Weather forecasting relies on vast arrays of satellite images and time series, where both small convective cells and planetary circulations matter. Existing methods, such as convolutional and recurrent networks, capture either space or time well, but rarely both at once. Transformers promise a solution, but their attention is quadratic in the number of tokens and loses detail at high resolution. SFANet fits into the search for efficient architectures for spatiotemporal sequences, alongside attempts to use sparse attention and Fourier transforms. The difference here is the explicit combination of the two domains through a dedicated module, rather than a simple mixing of features.
The authors build a network consisting of an encoder, a predictor, and a decoder. The encoder uses two types of token mixers: a pooling mixer, which averages neighboring features without learnable parameters, and a spatial mixer, which performs a Fourier transform for global mixing. Then the spatial-frequency attention module computes the relationship between features from the two domains and merges them. A convolution-based predictor forecasts future features, and the decoder reconstructs frames. The model was trained on 4 A100 GPUs for 100 epochs with the AdamW optimizer and a cosine learning rate schedule.
On the SEVIR dataset for precipitation forecasting, SFANet achieves an MSE of 2.7308 and CSI-16 of 0.7887, outperforming all compared methods, including Earthformer. The model has 16.4 million parameters and 185 GFLOPS, making it lighter than many counterparts. On the ICAR-ENSO dataset for El Niño forecasting, the improvement is 4.7% in C-Nino3.4-M and 8.2% in C-Nino3.4-WM, with MSE reduced by 3.5%. Ablation studies show that both mixers complement each other: pooling is better for light precipitation, the spatial mixer for heavy precipitation, and their combination yields the maximum effect. The attention module with a transformer adds about another 2% to CSI-M.
This work demonstrates that separate processing of space and frequencies, followed by attention between them, is more effective than trying to mix everything in a single layer. This sets SFANet apart from neighboring approaches that rely either on pure transformers or on convolutions with recurrence. For meteorology, this result means that more accurate models can be built without a sharp increase in computational costs. The method is also transferable to other spatiotemporal tasks where global correlations are important.
The direction may evolve toward adaptive threshold selection for the spatial mixer and more refined handling of rare extreme events. It may be possible to incorporate physical constraints into the frequency domain to improve interpretability. Applying the architecture to other geophysical data, such as wind or wave forecasting, is also promising.
The results will affect operational forecasting of precipitation and climate indices, where speed and accuracy are crucial. The spatial-frequency attention module could be used in related fields, such as medical image analysis or video prediction.
The next step will be to address data imbalance to reduce underestimation of heavy precipitation. It is also worth investigating error accumulation in long-term forecasting and ways to suppress it.
In the field of weather forecasting, open problems remain in accurately predicting extreme events and accounting for rare occurrences. This work contributes to solving these problems by improving overall accuracy, but does not fully resolve them. The connection to unresolved issues lies in the need for a deeper understanding of the interaction between spatial and frequency characteristics of atmospheric processes.
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
The paper introduces SFANet, a weather forecasting model that processes data simultaneously in the spatial and frequency domains and links them through a spatial-frequency attention module. The model consists of an encoder with two types of token mixers (a pooling mixer and a Fourier-based spatial mixer), an attention module, a predictor, and a decoder. On the SEVIR and ICAR-ENSO datasets, the model outperforms existing approaches in precipitation and sea surface temperature forecasting accuracy at comparable computational costs.
Where its strength is
The strength of the work lies in the explicit separation of spatial and frequency feature processing followed by attention between them, which effectively captures both local details and global patterns without a significant increase in computational complexity.
Where it could go next
Worth trying
Why this may matter
Extending SFANet to long-term forecasting and other spatiotemporal tasks could broaden the method's applicability and contribute to the development of accurate yet computationally efficient models for a wide range of geophysical applications.
Nearby works in our archive
The archive is sparse around this work: the closest review is only 0.52, against a usual 0.64. It looks like few have drilled nearby.
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.