| Abstract: |
| Extreme rainfall is one of the major natural disasters in Taiwan and it made significant environmental and societal impacts. The occurrence of rainfall extremes involves complicated multiscale interactions and hence the ability in Quantitative Precipitation Estimations(QPEs) / Quantitative Precipitation Forecasts (QPFs) is limited. In the past, we used the numerical model which based by the physical concepts to predict the rainfall amount. Limited by model resolution and traditional theoretical architectures, many simplifying assumptions and uncertainty in convection simulation lead to large prediction errors. By incorporating machine learning(ML) techniques, a new approach of multi-time scales QPEs and QPFs is proposed. The designed framework is based on our prior work on classifying weather events and was proved to outperform traditional objective analysis. For QPEs/QPFs tasks, we used objective weather events data, real-time surface/radar observations, and numerical model output as dependent variables. By combining the machine learning techniques and multiple data sources, we developed an extreme rainfall warning system. The preliminary result of using ML method to detect the heavy rainfall events in Taiwan shows the hit rate is about 79%-82% for different classifiers of all positive events. |
|