Clarifying Applications of Neighborhood Approaches to High-Resolution Forecasts

“Neighborhood approaches” are regularly employed to post-process and verify high-resolution numerical weather prediction model output in both deterministic and ensemble settings. However, there are multiple ways neighborhood approaches can be applied, which can have profound impacts on subsequent verification metrics and cause confusion.

For example, in ensemble frameworks, neighborhood approaches have been used in two primary manners that appear deceptively similar but in fact define events over different spatial scales and yield fields with different interpretations: the first produces probabilities interpreted as likelihood of event occurrence at the grid-scale, while the second produces probabilities of event occurrence over spatial scales larger than the grid-scale. These two neighborhood-based probabilistic fields, when verified with common metrics (e.g., the Brier score), provide very different indications of forecast quality. Unfortunately, some studies have confused the two methods, leading to challenges when attempting to synthesize findings across multiple published papers.

Similarly, three ways of populating neighborhood-based 2 x 2 contingency tables have been proposed to evaluate deterministic forecasts. Although the three methods only subtly differ, verification metrics computed from the various 2 x 2 tables give substantially different impressions of forecast performance.

Accordingly, this presentation will review the various flavors of neighborhood approach application to high-resolution forecasts. Furthermore, real-data examples of how the various neighborhood-based methods yield statistically significantly different objective conclusions about forecast performance will be provided, underscoring the critical need for thorough descriptions of how neighborhood approaches are implemented.