Enhancing Flood Prediction using Ensemble and Deep Learning Techniques

Though flooding is seen as a common environmental threat globally, it has dramatically increased recently due to climate change, impacting underdeveloped and developing countries dangerously. For example, in most developing countries like Ghana, flooding has affected over four million people in terms of property damage, loss of lives, income and spread of diseases, resulting in economic harm beyond USD780 million. At least one major flood disaster does occur yearly. The recurring incidences of flooding and associated calamitous socio-economic risks and anticipated increase of its prevalence soon in cities of developing countries such as Ghana have necessitated an intelligence system to offer efficient and early warning of its occurrence. In this study, we explore the potential of the machine learning (ML) computing paradigm to propose a flooding prediction model. Specifically, four state-of-the-art ML algorithms, namely long short-term memory (LSTM), extreme gradient boosting (XGBoost), random forest (RF) and extremely randomised trees (Extra Trees), are used to implement four different flood prediction models. We measure the performance of our developed models with multiple statistical performance evaluators. The experimental results show the potential of the developed models for efficient and effective prediction of flooding. The merit of this study lies in the fact that it is the first to the best of our knowledge to use a combination of environmental factors from Ghana and machine learning algorithms to develop intelligent flood models to help stakeholders make informed decisions.