Martin, R.D., Brabyn, L. & Potter, M.A. (2010). Sensitivity of GIS-derived terrain variables at multiple scales for modelling stoat (Mustela erminea) activity. Applied Geography.
Permanent Research Commons link: http://hdl.handle.net/10289/4610
This research combines automated Geographical Information Systems (GIS) and iterative logistic regression scripting to data-mine for terrain variables that predict stoat (Mustela erminea) visitation to tracking-stations in a New Zealand indigenous forest and explore the impact of spatial analysis scale on modelling outcomes. Variables such as curvature and density of tracks are dependent on the scale of the analysis window used to compute the values. With automated GIS analysis it is possible to derive a large number of terrain parameters that vary in computational scale. It is common for analysts to make nominal choices regarding the size and type of analysis windows when calculating these derived variables, without testing the statistical validity of that choice. Stoats are a significant pest in New Zealand and threaten extinction for a number of vulnerable native species. Field data on stoat activity, based on footprint tracking tunnels, were used to develop an Akaike Information Criterion (AIC) optimised predictive model applying stepwise model selection. Once the optimal model was determined, the sensitivity of the model to different terrain parameters was tested by systematically substituting each variable and calculating the difference this made to the model equation. The most dominant terrain predictors influencing stoat visitation were proximity to tracks, altitude, northerly and easterly aspect, mean curvature, and topographical position and slope. Proximity to tracks and mean curvature were the most sensitive variables to analysis scale. This paper demonstrates the importance of considering scale in developing predictive models and the need to test many ecologically sensible analysis scales in order to find the best predictive variables. The paper concludes that GIS-based spatial data extraction, combined with automated statistical data mining methods, has an important role in developing accurate animal activity models.