♫ I don’t care what the weatherman says when the neural network says it’s hailing
David Gagne, a postdoctoral researcher at the US National Center for Atmospheric Research (NCAR), developed a simple convolutional neural network model to forecast the chances of hailstorms.
In the last decade, severe storms caused about $14bn worth of damage and killed 94 people per year, Gagne said during a presentation at the GPU Technology Conference in San Jose, California.
Meteorologists begin warning people of severe weather conditions the day before a hail event, but it’s difficult to be precise. So Gagne wanted to see if deep learning could accurately identify the weather patterns leading up to hailstorms and reduce false alarms.
He fed the neural network a series of images showing storm activity over the US taken from NCAR’s Real Time Ensemble Forecasts. Details about the temperature, wind conditions, atmospheric pressure levels were added to create a profile for each storm.
Hail starts off as graupel – small icy particles that act as embryos for hail – Gagne told The Register. As they whirl around in a storm, they collect small droplets of water that freeze on its surface. Over time, they grow in size to become solid ice lumps that fall as hail. If hail reaches three inches in diameter, it can crack windscreens and leave walls and roofs pockmarked.
He trained the neural network to learn the features associated with storms in the images in the training data. Since the hail process requires the presence of a storm, the model was trained to look out for properties such as wind shear, a change of wind speed along different directions and height – a good predictor that a storm is brewing.
The goal was to highlight the same storm regions on radar images that a standard statistical model would pick out when monitoring the risk.
After training on 82,000 different storm profiles, Gagne tested the neural network on 32,000 storms and saw it was about 88 per cent accurate.
That’s a pretty decent result, considering the model is only three layers deep compared to state-of-the-art image classification models that have hundreds of layers. It took 15 hours to train the model on eight of Nvidia’s Tesla K40 GPU chips, an older model launched in 2013.
The model is a proof of concept and not complex enough to be used in real weather warning systems yet. It’ll be a while before deep learning takes over in meteorology since it requires better algorithms and a change of infrastructure to invest in more GPU supercomputer clusters. But more meteorologists are now interested in AI and deep learning than ever before, Gagne said. ®