Answers are grounded in this paper and cite their sources.
Precipitation nowcasting aims to predict rainfall intensity in the short term (0–2 h) based on historical observational data. Existing approaches often suffer from insufficient extraction of conditional information, inadequate capture of spatial-temporal dependences, and low adaptability to complex terrain regions. To address these challenges, we propose a U-Net-based diffusion transformer model, namely WaveCastNet, which consists of a multiscale enhancement (MSE) module and an efficient spatial-temporal dependence extraction (STDE) module. Specifically, feature representations of historical conditional information extracted by the MSE module simultaneously reflect the large-scale global patterns and small-scale local evolution. In addition, an efficient spatial-temporal mixture-of-head attention mechanism is introduced into the STDE module to thoroughly capture the temporal dynamic dependences and spatial correlations of precipitation data. To further enhance the ability of precipitation forecasting in complex mountainous scenarios, we construct the Southern Anhui Mountain Precipitation dataset. Extensive experiments are conducted on our presented dataset and two public datasets [storm event imagery repository (SEVIR) and Shanghai Radar]. Experimental results demonstrate that the proposed WaveCastNet model achieves state-of-the-art performance on Southern Anhui Mountain Precipitation and Shanghai Radar datasets, while showing very competitive results on the SEVIR dataset.
No takes yet. Share an insight, caveat, or question.
Zheng et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: