Investigating bird call identification uncertainty using data from processed audio recordings

To effectively monitor bird populations, accurate identification of species is critical. However, the reliability of species identification is rarely taken into account or quantified. For this study, bird call data was collected using automated acoustic recording devices (ARDs) over a 3-year period. We then compared the results from experienced ornithologists who independently identified bird calls from the same samples. Results were highly variable. The level of agreement between processors on identification for some species was high (e.g. tomtit Petroica macrocephala, 85.1%), whilst for others it was considerably lower (e.g. song thrush Turdus philomelos, 23.5%). There was no statistically significant difference in agreement between native and non-native species. However, there was some evidence for improvement in agreement for the third survey season, when compared to the first. In a more specific comparison of bellbird Anthornis melanura and tui Prosthemadera novaeseelandiae calls, our results showed that these two species were frequently confused. There were many instances where only one of the processors identified a species. Possible explanations for why calls were missed include differences in hearing ability and levels of concentration between processors, whilst false positives could have resulted from confirmation bias. These results have implications not only for data collected using recording devices but also field-based counts of birds conducted by observers.