TSPF-GAN: A lightweight temporal-spatial-pixel feature fusion GAN for radar-based precipitation nowcasting

Haiping Zhang, Benze Zhang, Baitian Liu, Dongjing Wang, Liming Guan, Chonglei Zhu, Zeng Zhang

Abstract: Precipitation nowcasting is a critical spatiotemporal prediction task that forecasts radar echo sequences based on current observational data. Accurate and reliable forecasting can significantly mitigate the risks associated with sudden precipitation events. Although existing methods have achieved progress in this domain, they generally suffer from two major issues: (1) insufficient sensitivity to heavy precipitation regions due to inadequate sample size and imbalanced distribution; (2) high computational and hardware resource demands caused by large model sizes. To address these challenges, we propose TSPF-GAN, a lightweight generative adversarial network (GAN) based on temporal-spatio-pixel feature fusion. We introduce three key innovations: (1) the Temporal-Spatial-Pixel Fusion Module (TSPFM), which employs a threshold-guided mechanism and multi-scale feature fusion to significantly enhance perception of heavy precipitation regions; (2) a lightweight GAN architecture with only 4.5 million parameters, substantially reducing hardware resource requirements; (3) a mixed loss function enhanced for heavy precipitation regions to optimize training and improve prediction accuracy. Evaluated on a real precipitation radar dataset from the Netherlands, TSPF-GAN demonstrates performance improvements across various precipitation thresholds compared to other models. Furthermore, in comparisons with diffusion-based methods on the Shanghai radar dataset, TSPF-GAN maintains comparable prediction performance while reducing parameters by 90%.

External IDs:dblp:journals/esi/ZhangZLWGZZ26