When a person listens to someone else's speech, they distinguish both the words and the intonation. If you remove either one, the meaning gets distorted. Weather forecasting works in a similar way: a satellite image contains both fine details, such as individual clouds, and large waves that span half a continent. Previously, neural networks were good at seeing either one or the other, but not both at once.

SFANet works like an attentive interpreter. First, it examines the image in parts, averaging neighboring patches so as not to get lost in details. Then it transforms the picture into another representation, where both large and small waves are visible, like in musical notation. Next, a special module compares the two 'translations' and decides which one to trust at each point.

Thanks to this approach, the model with 16.4 million parameters predicts precipitation more accurately than heavier competitors. For El Niño, the warm current in the Pacific Ocean that affects weather worldwide, the improvement is 4.7%. This means forecasts will become more reliable, and computational resources will be needed no more than before.