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

3.2. TESTED MODELS 31 CNN Series: 2 class FIXED SETTINGS input : rain learning rate : 0.001 batchsize : 64 epochs : 60 hidden : 64 nb filters : 64 (100 for one) TESTED SETTINGS nb-lags : 10,15 kernel size : 3, 5, 9, 13 weighting : True, False dropout : 0.0, 0.2, 0.4 rescale : True, False Over 11 tested networks, best results are obtained with 100 convolutional filters of 3 days size. Balanced accuray is 0.65, recall at 0.44 and precision 0.32. Compare to DNN, with quite same level of recall, they have less over-detection (so better precision) and confusion matrix: Predicted 0 1 0 805 133 GT 1 80 62 3.2.4 Regression Models: Generalised Linear Model (GLM) We test some GLM with either Poisson Distribution or Binomial (with normalising target Y by their maximum values). All this fitted models have difficulties to model seasonal activities: bursting period followed by calm one. For example, you can see predictions on train dataset in Figure 3.2 for the estimated model (rain lagged to 5 days) below: ============================================================================== Dep. Variable: nb No. Observations: 1512 Model: GLM Df Residuals: 1506 Model Family: Binomial Df Model: 5 Link Function: logit Scale: 1.0000 Method: IRLS Log-Likelihood: -92.975 Date: Mon, 08 Nov 2021 Deviance: 83.445 Time: 16:51:28 Pearson chi2: 149. No. Iterations: 7 Covariance Type: nonrobust ============================================================================== coef std err z P>|z| [0.025 0.975] ------------------------------------------------------------------------------ rain1 0.0183 0.005 3.420 0.001 0.008 0.029 rain2 0.0028 0.006 0.515 0.606 -0.008 0.014 rain3 0.0149 0.005 2.759 0.006 0.004 0.025 rain4 -0.0046 0.014 -0.335 0.738 -0.031 0.022