Segmenting Radar Reflectivity Data using Texture

A novel method of performing multiscale segmentation of radar reflectivity data using statistical properties within the radar data itself is introduced. The method utilizes a K-Means clustering of texture vectors computed within the reflectivity scan.

Splitting an image into several components, by assigning one of these components to each pixel in the image, is termed image segmentation. The traditional way to segment radar reflectivity images is to term contiguous areas within a reflectivity band, for example all adjoining gates with reflectivity values between 40 and 45 dBZ a ``cell'' or a region~\citep{scit}. There are numerous problems with such hard thresholds. These have typically been resolved heuristically, using runs with tolerances~\citep{scit} or using fuzzy logic~\citep{aires}.

In this paper, we present a more sophisticated approach that uses, besides the actual reflectivity value within a gate, the distribution of reflectivity values around that gate. This distribution is used to cluster similar pixels together~\citep{icip2000}. We demonstrate that the quality of the segmentation is noticably better.