The use of neural networks to classify echolocation calls of bats

The echolocation calls of bats contain substantial information about the animal producing the call. Previous studies have used discriminant function analysis (DFA) to classify calls to individual, to calling situation, to age, and to sex. DFA requires that the number of groups in the sample be known beforehand, and thus cannot be used to determine the group structure of novel sets of data. Echolocation calls (described by a variety of frequency, time, and amplitude variables) were given as inputs to artificial neural networks. The networks attempted to classify the calls without prior knowledge of group composition. A self-organizing map (SOM) neural network was used to estimate the number of bats that produced the set of calls on the basis of the variables describing the calls. Results suggest that this approach may yield a reasonable estimate of the number of bats in the sample even when a priori information is lacking. This may be useful for estimating the number of bats flying in an area based on recordings of their echolocation calls. Further analyses were conducted to determine the usefulness of using neural networks for other classifications. [Work supported by NIH.]