Tracking features in a time-series of images requires a consistent and stable clustering algorithm. If the clustering algorithm cannot produce similar results for similar images, in other words, the clustering output does not change continuously with changes in the input images, then any object-based tracking algorithms will be severely hindered. Similarly, another important issue with clustering algorithms is the impact of input parameters on resulting clusters. The clustering algorithm becomes difficult to use if a user cannot easily fine-tune the parameters to the desired results.

In this work, the Strong Point Analysis (SPA), which is a threshold-based clustering algorithm, is developed to produce consistent and stable clustering results. SPA uses a statistically-driven approach for determining thresholds, as well as a feature to dynamically re-evaluate the thresholds in various regions of the image being processed. Using statistics to determine the clustering thresholds allow for a more dynamically reactive algorithm, as opposed to using explicit thresholds. In addition, SPA essentially adapts the thresholds to regions within an image, as opposed to using the thresholds globally. This “sub-clustering” feature is useful for complicated images.

The goal of the algorithm is to consistently and reliably identify the salient features of an image. SPA has demonstrated some remarkable properties while tested upon radar data from several Weather Surveillance Radar 1988 Doppler (WSR-88D) sites and the Phased Array Radar (PAR) at the National Weather Radar Testbed (NWRT) in Norman, Oklahoma. Given that the field of radar reflectivities typically does not vary too much for each successive scan (both statistically and spatially), SPA can then produce consistent cluster features as input to tracking algorithms. Moreover, the clustering results changes fairly continuously with changes in the input parameters, which can then be useful for radar operators to fine-tune the clustering to more desirable results. These properties are essential for reliable and practical tracking of storm features over time.

The purpose of this project is to introduce SPA and to showcase its efficacy as applied to radar reflectivity data. This project also shows the independence of the algorithm from any particular source of radar data. The SPA was demonstrated using a wide range of events from localized thunderstorms to squall-lines to large-scale synoptic events. Also, the sources of radar data come from different radar stations, both geographically and technologically. The purpose of which is to demonstrate the versatility of SPA in being able to produce relevant clusters for whatever kinds of shapes and sizes of the features.