1National Severe Storms Laboratory, NOAA - U.S.A.
2University of Oklahoma - U.S.A.
3University of Wisconsin, Madison - U.S.A.
Segmentation of weather imagery is a problem fundamental to automated weather analysis, but one that has defied good solutions.
Weather images have certain characteristics that cause problems for traditional image processing algorithms — the textural nature of clouds, poor dynamic range and poor spatial resolution. In this paper, we describe a novel method of performing multiscale, hierarchical segmentation of images using texture properties that has been shown to perform consistently better than other untrained, unsupervised texture segmentation algorithms. The images are first requantized using contiguity-enhanced K-Means clustering. Morphological operations and region growing based on textural properties are used to arrive at the most detailed segmentation. Successively coarser segmentations are achieved by the use of inter-cluster distances in an dyadic, agglomerative technique. We demonstrate that this technique of texture segmentation outperforms other segmentation algorithms on radar and satellite weather images and show some advantages of the multiscale technique. We note some problems that are yet to be resolved, especially on weather satellite imagery.