Flood Prediction and Mining Influential Spatial Features on Future Flood with Causal Discovery

The development of accurate flood prediction model could reduce number of fatalities. In this paper, water level time series, spatio-temporal precipitation and hydrological data are used for flood prediction. Since our data is high dimensional and not all features are correlated to flood, our proposed algorithm is designed to find influential spatial features, or features at locations which are highly correlated to flood. With the idea that true causes, or highly correlated features to flood, should give accurate information about flood, our proposed flood prediction algorithm is based on Bayesian based causal discovery. The purpose of this paper is twofold. Firstly, we propose a new causal discovery algorithm which is Bayesian-based approach with an optimization function for maximizing mutual information. Secondly, the proposed algorithm is applied to real-world precipitation and hydrological data to find influential spatial features on future flood in North Texas area. Flood prediction models can then be learned from selected features. Experiments on synthetic data confirm that our proposed algorithm is more accurate in finding true causal relationships than two competitors, Group Lasso and Markov2P. Experiments on flood predictions show that our approach has the best accuracy in almost all six lead time predictions. The accuracy and result visualizations also suggest that our proposed algorithm can find influential features to flood.