Human vs. machine : identification of bat species from their echolocation calls by humans and by artificial neural networks

摘要: Automated remote ultrasound detectors allow large amounts of data on bat presence and activity to be collected. Processing such involves identifying species from their echolocation calls. identification has the potential provide more consistent, predictable, potentially higher levels accuracy than by hu- mans. In contrast, humans permits flexibility intelligence in identification, as well incorpora- tion features patterns that may difficult quantify. We compared with artificial neural networks (ANNs) ability classify short recordings calls variable signal noise ratios; these sequen- ces are typical those obtained automated recording systems often used large-scale ecological studies. presented 45 (1-4 calls) produced known bats ANNs 26 human participants 1 month 23 years experience acoustic bats. Humans correctly classified 86% genus 56% species; identified 92% 62%, respectively. There was no significant difference between performance humans, but performed better about 75% humans. little relationship classification rate. However, <1 year worse others. Currently, is suitable for research, after careful consideration biases. improvements they trained future increase beyond demonstrated Resume´ : Les detecteurs automatisesadistance d'ultrasons permettent de recolter des quantites importantes donnees sur la et l'activitedes chauves-souris. Pour traiter donnees, il est necessaire d'identifier les especes chauves-souris d'apres leurs cris d'echolocation. L'identification automatisee pourrait potentiellement per- mettre une precision plus uniforme, previsible probablement fine que l'identification faite par humains. En revanche, humains permet flexibiliteet du jugement dans le processus reconnaissance, ainsi l'incorporation caracteristiques patrons qui peuvent etre difficiles aquantifier. Nous comparons ca- pacites d'humains reseaux neurones artificiels a classifier courts enregistrements d'appels d'echolo- cation contenant divers rapports signal:bruit; ce sont sequences representatives celles obtenues systemes automatises d'enregistrement distance couramment utilises etudes ecologiques agrande echelle. avons presente ´ (de ` 4 appels) produits connues collaborateurs possedent entre mois ans d'experience acoustique ont classecorrectement 86 % au genre 56 l'espece; reussites correspondantes 92 62 %. Il n'y pas significative performances ANNs, mais donnent un meilleur rendement 75 y peu cor- relation l'experience leur taux classifications reussies. Neanmoins, ayant an reussite inferieur aux autres. Dans conditions actuelles, appels d'echolocation convient ecologiques, si l'on tient compte soigneuse- ment sources d'erreur. Cependant, amelioration qu'ils utilisent l'avenir accrooˆtre au-delade celle demontree (Traduit Redaction)