CPrecNet: Enhanced Nowcast of High‐Resolution Short‐Term Precipitation Using Deep Learning

Abstract Accurate short‐term precipitation nowcasting is essential for disaster prevention and water resource management. Traditional numerical weather prediction faces challenges in delivering high‐resolution nowcasts due to computational limitations. We presents CPrecNet, a deep learning model utilizing a Swin Transformer‐based architecture and high‐resolution radar data (500 m and 5 min) to improve nowcasting accuracy. Based on an observation that precipitation fields rapidly decorrelate in space, we divide the whole domain into smaller 128 × 128 km regions and train on adjacent areas, addressing network size and data scarcity issues. The model uses a residual add method to enhance rainfall prediction accuracy by focusing on the evolution of small‐scale components. Test results show that CPrecNet outperforms existing models such as Pysteps in root‐mean‐squared error (RMSE) and maintains competitive critical success index performance for various precipitation thresholds. These findings highlight the potential of deep learning based on physical considerations to enhance high‐resolution precipitation nowcasting for better decision‐making in weather‐sensitive applications.