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

equal importance to predicting and not predicting, which led to models that were not as safe as the operational, meaning the risk of driving the road would decrease if this model was used. Two models (decision tree 1:3 and bagging) were created with an importance 3 times greater to correctly predicting rock falls rather than correctly predicting no rock falls and those seem to be a satisfactory trade off. Those models have similar yet slightly higher indices. Even though those models do have a higher precision than the operational they do close the road a greater amount of times. It could be argued that the risk of the operational model is already sufficiently low and that as it closes the road less than the other models it would be preferable to the other models. The 7 models created are different and are superior to the operational on some aspects, but all of them have similar accuracy and balanced accuracy. This indicates there is a part of the data that can not be explained by the rain by itself. Other variables could be included and could help to explain a part of those rock falls or they could simply be random and no other explicative variable could improve the performance. Models Accuracy balanced accuracy precision recall F1 Operational 0.84 0.60 0.35 0.27 0.30 Nearest neighbors 0.87 0.59 0.53 0.21 0.30 decision tree 0.86 0.59 0.48 0.22 0.30 decision tree 1:3 0.84 0.62 0.38 0.32 0.35 bagging 0.84 0.63 0.38 0.34 0.36 DNN 0.78 0.63 0.28 0.42 0.33 DNN 3-classes 0.74 0.61 0.24 0.44 0.31 CNN 0.80 0.65 0.32 0.44 0.37 Table 10: Combination of result of different models When considering the application of an AI model to the operation of the road and comparing how it would affect drivers it is important to consider the frequency of the closures and their durations. Will the closures be punctual or will they be continuous? As rock falls are correlated with the rain which is a seasonal event the events will also be seasonal and be concentrated on certain moments of the year. This means there will be some continuity as is the case for the operational system. Figure 37 illustrates this continuity during the first months of 2004 when there is the greatest concentration of events. 45