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

falls it destabilizes the bedrock. The distribution of stresses is altered and this can cause other blocs to fall. This additional information did not improve the performance of the model. DNN 3-classes actual 0 1 2 predicted 0 737 66 14 1 164 29 14 2 32 4 15 Figure 34 : confusion matrix for DNN model using 3 classes for the test data DNN 3-classes actual 0 1 predicted 0 737 80 1 196 62 Figure 35 : confusion matrix for DNN model using 3 classes adapted to 2 classes DNN 3-class DNN 2-class Operational accuracy 0.74 0.78 0.84 balanced accuracy 0.61 0.63 0.60 precision 0.24 0.28 0.35 recall 0.44 0.42 0.27 F1 0.31 0.33 0.30 Table 8: indices for the DNN 3-classes model adapted to 2-classes and the Operational model Convolution Neural Network This method is more adapted to this data set compared to the other neural networks as it was capable of detecting at least as many rock falls as the others while incorrectly predicting less events. Comparing it to the operational model it is safer with a higher recall and a similar, yet lower precision. This model is safer than the operational code, but it will dictate the road stays closed for a longer period of time. 43