A translator at negotiations picks up not only the words but also the tone in which they are spoken. If they focus only on vocabulary, they might miss sarcasm or fatigue in the voice. If only on intonation, they could mistranslate the details of the contract. A similar story unfolds in weather forecasting: satellite images carry both local details, like thunderstorm cells, and global waves that stretch across the entire ocean.
Typical neural networks usually pick one or the other. Some examine the image piece by piece and see small clouds well, but they don't sense that a large cyclone is looming over the continent. Others look at the whole map at once, but then individual downpours blur into a murky patch. SFANet works differently: it processes the data twice, in two different 'languages'.
Then a spatial-frequency attention module comes into play. It compares the two 'translations' and decides which information to trust at each location in the forecast. For example, for light drizzling rain, the averaged picture matters more, while for a thunderstorm front, sharp gradients are key. This approach allows the model with 16.4 million parameters to outperform heavier competitors: on the SEVIR dataset, the MSE error drops to 2.7308, and the CSI-16 metric reaches 0.7887.
For El Niño forecasting, the warm current that every few years alters weather across the planet, the improvement is 4.7% on the key metric. That is significant, given that crop yields and water supplies depend on the accuracy of such forecasts. Moreover, the model was trained on just four A100 accelerators, making it accessible to operational weather services.
To return to the translator: they don't just listen to two channels; they understand when one matters more than the other. SFANet does the same with precipitation and temperature maps. Previously, weather models were like translators who either closed their eyes and listened intently to intonation, or plugged their ears and read lips. Now there is a way to hear and see simultaneously, and that is already changing forecast quality.
The model uses the Fourier transform, which was invented by Jean-Baptiste Joseph Fourier in 1822 to describe heat propagation. Today it helps predict rain.
In Stanisław Lem's novel 'Solaris', the planet's ocean exhibits complex spatiotemporal dynamics that are difficult to predict. SFANet, by processing data in two domains, vaguely resembles an attempt to understand such systems through different representations.