A Deep Learning Nowcasting Model for Convective Cell Occurrence in Taiwan

Abstract

Afternoon thunderstorms, mesoscale convective systems, and other short-duration rainfall events threaten property and transportation. Recent deep learning techniques have been proven effective in nowcasting for rainfall accumulation (rain maps), but predicting occurrences of intense convective cells can add additional value to decision-making procedures. This study develops a deep-learning model that predicts the locations of cell occurrences in the next 60 minutes. The training data include reflectivities from the Taiwanese radar network and convective cell trajectories from the System for Convection Analysis and Nowcasting (SCAN). The label is the SCAN cell occurrence (1 or 0) within a 7.5 × 7.5 km² area in the next hour. In addition to providing occurrence probabilities, the post-analysis procedure deploys a threshold mask to convert the probabilistic forecast into deterministic forecasts; it achieves a ∼40% improvement in the critical success index compared with the baseline method. Furthermore, the new model informs users about the risks under the chosen threshold selected based on their risk tolerance. This study provides proof of concept that replacing the predicting objectives (“cell occurrence” instead of “rainfall”) of the model may help forecasters' decisions and the integration of deep learning into operational forecasting.