Application of SOM technique for extreme rainfall forecasting during the Spring season in the Northern Vietnam delta region

Abstract

Extreme rainfall events pose significant challenges to disaster management, particularly in the densely populated Northern Vietnam delta region. This study aims to improve the forecasting skills of extreme rainfall events through the use of Self-Organizing Maps (SOM) and K-means clustering. Atmospheric data from ERA5 was used to train the SOM, while global forecasting system (GFS) products were utilized to test extreme rainfall predictions. The threshold for extreme rainfall was defined as 22.3 mm/day, corresponding to 90 % of cumulative probability. Four synoptic patterns responsible for extreme rainfall during the spring season were classified using objective pattern recognition techniques from ERA5 data, including: (i) A low-pressure trough over Northern Vietnam combined with a low-level cyclonic circulation extending up to 1500 m; (ii) Cold air or an eastward-slanted cold air mass compressing a low-pressure trough, combined with a low-level cyclonic circulation up to 1500 m; (iii) Wind convergence from the surface to 5000 m; (iv) Cold air accompanied by a strong surface front, combined with a high-altitude subtropical jet stream (500 mb); (v) Tropical cyclones. The SOM method was applied to bias-correct GFS 1 - 3 day lead forecasts, showing improved forecast accuracy, especially when the mean distance index (md_j) was less than 2.5. These results demonstrate the potential of SOM in enhancing the quality of short-term extreme rainfall forecasts.