<p>Precipitation nowcasting is a critical topic in meteorological research, holding significant importance for natural disaster warning, agricultural production, and urban drainage management. However, the sudden and localized nature of precipitation events poses a persistent challenge for achieving accurate forecasts. Advances in weather radar technology have enabled the provision of real-time datasets for data-driven deep learning models. While convolutional neural networks (CNNs) excel at extracting local features from radar echo images, their performance is often constrained by limited receptive fields, which restrict the capture of long-range spatial dependencies. To address this limitation, this study leverages Swin-Transformer’s expertise in global modeling to expand the effective receptive field of CNNs and capture long-distance feature interactions. We propose STAt-Former, a novel precipitation nowcasting model that integrates spatiotemporal channel attention and Transformer architecture to facilitate multi-scale spatiotemporal feature learning. The model employs dual coding pathways to separately extract local and global features from radar echo images. Additionally, a sophisticated feature fusion and decoding architecture is designed to achieve precise precipitation forecasting, thereby enhancing the capability to capture spatiotemporal correlations in hydrometeorological data analysis. Experimental validation using the Netherlands open precipitation dataset demonstrates that STAt-Former successfully achieves precipitation forecasts at 30, 60, 90, and 120-minute time scales. The results show that the proposed method effectively captures the deep spatiotemporal dynamics of radar echo sequences, demonstrates superior nowcasting performance, and outperforms baseline models in key prediction metrics (e.g., MSE, accuracy).</p>