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

Chapter 2 Toward Formalisation We can study rockfall phenomenon at 3 different level : 1. at temporal level , where events are spatially aggregated. Available variable are daily times series of rain and of the number of rockfall. 2. at spatial level , where events are temporally aggregated. Available variable are the number of rockfall, by binned location, protection zone and geological zone. 3. at spatio-temporal level , with all information. Available variable are daily times series of rain, but also size of event, its location PR. This latter give extra information like geological zone and protection zone. As resolution is better, the size of dataset should be much more than here (less than 1000 rockfall), Remark. At temporal level, when we aggregate spatially event. We could ask if there is a sense for other variables to sum size, to average location ? Different point of view : Mostly depending of sampling time of events, we can consider different mathematical objects to model for this phenomenon if the amount of data allow it: 1. When events occurs at irregular times, we speak about point process (with only time event, or with location and exogenous variable and size as markers). For these process, many recent developments have been done with neural network . Counting process are naturally associated with point process and they permit to view point process at a regular sampling. 2. When events occurs at regular times, time series are the more natural way to see data, like popular ARMA models with Auto-Regressive and Moving Average part. These process are linear but we could add non-linearity like in Non-linear Auto-Regressive process (NAR) and exogenous variable (NARX process). Here, we present 2 types of machine learning model, which have been tested in this work: 17