Scientists in meteorology have made progress in predicting weather for over a century using chalkboards, equations and powerful computers. However, despite all progress, they still face a major problem against a simple, yet vital element: humidity.
Humidity is an invisible fuel for hurricanes, floods and storms. The difference between a light rain and a summer storm that forces you to flee is. And, until now, satellites have not been able to hand us the details we need to sound the alarm before the sky opens.
A team from the University of Environmental and Life Sciences in Wroclaw, Poland may be able to help change this situation. In an article published in the magazine Navigation, researchers explain how deep learning can convert fuzzy images of the atmosphere into exact three-dimensional maps of humidity and reveal the hidden currents that shape local weather.
The secret to this is the use of a Generative Adversarial Network (GAN) - a well-known type of AI for converting blurry images into clear ones. Instead of movie stars or landscapes, scientists trained this network on global weather data and empowered it with NVIDIA graphics processors. The result is reading satellite's low-resolution inputs converted into high-resolution maps of humidity with less error.
This progress can have great implications. By feeding these high-resolution humidity maps into physical or artificial intelligence models, even more accurate, timely predictions can be made that can predict floods and storms before they occur. Communities at risk can gain valuable time.