Using Deep Learning to Extract Regions of Interest (ROI) in Real-Time from Geostationary Satellite Data

Atmospheric science and numerical weather models have a data problem: keeping up with the quantities of satellite data in real-time applications. These include but are not limited to alerts, now-casting, and data assimilation. Some of the biggest questions are how do we extract valuable information from satellite data quickly, is this something that a human would identify or something new to consider, and what do we do with it once its been identified? Using deep learning, we developed different UNET models that take inputs from the Geostationary Operational Environmental Satellites (GOES) water vapor channel and produce labeled ROI within one or a few seconds. In this presentation, I will discuss the conclusion of our work on extracting cyclone information from satellite data using deep learning. I propose that deep learning is a valuable tool for quick satellite analysis and image segmentation based off the success of the model in identifying both tropical and extratropical cyclone regions from only one satellite field and doing so in a very fast time. We address the issues of validating the model and quantifying success/failure as well as present alternative ways to measure success in an image segmentation problem when the ROI definitions are not always a binary “yes” or “no”. Further, we address challenges in training deep learning models to detect ROI of rare or extreme weather events and obtaining the labeled dataset necessary in supervised learning.