Data-driven sampling strategies for fine-tuning bird detection modelsa).

Area of Science:

• Ecology
• Bioacoustics
• Machine Learning

Background:

• Passive acoustic monitoring (PAM) is valuable for ecological data, especially bird population studies.
• Automated bird identification uses models like BirdNET, improved by fine-tuning with local acoustic data.
• Large PAM datasets require expert annotation, which is time-consuming and limits model improvement.

Purpose of the Study:

• To address the challenge of limited expert annotation in large-scale PAM datasets.
• To develop efficient strategies for selecting the most informative samples for model fine-tuning.
• To enhance the performance of automated bird identification models in specific acoustic environments.

Main Methods:

• A regularization technique was used to manage class imbalance during model fine-tuning.
• An "influence score" was developed to quantify individual training sample impact on model performance.
• A linear model estimated influence scores for generalization, and sampling strategies were compared using acoustic indices and model predictions.

Main Results:

• The proposed methods effectively identify informative samples for annotation.
• The influence score provides a data-driven approach to optimize sampling for model fine-tuning.
• Comparison of sampling strategies highlights efficient methods for large-scale PAM.

Conclusions:

• This work enables efficient annotation strategies for large-scale PAM, overcoming resource limitations.
• The developed methodology improves the cost-effectiveness of building accurate bird identification models.
• Optimized annotation strategies enhance the utility of PAM for ecological research and conservation.