Applications of Biomimetic Image Processing to Storm Segmentation and Tracking

Segmentation of storm objects within reflectivity and other meteorological feature data forms a foundation for a variety of analysis tasks, including the tracking of storms over time. The current approach to storm segmentation within the National Severe Storms Laboratory is based on a k-means clustering algorithm that makes use of a modified watershed transformation, and subsequent storm tracking uses spatially local template matching techniques that incorporate Kalman filter model estimates. Operational use of this methodology has revealed that it can provide poor results, particularly for fast moving storms with shapes that change quickly, split, or merge with other storms.

Although the existing segmentation and tracking methods have difficulty in several situations, it is usually obvious to a human observer where the storm objects are and how they have moved. As such, there is a need for either revision to the existing approach or the development of new segmentation and tracking algorithms.

Recent research in the field of computer vision has produced biomimetic paradigms that enable an interpretation of images that shares characteristics with biological vision systems. This presentation discusses the applicability of such image processing techniques to the problem of storm segmentation and any effect they may have on storm tracking accuracy. A summary of existing and potential storm segmentation techniques is also given.