Audio-Based Automatic Giant Panda Behavior Recognition Using Competitive Fusion Learning

Yuancheng Li, Yong Luo, Qijun Zhao, Mingchun Zhang, Yue Yang, Desheng LiSensors 2025 · DOI

TL;DR

This paper introduces a novel audio-based system for automatically recognizing giant panda behaviors, addressing the limitations of video-based methods. By developing a new dataset and a competitive fusion learning approach, they achieve high accuracy in distinguishing behaviors like eating, resting, and moving, paving the way for more efficient wildlife monitoring.

What Problem Does This Paper Address?

Monitoring giant pandas in the wild is crucial for their conservation, but traditional methods relying on expert manual analysis of video or audio data are time-consuming, labor-intensive, and prone to subjective bias. While video-based automatic behavior recognition has seen progress, audio-based methods have been largely overlooked. This is despite the potential for audio data, collected via collar-mounted devices, to offer a simpler and more continuous monitoring solution. However, audio data presents its own challenges: significant environmental noise, the complexity of panda vocalizations, and individual differences in sound production make accurate recognition difficult. Existing audio recognition methods struggle to effectively capture the nuanced bioacoustic features needed for reliable giant panda behavior recognition, especially for new individuals.

Methods

The core of this work is a novel audio-based giant panda behavior recognition (GPBR) method that leverages competitive fusion learning. The process begins with non-stationary noise reduction to clean the audio signal, followed by the extraction of Mel spectrograms to represent time-frequency characteristics.

The key innovation lies in the dual-branch neural network architecture built upon a ResNet50 backbone. This architecture is designed to extract richer, more complex acoustic features. During training, these two branches compete with each other, encouraging them to learn diverse feature representations and advanced combinations. This competition is guided by a custom loss function that includes:

  • Focal Loss: To focus on harder-to-classify samples.
  • Overlapped Activation Penalty (OAP): To minimize the overlap between the high-response regions of the two branches, forcing them to specialize.
  • Unbalanced Score Penalty (BSP): To ensure the branches produce balanced outputs and align as competition weakens.

Crucially, during inference, only one branch is needed, meaning there's no additional computational overhead compared to a single-branch model. The authors also explore the effect of a dynamic parameter ($\lambda$) that controls the strength of competition over training epochs, showing that a decaying $\lambda$ allows for convergence to similar outputs, enabling efficient single-branch inference.

Datasets

The paper introduces a new benchmark dataset called abPanda-5. This dataset comprises:

  • 18,930 audio samples
  • Collected from five individual adult female giant pandas
  • Covering five main behaviors: eating, resting, moving, nursing, and drinking.
  • Audio was recorded using Sony ICD-PX333M devices, stored as MP3 files with a 44,100 Hz sampling rate and 192 kbps compression.
  • The data was segmented into 1-minute clips and manually labeled.

Key Insights and Ideas

  • Audio as a Viable Alternative: The research strongly supports the feasibility of using audio data for giant panda behavior recognition, offering a complementary or alternative approach to video-based methods, especially in scenarios where video quality is compromised.
  • Competitive Fusion Learning for Nuance: The core idea of having two branches "compete" during training is a clever way to force the model to learn a more comprehensive set of features. By penalizing overlap, each branch is pushed to discover distinct acoustic cues, leading to a richer overall representation.
  • Inference Efficiency: The ability to use only one branch during inference without sacrificing performance is a significant practical advantage, making the method suitable for deployment on resource-constrained devices.
  • Noise Reduction is Critical: The ablation study highlights the substantial impact of noise reduction on performance, underscoring the importance of robust preprocessing for real-world bioacoustic data.
  • Nursing Misclassification Insight: The observation that nursing is often misclassified as resting, potentially due to the mother's quiet state and the blurring of cub vocalizations by MP3 compression, provides valuable context for future dataset improvements and model fine-tuning.

Key Results

  • The proposed audio-based method achieved an average accuracy of 92.98% ± 2.58% and an average F1-score of 92.99% ± 2.20% on the abPanda-5 dataset.
  • This performance outperformed several state-of-the-art deep learning models, including BiLSTM, MobileNetV3, ShuffleNetV2, VGG16, EfficientNetV2, ResNet50, ResNext50, and HTS-AT.
  • The method demonstrated robustness across different panda individuals, showing more stable classification performance compared to other methods.
  • Ablation studies confirmed the effectiveness of denoising and the competitive fusion learning components (Loap and Lbsp), with the full model achieving the best results.
  • Experiments with various CNN backbones showed that the proposed method consistently improved accuracy and F1-score by 1-2% on average, even with lightweight architectures.

Limitations and Future Work

  • Data Scarcity and Labeling: The authors acknowledge that the scarcity of labeled data, noise interference, and behavioral differences between adult and juvenile individuals remain challenges.
  • MP3 Compression Artifacts: The use of MP3 compression might blur certain vocalizations, potentially hindering accurate identification, particularly for subtle sounds like those of cubs.
  • Future Work: The authors plan to expand the dataset and refine the method to address these challenges, aiming for even better performance in audio-based GPBR. They also suggest that the approach is promising for deployment on mobile devices and edge computing, facilitating broader wildlife conservation efforts.

Summary updated

Abstract

Automated giant panda (<i>Ailuropoda melanoleuca</i>) behavior recognition (GPBR) systems are highly beneficial for efficiently monitoring giant pandas in wildlife conservation missions. While video-based behavior recognition attracts a lot of attention, few studies have focused on audio-based methods. In this paper, we propose the exploitation of the audio data recorded by collar-mounted devices on giant pandas for the purpose of GPBR. We construct a new benchmark audio dataset of giant pandas named abPanda-5 for GPBR, which consists of 18,930 samples from five giant panda individuals with five main behaviors. To fully explore the bioacoustic features, we propose an audio-based method for automatic GPBR using competitive fusion learning. The method improves behavior recognition accuracy and robustness, without additional computational overhead in the inference stage. Experiments performed on the abPanda-5 dataset demonstrate the feasibility and effectiveness of our proposed method.
BibTeX
@inproceedings{li2025audio,
  author = {Yuancheng Li and Yong Luo and Qijun Zhao and Mingchun Zhang and Yue Yang and Desheng Li},
  title = {{Audio-Based Automatic Giant Panda Behavior Recognition Using Competitive Fusion Learning}},
  booktitle = {Sensors},
  year = {2025},
  doi = {10.3390/s25133878},
}
Figure