Estimating risk of events using SOM models: A case study on invasive species establishment

The use of advanced modelling methods in ecology expands as ecological data accumulates and increases in complexity. Artificial neural networks (ANN), and in particular, the self-organising map (SOM), has become very popular for analysing particular kinds of ecological datasets. As SOM have become more utilised, it has become increasingly clear that the results of SOM models must be interpreted carefully. ve been used in a number of ecological studies to investigate the spatial distribution of species. When using presence–absence data of species distributions at given locations, the input vectors to a SOM are binary and the connection weights after learning are between 0 and 1. Using fuzzy set theory, we present an approach to the interpretation of these weights. Taking an example from invasive species research, we show that in the case of presence/absence data, a connection weight can be interpreted as a risk that an event will occur at a given location. was used to model the worldwide distribution insect pests to determine geographic patterns and define the species assemblages. The SOM weights were used as a measure of the risk of invasion for each species such that its potential to invade a geographic area could be evaluated. aper shows that while there are limitations concerning the interpretation of a model parameter, it is still possible to obtain relevant information when such limits are recognised. We emphasise however, that the interpretation of SOM weights must be appropriate to the particular study of interest.