Spatial verification using a true metric

Verifying high-resolution forecasts is challenging because forecasts can be considered good by their end-users even when there is no pixel-to-pixel correspondence between the forecast and the verification fields. Many of the verification methods that have been proposed to address the verification of high-resolution forecasts are based on filtering, warping or searching within a neighborhood of pixels in the forecast and/or the verification fields in order to retain the capability to use a simple metric. This is because it is necessary for a verification score to be a metric to allow comparisons of forecasts. In this paper, we devise a computationally simple scalar spatial verification metric that is capable of ordering forecasts without preprocessing the fields. The metric is based on the insight that in the verification problem, the observation field can be considered a reference field that forecast fields are ordered against. This new metric is demonstrated on synthetic and real model forecasts of precipitation.

► We describe a metric for evaluating model forecasts.
► The metric is easy to compute.
► There is no need to warp or filter the input fields.
► The metric can deal with displacement errors.
► Demonstrated on geometric, perturbed and real data.