Abstract
Bioacoustic monitoring is a critical tool for wildlife conservation, yet the manual annotation of audio data remains a significant bottleneck. This paper proposes a fully automated pipeline for bird species classification that integrates unsupervised image-based segmentation with Multiple Instance Learning (MIL). By treating audio spectrograms as images and recordings as "bags" of sound "instances," the framework eliminates the need for precise, manual timestamps. Experimental results on a diverse 13-species dataset show that unsupervised segmentation performs on par with labor-intensive supervised methods, offering a scalable solution for biodiversity monitoring.
1. Introduction
Monitoring wildlife through sound recordings provides valuable insights into ecosystem health. However, traditional classification methods typically require either exhaustive manual labeling of individual syllables or struggle with the noise inherent in recording-level analysis. This study aims to bridge this gap by developing a system that automatically extracts relevant sound units from recordings and identifies species using a Multiple Instance Learning approach, significantly reducing the human effort required for data preparation.
2. Methodology
The proposed framework consists of a multi-stage pipeline designed to process raw audio into species classifications without manual intervention at the segmentation level.
2.1 Unsupervised Segmentation
Instead of traditional signal processing, the system treats the spectrogram as a 2D image. The process involves:
- Denoising: A 2D Wiener filter is applied to smooth the spectrogram.
- Morphological Reconstruction: Image processing techniques isolate high-intensity regions, effectively suppressing background noise.
- Automated Thresholding: Otsu’s method is used to binarize the image, separating birdsong signals from the background.
- Instance Extraction: Connected components of signal pixels are extracted as individual "instances" or syllables.
2.2 Feature Extraction
Four distinct feature sets were evaluated for the extracted instances:
- Mask Descriptors (MD): Shape-based features such as area and perimeter.
- Profile Statistics (PS): Statistical properties of sound intensity.
- Histogram of Gradients (HOG): Texture-based features capturing local shape information.
- All-Features (AF): A combination of the above descriptors.
2.3 Classification via Multiple Instance Learning (MIL)
The system employs the MILES (Multiple Instance Learning via Embedded Instance Selection) algorithm. In this paradigm, a 10-second recording is treated as a "bag," and the segmented syllables are its "instances." The classifier learns to identify a species if at least one instance in the bag belongs to that species, providing robustness against noise and overlapping sounds.
3. Experimental Setup
The study utilized a public dataset (Briggs et al., 2012) containing 548 10-second recordings. The dataset includes 13 bird species and is characterized by significant environmental noise and simultaneous vocalizations from multiple species.
4. Results and Discussion
The performance was evaluated using the F-score. The best average F-score reached 0.821 with supervised segmentation and 0.812 with the unsupervised approach using the "All-Features" set.
Key findings include:
- Statistical Parity: Paired t-tests confirmed no significant difference between unsupervised and supervised approaches for PS, HOG, and AF feature sets.
- Consistency in Shape: For Mask Descriptors, the unsupervised method actually outperformed the supervised baseline, likely due to the higher consistency of automated boundary detection compared to human annotation.
- Robustness: The MIL framework successfully ignored "noise instances" (e.g., wind) within a bag, focusing on the relevant birdsong instances.
5. Conclusion
This research demonstrates that unsupervised segmentation is a highly effective alternative to manual annotation for birdsong classification. By leveraging MIL, the system achieves state-of-the-art performance while drastically reducing the time and expertise required for data labeling. Future work will focus on optimizing texture-based features like HOG and improving performance for species with limited training data.
Summary updated