Applying a Neighborhood Fractions Sampling Approach as a Diagnostic Tool

Abstract

he fractions skill score (FSS) belongs to a class of spatial neighborhood techniques that measures forecast skill from samples of gridded forecasts and observations at increasing spatial scales. Each sample contains the fraction of the predicted and observed quantities that exist above a threshold value. Skill is gauged by the rate that the observed and predicted fractions converge with increasing scale. In this study, neighborhood sampling is applied to diagnose the performance of high-resolution (1.67 km) precipitation forecasts over central Florida. Reliability diagrams derived from the spatial fractions indicate that the FSS can be influenced by small, low-predictability events. Further tests indicate the FSS is subtly affected by samples from points on and near the grid boundaries. Inclusion of these points tends to reduce the magnitude and sensitivity of the FSS, especially at large scales. An attempt to mine data from the set of neighborhood fractions was moderately successful at obtaining descriptive information about the precipitation fields. The width of the distribution of the fractions at each scale provided information concerning forecast resolution and sharpness. The rate at which the distribution of the fractions converged toward the domain mean with increasing scale was found to be sensitive to the uniformity of coverage of precipitation through the domain. Generally, the 6-h forecasts possessed greater spatial skill than those at 12 h. High-FSS values at 12 h were mostly associated with evenly distributed precipitation patterns, while the 6-h forecasts also performed well for several nonuniform cases.