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

3.2. TESTED MODELS 27 Predicted 0 1 2 0 300 50 0 GT 1 5 33 10 2 0 2 12 Predicted 0 1 Recall 0 TN FP TN TN + FP GT 1 FN TP TP TP + FN Precision TN TN + FN TP TP + FP Table 3.2: Confusion matrix for classification: an example for 3 classes on left, the general form for 2 classes (Positive/Negatives and True/False) on right. For 2 class version, row/column normalized versions of True Positives statistic, called Recall or Precision (of positive class) 3.2 Tested Models 3.2.1 Reference: Operating rules after year 2000 Some practical rules, based on rainfall statistics, have been set for RN1 Road. They are based on different threshold for rainfall quantities. Before major works , we can consider rules for road tipping ( Y t = 1 ) and road opened ( Y t = 0 ) on day t , defined in 2004 by, based on rain quantities R t : • if 15 ≤ R t < 30 , then a road tipping for the next day: Y t +1 = 1 • if R t ≥ 30 , then road tipping for 3 days: Y t +1 = 1 , Y t +2 = 1 and Y t +3 = 1 After major works in 2008, we used other rules defined by : • if 40 ≤ R t < 50 , then a road tipping for the next day: Y t +1 = 1 • if R t ≥ 50 , then road tipping for 2 days: Y t +1 = 1 and Y t +2 = 1 Open Close Recall Nothing 872 104 Rockfall 71 38 0.35 Precision 0.27 - Accuracy: 0.84 - Precision: 0.27 - Recall: 0.35 - Balanced accuracy: 0 . 62 = (0 . 89 + 0 . 35) / 2 Table 3.3: Operating Rules defined in 2004 applied on validation set (1085 days from 2005 to 2008)