RINA - Démonstrateur d’utilisation de l’intelligence artificielle pour la gestion opérationnelle des risques naturels gravitaires géologiques | FEREC 2020

Chapter 3 Experiments Here, we present different models tested on Réunion Data. As we have sparse data, we just investigate models at temporal level. We first present operational rules used during period 2000- 2020. Those expert rules applied on data spanning over the same period, give us a baseline for comparing 3 types of models with different variations: • Neural Networks : Dense ones (DNNs) or Convolutional ones (CNNs) • Regression models : General Linear Model (GLM) In all cases, targets are the daily number of rockfall NB t and inputs are previous rain quantities R t (and sometimes previous number of rockfalls) . 3.1 Experiments settings 3.1.1 Train/validation settings As events are more frequent before protection works, we focus on the first part of data from 2000 to 2007 (Table 3.1) . Split Train/Val: For this period, rockfall data are absent during year 2005. So we split data into train/val set around this year (See Table 3.1 for precision). Such split is permit because rockfall event are quite stationary in time. 3.1.2 Grouping label for classification As data are sparse, it would be very hard to predict exactly the number of rockfall at day t given the history. In first attempt, we group labels (which span from 0 to K max over the whole history) for classification task: • 2 classes with ’0’ no event, ’1’ one or more event: a simple binary model that math exactly expert rules 25