Census and monitoring based on individually identifiable vocalizations: the role of neural networks

Vocal individuality is widely suggested as a method for identifying individuals within a population. But few studies have explored its performance in real or simulated conservation situations. Here we simulated the use of vocal individuality to monitor the calling corncrake (Crex crex), a secretive and endangered land rail. Our data set contained 600 calls from 30 individuals and was used to simulate a population of corncrakes being counted and monitored. We tested three different neural network models for their ability to discriminate between and to identify individuals. Neural networks are nonlinear classification tools widely applied to both biological and non-biological identification tasks. Backpropagation and probabilistic neural networks were used to simulate the reidentification of members of a known population (monitoring) and a Kohonen network was used to simulate the counting of a population of unknown size (census). We found that both backpropagation and probabilistic networks identified all individuals correctly all the time, irrespective of sample size. Kohonen networks were more variable in performance but estimated population size to within one individual of the actual size. Our results indicate that neural networks can be used effectively together with recordings of vocalizations in census and monitoring tasks.