This post will present some new work that extends previous work documented in a couple of papers in the Canadian Geotechnical Journal, here: landslide susceptibility in eastern Canada (paper #1), and here: landslide hazard and risk in eastern Canada (paper #2). The intention will be to wrap these new findings into a new journal paper for submission shortly.
With further support from Transport Canada, CN and CPR, I’ve made some refinements to the landslide susceptibility model presented in paper #1 listed above, and I will discuss some of the changes here. Independent from that work, I’ve also done some further work on risk, extending some of the work I presented earlier here:road density versus population density, here: scale dependence, and here: St Lucia road vs pop density.
The grey area is the rough extent of marine invasion following the most recent glacial retreat, and therefore corresponds to the approximate extent of soils deposited in a marine environment, which thus limits the extent of particularly sensitive silt/clay soils, which are the subject of this work. The red boundary reflects the outline of watershed considered in the hydrological analysis. The black rectangle is the area where I have digital landslide data, so the statistical analysis for the whole model are based on spatial correlations developed within that selected part of the study area.
The landslide susceptibility model shown in paper #1 had several limitations we aimed to eliminate through further work. This included, generally:
– model only covered western half of affected area
– model did not consider several types of spatial data that might influence susceptibility, including earthquakes, stream gradient, stream tortuosity, river bank height, regional relief, etc
– model would benefit from expansion of the landslide inventory that serves as the basis for inferred spatial correlations, as it covers only part of the study area (NTS 31H) which does not include all geomorphological characteristics of the complete study area
I have not extended the landslide inventory. We have looked at several additional sets of spatial data. The original model was based on a bivariate statistical analysis of the following factors:
– soil type
– ground elevation
– land use
– overburden thickness, and
– flow accumulation in the stream networks.
Flow accumulation was by far the most important factor in the model. Of note, slope angle is irrelevant, in part because it cannot be properly related to existing landslides given morphology and data resolution, but also because it isn’t particularly important for these landslides, being far less important than riverbank height and brittleness of the involved materials.
The new model considered most of the same factors plus some new ones. The following factors were carefully examined, but found to have weak or no relationships with landslide occurrence:
– land use – while this factor has a weak correlation with landslide occurrence, its inclusion in the overall model neither improved not weakened the overall model, therefore it was excluded
– LANDSAT satellite imagery – several different combinations of the different LANDSAT bands were considered, but no useful relationships were found that merited the effort to obtain and manipulate the data
– bedrock geology – available mapping have limited resolution, and very little variability across the area of landslide inventory, hence no relationships could be inferred
– stream gradient – this factor is expected to have a good correlation with landslide incidence when considered in combination with streamflow (Bjerrum et al, 1969 did some excellent work relating landslide occurrence to the potential for active erosion in relation to the state of development of the stream), but unfortunately the stream data do not contain gradient information at a meaningful scale
– seismicity – many of the largest landslides in the study area are known to have been triggered by large earthquakes. I would suggest than perhaps 5 % of the naturally triggered landslides, being most of the largest, were triggered by earthquakes. By contrast, perhaps 95 % were triggered by stream erosion (or related stream dynamics). Unfortunately, it’s not possible to know which landslides in the digital inventory were triggered by earthquakes, and furthermore there is very little variation in seismic hazard across the landslide inventory area, so meaningful relationships would likely not emerge even if we could categorize triggering mechanisms. To illustrate the concern, this image shows historical seismicity (black dots) and inferred seismic hazard for the overall study area:
The digital landslide inventory covers the rectangle in the middle of the study area, which has little variability in seismic hazard relative to the overall study area.
The refined landslide susceptibility model considers the following factors:
– stream flow accumulation
– surficial geology:
– overburden thickness:
– ground elevation:
– stream sinuosity:
– bank height (local relief):
– slope angle:
Interestingly, slope angle correlates negatively with landslide occurrence, unlike most other cases where the landslide susceptibility model is largely driven by the positive correlation between slope angle and landslide occurrence (see even for example my post about landslide susceptibility for St Lucia, here: St Lucia landslide susceptibility).
I’ll not go into the details here, but the new model has significantly better predictive power than the previous version published in paper #1 above, giving me better confidence in its use. Note that it is intended to be used at an appropriate scale, with appropriate consideration of the associated uncertainties. It is not intended to be used for site-specific purposes, but rather only to aid in region-wide planning and analysis.
In my next post I will talk a bit about extending this into qualitative risk.