Snow melt modelling for improving the forecast of rainfall threshold-based landslide triggering

During recent years in the Northern Apennines (central Italy) a relevant number of landslides were triggered by rapid snow melt. Hence, the need to integrate snow precipitation and melting processes within existing statistical models for landslide prediction. For this purpose, a dynamical model is proposed which takes into account the buildup and melting of the snowpack in time. This model is based on two equations: the conservation of mass (input-output balance) and an empirical equation for modeling the snow density variation. From the conservation of mass, a differential equation of snowpack depth can be obtained and it is discretized, while the second equation is an empirical discretized function for the density variation in time. The model is then divided in two modules depending on whether a thresholds temperature is exceeded or not. The first module accounts for the accumulation of solid rainfall in the snowpack and the second one for the snow melting. The main originality of the model is the use of an empirical functional based on chemical kinetics depending on air temperature, rainfall amount and depth and density of the snowpack, while other factors like wind, air humidity, atmospheric pressure and radiation are not considered since not available in our case study. In the present form, the model depends on 13 empirical parameters including a threshold temperature between snowfall and rainfall and the density of newly fallen snow. To assess the optimum values of the empirical parameters, we used an heuristic optimization algorithm (optimized flexible modified simplex) to minimize the errors between outputs of the model and experimental measures retrieved from a network of sensors located in the study area. The study of static and dynamic sensitivity analysis shows a good robustness of the model. In the study area this work was coupled to a statistical models for landslide prediction based on rainfall thresholds and improvements were achieved: several landslides (36 in four years), triggered by snow melting (and thus missed by the standard rainfall based warning system), were correctly detected. In addition, false alarms were reduced due to re-distribution of water input in the ground.