Deep Learning for Recognizing Bat Species and Bat Behavior in Audio Recordings

Abstract

Monitoring and mitigating the continuous decline of biodiversity is a key global challenge to preserve the existential basis of human life. Bats as one of the most widespread species among terrestrial mammals are excellent indicators for biodiversity and hence for the health of an ecosystem. Typically, bats are monitored by analyzing ultrasonic sound recordings. Stateof-the-art deep learning approaches for automatic bat detection and bat species recognition commonly rely on audio spectrogram classification models based on fixed time segments, lacking exact call boundaries. While great progress has been made on bat species recognition using echolocation calls, little attention has been paid to bat behavior recognition that provides valuable additional information about bat populations. In this paper, we present a novel end-to-end approach for bat species recognition and bat behavior recognition based on a deep neural network for object detection. In contrast to state-of-the-art approaches, the presented model provides accurate call boundaries. It recognizes 19 bat species and distinguishes between three different behaviors: orientation (echolocation calls), hunting (feeding buzzes), and social behavior (social calls). Our experiments with two data sets show that our method clearly outperforms previous approaches for bat species recognition, achieving up to 86.2% mean average precision. It also performs very well for bat behavior recognition, reaching up to 98.4%, 98.3%, and 95.6% average precision for recognizing echolocation calls, feeding buzzes, and social calls, respectively.