Improving learning-based birdsong classification by utilizing combined audio augmentation strategies

Abstract

In ecology, changes in environmental conditions are often closely linked to shifts in species diversity. This relationship can be investigated by analyzing avian vocalizations, which are robust indicators of trends in biodiversity. Within this contribution, we explored various data augmentation techniques and deep learning strategies for the classification of birdsong within natural soundscapes. For this purpose, we employed three fundamental deep neural network architectures, such as vision transformers, to classify 397 different bird species. To improve both the accuracy and generalizability of our models, we incorporated up to 19 well-established data augmentation techniques commonly used in audio classification. This included an iterative selection process where only augmentations that enhanced classification performance were selected. The primary augmentation technique involved the integration of various noise samples and non-bird audio elements, which significantly improved model performance as assessed on the BirdCLEF 2021 data set. Individual augmentations achieved F1-scores from 48.0 % (vertical flip) to 72.6 % (primary background noise soundscapes). Through the strategic combination of key techniques – namely simulated pink noise, interspecies sound mixing, and loudness normalization – we achieved a top F1-score of 73.7%. Depending on the selected classification model, this corresponds to an improvement by 4.81 % to 10.5 %. Improvements and deteriorations of all applied augmentation techniques appeared to be robust across our three evaluated models. Therefore, our approach highlights the potential of sophisticated audio augmentations in refining the accuracy and robustness of birdsong classification models.

Topics & Concepts