Flood Modelling using Long Short-Term Memory Algorithm with Synthetic Minority Oversampling Technique

Abstract

Flood disasters have occurred quite frequently in Malaysia and has been considered one of the most dangerous natural disasters. Meteorologists and hydrologist are having difficulty in predicting flooding due to the weather changes and are opting to deep learning techniques to predict flooding. This study compares the performance of deep learning techniques for flood prediction, namely Long Short-Term Memory (LSTM) network. Furthermore, the effect of using Synthetic Minority Oversampling Technique (SMOTE) as a ‘treatment’ to treat imbalanced data was also investigated to ensure all LSTM models were able to predict accurate flooding. The experimental results revealed the treated dataset had a positive impact in predicting flood with higher accuracy. Additionally, to increase the accuracy of deep learning methods, future researchers could use more hidden layers as well as different hyperparameter settings which could help create a better predicting LSTM model.