Development of Radar-Based Nowcasting for Intense Precipitation in the Tropics

Tropical convective thunderstorms in Singapore are known to be localised, associated with rapid growth and decay, and are often accompanied by weak environmental winds, especially during inter-monsoon seasons. This frequently leads to thunderstorms back-building, a phenomenon that exceeds the capabilities of most radar-based nowcasting systems. Previous work has shown that current nowcasting techniques such as optical-flow extrapolation and pre-trained machine learning (ML) methods, among others, struggle to beat a persistence forecast at lead-times of 30- to 60-minutes. In this study, we propose developing and training our own ML method for the tropics. We selected the RainNet model (Ayzel et al., 2020) as a benchmark, which is based on U-net architecture, to improve the nowcasting of convective thunderstorms in Singapore as a starting point. Unlike traditional approaches that use only regression loss functions, our methodology incorporates spatial verification techniques, such as structure, amplitude, and location, into the loss function to train the model.
The model was trained using 5-minute precipitation estimates from the C- and S-band radar composite operated by Meteorological Services Singapore. Using the last 30 minutes of rain rate images the next frame (T+5m) is predicted, the predicted frame is then added to the existing stack of images to create a continuous rolling prediction that extends up to a 90-minute lead-time. A 3-year dataset (176820 sample sets) was available for training and a 1-year dataset (65936 sample sets) was used for validation within the training process. A further independent 1-year dataset was used to run contingency table and neighbourhood forecast verification scores on for further testing. Several experiments were setup by adding various combinations of regression and spatial errors scores within the loss function during the training process.
Preliminary results are promising. Upon examination of the forecast output, it appears that the complex motion and interaction between storms, such as merging and splitting, and growth and decay are captured, which is an improvement over most traditional nowcasting methods. However, thunderstorm initiation is not well captured, which limits the effective forecast lead-time within the tropics based solely on past radar imagery. A more objective analysis is underway and will be presented along with plans on future development to move beyond the RainNet model to more sophisticated physics-informed based approaches.