| Popis: |
Passive acoustic monitoring (PAM) is a tool with immense potential to evaluate the response of wildlife to ecosystem disturbances. PAM allows to evaluate wildlife dynamics by means of acoustic indices that estimate the diversity or complexity of sounds in a recording, as well as to study ecological aspects at the species level by training machine learning-based automatic acoustic detector. In this study, five deep learning approaches for automatic song detection were evaluated of the near-threatened Bachman's Sparrow in data scarcity scenarios, and then used this classifier to study the response of this bird to the number of days following a prescribed burn in six subtropical habitats in central Florida. At the same time, the response of avifauna acoustic activity to prescribed burning was quantified by means of three of the most used acoustic indices used in the literature (Acoustic Complexity Index, Acoustic Diversity Index and Bioacoustic Index). I found that it is possible to construct competitive birdsong detectors with small datasets using pre-trained models regardless. Furthermore, the use of data augmentation can lead to a detriment of the detector performance, especially of lower quality recordings, and that increasing the dataset does not necessarily increase the generalizability of the model. On the other hand, I found that unlike Acoustic Diversity Index, the Bioacoustic and Acoustic Complexity indices are negatively correlated with time after a burn, the same trend that Bachman's Sparrow presence showed, even though it was more influenced by habitat type than by the effect of the prescribed burns. This study shows the potential of tools including automatic song detection and acoustic indices to model at different scales the dynamics of avifauna in response to ecosystem disturbances. Their development can provide efficient tools for the study and conservation of both threatened wildlife species and the ecosystems they inhabit. |
