AUTOMATIC IDENTIFICATION OF SPECIES FOR ADVANCED AND EFFICIENT RAPTOR SURVEYS: IMPROVING ACCURACY BY NEURAL NETWORKS AND NOISE REDUCTION

Abstract

 Surveys of raptor habitat are conducted in environmental impact assessments and natural environment surveys associated with development projects. In general, raptors are surveyed visually in the field to determine their habitat and nesting sites. However, because raptors are often alarmed when researchers enter the forest and because a large number of workers are required for surveys, efficient survey techniques must be developed. This study aimed to develop a technique for discriminating between five species of raptors (osprey, oriental honey buzzard, goshawk, grey-faced buzzard, and eastern buzzard) and the Japanese night heron by call. A sound analysis was conducted using deep neural networks after the noise reduction processing of sound data recorded near nesting forests of each species. As a result, a discrimination accuracy of more than 90% was achieved for five of the six species; however, there were also cases of misidentification resulting from environmental noise, indicating technical issues for further improvements in accuracy.