The Real Time Rainfall Forecasting and Landslide Early Warning with Precipitation from Correcting Weather Radar Reflectivity Using Machine Learning

Abstract

To reduce the damage of landslide disasters and flooding caused by heavy rainfall the development of an accurate rainfall forecasting system is necessary This study uses radar reflectivity provided by Central Weather Bureau (CWB) in Taiwan combining with the information from 134 ground rainfall stations to correct and predict rainfall in future time steps The approach we used is machine learning including back propagation neural network (BPNN) and long short-term memory (LSTM) which is a special case of recurrent neural network (RNN) There are two purposes in this study one is to enhance the accuracy of radar estimated rainfall in potential landslide areas and preserve the advantages of high spatial and temporal resolution of radar reflectivity The other one is to predict rainfall accurately so that we can reduce the error of predicting disasters Results show that the root-mean-square error of corrected rainfall by BPNN can reduce at least 29 80% and at most 90 44% compares to traditional radar estimated rainfall Most of the correlation coefficients can increase the estimate accuracy to the range in between 0 80 and 0 97 When apply this corrected result to rainfall prediction which is processed by LSTM and it shows highly prediction accurate that when rainfall rate is under 10 mm/hr; however if the rainfall rate exceeds 20 mm/hr the deviation will be much large Through the experiments we also find that the most accurate and stable forecast is to make the prediction an hour before weather front reaches This rainfall prediction model could help in landslide analysis and also to construct a landslide early warning system to reduce losses caused by landslide disasters in Taiwan

Date of Award	2020
Original language	English
Supervisor	Ting-To Yu (Supervisor)