, and orographic precipitation features). The set of features found in the observed field could then be compared to the set of features found in the forecast field. The joint distribution of forecast and observed features could then be interrogated following the general verification framework developed by Murphy and Winkler (1987). One could also measure errors in displacement, amplitude, areal extent, orientation, etc. of the various types of events classified by this scheme. We plan to use the NCEP 22km Eta Model as well as a parallel version of the Eta Model using the Kain-Fritsch convective parameterization in order to develop these techniques, although the verification methods will be applicable to any model.
In order to develop an objective classification scheme, our initial work involves examining a small target data set containing relatively high-resolution (4km x 4km) observed hourly precipitation fields for 48 cases. To validate the objective classification, the dominant precipitation event for each case was subjectively classified into four categories (lines, cells, stratiform, orographic). To reduce the dimensionality of the problem, the distribution of precipitation amounts for each case was fitted to a gamma distribution using the maximum likelihood estimation technique of Wilks (1990). Cluster analysis using the parameters of the gamma distribution as attributes was then performed. The resulting clusters successfully separate the convective events (lines, cells) from the non-convective (stratiform, orographic). However, this simple analysis fails to distinguish between linear and cellular cases, or stratiform and orographic events. Current results from this work in progress will be presented at the conference.