Progress in Predicting Rock-Slope Failures

Abstract:

Recent research on predicting landslides has seen a massive increase in statistical and computational methods that are largely adapted from the fields of machine learning and data mining. Judging from a sample of some 150 recent scientific papers, the gross majority of the reported success rates of these statistical methods are overwhelmingly high and promising at between 71% and 98%. Perhaps surprisingly, though, the death toll and damage from landslides has remained elevated in the early 21st century, so that reliably predicting the occurrence of rock-slope failures without overfitting our models remains challenging. Here I review some of the recent advances in this field, and show how novel results from landslide seismology and landslide sedimentology have promoted our ability of detecting large rock-slope failures in mountainous terrain. Several new detailed investigations of the internal nature of large rockslide deposits, for example, help to reduce the confusion potential with macroscopically similar moraine debris, or microscopically similar fault breccia. I further outline some of the limitations of empirical models that use rainfall intensity-duration thresholds for landslide early warning, and of multivariate methods concerned with mapping landslide susceptibility at the regional scale. I conclude by discussing the occurrence of ‘black swans’ such as long-runout rock-ice avalanches in size distributions of rock-slope failures, and their implications for quantitative hazard appraisals.