Interpretation of radar data is often done in a subjective, qualitative manner. While features of interest can be identified and tracked in this manner, this method does not lend itself for automation or mass processing of radar data. Use of centroid tracking is one method to approach this problem, but it is limited to convective cores. We believe that more robust clustering techniques can be developed that will allow better identification of storm features. Some applications of these techniques include: identification and study of features other than just convective cores, development of object-oriented codes to build databases from large amounts of radar data, and automation to provide operational meteorologists with more comprehensive (and easily accessible) information.

Clustering techniques for segmentation of radar images is problematic; the clustering methods segment the image, but often replace a complex image with an even more complex set of complicated segments. The statistical relevance of any particular segment is not always clear. Analysis and interpretation in the presence of white noise is straightforward, but analysis of clusters in the presence of spatially correlated noise is much more complex. Our study in this area follows our earlier work on Coherent Structure Detector (CSD). The method is based on a definition of an incoherent signal and controls the test size without assumptions about the process. We apply this concept of coherency to multidimensional images for the detection and characterization of coherent image segments. Our recent analysis establishes the test size and provides an algorithm for calculating p-values.

Two viewpoints of incoherency are considered: spatially correlated noise (white noise passed through a linear spatial filter), and Poisson clutter. Both are appropriate for analysis of radar images. Spatially correlated noise will be present in the fine structure of high-resolution images due to clouds and other concentrated scatters (insects, birds and dust), inversion layers, and air density variations. Poisson clutter would follow from clutter (ground, aircraft, etc.). The model of an incoherent image is used to build rigorous statistical tests for the classification of image segments, in terms of p-values, as coherent image structures or the random results of incoherent noise. The techniques are non-parametric and no explicit knowledge is needed of the spatial noise spectrum or Poisson amplitude distribution. Proofs of test size are already complete, but power is a critical issue. Performance will be determined by noise characteristics, the image segment properties, and clustering algorithm.