Evaluating the performance of propagation models of flow-like landslides at regional scale

Abstract

AbstractPropagation models of flow-like landslides can be calibrated by comparing on-site evidence of past occurrences with the propagation paths and the deposition zones resulting from numerical simulations of the phenomena. Most typically, the performance of these models is evaluated considering the events independently from one another and, heuristically, i.e., subjectively assessing the fit between numerical results and available on-site data. At regional scale, however, storms often trigger, within a given area, multiple landslides of the flow type that occur more or less simultaneously. At this scale, a procedure that objectively quantifies the success, or the errors, of the numerical simulations of multiple landslides is lacking. In this study, such a quantitative calibration procedure is proposed, and assessed, considering the debris flows that occurred in Sarno in 1998 (Italy). The numerical model used is called Debris Flow Predictor (DFP), which is able to simulate the propagation paths and the accumulation depths of multiple debris flows, at regional scale, from a series of predefined triggering areas. The model employs a cellular automata method with a probabilistic behavioral rule, which is a function of the adopted digital elevation model and a series of parameters related to the erosional, the depositional, and the spreading capacity of the propagating soil mass. The numerical simulations were evaluated over the study area considering the entire set of debris flow events, as well as the individual debris flows, following a preliminary discretization of both the mapped footprints and the remaining portion of the territory. The relative and total operator characteristic curves, in addition to 6 indicators derived from a confusion matrix, have been used to quantify the performance of the simulations. The results show that the quantitative evaluation of the numerical results is essential to properly calibrate the adopted model, i.e., to discriminate among different simulations arising from different sets of model parameters.