Detecting Above-Anvil Cirrus Plumes on a Pixel Scale Using Deep Learning

In satellite imagery, above-anvil cirrus plumes (plumes for short) are the strongest indicator of potentially significant severe weather, appearing on average 30 minutes before severe weather is reported. Real-time plume identification could provide forecasters with information of the convective environment in areas where radar coverage is sparse. One current drawback of plume classification for severe weather prediction is hand labeled classifications can be taxing to produce. Not only are the features relatively small in imagery, but specialized knowledge can be necessary for ideal classifications. A solution to quickly identify plumes over large domains is through training deep learning (DL) models to output plume classifications based on expert drawn labels. With various satellite imagery as input, a UNET (type of DL model) can learn spatio-temporal patterns in data and output predictions on a pixel-scale in a few minutes. For this talk I will show findings of plume classification using different combinations of visible and infrared data for testing data from 2020 over the contiguous United States.