A new survey method using convolutional neural networks for automatic classification of bird calls

Habitat and reproduction surveys of raptors are common components of environmental impact assessments. Raptors are few in number and widely dispersed; therefore, raptor surveys require greater survey frequency compared to surveys for many other bird species, which, in turn, may result in extra costs as well as negative impact on raptors caused by observers. In this research, we propose a new, efficient, method for surveying raptors which is low-impact on the raptors themselves. The target species of this study was the Northern goshawk (Accipiter gentilis), categorized as ?Near Threatened? in the Japanese Red List, its population estimated at 5010?8950 in 2008. We developed a system which can automatically classify five classes of sounds, including goshawk calls, using a convolutional neural network. To establish the system?s applicability as a survey method, we then additionally verified three factors; (1) applicability of the method in different locales, (2) optimal distance from the nest to recording device, and (3) the ability to gauge reproductive states. We report an overall accuracy of 97.0% for this system. This system could classify, with high accuracy, goshawk calls ?kek-kek-kek? and ?whee-oo? across different locales. This system could classify sounds collected by recorders placed far from nests and within the forest, not placed on the nest itself. Some limitations of the system, notably limitation in data for verification, remain to be improved through further studies. This survey method can be used to judge whether an area was inhabited by goshawks or not, and their approximate reproductive state, based on a threehour recording of environmental sounds, with required fieldwork limited to the placing and collection of recorders, and minimal additional human input, due to the system classifying the sounds automatically.