GEOGRAPHIC INFORMATION-SYSTEMS AND ARTIFICIAL-INTELLIGENCE FOR PREDICTING THE PRESENCE OR ABSENCE OF MOUNTAIN REEDBUCK

Rule-based models were developed to predict the presence or absence of mountain reedbuck Redunca fulvorufula on a nature reserve in Karoo arid dwarf shrubland in South Africa. A geographic information system (GIS) was used to store, generate and retrieve model variables. Variables used as input to the models were mean, minimum and maximum slope, mean, minimum and maximum elevation, aspect, vegetation community, distance to drinking water, and terrain ruggedness as the number of contour lines crossing the grid. The models were constructed by inductive inference, using an artificial intelligence technique known as iterative dichotomizing. The technique used a complex sequence of dichotomous rules, similar to decision trees, to predict the presence or absence of mountain reedbuck at a site. The rules were converted to expert systems. The expert systems were validated by testing their ability to classify correctly independent data from the same study area. The classification success of the models ranged between 63 and 66%. In both instances this was significantly greater than 50%. The advantages and disadvantages of models created by inductive inference from GIS-based data are discussed. Iterative dichotomizing is a promising technique to develop models from existing data which are stored in information systems, provided that the appropriate data are available. Such data will vary according to the animal being studied and the scale of study.