By using a convolutional neural network (CNN), two Stanford professors are analyzing weather patterns that lead to heavy precipitation. This may change the way we prepare for devastating climate impacts
It’s been a year of weird weather. Heck, it’s been a decade. Or two. Or three. Fires, floods, droughts, and other natural calamities have escalated since the world took notice with the so-called Climate Movement in the 1990s. Meanwhile, young people today are increasingly worried about the existential threat posed by climate change.
Climate scientists have long struggled with predicting where and how large weather patterns will fall, which is why two Stanford professors have created a new model using machine learning to help governments and public servants better prepare for future catastrophes.
The mechanism was simple: feed data on large-scale circulation patterns for the past two decades in the Midwest into a convolutional neural network (CNN) to help it recognize when an extreme precipitation circulation pattern (EPCP) would lead to devastating levels of precipitation. Their CNN model correctly identified 91% of extreme-weather days, which is higher than other statistical methods. It also correctly identified 100% of days that exceeded higher precipitation thresholds as EPCPs.
The research reveals that types of atmospheric pressure patterns that lead to heavy precipitation have become more frequent in the Midwest, by about one extra each year. That may seem small, but the researchers, aided by their trained ML model, also noticed that when these atmospheric pressure patterns did swing through the Midwest, the volume of precipitation has definitely increased. The authors believe the implications can be huge. “Our method could be used to better understand changes in extreme events in the Midwest and in other regions of the world,” said the authors, Frances V. Davenport and Noah S. Diffenbaugh. They published their findings in Geophysical Research Letters.
Noah Brenowitz, the Senior Machine Learning Scientist for Climate Modeling at the Allen Institute for Artificial Intelligence, has reviewed other papers using similar methodology and thinks the authors’ approach was well done.
“Extreme rainfall and floods have huge impacts on people,” he said. “Overall we are confident that extreme precipitation will increase as the climate warms simply because warm air holds more water, but know very little about long term changes in a given region.” This study, Brenowitz adds, joins a growing body of machine learning work whose goal is not to make short-term predictions but rather to explain why extreme weather occurs. “The authors cleverly combine physical insight and deep learning to reveal the weather patterns driving extreme precipitation in the U.S. Midwest, a region with notoriously fickle weather. They find this pattern has become increasingly common in the observed record.”
Of course, it remains to be seen whether this technique could skillfully augment precipitation projections made by traditional climate models. But, says Brenowitz, “it is a promising start.”