A Patch-Wise Mechanism for Enhancing Sparse Radar Echo Extrapolation in Precipitation Nowcasting

Abstract

Precipitation nowcasting is pivotal for disaster prevention, urban planning, and various societal applications. Although deep learning-based radar echo extrapolation methods are widely adopted, their effectiveness is limited by challenges in handling sparse echo sequences with low pixel proportions. This limitation hinders accurate predictions of low-frequency heavy rainfall and localized precipitation events. To address these issues, a novel Patch-wise (PW) mechanism is proposed in this study. Specifically, radar echo frames are divided into patches, which are then encoded into sequences and processed using a local attention mechanism to enhance critical regional information extraction. Furthermore, multiscale convolutions tailored to the patch scale are employed to expand the receptive field, and a convolutional block attention module is introduced to capture the spatiotemporal dynamics of sparse echoes and intense rainfall. These improvements enable extraction of sparse spatiotemporal features in a PW manner, enhancing prediction accuracy for regional sparse precipitation. Experiments across Chinese regions demonstrate the effectiveness of the PW mechanism. Notably, integrating the PW into PredRNN model yields improvements in the critical success index by 1.24%, 1.37%, and 12.59% for precipitation intensity thresholds of 20, 30, and 45 dBZ, respectively. Spatial visualizations of radar echoes reveal PW's superior in predicting localized and intense precipitation. This study is expected to advance regional precipitation nowcasting by offering both practical insights and methodological innovations.