Efficient Masked Autoencoder for Birdsong Representation with Applications on Wild Bird Species Classification

Birds play a critical role in maintaining ecological balance and serve as key indicators of biodiversity. Observing bird behavior in natural environments poses significant challenges. However, identifying bird songs through sensor technology provides a non-invasive and environmentally friendly method for monitoring avian diversity. Nevertheless, bird songs in natural environments are often obscured by substantial noise, and supervised learning-based recognition methods depend on extensive manual data annotation. To address these challenges, we propose Contrastive Residual Masked AutoEncoder-BirdNET (CResMAE-BirdNET), a specialized network for bird song recognition capable of autonomously extracting features from vast amounts of unlabeled acoustic data, thereby significantly enhancing recognition performance. First, to mitigate environmental noise and enhance model robustness, we apply four audio enhancement techniques and introduce a time-frequency self-calibration fusion module (TFSC) that integrates spectral ripple features. Next, CResMAE-BirdNET combines contrastive learning with a masked autoencoder framework, integrating residual attention in the encoder and a residual multi-layer perceptron in the decoder, enhancing the ability to capture the relationship between local and global features for superior feature representation. Finally, extensive experiments on our self-built 40-class dataset (Bird40Song) and the public dataset (Birdsdata) validate the effectiveness of the proposed method, achieving recognition accuracies of 99.35% and 98.43%, along with F1-scores of 99.34% and 98.28%, respectively. The results highlight significant advancements in bird song recognition, demonstrating the potential of CResMAE-BirdNET to support large-scale ecological monitoring and biodiversity research. Code available at: https://github.com/xzq-okkkkkkk/CResMAE-BirdNET.