Comparative Presentation of Machine Learning Algorithms in Flood Prediction Using Spatio-Temporal Data

Abstract

The development of accurate flood prediction model could reduce number of fatalities by assisting local government in decision making. In this paper, three well-known machine learning algorithms, including Support Vector Machine, Decision Tree, and Lasso, are compared in terms of flood prediction accuracy. The selected algorithms are applied to learn flood prediction models for six U.S. Geological Survey gauges in North Texas. Three data sets from different sources had been used to learn flood prediction models. The data sets include Water level time series from gauges, spatio-temporal precipitation from National Weather Service’s Weather Surveillance Radar-1988 Doppler, and hydrological data from Hydrology Laboratory-Distributed Hydrologic Model. Although Support Vector Machine usually performs well in many applications, the results suggest that Lasso is the most appropriate for flood prediction while Support Vector Machine performs the worst.