SFNL-Former: A novel spatial-frequency non-local network for precipitation forecast

Abstract

Accurate, high-resolution precipitation forecast is vital for urban disaster mitigation and agriculture, underpinning fine-grained social management. Furthermore, as the primary input to the terrestrial hydrological cycle, its accurate quantification is essential for sustainable water resource management and efficient circulation. However, this task remains highly challenging due to the complex spatiotemporal and non-linear dynamical evolution inherent in atmospheric processes. Deep learning has been extensively applied to weather forecast through CNN-RNN or Transformer architectures, yet these methods often struggle to capture global meteorological motions while simultaneously preserving the fine-grained details of localized strong convection. Furthermore, they are prone to producing blurred predictions or exhibiting high false alarm rates when encountering extreme precipitation events characterized by long-tailed distributions. To address these issues, we propose the Spatial-Frequency and Non-Local Former Network (SFNL-Former), which accurately predicts complex precipitation evolution via explicit spatial-frequency domain feature separation and efficient long-range spatiotemporal modeling. A Spatial-Frequency Domain Attention (SFDA) encoder is designed to separately extract low-frequency global backgrounds and high-frequency local textures of precipitation fields. Furthermore, the Locality-Sensitive Hashing (LSH) based Non-Local Sparse-Aware Transformer (NLSAT) is introduced to capture long-range spatiotemporal dependencies with linear complexity. Experiments on real-world datasets demonstrate that SFNL-Former outperforms existing baseline methods across multiple metrics and spatial scales. Specifically, under challenging high-intensity precipitation thresholds, the model improves the Critical Success Index and substantially reduces the False Alarm Rate, while maintaining optimal prediction coherence. Consequently, the framework enables critical decision support for proactive disaster response, fostering more resilient social management systems under future extreme weather scenarios.