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

Figure 37: Comparison of predictions with time of the operational model and the model created by bagging with weight 5 times greater to predicting an event during the first part of the year 2004. Conclusion This study consisted in studying a cliff and the rock falls that occurred on the road parallel to it in previous years.The first step was selecting which site to study. The RN1 from la Réunion was chosen due to its detailed data set with many events. The data needed to be acquired from different partners. When the data was transferred it could then be formatted and then analysed. It is important to do a simple analysis to have a good understanding of the situation and to spot any early tendencies and possible incorrect data. Subsequently predictive models can be created and finally analysed in order to draw conclusions. 3 different methods were used to create models, nearest neighbors, decision trees and NNs. For each one models were tuned in order to produce an optimal performance for predictions over the test data set. The results were summarized on confusion matrices and the models were compared between each other and with the operational system. This study has two goals. The first is to better understand the causes of rock falls. By creating predictive models based on past events it is possible to understand which are the most important factors and at what intensities the blocks will fall. The models created by this study indicate that the rain is capable of explaining a part of the events but not all of them. There is a 46