With hurricane season starting, I spent my time on stage at the National Geographic Explorers Festival showing off NASA’s Deep Learning Hurricane Intensity Estimator. The AI works insanely fast. Last year the neural networks caught a huge uptick in Florence wind intensity — from a Category 2 to Category 3 — a full day before the National Hurricane Center advisory came out. That’s the difference between limbs snapping off trees and trees being completely uprooted.
Traditionally NOAA would classify storms every 6 hours. Meteorologists determine the strength of a hurricane by looking at the shapes of the clouds, the infrared imagery, then estimate windspeed every 6 hours. NOAA does this for every storm, every 6 hours. Over the last 50 years those meteorologist have every image with every storm classified. This massive dataset makes for an amazing AI training set of data that is scientifically validated, which Development Seed and NASA used to train their neural networks.
The AI processes imagery from the new NOAA GOES satellites launched in 2016 and 2018, which take a pictures of the western hemisphere every 5 minutes. Tack on an extra minute to downlink from the satellite, process the image, run it through the AI, and push a new forecast to the API. By using neural networks that compare an image to thousands of past, known observations, the Deep Learning Hurricane Intensity Estimator forecasts storm windspeed in 6 minutes, versus 6 hours.