ANN Based Rainfall Prediction—A Tool for Developing a Landslide Early Warning System

Abstract

The study area of this work, Nilgiris district, is located in the southern state of Tamil Nadu in India. It receives heavy rainfall during South-west and North-east monsoons. The laterite soil and the presence of a large number of cut slopes make the region a landslide prone area, highly susceptible to rainfall induced landslides. This paper proposes a reliable rainfall forecast mechanism using only temporal and spatial rainfall intensity data recorded at rain gauge stations located close to the landslide risk sections in Coonoor. Several artificial neural network (ANN) based rainfall forecasting models were developed to forecast rainfall one day in advance at Coonoor. Mean square error (MSE) and correlation coefficient (CC) are considered as the performance measures to compare the forecasting ability of the ANN models. Wavelet Elman model, which had all the input predictors, emerged as the best model. Time delay neural network (TDNN) resulted in high correlation coefficient when the number of input predictors was limited. Results prove that the proposed wavelet Elman network has a forecasting accuracy better than all other ANN models and is an appropriate network to choose when the number of input predictors increases. This paper also describes the procedure adapted to develop a novel landslide early warning system based on the rainfall predicted by the best performing model and the rainfall threshold that exists for the study area. The results demonstrate the successful generation of landslide early warning messages that coincide with the landslide incidences in Coonoor.