I’ve been doing some work trying to understand landslide distribution in Saint Lucia. This is related to some potential work, but the guts of what I’m going to present come from personal research outside work. First I should acknowledge the Government of Saint Lucia for allowing use of some of their data to support this research.

This post is to discuss landslide susceptibility mapping. I will have some separate findings about population density and road density that I will try to use to extend the findings to a qualitative risk map. That won’t be ready for a while.

This image shows the distribution of past landslides, as inventoried by unknown mappers, and included in the Saint Lucia Landslide Response Plan dated 31 January 2008. It is believed these landslide points were derived from landslide inventories in the literature, and were prepared from air photo analysis.

An inventory like this serves as the basis for a statistical analysis of correlation between landslide presence/absence and other spatial features to develop a landslide susceptibility map. The basic data used in this work include high resolution topography, illustrated here:

and soil type:

The statistical analysis reveals correlations between different spatial data and landslide presence/absence. I’ve found that there is only a very weak connection between soil type and landslide occurrence, so soil type has been excluded from the model. With the high resolution topography, I looked at elevation, aspect (i.e. orientation of slopes with the compass), slope angle (i.e. steepness) and slope curvature (total curvature, plan curvature and profile curvature). The following images show the weight factors (with red being a higher weight, or stronger correlation with landslide presence) for elevation, aspect, curvature and slope angle:

I’ll not go into the details of how the weights are calculated, but I used a modified “weights of evidence” approach, which is a bi-variate statistical approach. If one assumes each of the different spatial themes to be independent (they are not), then one can add the individual weights together to come up with a combined weight value, and this gives an indication of landslide susceptibility. If one were so inclined, they could use multivariate methods to sieve out some of the effects of non-independence, but in my research into this topic I’ve found very little material difference in the maps resulting from multivariate methods and more simple (yet admittedly somewhat erroneous) bi-variate methods. At the end of the day, we are trying to build a multi-dimensional “best fit” curve, and the effort involved in the multi-variate refinements isn’t merited by the marginal improvements in model “accuracy.”

Anyway, here is the preliminary outcome, a draft landslide susceptibility map:

Now for reasons I won’t go into in detail yet, I’m not overly happy with the result. I expected the model to yield “better” predictions than what it does (i.e. do a better job of capturing all landslides within a relatively smaller proportion of the overall island). I think the problem is in the accuracy of spatial location implied in the landslide inventory. We’ve extracted the landslide locations from a relatively poor resolution map, and initial investigations suggest that most landslides are ~ 50 – 200 m away from actual mapped locations, possibly more. This means that the analysis has compared landslide presence with incorrect spatial data. We are working on trying to get the landslides better located, and will then re-do the analysis.

More to follow,…