Impacts of Neighborhood Approaches for Verification of Gridded Products

Neighborhood approaches are often employed to remove the penalty for small translation errors between forecasts and observations. Common techniques involve the choice of a single value (such as the maximum or mean) to represent the neighborhood, square or circular shaped regions, and overlapping or non-overlapping neighborhoods. The impacts of these choices on the resulting verification metrics using real products were investigated by NOAA/ESRL/Global Systems Division. Climatological baselines were also computed to differentiate between methodological artifacts and true forecast skill. The results showed that using either the maximum or mean of neighborhood values has drawbacks, and the sensitivity to the choice of representative value depends on the characteristics of the gridded field (e.g. sparsity, variability, and bias). The neighborhood shape and overlap fraction were found to have a low impact, suggesting that computationally efficient methodologies may be preferred.