This paper proposes a spatial data mining approach as a new tool for verifying and understanding ensemble precipitation forecasts. With this approach, particular attention is given to the spatial distribution characteristics of precipitation. Two particular issues are addressed in this paper: the assessment of forecast quality through pattern recognition which can identify errors due to phase shift or displacement, rotation and deformation, and investigation of association among observations and forecasts. This approach, when applied to ensemble precipitation forecasts, can take into account the spatial distribution characteristics of individual ensemble members, which are otherwise difficult to describe, especially quantitatively with conventional composite charts or “spaghetti” charts. Based on the knowledge gained from the distributions, a further step is taken to derive association rules between the observations and forecasts from the database. Eventually, the approach provides information on the non-uniform distribution of individual ensemble members and reveals the relative spatial associations among the observations and forecasts with certain confidence.

As a first step, we test the proposed approach using a simulated dataset, in which observed and forecast precipitation regions are approximated by ellipses. A spatial aggregation technique is used to select patterns or features from the fields, and a density-based clustering algorithm is then used to extract the spatial precipitation distribution information. Finally, spatial data mining association rules are derived from the database to reveal the relationships among the observations and forecasts.