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

● The daily rain in the previous day is superior to 15 mm ● The daily rain in one of the previous 3 days is superior to 30 mm This can be simplified as follows: P 1≥1 5 ou P 2≥3 0 ou P 3≥3 0. With this model we can observe, for each day, the rain of the 3 previous days and know if the model would predict that there are or are not rocks falling on the road. A confusion matrix is built for the test set as shown by Figure 24. Out of the 1085 days of the test data set, the model predicted that 104 (in table) days would have an event and about ⅓ of those predictions were correct. Actual 0 1 Predicted 0 872 104 1 71 38 Figure 24: Confusion matrix for the operational model for the test data set. accuracy 0.84 balanced accuracy 0.60 precision 0.35 recall 0.27 F1 0.30 Table 2: Indices calculated based on the confusion matrix to quantify performance of the operational model. For this confusion matrix all the indices described in the previous section were calculated and are presented on table 2. We can observe that the accuracy is quite high, which is expected as the data is not balanced. During 87% of the days there are no rocks falling along the road. This means that even a dummy classifier could have a better score than the operational one. On the other hand, we know that it has its value as it is capable of correctly predicting 38 events. 5.2. Nearest Neighbors The first and simplest model created is nearest neighbors. This method predicts if there will be rocks falling along the road based on similar days in the past. In order to produce the best model the number of neighbors, that will be taken into account in order to predict new data points, will be optimized. This is done by varying this number and observing the result of the 33