Image: Earthquake-triggered landslide damage to a building in the Nepal-China border town of Kodari. Note how landslide debris has filled all three stories, and that a boulder has fallen through the roof and taken out all floors on the right hand side. Photo taken by Josh Jones, Oct 2019.
Landslides pose a major hazard to communities worldwide, particularly in dynamic mountainous regions. For example, in the Nepal Himalaya, over 74 people die on average per year due to monsoon-triggered landslides, whilst rarer, but more catastrophic earthquake-triggered landslides can also lead to hundreds of deaths and significant economic damage. This illustrates that there is a clear and pressing need to improve our ability to manage and mitigate landslide hazard, particularly in dynamic mountainous regions such as the Himalayas.
To better manage future landslide hazard, we fundamentally need to know where future landslides are likely to occur. This information is usually obtained from landslide susceptibility models, which typically use statistical techniques to forecast the likely locations of future landslides based on the distributions of past landslides.
These models are commonly used by both scientists and hazard managers to undertake disaster risk reduction, but there is a problem – these models work under the assumption that past landslide distributions will be the same as future ones. This assumption has historically been difficult to test, as without long-records of landslide data it is challenging to systematically quantify how actual landslide distributions change through time. However, thanks to the growing archive of satellite data, it is now possible to create long (> 30 year) datasets of landslide occurrences from which the temporal characteristics of landslide distributions can be assessed.
As such, one aspect of my PhD project was to do just this. During the first year of my PhD, I spent (literally countless) hours using Landsat imagery to map over 12,900 monsoon-triggered landslides in central-eastern Nepal across a 30-year period spanning 29 individual monsoon seasons between 1988 and 2018. I then used a number of statistical techniques to systematically assess how the distributions of these landslides changed through time with respect to variety of topographical, geological, and human landslide predisposing factors (e.g., elevation, slope angle, rock type, land use etc.).
Image: Landslide damage to road-infrastructure along the Araniko Highway, Nepal. Photos taken by Josh Jones, Oct 2019.
Interestingly, this showed that monsoon-triggered landslide distributions actually changed significantly through time, with particularly significant changes occurring in years impacted by either extreme outburst storms, floods, or earthquakes. Indeed, in 2015, monsoon-triggered landslides following the devastating Gorkha earthquake were found to occur higher up slopes and at greater slope angles than normal. This change is likely a result of a phenomenon called “earthquake preconditioning”, which describes how earthquakes cause temporary damage to parts of the landscape where shaking was most intense (which due to the behaviour of seismic waves is usually high on hill slopes near ridges), thus making subsequent landslides more likely in those locations.
But what are the implications of these changing landslide distributions? To answer this question, I used logistic regression techniques to develop landslide susceptibility models for 12 of the monsoon-seasons in my dataset, including all of the years impacted by outburst storms (1993, 2002), floods (2017) and earthquakes (2015). I then assessed how accurately each of the 12 models could forecast the landslide data from each of the other 11 modelled years. If the changing landslide distributions were having no affect on the resulting susceptibility models, then all 12 models would be expected to accurately forecast the landslide occurrences in all other years. However, this was very much not the case, with the susceptibility models developed for the years 1993, 2002, 2015, and 2017 being incapable of accurately and consistently forecasting the landslide occurrences in other years, and vice versa.
Image: Landslide damage to road-infrastructure along the Araniko Highway, Nepal. Photos taken by Josh Jones, Oct 2019.
Overall, this highlights that in dynamic mountainous regions with multiple landslide-affecting processes (e.g. monsoonal rain, outburst storms, floods, earthquakes), it is not appropriate to use typical susceptibility approaches that assume past landslide distributions reflect future landslide distributions. Instead, there is a need to move towards time-dependent modelling techniques that can account for transient variations in landslide occurrence. This should allow for more accurate forecasting of future landslide occurrences and thus lead to much-needed improvements in landslide hazard management and mitigation.
If you want to find out more, this work has recently been published in the Journal of Geophysical Research Earth Surface: https://doi.org/10.1029/2021JF006067
