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

percentage of them that are not linked to the rain. The highest percentage of events that were correctly predicted was 44% by the NN models. As developed in the introduction, a combination of factors cause the rocks to detach and reach the road and this study focuses on one, the rain. The second objective is to 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. 3 different types of predictive models were compared to the operational one, and the performances (Table 10) were similar. Some models were safer and some were less restrictive, but with the same input we can affirm that both provide similar performances. It is also important to consider the applicability of the models. The operational model that was adopted after 2004 was simple, making it easy to apply. The other models are more complex. It is harder for an operator to manually predict if the road will be closed or not, and even harder for the driver to do the same. The applicability of the model would also be harder and a greater inconvenience for drivers if the road was constantly opening and closing. As discussed in section 5.6. the openings and closures of the road with the different models are similar. One of the greatest improvements of the AI over the operational model is the time needed to produce a model. Both methods need an initial arduous task of preparing the data, but after it is concluded the process of creating an effective model is reasonably simple and quick to a data scientist. The time to adapt the model, created by artificial intelligence, to be operational is yet to be analysed. This study presented positive results that indicate that AI could be used in operational scenarios in order to optimize road control. The models could produce results at least as performant as the one that was adopted. It is quicker to create the AI model saving time. Even though the results are promising there are still limitations to this work and areas that could be expanded. As the different hyper-parameters were tuned for the different models, the goal was to maximize the balanced accuracy of the test set. This means they were tuned specially to the studied test set. In order to avoid this, a validation set is used. The model is trained using the train set, the hyperparameters are tuned to the test set and a validation set is used to verify the performance of the model to an unseen data set. This study focused mainly on the period between 2000 and 2007. Studying the period after 2008 would give us an alternative data set to test the performance of the different methods. As it has fewer events it would help analysing the capability of the model for data sets with a greater imbalance of classes. As explained in section 5.6., the rain is not capable of explaining the totality of the events. Studies including other possible explicative variables could help explain another part of the rock falls. During the discussions the applicability of the AI model is taken into account. It wouldn’t be ideal to apply the model directly. A future study would be needed to modify the model in a way it becomes easily applied to control the road circulation. 47