Rainfall Prediction in Flood Prone Area Using Deep Learning Approach

Abstract

Flood is a catastrophic event that contributes to the impact on socio-economic of a developing country. Flood prone area is also known as the location at risk as heavy rainfall attribute to flood events. This circumstance leads to the effective flood mitigation phase. One of the critical problems facing by responsible government agencies is to minimize future uncertainties. This study explores four deep learning methods with univariate rainfall temporal data in gauging station near to flood prone area. Four models tested are Multi-layer Perceptron MLP, Long Short Term Memory LSTM, Stacked-LSTM, and hybrid model Convolutional Neural Network CNN-LSTM. The aim of this paper is to compare and determine the best method for rainfall prediction of next day event. Model comparison is conducted by comparing the correlation coefficient, Root Mean Square error (RMSE) and Mean Absolute Error (MAE). Based on the selected locations, the results showed at Kuantan station generally underfitting, meanwhile the Kuala Krai station does not showed discrepancies between training and testing dataset. It could be concluded that by adding the complexity to the model, will not significantly improved the model prediction. The LSTM model with 16 memory blocks was outperformed in both locations. The potential of deep learning methods should be considered for rainfall amount prediction due to less complexity and much easier to be applied into dataset for model fitting purposes. It may assist to accommodate strategic precautions for flood mitigation phases in flood prone areas.