Speaker
Description
In the context of the decline of the European common spadefoot toad (Pelobates fuscus), long term bioacoustic monitoring of this secretive species that vocalizes underwater is highly relevant. In this study, we present a software detector that features superior performance compared to literature [1]. For this study, hydrophone data was collected in Denmark and Poland with a different recorder than in [1]. Two candidate detectors were designed: a spectrogram-based EfficientNet-B0 classifier and a waveform-based CNN-BiGRU one. Both were trained on existing French hydrophone data and the new Danish Hydromoth data with call-aware losses and post-processing. The spectrogram-based model reached TPR 84.5% at FPR 0.11% with 94 % accuracy. The raw-waveform model reached TPR 75% at FPR 0.02% with 93 % accuracy. The new detectors were used to turn long recordings into phenologies.