Automated surface feature selection using SALSA2D: An illustration using Elephant Mortality data in Etosha National Park

This analysis is motivated by the MIKE dataset in Etosha National Park (ENP). We use this dataset to show the development of an automated selection method for regression models to replace the model averaging used in the original CReSS paper. This method shows clear numerical and practical benefits over model averaging, and it's application to the elephant carcass data are of immediate and practical value to a range of stakeholders. We have developed SALSA 2D in a GLM/GAM regression framework but this paper shows the flexibility of this approach by applying it to presence only data and use a downweighted Poisson regression. Using SALSA2D for model selection provided a more realistic local/clustered intensity surface compared with the model average approach. The full analysis results showed high carcass intensity close to water holes and roads and in areas of the park with average rainfall. Some high risk areas were identified and these revelations are important for effective park management, particularly mitigation of poaching. It is impossible to patrol such a large area at random and these high intensity areas (particularly those accessed by a subset of roads and near some waterholes) can be targeted for more monitoring efforts than others.