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

2. Compare the performance of AI with other methods and its capabilities for predicting rock falls. This includes the precision of predictions, the applicability of the models and the time to produce the models. This project was a collaboration between different teams with those objectives. The groups involved were: CEREMA, GEOLITHE, BRGM, Université Savoie Mont Blanc. The first 3 groups mentioned had expertise on rock falls while the other was expert in artificial intelligence and data analysis. Hermann Courteille (USMB) [Courteille, 2021] produced the NN models and did a qualitative analysis of the data. I processed the initial data combining the data sets and eliminating superfluous information. I also analysed the cliff in order to describe it in a way that could be inserted on a data analysis and possibly a predictive model, additionally I created alternative predictive machine learning models, using decision trees and nearest neighbors, to compare with the deep leraning models created by Hermann Courteille. 2. Study case A choice of the study case needed to be made. A well documented site with years of data, and a good number of events is required. Initially 6 different locations were proposed: RN1 (la Réunion), Val d’Arly (between la Savoie and la Haute-Savoie), la Saulcette (Savoie), cliff from Saint-Eynard (Isère), Mont Faron (Var) and la Rochaille (Alpes de Hautes-Provence). The two first sites collect the data about the events that reached the road while the others monitor the cliff and observe when blocks break off. The RN1 (Route National 1) from the Réunion was finally chosen because of the numerous well documented rock falls that reach the road and the longest period of observation. It has also been studied previously by CEREMA and Geolithe and experts were able to help with the understanding of the cliff behavior. One of the difficulties of applying traditional methods to rock fall prediction is the rarity of the events. The RN1 is the location with the greatest intensity of rock falls, and this will provide a richer data set. 2.1. La Réunion La Réunion is a French island on the south east of Africa, as shown on Figure 3. This French department has a population of 850 000 and a surface area of 2 512 km² (“Découvrir La Réunion” 2020). It has a tropical climate with higher temperatures close to the coast and lower temperatures inland as the altitude increases Figure 4. On the coastal area mean temperatures vary from 30° in summer to 24° in winter. Temperatures, in this region, are always superior to freezing temperatures which indicates that freeze-thaw cycles will not be taken into account as a potential triggering factor. Rain is not equally distributed along the island as displayed by figure 5. The West has much higher volumes of rain. This is due to the volcanoes that block the clouds from continuing to the East. The topography of the island is strongly linked to volcanic activity as it was formed 8