Rainfall nowcasting based on neighboring rain gauge stations using learning machines

Abstract

Unpredicted rainfall, even mild, can cause severe disruption to daily schedules and businesses. Bangkok, Thailand, regularly experiences streets flooded after a spell of rain; therefore, rainfall nowcasting is highly desirable. The city has a high-density rain gauge network. This paper aims to explore the efficacy of using rain gauge data as the only source of prediction. We leverage state-of-the-art machine learning to identify rainfall patterns from short histories of rain at neighboring stations. Four different learning machines: Classification and Regression Tree (CART), Multilayer Perceptron (MLP), Random Forest (RF), and Support Vector Machine (SVM), are investigated. The results show that the nowcasting is feasible. From the experiments, the F1 performance at the 90-minute lead time, which is the critical time to decide future activities, is average at 0.69, 0.59, 0.73 and 0.63 when using CART, MLP, RF and SVM, respectively. We also found that the accuracy is higher when the lead time is shorter; a station has a shorter distance to its closest neighboring station; or a station is located farther away to the east or the north.