A CNN-ConvLSTM Deep Learning Model with Axial Attention for Radar Reflectivity-based Precipitation Nowcasting in Southern Vietnam

Abstract

Precipitation nowcasting plays a critical role in early hazard warning, especially for rapidly evolving convective systems. This study proposes a radar-based deep learning model that combines convolutional neural networks (CNN) for spatial feature extraction, a recurrent spatiotemporal module (ConvLSTM) for image sequences, and axial attention to capture spatiotemporal context. The experiments use reflectivity fields derived from the CAPPI radar product during the 2022 rainy season (April–October) at the Nha Be radar station. Training samples are constructed from six consecutive radar frames at 10-minute intervals (i.e., the preceding 60 minutes) to predict the target frame 10 minutes ahead. To support an operational 60-minute warning window, we additionally apply a recursive rollout scheme to generate forecasts at lead times of 10, 20, 30, 40, 50, and 60 minutes, and report lead-dependent skill. On the October 2022 test set, the model achieves high structural similarity at 10 minutes (SSIM ≈ 0.79) with strong threshold-based classification skill (CSI ≈ 0.70 at 0.5 mm/h and ≈ 0.42 at 10 mm/h). At 60 minutes, SSIM remains around 0.71 and CSI at 0.5 mm/h remains around 0.36, indicating the model still captures the large-scale evolution of rainfall despite degradation at heavy-rain thresholds. When independently tested on June 2023 data, performance remains stable at 10 minutes (SSIM ≈ 0.79; CSI ≈ 0.69 at 0.5 mm/h and ≈ 0.44 at 10 mm/h), and at 60 minutes SSIM remains around 0.68 with CSI at 0.5 mm/h around 0.35. Compared with baseline models trained and evaluated under the same data processing protocol and the same loss function, the proposed model consistently performs better, particularly at moderate-to-heavy rainfall thresholds. These results suggest the model can serve as a complementary module for operational precipitation nowcasting workflows.