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

3.2. TESTED MODELS 29 3d DNN Series: 2 class FIXED SETTINGS learning rate : 0.001 batchsize : 64 epochs : 60 TESTED SETTINGS input : rain, rain and nb rockfall nb-lags :5, 10, 15 hidden : 32 or 64 weighting : True, False dropout : 0.0, 0.2 rescale : True, False 3-class DNN classifier You can see in Annex A all experiments results: • first series (108 trainings) with 3 classes and 5 days rain for inputs, show globally poor results but slightly better with learning rate=0.001 and batchsize=64, and when dropout is less than 0.4. • second series (46 trainings) start with previous settings (+learning rate=0.001 and batch- size=64), and test for lag of inputs. Globally, we still have poor result. When sample classes are not weighted , we get degenerate confusion matrix, class 1 or 2 is not predict. When we weight classes, class performance are much balanced with 73% accuracy : - precision by class [0.90, 0.14, 0.29] - recall by class [0.79, 0.29, 0.30] and confusion matrix: Predicted 0 1 2 0 737 164 32 GT 1 66 29 4 2 14 14 15 2-class DNN classifier To improve performance, we train a third series (with 80 trainings) of binary classifier with different inputs and other hyper-parameter. Results are quite better than with 3 classes, especially when rescale=True, weighting=True, inputs = (rain,nb), lag =(10,10), lr=0.001, batchsize=64. Two experiments show better results (see Annex A for logs), one with hidden=64, dropout=0 and the other with hidden=32, dropout=0.2. For last one, we get 72% of accuracy with 0.24 for precision, 0.51 for recall and confusion matrix: Predicted 0 1 0 710 228 GT 1 69 73