A hackathon for flood map prediction from geospatial data with parsimonious machine learning models (Papers Track)
David Medernach (Capgemini); Cyril Lemaire (Capgemini); Eva Girousse (EDF); Julie Keisler (EDF); Julie Richon (Capgemini); Nicolas Brunel (ENSIIE)
Abstract
Flooding poses significant risks across various sectors in France. This paper presents the outcomes of a machine learning hackathon focused on predicting the extent of various types of floods by leveraging a combination of geospatial and climate data. A Convolutional Neural Network (CNN) emerged as the most effective model, achieving strong performance in predicting the temporal evolution of flood risk maps. The evaluation not only includes prediction accuracy but also incorporates robustness, frugality, and explainability, in line with the principles of trustworthy AI principles. A key feature of this challenge was the absence of streamflow data, allowing the models to predict floods in regions where such data is unavailable. This highlights the potential of machine learning to improve flood forecasting in data-scarce environments.