Avian vocalizations in Huangmaohai sea-crossing channel: Automatic birdsong recognition and ecological impact analysis based on deep learning

Abstract

Avian vocalization monitoring provides fundamental data for ecological environment monitoring and evaluation, effectively promoting the conservation of avian species and ecosystems. To assess the impact of the construction of the Huangmaohai Sea-crossing Channel on biodiversity conservation and the ecological environment, extensive passive acoustic monitoring was conducted in the area from June 2023 to February 2024 during the construction period. An automatic birdsong recognition framework using deep learning was developed to process the vast amount of recorded acoustic data with an accuracy of 75.26 % in complex real-world environments. A birdsong event detection model converted the recordings into valid data regions, followed by recognition using the ECAPA-TDNN model, which achieved the best classification results compared to other baseline models. Results revealed that the diurnal and nocturnal activities of birds in this area exhibit a common bimodal pattern. There were no significant differences in the species and vocalization counts of birds between the construction and non-construction areas, indicating that the cross-sea channel area maintains a good avian diversity. The study also found that Light-vented Bulbul, Yellow-bellied Prinia, and Common Greenshank significantly increased their minimum vocalization frequency in noisy environments, which may be a result of behavioral plasticity. The automated birdsong recognition framework developed in this study can effectively be utilized to assess bird distribution and abundance, spatiotemporal characteristics of birdsong diversity, and study the adaptive adjustments of birdsong to noise environments, thereby contributing to the recording of acoustic information of birds in the area and biodiversity conservation.