Rockfall forecasting models along the roads of northern Gaspésie (Eastern Canada)

Rockfalls are major natural hazard for road users and infrastructures in northern Gaspésie (Eastern Canada) where nearly 25 kilometers of road runs along 10 to 100 m high flysch rockwall. The Ministère des Transports du Québec (MTQ) has recorded more than 17 500 rockfalls that have reached the roadway since 1987, which represents a nearly permanent danger for users. In the late 90s, protective berms were erected to reduce the number of rocks reaching the roadway. Despite the efficiency of these infrastructures, more than a hundred events are still recorded each year. Based on previous studies showing that rock instabilities in this type of geology is strongly correlated with meteorological events, we used different machine learning methods (logistic regression, classification tree, random forest, neural network) to design the best operational rockfall prediction model. Three event variables based on different rockfall frequency and magnitude thresholds were created. 94 weather variables were used to explain and predict events. Results show that 24h to 120h mean daily temperature above 0^oC and thawing degree-days are the most effective variables explaining the occurrence of winter and spring rockfall events. In summer, rainfall intensity is the most effective explanatory variable. The performance of the models has been optimized using a testing data set and then tested in an operational context using Environment Canada 24h and 48h HRDPS (high resolution deterministic prediction system) weather forecast model. A rockfall danger scale based on the probability of occurrence of medium and large magnitude rockfalls is proposed. These models can be used operationally as decision support tools to predict days with high event probability.