Automated Detection of Radar Severe Weather Signatures

This paper introduces an unsupervised learning method for the automated detection of several severe weather radar signatures. Hand identification of weather event radar signatures is a tedious and impractical task given the large amounts of data produced by modern radar installations. Furthermore, data is usually sorted by date with no means of distinguishing between days containing desired features, and days that do not. Automating this process would more efficiently identify scans of interest in both real time and in data-mining applications. The method proposed starts by examining the detection of bow echoes and linear convective systems. Each radar scan is segmented using a hierarchical texture-based segmentation that operates on the radar products. Each segment consists of a distribution of pixels from which characteristics are derived. These characteristics include the variance, curvature, length, and the goodness of fit of a second order polynomial to the distribution. These texture and morphology parameters are used to identify squall lines and bow echoes within each scan. Results indicate that this method performs with a high detection rate and low false detection rate on the tested data. Furthermore, the results of this algorithm can be queried to provide a data-mining tool for scientists and educators. This work is an important step in allowing for semantic exploration of weather radar data. The ability for users to search for data by features rather than by the underlying temporal organization vastly increases the usefulness of current radar databases.