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

3.3. CONCLUSION 33 3.3 Conclusion Here, we sum up these preliminary studies to build predictive models for rockfall hazard based on some dependencies like rains quantities. These studies focuses at temporal level without taking into account spatial information (location PR, geological and topological factors ...). The objective was to investigate different models, their limitations and their performances. Here, we have simplified the initial problem by considering classifying target in 2 or 3 groups. On Table 3.4, we sum up best performances models for each type of relevant models: • Operational rules standing for baseline is a binary classifier that have a good accuracy (84%) but with less True Positive (TP is rockfall correctly detected) than other models. As classes are unbalanced, it predicts mostly the frequent one (no rockfall). Accuracy is not relevant metric in that case. • 3-classes problem is much more expressive but in practice, even if classes are weighted, it’s difficult to separate positive classes "1" and "2" while keeping good performances. • 2-classes models have more good predictions (TP) but also over-detection (False positives FP) than Operational Rules. In Table 3.4, we point out the two best 2-classes DNN models, one with only rain as input and other one with rain/number of rockfall. Latter one gives the best recall and TP from all models (54% and 77 TP) but also more over-detections (237 FP) • 2-classes DNN models shows a good compromise between good-detections (62 TP) and over- detections (133 FP) and so the best balanced accuracy (65%). Globally, these Neural Networks have slightly better performance than Operational Rules. Nonetheless, all this performances should be seen from an operational point of view because over- detections (FP : predicting rockfall while there is nothing) and miss-detections (FN: predicting nothing while there is a rockfall) have not the same consequences. After the potential perspectives written in previous chapter conclusion, one could add: • rockfall hazards seem to be a mix of "natural rockfall event" hardly predictable and "forced rockfall event" caused by exogenous factors (like rain, temperature, thaw-thin cycle..). If size of events could discriminate them (thresholding size), performances would be better. • dynamic of rockfall is very different in low or high activity period. Here, we consider global model but it could better to investigate switching models (from dry to rainy periods) with their owns parameters.