SimNowcasting: Enhancing Precipitation Nowcasting via Lightweight Multiscale Convolution and Channel-Wise Attention Mechanism

Precipitation nowcasting refers to the prediction of precipitation within 0-2 hours. Extreme precipitation events have been occurring frequently, therefore, precipitation nowcasting is significantly important. Existing deep learning-based Precipitation nowcasting methods usually use complexity network structures, resulting into low computational efficiency and insufficient generalization. We introduce SimNowcasting, a straightforward yet highly effective model for spatiotemporal sequence prediction. By developing the lightweight SimInception module, SimNowcasting extracts multi-scale features from spatiotemporal data with fewer training parameters than standard Inception-based approaches. This approach shows improved generalization on the evaluated dataset. Furthermore, Furthermore, the integration of the Spatial Encoder (SE) module allows for the precise capture of critical features within the channel dimension of radar echo sequences, thereby heightening the network’s sensitivity to key features. Experimental results on Shanghai radar data indicate that SimNowcasting achieves competitive performance in forecast accuracy compared to baseline methods, suggesting potential for practical precipitation nowcasting applications.