Hail can be among the most damaging of severe weather phenomena, but predicting whether a passing thunderstorm might start spitting pea-sized (or golf ball–sized) hailstones is notoriously difficult. A new approach using machine learning techniques related to facial recognition technology is giving meteorologists a new tool for mapping how various components of a storm might add up to dangerous hail conditions.
Some types of thunderstorms, such as supercells, are more likely to produce hail than others. But the sheer scale of thunderstorms, which can stretch for kilometers and contain multitudes of intrastorm interactions, makes it difficult for computers to accurately model and predict storm behavior, said David John Gagne, a machine learning scientist at the National Center for Atmospheric Research (NCAR) in Boulder, Colo., and lead author of the new study, published in Monthly Weather Review.
Existing severe weather forecasting models tend to focus on slices of a storm because of the computational complexity of considering the whole storm. “In the past, we have tended to focus on single points or just vertical profiles, but the whole structure of the storm is really important in determining whether or not it will produce hail,” Gagne said.
Drawing upon machine learning technology sometimes used to identify features of individual faces, Gagne and colleagues at NCAR trained a deep learning model called a convolutional neural network to recognize various storm features known to produce hail.
In facial recognition technology, a computer program assesses individual features and the arrangement of features in relation to one another to identify a person’s face.
“To apply this technology to hail forecasting, we fed a computer program lots of data and images of storms and asked it to look at the shape and various components and how they relate to one another,” Gagne said. “As the computer finds patterns, we teach it to associate those patterns with the probability of whether a given storm will produce hail.”
Machine Learning Techniques Expand Meteorological Forecasting
Supercell thunderstorms tend to produce hail because of the large, wide updrafts that carry ice particles from the lower atmosphere into the troposphere, Gagne said. “As these icy embryos travel long distances through the storm, they have time to grow into larger hailstones.”
Making direct observations of how ice particles develop into hailstones is often impossible because of the damage incurred by instruments in hailstorms, said Amy McGovern, a computer scientist at the University of Oklahoma in Norman who was not involved in the new study. “We don’t fully understand the physical process of hail development because we can’t fly planes or send instruments into severe hailstorms without them being destroyed.”
In recent years, machine learning techniques have “exploded in popularity in the field of meteorology” because of a combination of widely available data processing tool kits and cloud storage, McGovern said.
The new study is among the first to use deep learning techniques to look for patterns in hailstorms, said Ryan Lagerquist, a meteorologist at the University of Oklahoma who was also not involved in the new study.
“So far, efforts have focused on traditional machine learning, which has no awareness of the spatial or temporal relationships of the inputs,” Lagerquist said. “A deep learning algorithm can pick up on spatial structures inside the storm in three dimensions, resulting in a more complete picture than you can get with traditional learning models.”
Gagne and colleagues are in the process of transitioning the deep learning model into a real-time weather forecasting model that can be used by the National Weather Service to make predictions about whether a developing storm may produce damaging hail, Gagne said. “It’s still in experimental mode right now, but we’re hoping to have it fully operational in the next year.”
—Mary Caperton Morton (@theblondecoyote), Science Writer
Morton, M. C. (2019), Finding faces in hailstorms, Eos, 100, https://doi.org/10.1029/2019EO132755. Published on 13 September 2019.
Text © 2019. The authors. CC BY-NC-ND 3.0
Except where otherwise noted, images are subject to copyright. Any reuse without express permission from the copyright owner is prohibited.