Multi-label bird sound recognition based on multi-view learning and dynamic threshold adjustment

Bird species monitoring is crucial for conservation, but overlapping vocalizations in natural environments complicate multi-label classification, affecting model performance. To address these issues, we developed the Adaptive Multi-label Attention Threshold Network (AMAT-Net) as a bird sound classification framework. AMAT-Net employs a multi-view strategy, combining bidirectional gated recurrent unit (BiGRU)-attention networks to analyze temporal features and multi-scale convolutional neural networks to extract spectral features, enabling analysis of bird sounds. Given the differences between temporal and spectral features, time-domain features capture transient changes, whereas frequency-domain features reveal spectral trends. Balancing the essential features of both without losing details is difficult. Therefore, we designed the temporal–spectral attention feature fusion (TSAFF) module to optimize feature fusion. TSAFF employs an attention-based mechanism to fuse temporal and spectral features, enhancing the cross-domain feature complementarity. Binary classification is conducted between relevant and irrelevant labels, and threshold is determined based on the results. A score-based thresholding strategy called dynamic threshold scaling was then developed. A label correlation matrix is constructed using Pearson's correlation coefficients, and the classifier's scores for instance-label pairs with high inter-label correlations are adjusted accordingly during prediction. In addition, hierarchical cross-validation is used to search for the threshold that maximizes the F1 score, dynamically optimizing the decision boundary for each species to adapt to the actual label distribution. Experimental results on a synthesized dataset of 10 bird species (including cases of 2, 3, and 4 species vocalizing simultaneously) and the public BirdCLEF+2025 dataset demonstrate that AMAT-Net achieves an accuracy of 95.54% with a macro-F1 score of 91.26% on the synthesized dataset, and an accuracy of 98.75% with a macro-F1 score of 93.14% on BirdCLEF+2025.