RainHCNet: Hybrid High-Low Frequency and Cross-Scale Network for Precipitation Nowcasting

Abstract

Precipitation nowcasting, particularly predicting heavy rainfall, is a critical aspect of meteorological forecasting. Recent advancements in deep learning have led to the development of radar echo extrapolation methods. However, most convolutional neural network-based methods focus primarily on high-frequency information, neglecting essential low-frequency cues necessary for forecasting high-intensity rainfall. Although some methods introduce attention mechanisms to improve predictions, they often encounter computational challenges and suffer from information loss related to rainfall. To address these limitations, we propose RainHCNet, a streamlined novel precipitation nowcasting method built on the UNet architecture. RainHCNet incorporates a hybrid channel–spatial attention mechanism to effectively capture low-frequency information, overcoming the limitations of traditional CNN-based methods that are unable to model global dependencies. In addition, a cross-scale supervision module integrates multiscale features from both deep and shallow layers to mitigate information loss. Moreover, a dynamic adjustment strategy for loss weights is employed, focusing on low-frequency information and samples linked to high-intensity rainfall events. We present two variants of the proposed architecture: RainHCNet (6.78 M) and RainHCNet^$\dag$ (0.35 M), the latter being a lightweight version suitable for computation and memory-constrained environments. Extensive experiments on the KNMI, Shanghai, and SEVIR datasets demonstrate that both models outperform state-of-the-art methods, particularly in predicting high-intensity rainfall events.