If you wonder why we need a better way to predict hurricanes, just ask the people of Houston.
Authorities knew Hurricane Harvey was heading to south Texas, but forecasters couldn’t say precisely which areas would be hardest hit. So, most Houstonians stayed put. The consequences: more than 75 deaths, 30,000 people in shelters and tens of thousands who needed rescuing.
And Harvey was just the start. Irma, Jose, Maria, Nate and Ophelia — with more than five weeks to go, the 2017 Atlantic hurricane season has already been one of the worst on record. The year is the first since 1893 to see 10 storms in a row reach hurricane strength, and only the fourth in recorded hurricane history. Without knowing where the brunt of a powerful storm will strike, officials are often puzzled about where and when to evacuate.
Halfway around the world, a team of scientists in Korea is using GPU-accelerated deep learning to help keep people out of harm’s way.
“We can’t prevent natural disasters, but with the right information, we can minimize the risks,” said Minsu Joh, director of high-performance computing research at the Korea Institute of Science and Technology Information (KISTI).
To Evacuate, or Not to Evacuate?
Korea is a country beset by typhoons, which are the same fierce storms as hurricanes or tropical cyclones. The term used depends on geography — it’s a hurricane in the Atlantic and Northeast Pacific, a typhoon in the Northwest Pacific, and a cyclone in the South Pacific and Indian Ocean.
In addition to storms becoming more frequent, climate change is intensifying weather-related disasters like hurricanes, according NASA’s Earth Observatory. Some studies indicate that warming oceans will cause more intense typhoons.
Joh and the KISTI team are combining deep learning technology with more traditional forecasting methods — numerical weather models produced GPU-accelerated supercomputers — to ease the speed and accuracy of typhoon predictions. If the scientists can more precisely pinpoint a storm’s path and intensity, authorities would have the certainty they need to order an evacuation in time for people to get to a safe place.
It’s no small matter to evacuate a region of millions of people, so officials are hesitant to issue a what may turn out to be a false alarm. Moreover, evacuation can be more dangerous than staying in place. During Hurricane Rita in 2005, Houston’s evacuation turned deadly — dozens of people died enroute from heat stroke, accidents and a bus fire — and led to the worst gridlock in Houston history, with countless cars stuck in 100-mile-long traffic jams.
Better Predictions in Less Time
Today, meteorologists rely on numerical models to predict wind speed, precipitation, air pressure and other factors that indicate the path and intensity of a hurricane over its lifetime. Instead, the KISTI team used observed data from satellites and radars to train their two deep learning systems —GlobeNet, which predicts the track of a typhoon, and DeepRain, which predicts heavy precipitation.
The researchers used data from numerical models to train a third system, DeepTC, which predicts tropical cyclones.
“Although these three models are still experimental, so far we’ve boosted accuracy over existing methods,” said Sa-kwang Song, the lead scientist developing the KISTI deep learning systems.
The KISTI scientists trained their models using the Keras toolkit and TensorFlow deep learning frameworks with cuDNN running on the institute’s NVIDIA GPUs and also on our GPUs in the Amazon Web Services (AWS) cloud. They also used our GPUs in AWS for inference.
So far, the KISTI system can predict typhoons and their associated rainfall just one to two hours in advance. The team plans to increase that range to six hours next year, and eventually to three days, which could be a real life-saver.
Typhoon, Flood Predictions in Action
In Korea, their work will be used to predict flooding for the region around the Imjin River, home to 30 percent of the nation’s population. The Korean Meteorological Administration’s typhoon center is testing the KISTI system, and could deploy it as early as next year.
Although KISTI’s work was designed for Korea, the same methods could be used elsewhere.
“If we can get enough satellite and radar data, we could easily apply DeepRain and GlobeNet to North America,” Joh said.
In the image at the top of this post, NOAA’s GOES-15 satellite captures geocolor imagery of Hurricane Harvey on the verge of making landfall on the Texas coast.