Improving Historical Data Discovery in Weather Radar Image Data Sets Using Transfer Learning

Historical data discovery is a challenging task for any study in radar meteorology when the region of interest is recorded in weather observations. Weather radars exist in overlapping networks around the globe and are observing the atmospheric conditions around the clock. The observations are stored according to date, time, and location, as opposed to an indexing scheme based on the phenomena present in the scans themselves. Performing feature-based searches in these voluminous data sets is, at current, impossible. This research seeks to enable users seeking to study such phenomena as a mechanism for locating events of interest by leveraging recent progress from the fields of transfer learning and computer vision. Specifically, this work illustrates a methodology for performing image classification of precipitation regimes on colormapped weather radar scan images, as opposed to the raw data itself. This system reduces the data needed to perform this classification by two orders of magnitude, increasing throughput and democratizing usage of the deep learning tools for this task by allowing training and testing of the models on modest compute systems.