Extended Abstract (68K)7B.3
Development of an “events-oriented” approach to forecast verification
Michael E. Baldwin, CIMMS/Univ. of Oklahoma and NOAA/NSSL, Norman, OK; and S. Lakshmivarahan and J. S. Kain
The ultimate motivation for this paper is to develop new meteorological forecast verification techniques, in particular, to provide useful information on the accuracy of spatial fields that may contain high-amplitude, small-scale features, such as those which are often observed in rainfall data or predicted by high-resolution numerical weather prediction models. For spatial fields that contain high-amplitude, small-scale features, small forecast errors in phase, displacement, or time lag can produce very large differences between forecast and observed scalar variables at specific locations, leading to relatively poor traditional measures of forecast accuracy. Despite the potential for large errors at specific points in time and space, predictions that contain similar spatial structures, scales, and amplitudes to those that are observed, albeit with phase/displacement errors, may be of considerable value to certain users. Consequently, the value of forecasts that contain small-scale, high-amplitude features will not be accurately expressed when using traditional methods of analyzing forecast accuracy. In order to obtain more useful information on the accuracy of forecasts that contain high-amplitude small-scale features, we plan to develop techniques that mimic, as closely as possible, how a human subjectively assesses the skill of forecast fields. To begin with, we plan to expand the paradigm of “point-to-point” verification to the verification of “events” or “objects”, which can be defined as meteorological phenomena. An “event” on a weather map can be defined as a region containing similar meteorological attributes or similar statistical properties. The main challenge in this work is to develop an objective method to determine which regions within the spatial field posses similar attributes, and then categorize or classify those features or events found in forecast and observed fields. New measures of accuracy can be developed by examining the similarity between forecast and observed “events.” In order to determine groups or clusters of points with similar characteristics, we naturally turn to the discipline of data mining. Initial work has consisted of analyzing observed rainfall fields in order to classify features found within. A statistical distribution is fitted to the observed histogram of values found within a region on the rainfall map. Cluster analysis is then performed using the parameters of the statistical distribution as attributes. Our initial results have shown that this technique has some success in discriminating between convective and non-convective rainfall features, but no success at discriminating to finer sub-classes within those main classes (such as linear vs. cellular convective rainfall). Restuls from this work in progress will be presented at the conference.
Session 7B, Statistical Evaluation II
Wednesday, 14 August 2002, 10:30 AM-12:00 PM
Previous paper Next paper