Flood forecasting with Machine Learning in a scarce data layout

Abstract

Flooding is one of the major natural disasters occurring in the world, with climate change increasing their occurrence and severity. Reliable flood forecasting models are needed to have better insurance in the emergency services’ actions. This work reflects the capacities of Machine Learning models to improve discharge prediction results from empirical lag and route models based on hourly measured water level at gauge stations on the Garonne River. With scarce flooding data (30000 points) over the last 15 years, several learning algorithms have been implemented to predict floods in Toulouse at a 6-hour lead time from upstream stations providing hourly observations. A Linear Regression, a Gradient Boosting Regressor (Machine Learning) and a MultiLayer Perceptron (Neural Network, Deep Learning) are compared, using various strategies for learning, validating and predicting. Preliminary results show that the various strategies score as well as the empirical lag and route model. Further improvements are being investigated regarding the constitution of learning and validation data bases. This paper highlights how AI algorithms allow to improve the reliability of flood forecasts and how the layout of the limited volume of data influences their performance.