A comparison of supervised learning techniques in the classification of bat echolocation calls

摘要: Abstract Today's acoustic monitoring devices are capable of recording and storing tremendous amounts data. Until recently, the classification animal vocalizations from field recordings has been relegated to qualitative approaches. For large-scale studies, approaches very time-consuming suffer bias subjectivity. Recent developments in supervised learning techniques can provide rapid, accurate, species-level bioacoustics We compared performances four (random forests, support vector machines, artificial neural networks, discriminant function analysis) for five different tasks using bat echolocation calls recorded by a popular frequency-division detector. found that all classifiers performed similarly terms overall accuracy with exception analysis, which had lowest average performance metrics. Random forests advantage high sensitivities, specificities, predictive powers across majority tasks, also provided metrics determining relative importance call features distinguishing between groups. Overall each task was slightly lower than reported accuracies time-expansion detectors. Myotis spp. were particularly difficult separate; best when members this genus combined genus-level analyzed separately at level species. Additionally, we identified ranked contributions predictor classifier measurements frequency, total duration, characteristic slope be most important contributors success. recommendations maximize efficiency analyzing data, suggest an application automated contribute wildlife efforts.