To build Dark Sky, the duo rely on free online data from the National Oceanic and Atmospheric Administration, which operates more than 140 radar stations across the United States. They also make use of data from the National Weather Service, which boils down that information into color-coded weather images that are easier to work with.
To generate the local forecasts, they clean up those images to filter out noise like movement from migrating birds. Then they use computer vision algorithms that can create a map of velocity based on the radar data. And they use those maps to try to predict which way a storm is moving and how it will behave in the near future.
Mr. Grossman said that in trial runs, the application worked with consistent accuracy.
“It’s tricky to quantify how good weather prediction is,” he said. “But I was on vacation in Ohio last summer watching a storm roll in over Lake Erie, and opened the app. It said it would start raining in 17 minutes.”
Mr. Grossman said he pulled out his watch and waited. Sixteen minutes and 45 seconds later, he said. “It started raining.”