Comparison of random forests and support vector machine for real-time radar-derived rainfall forecasting

Pao Shan Yu, Tao Chang Yang, Szu Yin Chen, Chen Min Kuo, Hung Wei Tseng

Research output: Contribution to journalArticlepeer-review

65 Citations (Scopus)

Abstract

This study aims to compare two machine learning techniques, random forests (RF) and support vector machine (SVM), for real-time radar-derived rainfall forecasting. The real-time radar-derived rainfall forecasting models use the present grid-based radar-derived rainfall as the output variable and use antecedent grid-based radar-derived rainfall, grid position (longitude and latitude) and elevation as the input variables to forecast 1- to 3-h ahead rainfalls for all grids in a catchment. Grid-based radar-derived rainfalls of six typhoon events during 2012–2015 in three reservoir catchments of Taiwan are collected for model training and verifying. Two kinds of forecasting models are constructed and compared, which are single-mode forecasting model (SMFM) and multiple-mode forecasting model (MMFM) based on RF and SVM. The SMFM uses the same model for 1- to 3-h ahead rainfall forecasting; the MMFM uses three different models for 1- to 3-h ahead forecasting. According to forecasting performances, it reveals that the SMFMs give better performances than MMFMs and both SVM-based and RF-based SMFMs show satisfactory performances for 1-h ahead forecasting. However, for 2- and 3-h ahead forecasting, it is found that the RF-based SMFM underestimates the observed radar-derived rainfalls in most cases and the SVM-based SMFM can give better performances than RF-based SMFM.

Original languageEnglish
Pages (from-to)92-104
Number of pages13
JournalJournal of Hydrology
Volume552
DOIs
Publication statusPublished - 2017 Sep

All Science Journal Classification (ASJC) codes

  • Water Science and Technology

Fingerprint Dive into the research topics of 'Comparison of random forests and support vector machine for real-time radar-derived rainfall forecasting'. Together they form a unique fingerprint.

Cite this