Transient impacts of extreme events on landslide spatial distributions: implications for landslide susceptibility modelling

Abstract:

Landslide susceptibility models are a fundamental component of landslide risk assessments and early warning systems. One limitation of current statistical landslide susceptibility approaches is the implicit assumption that the spatial distributions of past landslides will be sufficiently similar to future landslides so as to facilitate accurate prediction. Until recently, this time-independent assumption was generally accepted to be valid. However, there are now several studies suggesting that rates of landslide occurrence can be transiently increased following extreme events. These studies raise the question of whether extreme events could also cause transient changes to the spatial distributions (rather than just rates) of subsequent landslides. This is an important as yet unanswered question that could further challenge static approaches to landslide susceptibility modelling.

Here, we use a new multitemporal landslide inventory to investigate whether the spatial distributions of monsoon-triggered landslides in Nepal vary through time in response to extreme events and, if so, to quantify the impact of this on resulting susceptibility model predictive power. We use a log-linear regression model implemented alongside a LASSO to develop 12 susceptibility models, each trained on a different year’s landslide data. We then use AUROC validation to assess how well each model predicted landslide distributions from other years. We find that following the 2015 Gorkha earthquake, excess topography and local relief became transiently more important in controlling monsoon-triggered landslide distributions, with coefficient magnitudes increasing by > 100% compared to other years. Furthermore, AUROC validation shows that the 2015-trained model performed worst in terms of predictive power, predicting only two other years with an AUROC > 0.75. Similarly, 2015 was the least predictable year, with no other models predicting it with an AUROC > 0.75 and only five predicting it with an AUROC > 0.70. These results show that extreme events can cause significant variations in landslide spatial distributions, to the extent that similarly triggered past landslide data cannot sufficiently model it. This highlights that static susceptibility models must be used carefully, particularly in tectonically active regions.