Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SFANet: Spatial-Frequency Attention Network for Weather Forecasting

Jiaze Wang, Hao Chen, Hongcan Xu, Jinpeng Li, Bowen Wang, Kun Shao, Furui Liu, Huaxi Chen, Guangyong Chen, Pheng-Ann Heng

TL;DR

SFANet tackles the challenge of accurate spatiotemporal weather forecasting in high-resolution data. It combines pooling-based token mixing, FFT-based spatial mixing, and a spatial-frequency attention module within an encoder-predictor-decoder architecture to model long-range and cross-modal dependencies. The method yields state-of-the-art results on SEVIR for precipitation nowcasting and on ICAR-ENSO for ENSO prediction, with clear gains over strong baselines across CSI and MSE metrics. This work offers a scalable, efficient framework that can improve forecast accuracy and risk-informed decision making in weather-sensitive applications.

Abstract

Weather forecasting plays a critical role in various sectors, driving decision-making and risk management. However, traditional methods often struggle to capture the complex dynamics of meteorological systems, particularly in the presence of high-resolution data. In this paper, we propose the Spatial-Frequency Attention Network (SFANet), a novel deep learning framework designed to address these challenges and enhance the accuracy of spatiotemporal weather prediction. Drawing inspiration from the limitations of existing methodologies, we present an innovative approach that seamlessly integrates advanced token mixing and attention mechanisms. By leveraging both pooling and spatial mixing strategies, SFANet optimizes the processing of high-dimensional spatiotemporal sequences, preserving inter-component relational information and modeling extensive long-range relationships. To further enhance feature integration, we introduce a novel spatial-frequency attention module, enabling the model to capture intricate cross-modal correlations. Our extensive experimental evaluation on two distinct datasets, the Storm EVent ImageRy (SEVIR) and the Institute for Climate and Application Research (ICAR) - El Niño Southern Oscillation (ENSO) dataset, demonstrates the remarkable performance of SFANet. Notably, SFANet achieves substantial advancements over state-of-the-art methods, showcasing its proficiency in forecasting precipitation patterns and predicting El Niño events.

SFANet: Spatial-Frequency Attention Network for Weather Forecasting

TL;DR

SFANet tackles the challenge of accurate spatiotemporal weather forecasting in high-resolution data. It combines pooling-based token mixing, FFT-based spatial mixing, and a spatial-frequency attention module within an encoder-predictor-decoder architecture to model long-range and cross-modal dependencies. The method yields state-of-the-art results on SEVIR for precipitation nowcasting and on ICAR-ENSO for ENSO prediction, with clear gains over strong baselines across CSI and MSE metrics. This work offers a scalable, efficient framework that can improve forecast accuracy and risk-informed decision making in weather-sensitive applications.

Abstract

Weather forecasting plays a critical role in various sectors, driving decision-making and risk management. However, traditional methods often struggle to capture the complex dynamics of meteorological systems, particularly in the presence of high-resolution data. In this paper, we propose the Spatial-Frequency Attention Network (SFANet), a novel deep learning framework designed to address these challenges and enhance the accuracy of spatiotemporal weather prediction. Drawing inspiration from the limitations of existing methodologies, we present an innovative approach that seamlessly integrates advanced token mixing and attention mechanisms. By leveraging both pooling and spatial mixing strategies, SFANet optimizes the processing of high-dimensional spatiotemporal sequences, preserving inter-component relational information and modeling extensive long-range relationships. To further enhance feature integration, we introduce a novel spatial-frequency attention module, enabling the model to capture intricate cross-modal correlations. Our extensive experimental evaluation on two distinct datasets, the Storm EVent ImageRy (SEVIR) and the Institute for Climate and Application Research (ICAR) - El Niño Southern Oscillation (ENSO) dataset, demonstrates the remarkable performance of SFANet. Notably, SFANet achieves substantial advancements over state-of-the-art methods, showcasing its proficiency in forecasting precipitation patterns and predicting El Niño events.
Paper Structure (12 sections, 3 equations, 3 figures, 3 tables)

This paper contains 12 sections, 3 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Method
  4. Preliminary
  5. Token Mixer
  6. Model architecture
  7. Experiments
  8. Experimental Setup
  9. Results
  10. Ablation Study
  11. Qualitative Analysis.
  12. Conclusion

Figures (3)

  • Figure 1: Illustration of the SFANet architecture. (a) Overview of the network. We adopt the encoder-predictor-decoder architecture for weather forecasting. The encoder consists of a 2D CNN for downsampling, $L$ pooling mixer blocks to generate features at the spatial domain, $L$ spatial mixer blocks to generate features at the frequency domain, and a spatial-frequency attention module to capture the correlations between spatial and frequency features. (b) The process of our spatial mixer. (i) Spatial token mixing via fast Fourier transform. (ii) Adaptive Fourier Neural Operator to mix in the Fourier domain. (iii) Inverse FFT for token demixing.
  • Figure 2: Prediction results of SFANet on the SEVIR dataset. Illustrated via vertically integrated liquid water contents, quantified on a scale ranging from 0 to 255 shown in the color bar.
  • Figure 3: Error Analysis of SFANet on the SEVIR dataset. For each pixel in prediction results (Predict) and ground truth (GT), we will assign it a label according to the thresholds ([16, 74, 133, 160, 181, 219]). Hits: $Predict=GT$. Miss: $Predict<GT$. False Alarm: $Predict>GT$.